Geodesia

Geosystems Mathematics
Series editors
W. Freeden
Kaiserslautern, Germany
M.Z. Nashed
Orlando, Florida, USA
This series provides an ideal frame and forum for the publication of mathematical
key technologies and their applications to geo-scientific and geo-related problems.
Current understanding of the highly complex system Earth with its interwoven
subsystems and interacting physical, chemical, and biological processes is not
only driven by scientific interest but also by the growing public concern about
the future of our planet, its climate, its environment, and its resources. In this
situation mathematics provides concepts, tools, methodology, and structures to
characterize, model, and analyze this complexity at various scales. Modern high
speed computers are increasingly entering all geo-disciplines. Terrestrial, airborne
as well as spaceborne data of higher and higher quality become available. This
fact has not only influenced the research in geosciences and geophysics, but also
increased relevant mathematical approaches decisively as the quality of available
data was improved.
Geosystems Mathematics showcases important contributions and helps to
promote the collaboration between mathematics and geo-disciplines. The closely
connected series Lecture Notes in Geosystems Mathemactics and Computing offers
the opportunity to publish small books featuring concise summaries of cutting-edge
research, new developments, emerging topics, and practical applications. Also PhD
theses may be evaluated, provided that they represent a significant and original
scientific advance.
Edited by
• Willi Freeden (University of Kaiserslautern, Germany)
• M. Zuhair Nashed (University of Central Florida, Orlando, USA)
In association with
• Hans-Peter Bunge (Munich University, Germany)
• Roussos G. Dimitrakopoulos (McGill University, Montreal, Canada)
• Yalchin Efendiev (Texas A&M University, College Station, TX, USA)
• Andrew Fowler (University of Limerick, Ireland & University of Oxford, UK)
• Bulent Karasozen (Middle East Technical University, Ankara, Turkey)
• Jürgen Kusche (University of Bonn, Germany)
• Liqiu Meng (Technical University Munich, Germany)
• Volker Michel (University of Siegen, Germany)
• Nils Olsen (Technical University of Denmark, Kongens Lyngby, Denmark)
• Helmut Schaeben (Technical University Bergakademie Freiberg, Germany)
• Otmar Scherzer (University of Vienna, Austria)
• Frederik J. Simons (Princeton University, NJ, USA)
• Thomas Sonar (Technical University of Braunschweig, Germany)
• Peter J.G. Teunissen, Delft University of Technology, The Netherlands and
Curtin University of Technology, Perth, Australia)
• Johannes Wicht (Max Planck Institute for Solar System Research, Göttingen,
Germany).
More information about this series at http://www.springer.com/series/13389

Willi Freeden • M. Zuhair Nashed
Editors
Handbook of Mathematical
Geodesy
Functional Analytic and Potential
Theoretic Methods
Editors
Willi Freeden M. Zuhair Nashed
Geomathematics Group Department of Mathematics
TU Kaiserslautern University of Central Florida
Kaiserslautern, Germany Orlando, FL, USA
ISSN 2510-1544 ISSN 2510-1552 (electronic)

Geosystems Mathematics
ISBN 978-3-319-57179-9 ISBN 978-3-319-57181-2 (eBook)
https://doi.org/10.1007/978-3-319-57181-2
Library of Congress Control Number: 2018940865
Mathematics Subject Classification (2010): 86A30, 86A20, 31A25, 47A52, 65A20, 60J45
© Springer International Publishing AG, part of Springer Nature 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Cover illustration: designed by Prof. Dr. Schreiner, Buchs
Printed on acid-free paper
This book is published under the imprint Birkhäuser, www.birkhauser-science.com by the registered company
Springer International Publishing AG part of Springer Nature.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Handbook of Mathematical Geodesy
Geosystems Mathematics, v–vi

c Springer International Publishing AG, part of Springer Nature 2018
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
W. Freeden
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
W. Freeden, T. Sonar, and B. Witte

Gauss as Scientific Mediator Between Mathematics and Geodesy
from the Past to the Present . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
M. Augustin, S. Eberle, and M. Grothaus

An Overview on Tools from Functional Analysis . . . . . . . . . . . . . . . . . . . . 165
W. Freeden and M.Z. Nashed

Ill-Posed Problems: Operator Methodologies of Resolution
and Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
W. Freeden and H. Nutz

Geodetic Observables and Their Mathematical Treatment
in Multiscale Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
F. Sansò
The Analysis of the Geodetic Boundary Value Problem:
State and Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
M. Grothaus and T. Raskop

Oblique Stochastic Boundary Value Problem . . . . . . . . . . . . . . . . . . . . . . . . 491
M. Augustin, W. Freeden, and H. Nutz

About the Importance of the Runge–Walsh Concept
for Gravitational Field Determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
W. Freeden, H. Nutz, and M. Schreiner

Geomathematical Advances in Satellite Gravity
Gradiometry (SGG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
M. Gutting
Parameter Choices for Fast Harmonic Spline Approximation . . . . . . . . 605
vi Contents
W. Freeden and M. Zuhair Nashed

Inverse Gravimetry as an Ill-Posed Problem in
Mathematical Geodesy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
C. Blick, W. Freeden, and H. Nutz

Gravimetry and Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
W. Freeden, V. Michel, and F.J. Simons

Spherical Harmonics Based Special Function Systems and
Constructive Approximation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753
C. Gerhards
Spherical Potential Theory: Tools and Applications . . . . . . . . . . . . . . . . . 821
C. Gerhards, S. Pereverzyev Jr., and P. Tkachenko

Joint Inversion of Multiple Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855
S. Leweke, V. Michel, and R. Telschow

On the Non-Uniqueness of Gravitational and Magnetic
Field Data Inversion (Survey Article) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 883
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 921
Geosystems Mathematics, vii–viii

Preface
Geodesy, as most other disciplines, spans activities ranging from theoretical to

applied border lines. In the twenty-first century, geodesy is strongly influenced by
two scenarios: First, the technological progress, in particular, space observation
has opened fundamentally new methods of measurements. Second, high speed
computers have led to a strong “mathematization”. As a consequence, geodesy
is in great shape. However, the width and depth of new geodetic challenges will
simultaneously require basic analysis and understanding of all technologically as
well as mathematically driven components. These requirements are inextricably
necessary to provide future improvements in diverse fields of geodetically involved
public concern for our planet such as climate environment, expected shortage of
natural resources, etc.
This “Handbook of Mathematical Geodesy” deals with mathematics as the
key technology for modeling purposes and analysis of today’s geodetic measure-
ments and observations. It supplies deep modern and cutting-edge mathematical
knowledge as transfer methodology from the reality space of measurements to the
model space of mathematical structures and solutions, and vice versa. Essential
interest is laid in studying the gravitational field usually in macroscopic sense,
where the quantum behavior of gravitation may not be taken in account. More-
over, in geodetically reflected Earth’s gravity work, velocities that are encountered
are considerably smaller than the speed of the light. As a consequence, Newtonian
physics can be safely used.
In detail, this Handbook is concerned with the following selection of topical
areas:
• functional analysis and geodetic functional models
• constructive polynomial, spline and wavelet approximations
• mathematical treatment of geodetic observables and multiscale integrated
concepts
• geodetic boundary value problems and oblique stochastic derivative problems
• Runge–Walsh mono-and multi-pole expansions on geodetic reference surfaces
such as sphere, ellipsoid, telluroid, geoid, real Earth’s surface
• regularization methods of ill-posed and inverse problems
• gravimetric and gradiometric (multiscale) modelling.
The objective of the handbook is twofold: on the one hand it serves as a
self-consistent collection of newsworthy material at the graduate-student level for
viii Preface
all members of the mathematical community interested in any of the diverse prob-
lems relevant in today’s geodesy. On the other hand, the book represents a valuable
reference for all geodesists facing innovative modeling supplies involving recently
measured datasets in their professional tasks. For both groups the Handbook pro-
vides important perspectives and challenges in crossing the traditional frontiers.
The Handbook consolidates the current knowledge by providing summaries
and concepts as a guide for geodetic transfer from reality space (“measurements”)
to virtuality space (“models”). All in all, the work is an authoritative forum of-
fering appropriate mathematical means of assimilating, assessing, and reducing to
comprehensible form the flow of measured data and providing the methological
basis for scientific interpretation, classification testing of concepts, modeling, and
solution of problems.
The editors wish to express their particular gratitude to the people who not
only made this handbook possible, but also made it extremely satisfactory:
• The contributors to the handbook, who dedicated much time, effort, and cre-
ative energy to the project. The handbook evolved continuously throughout
the recruitment period, as more and more facets became apparent, many
aspects were entirely new at the time of recruitment.
• The folks at Birkhäuser, particularly Clemens Heine, who initiated the whole
work and gave a lot of encouragement and advice.
• Helga Nutz, Geomathematics Group of the University of Kaiserslautern, for
reading most of the proofs and giving valuable comments.
Thank you very much for all exceptional efforts and support in creating a work
offering exciting discoveries and impressive progress. We hope that the “Handbook
of Mathematical Geodesy” will stimulate and inspire new research achievements
in geodesy as well as mathematics.
February 2017 Willi Freeden, Kaiserslautern

M. Zuhair Nashed, Orlando
Geosystems Mathematics, ix–xiv

Introduction
Willi Freeden
In natural extension to the classical definition due to F.R. Helmert [2], geodesy is
the science that deals with the measurement and modeling of the Earth, including
its gravity field. So, the basis of geodetic science is its measurements, i.e., scalar
numbers, vectors, tensors such as distances, angles, directions, velocities, acceler-
ations. In this respect, the relevance of the gravity field manifests itself in twofold
sense: from the need to handle heights and from the determination of the Earth’s
shape. Consequently, geodesy realizes a physical rather than a geometrical under-
standing of height by observing that a point is higher than another if water flows
from the first to the second. In other words, “geometric” obligations do not allow
to be separated from physical ones. The gravity field is still present, as the driving
force.
Nowadays, geodesy as a measuring discipline is in great shape. In fact, com-
puter facilities as well as measurement and observation methods open new research
areas and opportunities. However, it is geodetic trademark to present measured
values always together with a suitable modeling procedure for interpretation and
an appropriate knowledge and estimation about reliability and accuracy. Follow-
ing R. Rummel [6], this diligence demonstrates the geodesists role as notary of the
Earth. As an evident consequence, however, this notarial role explains that geodesy
is more than a discipline concerned only with measurements. Inherently, mathe-
matics is implied as key technology bridging the real world of measurements and
the virtual world of handling datasets, modeling geodetic quantities and processes,
and providing illustrations and interpretations. Once more, the result of measure-
ments are numbers, vectors, tensors, i.e., raw material. Mathematical handling and
approximation of datasets as well as modeling techniques are necessary to connect
the “reality space” with the “virtuality space”. In this sense, a model represents the
result of the transfer, it intends to be an image of the reality, expressed in math-
ematical language, so that an interaction between abstraction and concretization
is involved. The mathematic’s world of numbers and structures contains efficient
tokens by which we are able to describe the rule-like aspect of a real problem. This
description includes a simplification by abstraction: essential properties of, e.g.,
a certain geodetic problem are separated from unimportant ones and a solution
scheme is set up. The “eye for similarities” enables mathematicians to recognize a
x W. Freeden
posteriori that resulting solutions become applicable to multiple cases not only in
geodesy but also in other scientific disciplines after an appropriate adaptation.
Summarizing we are led to the following conclusion: Mathematical Geodesy
is characterized by a twofold interaction. An input problem from reality space
(“measurements”) reduced by abstraction and transferred into virtuality space
results in a mathematical output model which following a “circuit” (cf. Figure 1)
becomes a new “concrete” input problem in reality space (usually in geodesy, but
possibly also in other sciences).
As a consequence, the ideal process (circuit) for the solution process of geo-
detic problems (as proposed recently by R. Rummel [6]) canonically shows the
following steps to be handled (see also the approach sketched in [1]):
• Transfer from Reality to Virtuality Space: Measurements and observational
events in reality space lead to mathematical tokens and quantities as “row
material” for modeling and processing in virtuality space.
• Mathematical Modeling: The observational input is translated into the lan-
guage of the virtuality space, i.e., mathematics, requiring close cooperation
between application-oriented and mathematical geodesists.
• Development of a Mathematical Solution Method: Appropriate analytic, al-
gebraic, statistic, stochastic, and/or numeric methods must be taken into
account; if necessary, new solution techniques must be proposed.
• Retransfer from Virtuality to Reality Space: The mathematical model is vali-
dated, the aim is a good accordance of model and measurement. If necessary,
on the basis of new measurements, the model must be improved by use of
modified “raw material”.
Observation ( = measurement) Modeling ( = solution)

of events on the basis of datasets
in reality space in virtuality space
transfer (bridging process)

distances, potenƟal theory,
direcƟons, approximaƟon,
run Ɵmes, staƟsƟcs,
velociƟes, inverse problems,
acceleraƟons, numerics,
… ….
retransfer (interpretation, validation)
Figure 1. The circuit.

Introduction xi
Usually, the circuit must be applied several times in an iterative way in or-
der to get sufficient insight into the geodetic system. Obviously, the benefit of a
circuit is a better, faster, cheaper, and more secure problem solution on the basis
of the mentioned processes of modeling, simulation, visualization, and reduction,
decorrelation and denoising of large amounts of data. The more measurements are
available, the more one recognizes the causality between abstraction by mathe-
matical concepts and their impact and cross-sectional importance to reality.
Evidently, the circuit in its ideal manifestation (as illustrated in Figure 1)
has to follow an obligatory line, namely to provide an appropriate platform within
which mathematically/geodetically interrelated features can be better motivated
and understood, thereby canonically leading to an interdisciplinary palette of solu-
tion procedures in geodetic areas of application. In accordance with this intention,
criteria must be found relative to which the scope and limitations of the various
methods can be assessed. This is important both in theory and practice since there
generally is no cure-all method for most of geodetic problems.
The interaction between abstraction and concretization characterizes the his-
tory of geodesy and its efforts as an independent science.
The questions, however, are why
• today’s geodesists commonly restrict themselves to the reality space (“mea-
surements”) with a necessity to accept some “service fundamentals” of the
virtuality space,
• today’s mathematicians are interested only in rare exceptions in appropriate
handling of geodetically relevant obligations including specific model devel-
opments.
Following an article about the interconnecting roles of geodesy and mathe-
matics presented by H. Moritz [5], a prominent member of today’s geodesy, the
actual interrelationship shows a twofold appraisal from history:
• First, H. Moritz [5] states that the old days are gone when Carl Friedrich
Gauss himself developed his epoch-making theories inspired by his geodetic
concerns. Gone also are the days when Felix Klein (1849–1925), one of the
leading mathematicians of his time, called geodesy “that geometrical disci-
pline in which the idea of approximation mathematics has found its clearest
and most consequent expression”(see [4], p. 128). Gone are the times when
Henri Poincaré (1854–1912) investigated problems of astronomy and geodesy
and actively participated in geodetic life. So, we are led to the conclusion
that it apparently is the fault of today’s mathematicians that they provide
mathematics in an increasingly abstract way, without any regard to possible
geodetic applications and, so to say in the scheme of Figure 1, out of touch
with reality? Moritz’ opinion is as follows: “In part, certainly, they are out
of reality.”
• Second, H. Moritz [5] is deeply convinced that an increasing abstraction
is necessary to achieve progress, not only in mathematics, but also in to-
xii W. Freeden
day’s geodesy. What is frequently overlooked by potential geodetic users of

mathematical theory is that the modern abstract methods of mathematics,
if properly understood, provide an extremely powerful tool for the solution
of applied problems which could not be solved otherwise: the more abstract
a method is, the more it is sometimes suitable for a concrete problem. Thus,
we may also conclude that it apparently is the fault of modern geodesists to
be restricted to measurement tasks, without any regard to virtuality space
providing valuable mathematical concepts and, so to say, also out of touch
with virtuality?
As a consequence, in the sense of Moritz’ explications, today’s circuits should

follow the ideal way – at least to a considerable extent – that was initiated by Gauss
as one of history’s most influential mathematician and geodesist for an extremely
fruitful interdisciplinary exchange. The heritage of Gauss’s work has much to of-
fer even these days to build a strong scientific bridge between mathematics and
geodesy by the consequent continuation of the interplay between abstraction and
concretization.
However, it must be confessed that today’s circuits (in the sense as depicted
in Figure 1) turn out to be too complex in their transfer demands from reality
to virtuality space, and vice versa, as to be handled by only one ingenious geo-
scientist. In addition, geodetic changes have been accelerated dramatically. A last
“tour de force” for a consolidation of a circuit by a single scientist in the aforemen-
tioned classical sense probably was the work by L. Hörmander [3] on the “Geodetic
Boundary Value Problem”, but only a few years later geodetic space observation
by GPS made Hörmander’s deep model approach unrealistic, since he started from
the traditional assumption that the actual Earth’s surface was unknown. Nowa-
days, the appalling résumé is that a large number of geodetic problems in their
specific changes and modifications over the last years must be solved simultane-
ously. Interdisciplinary solutions are urgently required as answer to an increasingly
complex geodetic world. In the opinion of the author, the scientific challenge is
a “geodetic consortium”, in which mathematics should not stay for geodesists in
unloved external partnership and geodesy is of high quality problem-attractiveness
for applied mathematicians.
Indeed, the leading role of mathematics for obligations in virtuality space
must be acknowledged (again) within today’s geodesy, so that mathematicians
will become more enthusiastic about working on geodetic programs. A “geodetic
consortium” reflecting the cross-sectional demands in reality as well as virtual-
ity space is absolutely essential for a sustainable development in the future. No
doubt, as this handbook will show, mathematicians can and should be integrated
smoothly into the geodetic phalanx instead of restricting geodesists exclusively to
measurements and mathematicians to mere service functions. Only a “geodetic
consortium” consisting of scientists with equal standing, rights, and research po-
sition will be able to promote the significance of geodesy in its responsibility even
for society similarly to the Gaussian epoch.
Introduction xiii
This “Handbook of Mathematical Geodesy” (HbMG) aims at providing in-

novative mathematical instruments in virtuality space in concrete adaptation to
recent demands of gravity field reflected geodesy. It presents geodetically rele-
vant tools and techniques from functional analysis, potential theory, constructive
approximation, inverse theory, and numerics. A selected list of topics includes
geodetically oriented functional analysis, inverse problem strategies, Gauss’s un-
derstanding of least squares minimization and Nashed’s concept of generalized
inverse, harmonic reproducing kernel Hilbert space theory, the uncertainty prin-
ciple in constructive approximation and its consequences for modeling measured
datasets, Slepian function calculus, wavelet-based Meissl schemata of geodetic ob-
servables, fast spline multi-pole approximation, regularized functional matching
pursuit and its variants, mono-pole and spherically oriented multi-pole Runge–
Walsh approximation for use of, e.g., spherical instead of ellipsoidal harmonics even
on ellipsoids or more complex geodetically relevant surfaces, stochastic geodetic
boundary value problems, spectral tree regularization of “downward continuation
problems” such as gradiometry, mollifier techniques in gravimetry, etc.
The handbook is meant as a mathematical addendum to the foundations in

use within today’s virtuality space (cf. Figure 1). The goal of the handbook is
twofold:
• to make mathematicians aware of the particular mathematical developments
and calamities occurring in modern geodetic concretizations,
• to make geodesists conscious of new tools, means, structures, methods, and
procedures for handling recent measurements and observations by mathemat-
ical abstraction.
All in all, the handbook is understood as an essential step towards modern
manifestations of “geodetic consortia” realizing the cross-sectional demands and
requirements of today’s circuits in well-balanced interdisciplinary way.
References
[1] Freeden, W.: Geomathematics: Its Role, its Aim, and its Potential. Handbook of Geo-
mathematics, W. Freeden, M.Z. Nashed, and T. Sonar (Eds.), Vol. 1, 2nd Edition,
3–79, Springer, New York, Berlin, Heidelberg, 2015.
[2] Helmert, F.R.: Die Mathematischen und Physikalischen Theorien der Höheren Geo-
däsie 1. Vol. 1, Teubner, Leipzig, 1880.
[3] Hörmander, L.: The Boundary Problems of Physical Geodesy. The Royal Institute
of Technology, Division of Geodesy, Report 9, Stockholm, 1975.
[4] Klein, F.: Elementarmathematik III. Die Grundlagen der Mathematischen Wis-
senschaften, Band 16, Springer, Berlin 1928.
[5] Moritz, H.: Geodesy and Mathematics. Zeszyty Naukowe Akademii Görniczo-
Hutniezej I.M. Stanislawa Staszica, No. 780, Geodezja, 63: 38–43, Krakow, 1981.
xiv W. Freeden
[6] Rummel R.: Geodäsie in Zeiten des Wandels – Versuch einer Standortbestimmung.
Zeitschrift für Geodäsie, Geoinformation und Landmanagement (ZfV), 139: 211–216,
2014.
Willi Freeden
Geomathematics Group
University of Kaiserslautern
MPI-Gebäude, Paul-Ehrlich-Str. 26
D-67663 Kaiserslautern, Germany
e-mail: freeden@rhrk.uni-kl.de
Geosystems Mathematics, 1–163

Gauss as Scientific Mediator

Between Mathematics and Geodesy
from the Past to the Present
Willi Freeden, Thomas Sonar, and Bertold Witte
Abstract. The objective of the paper is to document the pioneer dimension of

Gauss’s ideas, concepts, and methods in a twofold direction based on selected
case examples, namely to demonstrate his mediation function between mathe-
matics and geodesy to explain the historic development over the past centuries
from the initial Gaussian ignition to modern characteristics and tendencies.
Keywords. Gauss, geometric number theory, numerical integration, integral

theorems and boundary value problems, least squares adjustment.
1. Introduction
1.1. C.F. Gauss (1777–1855): A brief look at his life
Carl Friedrich Gauss, born on 30 April 1777 in Brunswick (Braunschweig), came
from a humble background whose parents were only able to enroll him in a basic
writing and counting school. His mathematics teacher discovered his exceptional
arithmetic skills and became an advocate for the talented student to facilitate the
placement in a grammar school. At the age of fourteen the young Gauss was in-
troduced to the Duke of Brunswick who pledged to finance the education of the
extraordinarily talented boy. At the age of 15, Gauss had the genial idea to trans-
fer the principles of logarithm tables to the prime number theory. It was the first
time in the history of prime number research that prime number probability be-
came object of research for increasing number ranges. However, he was not able to
prove his prime number assumption that the number π(x) smaller than x behaves
asymptomatically as the quotient from x and log(x). Still this became the starting
point for a variety of number theoretical examinations of renowned mathemati-
cians until present. In the year 1795 Gauss started his studies at the University
of Göttingen. After a brief period of time he decided to study mathematics. In
2 W. Freeden, T. Sonar, and B. Witte
1796 Gauss discovered the solution to an old geometric problem, namely the an-
swer to the question “which straight lines, regular polygons can be exclusively
constructed by compass and straightedge”. The evidence for the constructability
of the regular hexadecagon led Gauss to continue groundbreaking research work,
the results of which he compiled to a significant number theoretical work already
during his study period which, alas, was only published in 1801 because it had to
be translated into Latin as it was customary at that time. This opus published
under the title “Disquisitiones arithmeticae” (cf. Figure 2.1) assured Gauss the
recognition, in particular, amongst the leading French mathematicians of the era
(Cauchy, Laplace, d’Alembert, Laguerre). He was attributed to the best mathe-
maticians of his time due to his number theory. The “Disquisitiones arithmeticae”
contain many significant research results such as, e.g., the celebrated Fundamental
Theorem of Gauss, or the Law of Quadratic Reciprocity of Legendre. The crowning
result of his contributions was the complete solution of binomial equations, and
a most unexpected achievement in placing the imaginary unit on a firm basis. He
actually was the first to use the imaginary symbol “i”, giving it the interpreta-
tion of a geometric mean (see [27] for more details). After having graduated from
Göttingen – upon the Duke’s request who continued to grant him financial sup-
port – he switched to the University of Helmstedt where he studied under Prof.
Pfaff and finished his thesis in 1799 in which he provided the exact evidence of the
“Fundamental Theorem of Algebra” (Latin title: “Demonstratio nova theorematis,
omnem functionem algebraicam rationalem integram unius variabilis in factores
reales primi vel secundi gradus resolvi posse”). In the following time Gauss almost
exclusively dealt with strictly mathematic or geometric questions. Applications
were secondary during this epoch although he had already been working on the
least squares method since 1794 with a multitude of practical studies. The deeper
he penetrated into mathematics, the more fully he was persuaded that its true
meaning lies in its application to practical life and natural science (cf. [27]). On
January 1st, 1801 the astronomer Piazzi from Palermo discovered a small planet
named Ceres. He could only track it for a time period of 41 days. Due to the low
number of measuring values, the known methods at that time did not allow to
calculate the precise orbit of Ceres in order to locate this planetoid again. The
interest in the new planet caused Gauss to temporarily lay aside his purely math-
ematical researches. Now he created applicable methods for orbit determinations.
He sought the orbit which would fit the observations as good as possible using an
ellipse by applying his method of least squares. In doing so, his way resulted in
the solution of two different problems: first, to find an approximate orbit; second,
to correct this orbit in such a way that it “satisfies” the observations as well as
possible (for more details the reader is referred to [27]). At the turn of the year
1801/1802 von Zach in Gotha and Olbers in Bremen rediscovered the planetoid, its
location agreeing exactly with the ephemeris computed by Gauss. The discovery of
Ceres introduced him to the world as an astronomer of the highest order (cf. [27]).
Gauss received a lot of praise and recognition for his scientific achievements;
in January 1802 the Petersburg Academy of Sciences nominated him Correspond-
Gauss as Scientific Mediator 3
Figure 1.1. Medal (headside (left), backside (right)) of the 150th an-
niversary of the death of C.F. Gauss, 2005 (from [122]).
ing Member and in September 1802 he was invited to teach at the Academy which
he declined as other offers later. In 1807 he accepted the call to Göttingen to be-
come a professor for astronomy and the director of the observatory. In the years
to follow Gauss mainly worked on astronomic topics but still on the theory of
numbers. In the meantime he refocused his research works on application-related
topics of mathematics which he found in astronomy as well as in geodesy. From
1818 until approximately 1830 geodesy became the center of his activities. Gauss
made many geodetic observations during the arc measurements and the land sur-
veying (1821–1844) in the Kingdom of Hanover in which he personally partici-
pated. As a matter of fact, Gauss alone managed the comprehensive calculations
for this surveying project. Geodesy derived large direct and indirect profits from
this practical activity which, for instance, was lamented by Bessel because of the
entailing large time commitment and physical strain. During these works of arc
measurements and geodetic surveying Gauss also proved to be a gifted practitioner:
Particular mention should be made to the heliotrope (cf. Figure 6.5) invented by
him which significantly facilitated and accelerated surveying works. The funda-
mental studies in classical differential geometry, in particular on the theory of
surfaces, and his contributions to potential theory and the further development of
the least squares method can be attributed to his practical activity as geodesist
and astronomer. The fundamental surface studies, the so-called “Disquisitiones
generales circa superficies curvas” (Gauss’s Works, Vol. IV), due to their content,
cannot be allocated to the actual geodetic works (see [141]) but more to the math-
ematical fundus. However, they are closely related to geodesy. The significance
for geodesy constitutes in the detailed elaboration on the theory of geodetic lines
for which preparatory studies had been done by Bernoulli, Euler, and Clairaut,
but not with the strong comprehensiveness and thoroughness Gauss applied. The
“Disquisitiones” laid the grounds for suitable arithmetic formulas and coordinate
systems on the geodetically relevant rotation ellipsoid. Gauss published practical
applications for geodetic questions in his “Studies on Topics of Higher Geodesy” in
1843 and 1846 (see Gauss’s Works, Vol. IV). Their importance for geodesy can be
concisely expressed by the following words: Gauss is the founder of higher geodesy.
Since the collaboration with the Göttingen physicist Weber in the year 1831,
Gauss published contributions to classical mechanics, Earth’s magnetism, geo-
metric optics, and electrodynamics. In mechanics, for example, he developed the
principle of least constraint which was named after him and which has the advan-
tage in comparison to other approaches that it can be demonstratively construed.
The theory of optical systems of Gauss for the first time shows the exact principles
of the passage of one light beam through a coaxial lens system when it is slightly
tilted to the axis. The Gauss ocular is still used for autocollimation nowadays,
e.g., for specific tasks in engineering surveying. Gauss’s magnetic studies are rec-
ognized as his most significant contributions to physics. Gauss and Weber jointly
developed the first electromagnetic telegraph.
The three areas of geomagnetism in which Gauss made great contributions
were those related to the absolute measurement of the field, the analysis in terms of
spherical harmonics, and the organization and equipping of magnetic observatories.
Because of the insufficiency of observations over the globe at the time he worked,
many of the investigations which he proposed on the basis of the spherical harmonic
analysis had in spite of their tremendous significance to await later workers. These
include the quantitative separation of internal and external sources, the effect of
the Earth’s ellipsoidal shape, the possible non-vanishing of the constant term in the
expression for the potential and the possible existence of a non-potential portion
of the field (see, e.g., [70, 179] and the references therein).
Following W.K. Bühler [17], Gauss did not venture deeply into electrodynamics.
In this field he is an outsider fascinated by interesting ideas and phenomena. His
legacy contains several interesting recordings, however, no reasonably rounded the-
ory. In fragments regarding the nature of the electromagnetic field, Gauss attempts
to describe the theory of long-distance effects – a theory which was then further
developed by Weber and Neumann and finally superseded by Maxwell’s theory of
electromagnetism.
Measured by his rich mathematic knowledge, Gauss published a relatively
small number of papers. After his death on 23 February 1855, large amount of
unpublished mathematic ideas was detected in his legacy, among others, record-
ings on non-Euclidean geometry. Gauss supposedly asked himself the question in
view of the different geometries, which geometry correctly depicts the physical
reality. The experience taught that if the dimensions are sufficiently small, the
Euclidean geometry applied regarding the measuring precision, based on which
Gauss assumed that the Euclidean geometry also applies to the “infinite small”,
but that deviations occur with larger dimensions. Gauss wished to know whether
our space is curved, i.e., non-Euclidean. For this purpose he measured the inner
angles in a large triangle (cf. Figure 6.1). More concretely, Gauss took measure-
ments from three mountains in Germany, Hohenhagen, near Göttingen, Brocken
in the Harz Mountains and Inselsberg in the Thüringer Wald to the south. The
three lines joining these locations form a great triangle, the angle at Hohenhagen
is close to a right angle, so the area of the triangle is close to half the product of
the two short sides. Gauss assumed that light propagates along geodesics. During
his high-precision measurement campaign he was not able to detect deviations
from the Euclidean internal angle amount, implying that the Euclidean geometry
can be applied for relatively large distances in the physical space. This result was
negative for Gauss which may have been the reason for his decision not to publish
his studies on non-Euclidean geometry. It is doubtful whether this interpretation
is correct; Gauss presumably knew that potential deviations would only become
obvious in triangles of astronomic dimensions (see [240] for more details). At a
later time, Riemann, who was inspired by Gauss, created the Riemann geometry,
the basis on which Einstein was able to build up his relativity theory which also
finds its application, e.g., in satellite geodesy (see also [224]).
In conclusion, it has to be stated that it was extremely fortunate for geodesy
that the world-renowned mathematician (“Princeps Mathematicorum”), Carl
Friedrich Gauss (cf. Figure 1.2), had been so much fascinated by geodesy from
early on in his mathematical career to which he dedicated a significant part of
this work time over his life. In a letter to the astronomer Olbers in January 1802
he expressed it as follows: “The most refined geometer and the perfect astronomer
these are two separate titles which I highly esteem with all my heart, and which I
worship with passionate warmth whenever they are united” (see [27, 141] for more
details).
1.2. Scientific bridge between mathematics and geodesy
Evidently, a contribution concerned with Gauss as scientific mediator between
mathematics and geodesy (as intended by this publication) has to follow an oblig-
atory line, namely to provide an appropriate platform within which mathemati-
cally/geodetically interrelated features can be better motivated and understood,
thereby canonically leading to an interdisciplinary palette of solution procedures
in diverse areas of application. In accordance with this intention, criteria must be
found relative to which the scope and limitations of the various methods can be
assessed. This is important both in theory and practice since there generally is no
cure-all method for most of the problems in reality; it is imperative to be able to
clarify why a certain method works in some context as well as when not to use
that method.
The authors of this work are convinced that Gauss’s suggestions and ideas
as exemplary work is still helpful to understand the intuitive principles and inter-
relations that underlie the various methods and procedures to be needed for the
solution of problems and desiderata. So, in summary, the present article may be
regarded as an attempt to justify this basic assertion.
1.3. Mathematical circuit: abstraction and concretization
What is it exactly that enables mathematicians to build a bridge between geodesy
and their discipline? What is exactly that enables the mathematicians to provide
the transfer from concrete geodetic measurements and observables to abstract
Figure 1.2. Johann Carl Friedrich Gauss, born, 30 April 1777 Bruns-
wick, Duchy of Brunswick-Wolfenbüttel, Holy Roman Empire, died, 23
February 1855 (aged 77) Göttingen, Kingdom of Hanover.
mathematical formalisms and models? What is exactly that enables the mathe-
maticians to bridge the time gap from historic nomenclature to modern notation?
Some answers should be given already at this early stage: The mathematics world
of numbers and structures contains efficient tokens by which the rule-like aspect
of geodetic problems can be described appropriately. In fact, this description in-
cludes as essential step a simplification by abstraction. Essential properties of the
problem are separated from unimportant ones, further specified, and afterwards
included into a solution scheme. The “eye for similarities” often enables mathe-
maticians to recognize a posteriori that an adequately reduced problem may also
arise from very different situations in various application areas, so that the result-
ing solutions may be applicable to multiple cases after an adequate adaptation or
concretization. Without this ingredient, the abstraction remains essentially useless.
The interaction between abstraction and concretization characterizes the history
of mathematics and its current development as a common language and indepen-

dent standards in a unified setup. A problem reduced by abstraction has to be
considered as a new “concrete” problem to be solved within a general framework,
that determines the validity of a possible solution, not only in geodesy, but also in
quite different sciences. The more examples and facets one knows, the more one
recognizes the causality between the abstractness of mathematical concepts and
their impact and cross-sectional significance.
As an immediate consequence, two important questions arise in a work mak-
ing the attempt to pursue geodetic ideas and concepts from Gaussian time up to
resulting mathematical settings in our days.
(i) What can be specified as mathematical structures and settings by a process
of abstraction, i.e., a reduction to essential features?
(ii) What can be achieved from the reduced context obtained by abstraction as
a new field of scientific interest, e.g., in geodesy or other sciences?
The purpose of this contribution is to discuss these questions for a selected
collection of case studies. The interplay between abstraction and concretization will
be explained for each of the problems, thereby offering new perspectives for future
challenges. Dilemmas and methodologies will be indicated during the resolution
process.
1.4. Specific strategies imposed on our work

The purpose of our work is not to describe the entire opus of the “science ti-
tan” Gauss in all its facets from the past until today. Such an approach would be
doomed to fail in view of the enormous significance and the immense amount of
ideas and works of Gauss. The objective of the authors is rather to document the
pioneer dimension of Gauss’s ideas, concepts, and methods in a twofold direction
based on selected case examples, to demonstrate his mediation function between
mathematics and geodesy firstly and secondly the historic development over the
past centuries from the initial ignition by Gauss to the modern characteristics and
tendencies in mathematics and/or geodesy. The authors consciously limited them-
selves to special topics in which they feel competent to focus on based on their
scientific formation. It includes lattice point number theory with its basics for the
development of modern sampling methods, Gauss integration with its impacts on
modern fast algorithms for global modeling of geodetically and geophysically rel-
evant quantities, the potential theory with the Gauss law as the initial point for
geodetic boundary value problems to determine the physically defined Earth’s fig-
ure and finally the least squares method in its canonic continuation into the issues
of present ill-posed and inverse problems, e.g., in terms of pseudodifferential equa-
tions. Due to the reference of this contribution to mathematical geodesy, Gauss’s
contributions to geomagnetics were not reflected here (in this conjunction, refer-
ence is made to, e.g., publications of G.D. Garland [70], K. Reich, E. Roussanova
[179] and their cross-references). In addition, a difficulty arose with translating
Gauss terms into modern mindsets. The mathematic/geodetic language and forms
of expression have developed and changed over the past two centuries. The authors,
alas, were not able to convey and suitably illustrate the development process of
the notation. For reasons of legibility, a standardized unified notation was selected
(except quotes) adjusted to the requirements of modern times as the last link of
the term-defining process.
2. From Gaussian circle problem to geosampling

There is no doubt that the theory of numbers was Gauss’s favourite subject. In a
much quoted dictum, he asserted that
“Mathematics is the Queen of the Sciences and the Theory of Numbers
is the Queen of Mathematics”.
Moreover, in the introduction to “Eisenstein’s Mathematische Abhandlungen”
Gauss wrote:
“The Higher Arithmetic presents us with an inexhaustible storehouse of
interesting truths – of truths, too, which are not isolated but stand in
the closest relation to one another, and between which, with each suc-
cessive advance of the science, we continually discover new and some-
times wholly unexpected points of contact. A great part of the theories of
Arithmetic derive an additional charm from the peculiarity that we eas-
ily arrive by induction at important propositions which have the stamp
of simplicity upon them but the demonstration of which lies so deep as
not to be discovered until after many fruitless efforts; and even then it is
obtained by some tedious and artificial process while the simpler methods
of proof long remain hidden from us.”
All this is well illustrated by what is perhaps Gauss’s most profound publi-
cation, namely his “Disquisitiones arithmeticae”(cf. Figure 2.1). It has been de-
scribed, quite justifiably, as the “Magna Carta of Number Theory”, and the depth
and originality of thoughts to be manifested in this work are particularly remark-
able considering that they were written when Gauss was only about eighteen years
of age. In view of the great impact Gauss had on large areas of modern number
theory, anything even approaching a comprehensive representation of their influ-
ence seems untenable. It is not surprising that there is a huge amount of literature
concerned with Gauss’s number theoretical results, and his influence on modern
mathematics is enormous. However, the obvious problem for our purposes here is
the question if the “Queen of Mathematics” is actually able to show impacts to
modern mathematical geodesy.
2.1. Lattice points inside circles
We start our bridge from Gaussian concepts of number theory to modern geode-
tically relevant sampling with a recapitulation of some results on the number of
lattice points inside circles
S1N = {x ∈ R2 : |x| = N } (2.1)
Figure 2.1. Title-page of “Disquisitiones arithmeticae”, original latin

edition by Gerhard Fleischer, Lipsiae (Leipzig) 1801 (668 pages), first
reprint as first issue of “Gesamtausgabe Carl Friedrich Gauss: Werke.
Band 1, Dieterich, Göttingen 1863”, reprint by Springer-Verlag, New
York Heidelberg 1986, ISBN 0-387 96254-9 (English translation by
Arthur A. Clarke, 1986, in revised form by William C. Waterhouse).
√
2
of radii N > 2 around the origin 0; more accurately, we deal with closed disks
B2N = {x ∈ R2 : |x| ≤ N } (2.2)

√
of radii N > 22 (for more background material and deeper number theoretical
concepts the reader is referred, e.g., to the monographs [47, 66]).
Figure 2.2. Lattice points inside a circle.
The problem of determining the total number of lattice points of Z2 inside and on
a circle with radius N , i.e., the determination of the quantity

Z2 B2N = (n1 , n2 )T ∈ Z2 : n21 + n22 ≤ N 2 (2.3)
reaches back to L. Euler [33]. In today’s nomenclature it can be equivalently ex-

pressed as a sum in the form

Z2 B2N = 1. (2.4)
n21 +n22 ≤N 2 ;
(n1 ,n2 )T ∈Z2
Figure 2.3. The polyhedral set P2N .

Gauss [1801] found a simple, but efficient method for its estimation (cf. Figures
2.2 and 2.3): associate to every square the Northwest edge as lattice point. The
union of all squares with lattice points inside B2N defines a polyhedral set P2N with
area
P2N = Z2 (B2N ) (2.5)
√
(cf. Figure 2.3). Since the diagonal of each square is 2, the geometry of Figure
2.3 tells us that
√ 2 √ 2
2 2
π N− ≤ Z2 B2N ≤ π N + . (2.6)
2 2

Therefore, Z2 B2N − πN 2 after division by N is bounded for N → ∞, which is
usually written with Landau’s O-symbol as

Z2 B2N = πN 2 + O(N ). (2.7)
In other words, the number of lattice points in B2N is equal to the area of that
circle plus a remainder of the order of the boundary. In particular,

Z2 B2N ∼ πN 2 (2.8)
so that a method of determining the irrational, transcendent number π becomes
obvious (for alternative approaches to π within the history of analysis the reader
is referred to [204]):
Z2 B2N
lim = π. (2.9)
N →∞ N2
C.F. Gauss [71] illustrated his result by taking N 2 = 100 000. In this case he
calculated
1 = 314 197. (2.10)
|g|2 ≤100 000;
g∈Z2
This calculation determines the number π up to three decimals after the comma.
2.2. Circle problem and Hardy’s conjecture

The formula (2.7) due to C.F. Gauss [71] allows the following representation in
the nomenclature of Landau’s O-symbols:

Z2 B2N = πN 2 + O(N ). (2.11)
The so-called circle problem is concerned with the question of determining the
bound
α2 = inf γ : Z2 B2√N = πN + O(N γ ) . (2.12)
Until now, we knew from (2.11) that α2 ≤ 12 . An improvement of the Gaussian
result, however, turns out to be very laborious, in fact, requiring a great effort.
A first remarkable result is due to W. Sierpinski [201], who proved by use of a

method of his teacher G. Voronoi [220] that
1
Z2 B2√N = πN + O N 3 , (2.13)
i.e., α2 ≤ 13 . The proof of Sierpinski is elementary (see, e.g., [47, 66] for more
details); it is a link between geometry and number theory.
By use of advanced methods on exponential sums (based on the work by,
e.g., H. Weyl [228], H.R. Chen [18], and many others) the estimate 13 could be
strengthened to some extent. It culminated in the publication by G. Kolesnik
[128], who had as his sharpest result with these techniques
139
Z2 B2√N − πN = O N 429 . (2.14)
M.N. Huxley [117] devised a substantially new approach (not discussed here); his
strongest result was the estimate
131
Z2 B2√N − πN = O N 416 . (2.15)
(note that 139 131

429 = 0.324009 . . . , while 416 = 0.315068 . . . ).
Hardy’s conjecture claims
1
Z2 B2√N − πN = O N 4 +ε (2.16)
for every ε > 0. This conjecture seems to be still a challenge for future work. How-
ever, in the year 2007, S. Cappell and J. Shaneson deposited a paper entitled “Some
Problems in Number Theory I: The Circle Problem” in the arXiv:math/0702613
1
claiming to prove the bound of O(N 4 +ε ) for ε > 0.
0.250000 Gauss (1801)

0.083333 . . . G. Voronoi (1903), Sierpinski (1906)
0.080357 . . . J.E. Littlewood, A. Walfisz (1924)
0.079268 . . . J.G. van der Corput (1928)
0.074324 . . . J.-R. Chen (1963)
0.074009 . . . G. Kolesnik (1985)
0.064903 . . . M.N. Huxley (2003)
Table 1. Incremental improvements for the value ε2 in the estimate (2.17).

Table 1 lists incomplete incremental improvements for the quantity ε2 of the upper
limit for the circle problem
1
Z2 B2√N − πN = O N 4 +ε2 . (2.17)
For all recent improvements, the proofs became rather long and made use of some
of the more heavy machinery in hard analysis.
Summarizing our results about lattice points inside circles (cf. [47]) we are
confronted with the following situation:
1 1
≤ α2 ≤ + ε2 (2.18)
4 4
and 1
Z2 B2√N − πN = O N 4 , (2.19)
1
Z2 B2√N − πN = O N 4 +ε2 , (2.20)
1
where 0 < ε2 ≤ 4 (for example, Huxley’s bound ε2 = 0.064903 . . .).
2.3. Variants of the circle problem

There are many perspectives to formulate variants of the Gaussian lattice point
problem for the circle. It already was the merit of E. Landau [135] to point out
particularly interesting areas, such as
• General two-dimensional lattices
Λ = {g = ng1 + mg2 : n, m ∈ Z} (2.21)
with g1 , g2 ∈ R2 linearly independent (see Figure 2.4) can be used instead of
the unit lattice Z2 .
g2
0 g1
Figure 2.4. Two-dimensional lattice Λ generated by g1 , g2 ∈ R2 .
• The remainder term can be represented as alternating series, called Hardy–

Landau series in terms of the Bessel function J1 of order 1 (for the different
facets of the proof see [103, 134, 160])
πN 2 πN 2 J1 (2π|h|N )
1= + lim , (2.22)
FΛ R→∞ FΛ π|h|N
|g|≤N 0<|h|≤R
g∈Λ h∈Λ−1
where Λ is an arbitrary lattice in R2 and

1 1
FΛ = x = x1 g1 + x2 g2 ∈ R2 : − ≤ xi < , i = 1, 2 (2.23)
2 2
is the fundamental cell of Λ ⊂ R2 with

FΛ = det ((gi · gj )i,j=1,2 ) (2.24)
as the area FΛ of FΛ . Moreover, the following convention
1
1= 1+ 1 (2.25)
2
|g|≤N |g|<N |g|=N
g∈Λ g∈Λ g∈Λ
is used in lattice point theory (note that the last sum only occurs if there is
a lattice point g ∈ Λ with |g| = N ).
• Lattice points can be affected by non-constant weights (see [35])

e2πiy·(a+g) F (a + g)
|a+g|≤N
g∈Λ
1 2πia·h
= lim e F (x) e−2πix·(h−y) dx, (2.26)
R→∞ FΛ |x|≤N
|h−y|≤R x∈R2
h∈Λ−1
where dx is the volume element, a, y ∈ R2 , F is twice continuously differen-

tiable in B2N , N > 0, and the following convention has been used analogously
to (2.25)
1
··· = ··· + ... . (2.27)
2
|a+g|≤N |a+g|<N |a+g|=N
Note that, for F = 1, this formula leads back to

πN 2 J1 (2π|h − y|N )
e2πia·y e2πig·y = lim e2πia·h . (2.28)
R→∞ FΛ π|h − y|N
|g+a|≤N |h−y|≤R
g∈Λ h∈Λ−1
For a = y = 0 we obtain the classical Hardy–Landau identity, i.e., the identity

πN 2 J1 (2π|h|N )
1 = lim . (2.29)
R→∞ FΛ π|h|N
|g|≤N |h|≤R
g∈Λ h∈Λ−1
holds true. Observe that J1 satisfies the asymptotic relation J1 (r) = r

2 +···,
so that
πN 2 πN 2 J1 (2π|h|N )
1= + lim . (2.30)
FΛ R→∞ FΛ π|h|N
|g|≤N 0<|h|≤R
g∈Λ h∈Λ−1
• Generalizations to lattices Λ ⊂ Rq and regular regions G ⊂ Rq , q ≥ 2, and

continuous functions on G = G ∪ ∂G can be formulated in Gaussian summa-
bility (see [47] for the proof and a more detailed study)

a+g∈G
g∈Λ

1 −τ π 2 h2 2πih·a
= lim e e F (x)e−2πix·(h−y) dx, a, y ∈ Rq , (2.31)
τ →0 FΛ G
τ >0 −1 h∈Λ
where a regular region G in Rq is understood to be an open and connected

set G ⊂ Rq , q ≥ 2, for which
(i) its boundary ∂G constitutes an orientable, piecewise smooth Lipschitzian
manifold of dimension q − 1,
(ii) the origin is contained in G,
(iii) G divides Rq into the “inner space” G and the “outer space” Rq \G, G =
G ∪ ∂G.
Clearly,

a+g∈G
g∈Λ
1
= F (x)e2πix·y dx
FΛ G

1 2
h2 2πih·a
+ lim e−τ π e F (x)e−2πix·(h−y) dx, a, y ∈ Rq .
τ →0 FΛ G
τ >0 0<|h|≤R
h∈Λ−1 (2.32)
The following abbreviation has been used consistently

··· = ··· + α(a + g) · · · (2.33)
a+g∈G a+g∈G a+g∈∂G
with α(a + g) denoting the solid angle subtended by ∂G at a + g (note that,

as geoscientifically relevant regular regions, we may choose the interior of the
(actual) Earth’s body or parts of it, the interior of geoscientifically relevant
surfaces such as the geoid, telluroid, etc., but also ball, ellipsoid, cube, poly-
hedral bodies, etc. are included in accordance with the above definition; in
potential-theoretic jargon, the solid angle will be explained by Definition 4.1
in a more detailed way).
2.4. Multivariate Shannon sampling

Let us continue with the observation that, for every y ∈ Rq , the (for dimensions
q ≥ 3 formally understood) series

1
a → lim e2πia·h F (x)e−2πix·(h−y) dx, a ∈ Rq , (2.34)
N →∞ FΛ
|h−y|≤N G
h∈Λ−1 FG∧ (h−y)
as well as the finite sum

a → e2πiy·(a+g) F (a + g), a ∈ Rq , (2.35)
a+g∈G
g∈Λ
show Λ-periodicity, i.e., as functions of the variable a ∈ Rq they are periodic with
respect to the lattice Λ ⊂ Rq . As a consequence (see [57] for the details), Shannon-
type sampling procedures can be obtained by formal integration of the lattice point
identity (2.31) over a regular region H that is not necessarily equal to G

e−2πiy·(a+g) F (a + g) da
H
a+g∈G
g∈Λ

=
F (x)e−2πiy·x dx
G∩ (((H∩(FΛ +{g }))−{g })+{g})
(FΛ +{g })∩H =∅ g∈Λ
g ∈Λ
= F∧ (y)
G∩ (((H∩(FΛ +{g }))−{g })+{g})
g∈Λ

1 −τ π 2 h2 −2πih·x
= lim e F (x)e dx e2πia·(h−y) da. (2.36)
τ →0 FΛ G H
τ >0 h∈Λ−1
= FG∧ (h) = KH (h−y)
The identity (2.36) has many interesting properties. For example, by virtue of the
Gaussian summability, the convergence of the cardinal-type series on the right-
hand side of (2.36) may be exponentially accelerated. All manifestations of over-
and undersampling can be explicitly analyzed by the finite sum of Fourier trans-
forms on the left side of the identity (see [57]), dependent on the geometric con-
figurations of the chosen regular regions G, H (note that the identity (2.36) also
seems to be unknown for the uni-variate case in this generality).
The Gaussian summability of the cardinal series on the right-hand side of
(2.36) is of great importance from numerical point of view; it enables a fast com-
putation of the series. Nonetheless, W. Freeden, M.Z. Nashed [57]) show that the
identity (2.36) additionally holds true in ordinary sense, i.e., we have

e−2πiy·(a+g) F (a + g) da
H
a+g∈G
g∈Λ

=
F (x)e−2πiy·x dx
G∩ (((H∩(FΛ +{g }))−{g })+{g})
(FΛ +{g })∩H =∅ g∈Λ
g ∈Λ
= F∧ (y)
G∩ (((H∩(FΛ +{g }))−{g })+{g})
g∈Λ
1
= F (x)e−2πih·x dx e2πia·(h−y) da. (2.37)
FΛ
h∈Λ−1 G H
= FG∧ (h) = KH (h−y)
A simple, but significant case of (2.37) in Euclidean space Rq showing no phenom-

enon of aliasing is obtained under the special choice G = H ⊂ FΛ leading to the
Shannon-type identity
1
FG∧ (y) = FG∧ (h) KG (h − y). (2.38)
FΛ
h∈Λ−1
In fact, the identity (2.38) is a multi-variate variant of the Shannon sampling

theorem (cf. [198]), but now for (geoscientifically relevant) regions G. The principal
impact of Shannon sampling on information theory is that it allows the replacement
of a bandlimited signal FG∧ related to G by a discrete sequence of its samples
without loss of any information. Also it specifies the lowest rate, i.e., the Nyquist
rate (cf. [57]), that it enables to reproduce the original signal. In other words,
Shannon sampling provides the bridge between continuous and discrete versions
of a bandlimited function.
The Shannon sampling theorem has many applications in engineering and
physics, for example, in signal processing, data transmission, cryptography, con-
structive approximation such as spectral analysis by Slepian functions (see, e.g.,
[202, 203]), partial differential equations such as Boltzmann equation (see, e.g.,
[171] and the references therein), and inverse problems such as the multi-variate
discussion of a Fourier transform over a (geodetically relevant) regular region (see,
e.g., [164] and the references therein for the uni-variate study of the antenna prob-
lem).
Finally, it should be remarked that sampling theory (in the sense of the rep-
resentation of an analog signal in terms of its samples) has attracted considerable
interest in the past three decades. Major advances in the theory and applications of
sampling expansion in different function spaces have been made by mathematicians
and engineers. There are several journals on signal processing and computational
harmonic analysis that publish papers on sampling theory and its applications.
The field also sports its own journal “Sampling Theory in Signal and Image Pro-
cessing” and has its own “professional society”. We quote from Wikipedia:
“SampTA (Sampling Theory and Applications) is a biennial interdisci-
plinary conference for mathematicians, engineers, and applied scientists.
The main purpose of SampTA is to exchange recent advances in sam-
pling theory and to explore new trends and directions in related areas of
applications. The SampTA conference series began as a small workshop
in 1995 in Riga, Latvia, but the meetings grew into full-fledged confer-
ences attracting an even mix of mathematicians and engineers as the
interest in sampling theory blossomed.”
2.5. Paley–Wiener spline interpolation

Under the aforementioned assumption that G is a regular region with G ⊂ FΛ ,
standard Fourier inversion (see, e.g., [161]) guarantees that

∧ ∧ 2πia·(x−y)
FG (y) = FG (x) e da dx
q
R G
= FG∧ (x) KG (x − y) dx (2.39)

R q

holds true for all y ∈ Rq , where Rq . . . is understood in the following sense:

. . . = lim ... . (2.40)
Rq N →∞ |x|≤N
x∈Rq
Hence, the Paley–Wiener space BG

−2πia·y 2
BG = y → e F (a) da, y ∈ R : F ∈ L (G) .
q
(2.41)
G
is a reproducing kernel Hilbert space with the uniquely determined kernel

KG (x − y) = e2πia·(x−y) da. (2.42)
G
The reproducing kernel framework enables us to realize minimum-norm-inter-

polation, i.e., spline interpolation, within the Paley–Wiener space BG . More con-
cretely, we are interested in finding a “smooth” spline interpolant to FG∧ ∈ BG
from a given finite dataset
{FG∧ (h) : h ∈ Ξ, Ξ ⊂ Λ−1 }. (2.43)
For that purpose, we consider the finite-dimensional space Spline Ξ

BG consisting of
all Paley–Wiener spline functions Ŝ of the form

Ŝ(y) = âh e2πia·(h −y) da, y ∈ Rq , âh ∈ C. (2.44)
h ∈ Ξ G
It is easy to deduce from the theory of Gramian determinants (see, e.g., [23]) that
there is one and only one spline in Spline Ξ F
BG , denoted by ŜF , with coefficients âh
satisfying the linear equations

F
âh e2πia·(h −h) da = FG∧ (h), h ∈ Ξ. (2.45)
h ∈ Ξ G
ŜF fulfills the minimum-norm-property

2 1/2 1/2
2
ŜF (y) dy = min |Ŝ(y)| dy .
G Ŝ∈BG G
Ŝ(h)=FG∧ (h)
h∈Ξ
As a consequence, suppose that the dataset

{FG∧ (h) = e−2πia·h F (a) da : h ∈ Ξ, Ξ ⊂ Λ−1 } (2.46)
G
is given such that ŜF is “close” to FG∧ , for example, if the nodal width of Ξ
is “small enough” (see the convergence theorems presented in [44]). Then, the
Fourier inversion formula (see, e.g., [49]) yields the identity

α(x) F (x) e2πix·y ŜF (y)dy
Rq

2πix·y
= âF
h e e2πia·(h −y) da dy
h ∈ Ξ Rq G

2πix·h
= âF
h e , x ∈ Rq , (2.47)
h ∈ Ξ
with coefficients âFh determined from the already known linear (spline) equations
(2.45). In other words, our spline interpolation technique approximately solves
the multi-variate inversion problem of determining the function αF ∈ L2 (G)
from a finite set of discrete values of FG∧ (note that combined spline interpola-
tion/smoothing (as proposed by W. Freeden, B. Witte [62]) can be used instead
of spline interpolation if the data are (partially) noisy).
Finally, a generalization of the multi-variate antenna problem should be ex-
plained in more detail: In antenna theory, the identification problem is the one
in which the far-field radiation pattern is known, either exactly or approximately,
by means of physical measurements. The problem is to find the aperture distribu-
tion producing this given pattern. The synthesis problems are ones in which we
are given a desired far-field pattern, specified exactly or partially (e.g., through
samples of the far field at a finite number of points or through measurements
contaminated by noise), and we wish to determine a source (a constrained or un-
constrained aperture distribution) whose far-field radiation pattern approximates
the desired pattern in some acceptable manner. In terms of the operator equation,

AG F (y) = e−2πia·y F (a) da = FG∧ (y), (2.48)
G
F represents the aperture distribution, G represents the far-field pattern, and AG

the operator which relates these two functions, characterizes the antenna structure.
For the convenience of the reader, we recast the operator-theoretic setting of
a finite area source antenna: Consider the problem, appropriately normalized, of a
linear aperture on G. Then, the aperture distribution F is related to the far field
FG∧ ∈ BG by means of an integral equation (2.48) of the form the pattern being
limited to a visible range containing the lattice points h ∈ Ξ, Ξ ⊂ Λ−1 . In fact,
assuming that FG∧ (h) is known for the lattice points h ∈ Ξ, Ξ ⊂ Λ−1 we are led
back to the aforementioned spline problem, and an approximation to the aperture
distribution is known from formula (2.47) to handle, for example, the multi-variate
antenna problem.
Summarizing our number theoretical excursion starting from the Gaussian
circle problem via the Hardy–Landau lattice point identities we resulted in new
geoscientifically relevant Shannon sampling procedures of high practical applicabil-
ity enabling Paley–Wiener spline interpolation of Fourier transforms from discrete
data, regularly located in lattice points.
3. From Gaussian integration to geocubature

Gauss’s famous method of approximate integration, almost immediately after its
discovery and throughout the whole 19th century, attracted the attention of the
leading mathematicians of the time. The story of Gaussian quadrature began with
Newton and Cotes. Following W. Gautschi [85], Newton, in 1676, was the first to
suggest a rule of approximate integration. Cotes, independently, arrived at similar
methods, and brought them into workable form after learning of Newton’s ideas.
In 1814, Gauss took the work of Newton and Cotes as a point of departure. He
combined it with his own knowledge on the hypergeometric series to develop his fa-
mous new method of integration. In more detail, if the set of (distinct) nodes in the
Newton–Cotes rules vary freely and the weights are computed in accordance with
the Newton–Cotes formula, Gauss’s problem was, what is the maximum degree of
polynomial exactness that can be achieved? How are the nodes to be selected in
order to realize this optimal situation? These were questions raised by C.F. Gauss
[76], and answered most elegantly by means of his theory of continued fractions
associated with hypergeometric series. The continued fraction was well known to
Gauss, being a special case of his general continued fraction for ratios of hypergeo-
metric functions (see [74]). Gauss’s breakthrough discovery first inspired Jacobi to
provide an elegant alternative derivation. Christoffel then significantly generalized
the method and subsequently extended it to arbitrary measures of integration.
Stieltjes established the legitimacy of the method, by proving its convergence,
while Markov endowed it with an error term. Thus, by the end of the 19th cen-
tury, the Gauss integration method became firmly entrenched in the repertoire of
numerical methods of approximation.
In spite of the huge literature about Gauss’s rules and its enormous appli-
cations there are still new aspects in the context of approximate integration. In
this work, we embed the Gaussian numerical integration in the theory of Legendre
operators and its inversion by means of mathematical physics, i.e., by use of the
theory of Green’s functions. This procedure (cf. [36]) leads us to new error terms
in adaptation to the integrand under consideration which finally contribute to the
convergence of geocubatures in terms of spherical harmonics.
3.1. Gaussian integration revisited
In order to derive our approach to Gaussian quadrature formulas it is necessary
to have some basic knowledge on classical Legendre polynomials. We start with
its explicit representation. Then we list some integral formulas involving Legendre
operators which turn out to be useful in Gaussian quadrature.
Legendre Polynomials. We begin with the explicit definition of the Legendre poly-
nomial (for more details see, e.g., [50] and the references therein).
Definition 3.1. The function Pn : [−1, 1] → R, n = 0, 1, . . . , defined by
n
2
(2n − 2s)!
Pn (t) = (−1)s tn−2s , t ∈ [−1, 1] (3.1)
s=0
2n (n − 2s)!(n − s)!s!
is called the Legendre polynomial.
Pn : [−1, 1] → R is uniquely determined by the properties:
(i) Pn is a polynomial of degree n on the interval [−1, 1],
1
(ii) Pn (t)Pm (t) dt = 0 for n = m,
−1
(iii) Pn (1) = 1.
This is easily seen from the usual process of orthogonalization. In particular, we
have, for n = 0, . . . , 4,
3 1
P0 (t) = 1, P1 (t) = t, P2 (t) = t2 − , (3.2)
2 2
5 3 35 4 15 2 3
P3 (t) = t3 − t, P4 (t) = t − t + . (3.3)
2 2 8 4 8
A graphical impression of the first Legendre polynomials can be found in Figure
3.1. Furthermore, an easy calculation shows that
1
2
Pn (t)Pm (t) dt = δnm . (3.4)
−1 2n +1
0.5
Ŧ0.5 P1
P2
P3
P4
Ŧ1
Ŧ1 Ŧ0.5 0 0.5 1
Figure 3.1. Legendre polynomials t → Pn (t), t ∈ [−1, 1], n = 1, . . . , 4.

The Legendre polynomial satisfies the estimate

|Pn(k) (t)| ≤ |Pn(k) (1)|, (3.5)
where
k
1 1
Pn(k) (1) = n(n + 1) ((n(n + 1) − 1 · 2) . . . (n(n + 1) − k(k − 1))) . (3.6)
2 k!
A straightforward calculation yields
2
2 d d
(1 − t ) − 2t + n(n + 1) Pn (t) = 0, t ∈ [−1, 1], (3.7)
dt dt
=L∧ (n)
=Lt
where Lt is called the Legendre operator. We therefore obtain the following lemma.
Lemma 3.1. The Legendre polynomial Pn is the only twice differentiable eigen-
function of the “Legendre operator” L as defined in (3.7) on the interval [−1, 1],
corresponding to the eigenvalues L∧ (n) = n(n + 1), n = 0, 1, . . . , and bounded on
[−1, 1] with Pn (1) = 1.
The differential equation (3.7) shows that Pn and Pn cannot vanish simulta-
neously such that Pn has no multiple zeros. Moreover, the Legendre polynomial
Pn has n different zeros in the interval (−1, 1).
Lemma 3.2 (Rodriguez Formula). For n = 0, 1, . . . ,
n
1 d
Pn (t) = n (t2 − 1)n , t ∈ [−1, 1]. (3.8)
2 n! dt
Integrating by parts we obtain the Rodriguez rule
1 1
1
F (t)Pn (t) dt = n F (n) (t)(1 − t2 )n dt (3.9)
−1 2 n! −1
for every F ∈ C (n) [−1, 1]. It is not hard to show that

Pn+1 (t) − tPn (t) = (n + 1)Pn (t), (3.10)
(t2 − 1)Pn (t) = ntPn (t) − nPn−1 (t), (3.11)
(n + 1)Pn+1 (t) + nPn−1 (t) − (2n + 1)tPn (t) = 0. (3.12)
The formulas (3.10)–(3.12) are known as the recurrence formulas for the Legendre
polynomials.
From [209] we borrow the following estimate of the Legendre polynomial.
Lemma 3.3. For n = 1, 2, . . . and t ∈ (−1, 1),
1/2
1 4
|Pn (t)| ≤ √ . (3.13)
π n(1 − t2 )
L2 -Orthonormalized Legendre Polynomials. The system {Pn∗ }n=0,1,... given by

∗ 2n + 1
Pn (t) = Pn (t), t ∈ [−1, 1] (3.14)
2
is orthonormal in the sense
1
Pn∗ (t)Pm
∗
(t) dt = δn,m . (3.15)
−1
The system {Pn∗ }n=0,1,... of orthonormal Legendre polynomials is closed in the

Hilbert space (L2 ([−1, 1]), ·, ·L2 ([−1,1]) ) and is closed in the Banach space
(C (0) ([−1, 1]), · C (0) ([−1,1]) ).
In consequence, any function F ∈ L2 ([−1, 1]) can be represented as orthogo-

nal expansion by means of {Pn∗ }n=0,1,...
⎛ 2 ⎞1/2
1 N 1

lim ⎝ F (t) − F (x)Pn∗ (x) dx Pn∗ (t) dt⎠ = 0. (3.16)
N →∞ −1 −1n=0

Our formulas arising orthonormal Legendre polynomials also give a different

and useful intuition to Gauss–Legendre sampling: The kernel Kn (·, ·) : [−1, 1] ×
[−1, 1] → R given by

n
Kn (t, x) = Pk∗ (t)Pk∗ (x), x, t ∈ [−1, 1] (3.17)
k=0
is a reproducing kernel of the space P ol0,...,n of (algebraic) polynomials of degree

≤ n, i.e., we have
(i) for each t ∈ [−1, 1], Kn (t, ·) is a member of P ol0,...,n ,
(ii) F (t) = Kn (t, ·), F L2 ([−1,1]) for all F ∈ P ol0,...,n .
The uniquely determined reproducing kernel (3.17) can be expressed in explicit way
(see, e.g., [209]): From the recursion relation (3.12) we easily obtain by elementary
manipulation
(k + 1)(Pk+1 (x)Pk (t) − Pk+1 (t)Pk (x)) − k(Pk (x)Pk−1 (t) − Pk (t)Pk−1 (t))
= (2k + 1)(x − t)Pk (x)Pk (t). (3.18)
Summation over k from 1 to n yields the identity

n
(x − t) (2k + 1)Pk (x)Pk (t)
k=1
= (n + 1)(Pn+1 (x)Pn (t) − Pn+1 (t)Pn (x)) − (x − t), (3.19)
where we have used that P0 (x) = 1, P1 (x) = x. Therefore, for x = t, we find the
so-called Christoffel–Darboux formula
n + 1 Pn+1 (x)Pn (t) − Pn+1 (t)Pn (x)
K(x, t) = . (3.20)
2 x−t
Equivalently, we have
∗
n+1 Pn+1 (x)Pn∗ (t) − Pn+1
∗
(t)Pn∗ (x)
K(x, t) = √ √ . (3.21)
2n + 1 2n + 3 x−t
We take xi,n+1 , i = 1, . . . , n + 1, to be the (distinct)
zeros of Pn+1 . Then,
n 2
K(xj,n+1 , xi,n+1 ) = 0 for i = j and K(xi,n+1 , xi,n+1 ) = k=0 Pk (xi,n+1 ) > 0
since the zeros are interlaced. As a consequence (cf. [167]), we have

n+1
K(xi,n+1 , t)
F (t) = F (xi,n+1 )
i=1
K(xi,n+1 , xi,n+1 )

n+1
Pn+1 (t)
= F (xi,n+1 ) . (3.22)
i=1
Pn+1 (xi,n+1 )(t − xi,n+1 )
Green’s Function for the Legendre Operator. As already known, the Legendre
operator L has a half-bounded and discrete eigenspectrum {L∧ (n)}n=0,1,... such
that
(Lt + L∧ (n))Pn (t) = 0, t ∈ [−1, 1], (3.23)
where
L∧ (n) = n(n + 1), n = 0, 1, . . . . (3.24)
∧ ∧
Thus, SpectL = {L (n) ∈ R : L (n) = n(n + 1), n = 0, 1, . . .} is the eigenspectrum
of the Legendre operator.
The Hilbert theory of Green functions (cf. [111]) leads to the following definition.
Definition 3.2. A function G(L + λ; ·, ·): [−1, 1] × [−1, 1] → R, (x, t) → G(L +
λ; x, t), λ ∈ R, is called Legendre (Green) function with respect to the operator
L + λ, λ ∈ R, if it satisfies the following properties:
(i) (Boundedness) For each fixed x ∈ [−1, 1], G(L + λ; x, ·) is a continuous func-
tion on (−1, 1) satisfying the conditions
|G(L + λ; x, 1)| < ∞, (3.25)
|G(L + λ; x, −1)| < ∞. (3.26)
(ii) (Differential equations) For each fixed x ∈ [−1, 1], G(L + λ; x, ·) is twice
/ SpectL and t ∈ [−1, 1]\{x}
continuously differentiable in [−1, 1]\{x}. For λ ∈
we have
(Lt + λ)G(L + λ; x, t) = 0, t ∈ [−1, 1]\{x}, (3.27)
while, for λ = L∧ (n) ∈ SpectL , we have
(Lt + λ)G(L + λ; x, t) = −Pn∗ (x)Pn∗ (t), t ∈ [−1, 1]\{x}. (3.28)
(iii) (Characteristic singularity)

t=x+0
d
(1 − x2 ) G(L + λ; x, t) = 1. (3.29)
dt t=x−0
(iv) (Normalization) For each x ∈ [−1, 1] and λ = L∧ (n) ∈ SepctL ,

+1
G(L + L∧ (n); x, t)Pn∗ (t) dt = 0. (3.30)
−1
The function Pn∗ is an eigenfunction of the Green function with respect to

the operator L in the sense of the linear integral equation
1
(1 − δλ,L∧ (n) )Pn∗ (t) = (λ − L∧ (n)) G(L + λ; t, u)Pn∗ (u) du. (3.31)
−1
The bilinear expansion of G(L + λ; ·, ·) therefore reads

∞
1
G(L + λ; x, t) = P ∗ (x)Pk∗ (t), (3.32)
λ − L∧ (k) k
k=0
L∧ (k) =λ
where ΣL∧ (k) =λ means that the summation is to be extended over all k ∈ N0 with
L∧ (k) = λ. Obviously, because of Pn (t) = O((n(1 − t2 ))−1/2 ), t ∈ (−1, 1), the
bilinear expansion (3.32) is absolutely and uniformly convergent both in x and t
for every compact subinterval of (−1, 1).
From the completeness of the system {Pn∗ }n=0,1,... we easily obtain that
G(L + λ; ·, ·) is uniquely determined by its defining properties (i)–(iv).
A particular role is played by the Legendre (Green) function with respect to
the operator L (i.e., λ = 0). It is explicitly available as elementary function (cf.
[36])

1 1
2 ln((1 + t)(1 − x)) + ln 2 − 2 , x≤t
G(L; x, t) = 1 1 (3.33)
2 ln((1 − t)(1 + x)) + ln 2 − 2 , t ≤ x.
In the sense of the Fredholm–Hilbert theory of linear integral equations we may
interpret the Green function G(L + λ; ·, ·) for λ = 0 as resolvent of the kernel
1
G(L + λ; x, t) = G(L; x, t) + λ G(L + λ; x, u)G(L; t, u) du
−1
1 1
− − Pn∗ (x)Pn∗ (t)δλ,L∧ (n) . (3.34)
2λ λ
Inserting the bilinear expansions we therefore obtain, for each λ = 0,
∞

∗ 1
G(L + λ; x, t) = G(L; x, t) − λ Pk∗ (x)Pk∗ (t)
(λ − L∧ (k))L∧ (k)
k=0
1 1
− − P ∗ (x)Pn∗ (t)δλ,L∧ (n) , (3.35)
2λ λ n
where the series on the right-hand side converges uniformly and absolutely both
in x and t on the interval [−1, 1], and the symbol Σ∗ means, that the summation
is extended over all k ∈ N0 satisfying (λ − L∧ (k))L∧ (k) = 0.
Next we introduce Legendre (Green) functions with respect to the iterated
operator (L + λ)p , p ∈ N.
Definition 3.3. For λ ∈ R, the function G((L + λ)p ; ·, ·), p = 2, 3, . . ., defined
recursively by
1
G((L + λ)p ; x, t) = G((L + λ)p−1 ; x, u)G(L + λ; t, u) du, (3.36)
−1
is called Legendre (Green) function with respect to the operator (L + λ)p .

G((L + λ)p ; ·, ·) is symmetric in its arguments, i.e., for all x, t ∈ [−1, 1]
G((L + λ)p ; x, t) = G((L + λ)p ; t, x). (3.37)
The bilinear expansion of G((L + λ)p ; ·, ·) reads as follows
∞
1
G((L + λ)p ; x, t) = P ∗ (x)Pk∗ (t). (3.38)
(λ − L∧ (k))p k
k=0
λ =L∧ (k)
For p = 2, 3, . . ., the series on the right-hand side is absolutely and uniformly

convergent both in x ∈ [−1, +1] and t ∈ [−1, 1].
If F is continuously differentiable at the point t ∈ (−1, 1), then
1
(Lt + λ) G(L + λ; t, u)F (u) du
−1
+1
= F (t) − δλ,L∧ (n) Pn∗ (t) Pn∗ (u)F (u) du. (3.39)
−1
In particular, we have
Lemma 3.4. For p = 2, 3, . . . and t = x
(Lt + λ)p−1 G((L + λ)p ; x, t) = G(L + λ; x, t). (3.40)
Integral Formulas for the Legendre Operator. Suppose that F : [−1, 1] → R is of
the class C (2) ([−1, 1]). Assume that λ ∈ R, x ∈ (−1, 1), and ε > 0 (sufficiently
small). Then partial integration, i.e., the so-called Green–Lagrange formula yields
x−ε
{F (t)(Lt + λ)G(L + λ; x, t) − G(L + λ; x, t)(Lt + λ)F (t)} dt
−1
1
+ {F (x)(Lt + λ)G(L + λ; x, t) − G(L + λ; x, t)(Lt + λ)F (t)} dt
x+ε

2 d d t=x+ε
= −(1 − t ) F (t) G(L + λ; x, t) − G(L + λ; x, t) F (t) . (3.41)
dt dt t=x−ε
Observing the differential equation and the characteristic singularity of the Green
function with respect to L + λ we obtain by letting ε → 0 the following integral
formulas.
Theorem 3.1 (Integral Formula for L + λ). Let x be a point in (−1, 1). Suppose
that F is of class C (2) ([−1, 1]). Then, for λ ∈/ SpectL ,
1
F (x) = G(L + λ; x, t)(Lt + λ)F (t) dt. (3.42)
−1
Furthermore, for λ ∈ SpectL , i.e., λ = L∧ (n) = n(n + 1),

1
F (x) = Pn∗ (x) F (t)Pn∗ (t) dt
−1
+1
+ G(L + λ; x, t)(Lt + λ)F (t) dt. (3.43)
−1
In particular, for λ = 0, we have

+1
1 +1
F (x) = F (t) dt + G(L; x, t) Lt F (t) dt. (3.44)
2 −1 −1
Repeated application of the Green–Lagrange formula yields

1
G(L + λ; x, t)(Lt + λ)F (t) dt
−1
1
= G((L + λ)p ; x, t)(Lt + λ)p F (t) dt (3.45)
−1
p = 2, 3, . . ., provided that F is sufficiently often differentiable. This leads us to

the following extension of Theorem 3.1.
Theorem 3.2 (Integral Formula for (L + λ)p ). Suppose that p ∈ N, λ ∈ R, x ∈
(−1, 1), and F ∈ C (2p) ([−1, 1]).
Then, for λ ∈
/ SpectL
1
F (x) = G((L + λ)p ; x, t)(Lt + λ)p F (t) dt. (3.46)
−1
Moreover, for λ ∈ SpectL , i.e., λ = L∧ (n) = n(n + 1),

1
∗
F (x) = Pn (x) F (t)Pn∗ (t) dt
−1
1
+ G((L + λ)p ; x, t)(Lt + λ)p F (t) dt. (3.47)
−1
Finally, for m ∈ N0 and λ ∈ R with λ = L∧ (m + k), k ∈ N, we let

G⊥
0,...,m ((L + λ) ; x, t) = G((L + λ) ; x, t) − G0,...,m ((L + λ) ; x, t),
p p p
(3.48)
where

m
1
G0,...,m ((L + λ)p ; x, t) = P ∗ (x)Pk∗ (t). (3.49)
(λ − L∧ (k))p k
k=0
λ =L∧ (k)
G⊥0,...,m ((L+λ) ; ·, ·) is called mth truncated Legendre (Green) function with respect
p
to the operator (L + λ)p . These functions admit an easy reformulation of Theorem

3.2 after application of partial integration.
Theorem 3.3. Suppose that p ∈ N, m ∈ N0 . Assume that λ ∈ R with
λ = L∧ (m + 1), L∧ (m + 2), . . . .
Then, for all x ∈ (−1, 1) and F ∈ C (2p) ([−1, 1]),
m +1

F (x) = F (t)Pn∗ (t) dt Pn∗ (x)
n=0 −1
1
+ G⊥ p p ⊥
0,...,m ((L + λ) ; x, t) (Lt + λ) F0,...,m (t) dt, (3.50)
−1
where we have used the abbreviation
m 1
⊥
F0,...,m =F − F (t)Pn∗ (t) dt Pn∗ . (3.51)
n=0 −1

=F0,...,m
Theorem 3.3 enables a comparison between the functional value F (x) at the point x
and the mth truncated orthogonal expansion of F in terms of Legendre polynomials
at x with explicit representation of the remainder term in integral form.
Later on, this formula will be used to formulate adaptive estimates of the
remainder terms in Gaussian numerical integration.
Gaussian Quadrature. Let x1,n , . . . , xn,n with x1,n < · · · < xn,n be the zeros of the
Legendre polynomial Pn of degree n. Then it is well known that there is precisely
one zero of the polynomial Pn+1 in each interval (−1, x1,n ), . . . , (xn,n , 1).
For later use we want to prove the following preparatory result.
Lemma 3.5. The matrix

A = Pk∗ (xj,n ) k=0,...,n−1 (3.52)
j=1,...,n
is non-singular.
Proof. Assume that the matrix is singular. Then there exist coefficients C0 , . . .,
Cn−1 with c = (C0 , . . ., Cn−1 )T = (0, . . . , 0)T such that cT A = 0, i.e., the following
polynomial of degree ≤ n − 1

n−1
Q(x) = Cl Pl∗ (x) (3.53)
l=0
has as zeros x1,n , . . . , xn,n . Hence, Q = 0. Now, for k = 0, . . . , n − 1,

1
0= Q(x)Pk∗ (x) dx = Ck . (3.54)
−1
But this is a contradiction. Thus, Lemma 3.5 must be true.
We are now interested in the “Gaussian integration formula” involving the
“Gaussian n-point formula” Gn
1
n
I(F ) = F (t) dt Gn (F ) = wi,n F (xi,n ), (3.55)
−1 i=1
where the knots xi,n , i = 1, . . . , n, are the zeros of the Legendre polynomial Pn
and the weights wi,n , i = 1, . . . , n, are determined by the linear system
n 1 √
wi,n P0∗ (xi,n ) = P0 (t)P0∗ (t) dt = 2,
i=1 −1
n 1
wi,n P1∗ (xi,n ) = P1 (t)P0∗ (t)dt = 0,
i=1 −1
.. ..
. .

n 1
∗
wi,n Pn−1 (xi,n ) = Pn−1 (t)P0∗ (t) dt = 0. (3.56)
i=1 −1
From Lemma 3.5 it is clear that the system (3.56) is uniquely solvable in the
unknown weights w1,n , . . . , wn,n .
Central for our considerations is the following well-known theorem (see, e.g.,
[206, 226]).
Theorem 3.4. Let x1,n , . . . , xn,n be the zeros of the Legendre polynomial Pn . Fur-
thermore, let w1,n , . . . , wn,n be the (unique) solution of the linear system (3.56).
(i) Then, the weights are positive, i.e., wi,n > 0 for i = 1, . . . , n, and we have
+1
n
P (t) dt = wi,n P (xi,n ) (3.57)
−1 i=1
for all P ∈ Pol0,...,2n−1 , i.e., for all algebraic polynomials of degrees ≤ 2n− 1,
(ii) Conversely, if (3.57) is valid for real numbers wi,n , xi,n , i = 1, . . . , n, then
the knots xi,n , i = 1, . . . , n, are the zeros of the Legendre polynomial Pn and
the weights wi,n , i = 1, . . . , n, satisfy the linear equations (3.56).
(iii) Finally, there exist no real numbers xi,n , wi,n , i = 1, . . . , n, such that (3.57)
holds true for all P ∈ Pol2n .
Proof. Consider a polynomial P ∈ Pol0,...,2n−1 . Then there exist Q, R ∈ P oln−1
with

n−1
Q= Ar Pr∗ (3.58)
r=0
and

n−1
R= Br Pr∗ (3.59)
r=0
such that
P = Pn∗ Q + R. (3.60)
Because of the orthogonality of the Legendre polynomials it is clear that
+1 1
P (t) dt = P (t)P0 (t) dt
−1 −1
=1
1 1
= Q(t)Pn∗ (t) dt + R(t)P0 (t) dt
−1 −1
1
= R(t)P0 (t) dt. (3.61)
−1
In connection with (3.59) this implies

1 √
P (t) dt = 2 B0 . (3.62)
−1
Moreover, because of Pn∗ (xi,n ) = 0, i = 1, . . . , n, we obtain

n
n
n
wi,n P (xi,n ) = wi,n Pn∗ (xi,n ) Q(xi,n ) + wi,n R(xi,n )
i=1 i=1
i=1
=0

n
n−1
= wi,n Br Pr∗ (xi,n )
i=1 k=0
√
= 2 B0 . (3.63)
This shows us that (3.57) holds true.
Inserting the polynomials T̃k ∈ Pol0,...,2n−1 , k = 1, . . . , n, given by

n
T̃k (x) = (x − xj,n )2 , x ∈ [−1, 1], (3.64)
j=1
j =k
into (3.57) we get for k = 1, . . . , n

+1
n
n
n
0< T̃k (t) dt = wi,n T̃k (xi,n ) = wi,n (xi,n − xj,n )2 , (3.65)
−1 i=1 i=1 j=1
j =k
hence, it is clear that

wi,n > 0, i = 1, . . . , n. (3.66)
This proves (i).
Next we verify (iii). Suppose that wi,n , xi,n , i = 1, . . . , n, are given in such a
way that (3.57) is valid for all P ∈ Pol0,...,2n . Choose H ∈ Pol0,...,2n of the form

n
H(x) = (x − xj,n )2 . (3.67)
j=1
Then 1
n
0< H(x) dx = wi,n H(xi,n ) = 0. (3.68)
−1 i=1
This is a contradiction.
In order to guarantee (ii) we apply the Gauss rule (3.57) especially to the
Legendre polynomials Pk∗ , k = 0, . . . , n − 1, such that
n 1 1
wi,n Pk∗ (xi,n ) = Pk∗ (x) dx = Pk∗ (x) P0 (x) dx. (3.69)
−1 −1
i=1
=1
In other words, the weights have to satisfy (3.56), i.e.,

√
n
2, k = 0,
∗
wi,n Pk (xi,n ) = (3.70)
i=1
0, k = 1, . . . , n − 1.
We now consider P = Pk∗ Pn∗ for k = 0, . . . , n − 1, and apply (3.57). This gives
1 n
0= Pk∗ (t)Pn∗ (t) dt = wi,n Pk∗ (xi,n ) Pn∗ (xi,n ) (3.71)
−1 i=1
for k = 0, . . . , n − 1. Consequently, the vector a = (A1 , . . . , An )T given by

(A1 , . . . , An )T = (w1,n Pn∗ (x1,n ), . . . , wn,n Pn∗ (xn,n ))T (3.72)
satisfies the linear system

n
0= Ai Pk∗ (xi,n ), k = 0, . . . , n − 1, (3.73)
i=1
where the coefficient matrix is non-singular. But this means that Ai = 0,
i = 1, . . . , n, hence, Pn (xi,n ) = 0, i = 1, . . . , n, as required.
Remainder Terms Involving Green’s Function. Now, the integral formula (Theo-
rem 3.3) comes into play. For all values λ ∈ R\{L∧ (2n), L∧ (2n + 1), . . .} and for
functions F ∈ C (2p) ([−1, +1]), p ∈ N, we have
n 1
wi,n F (xi,n ) = F (t) dt
i=1 −1

n 1
+ wi,n G⊥ p p
0,...,2n ((L + λ) ; xi,n , t) (Lt + λ) F (t) dt.
i=1 −1
(3.74)
In other words, the remainder

En (F ) = I(F ) − Gn (F ), (3.75)
when the integral
1
I(F ) = F (t) dt, F ∈ C (2p) ([−1, 1]), (3.76)
−1
is replaced by the “Gaussian n-point formula”

n
Gn (F ) = wi,n F (xi,n ), (3.77)
i=1
is expressible by means of the 2nth truncated Legendre (Green) function in integral

form
n 1
En (F ) = − wi,n G⊥ p p
0,...,2n ((L + λ) ; xi,n , t) (Lt + λ) F (t) dt (3.78)
i=1 −1
∧ ∧
for all λ ∈ R\{L (2n), L (2n + 1), . . .}.
Applying the Cauchy–Schwarz inequality to (3.78) we obtain

1
(p)
|En (F )| ≤ Aλ (n) |(Lt + λ)p F (t)|2 dt, (3.79)
−1
for λ ∈ R\{L∧ (2n), L∧ (2n + 1), . . .}, where we have used the abbreviation
(p)

n
n
Aλ (n) = wi,n wj,n G⊥ 2p
0,...,2n ((L + λ) ; xi,n , xj,n ) (3.80)
i=1 j=1
with
G⊥ 2p
0,...,2n ((L + λ) ; xi,n , xj,n )
1
= G⊥ p ⊥ p
0,...,2n ((L + λ) ; xi,n , t)G0,...,2n ((L + λ) ; t, xj,n ) dt. (3.81)
−1
Summarizing our results we obtain the following remainder estimate.
Theorem 3.5. Let xi,n , i = 1, . . . , n, be the zeros of the Legendre polynomial Pn ,
n ≥ 1, and let wi,n , i = 1, . . . , n, be the unique solution of the linear system
n 1
wi,n Pk (xi,n ) = P0 (t)Pk (t) dt, (3.82)
i=1 −1
k = 0, . . . , n − 1.
Then, for values λ ∈ R\{L∧ (2n), L∧ (2n + 1), . . .} and for functions F ∈
(2p)
C [−1, 1], we have

1
n 1
(p)
F (t) dt − wi,n F (xi,n ) ≤ Aλ (n) |(Lt + λ)p F (t)|2 dt, (3.83)
−1 −1
i=1
where

n
n ∞

(p) 1
Aλ (n) = wi,n wj,n P ∗ (xi,n )Pk∗ (xj,n ). (3.84)
i=1 j=1
(λ − L∧ (k))2p k
k=2n
Remainder Term Estimation. From (3.78), we get the estimate

1 2
(p)
|En (F )| ≤ Aλ (n) (Lt + λ)F ⊥ dt,
0,...,2n−1 (t) (3.85)
−1
where
1
2n−1
⊥
F0,...,2n−1 (t) = F (t) − F (u)Pn∗ (u) du Pn∗ (t), (3.86)
n=0 −1

=F0,...,2n−1
t ∈ [−1, 1].
Observing the estimate |Pn (t)| ≤ 1 for all n ∈ N0 and t ∈ [−1, 1] and the
properties of the Gaussian weights we get
∞ ∞
1
n n
2k + 1 2k + 1
|Apλ (n)| ≤ wi,n wj,n =2 (3.87)
2 i=1 j=1 (λ − L∧ (k))2p (λ − L∧ (k))2p
k=2n k=2n
for all λ ∈ R\{L∧ (2n), L∧ (2n + 1), . . .}.

Theorem 3.6. For F ∈ C (2n) ([−1, 1]),
1
n
F (t) dt − wi,n F (xi,n )
−1 i=1
1
n

= wi,n G⊥ n n ⊥
2n (L + λ) ; xi,n , t (Lt + λ) F0,...,2n−1 (t) dt, (3.88)
i=1 −1
⊥
where F0,...,2n−1 (t) is defined by (3.86).
Convergence of Gaussian Quadrature. From Theorem 3.5 we obtain

1
n 1
(p)
F (t) dt − wi,n F (xi,n ) ≤ A0 (n) (Lt F (t))2 dt, (3.89)
−1 −1
i=1
where

n
n ∞

(p) 1
A0 (n) = wi,n wj,n P ∗ (xi,n )Pk∗ (xj,n ). (3.90)
i=1 j=1
(k(k + 1))2p k
k=2n
The sequence ∞
2k + 1
2 (3.91)
(k(k + 1))2p
k=2n n∈N
is (strictly) monotonously decreasing and positive. Furthermore, we have the rough

estimate
∞
(p) 2k + 1
0 ≤ A0 (n) ≤ 2 . (3.92)
(k(k + 1))2p
k=2n
Therefore, we come to the conclusion.
(p)
Theorem 3.7. The sequence {Ao (n)}n∈N has the limit
lim Ap0 (n) = 0 (3.93)
n→∞
such that, for all F ∈ C (2p) ([−1, 1]),

lim En (F ) = 0, (3.94)
n→∞
i.e., the Gaussian quadrature rule is convergent:
n 1
lim wi F (xi,n ) = F (t) dt. (3.95)
n→∞ −1
i=1
Comparing different numerical integration techniques with the Gaussian method

we see that the last method assures the highest degree of polynomial precision. The
critical point of the estimate (3.85), however, is to find an estimate for (L+λ)2p F (t)
with (L+λ)2p being appropriately adapted to the specific properties of the function
such that the bound of |En (F )| becomes small. Moreover, it should be critically
mentioned that the pointsets {x1,n , . . . , xn,n }n∈N are not hierarchically distributed.
Gauss’s work was simplified by Jacobi and further developed by F.G. Mehler
[149], E.B. Christoffel [19], and others through much of the 19th century. Eventu-
ally, there emerged a coherent theory which received its first systematic expositions
by E.B. Christoffel [19], R. Radau [177], and E. Heine [106] in his book on spherical
functions. An outline of the developments in numerical quadrature that took place
in a period of approximately 200 years from Newton via Gauss to early develop-
ments in the 20th century can be found in [188], and a German edition of the four
principal memoirs (of Newton, Cotes, Gauss, and Jacobi) is due to A. Kowalewski
[129]. A more detailed survey on developments up to the second half of the 20th
century is given by W. Gautschi [85], culminating in the following comment:
“Gauss’s discovery must be rated as one of the most significant events of
the 19th century in the field of numerical integration and perhaps in all
of numerical analysis. The result not only has great beauty and power,
but also influenced many later developments in computing and approxi-
mation. It soon inspired contemporaries, such as Jacobi and Christoffel,
to perfect Gauss’s method and to develop it into new directions. Towards
the end of the century, it inspired K. Heun [109] to generalize Gauss’s
idea to ordinary differential equations, which in turn led to significant
developments in the numerical solution of differential equations, notably
the discovery of the Runge–Kutta method (cf. [133]). Gauss’s influence
continues into the 20th century and is still felt today, . . . .”
Whether or not the Gauss method had actually been widely used in practice
up to the seventieth of the last century is a matter of some doubt, since the
method requires the evaluation of functions at irrational arguments, hence, tedious
interpolation. All this changed when powerful digital computers entered the scene,
which generated a phase of renewed interest in Gaussian quadrature. The formulas
began to be routinely applied, and increased usage, in turn, led to important new
theoretical developments.
3.2. Periodic approximate integration
Next we are interested in trapezoidal sums based on Euler-type summation for-
mulas for approximate integration of one-dimensional periodic functions.
To this end we first list some obvious results on “τ -dilated lattices”
τ Z = {τ g : g ∈ Z, τ > 0}. (3.96)
The fundamental cell Fτ Z of the lattice τ Z is given by

1 1
Fτ Z = x ∈ R : − τ ≤ x < τ . (3.97)
2 2
A function F : R → C is called τ Z-periodical if F (x + g) = F (x) holds for all
x ∈ Fτ Z and g ∈ τ Z. The function Φh : R → C, h ∈ τ1 Z, given by
1
x → Φh (x) = √ e2πihx (3.98)
τ
is τ Z-periodical and satisfies the identity

1, h = h
Φh (x)Φh (x) dx = (3.99)
Fτ Z h .
0, h =
(m)
The space of all F ∈ C (m) (R) that are τ Z-periodical is denoted by Cτ Z (R), 0 ≤
(0)
m ≤ ∞. Clearly, the space L2τ Z (R) is the completion of Cτ Z (R) with respect to the
norm · L2τ Z (R) . The system {Φh }h∈ τ1 Z is orthonormal with respect to the L2τ Z (R)-
inner product. By convention, we say that λ is an eigenvalue of the lattice τ Z with
respect to the operator Δ of the second-order derivative (i.e., the one-dimensional
Laplace operator), if there is a non-trivial solution U of the differential equation
(Δ + λ)U = 0 satisfying the “boundary condition” of periodicity U (x + g) = U (x)
for all x ∈ Fτ Z and g ∈ τ Z. From classical Fourier analysis we know that the
operator Δ has a half-bounded and discrete eigenspectrum {Δ∧ (h)}h∈ τ1 Z ⊂ R
such that (Δx + Δ∧ (h)) Φh (x) = 0, x ∈ Fτ Z , with eigenvalues Δ∧ (h) given by
1
Δ∧ (h) = 4π 2 h2 , h ∈ τ1 Z, and eigenfunctions Φh (x) = τ − 2 e2πihx , h ∈ τ1 Z, x ∈ Fτ Z .
Consequently, the eigenspectrum of the operator Δ (with respect to τ Z) is
given by
1
SpectΔ (τ Z) = Δ∧ (h) : Δ∧ (h) = 4π 2 h2 , h ∈ Z . (3.100)
τ
Trapezoidal Integration Rule. For purposes of numerical integration of one-dim-
ensional periodic functions we introduce the definition of the τ Z-lattice Green
function with respect to the operator Δ + λ, λ ∈ R. It can be given in canonical

way in the sense of the classical Hilbert approach (cf. [111]):
Definition 3.4. A function G(τ Z; Δ + λ; ·) : R → R is called the Green function

for the Helmholtz operator Δ + λ, λ ∈ R, with respect to the lattice τ Z (in brief,
τ Z-lattice Green function for Δ + λ), if it fulfills the following properties:
(i) (Periodicity) G(τ Z; Δ + λ; ·) is continuous in R, and
G(τ Z; Δ + λ; x + g) = G(τ Z; Δ + λ; x) (3.101)
for all x ∈ R and g ∈ τ Z.

(ii) (Differential Equation) G(τ Z; Δ + λ; ·) is twice continuously differentiable for
all x ∈
/ τ Z with
(Δx + λ)G(τ Z; Δ + λ; x) = 0 (3.102)
provided that λ ∈ / SpectΔ (τ Z),
G(τ Z; Δ + λ; ·) is twice continuously differentiable for all x ∈
/ τ Z with
1
(Δx + λ)G(τ Z; Δ + λ; x) = − e2πihx (3.103)
τ ∧
λ−(Δ) (h)=0
h∈ τ1 Z
provided that λ ∈ SpectΔ (τ Z) (note that the summation on the right side of
(3.103) is to be taken over all lattice points h ∈ τ1 Z satisfying λ − (Δ)∧ (h)) =
0, i.e., 4π 2 h2 = λ).
(iii) (Characteristic Singularity)
1
x → G(τ Z; Δ + λ; ·) − x sign(x) (3.104)
2
is continuously differentiable for all x ∈ Fτ Z.
(iv) (Normalization) For all h ∈ τ1 Z with (Δ + λ)∧ (h) = 0,

G(τ Z; Δ + λ; x)e2πihx dx = 0. (3.105)
Fτ Z
The Fourier expansion reads

1 1
G(τ Z; Δ + λ; x) = e2πihx , x ∈ R. (3.106)
τ λ − (Δ)∧ (h)
λ−(Δ)∧ (h) =0
h∈ τ1 Z
By ordinary partial integration we obtain
Theorem 3.8 (τ Z-Euler Summation Formula for the Helmholtz Operator Δ + λ,

λ ∈ R). Let F be of class C(2) ([a, b]), a < b. Suppose that x is a point of Fτ Z .
Then, the following identity holds true:

b
F (g + x) = F (y)e2πihy dy e2πihx
a
g+x∈[a,b] Δ∧ (h)=λ
g∈τ Z h∈ τ1 Z
b
+ G(τ Z; Δ + λ; x − y)(Δy + λ)F (y) dy (3.107)
a
+ {F (y)(∇y G(τ Z; Δ + λ; x − y)) − (∇y F (y))G(τ Z; Δ + λ; x − y)} |ba ,
where the sum on the right side is to be taken over all points h ∈ τ1 Z for which
λ− (Δ)∧ (h) = 0. In case of λ− (Δ)∧ (h) = 0 for all h ∈ τ1 Z, this sum is understood
to be zero.
The case λ = 0 leads back to the concept for the Laplace operator, i.e., the
operator of the second derivative

1 b
F (g + x) = F (y) dy
τ a
g+x∈[a,b]
g∈τ Z
b
+ G(τ Z; Δ; x − y)(Δy )F (y) dy
a
+ {F (y)(∇y G(τ Z; Δ; x − y)) − (∇y F (y))G(τ Z; Δ; x − y)} |ba , (3.108)
(note that τ = 2π provides the Euler summation formula (see [33]) in its original
form).
As an immediate consequence, a (generalized) trapezoidal rule of the follow-
ing kind

τ F (g + x) − τ {F (y)(∇y G(τ Z; Δ; x − y)) − (∇y F (y)) G(τ Z; Δ; x − y)} |ba
g+x∈[a,b]
g∈τ Z
b
F (y) dy (3.109)
a
holds true, where the remainder term
b
τ G(τ Z; Δ; x − y)Δy F (y) dy (3.110)
a
is of the order O(τ 2 ). Note that the formula requires the knowledge of F (y) and
∇y F (y) at the endpoints a and b of the interval [a, b]. Especially, for x = 0, we
have the classical rule

τ F (g) − τ {F (y)(∇y G(τ Z; Δ; y)) − (∇y F (y)) G(τ Z; Δ; y)} |ba
g∈[a,b]
g∈τ Z
b
F (y) dy. (3.111)
a
Under the assumption λ ∈ R\{(Δ)∧ (m), (Δ)∧ (m + 1), . . . , m ∈ N} we intro-

duce
1 1
G⊥
0,...,m−1 (τ Z; Δ + λ; x) = G(τ Z; Δ + λ; x) − e2πihx .
τ λ − (Δ)∧ (h)
λ−(Δ)∧ (h) =0
h≤m−1
h∈ τ1 Z (3.112)
The Fourier series of G⊥ 0,...,m−1 (τ Z; Δ + λ; x), x ∈ R, is given by
1 1
G⊥0,...,m−1 (τ Z; Δ + λ; x) = ∧ (h)
e2πihx , x ∈ R. (3.113)
τ ∧
λ − (Δ)
(Δ+λ) (h) =0
h≥m
h∈ τ1 Z
By partial integration we therefore obtain from Theorem 3.8

Corollary 3.1 (Extended Trapezoidal Rule). Let x be an arbitrary point of R.
Suppose that F is of class C(2) ([a, b]). Then, for values
λ ∈ R\{(Δ)∧ (m), (Δ)∧ (m + 1), . . .}, m ∈ N,

1 b
F (g + x) = F (y) dy
τ a
b
g+x∈[a,b]
g∈τ Z
1
+ F (y)e2πihy dy e2πihx
τ a
0≤h≤m−1
Δ∧ (h)=λ
h∈ τ1 Z
! " b
+ F (y)∇y G⊥ 0,...,m−1 (τ Z; Δ + λ; x − y) |a
! " b
− ∇y F (y) G⊥ 0,...,m−1 (τ Z; Δ + λ; x − y) |a
b
+ G⊥
0,...,m−1 (τ Z; Δ + λ; x − y)(Δy + λ)F (y) dy. (3.114)
a
Quadrature of Periodic Functions. Next we turn to the well-known approximate

integration of 2π-periodic functions. More explicitly, we choose a = 0 and b =
(2)
2π. Moreover, we let F be of class C2π (R). Then we are able to deduce from
2π
Corollary 3.1 that, for τ = m , m ∈ N,
2π
2π
F (g) = F (y) dy
m 0
g∈[0,2π]
g∈ 2π
mZ 2π
+ F (y)e2πihy dy e2πihx
0≤h≤m−1 0
4π 2 h2 =λ
m
h∈ 2π Z
2π
2π
+ G⊥
0,...,m−1 Z; Δ + λ; y (Δy + λ)F (y) dy (3.115)
0 m
holds for all λ ∈ R\{Δ∧ (m), Δ∧ (m + 1), . . .}. Because of the 2π-periodicity of F
we are able to write

2π 2π
m
2π
F (g) = F (k − 1) . (3.116)
m m m
g∈[0,2π] k=1
g∈ 2π
m Z
Furthermore, we notice that the integration rule

2π
2π 2π
m m−1
2π 2π
F (y) dy = F (k − 1) = F k (3.117)
0 m m m m
k=1 k=0
holds true for all trigonometric polynomials F up to degree m − 1 in the interval

[0, 2π] (see, e.g., [206]). In order to guarantee this it is sufficient to verify that
(3.117) will be exact for the functions x → eilx , x ∈ R, l = 1, . . . , m − 1.
For m = 0, the assertion is evidently true. Choosing l ∈ {1, . . . m − 1} ⊂ N
we obtain 2π
1 2πil
eily dy = e −1 =0 (3.118)
0 il
and
m
2π
2π
eilm 2m − 1 e2πil − 1
eil((k−1) 2m ) = il 2π = il 2π = 0. (3.119)
k=1
e 2m − 1 e 2m − 1
As an immediate consequence of (3.115) we therefore find in combination with our
integral formulas that
2π 2π 2πk
m−1
F (y) dy = F
0 m m
k=0
2π
2π
− G⊥0,...,m−1 Z; Δ + λ; y (Δy + λ)F (y) dy (3.120)
0 m
(2)
is valid for all F ∈ C2π (R) and λ ∈ R\{m2 l2 , m2 (l + 1)2 , . . . , l ∈ N}.
The remainder term
2π
⊥ 2π
Em−1 (F ) = − G0,...,m Z; Δ + λ; y (Δy + λ)F (y) dy (3.121)
0 m
of the quadrature depends on the properties of the integrand F . It provides an
appropriate representation if the parameter λ can be chosen such that (Δ + λ)F
is small (with respect to a certain topology).
An estimate in the sense of A. Sard [190] is obtainable via the L22π ([0, 2π])-
topology
2π
⊥ (2)
2π
|Em−1 (F )| ≤ (G0,...,m ) Z; Δ + λ, 0 |(Δy + λ)F (y)|2 dy, (3.122)
m 0
where
2
2π m 1
(G⊥
0,...,m−1 )
(2)
Z; Δ + λ; 0 = eimlx , x ∈ R. (3.123)
m 2π λ − m2 l 2
m2 l2 =λ
l≥m
3.3. Latitude-longitude constituted spherical harmonics

We begin our considerations with some basic background on the convention-
ally used associated Legendre spherical harmonic system in physical geodesy (see,
e.g., [107], these and other types of spherical harmonic systems can be found in
[46, 161]): Let Hn be a homogeneous, harmonic polynomial of degree n with the
following properties:
(i) Hn (tx) = Hn (x) for all orthogonal transformations t leaving ε3 = (0, 0, 1)T
fixed, i.e., tε3 = ε3 ,
(ii) Hn (ε3 ) = 1.
Then it is known that Hn is uniquely determined by
Hn (x) = rn Pn (t), (3.124)
with
x = rξ, ξ ∈ Ω = {ξ ∈ R3 : |ξ| = 1}, (3.125)
(note that we use Ω instead of S2 for the unit sphere in R3 )

#
ξ = tε3 + 1 − t2 (cos λε1 + sin λε2 ), (3.126)
t = cos θ, t ∈ [−1, 1], λ ∈ [0, 2π), (3.127)
where

n
n−k
Pn (t) = C n−k (1 − t2 ) 2 tk (3.128)
2
k=0
and
0 , n − k odd,
C n−k = n−k (3.129)
2 (− 14 ) 2 (( n−kn!)!)2 k! , n − k even.
2
Equivalently, we have
2 n
l
1 (1 − t2 )l tn−2l
Pn (t) = n! − . (3.130)
4 (l!)2 (n − 2l)!
l=0
By a straightforward calculation we are able to introduce from

m
2 m/2 d
Pn,m (t) = (1 − t ) Pn (t), t ∈ [−1, 1], (3.131)
dt
the so-called associated Legendre function of degree n and order m

n−m
2
(2n − 2k)!
2 m/2
Pn,m (t) = (1 − t ) (−1)k tn−m−2k (3.132)
2n k!(n − k)!(n − m − 2k)!
k=0
(note that Pn is equal to Pn,0 ). In connection with Pn,m = 0 for m > n, the
preceding result leads to the following statement (see, e.g., [46]).
Lemma 3.6 (Associated Legendre Polynomial of Degree n and Order m). For
n = 0, 1, . . . , m = 0, 1, . . ., and t ∈ [−1, 1] we have
m

n
n−m−k
Pn,m (t) = (1 − t2 ) 2 Cm 2
n−m−k (1 − t )
2 tk , (3.133)
2
k=0
where the generating coefficients C m n−m−k of the associated Legendre polynomial of

2
degree n and order m are given by

(n+m)!
m ( 12 )m (n−m)!m! C n−m−k , n − m − k even, 0 ≤ k ≤ n − m
C n−m−k = 2
2 0 , otherwise. (3.134)
The associated Legendre functions Pn,m , n = 0, 1, . . ., m = 1, . . . , n, satisfy
the differential equation
2
d d m2
(1−t2 ) Pn,m (t)−2t Pn,m (t)+ n(n + 1) − Pn,m (t) = 0. (3.135)
dt dt 1 − t2
For every m = 1, 2, . . ., the system
1/2
2n + 1 (n − m)!
Pn,m
2 (n + m)!
n=m,m+1,...
2
is a complete orthonormal system in L ([−1, 1]).
The functions G, H defined by
G : t → G(t) = Pn,j (t), t ∈ (−1, 1), (3.136)

cos(jλ)
H : λ → H(λ) = , λ ∈ [0, 2π), (3.137)
sin(jλ)
respectively, satisfy the differential equations

2 j2
(1 − t )G (t) − 2tG (t) + n(n + 1) − G(t) = 0, (3.138)
1 − t2
H (λ) + j 2 H(λ) = 0. (3.139)
Therefore, the functions LR R
n,1 , . . . , Ln,2n+1 ∈ C
(∞) 2
(S ) given by

Pn,|j| (t) cos(jλ) , j = −n, . . . , 0
LR
n,j (ξ) = (3.140)
Pn,j (t) sin(jλ) , j = 1, . . . , n
satisfy the differential equation (for graphical illustrations see Figures 3.2, 3.3,
and 3.4),
(Δ∗ξ + (Δ∗ )∧ (n)) LR

n,j (ξ) = 0, ξ ∈ S2 , (Δ∗ )∧ (n) = n(n + 1), (3.141)
j = −n, . . . , n. In addition, the functions LR

n,j ∈ C
(∞)
(R3 ), j = −n, . . . , n, given by
LR n R
n,j (x) = |x| Ln,j (ξ), x = |x|ξ, ξ ∈ Ω, (3.142)
form homogeneous harmonic polynomials of degree n in R3 .
Definition 3.5. Let LR R

n,j , j = −n, . . . , n, be defined by (3.140). Then, Ln,j is called
associated Legendre (spherical) harmonic of degree n and order j. Correspondingly,
the system {Yn,j }j=−n,...,n given by
Yn,j = Cn,j LR
n,j , j = −n, . . . , n, (3.143)
with
2n + 1 (n − |j|)!
Cn,j = (2 − δj,0 ) (3.144)
4π (n + |j|)!
is called (fully) L2R (S2 )-orthonormal system of associated Legendre (spherical) har-
monics.
In terms of associated Legendre harmonics, the addition theorem allows the

following reformulation that is standard in all geosciences (see, e.g., [46, 159]).
Remark 3.1 (Addition Theorem for the system {Yn,j }). Suppose that ξ, η ∈ Ω are
given by

ξ= 1 − t2ξ cos λξ ε1 + 1 − t2ξ sin λξ ε2 + tξ ε3
− 1 ≤ tξ ≤ 1, tξ = cos θξ , 0 ≤ λξ < 2π, (3.145)

η = 1 − t2η cos λη ε1 + 1 − t2η sin λη ε2 + tη ε3
− 1 ≤ tη ≤ 1, tη = cos θη , 0 ≤ λη < 2π, (3.146)
respectively, so that

ξ · η = tξ tη +1 − t2ξ 1 − t2η (cos λξ cos λη + sin ϕξ sin λη )

= tξ tη + 1 − t2ξ 1 − t2η cos(λξ − λη ). (3.147)
Then we have
2n + 1
Pn (tξ tη + 1 − t2ξ 1 − t2η cos(λξ − λη ))
4π
1
= Pn (tξ )Pn (tη )
4π
2n + 1 (n − m)!
n
+ Pn,m (tξ )Pn,m (tη ) cos(m(λξ − λη ))
2π m=1 (n + m)!

n
= Yn,j (ξ)Yn,j (η). (3.148)
j=−n
In other words, summing up all spherical harmonics involving associated

Legendre functions via the addition theorem leads (apart from a multiplicative
factor) to the orthogonal invariant Legendre (kernel) functions.
Figure 3.2. Zonal (j = 0) spherical harmonics of different degrees

1, . . . , 6 (from left to right). The black and white color indicate the
zones of different signs of the function, respectively.
Figure 3.3. Tesseral (k = ±n) spherical harmonics of degree n = 4.

The black and white colors indicate the zones of different signs of the
function, respectively.
Figure 3.4. Sectorial (k = ±n) spherical harmonics of different de-

grees. The black and white colors indicate the zones of different signs of
the function, respectively.
Definition 3.6 (Complex-Valued Spherical Harmonics). Let n ∈ N0 , j ∈ Z, with

−n ≤ j ≤ n. The function

C 2n + 1 (n − j)!
ξ → Yn,j (ξ) = (−1)j
Pn,j (cos(θ)) eijλ
4π (n + j)!
is called (complex) associated Legendre spherical harmonic of degree n and order
j, where θ, λ are the spherical coordinates of ξ (note that i denotes the imaginary
unit with i2 = −1).
These spherical harmonics are orthonormal with respect to the canonical
scalar product of the space L2C (Ω) of complex-valued square-integrable functions
on the unit sphere Ω. Their addition theorem may be rewritten as follows

n
2n + 1
C (ξ)Y C (η) =
Yn,j Pn (ξ · η). (3.149)
n,j
j=−n
4π
For further details on this representation of spherical harmonics the reader is

referred to, e.g., [50] and [161].
For n ∈ N0 , j ∈ Z with j = −n, . . . , n,
⎧√
2n + 1 (n − |j|)! ⎨ 2 cos(jϕ) , j<0
ξ → Yn,j (ξ) = Pn,|j| (cos(θ)) √ 1 , j=0 (3.150)
4π (n + |j|)! ⎩
2 sin(jϕ) , j>0
can be represented in the form
⎧√
⎪
⎨ 2−δ0,j Y C (ξ) + Y C (ξ) ,
2 n,j n,j j ≤ 0,
ξ → Yn,j (ξ) = √ (3.151)
⎪ (−1)
⎩ 2i
j
2 C C (ξ) ,
Yn,j (ξ) − Yn,j j > 0,
for all ξ ∈ Ω, n ∈ N0 , and j ∈ Z with j = −n, . . . , n (note that in case of the

real-valued spherical harmonics the indexing with negative integers is just one
possibility to distinguish the two types with sine and cosine).
3.4. Latitude-longitude integration

As is well known longitude-latitude point sets on the unit sphere Ω ⊂ R3 enable
approximate integration weights being available in a comparatively easy and fast
way, thereby establishing spherical harmonics exact cubature formulas which are
of tremendous importance, e.g., in mathematical geodesy.
In what follows we are first concerned with the arrangement of equiangular
point sets. Systems of this type can be gained by suitably dividing [0, π) × [0, 2π]
into longitude-latitude grids. Clearly, there are numerous possibilities for a choice
of locations concerning the type and number of latitude as well as longitude dis-
tributions. Nonetheless, these choices are decisive for the quality of the resulting
integration formulas.
For application of longitude-latitude grid integration, suppose that the spher-

ical harmonics are represented in the following (complex) way

C 2n + 1 (n − j)!
Yn,j (ξ) = (−1)j Pn,j (cos(θ)) eijλ , (3.152)
4π (n + j)!

j
=Cn
√
n = 0, 1, . . . , j = −n, . . . , n and ξ = tε3 + 1 − t2 (cos λε1 + sin λε2 ), t = cos θ, t ∈
[−1, 1], λ ∈ [0, 2π).
On the one hand, taking into account that Y0,0 C
= √14π , we obtain
√
C 4π , if k = l = 0,
Yk,l (ξ) dS(ξ) = (3.153)
S2 0 , else
(dS denotes the surface element). On the other hand, it can be readily seen that
π 2π
C
Yk,l (ξ) dS(ξ) = Ckl Pk,|l| (cos(θ)) sin(θ) dθ (cos(lλ) + i sin(lλ)) dλ.
S2 0 0
(3.154)
The combination of (3.153) and (3.154) leads us to the equation
π 2π 2π
Ckl Pk,|l| (cos(θ)) sin(θ) dθ cos(lλ) dλ + i sin(lλ) dλ
0 0 0
√
4π , if k = l = 0,
= (3.155)
0 , else.
This identity is the point of departure for the derivation of exact integration for-
mulas. Together with C00 = √14π we easily see that
π 2π 2π
Pk,|l| (cos(θ)) sin(θ) dθ cos(lλ) dλ + i sin(lλ) dλ
0 0 0
√
4π , if k = l = 0,
= (3.156)
0 , else.
3.5. Cubature based on periodic integration

In accordance with (3.154) we are able to replace the integral on the whole unit
sphere by latitude-longitude integration. As a consequence, the desired cubature
rule on the sphere can be based on already known approaches to numerical inte-
gration from one dimension. In order to establish exact integration formulas we
pursue the following three-step strategy:
(i) Determine a number Nλ ∈ N, the nodes λ0 , . . . , λNλ −1 ∈ [0, 2π), and the
weights aλ0 , . . ., aλNλ −1 ∈ R, so that the integration formula
2π 2π λ −1
N λ −1
N
cos(lλ)dτ + i sin(lλ) dλ = aλj cos(lλj ) + i aλj sin(lλj )
0 0 j=0 j=0

2π , if l = 0,
= (3.157)
0 , else
is exact for l = 0, . . . , m. Evidently, (3.157) is fulfilled for l = 0 if the longitude

λ is discretized according to (i).This is why the integral
π
Pk,|l| (cos(θ)) sin(θ)dθ
0
is only significant for l = 0 and k = 0, . . . , m.

(ii) Concerning the discretization of the latitude θ it is sufficient to consider the
following integration rule:
Determine a number Nθ ∈ N, the nodes θ0 , . . . , θNθ −1 ∈ [0, π], and the weights
aθ0 , . . . , aθNθ −1 ∈ R, so that
π θ −1
N
Pk,0 (cos(θ)) sin(ϑ)dθ = aθj Pk (cos(θj )) (3.158)
0 j=0
=Pk (cos(θ))
is exact for k = 0, . . . , m.
(iii) Combine the latitude-longitude nodal systems to the product set
(θk , λj ) ∈ [0, π) × [0, 2π), k = 0, . . . , Nθ − 1, j = 0, . . . , Nλ − 1, (3.159)
and collect the integration weights aλ0 , . . . , aλNλ −1 and aθ0 , . . . aθNθ −1 .
We begin with the discretization with respect to the longitude that is well-known
from (3.120).
Lemma 3.7 (Longitude Integration). For Nλ ∈ N, assume that the nodes are given
2π
by λj = j N λ
∈ [0, 2π), j = 0, . . . , Nλ − 1. Then, the integration rule
2π λ −1
N
2π 2πj
T (λ) dλ = T (3.160)
0 j=0
Nλ Nλ
holds true for all trigonometric functions T of degree ≤ Nλ − 1.
In order to discretize the latitude θ, however, it is evident that the integration

weights cannot be expected to be equal for all samples. In fact, there are several
possibilities to determine point sets and weights which fulfill (3.157). The methods
presented here reflect straightforward numerical application.
In accordance with (3.156) we are interested in constructing an integral formula

of the form
π θ −1
N
Pk (cos(θ)) sin(θ) dθ = aθj Pk (cos(θj )) (3.161)
0 j=0
for k = 0, . . . , m. Since we already know that the set {P0 , . . . , PNθ }, Nθ ∈ N, rep-
resents a Tschebyscheff system in the uni-variate sense (see also [23]), integration
formulas of the type (3.157) exist if the angles θi , i = 0, . . . , Nθ − 1, are chosen
to be pairwise distinct. In particular, we can think of an equiangular distribution
of the samples on latitudes. As a matter of fact, we briefly explain an integra-
tion technique whose origin dates back to J.R. Driscoll, R.M. Healy [26]. For that
purpose we notice that, for an even Nθ ∈ N,
π
1 π
Pk (cos(θ)) sin(θ) dθ = Pk (cos(θ)) sin(θ)sgn(θ) dθ, k = 0, . . . , Nθ − 1.
0 2 −π
(3.162)
Remembering the well-known sign expansion (see, e.g., [144])
∞
4 1
sgn(θ) = sin((2l + 1)θ), θ ∈ [−π, π]. (3.163)
π 2l + 1
l=0
we obtain for k = 0, . . . , Nθ − 1
π ∞
2 π 1
Pk (cos(θ)) sin(θ) dθ = Pk (cos(θ)) sin(θ) sin((2s + 1)θ) dθ.
0 π s=0 −π
2s +1
(3.164)
It is clear that, for k = 0, . . . , Nθ − 1, the term Pk (cos(θ)) sin(θ) represents a
trigonometric function of degree ≤ Nθ . Due to the orthogonality relations we
readily find that
π
Nθ
Pk (cos(θ)) sin θ sin((2s + 1)θ) dθ = 0, s > − 1. (3.165)
−π 2
Hence, the series (3.164) reduces to
2 −1 π
Nθ
π
2 1
Pk (cos(θ)) sin(θ) dθ = Pk (cos(θ)) sin(θ) sin((2s + 1)θ) dθ.
0 π s=0 −π 2s + 1
(3.166)
The integrand on the right side is a trigonometric function of order 2Nθ − 1, hence,
it is not difficult to show that the identity
2 −1
Nθ π
2 1
Pk (cos(θ)) sin(θ) sin((2s + 1)θ) dθ (3.167)
π s=0 2s + 1
−π
Nθ −1
2 −1
Nθ

2 π π 1 π
= Pk cos j sin j sin (2s + 1) j
Nθ Nθ Nθ s=0 2s + 1 Nθ
j=−Nθ
holds true for k = 0, . . . , Nθ − 1. Moreover, the symmetry of the trigonometric

polynomials enables us to rewrite the first sum on the right-hand side of (3.167)
as follows
π
Pk (cos (θ)) sin(θ) dθ (3.168)
0
Nθ −1
2 −1
Nθ
2 π π 1 π
= Pk cos j sin j sin (2s + 1)j
Nθ j=0 N −θ N −θ s=0
2s − 1 Nθ
for k = 0, . . . , Nθ − 1. Finally, taking into account that the system of Legendre

polynomials Pk , for k = 0, . . . , Nθ − 1, forms a basis of the space of all polynomials
of degree less or equal to Nθ − 1 we arrive at following result.
Theorem 3.9 (Latitude Integration). Let Nθ ∈ N be an even number. Suppose that
θj = j Nπθ , j = 0, . . . , Nθ − 1. Then, the integration rule
π +1 θ −1
N
P (cos(θ)) sin(θ) dθ = P (t) dt = aθj P (cos(θj )) (3.169)
0 −1 j=0
with
Nθ
2 −1

4 π 1 π
aθj = sin j sin (2s + 1)j (3.170)
Nθ Nθ s=0 2s + 1 Nθ
holds true for all polynomials of degrees ≤ Nθ − 1.
Now, especially for Nθ = Nλ = m + 1, m ∈ N odd, the combination of the above
schemes for longitude-latitude distribution enables us to arrange an equiangular
longitude-latitude distributed grid (ELLG) integration rule, where the nodes and
weights, respectively, are given by
π 2π
(θj , λk ) ∈ [0, π)×[0, 2π), θj = j , j = 0, . . . , m, λk = k , k = 0, . . . , m,
m+1 m+1
(3.171)
and
m+1
2 −1

4 π 1 π
aθj = sin j sin (2s + 1)j , j = 0, . . . , m,
m+1 m+1 s=0
2s + 1 m+1
2π
aλk = , k = 0, . . . , m. (3.172)
m+1
Figure 3.5 gives a graphical illustration of the equiangular longitude-latitude grid
(ELLG) in the (θ, λ)-plane. As a consequence, any spherical harmonic Y of degree
≤ m, i.e., Y ∈ Harm 0,...,m can be integrated exactly in the form
m m
Y (η) dS(η) = aθj aλk Y (θj , λk ). (3.173)
Ω j=0 k=0
Figure 3.5. The ELLG longitude-latitude grid with m = 15.
In particular, if a product F Y is of class Harm 0,...,m , then the inner product

F, Y L2 (Ω) can be evaluated exactly by means of a proposed latitude-longitude
grid based on (m + 1)2 points.
3.6. Cubature based on Gaussian integration

From the identity (3.156) we know that integration involving spherical harmonics
along spherical longitudes and latitudes, respectively, can be treated independently
from each other. Thus, classical Gaussian quadrature can also be used to construct
an efficient cubature formula for latitude approximation.
We first choose the longitude λ ∈ [0, 2π) to be discretized in an equiangular way.
Our purpose is to reduce the number Nθ of the required latitudes in relation to the
already discussed equiangular latitude-longitude grid (ELLG) by use of Gaussian
quadrature.
Figure 3.6. Gaussian Grid (GG) with m = 15.
As a consequence, let Nθ and Nλ , as usual, denote the number of spherical latitudes

and longitudes, respectively, to be involved. Assume, especially, that m ∈ N is odd,
and Nθ and Nλ , are chosen (see Figure 3.6 for an example of a Gaussian grid (GG))
as follows:
Nλ = m + 1, (3.174)
1
Nθ = (m + 1). (3.175)
2
Let xθ1,Nθ < · · · < xθNθ ,Nθ be the zeros of the Legendre polynomial of degree Nθ ,
and suppose that the grid (GG) is given by
(θj , λk ) ∈ [0, π] × [0, 2π), (3.176)
with
1
θj = arccos(xθj,Nθ ), j = 1, . . . , Nθ = (m + 1), (3.177)
2
2π
λk = k , k = 0, . . . , m. (3.178)
m+1
(m
Then, for Y ∈ Harm 0,...,m = k=0 Harm k , we have
1
2 (m+1)

m
θ
Y (η) dS(η) = wj,N aλ Y (θj , λk )
θ k
Ω j=0 k=0
1
2 (m+1)

m
2π θ
= wj,N θ
Y (θj , λk ). (3.179)
m+1 j=0 k=0
Thus, for m ∈ N odd and Y, F ∈ Harm 0,..., m−1 , the product Y F is of class
2
Harm 0,...,m , hence, the inner product of Y and F

Y, F L2 (Ω) = Y (η)F (η) dS(η) (3.180)
Ω
can be calculated in exact way by use of the Gaussian grid (GG) with 12 (m + 1)2
points.
In particular, this result can be applied to the determination of the Fourier coef-
ficients of an arbitrary function F ∈ L2 (Ω). In more detail, if the Fourier series
expansion of a function F ∈ L2 (Ω)
∞
n
F ∼ F, Yk,l L2 (S2 ) Yk,l (3.181)
k=0 l=−n
with the Fourier coefficients given by the usual projection

F, Yk,l L2 (Ω) = F (η)Yk,l (η) dS(η). (3.182)
Ω
is replaced by its bandlimited version of degree m (for sufficiently large odd m)

m
n
F0,...,m = Yk,l , F L2 (Ω) Yk,l , (3.183)
k=0 l=−n
we are allowed to conclude that

F, Yk,l L2 (Ω) = F0,...,m , Yk,l L2 (Ω) , k = 0, . . . , m, l = −n, . . . , n. (3.184)
In connection with (3.179) we then obtain with Nθ = 12 (m + 1) the following
discrete version
⎛ 1
⎞

m n 2 (m+1)

m
F0,...,m = ⎝ 2π θ
wj,N Yk,l (θj , λk ) F (θj , λk )⎠ Ykl . (3.185)
θ
m + 1 j=0
k=0 l=−n k=0
Recently, a novel cubature formula based on the relation of spherical harmon-

ics to Wigner functions has been proposed by McEwen and Wiaux (for more de-
tails see [148]), that achieves the same degree of accuracy as the Gaussian formula
with (approximately) half as many points as the aforementioned Driscoll–Healy
approach (the details will not be discussed here).
3.7. Latitude-Longitude grids and Weyl’s equidistributions
Both the equiangular longitude-latitude grid (ELLG) and the Gauss grid (GG)
are constructed by a uniform distribution of points in the two-dimensional inter-
val [0, π] × [0, 2π). Unfortunately, this means that these pointsets are not equidis-
tributed on the sphere in the sense of Weyl (see [64] for more details). In fact,
the “density of points around the poles” on S2 is comparatively high, while larger
distances between the nodes can be observed around the equator. In the case of
(ELLG), even a certain number of points coincides with the poles. The Gauss-
ian grid (GG) achieves the same degree of accuracy with (approximately) half as
many points as the system (ELLG). However, the zeros of the Legendre polyno-
mials have to be included which is not as simple as for the grid (ELLG). Besides
the prescription of the location of nodal points where the measurements must be
done, the observation that the integrand is usually not sampled uniformly on the
sphere S2 is of disadvantage except, e.g., for satellite measurements, where the
data are acquired in similar way. No doubt, the facts that the approximate inte-
gration rules are explicitly known, easily implementable, and economical even for
a higher degree m, make latitude-longitude sampling attractive.
The cubature method presented here is only senseful for regular grids gener-
ated by partitioning of the two-dimensional interval [0, π) × [0, 2π). For scattered
data distributions this integration technique is not applicable. In this case, other
cubature formulas (for example, spline integration as proposed by W. Freeden
[38, 39]) have to come into play.
4. From Gaussian theorem to geoidal determination

In 19th century, the forces in physics were believed to be derived from potentials
which satisfy Laplace’s equation. This explains the term “potential theory”. Con-
sequently, in mathematical sense, potential theory was the study of functions that
could serve as (approximants of) potentials. Nowadays, we know that the equa-
tions that describe forces are systems of non-linear partial differential equations,
such as Einstein equations (see, e.g., [1, 146, 225]) and that the Laplace equation is
only valid as a limiting case. Nevertheless, the term potential theory has remained
as a convenient term for describing the study of functions satisfying the Laplace
equation and its generalizations.
4.1. Gauss’s role and influence to potential theory

In classical physical geodesy following the law proposed by I. Newton (1643–1727)
in 1687 the only forces considered were the forces of mutual attraction acting
upon two material particles of small size or two material points. These forces are
directly proportional to the product of the masses of these particles and inversely
proportional to the square of the distance between them. Thus, the first and the
most important problem from the point of view of physical geodesy was to study
the forces of attraction of a material point by a finite (regular) material body –
a spheroid and, in particular, an ellipsoid (since many celestial bodies have this
shape). After first achievements by Newton and others, studies carried out by J.L.
Lagrange (1736–1813) in 1773, A. Legendre (1752–1833) between 1784–1794 and
by P.S. Laplace (1749–1827) continued in 1782–1799 became of major importance.
Lagrange established that the field of gravitational forces, as it is called now, is a
(gradient) potential field. He introduced a function which was called in 1828 by
G. Green (1793–1841) a potential function and later in 1840 by C.F. Gauss just a
potential.
Already Gauss, Green, and their contemporaries discovered that the method
of potentials can be applied not only to solve problems in the theory of gravita-
tion but, in general, to solve a wide range of problems in mathematical physics,
in particular, in electrostatics and magnetism. In this connection, potentials be-
came to be considered not only for the physically realistic problems concerning
mutual attraction of positive masses, but also for problems with “masses of arbi-
trary sign”, or charges. Representative boundary value problems were defined in
potential theory, such as the Dirichlet problem and the Neumann problem, the
electrostatic problem of the static distribution of charges on conductors or the
Robin problem, and the problem of sweeping-out mass (balayage method). To
solve the aforementioned problems in the case of domains with sufficiently smooth
boundaries certain types of potentials turned out to be efficient, i.e., special classes
of parameter-dependent integrals such as volume potentials of distributed mass,
single – and double layer potentials, Green potentials, etc. Results obtained by
A.M Lyapunov (1857–1918) and V.A. Steklov (1864–1926) at the end of 19th cen-
tury were fundamental for the creation of strong methods of the solution of the
main boundary value problems. Studies in potential theory concerning properties

of different potentials have acquired an independent significance. In the first half
of the 20th century, a great stimulus for the generalization of the principal prob-
lems and the completion of the existing formulations in potential theory was made
on the basis of the general notions of a Radon measure, capacity, and generalized
functions. Nowadays, potential theory is closely related in its development to the
theory of analytic functions and to some extend to the probability theory (see,
e.g., [1, 225]). Together with further studies of classical boundary value problems
and inverse problems, the modern period of the development of potential theory is
usually characterized by the application of methods and notions of topology and
functional analysis, and the use of abstract axiomatic methods.
4.2. Vector analytic and potential theoretic tools
Newton’s law of universal gravitation, first published in his Principia in 1687,
asserts that the force v exerted on a point mass Q at x ∈ R3 by a system of
finitely many point masses qi at yi ∈ R3 , i = 1, . . . , N, is equal to
N
Cqi Q x − yi
x → v(x) = , x = yi , i = 1, . . . , N, (4.1)
i=1
|x − yi |2 |x − yi |
with a constant C < 0 (like masses attract). The same law of interaction between
point charges was discovered experimentally by C.A. de Coulomb (1736–1806) and
announced in 1785, now with C > 0 (like charges repel). Note that the numerical
value of the constant C depends on the unit system one is using to measure force,
mass (or charge), and distance. After the introduction of the function

N
Cqi
x → V (x) = , x = yi , i = 1, . . . , N, (4.2)
i=1
|x − yi |
into the theory of gravitation by D. Bernoulli in 1748, J.-L. Lagrange noticed in
1773 that
v(x) = Q ∇V (x), x = yi , i = 1, . . . , N. (4.3)
Hence the function V completely describes the gravitational (or electrostatic) field.
For a continuous distribution of charges with density ρ, vanishing outside G,
the potential becomes

C 1
V (x) = ρ(y) dy, x ∈ G c , (4.4)
4π G |x − y|
where dy is the volume element and G c = R3 \G is the outer space of G. As observed
by P.S. Laplace in 1782, the function G(Δ; | · −y|), y ∈ R3 , given by
1 1
G(Δ; |x − y|) = , x = y, (4.5)
4π |x − y|
in today’s jargon called the fundamental solution of the Laplace equation (4.6) in
R3 \{y}, satisfies
Δx G(Δ; |x − y|) = 0, x ∈ R3 \{y}. (4.6)
Later, the solutions of the Laplace equation came to be known as harmonic func-
tions. It should, however, be remarked that the Laplace equation had been also
considered by Lagrange in 1760 in connection with his study of fluid flow problems.
Laplace’s result was completed by his student S.D. Poisson (1781–1840) in 1813,
when he showed that ΔV = −Cρ for smooth enough densities ρ.
We summarize these classical results in today’s mathematical formalism:
Theorem 4.1. Suppose that G is a regular region in R3 (as introduced earlier in

Section 2).
(1) Let F : G → R be an integrable bounded function. Then

V (x) = C ρ(y) G(Δ; |x − y|) dy, x ∈ G c , (4.7)
G
satisfies

Δx C ρ(y) G(Δ; |x − y|) dy = 0 (4.8)
G
for all x ∈ G c , i.e., V is harmonic in G c .

(2) Let ρ : G → R be of class C (0) (G). Then V as defined by the volume integral
in (4.7) is of class C (0) (G). Furthermore, we have

∇x V (x) = C ρ(y) ∇x G(Δ; |x − y|) dy, x ∈ G. (4.9)
G
Moreover, the so-called Poisson equation under the assumption of μ-Hölder conti-
nuity, μ ∈ (0, 1], can be formulated as follows:
Theorem 4.2. If ρ is of class C (0,μ) (G), μ ∈ (0, 1], then the Poisson differential
equation

−Δx C ρ(y) G(Δ; |x − y|) dy = Cρ(x) (4.10)
G
holds true for all x ∈ G.
In Theorem 4.2 the assumption of μ-Hölder continuity of ρ, μ ∈ (0, 1], is needed for
its proof. As a matter of fact, H. Petrini [173] showed that the μ-Hölder continuity
of ρ, μ ∈ (0, 1], is necessary to imply the second continuous differentiability of the
Newton volume potential.
The Gauss integral theorem (from 1813) and the related Green formulas (cf.
[91]) are among the basic tools of potential theory. They are also indispensable for
a variety of problems in physical geodesy (cf. [107, 112, 159]).
Theorem 4.3 (Gauss’s Integral Theorem). Let G be a regular region. Let F : G → R

be a scalar field, f : G → R3 a vector field, that is continuous on G and differen-
tiable in G, respectively. Then

F (y)ν(y) dS(y) = ∇y F (y) dy, (4.11)
∂G G
f (y) · ν(y) dS(y) = ∇y · f (y) dy, (4.12)
∂G G
ν(y) ∧ f (y) dS(y) = ∇y ∧ f (y) dy, (4.13)
∂G G
provided that the integrand on the right-hand side is Lebesgue-integrable on G. The

vector field ν : ∂G → R3 is the (unit) normal field pointing into the exterior of G
(dS is the surface element).
The identities (4.12) and (4.13) are valid for all vector fields, whatever their
physical meaning is. Of special interest is the case (4.12) in which f may be
understood to be the velocity vector of an incompressible fluid. Inside the surface
∂G there may be sources in which the fluid is generated or sinks in which the fluid
is annihilated. The divergence
∇·f measures the strength of the sources and sinks.
The volume integral G ∇ · f (y) dy is the total amount of the fluid generated in

unit time. The surface integral ∂G f (y) · ν(y) dS(y) is the total amount of fluid
flowing in unit time across the surface ∂G. Therefore, the Gauss formula expresses
a balance equation, namely the evident fact that both integrals in (4.12) are equal.
Gravitational Interpretation. In the case where f is the vector of the gravitational

force, i.e., we especially choose instead of f the field v = ∇V , the intuitive inter-
pretation of the Gauss integral theorem is not so obvious, but the analogy to the
balance equation of fluid flow is often helpful. In gravitation we can take advantage
of the Poisson equation
∇ · v = ΔV = −Cρ. (4.14)
This equation (cf. [107]) can be interpreted to mean that the masses are the sources
of the gravitational field; the strength or the sources, ∇ · v, is proportional to the
mass density ρ. The right-hand side of (4.14) is called the flux of force, in our case
gravitational flux, also in analogy to the fluid flow.
Next we come to the interior Green formulas for regular regions G ⊂ R3 .

Suppose that f = ∇F , where F ∈ C (1) (G)∩C (2) (G), i.e., F : G → R is continuously
differentiable in G and F |G is twice continuously differentiable in G. Let ΔF be
Lebesgue-integrable in G. Then we obtain from the Gauss theorem (4.12)

∂F
(y) dS(y) = ΔF (y) dy, (4.15)
∂G ∂ν G
∂
where, as always, ∂ν = ν · ∇ denotes the derivative in the direction of the outer
(unit) normal field ν.
Under the special choice f = F ∇G the Gauss Theorem yields

Theorem 4.4 (Interior First Green Theorem). Suppose that G ⊂ R3 is a regular
region. For F ∈ C (1) (G), G ∈ C (1) (G) ∩ C (2) (G) with ΔG Lebesgue-integrable on
G we have

∂G
(F (y)ΔG(y) + ∇F (y) · ∇G(y)) dy = F (y) (y) dS(y). (4.16)
G ∂G ∂ν
Taking f = F ∇G − G ∇F we finally obtain
Theorem 4.5 (Interior Second Green Theorem). Suppose that G ∈ R3 is a regular
region. For F, G ∈ C (1) (G) ∩ C (2) (G) with ΔF, ΔG Lebesgue-integrable on G we
have
(G(y)ΔF (y) − F (y)ΔG(y)) dV (y)
G
(4.17)
∂F ∂G
= G(y) (y) − F (y) (y) dS(y).
∂G ∂ν ∂ν
In what follows, we collect some basic material well-known from classical
potential theory in the Euclidean space R3 . First we have a closer look at the
fundamental solution of the Laplace equation. Observing its specific properties we
are able to formulate the third interior Green formula. Mean value theorems and
maximum/minimum principle are the canonical consequences. Harmonic functions
are recognized to be analytic in their harmonicity domain. The Kelvin transform
enables us to study harmonic functions which are regular at infinity. Keeping the
regularity at infinity in mind we are finally led to exterior Green formulas. The
third exterior Green formula is formulated in analogy to its interior counterpart
(for the proof the reader is referred to, e.g., [49]). Essential tools are the limit
relations

x−y
lim F (y) · ∇y G(Δ; |x − y|) dS(y) = F (x), (4.18)
r→0+ Ωr (x) |x − y|

lim F (y) G(Δ; |x − y|) dS(y) = 0 (4.19)
r→0+ Ωr (x)
for a continuous functions F on Bρ (x), ρ > 0, x ∈ R3 . In fact, from (4.18), (4.19),

we easily obtain in combination with the Second Green Theorem.
Theorem 4.6 (Interior Third Green Theorem). Let G ⊂ R3 be a regular region
with continuously differentiable boundary ∂G. Suppose that F : G → R is of class
C (1) (G) ∩ C (2) (G) with ΔF Lebesgue-integrable on G. Then

∂ ∂
G(Δ; |x − y|) F (y) − F (y) G(Δ; |x − y|) dS(y)
∂G ∂ν(y) ∂ν(y)

− G(Δ; |x − y|)ΔF (y) dy) = α(x)F (x), (4.20)
G
where α(x) id solid angle α(x) subtended by the boundary ∂G at the point x ∈ R3
Especially, for a continuously differentiable boundary ∂G, we obtain

Corollary 4.1 (Interior Third Green Theorem). Let G ⊂ R3 be a regular region
with continuously differentiable boundary ∂G. Suppose that F : G → R is of class
C (1) (G) ∩ C (2) (G) with ΔF Lebesgue-integrable on G. Then

∂ ∂
G(Δ; |x − y|) F (y) − F (y) G(Δ; |x − y|) dS(y)
∂G ∂ν(y) ∂ν(y)
⎧
⎪ F (x), x ∈ G,
⎨
1
− G(Δ; |x − y|)ΔF (y) dy) = 2 F (x), x ∈ ∂G, (4.21)
G ⎪
⎩
0, x∈G .c
As special case we obtain for continuously differentiable functions F in G which

are harmonic in G the so-called Interior Fundamental Theorem.
Corollary 4.2. Suppose that F : G → R is of class C (1) (G) ∩ C (2) (G) with ΔF = 0
on G. Then

∂F ∂
G(Δ; |x − y|) (y) − F (y) G(Δ; |x − y|) dS(y)
∂G ∂ν(y) ∂ν(y)
⎧
⎪ F (x), x ∈ G,
⎪
⎨
1
= F (x), x ∈ ∂G, (4.22)
⎪
⎪
⎩2
0, x∈G . c
These formulas, which turn out to be the point of departure for the limit and
jump relations in potential theory (see, e.g., [49, 121, 191]), are also due to Gauss.
Letting F = 1 in G, we obviously find in connection with (4.5) and Corol-
lary 4.2 the following definition (cf. (2.33)).
Definition 4.1 (Solid Angle). Let G ⊂ R3 be a regular region. Then the solid angle
α(x) subtended by the boundary ∂G at the point x ∈ R3 is given by

∂
α(x) = − G(Δ; |x − y|) dS(y). (4.23)
∂G ∂ν(y)
Note that we have

⎧
⎪
⎨ 1 , x∈G
1
a(x) = 2 , x ∈ ∂G (4.24)
⎪
⎩
0 / G.
, x∈
provided that G ⊂ R is a regular region with continuously differentiable boundary
q
∂G (cf. Figure 4.1). In the case of the cube G = (−1, 1)3 ⊂ R3 we especially have
(i) α(x) = 1 if x is located in the open cube G, (ii) α(x) = 12 if x is located on
one of the six faces of the boundary ∂G of the cube G but not on an edge or in
a vertex, (iii) α(x) = 14 if x is located on one of the eight edges of ∂G but not
®(x) = 0
Figure 4.1. Solid angle subtended at x ∈ R3 by the surface ∂G of a

regular region G with “smooth boundary“.
3
6
)y* ?
3
4
)y* ?
)y* ? 2 )z* ? 3
3
)y* ? :
Figure 4.2. Solid angle subtended at x ∈ R3 by the surface ∂G of the

“non-smooth“ cube G = (−1, 1)3 .
in a vertex, (iv) α(x) = 18 if x is located in one of the eight vertices of ∂G (cf.

Figure 4.2).
It should be remarked that the divergence theorem first appeared in La-
grange’s 1860 posthumous work, and it was proved in a special case already by
Gauss in 1813. The general three-dimensional case was treated by M.V. Ostro-
gradsky in 1826. In a preliminary section of his groundbreaking 1828 essay, George
Green proved several reductions of three-dimensional volume integrals to surface
integrals, similar in spirit to the divergence theorem, and independently of M.V.
Ostrogradsky. Nowadays, those are called Green’s identities and best viewed as
consequences of the Gauss integral theorem.
Next we are concerned with the mean value theorem (in today’s mathematical
language) that dates back to C.F. Gauss (1840).
Theorem 4.7. Let G ⊂ R3 be a regular region. Then the following statements are
equivalent:
(a) U : G → R is harmonic in G, i.e., U ∈ C (2) (G) and ΔU = 0 in G,
(b) U : G → R possesses the Mean Value Property on G, i.e., U is of class

C (0) (G) and for all x ∈ G and all r > 0 with Br (x) = {z ∈ R3 : |z| < r} G

1
U (x) = U (y) dS(y), (4.25)
4πr2 |x|=r
(c) U is of class C (0) (G) and for all r > 0 with Br (x) G

(U (x) − U (y)) dy = 0. (4.26)
|x|≤r
A central result in the theory of harmonic functions is the Maximum/Minimum

Principle. Essential tool is the Mean Value Property.
Theorem 4.8 (Maximum/Minimum Principle). Let G ⊂ R3 be a regular region.
Suppose that U is harmonic in G and non-constant. Then U does not reach its
minimum or maximum in G. If, in addition, U is of class C (0) (G), then U reaches
its minimum and maximum in G, and the extremal points are lying on ∂G. More
precisely,
sup |U (x)| ≤ sup |U (x)|. (4.27)
x∈G x∈∂G
A direct consequence of the Maximum/Minimum Principle is the following stability

theorem.
Theorem 4.9. Let G ⊂ R3 be a regular region. Suppose that U and V are of class
C (0) (G) ∩ C (2) (G), and harmonic in G. Let ε be an arbitrary positive number. If
sup |U (x) − V (x)| ≤ ε, (4.28)
x∈∂G
then
sup |U (x) − V (x)| ≤ ε. (4.29)
x∈G
Now we are prepared to establish the (real) analyticity of harmonic functions.

Theorem 4.10 (Analyticity). Let G ⊂ R3 be a regular region. Suppose that U is
harmonic on G. Then U is (real) analytic, i.e., for x0 ∈ G there exists ρ > 0 such
that
∞
1
U (x0 + h) = ((h · ∇)j U )(x0 ) (4.30)
j=0
j!
for all h ∈ R3 with |h| < ρ.
The Newton (volume) potential extended over G is harmonic in the exterior G c =
R3 \G. This is the reason why potential theory under geoscientifically relevant as-
pects essentially aims at concepts in the outer space of a regular region. The
treatment of the outer space in the Euclidean space R3 , however, includes the
discussion at infinity. As a consequence, Green’s integral theorems must be for-
mulated under geophysically relevant conditions imposed on harmonic functions
at infinity. Mathematically (see, e.g., [49]), the “regularity at infinity” can be de-
duced via the Kelvin transform by a transition from functions harmonic in the
inner space to their counterparts in outer space, and vice versa.
Theorem 4.11. If U is harmonic in G c and U converges to zero for |x| → ∞

uniformly with respect to all directions, then |x||U (x)| and |x|2 |∇U (x)| are bounded
for |x| → ∞.
Theorem 4.11 leads us to the definition of the “regularity at infinity”.
Definition 4.2. A function U : G c → R is called regular at infinity, if U satisfies

the asymptotic relation |U (x)| = O(|x|−1 ) and |∇U (x)| = O(|x|−2 ), |x| → ∞,
uniformly with respect to all directions x/|x|.
Now we are prepared to discuss exterior versions of the Green identities involving
harmonic functions being regular at infinity. All these identities can be obtained
by first considering the auxiliary set GR
c
(0) = G c ∩ BR (0) (with R sufficiently large
such that G BR (0), i.e., G ⊂ BR (0) and dist(∂G, ∂BR (0)) > 0) and afterwards
letting R tend to infinity (note that GR c
(0) as the difference of the two regular
regions BR (0) and G allows the application of the interior Green formulas).
Theorem 4.12 (Exterior First Green Theorem). Let F be a function of class

C (2) (G c ) ∩ C (1) (G c ) such that F is harmonic in G c and regular at infinity. Suppose
that the function H ∈ C (1) (G c ) satisfies the asymptotic relations
|y|2 |F (y)∇H(y)| = O (1) (4.31)
and

1
|∇F (y) · ∇H(y)| = O , ε > 0. (4.32)
|y|3+ε
Then
∂H
∇F (y) · ∇H(y) dy = F (y) (y) dS(y), (4.33)
Gc ∂G ∂ν
where ν is the outer unit normal field to G c , i.e., the inner unit normal field to G.
Theorem 4.13 (Exterior Second Green Theorem). Let the function
F, G ∈ C (1) (G c ) ∩ C (2) (G c )
be harmonic in G c and regular at infinity. Then

∂ ∂
F (y) H(y) − H(y) F (y) dS(y) = 0. (4.34)
∂G ∂ν ∂ν
Theorem 4.14 (Exterior Third Green Theorem). Suppose that G is a regular region
with continuously differentiable boundary ∂G. Let U be of class C (1) (G c ) ∩ C (2) (G c )
such that U is harmonic in G c and regular at infinity. Then

∂ ∂
G(Δ; |x − y|) U (y) − U (y) G(Δ; |x − y|) dS(y)
∂G ∂ν(y) ∂ν(y)
⎧
⎨ U (x) , x ∈ G c
1
= U (x) , x ∈ ∂G (4.35)
⎩ 2
0 , x ∈ G,
where ν is the outer unit normal field to G c , i.e., the inner unit normal field to G,
and α(x) is the solid angle subtended by the boundary ∂G at x ∈ R3 .
4.3. Gravitational potential

Since the time of Newton, the theories for the precise determination of the attrac-
tion of the Earth, exerting on a point of the Earth’s surface, did have a special
significance for geodesy. These investigations are especially connected to the names
of Huygens, MacLaurin, d’Alembert, Lagrange, Legendre, and Laplace as well as to
their scholars Poisson and Plana (cf. [216]). In his treatise about his investigations
on potential theory concerning homogeneous ellipsoids (Theoria attractonis corpo-
rum sphaeroidorum ellipticorum homogeneorum methodo nova tractata, Vol. II,
Göttingen, [1813]),, presented to the “Königliche Gesellschaft der Wissenschaften
zu Göttingen” (Royal Society of Sciences at Göttingen), in which Gauss derives
the well-known his name bearing “Gauss’s Integral Theorem” (divergence theo-
rem) about the transformation of a volume integral in a surface integral, he writes:
“The solution of Laplace deserves because of its elegance and its sagac-
ity common admiration; but just the fact that it was necessary to use
special finesses and artifices in order to solve the problem, may among
the geometricians awake the desire for a simpler, less complicated and
more direct solution. This desire did not completely come true by a new
proof of the main sentence given by Legendre (Hist. de l’acad. roy. des
sc. 1788, Sur les intégrales doubles), though the here shown extraor-
dinary analytical skilfulness was by all geometricians recognized. Later
Biot and Plana have tried to simplify the solution (Mm. de l’institut T.
VI; Memorie di matematica e di fisica della societ. italiana T. XV). But
also these two solutions belong to one of the most complicated applica-
tions of analysis. We hope that a new solution will not be unwelcome to
mathematicians and astronomers, which will use a quite different way.”
As a consequence of the historic prework, the determination of the Earth’s
shape and its gravity potential are considered to be main tasks of physical geodesy.
Both problems relate to the work of G.G. Stokes [207]. Nowadays, they can be
tackled simultaneously by the so-called Molodensky problem (cf. [151]). In order to
understand Molodensky’s setup, we have to start once more with the gravitational
potential V of the Earth, related to its mass by the Newton volume integral (4.4).
As a consequence of the Gauss Theorem we already know that V is harmonic in
the outer space, i.e., ΔV = 0 in G c . Moreover, V is regular at infinity. As the Earth
is rotating, we can also assign a centrifugal potentialΦ to the rotating body. This
results in the so-called gravity potential W given by
W = V + Φ. (4.36)
The gradient of the gravity potential is the gravity vector
w = ∇W. (4.37)
The magnitude of w, in geodesy usually denoted by g = |w| and simply called the
gravity (intensity), can be obtained on the surface ∂G of the Earth from gravimet-
ric measurements. The direction of w gives the direction of the plumb line and can
be obtained from astronomic observations and today also from satellite measure-
ments. Plumb lines are not straight, but intersect each equipotential surface of W
normally, such that the gravity vector w at any given point is tangential to the
plumb line at this point.
Combining leveling with the gravimetric and astronomic measurements which
determine w allows us to get W on the surface up to an additional constant which
can be determined from additionally knowing at least one distance. All data sets
are assumed to be corrected for influences like gravitational potentials of other
celestial bodies or the Earth’s precession and atmosphere.
4.4. Geodetic boundary value problems

L. Hörmander [113] in his seminal work stated the situation related to Moloden-
sky’s problem as the following idealized setting (where we follow almost literally
the approach described in [3]):
(i) The Earth G is a rigid body rotating with a known constant angular speed ω
around a fixed axis, which we choose as the x3 -axis. The centrifugal potential
is given by
1 2 2
Φ(x) = ω (x1 + x22 ), x = (x1 , x2 , x3 )T . (4.38)
2
(ii) The center of gravity is the origin 0 of our (Earth’s fixed) system of coordi-
nates.
(iii) The gravity vector w is known at every point P of ∂G.
(iv) The gravity potential W is known at every point P of ∂G.
(v) ∂G can be mapped to the unit sphere Ω = {x ∈ R3 : |x|2 = 1}, i.e., there is
a differentiable embedding ς : Ω → R3 such that ς(Ω) = ∂G.
In the Molodensky problem, we are looking for the unknown gravity potential W in
the whole space R3 and the unknown embedding ς. As the gravitational potential
V is a harmonic function, i.e., ΔV = 0 outside the Earth, we obtain
ΔW = Δ(V + Φ) = ΔV + ΔΦ = 2ω 2 (4.39)
outside ∂G. The assumption of having the barycenter at the origin further on
imposes on V that the asymptotic relation
C
V (x) = + O(|x|−3 ), |x| → ∞, (4.40)
|x|
holds true, which can be seen by taking a multipole expansion into account.
All in all, Molodensky’s problem is a non-linear free-boundary problem and,
therefore, hard to solve mathematically. As a consequence, a variety of approximate
methods exist. Usually, Molodensky’s approach is modified via linearization. In
fact, linearization and a sophisticated iterative process which avoids the loss of
regularity are essential in treating Molodensky’s problem (note that we do not
cover here the iteration procedures as proposed by L. Hörmander [113]).
Linearizing Molodensky’s problem amounts to the introduction of two ingredients:
(α) an approximate surface ∂T for the Earth, called telluroid,
(β) an approximate potential U , called normal potential.
Classically, in the geodetic context, the following steps are carried out:
(1) Choose a reference body E, usually an ellipsoid, which lies completely inside
the Earth and rotates with the same angular speed ω as the Earth around
the x3 -axis.
(2) For every point P on ∂G, find a point Q0 on the surface ∂E such that xP −xQ0
is normal to ∂E. xP and xQ0 are the position vectors of P and Q.
(3) Determine the normal potential U such that
– its gravitational potential part is caused by a mass identical to the
Earth’s mass,
– its centrifugal potential part is identical to (4.38),
– the reference surface ∂E is an equipotential surface of U .
Note that U can be calculated explicitly as its centrifugal part is known and
its gravitational part is the solution to an exterior Dirichlet problem with
boundary values given on ∂E (see, e.g., [94, 107]).
(4) Compute the gradient ∇U of U , called normal gravity vector field and denoted
by u = ∇U with the magnitude γ = |u| called normal gravity. As ∂E is an
equipotential surface of U , u is normal to ∂E for every point on ∂E, i.e.,
u(xQ0 ) is parallel to xP − xQ0 .
(5) For every point P on the real Earth surface ∂G choose a point Q according
to one of the following conditions:
(A) Q lies on the line between P and its corresponding point Q0 on the
surface ∂E so that
W (xP ) = U (xQ ), (4.41)
w(xP ) u(xQ )
(B) = (4.42)
|w(xP )| |u(xQ )|
and
W (xP ) = U (xQ ), (4.43)
(C) w(xP ) = u(xQ ). (4.44)

All points Q chosen in this way make up the telluroid ∂T . Each of the above
conditions gives a slightly different telluroid. None of them is an equipotential
surface, neither for W , nor for U . Condition (A) is the most popular one
and the one originally used by Molodensky, whereas (B) is theoretically more
correct and (C) seems to be better adapted to the so-called Marussi condition
(cf. [113, 131]). The Marussi condition says that the Jacobi matrix of u, i.e.
the Hessian of U, should be non-singular at every point Q, i.e.,
3
∂ 2 U (x)

det = 0 (4.45)
∂xi ∂xj i,j=1
x=xQ
for all points Q on the telluroid ∂T . We shall see later why this condition is
needed.
In principle, the telluroid can be chosen by any surface as long as there is a one-
to-one mapping between ∂G and ∂T (in this respect it should be remarked that
the geoid may be assumed to be known with an accuracy of about one centimeter
or less). In order to be suitable as a point of departure in the context of lineariza-
tion, ∂T should be close to ∂G (in some sense) and chosen in a way that brings
advantages during the following process of linearization (note that a more correct
notation is to write Q0 (P ) and Q(P ) as the points Q0 and Q are dependent on P .
We will do so whenever it may help to avoid any confusion).
The introduction of the normal gravity field u also suggests the definition of
so-called normal plumb lines, i.e., lines which intersect each equipotential surface of
u normally, such that the normal gravity vector u at any given point is tangential
to the normal plumb line at this point.
The normal potential has been well established in geodesy long before Molo-
densky, whereas other surfaces have been in use for a long time before the telluroid,
e.g., the geoid ∂G. As mentioned above, ∂E is an equipotential surface of the nor-
mal potential U . Denoting the value of U on ∂E by U0 , the geoid ∂G is defined as
the equipotential surface of W for which we have W = U0 on ∂G.
Points on the geoid ∂G can be related to points on ∂G. We denote by P0 ∈ ∂G
the point related to P ∈ ∂G. We can determine P0 from P by moving along the
plumb line from P towards the center of the Earth until we reach the geoid.
Another possibility would be to use the normal plumb line to reach the geoid or
to choose P0 as the point on the geoid that lies also on the line between P ∈ ∂G
and the corresponding Q0 ∈ ∂E. With an appropriate choice of ∂E and the normal
potential U , all of these methods yield almost the same point P0 . We define the
distance vector between P0 (P ) and Q0 (P ) as
d(P ) = xP0 (P ) − xQ0 (P ). (4.46)
Its magnitude is the geoidal undulation, in the geodetic context denoted by
N (P ) = |d(P )| = |xP0 (P ) − xQ0 (P )| . (4.47)
Linearization of Molodensky’s Problem. In order to linearize Molodensky’s prob-

lem, we start from the assumption (due to L. Hörmander [113]) that instead of
one embedding ς : Ω → R3 we have a family of smooth embeddings depending on
a parameter χ ∈ [0, 1], i.e.,
S : Ω × [0, 1] → R3 , such that S(Ω, 0) = ∂T , S(Ω, 1) = ∂G (4.48)
smooth with respect to χ, too. Moreover, we let
W : R3 × [0, 1] → R (4.49)
be a family of potentials such that W(·, 0) = U and W(·, 1) = W . We assume that
W(·, χ) = V(·, χ) + Φ, (4.50)
i.e., the gravitational part v of W(·, χ) depends on χ and the centrifugal potential
part Φ is independent of χ (this is in line with the assumptions on U and W ).
Corresponding to W, we arrive at a family of gravity vectors
g = ∇W such that g(·, 0) = u and g(·, 1) = w. (4.51)
We are now able to discuss the composition
W = W ◦ S : Ω × [0, 1] → R, (ξ, χ) → W(ξ, χ) = W(S(ξ, χ), χ). (4.52)
For χ = 0, we find W(S(ξ, 0), 0) = U (xQ ) for Q ∈ ∂T , i.e., boundary values of U
on the telluroid. For χ = 1, we obtain W(S(ξ, 1), 1) = W (xP ) for P ∈ ∂G, i.e.,
boundary values of W on the real Earth’s surface. Analogously, we introduce
g = g ◦ S : Ω × [0, 1] → R, (ξ, χ) → g(ξ, χ) = g(S(ξ, χ), χ). (4.53)
Note that W and w are supposed to be known on ∂G whereas U and u can be
calculated analytically outside ∂E. Now, the linearization can be performed by
differentiation with respect to χ, denoted by a dot. We first notice that
∂W ∂V
Ẇ = = = V̇. (4.54)
∂χ ∂χ
Thus, Ẇ is harmonic and satisfies (4.40). For boundary values we obtain
Ẇ(ξ, χ) = Ẇ(S(ξ, χ), χ) + ∇W(S(ξ, χ), χ)T Ṡ(ξ, χ)

= Ẇ(S(ξ, χ), χ) + g(S(ξ, χ), χ)T Ṡ(ξ, χ), (4.55)
In the same way we arrive at
ġ(ξ, χ) = ġ(S(ξ, χ), χ) + J(g)(S(ξ, χ), χ)T Ṡ(ξ, χ)
= ġ(S(ξ, χ), χ) + Hess(W)(S(ξ, χ), χ)Ṡ(ξ, χ), (4.56)
with J(g) being the Jacobian of g and Hess(W) the Hessian of W. As immediate
results we are led to the aforementioned Marussi condition
3
∂ 2 W(S(ξ, χ), χ)
det = 0 (4.57)
∂xi ∂xj i,j=1
for all (ξ, χ) ∈ Ω × [0, 1]. Rewriting the equations (4.55), (4.56) in composition
form, we see that
Ẇ = Ẇ ◦ S + (g ◦ S)T Ṡ, (4.58)
ġ = ġ ◦ S + (Hess(W) ◦ S) Ṡ. (4.59)
From (4.59), we are able to deduce that
−1
Ṡ = (Hess(W) ◦ S) ġ − ġ ◦ S (4.60)
In connection with (4.58) we therefore obtain from (4.60)
W˙ = Ẇ ◦ S + (g ◦ S)T (Hess(W) ◦ S)−1 ġ − ġ ◦ S
T −1
= Ẇ ◦ S + ġ − ġ ◦ S (Hess(W) ◦ S) (g ◦ S)
T
−1
= Ẇ ◦ S + ġ − ġ ◦ S (Hess(W)) g ◦ S
T
−1
= Ẇ ◦ S + ġ − ġ ◦ S (J (g)) g ◦ S, (4.61)

−1
where we have used the fact that the Hessian is symmetric. (J (g)) g is the
tangent of the curve along which the gravity field has a fixed direction. Such lines
are called isozenithals in geodesy. We introduce

−1
M = (J (g)) g (4.62)
to gain the more compact notation. As a matter of fact, we are able to detect the
following equivalencies
W˙ = Ẇ ◦ S + (M ◦ S)T ġ − ġ ◦ S (4.63)
⇔W ˙ − (M ◦ S)T ġ = Ẇ ◦ S − (ġ ◦ S)T (M ◦ S) (4.64)

T
⇔ Ẇ − (M ◦ S) ġ = Ẇ − ġT M ◦ S, (4.65)
such that
T
Ẇ = Ẇ ◦ S + (M ◦ S) ġ − ġ ◦ S (4.66)

˙ − (M ◦ S)T ġ = Ẇ − ∇Ẇ T M ◦ S.
⇔W (4.67)
Looking at our considerations in more detail we notice that we have specified a

boundary condition for Ẇ. Since the values on the left-hand side of (4.67) are
supposed to be known. The boundary condition is of Robin-type with an oblique
derivative due to the occurrence of the vector M which is in general not normal
to the boundary surface S(Ω, ·).
Geodetic Quantities. Although our derivation as presented above is mathemati-
cally sound, it looks rather unfamiliar to the reader with a geodetic background.
However, the quantities above can be easily related to well-known and often used
quantities from geodesy.
We formally approximate derivatives with respect to χ by differences. As results

we are able to specify the following quantities:
(1) the potential anomaly by
W(S(ξ, 1), 1) − W(S(ξ, 0), 0)
Ẇ(x, χ) ≈ = W (xP ) − U (xQ ) = ΔW (xP ), (4.68)
1
(2) the gravity anomaly by
g(S(ξ, 1), 1) − g(S(ξ, 0), 0)
ġ(x, χ) ≈ = w(xP ) − u(xQ ) = Δw(xP ), (4.69)
1
(3) the disturbing potential by
W(x, 1) − W(x, 0)
Ẇ(x, χ) ≈ = W (x) − U (x) = T (x), (4.70)
1
(4) the gravity disturbance vector δg
g(x, 1) − g(x, 0)
ġ(x, χ) ≈ = w(x) − u(x) = δw(x). (4.71)
1
By comparison we are immediately led to ∇T = δw and the harmonicity of T
outside ∂G. The settings (4.68)–(4.71) also show, that T and δw can be understood
for all x ∈ R3 , whereas the anomalies ΔW and Δw can only be defined for pairs
of corresponding points on certain surfaces.
It is common practice in geodesy to work on the geoid instead of the real sur-
face of the Earth and the ellipsoid instead of the telluroid, although there is no need
mathematically. In this case, the anomalies are defined with respect to P0 ∈ ∂G
and Q0 ∈ ∂E. Moreover, a remove-compute-restore or gravity reduction technique
is required that removes all masses outside the geoid by condensing the topography
such that the disturbing potential can be treated as being harmonic outside the
geoid, computes this disturbing potential and then restores the removed masses
while adapting the potential (see, e.g., [112, 159] for a more detailed discussion of
remove-restore procedures).
In order to approximate the vector M, we go back to u in (4.62). Introducing
the (local) unit vector in (local) direction of the isozenithal as ετ and writing
∂
derivatives in direction of the isozenithal as ∂τ , it can be shown (see, e.g., [131,
158]) and the references therein) that
−1
1 ∂γ
M≈− ετ . (4.72)
γ ∂τ
Introducing −Δg (x) = −ΔwT (x)ετ as the component of the gravity vector in the
downward direction of the isozenithal, the identity (4.67) shows that
) *
1 ∂γ(x) 1 ∂γ(x) ∂T (x)
ΔW (x) − Δg (x) = T (x) − . (4.73)
γ(x) ∂τ γ(x) ∂τ ∂τ x=xP
In the case of a non-rotating sphere (see, e.g., [158]), we get with |x| = r
GM GM
u(x) = − 3 x=− εr , (4.74)
|x| r2
1 ∂γ(x) 1 ∂γ(x) 2
= =− . (4.75)
γ(x) ∂τ γ(x) ∂r r
where G is the gravitational constant and M is the (mean) Earth’s mass.
Given the reference ellipsoid with the semi-principal axes a, a and b, a > b,
consider a point P with the ellipsoidal (or geodetic) coordinates (h, θ, λ), whereas
h is the height above the ellipsoid, − π2 ≤ θ ≤ π2 the latitude and 0 ≤ λ < 2π the
longitude. xP can be written (see, e.g., [89] and the references therein) as
⎛ ⎞
√ a2
⎜ 2 2 2 2
+ h cos(θ) cos(λ)⎟
⎜ a cos (λ)+b sin (λ) ⎟
⎜ ⎟
xP = ⎜⎜ √ a2
+ h cos(θ) sin(λ) ⎟
a2 cos2 (λ)+b2 sin2 (λ) ⎟. (4.76)
⎜ ⎟
⎝ 2 ⎠
√ 2 2 b 2 2 + h sin(θ)
a cos (λ)+b sin (λ)
In spherical approximation, P is mapped on a point P with the spherical

coordinates (r, θ, λ), i.e.,
⎛ ⎞
r cos(θ) cos(λ)
xP = ⎝ r cos(θ) sin(λ) ⎠ (4.77)
r sin(θ)
√
3
whereas r = a2 b + h. Hence, (4.73) reduces to
) *
∂T (x) 2 2
+ T (x) = −ΔwT (x)er + ΔW (x) . (4.78)
∂r r r x=x P
The boundary value problem consisting of the Laplace equation

Δx T (x) = 0, x ∈ Gc, (4.79)
and (4.78) is calledsimple Molodensky problem.
If the telluroid is chosen according to condition (a), ΔW vanishes on the
Earth’s surface. Let us further assume that r can be chosen to take a value R
identical for all points. The direction of the isozenithal is then identical to the
direction of the radius vector er . We also approximate
ΔwT (xP )er ≈ Δg(xP ) = g(xP ) − γ(xQ ). (4.80)
Δg is called gravity anomaly. It is approximately, but not exactly, the magnitude
of the gravity anomaly vector Δw. The gravity anomaly Δg can be computed
directly from measurements of g as γ can be calculated due to the definition of U .
With these assumptions, we arrive at the classical (exterior) Stokes boundary

value problem (see [207])
Δx T (x) =0, |x| > R, (4.81)

∂T 2
(x) + T (x) = − Δg(x), |x| = R, (4.82)
∂r R
C
T (x) = + O(|x|−3 ), |x| → ∞. (4.83)
|x|
As the derivative with respect to r coincides on the sphere with the normal deriv-
ative, this is no longer an oblique-derivative problem.
Stokes Molodensky Koch, Pope
Earth's surface
height anomaly
telluroid
geoid
geoidal height
ellipsoid
Figure 4.3. Different realizations of the geodetic boundary value prob-

lem (modified illustration following R. Rummel [185], see also [3]).
The Stokes problem links the disturbing potential T to gravity anomalies

Δw. On the other hand, we have seen that ∇T = δw. On the sphere S2R around
the origin with radius R, the normal component of this relation is given by
∂T
= δwT εr . (4.84)
∂r
Using an approximation of the form
δwT (x)εr ≈ δg(x) = g(x) − γ(x), (4.85)

thus introducing the gravity disturbance δg, we obtain the exterior Neumann
boundary value problem (see, e.g., [112])
Δx T (x) = 0, |x| > R, (4.86)

∂T
(x) = δg(x) |x| = R, (4.87)
∂r
c
T (x) = + O(|x|−3 ), |x| → ∞. (4.88)
|x|
In the same way as before, δg is approximately the magnitude of the gravity

disturbance vector δw. If the approximation (4.85) is not admissible, which is also
the case if we do not use spherical approximation, the relation evaluation ∇T = δw
on a boundary surface yields an oblique derivative problem.
A historical remark involving Neumann’s problem is due to M.I. Yurkina
(cf. [112]): The Neumann problem is a classical problem of potential theory, with
a long history. Neumann’s problem is named after Carl Neumann, who edited his
father’s (Franz Neumann) lectures from 1850s. The external spherical Neumann
problem also occurs in [121]. It is again found in [114]. In future, because of
GNSS (Global Navigation Satellite System such as GPS, GLONASS, BEIDOU,
GALILEO), gravity disturbances may be expected to be much more available than
gravity anomalies. This observation implies a renewed importance of the Neumann
problem.
Another quantity that can be used to determine the disturbing potential is the
deflection of the vertical. The deflection of the vertical is the difference between
the direction of the reference normal vector ν (x) = u(x)/|u(x)|−1 , associated
with the reference potential U and the reference gravity vector w, and the normal
vector ν(x) = −w(x)/|w(x|, associated with the (actual) gravity potential W and
the (actual) gravity vector w. There are different definitions based on whether ν
is evaluated at the real surface of the Earth or at the geoid and whether ν is
evaluated at the real surface of the Earth, the telluroid, or the reference ellipsoid.
Considering a point P0 on the geoid ∂G, we can decompose ∇T into a normal
part and a surface part, defining the latter as
∇S T (xP0 ) = ∇T (xP0 ) − (∇T (xP0 ) · ν(xP0 )) ν(xP0 ) (4.89)
as ∂G is an equipotential surface of W and, thus, ν(xP0 ) is normal to ∂G. The

reference gravity vector u can be decomposed similarly by defining
uS (xP0 ) = u(xP0 ) − (u(xP0 ) · ν(xP0 )) ν(xP0 )

= −u(xP0 )ν (xP0 ) + u(xP0 ) (ν (xP0 ) · ν(xP0 )) ν(xP0 )
= −uxP0 ) [ν (xP0 ) − (ν (xP0 ) · ν(xP0 )) ν(xP0 )] . (4.90)
Note that w(x) = −g(x)ν(x) by definition, i.e., there are no surface components
of w(xP0 ) at the geoid. Therefore, we get for the surface components
uS T (xP0 ) = −uS (xP0 )
= −|u(xP0 )| [ν (xP0 ) − (ν (xP0 ) · ν(xP0 )) ν(xP0 )]
= −γu(xP0 )Θ(xP0 ) (4.91)
with the (vectorial) deflection of the vertical defined as
Θ(xP0 ) = ν (xP0 ) − (ν (xP0 ) · ν(xP0 )) ν(xP0 ). (4.92)
We can also use Bruns’s formula (cf. [15]) in the form
T (xP0 ) = γ(xQ0 )N (xP0 ) (4.93)
with Q0 being the point on the reference ellipsoid associated with P0 and N the
geoidal undulation (leading the approach given by F.A. Vening Meinesz [217].
In spherical approximation (see, e.g., [159]), we can estimate u(xQ0 ) by the
constant value u0 to arrive at
γ(xP0 )
∇S N (xP0 ) = − Θ(xP0 ). (4.94)
γ0
γ(x )
Traditionally, it is also assumed that γ0P0 ≈ 1 and that ν (xP0 ) can be approxi-
mated by the ellipsoidal normal, i.e., ν (xQ0 ).
In spherical notation, we have
ξ = sin θε3 + cos θ(cos λε1 + sin λε2 ), 0 ≤ λ < 2π, 0 ≤ θ ≤ π (4.95)
(λ: spherical longitude, θ: spherical latitude), where ε1 , ε2 , ε3 , respectively, form
the (canonical) orthonormal basis in R3 and a moving orthonormal triad on the
unit sphere Ω is given in the form
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
cos λ cos θ − sin λ − cos λ sin θ
εr = ⎝ sin λ cos θ ⎠ , ελ = ⎝ cos λ ⎠ , εθ = ⎝ sin λ sin θ ⎠ , (4.96)
sin θ 0 cos θ
so that (4.91) and (4.94) lead to
1 ∗ GM
∇ T (Rξ) = − 2 Θ(Rξ) (4.97)
R ξ R
1 ∗
⇐⇒ ∇ N (Rξ) = −Θ(Rξ). (4.98)
R ξ
The surface gradient ∇∗ξ in local spherical coordinates is given by
1 ∂ ∂
∇∗ξ = ελ + εθ , (4.99)
cos θ ∂λ ∂θ
Note that G is the gravitational constant and R is the (mean) Earth’s radius used
in spherical approximation. The particular representation of ∇∗ξ yields the scalar
equations
1 ∂N
− (λ, θ) = NSC(λ, θ), (4.100)
R ∂λ
1 1 ∂N
− (λ, θ) = EWC(λ, θ) (4.101)
R cos λ ∂λ
with NSC(λ, ϑ), the north-south component, and EWC(λ, ϑ), the east-west com-
ponent of Θ. The difference between (4.98) on the one hand and (4.100) and (4.101)
on the other hand is, that the former is an isotropic vectorial differential equation
whereas the latter is an anisotropic system of two scalar differential equations.
The solution of the isotropic vectorial differential equation (and its multi-scale
approximation) can be found in [61], while the solution of the scalar anisotropic
differential equations is due to F.A. Vening Meinesz [217].
For Dirichlet, Neumann, and Stokes problems, there exist solution procedures
based on an integral representation of T (see, e.g., [159]). However, evaluating these
integrals can be cumbersome. Therefore, it is necessary to consider suitable inte-
gration and approximation formulas to derive a numerical procedure that allows
a fast and precise determination of the disturbing potential T from given bound-
ary data (numerical realizations by locally supported wavelets have been recently
proposed by W. Freeden, M. Schreiner [46], W. Freeden, K. Wolf [63], W. Freeden
et al. [65], W. Freeden, C. Gerhards [49]). This leads us to the consideration of
finite pointset methods on spheres and low-pass filtering involving truncated series
of spherical harmonics. Even harmonic spline Runge methods can be performed,
e.g., after a suitable Kelvin transform (see [3, 20, 49, 55, 56]).
Simplified Geodetic Models. If the normal potential U is determined as above, it

has the same monopole component as W , as the mass of the reference ellipsoid is
assumed to be equal to the mass of the Earth. As a consequence, the monopole
component of T has to vanish, such that we actually have T (x) = O(|x|−3 ) as
|x| → ∞. Even more, the traditional approach to be realized in standard textbooks
of geodesy (see, e.g., [107, 112]) is based on the Pizzetti oriented concept (see, e.g.,
[174]) such that the following assumptions may be supposed to be valid:
(i) The mass within the reference ellipsoid for establishing the disturbing poten-
tial F is equal to the mass of the Earth.
(ii) The center of the reference ellipsoid coincides with the center of the Earth.
(iii) The value of the potential on the geoidal surface and the value of the normal
potential on the reference ellipsoidal surface are the same.
(iv) There are no masses outside the geoid (remove-restore-principle from masses
outside the geoid).
(v) The constructive approximation is simplified for reasons of computational
economy from an ellipsoidal to a spherical framework.
In the Pizzetti oriented concept we finally arrive at the following manifestations

T (x) dS(x) = 0, T (x)(εk · x) dS(x) = 0, k = 1, 2, 3. (4.102)
|x|=R |x|=R
resulting in the identities

Δg(x) dS(x) = 0, Δg(x)(εk · x) dS(x) = 0, k = 1, 2, 3, (4.103)
|x|=R |x|=R

δg(x) dS(x) = 0, δg(x)(εk · x) dS(x) = 0, k = 1, 2, 3. (4.104)
|x|=R |x|=R
To some extent, the conceptional vagueness of disturbing potential and geoid

via boundary problems based on heterogeneously distributed boundary data can be
overcome by combining terrestrial and satellite data. With precise positioning by
satellites, for example, intercontinental height links can be established between lo-
cal geopotential realizations. Moreover, nowadays, GNSS can be used to determine
the real Earth’s surface. This fact led K.R. Koch, A.J. Pope [125] to reformulate
the geodetic boundary value problem as an exterior oblique derivative problem
without any need to introduce an auxiliary surface such as telluroid or ellipsoid.
Conventionally, a solution of an oblique derivative problem corresponding to the
boundary values
∂V ∇V
= F, l= (4.105)
∂λ |∇V |
and
inf (l(x) · ν(x)) > 0, (4.106)
x∈∂G
is represented by a strongly singular (Fredholm) integral equation based on a
potential of a single layer as ansatz function that causes difficulties for numer-
ical realizations. W. Freeden, H. Kersten [51, 52], W. Freeden, V. Michel [56],
W. Freeden, C. Gerhards [49] show that the geodetic boundary value problem us-
ing the gravitational vector (4.105) on the known Earth’s surface can be deduced
from discrete data sets by a constructive Runge theorem in terms of, e.g., mono-
and/or multipoles situated on an arbitrary closed surface lying completely inside
the Earth.
4.5. Concluding remarks about the Earth’s figure
In the 18th century it was well established that deviations between measurement
and ellipsoidal theory were existent, which could not been explained by measure-
ment errors (i.e., inaccurate measurements). The geodetic arc measurements as
well as the pendulum measurements pointed to the problem that local deflec-
tions of the vertical occur because of the different composition of the masses of
the Earth. The actual composition of the Earth’s masses could not correspond to
the presumed assumption. Already Ch. Maire, R.J. Boscović [145] discussed the
gravity changes as local distortions, which they traced back to the influence of
mountains, a simple conclusion of Newton’s potential theory. Also the strongly
different values for the flattening of the ellipsoid coming from different geodetic
arc measurements questioned the regular curvature of the meridians. P.S. Laplace
[136] concluded in his “Traité de Mécanique Céleste” that the Earth is noticeably
different from an elliptic shape.
On the basis of the well-known scientific findings at that time and his own
knowledge about the significance of the deflections of the vertical, which he traced
back to the visibility of the irregularity of the masses as well as to the different den-
sities below the Earth’s surface, C.F. Gauss came as a first scientist to a compre-
hensive definition of the mathematical figure of the Earth which was fundamental
in physical geodesy. This definition has been published in 1828: “Bestimmung des
Breitenunterschiedes zwischen den Sternwarten von Göttingen und Altona” (De-
termination of the latitude difference between the observatories of Göttingen and
Altona), (Gauss Werke vol. IX, p. 49). In this treatise one can find the following
sentences:
“In our opinion the topic is viewed from a wrong angle, if such phenom-
ena are always considered as local deviations of the plumb line, and these
deviations are regarded as it were only an exception. What we call in
a geometric sense the surface of the Earth is nothing else as the same
surface, which everywhere intersects the direction of gravity at right an-
gles and part of it coincides with the surface of the world’s oceans. The
direction of the gravity is determined at every point by the shape of the
part of the solid Earth and its unequal density. At the exterior rind of
the Earth, from which we alone know anything, its shape and density
appear as extreme irregular; the irregularity of the density may extend
fairly deep under the outer rind and cannot be computed, because there
is nearly no data available. The geometrical surface is the product of the
total effect of these unequal distributed elements. In consideration of this
situation nothing can prevent us to view the Earth as a whole as a spher-
oid of revolution, from which its real (geometrical) surface will overall
deflect almost by stronger, weaker, shorter or longer undulations.”
For this definition of the geometrical surface of the Earth, Gauss’s scholar
J.B. Listing [143] chose the term “geoid”, which is up to date used, and is the
reference surface for all heights above sea level. Gauss explains the term “height”
in a letter to Olbers (Gauss works, vol. IX, p. 375):
“This surface (the surface of the oceans) is called a horizontal surface
(couche de niveau); equal heights are given to the points of this surface
without caring by no means about whether or how much these points
deviate from an ellipsoidal spheroid. The heights above this surface can
be measured by a barometer as well as trigonometrically so that both
must correspond to each other.”
According to H. Moritz [155] this surface of the Earth, in principle defined
by Gauss, can in its importance for geodesy be considered as a change like that
of Copernicus, because the simple model of the surface of an ellipsoid as an ideal
was replaced by the physically interpreted Earth. Since that time the ellipsoid
serves in geodesy mainly as surface for computations in geodetic surveying and

as a reference surface for the geoid and the Earth’s gravity field. The Gaussian
definition solves the problems with complicated correction models, which occur
by reducing measurements for the determination of deflections of the vertical. The
surface of an Earth representing ellipsoid of revolution cannot be exactly horizontal
because of gravitational pull of the mountains, that is vertical to the plumb line.
The Gaussian definition does not imply a simple geometrical surface, but it is
defined by specifying all locations in Euclidean space R3 satisfying the simple
equation W = W0 = const . In other words, W is equated with the constant W0 ,
with that the geoid is explained as a level (equipotential) surface of the Earth’s
gravity field. Gauss introduced with his definition a principally different model,
which is defined through the observations and represents a reversal of perspective
(Gauss says that this definition is a change of the point of view).
A generalization of the Gaussian definition was achieved by H. Bruns [15]:
“The problem of scientific geodesy is the investigation of the potential
function of the Earth.”
Instead of looking at a special equipotential surface, namely the geoid, all possible
equipotential surfaces are considered in this definition as well as the gravity poten-
tial. This definition by Bruns can especially be used in satellite geodesy, because a
satellite is moving in the gravity field of the Earth so that its orbit is determined
by the potential W .
In today’s geodesy concerned with Earth’s figure, a result first motivated by
C. Runge [187] in one-dimensional complex analysis and later generalized, e.g., by
J.L. Walsh [222], I.N. Vekua [218], and L. Hörmander [113] to potential theory in
three-dimensional Euclidean space R3 is of basic interest. For geodetically relevant
application (see, e.g., [130, 155, 159, 189] and the references therein), the Runge–
Walsh theorem may be formulated as follows (cf. [156]): Let the Earth’s interior
G ⊂ R3 be a regular region, i.e., a bounded region G ⊂ R3 dividing R3 uniquely
in G and the outer space G c = R3 \G, G = G ∪ ∂G, such that G contains the origin
and the boundary ∂G is an orientable smooth Lipschitzian manifold of dimension
2. The Earth’s gravitational potential in G c that is harmonic in G c and regular
at infinity can be approximated by a function that is harmonic outside an arbi-
trarily given Runge (in geodesy called Bjerhammar) ball A G, i.e., A ⊂ G with
dist(A, ∂G) > 0 in the sense that, for any given ε > 0, the absolute error between
the two functions is smaller than ε for all points outside and on any closed surface
completely surrounding ∂G in its outer space. The value ε may be arbitrarily small,
and the surrounding surface may be arbitrarily close to the surface.
Obviously, the Runge–Walsh theorem in the preceding formulation represents
a pure existence theorem. It guarantees only the existence of an approximating
function and does not provide a constructive method to find it.
The situation, however, is completely different if spherical geometrics are ex-
clusively involved in the Runge concept. Assuming that both A, G are concentric
balls around the origin with A G, a constructive approximation of a potential
in the outer space G c is available, e.g., by outer harmonic (orthogonal) expansions

(see, e.g., [81, 121, 168, 223]). More concretely, within the classical context of a
twofold spherical configuration, a constructive version of the Runge–Walsh the-
orem can be guaranteed by finite truncations of Fourier expansions in terms of
outer harmonics, where the L2 (∂G)-convergence of the Fourier series implies uni-
form converges on any pointset K G c . The Fourier coefficients are obtained by
integration over the sphere ∂G. The gravitational potential is available (in spec-
tral sense) by tables of the Fourier coefficients. Nowadays, in fact, outer harmonic
expansions constitute the conventional geodetic tools in globally reflected approx-
imation of the Earth’s gravitational potential and its observables.
From a superficial point of view, one could suggest that approximation by
truncated series expansions in terms of outer harmonics is closely related to spher-
ical geometries ∂A, ∂G. W. Freeden [37], however, showed that the essential steps
to a constructive Fourier approach can be extended to any regular (i.e., not-
necessarily spherical) region G and to any regular (i.e., not-necessarily spherical)
Runge domain A G. The Runge–Walsh theorem in this formulation avoids any
difficulty with the convergence to the gravitational potential by the generalized
Fourier series for arbitrary sets K G c . In analogy to the spherical case, however,
it likewise does not help to specify convergence inside Ac \G c , so that any attempts
(see [10]) to reduce gravitational information via infinite Fourier series downward
from ∂G to the Runge surface ∂A are not justifiable by the Runge–Walsh frame-
work.
In summary, the Runge–Walsh concept reflects constructive approximation
capabilities of the Earth’s gravitational (and not gravity) potential even if geosci-
entifically realistic (i.e., not necessarily spherical) geometries come into play. For
numerical computations, the Runge concept may be regarded as the justification
why boundary value problems in geodesy have to be solved without any need to
use specific telluroidal reflected trial function systems such as numerically more
difficult to handle ellipsoidal harmonics, Lamé functions, etc. Instead outer spher-
ical harmonics (i.e., multi-poles) and/or mono-poles showing a larger harmonicity
domain than the exterior of a telluroid (in linear Molodensky problem) or the
actual Earth (in oblique derivative problem) can be taken into account.
5. From Gaussian least squares adjustment to inverse multi-scale

regularization
5.1. Gauss’s historic role and influence
Mathematicians interested in geodesy and astronomy tried to use the probability
theory since the middle of the eighteenth century for the evaluation of redun-
dant measurements. Like in the theory of games, values of measurements can
be regarded as random samples of a random process. Already in the year 1755,
Simpson pointed out that the mean value of a number of measurements can be ad-
vantageously used (see [7] for more details). Especially in the works of the French
mathematicians Lagrange and Laplace, important theoretical foundations were
treated. The actual reasoning for the adjustment theory with its core, the method
of least squares, was done by Gauss and Legendre around 1800. Gauss probably
hit on the fundamental idea in the autumn 1794 reading a publication of Lam-
bert concerning applications of mathematics (see [69]), but published his method
not before 1809 (cf. “Theoria motus corporum coelestium in sectionibus conicis
solem ambientium”). He closed this publication by giving an explanation of his
method (“Determinatio orbitae observationibus quotcumque quam proximae sat-
isfacientis”). Further works followed in the year 1810 (“Disquisitio de elementis
ellipticis Palladis”) in 1816 (“Bestimmung der Genauigkeit der Beobachtungen”
1880) and finally, in 1821 and 1823 in a systematic presentation (“Theoria combi-
nationis observationum erroribus minimis obnoxiae, pars prior et pars posterior”).
Three years later these works were completed by the note “Supplementum theoriae
combinationis observationum erroribus minimis obnoxiae”.
Actually, in the year 1806, Legendre first published this adjustment method
in a supplement of his treatise “Nouvelles méthodes pour la détermination des
orbites comètes” and termed this principle “Sur la méthode des moindres carrés”
(cf. [139]). This term was also taken by Gauss. Legendre explains in his paper
that there is among all principles no simpler one for the evaluation of observations
than to minimize the sum of the squared errors in order to estimate an unknown
quantity.
Already about 1798 Gauss succeeded in giving his approach a foundation based
on the probability theory. It is possible that the notice in his diary from June 17th
of that year refers to this achievement: “Calculus probabilitatis contra Laplace
defensus” (Gauss Werke X, p. 533). Since 1801 Gauss applied this method nearly
daily for his astronomic computations (Gauss Werke, Vol. IV, p. 98). A well-known
application during this time was the computation of the orbit of the Planetoid
Ceres (cf. [27]), which could only be observed by Piazzi over 90◦ of its orbit. In spite
of the small number of observations Gauss was able to compute the ephemeris of
Ceres so precisely that Ceres could be redetected at the predetermined place. The
discovery of the planetoid Ceres introduced Gauss to the world as a theoretical
astronomer of the highest reputation ([27]). Why Gauss published his “Theoria
motus corporum coelestium . . . ” so late can probably only be explained by the
endeavour of Gauss to give his works “an inner perfection”. Gauss’s motto was:
“pauca sed matura”.
This is probably the reason that Legendre could publish before him in the year
1806 the method without giving a thorough reasoning. R. Adrain [2] (see also the
comments by E. Hammer [97]) introduced this principle, too.
Legendre characterized this method of least squares as the most general, most
accurate, and very easy to apply adjustment principle and proved that this method
leads for one unknown to the arithmetic mean and for two or three unknowns to the
center point of the system. According to Legendre a kind of equilibrium between
the deviations of the measurements will be fixed, comparable to the determination
of a space point by measurements, which lies in the center point of the system.
Independently from Legendre, R. Adrain [2] derived in his publication “Research
concerning the probabilities of the errors which happen in making observations”
the function of the error probability and the method of least squares resulting
out of it. The correspondence of great mathematicians of that time, e.g., between
Gauss and Laplace, supports the impression that there was no dispute concerning
the priority. Rather the correspondence proves that this simple principle of the
method of least squares was highly regarded by the contemporaries (see also [87]).
According to V. Bialas [7], the method of least squares can be regarded as a
characteristic example of the increasing importance of mathematics in geodesy
in the first half of the 19th century. At the same time this method also is an
expression of how the unity of observational practice, the evaluation of measuring
results, and the corresponding theory is growing together.
Reasons for the Least Squares Method.The introduction of the “mean square er-
ror” and its definition as an independent measure of precision was an important
assumption for the foundation of the least squares method, which can be written
by the following objective function using today’s representation (see, e.g., [238]):
v T P v = min . Here P is a weight matrix, e.g., a diagonal matrix, in the Gauss-
ian approach, and v is the vector of residuals. Besides the complete definition of
this measure of precision Gauss also demonstrated in his “Theoria combinationis
observationum erroribus minimis obnoxiae” how to get the mean square errors of
the unknowns as well as the weight of a function of the adjusted unknowns in the
course of an adjustment of observation equations (see [77] and the comments by
G.W. Stewart [205]).
Approximate Reasoning. Following a publication by R. Dedekind [25], Gauss did
argue in his lecture about the least squares method against the adjustment prin-
ciple of minimizing the absolute sum of errors (deviations) and the algebraic sum
being zero. This principle was first applied by Ch. Maire, R.J. Boscović [145] and
later by P.S. Laplace [136]
n
|vi | = min, (5.1)
i=1
under the constraint

n
vi = 0. (5.2)
i=1
n
Nowadays, geodesy calls the principle i=1 |vi | = min together with the con-
straint (5.2) the Boscovic̀–Laplace method. It seems that F.Y. Edgeworth [28] used
this method primarily but without the constraint (5.2) for the estimation of many
unknowns. A solution suitable for practical geodetic computations, which may
be used to determine the most favorable weights in base extensional networks,
was achieved by K. Friedrich [67]. About 50 years earlier, O. Schreiber [194]
installed this method to solve similar problems. Mathematically, the Boscović-
Laplace method (5.1) and (5.2) may be regarded as an 1 -method under con-
straints. No doubt, 1 -type approaches are important for geodesy. It is especially
appropriate to detect gross errors (outliers) in observations. It may also be under-
stood as a problem of linear programming. Using the 1 -method a decomposition
in consistent and non-consistent observation equations can be deduced so that an
outlier test can be developed. Concerning this outlier test the question, if there
exist a primal or dual degeneration of solutions, is of great numerical significance.
One of Gauss’s arguments against the Boscović–Laplace method shall be
mentioned: With a greater number of observations a bigger error would not ex-
ert a stronger impact on the results than many smaller errors, of which the ab-
solute values have the same sum. An error, which occurs n-times, must have a
stronger impact than n-single errors. Instead of the errors themselves the squares
are therefore taken to estimate the usefulness of a hypothesis of smallness of their
corresponding sum of squared errors. Gauss rejected errors with a higher power,
which have always to be even, because the computations would than be extremely
complicated so that the treatment would not be worth the effort (cf. [25]).
Today, for practical purposes, e.g., in engineering surveying, the estimation
procedures for the least squares method and the 1 -method are usually applied
in simultaneous combination, because the corresponding programs are available.
Specifically with the help of the 1 -method outliers are detected, localized, and
then eliminated. In some cases the 1 -method turns out to be insufficient, to detect
reliable outliers, so that other robust estimation methods come into play (see, e.g.,
[98, 115, 119, 123, 124, 142, 147, 192] and the references therein for more details).
Probabilistic Reasoning: Gauss’s work concerning the theory of probability starts
with a justification for the least squares method. In his “Theoria motus corporum
coelestium” the problem was set in such a way that, with measurements of equal
precision, the accidental errors exhibit a differentiable density of the probability
distribution. The distribution is determined under the supposition that the most
probable value for the measured quantity with any number of observations is equal
to the arithmetic mean of the measured values (cf. [87]). Under this assumption
Gauss was led to the well-known formula (cf. Figure 5.1) for the (Gaussian) prob-
ability distribution (i.e., the bell curve).
Figure 5.1. Gaussian distribution function with μ mean or expecta-

tion of the distribution (and also its median and mode), σ standard
deviation, σ 2 variance.
The formula provides a measure for the precision of the observations. Gauss
was aware of certain arbitrariness of this quantity. In fact, he wrote in Articles 178
and 179 of his publication “Theoria motus corporum coelestium: Functio modo
eruta omni quidem igore errorum probilitatis exprimere certo non potest. . . ”.
“Hoc principium, quod in omnibus applicationibus mathesis ad philosoph-
iam naturalem usum frequentissimum.”
The function just found cannot, it is true, express rigorously the probabilities of
the errors: For since the possible errors are in all cases confined within certain
limits, the probability of errors exceeding those limits ought always to be zero,
while our formula always gives some value. However, this defect, which every ana-
lytical function must, from its nature, labor under, is of no importance in practice,
because the value of our function decreases so rapidly, when it has acquired a con-
siderable magnitude, that it can safely be considered as vanishing. This principle,
which promises to be of most frequent use in all applications of the mathematics
to natural philosophy, must, everywhere, be considered an axiom with the same
propriety as the arithmetical mean of several observed values of the same quantity
is adopted as the most probable value (translated by P.J. Davis [23]).
After Gauss has determined the probability distribution for the measuring
errors (errors of observations) in such a way, he concluded in one of the next
paragraphs: The probability density of a given totality of observations will attain
its maximum under the condition that the sum of the squares of the deviations
of the observed quantities with respect to the true values of the quantities, which
have to be measured, will become a minimum. This principle can also be applied to
observations of different precision. If the mean square error is however unknown
and the deviations of the approximate values from the true values have to be
estimated, then the student distribution should be chosen according to the number
of degrees of freedom. Especially with a small number of observations the results
will be different to the ones determined with the procedure given by Gauss (see,
e.g., [87]).
In case that the Gauss’s error law and the axiom of the arithmetic mean
underlying this law would not be valid, then the unknowns determined by the
method of least squares are not any more the most probable values. Gauss came
to a second justification presumably on the basis of an article by Laplace, who
proved in his “Théorie analytique des Probabilités” ([138]) that the unknowns
determined by the method of least squares have the smallest average errors. This
proof is independent of the form of the error law, if for all observations the same
error law is valid and if positive as well as negative errors of equal absolute value are
equal probable. Laplace furnished this proof for two unknowns under the condition
that the number of the observations is infinite. Gauss did not take this assumption
into account and he proved this in his “Theoria observationum erroribus minimis
obnoxiae”. In the “Anzeigen” ([77]) the following sentence can be found:
“. . . , die Funktion für die Wahrscheinlichkeit der Fehler sei, welche sie
wolle, und die Anzahl der Beobachtungen möge gross oder klein sein.”
(. . . , the function for the probability of the errors may be as it likes to
be and the number of observations may be large or small).
In consistency with G.W. Stewart [205] we present a list of what was new in
Gauss’s treatment of random errors:
(1) The careful distinction between systematic and random errors.
(2) The use of the first moment of a distribution to measure its center.
(3) The use of the second moment to measure precision.
(4) A Chebyshev-like inequality.
(5) The correct formula for the expectation of a function of a random variable.
(6) The rate of convergence of the sample mean and variance.
(7) The correct formula for estimating the precision of observations from the
residual sum of squares.
Gaussian Adjustment Theory. For geodetic purposes, even in our days, the adjust-
ment of observation equations (Gauss–Markoff model) is in principle applied in
such a way as Gauss has proposed it. H. Wolf [236] explains: After linearization of
the observation equations, usually by Taylorizing, the system of normal equations
is set up and solved by use of the original Gaussian algorithm. Weight reciprocals
and weighting coefficients as well as the partial derivatives of the observations
with respect to the unknowns are computed. The error calculation consists of the
computation of the mean square error of unit weight, the observations, the un-
knowns, and their functions. The well-known control formula for the computation
of the sum of the squared residuals can also be dated back. Further, the problem is
solved to specify the alterations in the unknowns, which are caused by an addition
of a further observation equation or by a change of a single weight. Also the way
of getting homogeneity by multiplying with the square root of the weight can be
found in Gauss’s work.
Figure 5.2. Title-page of “Theoria motus corporum coelestium in sec-

tionibus conicis solem ambientium”, original latin edition by Friedrich
Perthes and I.H. Besser, Hamburg, 1809, also: “Carl Friedrich Gauss:
Werke, herausgegeben von der (Königlichen) Gesellschaft der Wis-
senschaften zu Göttingen” (Göttinger Digitalisierungszentrum).
Examples of Application by Gauss and Others. Starting with the successful com-
putation of the orbit of the planetoid Ceres (dwarf planet) up to the computation
of the geodetic arc measurements in the Kingdom of Hanover Gauss validated the
practical usefulness of his method. Besides the justification and development of his
least squares method Gauss proposed numerous approximate adjustments, which

have naturally to be free of arbitrariness and contradiction as well as qualified for
adjustments in successive steps and in groups. In order to adjust the angle obser-
vations taken on the single stations of the net for his arc measurements (station
adjustments) Gauss used partly angle unknowns and partly bearing unknowns.
To solve the adjustments he developed the successive approximation method (cf.
[141]), which was applied by an adjustment of observation equations as well as of
condition equations (cf. [86]). After getting the normal equation system with each
step only the residual for one unknown is determined. For this step the residuals
of the other unknowns are however set to zero. Gauss adjusted trigonometric nets
partly with angle residuals and partly with bearing residuals using the condition
equation method. The correlations coming from the station adjustments were not
regarded (cf. [236]). However, C.F. Gauss [80] pointed out that, for this purpose,
the observations to be adjusted must be independent of each other.
For the adjustment of condition equations it is not necessary to put up the
normal equation system, if directly the condition equations are step by step ful-
filled. This possibility for the application of the Gaussian successive approximation
procedure with height nets was firstly pursued by C.A. Vogler [219]. For the ad-
justment of the geodetic arc measurement net in the kingdom of Hanover with
condition equations Gauss used a successive approximation procedure, which he
did not publish in detail, but it was later reconstructed by L. Krüger [132]. Sub-
sequently, Krüger developed this procedure to the so-called Krüger method by
two groups: C.F. Gauss [80] recommended only in Article 20 of his publication
“Supplementum theoria combinationis . . .” to divide the condition equations in
two groups for the adjustment of large triangulation nets. More explicitly, Gauss
writes in Article 20 (due to G.W. Stewart [205]):
“Quoties multitudo aequationum conditionalium permagna est, determi-
natio correlatorum . . .” (“When the number of conditional equations is
very large, the calculation of the correlates A, B, C, etc. by direct elimi-
nation becomes so laborious that the endurance of the calculator is not
equal to the task. In such cases it is often better to use the theory of the
preceding subsection to compute the complete adjustment by successive
approximation . . .”).
A further important geodetic application example was the computation of the
dimensions of the Earth’s ellipsoid by the Gauss-scholar Schmidt, who at Gauss’s
suggestion adjusted the geodetic arc measurements of good quality known at that
time. For the polar flattening he actually came to the result of 1 : 298,39, an
amazing result, if compared to the corresponding value of the World Geodetic
System (WGS 84) of 1 : 298, 257.
The Gaussian Influence on More Recent Developments. It is possible to relate var-

ious settings of today directly to the Gaussian conceptions, e.g., the covariance.
In fact, the essence inherent in the definition of the covariance can be deduced
from Articles 17 and 18 of Gauss’s publication “Theoria Combinationis observa-

tionum. . . , Pars Prior”([77]). We present Article 17 here in the the original Latin
form and Articles 17 and 18 in their English translation:
Article 17 of Gauss’s Latin publication “Theoria Combinationis Observa-
tionum. . . , Pars Prior” in Latin language ([77]:
17.
Si valor quantitatis, quae ab alia quantitate incognita pendent, per ob-
servationem praecisione absoluta non gaudentem determinata est, valor
incognitae hinc calculatus etiam errori obnoxius erit, sed nihil in hac
determinatione arbitrio relinquiter. At si plures quantitates ab eadem
incognita pendentes per observationes haud absolute exactas innotue-
runt, valorum incognitae vel per quamlibet harum observationum eruere
possumus, vel per aliquam plurium observationum combinationem, quod
innitis modis diveris er potest. Quamquam vero valor incognitae tali
modo prodiens errori semper obnoxius manet, tamen in alia combina-
tione maior, in alia minor error metuendus ert. Similiter res se habebit,
si plures quantitates a pluribus incognitis simul pendentes sunt obser-
vatae: prout observationum multitudo multitudini incognitarum vel ae-
qualis, vel hac minor, vel maior fuerit, problema vel determinatum, vel
indeterminatum, vel plus quam determinatum erit (generaliter saltem
loquendo), et in casu tertio ad incognitarum determinationem observa-
tiones innitis modis diversis combinari poterunt. E tali combinationum
varietate eas eligere, quai maxime ad rem faciant, i.e., quae incogni-
tarum valores erroribus minimis obnoxios suppeditent, problema sane
est in applicatione matheseos ad philosophiam naturalem longe gravis-
simum.
In Theoria motus corporum coelestium ostendimus, quomodo val-
ores incognitarum maxime probabiles eruendi sint, si lex probabilitatis
errorum observationum cognita sit; et quum haec lex natura sua in om-
nibus fere casibus hypothetica maneat, theorem illam ad legem maxime
plausibilem applicavimus, ubi probabilitas erroris x quantiti exponentiale
e mit Exponent proportionalis supponitur, unde methodus a nobis dudum
in calculis praesertim astronomicis, et nunc quidem a plerisque calcu-
latoribus sub nomine methodi quadratorum minimorum usitata dem-
anavit.
Postea ill. Laplace, rem alio modo aggressus, idem principium om-
nibus aliis etiamnum praeferendum esse docuit, quaecumque fuerit lex
probabilitatis errorum, si modo multitudo sit permagna. At pro multi-
tudine observationum modica, res intacta mansit , ita ut si lex nostra
hypothetica respuature, methodus quadratorum minimorum eo tantum
nomine prae aliis commendabilis habenda sit, quod calculorum concin-
nitati maxime est adaptata.
Geometris itaque gratum fore speramus, si in hac nova argumenti

tractatione docuerimus, methodum quadratorum minimorum exhibere
combinationem ex omnibus optimam, non quidem proxime, sed abso-
lute, quaecumque fuerit lex probabilitatis errorum, quaecumque observa-
tionum multitudo, si modo notionem erroris medii non ad menterm ill.
Laplace set ita, ut in artt. 5 et 6 a nobis factum est, stabiliamus.
Ceterum expressis verbis hic praemonere convenit, in omnibus dis-
quisitionibus sequentibus tantummodo de erroribus irregularibus atque a
parte constante liberis sermonem esse, quum proprie ad perfectam artem
observandi pertineat, omnes errorum constantium causeas summo studio
amovere. Quaenam vero subsidia calculator tales observationes tractare
suscipiens, quas ab erroribus constantibus non liberas esse iusta suspi-
cio adest, ex ipso calculo probabilium petere possit, disquisitioni peculiari
alia occasione promulgandae reservamus.
Articles 17 and 18 of Gauss’s Latin publication “Theoria Combinationis Ob-
servationum. . . , Pars Prior” in English language [translated by G.W. Stewart
[205]]:
17.
Suppose a quantity that depends on another unknown quantity is esti-
mated by an observation that is not absolutely precise. If the unknown
is calculated from this observation, it will also be subject to error, and
there will be no freedom in this estimate of it. But if several quantities
depending on the same unknown have been determined by inexact obser-
vations, we can recover the unknown either from one of the observations
or from any of an infinite number of combinations of the observations.
Although the value of an unknown determined in this way is always
subject to error, there will be less error in some combinations than in
others.
A similar situation occurs when we observe several quantities de-
pending on several unknowns. The number of observations may be equal
to, less than, or greater than the number of unknowns. In the first case
the problem is well determined; in the second it is indeterminate. In the
third case the problem is (generally speaking) overdetermined, and the
observations can be combined in an infinite number of ways to estimate
the unknowns. One of the most important problems in the application of
mathematics to the natural sciences is to choose the best of these many
combinations, i.e., the combination that yields values of the unknowns
that are least subject to the errors.
In my Theory of the motion of heavenly bodies I showed how to
calculate most probable values of the unknowns, provided the probability
law of the observation errors is known. But in almost all cases this law
can only be hypothetical, and for this reason I applied the theory to the
most plausible law, in which the probability of an error x is proportional
e−hhxx . From this supposition came a method which I had already used
for some time, especially in astronomical calculations. It is now used by
many calculators under the name of the method of least squares.
Later Laplace attacked the problem from a different angle and
showed that if the number of observations is very large then the method
of least squares is to be preferred, whatever the probability law of the er-
rors. But for a modest number of observations, things are as they were,
and if one rejects my hypothetical law, the only reason for recommend-
ing the method of least squares over other methods is that it lends itself
to easy calculation.
I therefore hope that mathematicians will be grateful if in this new
treatment of the subject I show that the method of least squares gives the
best of all combinations – not approximately, but absolutely, whatever
the probability law of the errors and whatever the number of observations
– provided only that we take the notion of mean error not in the sense
of Laplace but as in Arts. 5 and 6.
Here we should say that in the sequel we will be concerned only
with random errors having no constant part, since the craft of taking
observation requires that we take pains to remove all causes of constant
errors. On another occasion I will give a special treatment about what
help a calculator can expect from the calculus of probabilities when he
undertakes to treat observations he suspects are not free of constant
errors.
18.
Problem. Given a function U of the unknown quantities V, V , V , etc.,
find the mean error M to be feared in estimating U when, instead of
the true values of V, V , V , etc. one uses independently observed values
having mean errors m, m , m , etc.
Solution. Let e, e , e , etc. denote the errors in the observed val-
ues of V, V , V , etc., and let λ, λ , λ , etc. be the differential quotients
e, e , e etc. at the true values of V, V , V , etc. Then the resulting error
in U can be represented by the linear function
λe + λ e + λ e + etc. = E
provided the observations are precise enough so that we can neglect
squares and products of the errors. From this it follows first that the
mean value of E is zero, since the observation errors are assumed to
have no constant parts. Moreover, the mean error to be feared in this
value of U is the square root of the of the mean value of EE; that is,
M M is the mean value of the sum
λλee + λ λ e e + λ λ e e + ect. + 2λλ ee + 2λλ ee + 2λ λ + etc.
Now the mean value of λλee is λλmm, the mean value of. λ λ e e is
λ λ m m , etc. The mean values of the products 2λλ ee , etc. are all
zero. Hence it follows that

√
M = λλmm + λ λ m m + λ λ m m + etc.
It is appropriate to append some comments to this solution.
I. Since we have taken the observation errors to be quantities of the
first order and have neglected quantities of higher orders, we may
dU
use the values of the differential quotients dV , etc. that come from

the observed quantities V, V , V , etc. to evaluate our formula in-
stead of λ, λ , λ , etc. Obviously this substitution makes no differ-
ence at all when U is a linear function.
II. Let p, p , p , etc. be the weights of the observation errors with re-
spect to an arbitrary unit, and let P be the weight of the estimate of
U derived from the observed quantities V, V , V , etc. If we prefer
to work in terms of these quantities rather than the mean errors,
then we have
1
P = λλ λ λ λ λ
.
p + p + p + etc.
III. Let T be another function of the quantities V, V , V , etc., and for

the true value of these quantities let
dT dT dT

= κ, = κ , = κ , etc.
dV dV dV
Then the error in the estimate for T obtained from the observed val-
ues V, V , V , etc. is E =√κe+κ e +κ e +etc., and the error to be
feared in this estimate is κκmm + κ κ m m + κ κ n m + etc.
The errors E and E are clearly not independent, and, unlike the
products ee , the mean value of EE is not zero but κλmm +
κ λ m m + κ λ m m + etc.
IV. Our problem also extends to the case where the quantities V, V , V ,
etc. are not obtained directly from observations but are derived from
arbitrary combinations of observations. However, the individual
quantities must be mutually independent, i.e., based on different
observations. If this condition does not hold, then the formula for
M will be in error. For example, if some observation that was in-
volved in the calculation of V is also used in the calculation of V ,
the errors e and e will no longer be independent, and the mean
value of the product ee , will not be zero. However, if we can ascer-
tain the relation of V and, V with the simple observations from
which they were derived, we can determine the mean value of the
product ee by the methods of comment III and hence give the cor-
rect formula for M .
Summarizing, we are allowed to say that Gauss was, in principle, acquainted with
the essence of covariance.
The Chebyshev Principle: The adjustment principle, established by Chebyshev in

1853, to make the biggest residual correction as small as possible, |vmax | = min,
also follows from the Gaussian principle. This perception can be already found in
Gauss’s publication (see Figure 5.2 for its title-page): “Theoria motus corporum
coelestium in sectionibus conicis solem ambientium” [1809]. Gauss writes in Book
2, Article 186 (in translation by P.J. Davis [23], p. 270):
“If we were to adopt a power with an infinite even exponent, we should
be led to that system in which the greatest differences become less than
in any other system”.
Hansen’s Relationship to Pseudoinverses. According to H. Wolf [237], P.A. Han-

sen (1795–1874) used the so-called Bessel zero point correction (in an adjustment
of a triangulation net) to go from the Bessel angle unknowns on to the bearing
unknowns. The arising singularity of the station adjustment removed Hansen by
an addition of a so-called restriction equation. Based on this procedure – already
in the year 1867 – a method was developed, that is used today for the (numerical)
computation of pseudoinverses.
After a watch-maker’s apprenticeship Hansen because of his astronomical in-
terests became a collaborator of Heinrich Christian Schumacher, the director of the
Altona observatory (Altona is nowadays a suburb of Hamburg). Here, P.A. Hansen
was able to gain know-how and experience in astronomical computations and ob-
servation techniques also in solving geodetic problems, which he could acquire
during the Danish arc measurement of that time. In 1825 he was on recommen-
dation of Gauss and Schumacher appointed as director of the Seeberg observatory
near to Gotha, Duchy of Sachsen-Coburg-Gotha in Thuringia. Here, he was in
charge of the surveying of the Duchy of Sachsen-Coburg-Gotha. The results of
this triangulation served as a basis for the measurements to establish a cadastre.
For this purpose the net, represented in Figure 5.3, had to be condensed by a low
order triangulation and by traverses. By the following measurements of parcels a
cadastral map was the final result. The triangulation net necessary for this task
had to be adjusted by the method of least squares. During this work he hit on the
problem to overcome the singularity of the station normal equations. His works
of this period and the further development of Gauss’s method of least squares
were published in “Abhandlungen der Mathematisch-Physikalischen Klasse der
Königlich-Sächsischen Gesellschaft der Wissenschaften, Leipzig”, (see [99–102]).
Robinson’s Thinking about the Origin of Pseudoinverses. The idea of a pseudo-
(or generalized) inverse may be briefly motivated as follows: Suppose that
A ∈ Rm×n , A = (Aj,l )j=1,...,m (5.3)
l=1,...,n
is understood as a linear mapping from Rn into Rm . If y = (y1 , . . . , ym )T in Rm is

known, then the problem is to find x = (x1 , . . . , xn )T in Rn such that
x → y = Ax (5.4)
Figure 5.3. The Hansen triangulation net of the Duchy of Sachsen-

Coburg-Gotha; point 21 (Inselsberg) is identical with the same point of
the Gaussian arc measurement net (for comparison see Figure 6.1), point
11 is the location of the Seeberg observatory. In contrast to Gauss’s mea-
surement campaign Hansen’s net served as a basis for the establishment
of a cadastre.
holds true. If n = m and A ∈ Rm×n is invertible, with inverse A−1 , then the
solution clearly is x = A−1 y. But, in case of n = m, the problem is a proper
understanding of the “invertibility” of Ax = y, A ∈ Rm×n , x ∈ Rn , y ∈ Rm . The
question is if there will be a mapping A† , called a generalized inverse (usually
called, pseudoinverse) of A, such that x = A† y (so that A† = A−1 in case of
n = m)?
During the last decades of the last century, the concept of a pseudo- (or
generalized) inverse has rated considerable attention in the mathematical as well
as geodetic literature (a bibliography, for example, listing over 1700 references on
the subject is due to M.Z. Nashed [163]). Early interest in the first half of the
last century in the subject of pseudoinverses was initiated by a paper on matrices
by R. Penrose [172]. However, this concept had been considered somewhat earlier.
For example, E.H. Moore [152] presented a development of the notion (see also
[4, 9, 68, 153, 169, 178, 200]). Moreover, in the setting of integral and differential
operators the concept was considered even earlier by I. Fredholm [34] and W.A.
Hurwitz [116], and by D. Hilbert [111] (see [180] for a discussion of generalized
inverses in classical analysis, and see also [5, 14, 163] for brief historical sketches
of the subject).
The relation between Gauss’s ideas and the concept of the pseudoinverse
was discussed by D.W. Robinson [183]. His paper attempts to show that although
Gauss did not formalize the notion of a pseudoinverse, he provided the essential
ingredients to produce one. Next we follow this approach almost literally, however,
formulated within a today’s notational framework of linear algebra.
The point of departure for a mathematical concretization of Gauss’s role is
what is usually called today the full-rank linear model, which can be described as
follows: Given a set L1 , . . . , Ln of linear functionals on Rn , a set y = (y1 , . . . , ym )T
of observations, a set e = (e1 , . . . , em )T of errors, and a set w = (w1 , . . . , wm )T of
positive numbers. The problem is to find x = (x1 , . . . , xn )T that makes the errors
e = (e1 , . . . , em )T = (L1 x − y1 , . . . , Lm x − ym )T
in a certain metric, as small as possible, subject to the condition that the weights
constituting the metric influence the precision of the respective equations. It should
be noted that Gauss considered the adoption of several possible principles to solve
this problem, but finally argues for the minimization of “the sum of the squares
of the differences between the observed and the computed values multiplied by
numbers which measure the degree of precision” (see [72]). In more detail, Gauss
sought to minimize the sum

m
m
wj e2j = wj (Lj x − yj )2 (5.5)
j=1 j=1
of the weighted squares of the errors.

In terms of modern notation we are led to the following context: Let Rm be
equipped with an inner product defined by

m
d, z = wj dj zj , d ∈ Rm , z ∈ Rm , (5.6)
j=1
with associated norm

⎛ ⎞1/2
m
d = d, d1/2 =⎝ wj d2j ⎠ . (5.7)
j=1
If y = (y1 , . . . , ym )T is a given element of Rm , and if L1 , . . . , Lm is used to provide

the linear mapping A : Rm → Rn , x → Ax = (L1 x, . . . , Lm x)T , by letting

n
Lj x = Aj,l xl , j = 1, . . . , m, (5.8)
l=1
then the problem is to minimize

n 2

m
Ax − y = Ax − y, Ax − y1/2 = wj Aj,l xl − yj . (5.9)
j=1 l=1
C.F. Gauss [72] provided what he termed a “very expeditious algorithm” for
solving this problem. He argued that if AT ∈ Rn×m means what is now called the
adjoint of A, i.e., AT = (Al,j ) l=1,...,n, , then the minimization problem is equivalent
j=1,...,m
to the solution of the system
AT Ax = AT y (5.10)
with A A ∈ R
T n×n T T
. These equations A Ax = A y are known to be the normal
equations, which Gauss solved by the process of elimination to obtain the unique
solution denoted by b = (b1 , . . . , bn )T . Thus, the component bi was Gauss’s choice
for the best estimate of xi , i = 1, . . . , n. Although Gauss did not formalize the
notion of linear rank, it is clear from his context that he wished to consider only
the case, where the rank of A is so that n ≤ m.
At this stage, D.W. Robinson [183] comes to the following conclusion: In
the “Theoria motus”(1809), Gauss did not hint at the idea of a pseudoinverse.
However, this work was only his first on the subject of leastsquares. In 1821 he
presented to the Royal Academy of Sciences in Göttingen the first part of his
“Theoria combinations”, followed by the second part in 1823 and a supplement in
1826. His purpose in preparing this lengthy paper was to improve the foundations
of the theory of least squares ([77], Art. 17). He accomplished this by introducing
the dual inner product in the space (Rn )∗ of linear functionals. As a consequence,
the functional Pj , defined by
Pj : Rm y → Pj y = yj , j = 1, . . . , m (5.11)
has the norm

1
Pj = Pj , Pj = √ . (5.12)
wj
Thus, by considering once more the full-rank linear model n ≤ m, yj = Pj y is an
estimate of

n
Lj x = Aj,l xl = yj = Pj y (5.13)
l=1
−2
having the “weight” wj = Pj . In other words, the estimate of Lj x is given as
the image of y under the linear functional Pj , which satisfies Pj A = Lj with the
additional property that
wj = Pj −2 , j = 1, . . . , m. (5.14)
The idea is to require that, for every linear functional J in (Rn )∗ , the best estimate
of Jx be the image under y of the linear functional K in (Rn )∗ satisfying KA = J
and having the maximum weight K−2 .
More concretely, Gauss formulated what he termed the “problem”, which
may be rephrased in the language of inner product spaces as follows: Given Ji
satisfying Ji x = xi , find, among the linear functionals K : Rn → R with KA = Ji ,
the one having minimum norm. Gauss successfully solved this problem. In fact, he
obtained the explicit solution in the form
Ki = Ji (B T B)AT , i = 1, . . . , m, (5.15)
where B T B is the result of transforming by elimination z = AAT x into x =
(B T B)z (cf. [77], Art. 20 Eq. (4)). Thus, the best estimate of xi = Ki y was taken
to be Ki y = Ji (B T B)AT , i = 1, . . . , m. Furthermore, Gauss argued that this best
estimate Ki y was equal to the value bi obtained by the method of least squares
([77], Art. 21). By composing these functionals together, if
A† : Rm → Rn , y → (K1 y, . . . , Kn y)T , (5.16)
then b = A† y minimizes Ax − y. In this sense, A† may be recognized as a
pseudoinverse of A.
In conclusion, D.W. Robinson [183] was led to the following statement: Gauss
did not formally display A† in his 1821 paper. The ingredients for the construction
of a pseudoinverse were essentially available to him, but he did not use them toward
this end. Indeed, there appears to be no evidence that he was inclined to proceed
in that direction. On the other hand, his approach to the problem of determining
best estimates is certainly in the spirit of pseudoinverses. This is suggested by the
diagrams (again a contemporary device), when A is associated to (L1 , . . . , Lm )
and A† with (K1 , . . . , Kn ):
‫ܣ‬ ‫ܣ‬ற
௡ ௠ ௡
Թ Թ Թ Թ௠
ࣦ௝ ࣪௜ ‫ܬ‬௜ ࣥ௜
Թ Թ
Figure 5.4. Diagram illustrating the roles of A and A† in Gauss’s

approach. Specifically, the idea of “mapping back” is suggested in Figure
5.4 (cf. [183]).
Moreover, in the supplement to “Theoria combinationis”, C.F. Gauss (1826)

showed that the solution b which minimized Ax − y was expressible as the image
of y under a mapping depending only upon A.
In the opening paragraph of Article 8 of this supplement, Gauss stated his
objective:
“When one wants to find the most likely values of several unknowns,
depending on the same observations, or when one does not know which
unknowns it is preferable to derive from the observations, it is convenient
to proceed in a different way” (translation from D.W. Robinson [183]).
The “different way” was to solve the normal equations AT Ax = AT d ex-
plicitly. He did so by letting z = AT Ax and obtained x = B T Bz by elimination.
Although he did not use the formal “inverse”, he appreciated that B T B was a
function of AT A and noted, in particular, that AT A was symmetric. He then ex-
pressed the solution of the normal equations in the form b = (B T B)AT y (see the
supplement to [77], Art. 8). Consequently, his objective was satisfied: b was the
image of y under (B T B). In other words, interpreted in contemporary language,
(B T B)AT is a pseudoinverse of A. Thus Gauss’s approach to this problem was con-
sistent with the objectives of the theory of pseudoinverses, and his explicit solution
is readily identified with the generalized inverse representation b = (AT A)† AT y
used today (see, for example, [163]).
Finally, once more following D.W. Robinson [183], the two pseudoinverses
identified above are the same; in fact, they are equal to what is called today
the pseudoinverse (or Moore–Penrose) inverse. Indeed, by using the standard
definition, for an inner product of two linear mappings, Gauss’s problem may
be viewed as a functional version of the following problem: Given the identity
map I : Rn → Rn , find among the linear mappings K : Rm → Rm , satisfying
KA = I, the one having minimum norm. Since A† is the solution to this prob-
lem, it is what is called the best approximate solution of the equation KA = I,
and in this case it is the pseudoinverse (generalized or Moore–Penrose) inverse
of A (cf. [172]). Moreover, it follows from Gauss’s explicit solution Ki of the best
linear estimate problem and the fact that (J1 x, . . . , Jn x) = x, that, for every y
in Rm , A† y = (K1 y, . . . , Kn y)T . That is A† = (B T B)AT . Consequently, we are
able to conclude that, under the conditions of the Gaussian full-rank linear model,
A† , (B T B)AT , and the Moore–Penrose inverse of A are one and the same operator.
All in all, the conclusion of D.W. Robinson [183] is as follows:
“While generalized inverses were not a part of Gauss’s vocabulary, equiva-
lent expressions may be found in his writings. Specifically, in his solution
of the problem of least squares, Gauss established explicit formulas which
may readily be identified with generalized inverses. Indeed, by translat-
ing his analytical formulation of the problem into the more geometrical
framework of vector spaces and linear mappings, the features of gener-
alized inverses are easily recognized. In particular, Gauss’s development
of best linear estimates was in the spirit of generalized inverses. This
observation suggests the possibility that Gauss’s view of least squares
in 1826 may have been more geometrical than the analytical form of
his presentation could express. He may well have conceived best linear
estimates in some mode of geometrical thought, but communicated the
results in the accepted and more rigorous analytical mode of the day,
since the geometrical tools of the early 19th century were limited, and
the conceptual framework needed to develop a theory of generalized in-
verses was not available to him.”
Hints for Further Developments. Next we list some areas of today’s mathematics
which can be essentially traced back to Gauss’s work:
(a) The probability distribution function (“bell curve”) introduced by Gauss has
as normal distribution obtained a special significance in statistics, because
theory and practice of confidence intervals and the resulting hypothesis tests
were not possible without the Gaussian distribution (see [236]).
(b) According to W. Jordan, the least squares method can be used to solve opti-
mization problems by an iterative process, if the observations are redundant,
e.g., for the detection of optimal weight distributions, a problem, which can
primarily be settled by use of the Boscović–Laplace method respectively, the
simplex algorithm of linear programming (cf. [236]).
(c) Also the procedures of spherical and spherical-harmonic prediction and col-
location, which made their headway by H. Moritz [154, 155] are based on the
method of least squares (note that “least squares collocation” as it is called in
geodesy is just named “minimum norm interpolation” in mathematical lan-
guage, while “least squares adjustment” as understood in geodesy is nothing
more than “smoothing”, e.g., in the sense of C. Reinsch [181]). Furthermore,
the prediction method can be considered as a kind of inter- or extrapolation
to determine a trend function.
These approximation methods were transferred and widely extended to
the spline context by W. Freeden and many others [39, 40, 42, 43, 62, 64, 95,
96, 182, 199, 221].
Clearly, the list of examples of Gaussian ideas as presented here is rather
incomplete. Furthermore, our approach can be deepened by far, which will be
done partly later in a functional analytic jargon.
5.2. Bridge between least squares solutions and pseudoinverses

Until now, our historic survey on least squares and pseudoinverse facets has been
formulated within the context of the original framework. In what follows we would
like to answer the question what exactly enabled mathematicians to build the
bridge from Gaussian least squares theory to essential settings of today’s theory
of inverse problems. To this end, the following two question (Q1) and (Q2) have
to be answered:
(Q1) What can be specified as mathematical structures and settings from least
squares by a process of abstraction, i.e., a reduction to essential features?
(Q2) What can be achieved from the reduced context obtained by abstraction from
least squares theory as a new field of scientific interest, e.g., in geodesy or
other sciences?
The answers can be given in a three step (S1), (S2), (S3) procedure:
(S1) We specify the transfer from least squares solutions to the theory of the
pseudoinverse in a finite-dimensional matrix calculus. The pseudoinverse is
seen, comparably, but more generally to the Gaussian approach, as the best-
approximate solution within the set as least squares solutions in finite-dim-
ensional settings.
(S2) In turn, the reduction of the pseudoinverse to its specific properties involving
functional analytic means based on operator theoretical background leads us
to the concretization of the pseudoinverse in finite-dimensional spaces that
can be extended to infinite-dimensional spaces in a straightforward way.
(S3) As a consequence, we shall be able by use of an infinite-dimensional opera-
tor calculus to solve ill-posed problems for (compact) operator equations by
regularization. In particular, dilemmas and methodologies of resolution of ill-
posed problems become obvious with particular reference to the problem of
finding minimum norm least squares solutions of first kind integral equations
(and, more generally, of linear operator equations with non-closed range).
5.3. Pseudoinverse for finite-dimensional matrix equations
Since any numerical approximation procedure usually leads to finite-dimensional
problems involving a singular functional analytic context, we first consider finite
systems of linear equations (see, e.g., M.Z. Nashed [163, 164] for more details).
After having treated the finite-dimensional situation, we turn to the analogous
theory in operator framework.
Spectral Matrix Representation. We start with a linear matrix equation of the
form
Ax = y, (5.17)
with A ∈ K n×n
being a Hermitian positive definite matrix with n rows and n
columns. From the spectral theory (see, e.g., [226]) it is well known that there
exist eigenvalues 0 < λ1 ≤ · · · ≤ λn and a corresponding unitary matrix U =
(u1 , . . . , un ) of eigenvectors ui ∈ Cn \{0} (i.e., uH
i uj = δij ) such that A has a
representation of the form
⎛ ⎞
λ1 0
⎜ .. ⎟
⎜ . ⎟
H
U AU = ⎜ ⎜ ⎟ = diag(λ1 , . . . , λn ). (5.18)
. ⎟
⎝ .. ⎠
0 λn

n×n
The condition number of A is given by the quotient of the largest and smallest
eigenvalue, i.e., κ = λλn1 (note that λ1 > 0). For the sake of simplicity and coherence
with the analysis for the infinite-dimensional case below, we shall assume here that
the scaling is such that λn = 1, so that κ = λ−11 . The condition number is a measure
for the stable solvability of the problem (5.17).
Ill-Conditioned Matrix Equations. Assume that we have noisy data y ε instead of

y, which satisfy the Euclidean norm estimate
y ε − y ≤ ε (5.19)
in the Euclidean norm on K . Let x denote the solution with right-hand side y ε .
n ε
Then it follows from the spectral representation that

1 1
x − x = U diag
ε
,..., U H (y ε − y). (5.20)
λ1 λn
Hence, observing the orthogonality of eigenvectors we are led to the estimate

n
xε − x2 = λ−2 H ε 2 −2 ε 2
i |ui (y − y)| ≤ λi y − y . (5.21)
i=1
In other words, we have

xε − x ≤ κy ε − y ≤ κ ε. (5.22)
The sharpness of this estimate can be immediately seen for y − y = εu1 . It is clear
ε
that with increasing condition number of the matrix A, the noise amplification
increases in the worst case. For large κ one therefore speaks of an “ill-conditioned
problem” (IPP). But it should be remarked that a finite-dimensional linear problem
is never ill-posed (in the sense that the third condition in Hadamard’s classification
is violated), but for κ large one certainly comes close to this case.
We also observe that errors in low frequencies (i.e., corresponding to eigen-
vectors with large eigenvalues) are amplified less. Following our nomenclature we
see that an error in the lowest frequency, i.e., y ε − y = ε un , is not amplified at
all. In fact, we just obtain xε − x = ε from the spectral representation. This
is a typical effect for inverse problems. It means that not all possible versions of
noise of the same size are equally bad, high-frequency noise corresponding to low
eigenvalues is always worse than low-frequency noise. However, in practice, we
are able to make any assumption on the noise only in rare exceptions, so that a
regularization method has to deal with arbitrary noise.
Until now, we have assumed that the matrix A is Hermitian positive definite,
i.e., the minimal eigenvalue is positive. If this is not the case, the matrix has a
non-trivial null space. If λr denotes the minimal non-zero eigenvalue, then the
solution formula becomes
n
x= λ−1 H
i ui ui y, (5.23)
i=r
and the problem is solvable if and only if uH i y = 0 for i < r. If the data set is
noisy, i.e, instead of y we have y ε , we are led to use the projection P y ε onto the
range of A. In doing so we obtain for the corresponding solution xε with data P y ε
that
n
xε − x = λ−1
i ui ui (P y − y).
H ε
(5.24)
i=r
i P y = ui y for i ≥ r we thus can estimate similarly as described above

Since uH ε H ε
xε − x ≤ λr ε. (5.25)
Consequently, there is no error propagation in the null space components and the
noise amplification is actually determined by the minimal nonzero eigenvalue.
Matricial Pseudoinverse (Generalized Inverse, Moore–Penrose Inverse). Let A ∈
Kn×m be a matrix with n rows and m columns, y ∈ Km . Note that A is not
required to be square, no rank assumptions are made at this stage. Then we know
that the linear system
Ax = y, x ∈ Kn , (5.26)
needs not have a (unique) solution. If (5.26) is unsolvable, a reasonable generalized
notion of a solution is a “least square solution”, which minimizes the residual Ax−y
in the Euclidean norm (once more, note that stands for the Euclidean norm
in this subsection): A vector x ∈ Kn is called a
(1) least squares solution of (5.26) if and only if
Ax − y = inf{Az − y : z ∈ Kn }, (5.27)
(2) best-approximate-solution (or minimal norm solution) of (5.26) if and only
if x is a least squares solution and
x = inf{z : z is a least square solution}. (5.28)
As already mentioned, one could also use other norms in (5.27) and in (5.28),
which would lead to different notions of a generalized solution. Also, instead of
minimizing z in (5.28), it is often of interest to minimize T z for some pre-
scribed matrix T .
The following results are well known from classical linear algebra (see, e.g.,
[226]):
(i) A vector x∗ is a solution of (5.27) if and only if the “normal equations”
AH Ax∗ = AH y are satisfied.
(ii) The problem (5.27) possesses a unique solution if and only if A has full rank.
Our purpose is to show that a best-approximate solution in the sense of (2)
always exists and is unique such that the following definition makes sense:
Definition 5.1. If A† ∈ Km×n is understood as the matrix which assigns to each
y ∈ Kn the best-approximate solution of (5.26), then it is called the pseudoinverse
(Moore–Penrose or generalized inverse) of A.
Our aim is to construct A† and, hence, best-approximation solutions via the
so-called singular value decomposition (SVD) of A. To this end we first recall the
definition of the singular values of a matrix A.
Definition 5.2. Let σ1 , . . . , σr > 0 be such that σ12 ≥ σ22 ≥ · · · ≥ σr2 > 0 are the
positive eigenvalues of the matrix AH A (each one written down as often as its
multiplicity is). Then σ1 , . . . , σr are called the “(non-zero) singular values of A”.
This definition makes sense, since AH A is positive semidefinite. Obviously,

r ≤ min{n, m}, where r is the rank of A.
We know that a Hermitian matrix can be diagonalized, where the diagonal
elements are its eigenvalues. The following theorem generalizes this result to the
non-Hermitian case.
Theorem 5.1. Let σ1 ≥ σ2 ≥ · · · ≥ σr > 0 be the singular values of A. Then there
exist unitary matrices U ∈ Km×m and V ∈ Kn×n such that
⎛ ⎞
σ1 0
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ σr ⎟
V H AU = ⎜
⎜
⎟
⎟ (5.29)
⎜ 0 ⎟
⎜ .. ⎟
⎝ . ⎠
0 0

m×n
holds true. The columns of U and V are eigenvectors of AH A and AAH , respec-
tively. The expression (5.29) is called “singular value decomposition (SVD) of A”.
The singular value decomposition (SVD) is not unique, since the unitary
matrices U and V are not. Obviously, from (5.29), we obtain
⎛ ⎞
σ1 0
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ σr ⎟ H
A=V ⎜ ⎜
⎟U ,
⎟ (5.30)
⎜ 0 ⎟
⎜ .. ⎟
⎝ . ⎠
0 0
since V and U are invertible, and the rank of A is r, where r is the number of
non-zero singular values (counted with multiplicity).
Remark 5.1. In accordance with Theorem 5.1 one has to compute the singular
values of A, e.g., as the positive square roots of the eigenvalues of AH A, which
can be done, e.g., by the QR-algorithm. However, since AH A usually has a worse
condition than A, one should (in critical cases) use a variant of the QR-method
that does not use AH A explicitly.
Note that with U = (u1 , . . . , um ), V = (v1 , . . . , vn ), we have for i ∈ {1, . . . , r}
Aui = σi vi (5.31)
and
AH vi = σi ui , (5.32)
which follows from the singular value decomposition (5.29) via multiplication by V
and U H , respectively. The system {(σi ; ui , vi ) : i ∈ {1, . . . , r}} is called a “singular
system” for A. Since, as we have seen in the proof of Theorem 5.1, {v1 , . . . , vr } is
an orthonormal basis for the range R(A) = {y : Ax = y}, for any x ∈ Kn ,
r
r
Ax = Ax, vi vi = x, AH vi vi , (5.33)
i=1 i=1
which implies in connection with (5.32) that
r
Ax = σi x, ui vi (5.34)
i=1
holds true. Analogously, for all y ∈ K , m
r
AH y = σi y, vi ui . (5.35)
i=1
Remark 5.2. The notion of a singular system and the expansion (5.34) and (5.35)
generalize to compact operators on infinite-dimensional spaces, e.g., integral oper-
ators, as we will see later on.
Remark 5.3. Note that if A has real entries, so U and V have.
Now we relate the SVD to the Moore–Penrose inverse. Moreover, we show
that SVD can be used to compute the best-approximate solution.
Theorem 5.2 (Pseudoinverse). Let A have the SVD (5.29). Then
⎛ 1 ⎞
σ1 0
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ 1 ⎟ H
A† = U ⎜
⎜
σr ⎟V .
⎟ (5.36)
⎜ 0 ⎟
⎜ .. ⎟
⎝ . ⎠
0 0

n×m
Theorem 5.2 also implies the existence and uniqueness of a best-approximate

solution.
Since A† b is the least squares solution of minimal norm, we obtain that A† b
is a solution of the normal equations AH Ax = AH b with minimal norm, i.e.,
A† y = (AH A)† AH y. (5.37)
†
This means that in order to approximate A b we may as well compute an ap-
proximation to the minimal-norm solution in the normal equations, a fact we will
heavily use in the construction of regularization methods, later on.
For the case of overdetermined linear systems, i.e., (5.26) with m ≥ n the null
space N (A) = {x : Ax = 0} only consists of {0}, so that the best-approximation
solution of (5.26) can effectively be computed via Householder transformations
(see, e.g., [227]) in the following way: One transforms A by n − 1 Householder
(or other orthogonal, e.g., Givens) transformations to a n × n-triangular matrix R

(and 0 from the (n + 1)st row on) and applies the same transformations to b. If Q
is the unitary matrix representing these transformations, we are led to

R h1
QA = , Qb = . (5.38)
0 h2
For any x ∈ C n ,
- -
- Rx − h1 -
Ax − b = Q(Ax − b) = -
- −h2
-.
- (5.39)
Consequently, the best-approximate solution can be computed in the form A† b =

R−1 h1 (since R is triangular, R−1 can be computed easily by substitution).
Continuous Dependence While for the case of an invertible matrix A, A† b = A−1 b
depends continuously not only on b, but also on A, this is not the case in general:
Consider, for example, the matrix

1 0
A() = ; (5.40)
0
then
1 0
A()† = (5.41)
0 1
for = 0, and we have

† 1 0
A(0) = , (5.42)
0 0
so that, e.g.,

† 0 † 0
A() A(0) (5.43)
1 1
as → 0.
Truncated Singular Value Regularization. The identity (5.36) also shows how er-
rors in y affect the result A† y: Errors in components of y corresponding to small
singular values are amplified by the large factor of the singular value, so that such
data errors are dangerous. This explains the numerical instability of (5.36), if A has
small singular values. Although the problem of computing the best-approximate
solution is well posed, it is then numerically unstable. The first idea to reduce this
instability is to replace (5.36) by
r
y, vi
xα = ui (5.44)
i=1
σi
σi2 ≥a
with an appropriately chosen value α > 0; this truncation is the first example of
a “regularization”, where the original problem is replaced by a neighboring one,
which is more stable. However, the choice of the “regularization parameter” α is
quite crucial.
If we use (5.44) with perturbed data y ε (with y − y ε ≤ ε), we obtain as the

“regularized solution”
r
y ε , vi
xεα = ui . (5.45)
i=1
σi
σi2 ≥α
We estimate the total error between xεα and the sought-for quantity A† y:
- - - -
- y, vi - - r -
- - y, vi − y ε , vi -
r r
† - y, vi
xα − A y ≤ -
ε
ui − ui - + - ui - .
- σi σi - - σi -
i=1 i=1 i=1
σi2 ≥α σi2 ≥α
(5.46)
For sufficiently small α, the first summand on the right-hand side is empty. For
the second summand we have
- r -
- y, v − y ε , v -2 r
|y − y ε , vi |2 1
r
ε2
- -
|y − y ε , vi |2 ≤ .
i i
- ui - = 2 ≤
- σi - σi α i=1 α
i=1 i=1
σi2 ≥α σi2 ≥α
(5.47)
The sum (5.45) is called truncated singular value expansion. It can be inter-
preted as applying a low-pass filter to the data.
Tikhonov Regularization. Another way of making (5.36) more stable would be to
replace it by the sum
r
σi
xεα = 2 + α y , vi ui .
ε
(5.48)
σ
i=1 i
This is a famous Tikhonov regularization method (cf. [211–215]), which we shall
consider in more detail in infinite dimensions, too. It is helpful to characterize it
in a different way: If xεα is defined by (5.48), then it is not hard to see that
H
A A + αI xεα = AH y ε , (5.49)
which is an alternative characterization of the Tikhonov regularization. From this
“regularized normal equation”, we can obtain still another characterization of xεα ,
namely as the unique minimizer of the so-called Tikhonov functional
x → Ax − y ε 2 + αx2 , (5.50)
which can be seen by putting the first derivative of the functional in (5.50) to 0,
resulting exactly in the linear equation (5.49). The minimization of (5.50) can be
seen as a combination of the two minimizations that appear in the definition of a
best-approximate solution. It has also interpretations as a penalty method, e.g.,
via Lagrange multipliers.
The computation explained above can also be performed for α = 0 (with
(5.30) replaced by (5.36)). In this case it shows that x is the solution of minimal
norm of the normal equations
AH Ax = AH y, (5.51)
which was already attacked by C.F. Gauss [72], (see also the contributions by R.L.
Plackett [175], D.W. Robinson [183]). If AH A is invertible (and hence positive
definite), the normal equations (5.51) can be solved by standard Cholesky decom-
position, which leads to an alternative method for computing best-approximation
solutions, for which no SVD is needed. However, as mentioned above, forming
AH A may seriously decrease the numerical stability. Hence, it should be avoided
in ill-conditioned cases.
5.4. Today’s functional analytical nomenclature
The following nomenclature is part of any functional analytical textbook (see, e.g.,
[120, 210, 241]). A mapping A : X → Y between two normed spaces X and Y is
called an operator. With
N (A) = {x ∈ X : Ax = 0} ⊂ X (5.52)
and
R(A) = {y = Ax : x ∈ X} ⊂ Y (5.53)
we denote the kernel (or the null space) of A and the image of A, respectively.
Definition 5.3. Let X and Y be normed spaces. The operator A : X → Y is called
linear, if
(1) A(x + y) = Ax + Ay for all x, y ∈ X,
(2) A(αx) = αAx for all x ∈ X and α ∈ K.
The operator A is called bounded, if there exists a constant C ≥ 0 such that
AxY ≤ C xX for all x ∈ X.
Theorem 5.3. Let X and Y be normed spaces and A : X → Y be a linear operator.
Then the following statements are equivalent:
(1) A is continuous on X.
(2) A is continuous in 0 ∈ X.
(3) A is bounded on X.
The space of all continuous linear operators between X and Y is denoted by
L(X, Y ). If Y = X we set L(X) = L(X, X). A norm on L(X, Y ) is given by
AxY
A = sup = sup AxY = sup AxY . (5.54)
x =0 xX xX ≤1 xX =1
In order to be more precise we often use the notation AX→Y for the norm of an
operator A : X → Y .
Theorem 5.4. Together with the norm (5.54) the space L(X, Y ) is a normed space.
If X is a normed space and Y is a Banach space, then L(X, Y ) is a Banach space.
For combinations of two linear operators we have the following result.
Theorem 5.5. Let A ∈ L(X, Y ) and B ∈ L(Y, Z), then BA ∈ L(X, Z) and we have
BAX→Z ≤ BY →Z AX→Y . (5.55)
For A ∈ L(X) we get iteratively

n
An ≤ A (5.56)
for all n ∈ N.
For sequences of operators two different terms of convergence are used.
Definition 5.4. Let {An }n∈N ⊂ L(X, Y ) be a sequence of operators between X and
Y and let A ∈ L(X, Y ).
(1) A sequence of operators {An }n∈N is called pointwise convergent to an operator
A if
lim An x − AxY = 0 (5.57)
n→∞
for all x ∈ X.
(2) A sequence of operators {An }n∈N is called uniformly convergent to an oper-
ator A if
lim An − AX→Y = 0. (5.58)
n→∞
Uniform convergence implies pointwise convergence but the converse is in

general not true.
Since we mainly study inverse problems, the inversion of operators are of
particular interest. For the existence of a continuous inverse of an operator we
have the following equivalent criterion.
Theorem 5.6. The linear operator A : X → Y between the normed spaces X, Y
has a continuous inverse A−1 ∈ L(R(A), X) if and only if there exists a constant
c > 0 such that
c xX ≤ AxY for all x ∈ X . (5.59)
In this case, - −1 -
-A - ≤ c−1 . (5.60)
As for the existence of a bounded inverse of an operator A we also have an
equivalent criterion for the non-existence of a bounded inverse.
Theorem 5.7. An operator A ∈ L(X, Y ) does not have a continuous inverse if
and only if there exists a sequence {xn } ⊂ X with xn X = 1 for all n ∈ N and
limn→∞ Axn = 0.
Since K = R or K = C, the space L(X, K) is clearly a Banach space. It is
symbolized by X ∗ and called dual space of X. The elements of X ∗ are called linear
functionals.
Well-Posedness in the Sense of Hadamard. A mathematical model is a mapping
A : X → Y from the set (of causes) X to the set (of effects) Y . D(A) = X,
R(A) = {y = Ax : x ∈ X} and N (A) = {x ∈ X : Ax = 0} are the domain (of
definition) of A, the range (or image) of A, and the null space (or kernel) of A,
respectively. We are interested in investigating the operator equation
Ax = y, x ∈ X, y ∈ Y. (5.61)
A direct problem amounts to the description and evaluation of y, i.e., Ax. An

inverse problem is the task to find, for a given element y ∈ Y , an element x ∈ X
such that (5.61) holds true.
Definition 5.5 (Hadamard’s Classification). Let X, Y be metric spaces. Suppose

that A : X → Y is a mapping from X to Y . The problem (A; X, Y ) is called well
posed (in the sense of Hadamard) if
(H1) For every y ∈ Y there exists an x ∈ X with Ax = y (Existence of the Inverse).
(H2) For every y ∈ Y there exists one and only one x ∈ X with Ax = y (Uniqueness
of the Inverse).
(H3) The inverse mapping A−1 : Y → X is continuous, i.e., the solution x ∈ X of
Ax = y depends continuously on y ∈ Y (Continuous Dependence).
If one of the three properties is violated, then the problem is called ill-posed (in
the sense of Hadamard).
Using a singular system in finite-dimensional context, x can also be written

r
as the sum σi−1 y, vi ui , such that
i=1

r
y, vi
A† y = ui , (5.62)
i=1
σi
which will also generalize to the infinite-dimensional setting.

The matrix AH A is invertible if N (AH A) = N (A) = {0}. Then (cf. [152,
153]) we have
A† = (AH A)−1 AH . (5.63)
The SVD of A immediately lead us to the famous “Moore–Penrose equations”,
which also characterize A† :
A† AA† = A† , (5.64)
†
AA A = A, (5.65)
(A† A)H = A† A, (5.66)
† H †
(AA ) = AA . (5.67)
It can also be seen that A† A and AA† are orthogonal projectors onto N (A)⊥ and
R(A), respectively. These facts, in turn, can be used to characterize A† in yet
another way, namely as the unique linear operator satisfying
. /−1
A† |R(A) = A|N (A)⊥ (5.68)
and
N (A† ) = R(A)⊥ . (5.69)
In other words, the Moore–Penrose (generalized) inverse is the unique linear ex-
0 1−1
tension of A|N (A)⊥ to
A† = R(A) ⊕ R(A)⊥ (5.70)
satisfying (5.69). Moreover, we have
Theorem 5.8 (Pseudosolution). For each y ∈ D(A† ), the equation (5.26) has a
unique best-approximate-solution (pseudosolution) given by
x† = A† y, (5.71)
†
where the set of all least squares solutions is given by {x } + N (A).
5.5. Pseudoinverse for infinite-dimensional operator equations
Next our goal is to extend some material known from matrix analysis in (finite-dim-
ensional) Euclidean spaces to linear operator framework between Hilbert spaces.
Central in our considerations is the introduction of the pseudoinverse. We begin
with some preparatory remarks clarifying the functional analytic background in
order to make the Hilbert space (HS) context of the theory of ill-posed problems
(IPP) more transparent:
Let X and Y be Hilbert spaces and let A : X → Y be a bounded linear
operator whose range R(A) is not necessarily closed. Then we have the orthogonal
decompositions
X = N (A) ⊕ N (A)⊥ , A = R(A) ⊕ R(A)⊥ (5.72)
and
N (A∗ ) = R(A)⊥ , (5.73)
where N (A) is the null space of A, R(A) is the closure of the range of A, and A∗
is the adjoint operator of A, i.e., Ax, y = x, A∗ y for all x ∈ X and y ∈ Y .
Four (mutually exclusive) situations (S1), . . ., (S4) arise in considering the
operator equation (cf. [165])
Ax = y, x ∈ X, y ∈ Y, (5.74)
involving Hilbert spaces X, Y :
(S1) R(A) is dense in Y , (so N (A∗ ) = {0}), and y ∈ R(A);
(S2) R(A) is dense in Y , and y ∈/ R(A);
(S3) R(A) is a proper subspace of Y , and y ∈ R(A) ⊕ R(A)⊥ ;
(S4) R(A) = Y , and y ∈ / R(A) ⊕ R(A)⊥ .
In case (S1), the operator equation (5.74) has, of course, a solution in the classical
sense; in cases (S2) and (S4), a classical solution does not exist, while in case (S3)
a solution need not exist. Later on, in analogy to finite-dimensional settings, we
shall discuss that x is a “least squares solution” of the operator equation (5.74) if
inf{Az − y : z ∈ X} = Ax − y. Since Az − y2 = Az − Qy2 + y − Qy2 ,
where Q is the orthogonal projector PR(A) of Y onto R(A), we are led to the fact
that a least squares solution exists if and only if y ∈ R(A) ⊕ R(A)⊥ , which is a
dense set in Y . For such y the set of all least squares solutions of Ax = y, denoted
by L(y), is a non-empty closed convex set (indeed, L(y) is the translate of N (A)
by a fixed element of L(y)), hence, has a unique element of minimal norm, denoted
by A† y.
Hadamard’s and Nashed’s Classifications. It turns out that the statements (H1),
(H2), and (H3) of Hadamard’s classification (Definition 5.5) are satisfied by the
Moore–Penrose inverse A† if and only if R (A) = R (A). This observation leads to a
new notion of well-posedness that goes back to Nashed (for more details see [166]).
Definition 5.6 (Nashed’s Classification). The problem (A; X, Y ) is called well posed
in the sense of Nashed, if R (A) is closed in Y . Otherwise, i.e., if R (A) is not closed
in Y , the problem (A; X, Y ) is called ill posed in the sense of Nashed.
In accordance with M.Z. Nashed [164], we are led to say that the operator
equation (5.74) is well posed in the least squares (relative to X and Y ) if, for each
y ∈ Y , it has a unique least squares solution (of minimal norm), which depends
continuously on y; otherwise the problem is ill posed. The advantage of adopting
this notion of well-posedness is that it focuses on infinite-dimensional problems
(e.g., an inconsistent finite system of linear algebraic equations will not be ill
posed in our sense, while it is ill posed in the sense of Hadamard). Furthermore,
we are led to show by functional analytic means that the following statements are
equivalent:
(a) the operator equation (5.74) is well posed;
(b) R(A) is closed;
(c) A† is bounded.
As a consequence, it turns out that the pseudoinverse A† is the linear oper-
ator which assigns to each y ∈ D(A† ) = R(A) ⊕ R(A)⊥ , the unique element in
L(y) ∩ N (A)⊥ , so that L(y) = A† y + N (A). A† y is the minimal-norm solution
(equivalently the unique solution in N (A)⊥ ) of the normal equations
A∗ Ax = A∗ y (5.75)
(the equation obtained with Q = PR(A) by setting the first variation of Ax − y2
equal to zero). A† = (A/N (A)⊥ )−1 Q so that A† can be characterized as the linear
operator with the function-theoretic properties: D(A† ) = R(A)⊕R(A)⊥ , N (A† ) =
R(A)⊥ = N (A∗ ), and R(A† ) = N (A)⊥ .
In fact, we are led to the following conclusions: In case (S1) above, A† indeed
gives to the minimal-norm solution of Ax = y. In case (S3), Ax = y has a least
squares solution (which is unique if and only if N (A) = {0}). In both cases, the
infimum is attained and is equal to zero and y − Qy, respectively. Cases (S2)
and (S4) are pathological and are of no deeper interest in pseudoinverse theory
and since in both cases y ∈ / D(A† ), and the infimum is not attained.
After the general remarks we are now prepared to characterize least squares
solutions of operator equations in more detail: Let X and Y be Hilbert spaces
and A ∈ L (X, Y ), i.e., A is linear and bounded. Our interest is to determine an

element x ∈ X for a given element y ∈ Y such that the operator equation
Ax = y, x ∈ X, y ∈ Y (5.76)
becomes attackable (in least squares sense). An element x ∈ X is called a
(i) least squares solution of (5.76), if
Ax − yY = inf{Az − yY : z ∈ X} (5.77)
(ii) best-approximate solution (or minimal norm solution) of (5.76), if x ∈ X
solves (5.77) and
xX = inf {z : z is least squares solution of Ax = y} .
Obviously, the problem in the framework of L (X, Y ) involving Hilbert spaces
X and Y is to minimize the residual.
Theorem 5.9. Suppose that y ∈ Y and A ∈ L (X, Y ) with X, Y Hilbert spaces.
Then, the following statements are equivalent:
(1) x fulfills
Ax = PR(A) y, (5.78)
where PR(A) denotes the orthogonal projection on R (A).
(2) x minimizes the residual, i.e.,
Ax − y ≤ Ax − y for all x ∈ X. (5.79)
(3) x ∈ X solves the so-called normal equations
A∗ Ax = A∗ y. (5.80)

Proof. (1) ⇒ (2): Suppose that x ∈ X and y ∈ Y . Then we obtain PR(A) y − y ∈
⊥
R (A) , and we have
- -2 - -2 2 3
2 - - - -
Ax − y = -Ax − PR(A) y - + -PR(A) y − y - + 2 Ax − PR(A) y, PR(A) y − y
(5.81)
such that the Pythagorean theorem tells us that
- -2
2 - -
Ax − y = -Ax − PR(A) y - + Ax − y2
2
≥ Ax − y for all x ∈ X. (5.82)
2
(2) ⇒ (3): Suppose that x ∈ X and F (λ) = A (x + λx ) − y . The condi-
tion (2) tells us that F possesses a minimum for λ = 0. Therefore, we obtain

∂F ∂
0= (0) = (Ax + λAx − y, Ax + λAx − y)
∂λ ∂λ λ=0
= 2 Ax , Ax − y
= 2 x , A∗ Ax − A∗ y (5.83)
for all x ∈ X. Hence, we are able to conclude that

A∗ Ax − A∗ y = 0. (5.84)
(3) ⇒ (1): For y ∈ Y , let x ∈ X satisfy the equation
A∗ Ax = A∗ y ⇔ A∗ (Ax − y) = 0. (5.85)
Then, from functional analysis, it follows that
⊥
Ax − y ∈ N (A∗ ) = R (A)
⇒ PR(A) (Ax − y) = 0
⇒ PR(A) Ax = PR(A) y. (5.86)
In other words,
Ax = PR(A) y. (5.87)
This completes our proof.
Remark 5.4. The normal equations (5.80) owe their name to the property that
⊥
Ax − y ∈ R (A) , i.e., Ax − y ∈ N (A∗ ).
Theorem 5.10. Assume that y belongs to Y . Then the following statements are
true:
(1) The set of solutions of the normal equations
L (y) = {x ∈ X : A∗ Ax = A∗ y} (5.88)
⊥
is non-empty if and only if y ∈ R (A) ⊕ R (A) .
(2) L (y) is closed and convex.
Proof. (1) Assume that x ∈ L (y). Then we have y = Ax + (y − Ax) ∈ R (A) ⊕
⊥ ⊥
R (A) . Moreover, suppose that y ∈ R (A) ⊕ R (A) . Then there exists elements
⊥
x ∈ X and y4 ∈ R (A) satisfying
y = Ax + y4 (5.89)
and
PR(A) y = PR(A) Ax + PR(A) y4 = Ax + 0.
Thus, x satisfies the normal equations, hence, x ∈ L (y).
(2) Let {xn }n be a sequence in L (y) which converges to x ∈ X. Both opera-
tors A and A∗ are continuous. Therefore, we obtain
A∗ y = A∗ Axn , (5.90)
for all n ∈ N. Furthermore, for n → ∞,
A∗ y = A∗ Ax, (5.91)
which shows us that x ∈ L (y), i.e., L (y) is closed. Suppose now that x, x ∈ L (y)
and 0 ≤ λ ≤ 1. Then the identities
A∗ A (λx + (1 − λ) x ) = λA∗ Ax + (1 − λ) A∗ Ax
= λA∗ y + (1 − λ) A∗ y
= A∗ y, (5.92)
imply that L (y) is convex.
Remark 5.5. Generally, we have

⊥
R (A) ⊕ R (A) = Y. (5.93)
⊥
However, it should be noted that R (A) ⊕ R (A) = Y if the dimension of Y is
finite.
⊥
Lemma 5.1 (Pseudosolution). Suppose that y belongs to R (A) ⊕ R (A) . Then
L (y) possesses a unique element x† ∈ L (y) satisfying
- †-
-x - < x (5.94)
! †"
for all x ∈ L (y) \ x , i.e., there exists an element of minimal norm.
Proof. L (y) is non-empty, closed and convex. x† is the best-approximate element

to 0 ∈ X in L (y).
Remark 5.6. The reason for the proof of Lemma 5.1 is the theorem of best-
approximate elements. Let ∅ = U ⊂ X be closed and convex and x ∈ X. Then
there exists one and only one u ∈ U such that
x − uX < x − u X (5.95)
for all u ∈ X \{x}. The proof can be found in any standard textbook on functional
analysis (see e.g., [110]).
Pseudoinverse (Moore–Penrose Inverse, Generalized Inverse). After these con-
siderations concerned with least squares approaches we are in position to intro-
duce the pseudoinverse (Moore–Penrose inverse, generalized inverse) of operators
A ∈ L(x, y), where X and Y are Hilbert spaces.

Definition 5.7. The operator A† : R (A) ⊕ R(A)⊥ → X (i.e., D A† = R (A) ⊕
⊥
R (A) ⊂ Y ), that maps each element y ∈ D A† to the unique element x† ∈
L (y) with minimal norm, is called the pseudoinverse or Moore–Penrose inverse,
generalized inverse of A.

Theorem 5.11. Let y belong to D A† . Then x† = A† y is the best-approximate
solution of (5.76). It represents the unique solution of the normal equations in
⊥ ⊥
N (A) , i.e., x† = A† y if and only if A∗ Ax† = A∗ y and x† ∈ N (A) .
Proof. “⇒” Suppose that x† = A† y. Then x† solves the normal equations and has
⊥
minimal norm by Definition 5.7. It remains to show that x† ∈ N (A) : Assume
that x† ∈ N (A) and x† = 0. Then it follows that
A∗ Ax† = A∗ 0 = 0 = A∗ y, (5.96)
A∗ A 0 = 0 = A∗ y. (5.97)
Consequently, 0 ∈ L(y). However, 0 = 0 < x† which forms a contradiction to

⊥ ⊥
the minimal norm property of x† . Thus, x† ∈ N (A) since X = N (A) ⊕ N (A) .
⊥
“⇐” Suppose that x† ∈ N (A) and A∗ Ax† = A∗ y. Then, x† ∈ L(y). We
have to show that x† has minimal norm. Let x ∈ L(y) be arbitrary. It follows that
A(x† − x ) = Ax† − Ax = PR(A) y − PR(A) y = 0. (5.98)
Thus, x† − x ∈ N (A). Now we consider the norm of x :

x 2 = x† + (x − x† )2 = x† 2 + x − x† 2 + 2 x† , x − x† ≥ x† . (5.99)

=0
⊥
Note that the scalar product is 0 since x† ∈ N (A) , whereas x − x† ∈ N (A). As
a consequence, we have x† = A† y.
Theorem 5.12 (Properties of the Pseudoinverse). The generalized inverse A† pos-

sesses the following properties:

(1) D A† = Y if and only if R (A) is closed.
⊥
(2) R A† = N (A) .
†
(3) A is linear.
(4) A† is continuous if and only if R (A) is closed.

Proof. (1) D A† = Y is equivalent to R (A) ⊕ R (A)⊥ = Y , i.e., R (A) is closed.
⊥ ⊥
(2) We have R A† ⊂ N (A) due to Theorem 5.11. Choose x ∈ N (A)
and set y = Ax. Then we obtain
PR(A) y = Ax ⇒ x ∈ L (y) . (5.100)
⊥
Using Theorem 5.11 we are led to N (A) ⊂ R A† .

(3) Suppose that y, y ∈ D A† . Then we have
AA† y = PR(A) y, (5.101)

AA† y = PR(A) y . (5.102)
Thus it is clear that

A A† y + A† y = PR(A) (y + y ) = AA† (y + y ) (5.103)
and
†
A y + A† y − A† (y + y ) ∈ N (A) , (5.104)
† ⊥
A y + A† y − A† (y + y ) ∈ R A† = N (A) , (5.105)
so that
A† y + A† y − A† (y + y ) = 0. (5.106)
† †
†
Analogously we are able to show that A (αy) = αA y for all y ∈ D A and
α ∈ C.
⊥
(4) Assume that A† is continuous. Then, D A† = R (A)⊕R (A) is dense in
†
Y and A can be extended continuously to all of Y by B ∈ L (Y, X) (in accordance
with the continuous extension of operators, see, e.g., [110]). It follows that
ABy = PR(A) y (5.107)
for all y ∈ Y , i.e.,

R (A) = R PR(A) ⊂ R (A) . (5.108)
Therefore,
R (A) = R (A) . (5.109)
On the other hand, let R (A) be closed. Consider the operator Â given by
⊥
Â : N (A) → R (A) , x → Ax. (5.110)
Â is bijective and bounded. Due to the inverse mapping theorem of functional
analysis (see, e.g., [110]) Â−1 is also bounded, and we obtain
- † - - -
-A y - = -Â−1 ÂA† y -
X X
- - - -
≤ -Â−1 -Y →X -ÂA† y -Y
- - - -
= -Â−1 -Y →X -AA† y -Y (5.111)
†
for all y ∈ D A = Y . Furthermore,
- - - - - † - - −1 -−1 - † -
-y - ≥ -P - - - - - -A y -
Y R(A) y Y = AA y Y ≥ Â Y →X X
(5.112)
so that
A† yX - -
≤ -Â−1 - (5.113)
yY
for all y ∈ Y . Altogether, we have
- †- - -
-A - ≤ -Â−1 -Y →X . (5.114)
Y →X
Thus, A† is bounded.
An equivalent way to introduce the generalized inverse is the introduction

via the four Moore–Penrose conditions listed below.
Theorem 5.13 (Moore–Penrose Conditions). The generalized inverse is uniquely

determined by the four Moore–Penrose conditions:
(1) AA† A = A,
(2) A† A = PR(A† ) ,
(3) A† AA† = A† ,
(4) AA† = PR(A) ,
where the last property is restricted to D(A† ).
Proof. The proof is split into two parts:

(a) A† satisfies the conditions (1) to (4),
(b) if B ∈ L(Y, X) and C ∈ L(Y, X) satisfy the Moore–Penrose conditions (1) to
(4), then B = C.
⊥
We begin with (a): Clearly, R(A† ) = N (A) and R(A† ) = R(A† ). Let y ∈ D(A† ).
Then we are able to conclude
AA† y = Ax with x = A† y ⇒ A∗ Ax = A∗ y ⇒ Ax = PR(A) y ⇒ (4).

(5.115)
Set x = x1 + x2 = PN (A)(A)⊥ x + PN (A) x. Then, it follows that

∈N (A)⊥ ∈N (A)
A† Ax2 = A† 0 = 0 = PN (A)⊥ x2 (5.116)
and
z = A† Ax1 ⇒ Az = PR(A) Ax1 = Ax1 ⇒ A(z − x1 ) = 0 ⇒ z − x1 ∈ N (A) .

(5.117)
But it should be noted that
⊥ ⊥
z = A† Ax1 ⇒ z ∈ R(A† ) = N (A) ⇒ z − x1 ∈ N (A) . (5.118)
Thus, from (5.117) and (5.118), it can be deduced that z = x1 = PN (A)⊥ x. This
means that
A† Ax = A† Ax1 + A† Ax2 = PN (A)⊥ x = PR(A† ) x ⇒ (2). (5.119)

=0
Now, we easily arrive at the other two Moore–Penrose conditions for A† :
A† AA† = PR(A† ) A† = A† ⇒ (3), (5.120)

†
AA A = PR(A) A = A ⇒ (1). (5.121)
This concludes the part (a) of the proof.

We continue with part (b): Assume that B, C ∈ L(Y, X) satisfy (1) to (4).
Then we have
B = BAB
= BAC (note that AB = PR(A) = AC)
= CAC (note that BA = PR(B) = PN (A)⊥ = PR(C) = CA)
= C, (5.122)
⊥
where we still need to prove that R(B) = N (A) (which also shows that R(C) =
⊥
N (A) since B and C can be exchanged). Let x ∈ R(B). Assume that x ∈ N (A),
x = 0. Then it follows that
⊥
Ax = 0 ⇒ BAx = 0 ⇒ PR(B) x = 0 ⇒ x ∈ R(B) , (5.123)
⊥
which is a contradiction. Thus, x ∈ N (A) .
⊥
On the other hand, suppose that x ∈ N (A) . Moreover, assume that x ∈
⊥
R(B) . Then
0 = PR(B) x = BAx ⇒ Ax = ABAx = A0 = 0 ⇒ x ∈ N (A) , (5.124)
which again is a contradiction and, therefore, x ∈ R(B).
Summarizing our results we obtain
⊥
R(B) = N (A) , (5.125)
which yields B = C, i.e., the Moore–Penrose conditions uniquely determine the
pseudoinverse concludes the part (b) of the proof.
Compact Operators. Next we discuss a certain set of operators, viz. compact op-
erators, that turn out to be prototypes for generating a large class of ill-posed
problems.
Definition 5.8. Let X, Y be normed spaces. An operator A : X → Y is called
compact, if one of the following equivalent conditions is fulfilled:
(1) Every bounded subset U ⊂ X possesses an image in Y , which is relatively
compact, i.e., A(U ) is a compact set.
(2) For every bounded sequence {xn }n ⊂ X the sequence {Axn }n possesses a
subsequence that converges in Y .
By convention, we introduce the following settings:
K (X, Y ) = {A : X → Y : A is linear and compact} (5.126)
and
K (X) = K (X, X) . (5.127)
Lemma 5.2. Let X, Y be normed spaces. Then the following statements hold true:
(1) K (X, Y ) ⊂ L (X, Y ) .
(2) If A ∈ L(X, Y ) with dim R (A) < ∞, then A is compact.
Proof. (1) Assume that A belongs to K(X, Y ). Clearly, the closed unit ball B1 (0) ⊂
X is bounded. Therefore, A(B1 (0)) is relatively compact and A(B1 (0)) is bounded,
i.e., Ax ≤ C for all x ∈ B1 (0). This implies supx∈B1 (0) Ax ≤ C, so that A is
continuous.
(2) A : X → Y , dim R(A) < ∞. Therefore, each closed and bounded subset
of R(A) is compact. This means that each bounded subset of R(A) is relatively
compact. Now, let U ⊂ X be bounded. Then A(U ) is also bounded and A(U ) ⊂
R(A) is relatively compact. Thus, A is compact.
Theorem 5.14. Let X, Y, Z be normed spaces.
(1) If A : X → Y is compact and B : Y → Z is continuous or A : X → Y is
continuous and B : Y → Z is compact, then AB : X → Z is compact.
(2) The identity operator I : X → X is compact if and only if X is finite-
dimensional.
(3) Let Y be a Banach space. Then K (X, Y ) is closed, i.e., a sequence of compact
operators {An }n ⊂ K (X, Y ) with limn→∞ An − AX→Y = 0 has a compact
limit, i.e., the limit operator A is compact.
(4) If A is compact and invertible and X is not finite-dimensional, then A−1 is
not continuous.

Example. Let G ⊂ Rq be a regular region and let K be of class C (0) G × G . We

introduce the integral operator A : C (0) G → C (0) G by letting

(AF ) (x) = K (x, y) F (y) dy, F ∈ C (0) G . (5.128)
G
The proof of the compactness of A can be based on a well-known theorem of clas-
sical analysis, namely the Theorem of Arzelà–Ascoli. This theorem (cf. Theorem
5.15) provides
two equivalent properties to the relative compactness of a subset
U ⊂ C (0) G .
(Theorem of Arzelà–Ascoli). Let ∅ = G ⊂ R be regular. A subset

q
Theorem 5.15
(0)
U ⊂C G is relatively compact if and only if the following two statements are
valid:
(1) U is equicontinuous, i.e., for every ε > 0 there exists δ (ε) > 0, such that for
all F ∈ U
|F (x) − F (y)| < ε (5.129)
for all x, y ∈ G with x − y < δ (ε) .
(2) U is bounded, i.e., there exists an M > 0 with F C (0) (G ) < M for all F ∈ U.
In accordance with the theorem of Arzelà–Ascoli (Theorem 5.15) we are now

interested in applying the conditions (1) and (2) to the context of the integral
operator introduced in (5.128):
(1) K is uniformly continuous in G × G, i.e., for all ε > 0 exists a δ > 0 such that
for all x, y, z ∈ G with x − y < δ holds |K (x, z) − K (y, z)| < M·G
ε
. Thus
we get, for every bounded subset U ⊂ C (0) (G),

|AF (x) − AF (y)| = K (x, z) F (z) dz − K (y, z) F (z) dz
G G

≤M |K (x, z) − K (y, z)| dz
G
ε
< M G
M G
= ε, (5.130)
provided that x − y ≤ δ. Therefore, AU is equicontinuous.
(2) Suppose that U ⊂ C (0) (G) is bounded. Assume that F is of class U . Then we
have

|AF (x)| = K (x, y) F (y) dy ≤ M max |K (x, y)| G < ∞. (5.131)
G x,y∈G
In other words, AU is bounded.

As a consequence, by virtue of Theorem 5.15, we are able to conclude that AU is
relatively compact.
Theorem 5.16. The operator A defined by (5.128) is compact.

Example. Let G ⊂ Rq be a regular region. We introduce the operator A : L2 G →

L2 G given by

AF (x) = K (x, y) F (y) dy, F ∈ L2 (G). (5.132)
G
If K is continuous for x = y and weakly singular, i.e., there exist a value α ∈ (0, q)
and a constant C > 0 such that
1
|K (x, y)| ≤ C , (5.133)
|x − y|q−α
then A is compact (for the proof the reader is referred to standard textbooks about
integral equations, e.g., [30, 110, 120]). As a consequence, the Newton volume
integral occurring in the theory of Earth’s gravitation forms a compact operator
A.
Singular Value Decomposition. Next we are concerned with the introduction of
eigenvalues and eigenfunctions corresponding to an operator A ∈ L(X).
Definition 5.9. Let X be a normed space. Suppose that A is of class L (X).
(1) λ ∈ C is called a regular value of A if and only if λI −A is continuously invert-
ible. ρ (A) = {λ ∈ C : λI − A continuously invertible} is called the resolvent
set.
(2) σ (A) = C \ ρ (A) is called the spectrum of A.
(3) λ ∈ σ (A) is called an eigenvalue of A if N (λI − A) = {0}. The elements
of N (λI − A) \ {0} are called eigenvectors of A corresponding to the eigen-
value λ.
We are now in position to verify the following result.

Theorem 5.17. Let X be a Banach space.
(1) If λ ∈ σ (A), then |λ| ≤ A, i.e., the spectrum is bounded.
(2) σ (A) ⊂ C is compact.
Theorem 5.18. Let X be a normed space. Suppose that A is a compact operator
on X (i.e., A ∈ K (X)).
(1) If λ ∈ σ (A) \ {0} , then λ is an eigenvalue of A.
(2) If λ is an eigenvalue of A, then dim N (λI − A) < ∞.
(3) σ (A) is at most countable. Furthermore, 0 ∈ σ (A).
(4) 0 is the only accumulation point of σ (A).
The proofs are standard. They can be found in any textbook of functional analysis
(e.g., [110]).
Central in our considerations about compact operator is the following spec-
tral theorem that can be seen in parallel to the finite-dimensional case of matrix
operators.
Theorem 5.19 (Spectral Theorem for Compact Self-adjoint Operators). Let X be
a Hilbert space. Assume that A is of class K (X) and that A is self-adjoint, i.e.,
A∗ = A. Then there exists an orthonormal system {xi }i∈N ⊂ X and a sequence
{μi }i∈N ⊂ R (finite or countable) with |μ1 | ≥ |μ2 | ≥ · · · > 0, such that
∞

Ax = μi x, xi xi (5.134)
i=1
holds true for all x ∈ X.
Proof. Once again, the proof is standard.
Suppose now that X and Y are Hilbert spaces. Furthermore, let A be of class
K(X, Y ). Then A∗ A is also compact and obviously self-adjoint. Due to the spectral
theorem (Theorem 5.19) there exist a sequence {λi }i∈kN ⊂ R and an orthonormal
system {xi }i∈N ⊂ X such that
∞

A∗ Ax = λi x, xi xi , x ∈ X. (5.135)
i=1
Suppose that λi ∈ σ (A∗ A)\ {0} and denote, as usual, by xi its corresponding
eigenvector. It follows that
2
λi xi = λi xi , xi = λi xi , xi X = A∗ Axi , xi X
2
= Axi , Axi Y = Axi Y . (5.136)
Therefore we are able to conclude that λi > 0.
In the sequel, we assume that the eigenvalues are listed in#the chronological
order as follows: λ1 ≥ λ2 ≥ · · · ≥ λi ≥ λi+1 ≥ · · · ≥ 0. Set σj = λj . Moreover, let

yi = σ1i Axi i.e., Axi = σi yi , i ∈ N, and A∗ yi = A∗ σ1i Axi = σ1i A∗ Axi = σ1i λi xi =
σi xi . Furthermore, we have yi , yk Y = σi1σk Axi , Axk Y = σi1σk A∗ Axi , xk X =
δi,k . Thus, {yi }i∈N ⊂ Y forms a complete orthonormal system (ONS) in R(A), so
⊥
that the system {xi }i∈N is a complete ONS in N (A) . Now, assume that x is a
⊥
member of N (A) . Then it follows that
∞
x= x, xi X xi (5.137)
i=1
and
∞
∞

Ax = x, xi X Axi = σi x, xi X yi (5.138)
i=1 i=1
for all x ∈ N (A)⊥ .
Definition 5.10. Let X, Y be Hilbert spaces. The set {σi ; xi , yi }i∈N ⊂ (0, ∞)×X×Y
is called the singular system of an operator A ∈ K (X, Y ). The values σi are
called the singular values of A. The elements xi , yi are called the singular vectors.
Furthermore, the series
∞
Ax = σi x, xi X yi , x ∈ X (5.139)
i=1
is called the singular value decomposition (SVD) of A.
Picard Condition. The following condition plays an essential role in the solvability
of inverse problems.
Theorem 5.20 (Picard Condition). If A : X → Y is compact with singular value
decomposition (SVD) {σi ; xi , yi }i∈N , then y ∈ R (A) is an element of R (A) if and
only if
∞
|y, yi |2
(5.140)
i=1
σi2
is convergent.
Proof. “⇒” Suppose that y belongs to R (A). Then there exists a member x ∈ X
with Ax = y such that
y, yi Y = Ax, yi Y = x, A∗ yi X (5.141)
and
∞
∞

|y, yi |2 Bessel
= |x, xi |2 ≤ x2X < ∞. (5.142)
i=1
σi2 i=1

∞
|y,yi |2
“⇐” Suppose that y belongs to R (A). Assume that the series σi2
i=1
converges. We let
∞
y, yi
x= xi . (5.143)
i=1
σi
Then it follows that

∞

2 |y, yi |2
xX = <∞ (5.144)
i=1
σi2
and
∞
∞

y, yi y∈R(A)
Ax = Axi = y, yi yi = PR(A) y = y. (5.145)
i=1
σi i=1
Therefore, y is an element of R (A).
Remark 5.7. Obviously, from the Picard condition, it follows that →0 σi−2 |y, yi |2
if i → ∞ so that information about the decay of the Fourier coefficients of an
element y becomes available.
Theorem 5.21 (Generalized or Pseudoinverse of a Compact Operator). Let A be
a compact operator (i.e., A ∈ K (X, Y )) with SVD {σi ; xi , yi }i∈N . Then we have
∞
1
A† y = y, yi Y xi (5.146)
σ
i=1 i

for all y ∈ D A† .

Proof. Suppose that y ∈ D A† = R (A) ⊕ R (A)⊥ such that
y = Ax + ỹ. (5.147)
By definition, we set
∞
1
x̃ = y, xi Y xi . (5.148)
σ
i=1 i
Then it follows that
∞ ∞ ∞
1 1
x̃ = Ax, yi Y xi = x, A∗ yi X xi = x, xi xi . (5.149)
σ
i=1 i
σ
i=1 i i=1
⊥ ⊥
In other words, x̃ ∈ X. Moreover, x̃ ∈ N (A) , since {xi } ⊂ N (A) . It is not
hard to see that
∞ ∞
1 1
A∗ Ax̃ = y, yi A∗ Axi = y, yi σi2 xi
i=1
σi i=1
σi
∞
∞

= σi y, yi xi = y, yi A∗ yi = A∗ y. (5.150)
i=1 i=1
Consequently, x̃ satisfies the normal equations, and x̃ ∈ N (A)⊥ . Thus, x̃ is equal

to A† y.
Corollary 5.1. If R (A)
is finite dimensional (i.e., there exist only finitely many
element yi ), then R A† < ∞. Therefore, A† is compact. In particular, A† is
continuous.
∞ 1
Proof. If there exist only finitely many elements yi , then y, yi Y xi is a
i=1 σi
finite sum, hence, R A† is finite dimensional.
The representation of the generalized inverse in terms of the singular value
decomposition (5.146) opens the perspective to classify ill-posed problems. Indeed,
the summands σi−1 y, yi Y xi occurring in the series (5.146) depend closely on the
singular values. If the values σi are small, then the contribution by the series
(5.146) becomes large. The existence of SVD can be guaranteed for all compact
operators. Nevertheless, the concrete knowledge of SVD is critical. Only in rare
exceptions, SVD is explicitly known.
Remark 5.8. Singular systems are theoretically nice and easy, but their calculation
might be rather tricky.
The specific amount of the growth of the singular values finally leads us to
classify ill-posedness.
Definition 5.11 (Classification of Ill-posed Problems for Compact Operators). Let
A be a compact operator (i.e., A ∈ K(X, Y )) with SVD {σi ; xi , yi }i∈N .
(1) If there exists α > 0 such that
σi = O(i−α ), (5.151)
then the operator A is called ill posed of order α.
(2) If there exists ρ > 0 such that
| ln σi | ≥ c iρ , (5.152)
then the operator A is called exponentially ill posed.
Remark 5.9. Note that the aforementioned classification is senseful only for linear
problems reflecting the representation of A† y by its superposition in terms (5.146).
Truncated Singular Value Regularization. Obviously, the first two criteria (H1)
and (H2) determining a well-posed problem in the sense of Hadamard, can always
be enforced by considering the generalized inverse A† . A violating of the third
point, i.e., instability arises if the spectrum of the operator A is not bounded away
from zero. Thus, it seems to be natural to construct regularizing approximations
via modifying the smallest singular values.
Indeed, in accordance with the singular value decomposition of the gener-
alized inverse, it follows that such a modification of small values and, hence, a
construction of regularization operators can be obtained in the form
∞

xα = Rα y = σi Fα (σi2 )y, yi Y xi y ∈ Y, (5.153)
i=1
with some function Fα : R+ → R+ such that

1
Fα (λ) → , λ > 0, α → 0. (5.154)
λ
Such an operator Rα as defined by (5.153) and (5.154) may be understood as a

regularization operator if
λ|Fα (λ)| ≤ CFα < ∞, λ > 0. (5.155)
If (5.155) is satisfied, then we are able to see that
∞ ∞

2 2 2 2 2
Rα yX = σi (Fα (σi )) |y, yi Y | ≤ CFα |y, yi Y |2 ≤ CF2 α y2Y , (5.156)
i=1 i=1
where CF is a bound for the norm of Rα . Note that the pointwise convergence of
Fα immediately implies the pointwise convergence of Rα to A† .
Let A be a compact operator (i.e., A ∈ K (X, Y )) with SVD {σi ; xi , yi }i∈N .
The main idea of truncated singular value decomposition is to ignore all singular
values below a certain threshold value, which we can identify with the regulariza-
tion parameter α, hence, the representation of the regularized solution is given by
1
, λ ≥ α,
Fα (λ) = λ (5.157)
0 , λ < α,
i.e.,
1
xα = Rα y = y, yi Y xi , y ∈ Y, (5.158)
σi
σi ≥α
which explains the name truncated singular value decomposition, since all terms
in the sum corresponding to small singular values are truncated. Since 0 is the only
accumulation point of the singular values of a compact operator, the sum in (5.158)
is always finite for α > 0. In particular, only a finite number of singular values
and singular vectors has to be computed in order to realize this method. On the
other hand it should be mentioned that, for α being small, the number of singular
values that need to be computed can increase strongly. Obviously, CFα = α−1 .
Tikhonov Regularization. The regularized solution is given by
1
Fα (λ) = , λ > 0, α > 0, (5.159)
λ+α
so that ∞
σi
xα = Rα y =
σ 2 + α y, yi Y xi , y ∈ Y. (5.160)
i=1 i
As in the case of Lavrentiev regularization, we can compute xα defined by (5.160)
without any knowledge of the singular system. In fact, it is easy to see that
(A∗ A + αI) xα = A∗ y (5.161)
and, hence, we can solve a well-posed linear system to obtain xα . From this rep-
resentation it also follows that Tikhonov regularization is just Lavrentiev regular-
√
ization applied to the normal equations. It is not hard to see that λ2 + α ≥ 2λ α,
hence, CFα can be chosen as 2α−1/2 .
We are interested in an estimate of the approximation error, which is inde-
pendent of the noise level ε:
Theorem 5.22. Let Fα : R+ → R+ be a piecewise continuous function satisfying

the assumptions
1
Fα (λ) → , λ > 0, α → 0, (5.162)
λ
|Fα (λ)| ≤ Mα < ∞, λ > 0, (5.163)
and
sup(λFα (λ)) ≤ CF < ∞ (5.164)
α,λ
for some constant CF > 0. Moreover, let the regularization operator be defined by
(5.153). Then for all y ∈ D(A† ),
Rα y → A† y, α → 0. (5.165)
Proof. From the singular value decomposition (SVD) we obtain
∞
† 2 1
Rα y − A y = σi Fα (σi ) − y, yi Y xi
i=1
σi
∞

= (σi2 Fα (σi2 ) − 1)x† , xi X xi . (5.166)
i=1
Under the assumptions imposed on Fα we easily see that

|(σi2 Fα (σi2 ) − 1)x† , xi ≤ (CF + 1)x† . (5.167)
Hence, we are able to deduce that
∞

lim sup Rα y − A† y2 ≤ lim sup (σi2 Fα (σi2 ) − 1)2 x† , xi 2X
α→0 α→0
i=1
∞
2
≤ lim (σi2 Fα (σi2 )) − 1 x† , xi 2X . (5.168)
α→0
i=1
=0
From the pointwise convergence limα→0 (λFα (λ)) − 1 = 0 we obtain the de-
sired result. y
Remark 5.10. The function t → Fα (t), t ∈ R+ , converges pointwise to the function

0 , t > 0,
F (t) = (5.169)
1 , t = 0.
Due to the discontinuity at zero, the convergence of tFα (t) − 1 to zero is becoming
slower and slower as t decreases to zero. Since it is allowed to specify an arbitrarily
small singular value σi and the minimal norm solution x† = xi , the convergence of
regularized solutions is arbitrarily slow. On the other hand, we observe from the
proof that there is a possibly faster convergence if the components x† , xi X decay
sufficiently fast compared to the eigenvalues. For example, if we have |x† , xi X | ≤

cσiμ for some constant c > 0 and μ > 0, then it follows
∞

lim sup Rα y − A† y2 ≤ lim sup c2 (σi Fα (σi ) − 1)2 σi2μ
α→0 α→0
n=1
∞

≤ c2 lim(σi1+μ Fα (σi ) − σiμ )2 . (5.170)
α
n=1
In other words, one has to consider the limit of the function t → |t1+μ Fα (t)−tμ | as
t → ∞ instead, which is usually much faster. For example, in case of the truncated
singular value decomposition, we obtain

0, t ≥ α,
|t1+μ Fα (t) − tμ | = (5.171)
tμ , t < α.
If the singular values of the operatordecay sufficiently fast (which is the typi-
∞
cal case for ill-posed problems), e.g., n=1 σiμ < ∞, we are confronted with the
situation
2μ ∞
Rα y − A† y2 ≤ c2 σi ≤ c2 αμ σiμ (5.172)
σi <α i=1
so that Rα y − A†y μ/2

is of order α . Consequently, we somehow need smoothness
of the solution (in terms of the smoothing properties of the operator) in order to
obtain a convergence rate in terms of α. We shall pursue this idea by introducing
spaces of smoothness involving the absolute value of a compact operator.
Next we are concerned with the propagation of the data error through the
regularization.
Theorem 5.23. Let Fα and CF be as in Theorem 5.22, and let xα = Rα y, xεα =

Rα y ε . Then the estimates
Axα − Axεα ≤ CF ε, (5.173)
and
xα − xα
α ≤ Mα ε (5.174)
are valid.
Proof. From the singular value decomposition it follows directly

∞
2 2
Axα − Axεα 2Y ≤ σi Fα (σi2 ) |y − y δ , yi Y |2
i=1
∞

≤ CF2 |y − y ε , yi Y |2 = CF2 y − y ε 2 ≤ (CF ε)2 , (5.175)
n=1
so that (5.173) is valid. In the same way we obtain

∞

xα − xεα 2 ≤ (Fα (σi ))2 |y − y ε , yi Y |2
i=1
∞

≤ Mα2 |y − y ε , yi Y |2 = Mα2 y − y ε 2 ≤ (Mα ε)2 , (5.176)
i=1
so that (5.174) is implied (note that (5.174) estimates the norm of Rα by CF ).

As a consequence, the error can be split in the following form
xεα − xX = Rα y ε − xX
≤ Rα y ε − Rα yX + Rα y − xX
≤ Rα Y →X y ε − yY + xα − xX , (5.177)
such that
xα,ε − xX ≤ Rα Y →X ε + xα − xX . (5.178)
We see that the error between the exact and the approximate solution consists
of two parts: The first term is the product of the bound for the error in the data
and the norm of the regularization parameter Rα . This term will usually tend to
infinity for α → 0 if the inverse A−1 is unbounded and A is compact. The second
term denotes the approximation error (Rα − A−1 )yX for the exact right-hand
side y = Ax. This error tends to zero as α → 0 by the definition of a regularization
strategy. Thus, both parts of the error show a diametrically reflected behavior. A
typical picture of the errors in dependence on the regularization parameter α is
sketched in Figure 5.5. Thus, a strategy is needed to choose α dependent an ε in
error
total error
Rα Ax − xX
ε Rα Y →X
α
Figure 5.5. Typical behavior of the total error in a regularization process.
order to keep the error as small as possible, i.e., we would like to minimize
Rα Y →X ε + Rα Ax − xX . (5.179)
In principle, we distinguish two classes of parameter choice rules: If α = α(ε)
does not depend on ε, we call α = α(ε) an a priori parameter choice rule. Otherwise
α depends also on y ε and we call α = α(ε, y ε ) an a posteriori parameter choice
rule. It is conventional to say that a parameter choice rule is convergent, if for
ε → 0 the rule is such that
lim sup{Rα(ε,yε ) y ε − A† yX : y ε ∈ Y, y ε − yY ≤ ε} = 0 (5.180)
ε→0
and
lim sup{α(ε, y ε ) : y ε ∈ Y, y − y ε Y ≤ ε} = 0. (5.181)
ε→0
All in all, numerous methods have been proposed for treating and regularizing
various types of ill-posed problems. The rationale in most methods for resolution
(approximate solvability) of ill-posed problems is to construct a “solution” that is
acceptable physically as a meaning field approximation and is sufficiently stable
from computational standpoint. The main dilemma of modeling ill-posed problems
is that the closer the mathematical model describes the ill-posed problem the worse
is the “condition number” of the associated computation problem (i.e., the more
sensitive to errors, see [165]). A way out can only be found by additional “exterior”
information about the problem to be solved.
5.6. Multi-scale solutions of inverse pseudodifferential equations
All gravitational information under discussion in physical geodesy leads to operator
equations relating the disturbing potential to geodetically relevant observables.
The most important operators are listed in the so-called “Pocket Guide of Physical
Geodesy” (see, e.g., [170, 186]).
Pocket Guide of Physical Geodesy. In physical geodesy, one can think of observ-
ables as operating on an “input signal” F (e.g., the disturbing potential) to produce
an “output signal” of the form
ΛF = G (5.182)
(for example, geoidal undulation, gravity anomaly, radial derivative), where Λ is
a certain operator (note that we use capital letters F, G, . . . in this subsection to
characterize geodetic quantities). Fortunately, it is the case in geodetic applica-
tions involving the disturbing potential that large portions of interest can be well
approximated by operators that represent linear, rotation-invariant pseudodiffer-
ential operators.
The standard pseudodifferential operators Λ occurring in physical geodesy
(cf. [208]) have to reflect the aforementioned Pizzetti concept. As an immediate
consequence, for the operator equation relating a geodetic observable G = ΛF to
the disturbing potential F (see, e.g., [107, 150]), we are led to an operator equation
ΛF = G which links F and the input function G under the following constraint:
ΛYn,k = Λ∧ (n) Yn,k , n = 0, 1, 2, . . . , k = 1, . . . , 2n + 1, (5.183)
such that
Λ∧ (0) = Λ∧ (1) = 0 (5.184)
and
Λ∧ (n) = 0, n ≥ 2, (5.185)
∧
where {Λ (n)}n=2,3,... is a sequence of real values (note that {Yn,k } is assumed
to constitute a (real) complete system of spherical harmonics in L2 (Ω)). Conse-
quently, we have to discuss the invertibility of the operator Λ on the space L22,...(Ω)
defined by
1
5
L22,... (Ω) = L2 (Ω) \ Harm n = L2 (Ω) \ Harm 0,1 . (5.186)
n=0
where Harm n is the linear space of all spherical harmonics of degree n.

In other words, we have the requirement that the spherical harmonics of
degrees n ≥ 2 are the eigenfunctions of the operator Λ, and the invertibility has
to be controlled by the invertibility of the values Λ∧ (n), n ≥ 2.
Definition 5.12 (Definition of Pseudodifferential Operators). Let {Λ∧ (n)}n=2,3,...
be a sequence of real numbers Λ∧ (n) satisfying
|Λ∧ (n)|
lim = const . = 0 (5.187)
n→∞ (n + 1 )t
2
for some t ∈ R. Then the operator Λ defined by
∞ 2n+1

Λ(F ) = Λ∧ (n) F (η))Yn,j (η) dS(η) Yn,j (5.188)
n=0 j=1 Ω
=F ∧ (n,j)
is called (invariant) pseudodifferential operator of order t. {Λ∧ (n)}n=2,3,... is called

spherical symbol of Λ. Moreover, if
|Λ∧ (n)|
lim =0 (5.189)
n→∞ (n + 1 )t
2
for all t ∈ R, then the operator Λ is called pseudodifferential operator of order −∞.
The spherical symbol has many appealing properties: It is easily seen that
(Λ + Λ )∧ (n) = (Λ )∧ (n) + (Λ )∧ (n), (5.190)
∧ ∧ ∧
(Λ Λ ) (n) = (Λ ) (n)(Λ ) (n) (5.191)
for all n = 2, 3, . . ..
As any “output function” (output signal) can be expanded into an orthogonal
series of surface spherical harmonics
∞ 2n+1
∞ 2n+1

G = ΛF = Λ∧ (n)F ∧ (n, k)Yn,k = G∧ (n, k)Yn,k (5.192)
n=2 k=1 n=2 k=1
in the sense of · L2 (Ω) , we are confronted with a spectral representation of the
form
G∧ (n, k) = (ΛF )∧ (n, k) = Λ∧ (n) F ∧ (n, k), n = 2, 3, . . . , k = 1, . . . , 2n + 1.
(5.193)
This means that the “amplitude spectrum” {G∧ (n, k)} of the response of Λ is
described in terms of the amplitude spectrum of functions (signals) F by a simple
multiplication by the “transfer” Λ∧ (n). If a comparison of the “output function”
with the actual value is done, discrepancies would be observed. A mathematical
description of these discrepancies has to follow the laws of probability theory in a
stochastic model. According to this approach we again assume that we have
Gε = G + ε = ΛF + ε, (5.194)
where ε is the observation noise.
Table 2 shows the so-called Integrated Spherical Harmonic Model of Physical
Geodesy or “Meissl Scheme”, see [150, 170, 185] (earlier already called “Pocket
Guide of Physical Geodesy”). Herein, R designates the Earth’s mean radius, H is
the satellite height.
operator/quantity Λ term Λ∧ (n) order

n−1
gravity anomaly ΛA R 1
geoid undulations ΛU R2 0
Stokes operator ΛSt R
n−1 −1
single layer ΛS R
n+ 12
−1
double layer ΛD − 2n+1
R
−1
first radial derivative ΛF N D − n+1
R 1
(n+1)(n+2)
second radial derivative ΛSN D R2 2
n+1
upward continuation ΛUP C R
R+H −∞
n+1
(n+1)(n+2)
satellite gravity gradiometry ΛSGG R
R+H (R+H)2 −∞
Table 2. Geodetic Operators in Spherical Nomenclature.
In spherically reflected satellite problems, the orbits are quite attractive for
mathematical modeling: A circular orbit implies that the data are lying on a sphere;
the measurements offer a global data coverage and an extremely dense and uniform
distribution; the measurements (achieved by employing the significant principles
of, e.g., satellite gravity gradiometry (SGG)) provide global information about
the second radial derivatives of the gravitational potential at a moderate altitude.
Table 2 shows that the radial derivatives on spherical orbits are representable by
rotation-invariant pseudodifferential equations (for more details on pseudodifferen-
tial operators on the sphere see, e.g., [208]; modeling concepts in modern satellite
problems are described in [45]).
Multi-scale Regularizations of Inverse Geodetic Pseudodifferential Equations.
Next we are interested in discussing a wavelet sampling solution of pseudodifferen-
tial equations. Since well-posed problems can be solved in a more or less straight-
forward way, we restrict ourselves to the solution of ill-posed pseudodifferential
equations. We shall see that a sufficiently fast decay of the Legendre transform
of the scaling function leads to regularization strategies for ill-posed problems in-
volving pseudodifferential operators of finite order. For the exponentially ill-posed
problems it is particularly adequate to use bandlimited scaling functions.
Given G ∈ L22,... (Ω), find F ∈ L22,... (Ω) so that
ΛF = G, (5.195)
where Λ : L22,...(Ω) → L22,...(Ω)
is an isotropic pseudodifferential operator of order
s < 0 or s = −∞.
We assume in the following that Λ∧ (n) = 0 for all n = 2, 3, . . . , which makes
Λ injective. So, Λ is a linear bounded injective compact operator so that problem
(5.195) is ill-posed. In the nomenclature of the theory of ill-posed problems we are
able to say that Λ possesses the singular system (σn,k ; Yn,k , Yn,k )n=2,3,...,k=1,...,2n+1
with σn,k = Λ∧ (n).
We start our considerations with the regularization of problem (5.195).
Roughly speaking we call a regularization a family of bounded linear opera-
tors Rα : L22,... (Ω) → L22,...(Ω) which approximates the inverse Λ−1 .
In more detail,
Definition 5.13 (Regularization Strategy). A regularization strategy for the prob-
lem (5.195) is a family of linear bounded pseudodifferential operators
Rα : L22,... (Ω) → L22,... (Ω), α > 0, (5.196)
∧
with symbol {(Rα ) (n)}n=2,3,..., so that
lim Rα ΛF = F (5.197)
n→∞
for all F ∈ L22,...(Ω), i.e., the operators Rα Λ converge in pointwise sense to the
identity operator in L22,... (Ω).
The following result immediately follows from the theory of inverse problems.
Theorem 5.24 (Tikhonov Regularization Strategy). Suppose that the pseudodif-
ferential operator Λ of type (5.195) is of order s < 0 or −∞. Assume that the
(non-bandlimited) Tikhonov kernel Φj is given by
∞
2n + 1
Φj (ξ · η) = (Φj )∧ (n)Pn (ξ · η), ξ, η ∈ Ω (5.198)
n=0
4π
with
(Λ∧ (n))2
(Φj )∧ (n) = , n ∈ N, j ∈ N0 , (5.199)
(Λ∧ (n))2 + γj2
where {γj }, j ∈ N0 is a sequence of real numbers satisfying limj→∞ γj = 0. Then
the operators
Rj = Φj ∗ Λ−1 (5.200)
constitute a regularization strategy in L22,... (Ω).
Since the boundedness of regularization operators is an important property,
we are led to the following characterization within the framework of L22,... (Ω).
Lemma 5.3. The pseudodifferential operator Λ : L22,...(Ω) → L22,... (Ω) with symbol
{(Λ)∧ (n)}n=2,3,... is bounded, if
∞
2n + 1 ∧
|Λ (n)|2 < ∞. (5.201)
n=2
4π
Proof. Let F be of class L22,...(Ω), i.e.,

∞ 2n+1

F 2L2 = |F ∧ (n, m)|2 < ∞. (5.202)
2,... (Ω)
n=2 m=1
Now,
(ΛF )∧ (n, m) = Λ∧ (n)F ∧ (n, m). (5.203)
Thus, for N ≥ 2, we are able to deduce from the Cauchy–Schwarz inequality that
N 2n+1
N 2n+1 N 2n+1

|Λ∧ (n)F ∧ (n, m)|2 ≤ |Λ∧ (n)|2 |F ∧ (n, m)|2
n=2 m=1 n=2 m=1 n=2 m=1

N
≤ (2n + 1)|Λ∧ (n)|2 F 2L2 . (5.204)
2,... (Ω)
n=2
Taking the limit N → ∞ we get the desired result.
We are interested in regularizations by use of isotropic filtering, i.e., the filter

Qα is chosen as a pseudodifferential operator with symbol {(Qα )∧ (n)}n=2,3,... (see
Figure 5.6 for an example of SGG). The regularization operator Rα = Qα Λ−1 has
to be bounded, i.e., in view of Lemma 5.3 we arrive at the condition
∞ 2
2n + 1 (Qα )∧ (n)
< ∞. (5.205)
n=2
4π Λ∧ (n)
Theorem 5.25. Let Λ : L22,...(Ω) → L22,...(Ω) be a pseudodifferential operator of

order s < 0 or of order −∞. Assume that the family of pseudodifferential operators
Qα with symbol {(Qα )∧ (n)}n=2,3,... , α > 0 has the following properties:
15
10
SGG−symbol
Scale 0
10
Scale15
10 Scale 30
Scale 45
Scale 60
5
10
0
10
−5
10
−10
10
−15
10
0 100 200 300 400 500 600 700 800
Figure 5.6. Symbol of the SGG-operator ΛSGG with R = 6378.127

[km], H = R+200 [km] (black) and the operators Rj in case of Tikhonov
regularization for different dyadic scales γj = 2−j .
(i) |(Qα )∧ (n)| ≤ 1 for all α > 0 and n = 2, 3, . . .

(ii) For every α > 0,
∞ 2
2n + 1 (Qα )∧ (n)
< ∞. (5.206)
4π Λ∧ (n)
n=2
(iii) For every n = 2, 3, . . .,
lim (Qα )∧ (n) = 1. (5.207)
α→0
Then the operator Rα = Qα Λ−1 is a regularization strategy for Λ.

Proof. From the estimate (5.206) and Lemma 5.3 we are able to conclude that the
operators Rα : L22,... (Ω) → L22,... (Ω) are bounded. For F ∈ L22,... (Ω), we have
Rα ΛF = Qα Λ−1 ΛF = Qα F. (5.208)
By virtue of the theory of singular integrals it thus follows that
lim Rα ΛF = F. (5.209)
α→0

The proof of Theorem 5.25 demonstrates the close relationship between the con-
cepts of regularization and singular integrals.
In order to deal with pseudodifferential operators Λ of order −∞ we are concerned
with the fully discrete wavelet transform as presented in [60], where the following
properties imposed on functions ϕ0 : [0, ∞) → R, ψ0 : [0, ∞) → R, respectively,

are given to establish spherical scaling – and wavelet functions:
(i) ϕ0 (0) = 1,
(ii) ϕ0 is monotonously decreasing,
(iii) ϕ0 is continuous at 0.
(iv) ϕ0 : [0, ∞) → R has a local support, i.e., supp ϕ0 ⊂ [0, 1].
(v) The generator ψ0 : [0, ∞) → R of the mother wavelet given by the so-called
refinement equation
2
2 t 2
(ψ0 (t)) = ϕ0 − (ϕ0 (t)) , t ∈ [0, ∞), (5.210)
2
also possesses a local support, i.e.,
supp ψ0 ⊂ [0, 1]. (5.211)
So we are canonically led to the scale-discrete (zonal) scaling and wavelet
functions, respectively,
∞
2n + 1
Φj (ξ · η) = (Φj )∧ (n) Pn (ξ · η), ξ, η ∈ Ω, (5.212)
n=0
4π
=ϕ0 (2−j n)
and
∞
2n + 1
Ψj (ξ · η) = (Ψj )∧ (n) Pn (ξ · η), ξ, η ∈ Ω, (5.213)
n=0
4π
=ψ0 (2−j n)
where
0 ≤ (Φj )∧ (n) = ϕ0 (2−j n) ≤ 1, n ∈ N0 , (5.214)
and
0 ≤ (Ψj )∧ (n) = ψ0 (2−j n) ≤ 1, n ∈ N0 (5.215)
with
lim (Φj )∧ (n) = lim ϕ0 (2−j n) = 1, n ∈ N0 , (5.216)
j→∞ j→∞
and
lim (Ψj )∧ (n) = lim ψ0 (2−j n) = 0, n ∈ N0 , (5.217)
j→∞ j→∞
such that the “approximate identity”

Φj ∗ F = Φj ( ·η)F (η) dS(η) → F, j → ∞, (5.218)
Ω
holds for F being of class L22,...(Ω) (in the topologies ·L2 (Ω) and ·C (0) (Ω) ).
Note that the compact support of ϕ0 implies that only finitely many
(Φj )∧ (n) are different from 0. This analogously holds true for ψ0 .
Moreover, for the functions ϕj and ψj , defined by ϕj = ϕ0 (2−j ·) and

ψj = ψ0 (2−j ·), respectively, we have
supp ϕj ⊂ [0, 2j ], (5.219)

supp ψj ⊂ [0, 2 ]. j
(5.220)
Since there are only a few conditions for a function ϕ0 : [0, ∞), there are
various possibilities for its bandlimited as well as non-bandlimited realizations (cf.
[60]). In our geodetically oriented framework we restrict ourselves to bandlimited
cases.
Example. The generator of the Shannon scaling function is given by

1, for t ∈ [0, 1),
ϕ0 (t) = (5.221)
0, for t ∈ [1, ∞),
so that
1, for t ∈ [0, 2j ),
ϕj (t) = (5.222)
0, for t ∈ [2j , ∞).
It is easy to see that all conditions for ϕ0 to be a generator of a scaling function

are fulfilled. We have
j
2 −1
2n + 1
Φj (t) = Pn (t) , t ∈ [−1, +1]. (5.223)
n=0
4π
A remarkable property is that Φj coincides with its iterations, i.e.,

(k) (k−1) (k−1)
Φj (ξ · η) = Φj ∗ Φj (ξ · η) = Φj (ξ · ζ)Φj (ζ · η) dS(ζ), ξ, η ∈ Ω , (5.224)
Ω
(2)
k = 2, 3, . . . , in particular, Φj (ξ · η) = Φj (ξ · η), ξ, η ∈ Ω. The construction of
wavelets is straightforward (cf. Figure 5.7).

1, for x ∈ [2j , 2j+1 ),
ψj (t) = (5.225)
0, elsewhere.
Hence,
2j+1
−1 2n + 1
Ψj (t) = Pn (t) , t ∈ [−1, +1]. (5.226)
4π
n=2j
Example. We consider a somehow “smoothed” version of the generator of the Shan-

non wavelets, called de la Vallée Poussin generator, (dependent on a parameter
20
15
10
−5
−3 −2 −1 0 1 2 3
Figure 5.7. Shannon wavelets Ψj (cos ϑ), ϑ ∈ [−π, π], j = 0, . . . , 3

(sectional illustration).
h ∈ (0, 1)) ⎧
⎨ 1 , for t ∈ [0, h),
1−t
ϕ0 (t) = 1−h , for t ∈ [h, 1), (5.227)
⎩
0 , for t ∈ [1, ∞).
With the definition (5.227) the “dilates” have the form
⎧
⎨ 1 , for t ∈ [0, 2j h),
−j
ϕj (t) = 1−2 t
, for t ∈ [2j h, 2j ), (5.228)
⎩ 1−h
0 , for t ∈ [2j , ∞),
j ∈ N0 . For the formulation of the wavelets corresponding to the “de la Vallée
Poussin generator” we distinguish three cases:
• 0 < h < 12
⎧
⎪
⎪ 0, for n < 2j h,
⎪
⎪ 2 1/2
⎪
⎪ 1−2−j n
⎪
⎪ 1 − , for 2j h ≤ n < 2j+1 h,
⎪
⎨ 1−h
Ψ∧ 2 2 1/2
j (n) = 1−2−j−1 n 1−2−j n
⎪
⎪ − , for 2j+1 h ≤ n < 2j ,
⎪
⎪ 1−h 1−h
⎪
⎪ 1−2−j−1 n
⎪
⎪ , for 2j ≤ n < 2j+1 ,
⎪
⎩ 1−h
0, for 2j+1 ≤ n < ∞.
(5.229)
1
• 1>h> 2
⎧
⎪
⎪ 0, for 0 ≤ n < 2j h,
⎪
⎪ 2 1/2
⎪
⎪
⎪
⎨ 1 − 1−2
−j
n
, for 2j h ≤ n < 2j ,
1−h
Ψ∧
j (n) = (5.230)
⎪
⎪ 1, for 2j ≤ n < 2j+1 h,
⎪
⎪ 1−2−j−1 n
⎪
⎪ , for 2j+1 h ≤ n < 2j+1 ,
⎪
⎩ 1−h
0, for 2j+1 ≤ n < ∞.
1
• h= 2
⎧
⎪
⎪ 0, for 0 ≤ n < 2j−1 ,
⎨
⎪ 2
1/2
Ψ∧ 1 − 2 − 2j+1 n , for 2j−1 ≤ n < 2j ,
j (n) = (5.231)
⎪
⎪ − −j
2j ≤ n < 2j+1 ,
⎪
⎩
2 2 n, for
0, for 2j+1 ≤ n < ∞.
Compared with the Shannon wavelets there generally are more non-vanishing
Legendre coefficients of Ψj . This explains the suppressing frequency effect. An
illustration is given by Figure 5.8.
12
10
−2
−4
−3 −2 −1 0 1 2 3
Figure 5.8. De la Vallée Poussin wavelets Ψj (cos ϑ), ϑ ∈ [−π, π], j =

0, . . . , 3, h = 0.5 (sectional illustration).
Theorem 5.26. Let Λ : L22,...(Ω) → L22,...(Ω) be a pseudodifferential operator of

order s < 0 or of order −∞. Assume that the bandlimited kernels Φj are defined via
a generator ϕ0 satisfying the properties (i)–(iv) as stated above. Then the operators
Rj = Φj ∗ Λ−1 (5.232)
constitute a regularization strategy in the following sense: Rj : L22,... (Ω) → L22,...(Ω)
is bounded, and the limit relation
lim Rj ΛF = F (5.233)
j→∞
holds true for all F ∈ L22,... (Ω).

(2)
Remark 5.11. Analogously, {Φj ∗ Λ−1 }j∈N0 constitutes a regularization strategy,
provided that {Bj }j∈N0 defines a (scale-discrete) scaling function.
5.7. Multi-scale signal-to-noise ratio and tree sampling

Usually, observations in geosciences are looked upon as a function G on the sphere
Ω so that
G = G + ε, (5.234)
where ε is the observation noise. We suppose the covariance to be known
E [ε̃(ξ), ε̃(η)] = K(ξ · η), (ξ, η) ∈ Ω × Ω, (5.235)
where the following conditions (cf. [54]) are imposed on the symbol {K ∧ (n)}n=0,1,
of the kernel K : Ω × Ω → R:
(C1) K ∧ (n, k) ≥ 0 for all n = 0, 1, . . . , k = 1, . . . , 2n + 1,
∞ ∧ 2
(C2) n=0 2n+1 4π K (n) < ∞.
Condition (C2), indeed, implies the (2) -summability of the symbol, i.e.,
∞
2n + 1 ∧
(K (n))2 < ∞. (5.236)
n=0
4π
The error-affected J-scale approximation provided by a bandlimited/non-

bandlimited regularization strategy of aforementioned type is given by

J−1
ε
RJ (G )(ξ) = RJ0 (ξ · ζ)G (ζ) dS(ζ) +
ε
Sj (ξ · ζ)Gε (ζ) dS(ζ), J > J0 ,
Ω j=J0 Ω
(5.237)
where Sj designates the difference
Sj = Rj+1 − Rj . (5.238)
Evidently, the computation of the occurring integrals will require methods of nu-
merical cubature. We base the integration on approximate formulas associated to
Nj Nj
known weights wi ∈ R and knots ηi ∈Ω

NJ0
NJ0 NJ0 NJ0
RJ0 (ξ · ζ) Gε (ζ) dS(ζ) wi RJ0 (ξ · ηi ) Gε (ηi ), (5.239)
Ω i=1

Nj
N N N
Sj (ξ · ζ) Gε (ζ) dS(ζ) wi j Sj (ξ · ηi j ) Gε (ηi j ), j = J0 , . . . , J − 1
Ω i=1
(5.240)
(the symbol “” always means that the error is assumed to be negligible, even
better, in case of a bandlimited regularization strategy the integration error can
be guaranteed to be 0). Since the “true” coefficients of (5.239) and (5.240) are
the ones that should be included in a selective reconstruction of G from Gε , in
estimating the unknown function F it is natural to include only coefficients larger
than some specified threshold value. The threshold value is understood to be the
N
scale and space error covariance at ηi j with respect to the (scale discrete) wavelet
function {Sj }j∈N0

S N N
Cov j Nj (K) = K(ξ · ζ) Sj (ξ · ηi j ) Sj (ζ · ηi j ) dS(ξ) dS(ζ) (5.241)
j,ηi Ω Ω

Nj

Nj
N N
wpNj wrNj K(ηpNj · ηrNj ) Sj (ηpNj · ηi j ) Sj (ηrNj · ηi j ).
p=1 r=1
N
We compare the scale and space error covariance at ηi j with the scale and space
N
error variance of Gε at ηi j with respect to the (scale discrete) scaling function
{Φhj }j∈N0

Sj N N
ε
Var Nj (G ) = Gε (ξ) Gε (ζ) Sj (ξ · ηi j ) Sj (ζ · ηi j ) dS(ξ) dS(ζ) (5.242)
j,ηi Ω Ω

Nj

Nj
N N
wpNj wrNj Gε (ηpNj )Gε (ηrNj ) Sj (ηpNj · ηi j ) Sj (ηrNj · ηi j ).
p=1 r=1
Signal and noise scale “intersect” at the so-called scale and space resolution set
S
Zresj , j = J0 , . . . , J − 1. We distinguish the following cases for signal-to-noise ratio:
(i) Signal dominates noise

Sj Sj N
Var Nj (Gε ) ≥ Cov Nj (K), (j, ηi j ) ∈ Zres
Sj
, i ∈ {1, . . . , Nj }.
j,ηi j,ηi
(ii) Noise dominates signal

Sj Sj N
Var Nj (Gε ) < Cov Nj (K), (j, ηi j ) ∈ Zres
Sj
, i ∈ {1, . . . , Nj }.
j,ηi j,ηi
An estimator of the “true” solution F = Λ−1 G can be determined via the indicator
function I in the form
RJest (Gε )(ξ) (5.243)

NJ0
R R N N N
= I Var J0 NJ0 (Gε ) ≥ Cov J0 NJ0 (K) wi J0 RJ0 (ξ · ηi J0 ) Gε (ηi J0 )
J0 ,ηi J0 ,ηi
i=1

J−1 Nj
Sj Sj N N N
+ I Var Nj (G ) ≥ Cov Nj (K) wi j Sj (ξ · ηi j ) Gε (ηi j ) .
ε
j,ηi j,ηi
j=J0 i=1
In other words, the large coefficients are kept intact and the small coefficients are
set to zero. The thresholding estimator of “true” coefficients are thus character-
ized by

RJ0 (ξ · ζ) Gε (ζ) dS(ζ)
Ω

NJ0
δ hardRJ0
RJ0 NJ0 NJ0 NJ0
(Var NJ (Gε )) wi RJ0 (ξ · ηi ) Gε (ηi ), (5.244)
Cov NJ J0 ,ηi 0
i=1 J0 ,η 0
i
Sj (ξ · ζ) Gε (ζ) dS(ζ)
Ω

Nj
δ hardSj
Sj N N N
(Var Nj (Gε )) wi j Sj (ξ · ηi j ) Gε (ηi j ), (5.245)
Cov Nj j,ηi
i=1 j,η
i
j = J0 , . . . , J − 1, where the function δλhard is the hard thresholding function

hard 1, if |x| ≥ λ,
δλ (x) = (5.246)
0, otherwise .
The “keep or kill” hard thresholding operation is not the only reasonable way
of estimating the coefficients. Recognizing that each coefficient consists of both
a signal portion and a noise portion, it might be desirable to attempt to isolate
the signal contribution by removing the noisy part. This idea leads to the soft
thresholding function

soft max{0, 1 − |x|
λ
}, if x = 0,
δλ (x) = (5.247)
0, if x = 0,
which can also be used in the coefficients of (5.244) and (5.245). When soft thresh-
olding is applied to a set of empirical coefficients, only coefficients greater than
the threshold (in absolute value) are included, but their values are ‘shrunk’ to-
ward zero by an amount equal to the threshold λ. In other words, an estimator
RJest (Gε )(ξ) of the “true” solution F is first approximated by a thresholded version
of (5.244), which represents the trend (smooth) components of the data. Then the
coefficients at higher resolutions are thresholded, so that the noise is suppressed

but the fine-scale details are included in the calculation.
Tree Sampling. Let us again consider an ill-posed pseudodifferential equation of
the form Λ : L22,... (Ω) → L22,...(Ω), ΛF = G, with a given error-affected right-
hand side Gε instead of G ∈ L22,... (Ω). We assume that the operator Λ is of finite
order s < 0 or of order −∞. We have seen in Subsection 5.6, that a bandlimited
regularization strategy based on a scaling function (5.212) can be given by Rj =
(2)
Φj ∗ Λ−1 , j = 1, 2, . . .. For more the decorrelation of the structural content in
the solution F it is important to become more detailed insight into the regularized
solutions Fj = Rj G at many levels j. To this end, we present a tree algorithm
which allows an efficient estimation of Fj at different scales from Gε .
Once more, the assumptions on the generator ϕ0 of the bandlimited scale
discrete scaling function allow a refinement equation of the form
(2) (2)
Φj = Ξj ∗ Φj+1 , (5.248)
where
2n + 1 ϕ20 (2j n)
Ξj = Pn (5.249)
4π ϕ20 (2j+1 n)
n∈N0
ϕ20 (2j n) =0
(note that the monotonicity and the compactness of ϕ0 imply ϕ20 (2j n) = 0 for
n ∈ N0 provided that ϕ20 (2j+1 n) = 0 for that n, hence, the kernel (5.249) is well
defined as finite sum). This observation enables us to realize a tree algorithm for
the decomposition of a signal Gε ∈ L22,...(Ω): Starting from a (sufficiently large)
J ∈ N, such that the sampling formula

(2)
F ε (ξ) RJ Gε = ΦJ ∗ Λ−1 ∗ Gε (ξ)
(2mJ +1)2
(2)
= wiJ ΦJ ∗ Λ−1 (ηiJ · ξ), ξ ∈ Ω, (5.250)
i=1
with
wiJ = aJi Gε (ηiJ ), i = 1, . . . , (2mJ + 1)2 , (5.251)
is valid, we are able to conclude that the coefficient vectors
wj = (w1 , . . . , w(2mj −1)2 )T , j = 0, . . . , J − 1 (5.252)
ε
(being, of course, dependent on the bandlimited “replacement” of G ) can be
calculated in the following way:
(i) The vectors wj , j = 0, . . . , J − 1, with

(2)
wij = aji ΦJ ∗ Λ−1 ∗ Gε (ηij ), i = 1, . . . , (2mj + 1)2 , (5.253)
are subsequently obtainable by recursion from the values wiJ in (5.251).

(ii) For j = 0, . . . , J − 1, we have

2
(2mj +1)

(2) −1 (2)
Φj ∗ Λ ∗ G (ξ) =
ε
wij Φj ∗ Λ−1 (ηij · ξ), ξ ∈ Ω. (5.254)
i=1
Our approach is divided into two parts, viz. an initial sampling step concern-
ing the (sufficiently large) scale level J and the recursion step:
The initial sampling step: We just read in the given data wiJ = aJi G̃(ηiJ ), i =
1, . . . , (2mJ + 1)2 , to get the representation (5.250).
The recursion step: For j = 0, . . . , J − 1 it follows that

(2)
wij = aji Φj ∗ Λ−1 ∗ Gε (ηij )

(2)
= aji Ξj ∗ Φj+1 ∗ Λ−1 ∗ Gε (ηij )
(2mj+1 +1)2
(2)
= aji wij+1 Ξj (ηij · ηij+1 )((Φj+1 ∗ Λ−1 ) ∗ Gε ) (ηij+1 )
i=1
(2mj+1 +1)2

= aji wij+1 Ξj (ηij · ηij+1 ) wij+1 . (5.255)
i=1
In other words, the coefficients wiJ−1 can be calculated recursively starting from
the data wiJ for the initial level J, wiJ−2 can be deduced recursively from wiJ−1 ,
etc. Moreover, the coefficients are independent of the special choice of the kernel
(observe that (5.250) is equivalent to
(2mj +1)2
N
G ε∧
(n, k) = wij Yn,j (ηi j ) (5.256)
i=1
for n = 0, 1, . . . , k = 1, . . . , 2n + 1). This fact finally leads us to the formulas
(2mJ +1)2
(2)
(2)
−1
(ΦJ ∗Λ ) ∗ G (ξ) =
ε
wiJ (Φj ∗ Λ−1 )(ηiJ · ξ), ξ ∈ Ω, (5.257)
i=1
(2mJ +1)2

(ΦJ ∗ Λ−1 ) ∗ Gε (ξ) = wiJ (Φj ∗ Λ−1 )(ηiJ · ξ), ξ ∈ Ω, (5.258)
i=1
and
(2mJ +1)2

−1
(ΨJ ∗ Λ ) ∗ G (ξ) =
ε
wiJ (Ψj ∗ Λ−1 )(ηiJ · ξ), ξ ∈ Ω, (5.259)
i=1
(2mJ +1)2

((ΨJ ∗ Ψ̃j ) ∗ Λ−1 ) ∗ Gε (ξ) = wiJ ((ΨJ ∗ Ψ̃j ) ∗ Λ−1 )(ηiJ · ξ), ξ ∈ Ω,
i=1 (5.260)
for j = 0, . . . , J with coefficients wij given by (5.251) and (5.255).
The recursion step leads to the following decomposition scheme:

G ε
→ wJ → wJ−1 → ··· → w0
↓ ↓ ↓
(ΦJ ∗ Λ−1 ) ∗ Gε (ΦJ−1 ∗ Λ−1 ) ∗ Gε (Φ0 ∗ Λ−1 ) ∗ Gε .
The coefficient vectors w0 = (w10 , . . . , w(2m0 +1)2 )T , w1 = (w11 , . . . , w(2m1 +1)2 )T , . . .
allow the following reconstruction scheme of F :
w0 w1 w2
↓ ↓ ↓
(Ψ0 ∗ Λ−1 ) ∗ Gε (Ψ1 ∗ Λ−1 ) ∗ Gε (Ψ2 ∗ Λ−1 ) ∗ G̃

(Φ0 ∗ Λ−1 ) ∗ Gε → + → (Φ1 ∗ Λ−1 ) ∗ Gε → + → (Φ2 ∗ Λ−1 ) ∗ Gε → + → · · · .
N T
Scale Thresholding. As we have seen, the coefficient vectors wj = w1j , . . . , wj j ∈
RNj j = J0 , . . . , J − 1, consists of the components

(2)
wij = aji Φj ∗ Λ−1 ∗ Gε (ηij ), j = J0 , . . . , J − 1 (5.261)
with wij , ηij being known weights and knots of the applied approximate integration
formula.
Since the large “true” coefficients are the ones that should be included in a
selective approximation, in estimating an unknown function it is natural to include
only coefficients larger than some specified threshold value.
In our context a “larger” coefficient is taken to mean one that satisfies for
j = J0 , . . . , J and i = 1, . . . , Nj
2 2
(2)
Φj ∗ Λ−1 ∗ Gε (ηij )
N
wi j = aji

Nj 2 (2)
Gε (ξ) Gε (ζ) Φj ∗ Λ−1 (ξ, ηi j )
N
= (ai )
Ω Ω

(2)
Φj ∗ Λ−1 (ζ, ηi j ) dS(ξ) dS(ζ)
N

(2)
≥ (ai j )2 K(ξ · ζ) Φj ∗ Λ−1 (ξ, ηi j )
N N
Ω Ω

(2)
Φj ∗ Λ−1 (ζ, ηi j ) dS(ξ) dS(ζ)
N
= (kij )2 . (5.262)
For the given threshold values kij such an estimator can be written in explicit
form:

NJ0
(2)
F̂J = I{(wJ0 )2 ≥(kJ0 )2 } ΦJ0 ∗ Λ−1 (·, ηiJ0 ) wiJ0
i i
i=1

J−1 Nj
(2)
I{(wj )2 ≥(kj )2 } ΨJ0 ∗ Λ−1 ·, ηi j
N
+ wij . (5.263)
i i
j=J0 i=1
We conclude our considerations on multi-scale approximation with the decor-

relation of the geoidal undulations into wavelet signatures for the Earth’s Gravi-
tational Model EGM96) via a tree algorithm using Shannon scaling functions
of scales 8, 7, 6, 5, 4 (left) and the Shannon wavelet functions of scales 7, 6, 5, 4, 3
(right) in [m] (see Figure 5.9). The illustrations (left) provide low-pass filtered
geoidal heights from a fine (j = 8) down to a rough (j = 4) resolution, while
the illustrations (right) show the corresponding band-pass filtered geoidal heights.
Each band-pass filtered illustration yields detail information contained in the low-
pass filtered version of scale j + 1, but not in the low-pass filtered version of scale
j for 7, 6, 5, 4, 3.
Figure 5.9 shows that geoidal undulations (and, by virtue of Bruns’s formula,
the Earth’s disturbing potential) are “smooth” functions for large parts, so that
they can be approximated efficiently and economically by a multi-scale procedure
in form of a “read in” (tree) algorithm (see the low-passed filtered version of scale
j = 8, i.e., the topmost illustration (left)). However, it also becomes obvious from
Figure 5.9 that parts of particular geodetic interest, e.g., subduction zones, oro-
genetic areas, etc. are not sufficiently reflected by the global (spherical harmonic)
EGM96-model (see the band-passed filtered version of scale j = 7, i.e., the topmost
illustration (right) characterizing the non-green areas).
Geoid undulations (cf. Figure 5.10) may be understood as a measure for the
perturbations of the Earth from a hydrostatic equilibrium (see, e.g., [185] for a
more detailed geodetic interpretation). They form the deviations of the equipo-
tential surfaces at mean sea level from the reference ellipsoid. Geoid undulations
show no essential correlation to the distributions of the continents. They seem to
be generated by density contrasts deeper inside the Earth.
Figure 5.9. Multi-scale decomposition of the geoidal undulations

(from Earth’s Gravitational Model EGM96) via a tree algorithm using
Shannon scaling functions of scales 8, 7, 6, 5, 4 (left) and the Shan-
non wavelet functions of scales 7, 6, 5, 4, 3 (right) in [m] (Illustration
taken from W. Freeden, M.Z. Nashed, M. Schreiner (2018): Spherical
Sampling, Geosystems Mathematics, Birkhäuser, Basel).
Figure 5.10. 3D (left) and 2D (right) graphical illustrations of the

EGM96 – geoidal surface (taken from [48]).
6. Conclusions: Gaussian geometry and geodetic surveying

We have singled out only four fields of the wide spectrum of Gauss’s works which
are still relevant in mathematics as well as in geodesy up to our times. These four
fields serve us as examples for building bridges between Gauss’s world of ideas
and modern scientific developments. A further important field, the foundation of
surveying, will only be touched upon here. These important foundations for geo-
detic practice even today rest essentially on the mathematical foundations created
by Gauss who was led fairly early to work on the theory of surfaces. Gauss had
planned a summarizing publication on “Higher Geodesy” but he did eventually not
finish it, as also a work on his trigonometric survey in the Kingdom of Hanover
never saw the light of day (Gauss Werke, Vol. VIII, p. 400 and Vol. IX, p. 401).
6.1. The geodesic
Gauss’s investigations concerning the theory of surfaces, the Disquisitiones gen-
erales circa superficies curvas (Gauss Werke, Vol. IV) were published in 1828 and
are dealing extensively among other things with the theory of the geodesic, which
is directly of great importance for geodetic surveys. Indeed, mathematicians like
Bernoulli, Euler, and Clairaut had dealt with this topic before Gauss, but nobody
so completely and profoundly. The findings in this publication are the basis for
the Untersuchungen über Gegenstände der Höheren Geodäsie, published in 1843
and 1846, which give plenty of applications for geodetic tasks and computations
on the ellipsoid of rotation. In the second publication Gauss explains above all the
transfer of geographic coordinates on the ellipsoid from one given point to another
one (the direct problem) using the azimuth and both of the points connecting a
geodesic. In 1806 Legendre was the first to use the geodesic for the transfer of
geographic coordinates. For the solution of this task he developed the so called
Legendre series. Gauss’s solution employs arguments of the means for the series
so that the Taylor-type series converge faster and therefore less elements have to
be computed. For the inversion of the direct problem the determination of the arc
length of the geodesic and its azimuths from the given geographic coordinates of
both points Gauss’s idea proved to be very fertile. His formulas are distinguished
from other solutions due to their superior convergence, but also because Gauss had
proven their accuracy and the size of the neglected terms. Gauss derived these for-
mulas in two different ways: through a conformal mapping of the ellipsoid to the
sphere and through power series expansions stemming from the differential equa-
tions of the geodesic on the ellipsoid of rotation. Gauss, however, did not present
the formulas. This was done by F.R. Helmert [108] based on Gauss’s formulas of
the mean width (cf. [141]). A disadvantage of the series ansatz using arguments of
means lies in the fact that the coefficients of the series have to be newly computed
for each new pair of points, while the coefficients of the Legendre series have to be
computed only once if the coordinates of multiple points have to be determined
with respect to the same pole.
Approximately since 1965 the possibilities of numerical methods to solve the
main tasks of geodesy, provided by standard methods of numerical mathematics,
can be exploited due to the beginning development of electronic data process-
ing computers. The variety of potential solutions can be divided roughly in four
categories (according to B. Heck [105]):
(i) The first kind of approach rests on classical solutions of O. Schreiber [196] and
H. Boltz [12]. Since poles are singular points of the geographical coordinate
system Legendre’s series cannot be used in the vicinity of poles. Moreover,
these series show slow convergence in case of large distances so that this
approach makes sense only in case of a short geodesic up to 150 km and up
to 400 km if Gauss’s formulas are used.
(ii) The second category rests on Bessel’s approach (cf. [6]) of exploiting elliptic
integrals which are today directly computed by numerical quadrature rules
(see, e.g., [193]). This approach can also be used in the vicinity of poles and
is accurate and advisable even in case of large distances of > 500 km.
(iii) The solution approaches in the third category are based on Gauss’s confor-
mal mapping of the ellipsoid to the sphere or the plane (cf. [84]) which are
considered as auxiliary surfaces on which the main tasks of geodesy can be
accomplished by means of closed formulas from spherical and plane trigonom-
etry, respectively. Subsequently, the inverse mapping may be used to transfer
back to the ellipsoid.
(vi) In the fourth category we find methods which were already used by A.M.
Legendre [139] in a special case, namely that a given geodetic line is replaced
by another line connecting start and end point of the geodesic. This other line
may be a spatial chord, for example. Due to geodetic use of Earth satellites
this three-dimensional ansatz has grown in importance.
6.2. Gaussian conformal mapping of the Earth’s surface
For quite some time Gauss had dealt with the question of the method best suited to
coordinate triangulation points because the knowledge of geographical coordinates
is not sufficient. It seemed most convenient to him to exploit the coordinates with
which every point could be represented in a plane (Gauss Werke, Vol. V, p. 367).
About 1815 he recognized the conformal mapping of the triangular points onto the
plane as being the most appropriate solution to the problem. The term “konform”
was used by him since 1843. The task formulated by Gauss, to project a given
surface onto another one in a fairly general manner so that the image becomes
similar to the preimage in its smallest parts was formulated as a contest question
for the year 1821 by the Scientific Society in Copenhagen and reformulated the
next year. Gauss submitted his solution (cf. [78]) on the 11, December 1822 and
won the prize. His solution was distinguished from other known work in that it was
valid in case of arbitrary surfaces and that he had fully laid open the conditions of
conformity, the constancy of the augmenting relation in a certain point, and the
equal angle condition (cf. [141]). Gauss applied his general solution to the following
particular cases:
• Conformal mapping between two planes. As Gauss himself remarked (cf. Art.
8 in [78]), this is a useful method for the geodetic praxis, in order “to trans-
form a map based on mediocre measurements which may be good in small
details but is generally somehow distorted into a better one, if one knows the
correct loci of a number of points.”
• Conformal mapping of the sphere by transversal Mercator’s projection or
other mappings, e.g., the conforming Lambert mapping.
• Conformal mappings of the ellipsoid of rotation onto the sphere. In this map-
ping one can choose a suitable radius of the sphere to keep the differences
between the ellipsoidal and spherical geometries relatively small. Hence, for
practical applications spherical relations can be used which allow closed form
solution. In case of local computations reasonable accuracy can be gained by
letting the radius of the sphere depend on the geographic latitude. To this end
two “replacement spheres” have proved their worth, namely the image sphere
of Soldner, formerly used in Bavarian land surveying, and the Gaussian oscu-
lating sphere. The latter was used by Gauss (see [84]) with the radius M0 N0
with regard to a point P0 , M0 being the radius of curvature of the meridian
and N0 being the oblique radius of curvature.
Since the ellipsoid shows the same measure of curvature in the central
point P0 in this mapping the differences in scale between ellipsoid and sphere
stay small for regions not too large. As Gauss has shown the metric on a
surface is determined by the Gaussian curvature. Hence, line segments and
angles agree on two surfaces with the same Gaussian curvature. The surface
of an ellipsoid, however, shows a variable Gaussian curvature so that the
relationship of curvatures are only equivalent in the small (cf. Figures 6.2
and 6.3).
Under all mappings associated with Gauss’s name the so-called Gauss–Krüger
projection (and similar ones) have enjoyed wide distribution. In this projection the
main meridian is mapped length-preserving to a straight line and is the abscissa
of the plane system. Gauss had chosen this projection for the land surveying of
Hanover (1828–1844) (cf. [141]). The size of the distortion depends only on the
Figure 6.1. Triangulation of Hanoverian arc measurement between

Göttingen and Altona (now a suburb of Hamburg) carried out under
Gauss’s supervision 1821–1825, Collected Works, Vol. 9, p. 347 (Illus-
tration taken from [118]).
distance between the points and the main meridian and is independent from the
north-south dilatation. Today one follows the idea that meridian strips of 6◦ width
with a difference in longitude of 3◦ from the main meridian to the boundary
meridians suffice for tolerable distortions (this is today’s UTM-coordinate system).
Gauss’s work on surface theory was important for the later nineteenth century
in the sense of Sophus Lie, while until his time in geometry only finite groups of
transformations had been considered. Gauss paved the way for the general theory
Figure 6.2. Ellipsoidal orthogonal coordinates (Illustration taken from [233]).
Figure 6.3. Differential projection distortions (Illustration taken from [233]).
of the multiply extended manifolds, or n-dimensional space (see [27] for more
details).
Figure 6.4. A detail of the triangulation of Hanover carried out under

Gauss’s supervision 1821–1825 (Collected Works, Vol. 9, p. 347). Note
that the three lines joining the locations Hohenhagen, Brocken, and
Inselsberg (this point does not appear here, only a part of the connecting
lines is displayed) form a great triangle, the angle at Hohenhagen is
close to a right angle, so that the area of the triangle is close to half the
product of the two short sides.
The Hanover surveying work also stimulated Gauss’s interest in the study of
curves and surfaces in three-dimensional differential geometry in Euclidean space
(today usually called Gaussian geometry). Gauss also was led to the Gaussian
curvature (an intrinsic measure of curvature, dependent only on how distances are
measured on the surface, not on the way it is embedded in space).
6.3. Claims and perspectives

All in all, while engaged on a surveying task for the Royal House of Hanover in the
years after 1818, Gauss was also concerned with the shape of the Earth. He started
to formulate revolutionary ideas, like the concept of the geoid. He questioned one
of the central tenets of the whole of mathematics, Euclidean geometry, which
was clearly premised on a flat, and not a curved, universe. He later claimed to
have considered a non-Euclidean geometry, which was internally consistent and
free of contradiction. Unwilling to court controversy, however, Gauss decided not
to pursue or publish any of his far-reaching concepts in non-Euclidean geometry.
János Bolyai independently discovered non-Euclidean geometry in 1829; his work
was published in 1832. After seeing it, Gauss wrote to Farkas Bolyai, the father of
János Bolyai:
“To praise it would amount to praising myself. For, the entire content of
the work . . . coincides almost exactly with my own meditations which
have occupied my mind for the past thirty or thirty-five years.”
The long history of the discussion of the parallel postulate started probably
sooner than Euclid published his Elements about 300 BC as his fifth postulate
[32], p. 155:
“That, if a straight line falling on two straight lines make the interior
angles on the same side less than two right angles, the two straight
lines, if produced indefinitely, meet on that side on which are the angles
less than the two right angles.”
Many modern authors have commented on the meaning of this postulate; we men-
tion only R. Bonola [13], J. Gray [90], B.A. Rosenfeld [184]. In fact, even in
our time the fifth Euclidean postulate seems odd and we may understand why so
many authors after Euclid tried to prove the fifth postulate (and thus making it a
theorem) from the other axioms and postulates. Already Proclus (412–485), who
wrote a Commentary on the First Book of Euclid’s Elements, tried to prove the
fifth postulate exploiting an argument given by Aristoteles to show the finiteness
of the universe [176], p. 291, and failed, cf. [90], p. 39. Proclus informs us, that the
first of the Ptolemies under whose reign Euclid lived, wrote himself a book on the
fifth postulate and proved it, but Proclus points out the fallacy of this attempt.
John Playfair (1748–1819) gave an alternative formulation of the fifth postulate
in 1795 which is now known as Playfair’s Axiom but which he himself attributed
to William Ludlam (1717–1788). This formulation is the one which most people
refer to nowadays when speaking of the “axiom of parallels” [239, p. 16]:
“Given a line a and a point A not lying on a, then there exists, in the
plane determined by a and A, one and only one line which contains A
but not any point of a.”
Or, even shorter,
“Through a given point can be drawn only one parallel to a given line.”
Many mathematicians tried to get hold of a proof of the fifth postulate; famous
names over the centuries being Nasir al-Din al-Tusi (1201–1274) in Persia, John
Wallis (1616–1703) in England, Giovanni Girolamo Saccheri (1667–1733) in Italy,
the Swiss mathematician Johann Heinrich Lambert (1728–1777) in Germany, and
Adrien-Marie Legendre (1752–1833) in France. At the beginning of the 19th cen-
tury the invention of a Non-Euclidean Geometry seemed to have been in the air.
If so many and different attempts to prove the fifth postulate a mere conclusion
of other axions had failed the question whether there exists a geometry in which
the fifth postulate was wrong became sensible. This may explain why it was not
only Gauss who found the key to this new geometry but also Nikolai Lobachevsky
(1792–1856) and János Bolyai (1802–1860) independently. However, Gauss was
the first, although he did not publish his results. It is now known that it was he
who coined the denotation “Non-Euclidean” for the new geometries which emerge
out of a neglect of the fifth postulate. Already in 1804 Gauss received a letter
of his friend Farkas Bolyai (1775–1856), father of János Bolyai, in which Bolyai
presented a proof that the fifth postulate could be deduced from other axioms of
geometry. Gauss praised the work of his friend, but found a flaw in the argument.
He wrote (cf. [16, p. 100]):
“You want to know my sincere and frank opinion. And this is that your
method does not satisfy me. I will try to make the critical point (which
belongs to the same kind of obstacles which made my own efforts so
futile) as clear as I can. I still hope that these cliffs will be navigated
eventually, and this, before I die. For now, I am, however, extremely
busy with other things. . . ”
Although Gauss remarked in 1846 that he knew about the existence of Non-
Euclidean geometries for the last fifty years it is not before 1816 that we see
written evidence. In that year Gauss reflected on different attempts to prove the
fifth postulate in a book review. Although he was too careful to express his own
opinion we can be sure that the reason why he reviewed these attempts can be
seen in the fact that Gauss was already convinced in 1816 that Non-Euclidean
geometries existed.
Perhaps the Non-Euclidean geometry which can be most easily understood
is the spherical geometry. The navigators, map makers, and naval mathematicians
of the 16th century were very well aware of this type of geometry in which every
triangle (build from parts of three great circles on the sphere) shows a sum of
inner angles of larger than 180◦ and is thus today seen as a simple model of elliptic
geometry. The Non-Euclidean geometries by Gauss, Lobachevsky, and Bolyai are
nowadays classified as being hyperbolic geometries, in which a triangle shows an
angular sum of less than 180◦ .
Spherical geometry was Gauss’s bread-and-butter occupation as early as 1803
when he started to survey the Duchy of Brunswieck by means of a triangulation
and met his wife to be, Johanna Osthoff. After Heinrich Christian Schumacher
(1780–1850), astronomer and geodesist, informed Gauss of the Danish arc mea-
surement and stimulated a surveying of the Hanoverian lands to fit the Danish
measurements in the north, Gauss became enthusiastic. Eventually, Schumacher
succeeded in persuading the English king Georg IV, who was head of the house
of Hanover, to authorize Gauss with the surveying work, and Gauss started field
work in 1821 (see Figures 6.1 and 6.4) concerned with measurements from three
mountains in Germany, Hohenhagen, near Göttingen, Brocken in the Harz Moun-
tains and Inselberg in the Thüringer Wald to the south. The three lines joining
these locations form a great triangle, the angle at Hohenhagen is close to a right
angle, so the area of the triangle is close to half the product of the two short
sides). Gauss made the important invention of the heliotrope (cf. Figure 6.5) to
ensure measurements of hitherto unknown accuracy. In fact, heliotropes were used
Figure 6.5. Heliotrope (Geophysical Institute of Göttingen University).
in surveys from Gauss’s survey in Germany in 1821 through the late 1980s, when
GPS measurements replaced the use of the heliotrope in long distance surveys.
Surveying field work lasted until 1825 when Gauss withdraw from working in the
countryside. However, he oversaw the further surveying work going on in 1828 and
lasting until 1844 and did all the number crunching necessary. Friends, including
Friedrich Wilhelm Bessel (1784–1846), criticized him for wasting his time in these
computations instead of creating new theorems, but Gauss saw deeper. In a letter
to Bessel dated March 14, 1824, he wrote (cf. [8]):
“. . . you accused me of loosing my time and wished me luck that the loss
of time might be over soon. Great God, how wrong you judge me. . . .
Certainly, I also think like you in that matter. All measurements in the
world do not outweigh a single theorem, with which the science of eternal
truths will be truly advanced. But you should not judge over the absolute,
but over the relative worth [of measurements]. . . . And however small
you estimate this worth, in my eyes it is higher than those concerns
which are interrupted by them.”
Surveying problems also motivated Gauss to develop his thoughts on least
squares and more general problems of what is now called mathematical statistics.
The result was the definitive exposition of his mature concepts in the note “Theoria
combinationis observationum erroribus minimis obnoxiae” (1823, with supplement
in 1826). In “Bestimmung des Breitenunterschiedes zwischen den Sternwarten von
Göttingen and Altona durch Beobachtungen am Ramsdenschen Zenithsector” of
the year 1828 he summed up his ideas on the figure of the Earth, instrumental
errors, and the calculus of observations. Furthermore, his publication “Disquisi-

tiones generales circa superficies curvas” (1828), which grew out of his meditations
in surveying and geodesy of three decades, represented the seed of more than a
century of work on differential geometry.
Finally it should be mentioned that resulting research led to, among other
things, Einstein’s theory of general relativity, which describes the universe as non-
Euclidean.
References
[1] Adams, D.R., Hedberg, L.I.: Function Spaces and Potential Theory. Grundlehren
der Mathematischen Wissenschaften, 314, Springer, Berlin, (1996).
[2] Adrain, R.: Research Concerning the Probabilities of the Errors which Happen in
Making the Observations. The Analyst of Mathematical Museum, Vol. I, W.P. Far-
rand and Co., Philadelphia, (1808).
[3] Augustin, M., Freeden, W.: A Survey on Classical Boundary Value Problems in
Physical Geodesy. Encyclopedia of Geodesy, E.W. Grafarend, ed., Springer Int.
Publ. Switzerland, DOI 10.1007/978-3-319-02370-0-1117-1, (2015).
[4] Baer, R.: Linear Algebra and Projective Geometry. Academic Press, New York,
(1952).
[5] Ben-Israel, A., Greenville, T.N.E., Generalized Inverses: Theory and Applications.
New York: Wiley-Interscience, (1974).
[6] Bessel, W.: Über die Berechnung der geographischen Längen und Breiten aus geo-
dätischen Vermessungen. Astronomische Nachrichten, Bd. 4, Nr. 86, Altona, (1826).
[7] Bialas, V.: Erdgestalt, Kosmologie und Weltanschauung. Konrad Wittwer Stuttgart,
(1982).
[8] Biermann, K.-R: Carl Friedrich Gauss – Der “Fürst der Mathematiker” in Briefen
und Gesprächen. Verlag C.H. Beck, München, (1990).
[9] Bjerhammar, A.: Rectangular Reciprocal Matrices with Special Reference to Geo-
detic Calculations. Bulletin Géodesique, 188–220, (1951).
[10] Bjerhammar, A.: Gravity Reduction to an Internal Sphere. Division of Geodesy,
Stockholm, Sweden, (1962).
[11] Börsch, A.: Simon, P.: Abhandlungen zur Methode der kleinsten Quadrate von Carl
Friedrich Gauss (in deutscher Sprache herausgegeben). Berlin, (1887).
[12] Boltz, H.: Formeln und Tafeln zur numerischen (nicht logarithmischen) Berechnung
geographischer Koordinaten aus den Richtungen und Längen der Dreiecksseiten er-
ster Ordnung. Veröff. Geod. Inst. Potsdam, Neue Folge, Nr. 110, (1942).
[13] Bonola, R: Non-Euclidean Geometry – With a Supplement Containing ‘The Theory
of Parallels’ by Nicholas Lobachevski and ‘The Science of Absolute Space’ by John
Bolyai. Engl. transl. of the first edition 1912. Dover Publication, New York, (1955).
[14] Boullion, T.L., Odell, P.L.: Generalized Inverse Matrices. Wiley-Interscience, New
York, (1971).
[15] Bruns, H.: Die Figur der Erde “Ein Beitrag zur europäischen Gradmessung”. P.
Stankiewicz, Berlin, (1878).
[16] Bühler, W.K.: Gauss. A Biographical Study. Springer-Verlag, Berlin, Heidelberg,

New York, (1981).
[17] Bühler, W.K.: Gauss – Eine Biographische Studie. Springer, Berlin. Heidelberg,
(1987).
[18] Chen, J.-R.: The Lattice Points in a Circle. Sci. Sinica, 12:633–649, (1963).
[19] Christoffel, E.B.: Über die Gaussische Quadratur und eine Verallgemeinerung der-
selben. J. Reine Angew. Math., 55:61–82. (Ges. Math. Abhandlungen I, 65–87),
(1858).
[20] Cui J., Freeden W.: Equidistribution on the Sphere. SIAM J. Sci. Stat. Comp.,
18:595–609, (1997).
[21] Davis, C.H.: Theory of the Motion of Heavenly Bodies Moving about the Sun in
Conic Sections, translation of “Theoria motus corporum coelestium in sectionibus
conicis solem ambientium” by C.F. Gauss, reprinted Dover, New York, (1963).
[22] Davis, C.H., Robinson, D.W.: Generalized Inverses of Morphisms, Linear Algebra
and Its Applications, 5:319–328, (1972).
[23] Davis, P.J.: Interpolation and Approximation. Blaisdell, New York, (1963).
[24] Davis, P.J., Rabinowitz, P.: Numerical Integration. Blaisdell, Toronto, London,
(1967).
[25] Dedekind, R.: Gauß in seiner Vorlesung über die Methode der kleinsten Quadrate.
Festschrift zur Feier des hundertfünfzigjährigen Bestehens der Königlichen Gesell-
schaft der Wissenschaften zu Göttingen, Weidmannsche Buchhandlung, (1901).
[26] Driscoll, J.R., Healy, R.M.: Computing Fourier Transforms and Convolutions on
the 2-Sphere. Advances in Applied Mathematics, 15:202–250, (1994).
[27] Dunnington, G.W.: Carl Friedrich Gauss: Titan of Science A Study of his Life and
Work. Exposition Press, New York, (1955).
[28] Edgeworth, F.Y.: On Observations Relating to Several Quantities. Hermathena,
6:279–285, (1887).
[29] Engl, H.: Discrepancy Principles for Tikhonov Regularization of Ill-posed Prob-
lems, Leading to Optimal Convergence Rates. J. Optim. Theory Appl., 52:209–215,
(1987).
[30] Engl, H.: Integralgleichungen. Springer, Berlin, (1997).
[31] Eisenhart, C., Carl Friedrich Gauss. International Encyclopedia of Social Sciences,
New York, Vol. VI, 74–81, (1986)
[32] Euclid: The Thirteen Books of The Elements. Vol. 1. 2nd Edition, translated with
introduction and commentary by Sir Thomas L. Heath. Dover Publication, New
York, (1956).
[33] Euler, L.: Methodus universalis serierum convergentium summas quam proxime in-
veniendi, Commentarii Academiae Scientiarum Petropolitanae. 8:3–9, Opera Omnia
(XIV), 101–107, (1736).
[34] Fredholm I.: Sur une classe d’équations functionelles. Acta Mathematica 27:365–
390, (1903).
[35] Freeden, W.: Über eine Verallgemeinerung der Hardy–Landauschen Identität.
Manuscr. Math., 24:205–216, (1978).
[36] Freeden, W.: Über die Gaußsche Methode zur angenäherten Berechnung von Inte-
gralen. Math. Method. Appl. Sci., 2:397–409, (1980).
[37] Freeden, W.: On the Approximation of External Gravitational Potential with
Closed Systems of (Trial) Functions. Bull. Géod., 54, 1–20, (1980a).
[38] Freeden, W.: On Integral Formulas of the (Unit) Sphere and Their Application to
Numerical Computation of Integrals. Computing, 25:131–146, 1980b.
[39] Freeden, W.: On Spherical Spline Interpolation and Approximation. Math. Meth.
in the Appl. Sci., 3:551–575, (1981a).
[40] Freeden, W.: On Approximation by Harmonic Splines. Manuscr. Geod., 6:193–244,
(1981b).
[41] Freeden, W.: Least Squares Approximation by Linear Combinations of (Multi)-
Poles. Report 344, Department of Geodetic Science and Surveying, The Ohio State
University, Columbus, (1983).
[42] Freeden, W.: Spherical Spline Interpolation: Basic Theory and Computational As-
pects. J. Comput. Appl. Math., 11:367–375, (1984).
[43] Freeden, W.: A Spline Interpolation Method for Solving Boundary Value Prob-
lems of Potential Theory From Discretely Given Data. Math. Part. Diff. Equations,
3:375–398, (1987).
[44] Freeden, W.: Interpolation by Multidimensional Periodic Splines. J. Approx. Theory
55:104–117, (1988).
[45] Freeden, W.: Multiscale Modelling of Spaceborne Geodata. B.G. Teubner, Stuttgart,
Leipzig, (1999).
[46] Freeden, W., Schreiner, M.: Spherical Functions of Mathematical Geosciences – A
Scalar, Vectorial, and Tensorial Setup. Springer, Heidelberg, (2009).
[47] Freeden, W.: Metaharmonic Lattice Point Theory. CRC Press, Taylor & Francis
Group, Boca Raton, (2011).
[48] Freeden, W.: Geomathematics: Its Role, Its Aim, and Its Potential. in: Handbook of
Geomathematics, Vol. 1, 2nd ed., W. Freeden, M.Z. Nashed, T. Sonar, Eds., 3–79,
Springer, New York, Berlin, Heidelberg, (2015).
[49] Freeden, W., Gerhards, C.: Geomathematically Oriented Potential Theory. Chap-
man and Hall, CRC Press, Boca Raton, London, New York, (2013).
[50] Freeden, W., Gutting, M.: Special Functions of Mathematical (Geo)Physics. Birk-
häuser, Basel, (2013).
[51] Freeden, W., Kersten, H.: The Geodetic Boundary-Value Problem Using the Known
Surface of the Earth. Veröff Geod Inst RWTH Aachen, 29, (1980).
[52] Freeden, W., Kersten, H.: A Constructive Approximation Theorem for the Oblique
Derivative Problem in Potential Theory. Math. Meth. Appl. Sci., 4:104–114, (1981).
[53] Freeden, W., Kersten, H.: An Extended Version of Runge’s Theorem. Manuscr.
Geod., 7:267–278, (1982).
[54] Freeden, W., Maier, T.: On Multiscale Denoising of Spherical Functions: Basic
Theory and Numerical Aspects. Electronic Transactions on Numerical Analysis
(ETNA), 14:40–62, (2002).
[55] Freeden, W., Mayer, C.: Multiscale Solution for the Molodensky Problem on Reg-
ular Telluroidal Surfaces, Manuscr. Math., Acta Geod. Geophys. Hung., 41:55–86,
(2006).
[56] Freeden, W., Michel, V.: Multiscale Potential Theory (With Applications to
Geoscience). Birkhäuser, Boston, (2004).
[57] Freeden, W., Nashed, M.Z.: Multi-Variate Hardy-Type Lattice Point Summation
and Shannon-Type Sampling. Int. J. Geomath., 6:163–249, (2015).
[58] Freeden, W., Schneider, F.: Regularization Wavelets and Multiresolution. Inverse
Problems, 14:225–243, (1998).
[59] Freeden, W., Schneider, F.: Runge–Walsh Wavelet Approximation for the Helm-
holtz Equation. J. Math. Anal. Appl., 235, 533–566, (1999).
[60] Freeden, W., Schreiner, M.: Orthogonal and Non-Orthogonal Multiresolution Anal-
ysis, Scale Discrete and Exact Fully Discrete Wavelet Transform on the Sphere.
Const. Approx., 14:493–515, (1998).
[61] Freeden, W., Schreiner, M.: Local Multiscale Modelling of Geoid Undulations from
Deflections of the Vertical. J. Geod., 79:641–651, (2006).
[62] Freeden, W. and Witte, B.: A Combined (Spline-)Interpolation and Smoothing
Method for the Determination of the Gravitational Potential from Heterogeneous
Data. Bull. Geod., 56:53–62, (1982).
[63] Freeden, W., Wolf, K.: Klassische Erdschwerefeldbestimmung aus der Sicht moder-
ner Geomathematik. Math. Semesterb., 56: 53–77, (2008).
[64] Freeden, W., Gervens, T., Schreiner, M.: Constructive Approximation on the Sphere
(With Applications to Geomathematics). Oxford Science Publications, Clarendon,
Oxford, (1998).
[65] Freeden, W., Fehlinger, T., Klug, M., Mathar, D., Wolf, K.: Classical Globally Re-
flected Gravity Field Determination in Modern Locally Oriented Multiscale Frame-
work. J. Geodesy, 83:1171–1191, (2009).
[66] Fricker, F.: Einführung in die Gitterpunktlehre, Birkhäuser, Basel, Boston, Stutt-
gart, (1981).
[67] Friedrich, K.: Allgemeine für die Rechenpraxis geeignete Lösung für die Aufgaben
der kleinsten Absolutsumme und der günstigsten Gewichtsverteilung, Zeitschrift
für Vermessungswesen, 337–358, (1937).
[68] Friedrichs, K.O.: Functional Analysis and Applications. pp. 144–147, 151–153, New
York University, mimeographed lecture notes.
[69] Galle, A.: Über die geodätischen Arbeiten von Gauß, I. Abschnitt: Die Methode
der kleinsten Quadrate. Berlin, (1924), (Gauß Werke Bd. XI 2. Abteilung).
[70] Garland, G.D.: The Contributions of Carl Friedrich Gauss. Historia Mathematica,
6:5–29, (1979).
[71] Gauss, C.F.: Untersuchungen über Höhere Arithmetik (Disquisitiones Arithmeticae
(Latin)), Gerhard Fleischer jun., Leipzig, auch: Gauss Werke, Band 1, (1801).
[72] Gauss, C.F.: Theoria motus corporum coelestium in sectionibus conicis solem ambi-
entium, Hamburg, (1809), Werke 7. Translated into English by C.H. Davis, (1963).
[73] Gauss, C.F.: Disquisitio de elementis ellipticis Paladis ex opositionibus annorum
1803, 1804, 1805, 1807, 1808, 1809, Göttingen, (1810).
[74] Gauss, C.F.: Disquisitiones generales circa seriem infinitam 1+((α·β)/ (1−γ))x+· · ·
etc., “Commentationes Societatis Regiae Scientiarium Gottingensis Recentiores”, 2.
(Werke Ill, 123–162], (1812).
[75] Gauss, C.F.: Theoria attractionis corporum sphaeroidicorum ellipticorum homo-
geneorum methodo nova tractata, Commentationes societatis regiae scientiarium
Gottingensis recentiores, 2, 355–378, (1813).
[76] Gauss, C.F.: Methodus nova integralium valores per approximationem inveniendi.
Commentationes Societatis Regiae Scientiarium Gottingensis Recentiores 3, (Werke
III, 163–196), (1814).
[77] Gauss, C.F.: Theoria combinationis observationum erroribus minimis obnoxiae, 1.
Teil, Göttingen, (1821).
[78] Gauss, C.F.: Allgemeine Auflösung der Aufgabe: Die Teile einer gegebenen Fläche
auf einer andern gegebenen Fläche so abzubilden, dass die Abbildung dem Abge-
bildeten in den kleinsten Theilen ähnlich wird. In H.C. Schumacher (Hrsg.): As-
tronomische Abhandlungen, Heft 3, Altona 1825, S. 5–30, (1822).
[79] Gauss, C.F.: Theoria combinationis observationum erroribus minimis obnoxiae. 2.
[80] Gauss, C.F.: Supplementum theoriae combinationis observationum erroribus min-
imis obnoxiae. Göttingen, (1826).
[81] Gauss, C.F.: Allgemeine Theorie des Erdmagnetismus. Resultate aus den Beobach-
tungen des magnetischen Vereins, Göttingen, (1838).
[82] Gauss, C.F.: Werke, Vierter Band. Hrsg. Königl. Gesellschaft der Wissenschaft zu
Göttingen, 109–117, B.G. Teubner, Leipzig, (1880).
[83] Gauss, C.F.: Werke, Neunter Band, Hrsg. Königl. Gesellschaft der Wissenschaften
zu Göttingen, B.G. Teubner, Leipzig, (1903).
[84] Gauss, C.F.: Untersuchungen über Gegenstände der Höheren Geodäsie. Hrsg. J.
Frischauf, Ostwald’s Klassiker der exakten Wissenschaften, Nr. 177, Leipzig, (1910).
[85] Gautschi, W.: A Survey of Gauss–Christoffel Quadrature Formulas. In: E.B.
Christoffel: The Influence of His Work on Mathematics and the Physical Sciences.
P.L. Butzer and F. Feher, Eds., Birkhäuser, Basel, (1981).
[86] Gerling, C.L.: Ausgleichungsrechnungen der praktischen Geometrie oder die Meth-
ode der kleinsten Quadrate. Hamburg und Gotha, (1843).
[87] Gnedenko, B.W.: Über die Arbeiten von C.F. Gauß zur Wahrscheinlichkeits-
rechnung, 193–204, in: C.F. Gauß Leben und Werk, Hrsg. H. Reichardt, Haude
& Spenersche Verlagsbuchhandlung, Berlin, (1960).
[88] Goldstine, H.H.: A History of Numerical Analysis from the 16th Through the 19th
Century. Springer-Verlag, New York, (1977).
[89] Grafarend, C.F., Klapp, M., Martinec, M.: Spacetime Modelling of the Earths Grav-
ity Field by Ellipsoidal Harmonics. Handbook of Geomathematics, W. Freeden, Z.
Nashed, T. Sonar, Eds., Springer, Heidelberg, 2nd Edition, 381–496, (2015).
[90] Gray, J: Ideas of Space. Euclidean, Non-Euclidean, and Relativistic. Clarendon
Press, Oxford, (1979).
[91] Green, G.: An Essay on the Application of Mathematical Analysis to the Theories
of Electricity and Magnetism. T. Wheelhouse, Nottingham, England, (1838).
[92] Groetsch, C.W.: Generalized Inverses of Linear Operators. Marcel Dekker. Inc.,
New York, (1977).
[93] Groetsch, C.W.: Inverse Problems in the Mathematical Science. Vieweg, Braun-
schweig, (1993).
[94] Groten, E.: Geodesy and the Earth’s Gravity Field I + II. Dümmler, Bonn, (1979).
[95] Gutting, M.: Fast Multipole Methods for Oblique Derivative Problems. Ph.D.-thesis,
Geomathematics Group, University of Kaiserslautern, (2007).
[96] Gutting, M.: Fast Multipole Accelerated Solution of the Oblique Derivative Boundary
Value Problem. GEM Int. J. Geomath., 3:223–252, (2012).
[97] Hammer, E.: Beitrag zur Geschichte der Ausgleichungsrechnung. Zeitschrift für
Vermessungswesen, 613–628, (1900).
[98] Hampel, F.R., Ronschetti, E.M, Rousseeuw, P.R., Stahel, W.A.: Robust Statistics.
Wiley, New York, (1986).
[99] Hansen, P.A.: Von der Methode der kleinsten Quadrate im Allgemeinen und in
ihrer Anwendung auf die Geodäsie. Abhandlungen der Mathematisch-Physikalischen
Klasse der Königlich-Sächsischen Gesellschaft der Wissenschaften, Band 8, 571–
806, S. Hirzel, Leipzig, (1867).
[100] Hansen, P.A: Kurz gefaßte rationelle Ableitung des Ausgleichsverfahrens eines
Dreiecksnetzes nach der Methode der kleinsten Quadrate. Berichte der Mathema-
tisch-Physikalischen Klasse der Königlich-Sächsischen Gesellschaft der Wissenschaf-
ten, Band 20, 129–150, S. Hirzel, Leipzig, (1868).
[101] Hansen, P.A: Fortgesetzte geodätische Untersuchungen bestehend aus zehn Supple-
menten zur Abhandlung von der Methode der kleinsten Quadrate. Abhandlungen
der Mathematisch-Physikalischen Klasse der Königlich-Sächsischen Gesellschaft
der Wissenschaften, Band 9, 1–185, S. Hirzel, Leipzig, (1871).
[102] Hansen, P.A.: Entwicklung eines neuen veränderten Verfahrens zur Ausgleichung
eine Dreiecksnetzes mit besonderer Beachtung des Falles, in welchem gewisse Winkel
im voraus bestimmte Werte erhalten sollen. Abhandlungen der Mathematisch-
Physikalischen Klasse der Königlich-Sächsischen Gesellschaft der Wissenschaften,
Band 9, 185–289, S. Hirzel, Leipzig, (1871).
[103] Hardy, G.H.: On the Expression of a Number as the Sum of Two Squares. Quart.
J. Math. (Oxford), 46:263–283, (1915).
[104] Harter, H.L.: Nonuniqueness of Least Absolute Values Regression. Commun.
Statist. Theor. Meth., A6 (9):829–838, (1977).
[105] Heck, B.: Rechenverfahren und Auswertemodelle der Landesvermessung Klassische
und moderne Methoden. 3. Auflage, Wichmann Verlag, Heidelberg, (2003).
[106] Heine, E.: Anwendungen der Kugelfunctionen und der verwandten Functionen. 2nd
Edition, Reimer, Berlin. (I. Theil: Mechanische Quadratur, 1–31), (1881).
[107] Heiskanen, W.A., Moritz, H.: Physical Geodesy. Freeman, San Francisco, (1967).
[108] Helmert, F.R.: Lotabweichungen. Heft 1, Veröff. d. Kgl. Preuss. Geod. Inst., Berlin,
(1886).
[109] Heun, K.: Neue Methode zur approximativen Integration der Differentialgleichun-
gen einer unabhängigen Veränderlichen. Z. Math. Phys., 45:23–38, (1900).
[110] Heuser, H.: Funktionalanalysis. 4. Auflage, Teubner, (1975).
[111] Hilbert, D.: Grundzüge einer allgemeinen Theorie der linearen Integralgleichungen.
Teubner, Leipzig, (1912).
[112] Hofmann-Wellenhof, B., Moritz, H.: Physical Geodesy. Springer, Wien, New York,
2nd Edition, (2006).
[113] Hörmander, L.: The Boundary Problems of Physical Geodesy. The Royal Institute
of Technology, Division of Geodesy, Report 9, Stockholm, (1975).
[114] Hotine, M.: Mathematical Geodesy. ESSA Monography 2, U.S. Department of Com-
merce, Washington (Reprint 1992 by Springer), (1969).
[115] Huber, P.J.: Robust Statistics. Wiley, New York, (1981).
[116] Hurwitz, W.A.: On the Pseudo-Resolvent to the Kernel of an Integral Equation
Transactions of the American Mathematical Society 13: 405–418, (1912).
[117] Huxley, M.N.: Exponential Sums and Lattice Points III. Proc. London Math.,
87:591–609, (2003).
[118] Jordan, W., Eggert, O.: Handbuch der Vermessungskunde. Erster Band, Ausglei-
chungsrechnung nach der Methode der kleinsten Quadrate, 9. Aufl., J.B. Metzler,
Stuttgart, (1948).
[119] Kanani, E.: Robust Estimators for Geodetic Transformations and GIS. Ph.D.-thesis,
ETH No. 13521, Institut für Geodäsie und Photogrammetrie, Eidg. Technische
Hochschule Zürich, Mitteilung Nr. 70, Zürich, (2000).
[120] Kantorowitsch, L.W., Akilow, G.P.: Funktionalanalysis in normierten Räumen.
Akademie-Verlag, Berlin, (1964).
[121] Kellogg, O.D.: Foundations of Potential Theory. Frederick Ungar Publishing Com-
pany, New York, (1929).
[122] Kertscher, D.: Carl Friedrich Gauss – “Genie, Gigant, Titan”. Z. f. Vermessungswe-
sen, 130. Jg., 6:372–378, (2005).
[123] Koch, K.R.: Parameterschätzung und Hypothesentests. Dümmler Verlag, Bonn
(1997).
[124] Koch, K.R.: Parameter Estimation and Hyothesis Testing in Linear Models. 2nd
ed., Springer, Berlin, (1999).
[125] Koch, K.R., Pope, A.J.: Uniqueness and Existence for the Geodetic Boundary Value
Problem Using the Known Surface of the Earth. Bulletin Géodésique, 106:467–476,
(1972).
[126] Koch, K.R., Witte, B.: The Earth’s Gravity Field Represented by a Simple Layer
Potential From Doppler Tracking of Satellites. NOAA Technical Memorandum NOS
9, National Oceanic and Atmospheric Administration, National Ocean Service,
Rockville, Md., (1971).
[127] Koch, K.R., Witte, B.: The Earth’s Gravity Field Represented by a Simple Layer
Potential From Doppler Tracking of Satellites. Journal of Geophysical Research,
76(35): 8471–8479, The American Geophysical Union, (1971).
[128] Kolesnik, G.: On the Method of Exponent Pairs. Acta Arith., 45: 115–143, (1985).
[129] Kowalewski, A.: Newton, Cotes, Gauss, Jacobi: Vier grundlegende Abhandlungen
über Interpolation und genäherte Quadratur. von Veit, Leipzig, (1917).
[130] Krarup, T.: A Contribution to the Mathematical Foundation of Physical Geodesy.
Danish Geodetic Institute, Report No. 44, Copenhagen, (1969).
[131] Krarup, T.: Letters on Molodensky’s Problem I–IV. Communication to the members
of the IAG Special Study Group 4.31, (1973).
[132] Krüger, L.: Über die Ausgleichung von bedingten Beobachtungen in zwei Gruppen.
Veröff. d. Kgl. Preu. Inst., N. F., Nr. 18, Potsdam, (1905).
[133] Kutta, W.: Beitrag zur näherungsweisen Integration totaler Differentialgleichungen.
Z. Math. Phys., 46: 435–453, (1901).
[134] Landau, E., Über die Gitterpunkte in einem Kreis (Erste Mitteilung). Gött. Nachr.,
148–160, (1915).
[135] Landau, E.: Über die Gitterpunkte in einem Kreis IV. Gött. Nachr., 58–65, (1924).
[136] Laplace, P.S.: Traité de mécanique céleste. tome 2, Paris, (1799).
[137] Laplace, P.S.: Théorie analytique des probabilités. Livre II, Chap. IV. Paris, (1812).
[138] Laplace, P.S.: Théorie analytique des probabilités. Œuvres, tome VII, p. 353, (1812).
[139] Legendre A.M.: Nouvelles méthodes pour la détermination des orbites des comètes.
Paris, (1806).
[140] Legendre, A.M.: Analyse des triangles tracés sur la surface d’un sphéroide. Tome
VII de la I Série des mémoires de l’Académie des Sciences, Paris, 131, (1806).
[141] Lehmann, G.: Gauss theoretische Arbeiten. In C.F. Gauss und die Landesvermes-
sung in Niedersachsen, herausgegeben von der Niedersächsischen Vermessungs- und
Katasterverwaltung, (1955).
[142] Lehmann, R., Lösler, M.: Multiple Outlier Detection: Hypothesis Tests versus
Model Selection by Information Criteria. Journal of Surveying Engineering, 142:
DOI: http://dx.doi.org/10.1061/(ASCE)SU.1943-5428.0000189, (2016).
[143] Listing J.B.: Über unsere jetzige Kenntnis der Gestalt und Größe der Erde. Diet-
richsche Verlagsbuchhandlung, Göttingen, (1873).
[144] Magnus, W., Oberhettinger, F., Soni, R.P.: Formulas and Theorems for the Special
Functions of Mathematical Physics. In: Die Grundlehren der mathematischen Wis-
senschaften in Einzeldarstellungen, Band 52, Springer, Berlin, 3. Auflage, (1966).
[145] Maire, Ch., Boscović, R.J.: De Litteraria Expeditione per Pontificiam Ditionem ad
Dimentiendos duos Meridiani Gradus, Rom, 409–503, (1755).
[146] Markina, I., Vodopyanov, S.K.: Fundamentals of the Nonlinear Potential Theory
for Subelliptic Equations. I, II. Siberian Adv. Math., 7:32–62, (1997).
[147] Marx, C.: Outlier Detection by Means of Monte Carlo Estimation Including Resis-
tant Scale Estimation. J. of Applied Geodesy, 9:123–141, (2015).
[148] McEwen, J.D., Puy, G., Thiran, J.-P., Vandergheinst, P.: Sampling Theorems and
Compressive Sensing on the Sphere. Wavelets and Sparsity XIV, Papadakis, M.,
van de Ville, D., Goyal, V.K., Eds., Proc. of SPIE, Vol. 8138, F1–F9, (2011).
[149] Mehler, F.G.: Bemerkungen zur Theorie der mechanischen Quadraturen. J. Reine
Angew. Math., 63:152–157, (1864).
[150] Meissl, P.: On the Linearization of the Geodetic Boundary Value Problem. Depart-
ment of Geodetic Science, No. 152, The Ohio State University, Columbus, OH,
(1971).
[151] Molodensky, M.S., Eremeev, V.F., Yurkina, M.I.: Methods for Study of the Exter-
nal Gravitational Field and Figure of the Earth. Trudy TSNIIGAiK, Geodezizdat,
Moscow, 131, (1960), English translat.: Israel Program for Scientific Translation,
Jerusalem, (1962).
[152] Moore, E.H.: On the Reciprocal of the General Algebraic Matrix. Bulletin of the
American Mathematical Society 26:394–395, (1920).
[153] Moore, E.H.: General Analysis. Memoirs of the American Mathematical Society,
1:197–209, (1935).
[154] Moritz, H.: Statistische Methoden in der gravimetrischen Geodäsie. Zeitschrift für
Vermessungswesen, 409–416, (1963).
[155] Moritz, H.: Der Begriff der mathematischen Erdgestalt seit Gauss. Allgemeine Ver-
messungs-Nachrichten, 133–138, (1977).
[156] Moritz, H.:Advanced Physical Geodesy. Wichmann Verlag, Karlsruhe, (1980).
[157] Moritz, H.: Geodesy and Mathematics. Zeszyty Naukowe Akademii Görniczo-
Hutniczej IM. Stanislawa Staszica, No. 780, Geodezja, 63, Kraków, 38–43, (1981).
[158] Moritz, H.: Advanced Physical Geodesy, Wichmann, Karlsruhe, 2nd Edition, (1989).
[159] Moritz, H.: Classical Physical Geodesy. Handbook of Geomathematics, W. Freeden,
Z. Nashed, T. Sonar, Eds., Springer, Heidelberg, 2nd Edition, 253–290, (2015).
[160] Müller, C.: Eine Verallgemeinerung der Eulerschen Summenformel und ihre Anwen-
dung auf Fragen der analytischen Zahlentheorie. Abh. Math. Sem. Univ. Hamburg,
19:41–61, (1954).
[161] Müller, C., Analysis of Spherical Symmetries in Euclidean Spaces. Springer, New
York, Berlin, Heidelberg, (1998).
[162] Nashed, M.Z.: Generalized Inverses, Normal Solvability and Iteration for Singular
Operator Equations. In: Nonlinear Functional Analysis and Applications, L.B. Rall,
ed., Academic, New York, 311–359, (1971).
[163] Nashed, M.Z.: Aspects of Generalized Inverses in Analysis and Regularization. In:
Generalized Inverses and Applications. Academic Press, New York, 193–244, (1976).
[164] Nashed, M.Z.: Generalized Inverses and Applications. Academic Press, New York,
(1976).
[165] Nashed, M.Z.: Operator-Theoretic and Computational Approaches to Ill-Posed
Problems with Applications to Antenna Theory. IEEE Transactions on Antennas
and Propagation, 29:220–231, (1981).
[166] Nashed, M.Z.: Inner, Outer, and Generalized Inverses in Banach and Hilbert Spaces.
Numer. Funct. Anal. Optim., 9:261–326, (1987).
[167] Nashed, M.Z., Walter, G.G.: General Sampling Theorems for Functions in Repro-
ducing Kernel Hilbert Spaces. Math. Control Signals Systems, 4:363–390, (1991).
[168] Neumann, F., Vorlesungen über die Theorie des Potentials und der Kugelfunktionen,
Teubner, Leipzig, 135–154, (1887).
[169] Neumann, J.V.: On Regular Rings. Proceedings of the National Academy of Science
of the United States of America, 22:707–713, (1936).
[170] Nutz, H.: A Unified Setup of Gravitational Field Observables. Ph.D.-thesis, Geo-
mathematics Group, University of Kaiserslautern, Shaker, Aachen, (2002).
[171] Palczewski, A., Schneider, J., Bobylev, A.: A Consistency Result for a Discrete
Velocity Model of the Boltzmann Equation. SIAM J. Numer. Anal., 34:1865–1883,
(1997).
[172] Penrose, R.: A Generalized Inverse for Matrices. Proceedings of the Cambridge
Philosophical Society 51, 406–413, (1955).
[173] Petrini, H.: Sur l’existence des dérivées secondes du potentiel. C.R. Acad. Sci. Paris,
130:233–235, (1900).
[174] Pizzetti, P.: Sopra il Calcoba Tesrico delle Deviazioni del Geoide dall’Ellissoide.
Att. R Accad. Sci. Torino, 46:331–350, (1910).
[175] Plackett, R.L.: A Historical Note on the Method of Least Squares. Biometrika,
36:458–460, (1949).
[176] Proclus: A Commentary on the First Book of Euclid’s Elements. Translated with
Introduction and Notes, by Glenn R. Morrow. Princeton University Press, (1970).
[177] Radau, R.: Étude sur les formules d’approximation qui servent à calculer la valeur
numérique d’une intégrale définie. J. Math. Pures Appl., 3(6):283–336, (1880).
[178] Rado, R.: Note on Generalized Inverses of Matrices. Proceedings of the Cambridge
Philosophical Society 52:600–601, (1956).
[179] Reich, K., Roussanova, E.: Visualising Geomagnetic Data by Means of Correspond-
ing Observations. GEM International J. Geomath., 4:1–25, (2013).
[180] Reid, W.T.: Generalized Inverses of Differential and Integral Operators. Theory
and Applications of Generalized Inverses of Matrices (T.L. Boullion and P.L. Odell
1–25). Symposium Proceedings, Texas Tech University Mathematics Series, Vol. 4.,
Lubbock; Texas Tech Press, (1965).
[181] Reinsch, C.: Smoothing by Spline Functions. Numer. Math., 10:177–183, (1967).
[182] Reuter, R.: Über Integralformeln der Einheitssphäre und harmonische Splinefunk-
tionen. Ph.D.-thesis, Veröff. Geod. Inst. RWTH Aachen, Report No. 33, (1982).
[183] Robinson, D.W.: Gauss and Generalized Inverses. Historia Mathematics, 7:118–125,
(1980).
[184] Rosenfeld, B.A.: A History of Non-Euclidean Geometry – Evolution of the Concept
of a Geometric Space. Springer Science+Business Media, (1988).
[185] Rummel, R.: Geodesy. Encyclopedia of Earth System Science, W.A. Nierenberg,
ed., Academic Press, 2:253–262, (1992).
[186] Rummel, R.: Spherical Spectral Properties of the Earth’s Gravitational Potential
and Its First and Second Derivatives. In: Lecture Notes in Earth Science, Springer,
Berlin, Vol. 65:359–404, (1997).
[187] Runge, C.: Zur Theorie der eindeutigen analytischen Funktionen, Acta Math. 6,
229–234, (1885).
[188] Runge, C., Willers, F.A.: Numerische und graphische Integration. Encyklopädie der
Math. Wiss., 2(3):47–176, Leipzig, (1915).
[189] Sanso, F.: A Note on Density Problems and the Runge Krarup’s Theorem. Bolletino
di Geodesia e Science Affini 41, 422–477, (1982).
[190] Sard, A.: Integral Representations of Remainders. Duke Math. J. 15:333–345,
(1948).
[191] Schauder, J.: Potentialtheoretische Untersuchungen, Math. Z., 35:536–538, (1931).
[192] Schlossmacher, E.J.: An Iterative Technique for Absolute Deviations Curve Fitting.
J. Americ. Statistical Association, 68:857–859, (1973).
[193] Schmidt; H.: Lösung der geodätischen Hauptaufgaben auf dem Rotationsellipsoid
mittels numerischer Integration. Zeitschrift für Vermessungswesen 123:121–128,
(1999).
[194] Schreiber, O.: Anordnung der Winkelbeobachtungen im Göttinger Basisnetz.
Zeitschrift für Vermessungswesen, 129–161, (1882).
[195] Schreiber, O.: Die konforme Doppelprojektion der trigonometrischen Abteilung der
Königlich Preussischen Landesaufnahme. Formeln und Tafeln. Selbstverlag, Berlin,
(1897).
[196] Schreiber, O.: Formeln und Tafeln zur Berechnung der geographischen Coordinaten
aus Richtungen und Längen der Dreiecksseiten, Königl. Preuss. Landesaufn., Berlin,
(1878).
[197] Schreiner, M.: Wavelet Approximation by Spherical Up Functions. Habilitation The-
sis, Geomathematics Group, University of Kaiserslautern, Shaker, Aachen, (2004).
[198] Shannon, C.E.: Communication in the Presence of Noise. Proc. Institute of Radio
Engineers, 37:10–21, (1949).
[199] Shure, L., Parker, R.L., Backus, G.E.: Harmonic Splines for Geomagnetic Modelling.
Phys. Earth Planet. Inter. 28:215–229, (1982).
[200] Siegel, C.L.: Über die analytische Theorie der quadratischen Formen, III. Annals
of Mathematics 38:212–291, (1937).
[201] Sierpinski, W.: O pewnem zagadnieniu z rachunku funckcyj asmptotycznych (Über
ein Problem des Kalküls der asymptotischen Funktion (polnisch)), Prace Math.-
Fiz., 17:77–118, (1906).
[202] Simons, F.J., Plattner, A.: Scalar and Vector Slepian Functions, Spherical Sig-
nal Estimation and Spectral Analysis. Handbook of Geomathematics (W. Freeden,
M.Z. Nashed, T. Sonar, Eds.), 2nd Edition, Vol. 3, 2563–2608, (2015).
[203] Slepian, D.: Some Comments on Fourier Analysis, Uncertainty and Modeling. SIAM
Rev., 25(3):379–393, (1983).
[204] Sonar, T., 3000 Jahre Analysis. Springer, Berlin, Heidelberg, New York, (2011).
[205] Stewart G.W.: Theoria Combinationis Observationum Erroribus Minimis Obnoxiae.
Pars Prior et Pars Posterior and Supplementum, Theory of the Combination of
Observations Least subject to Errors, Part One, Part Two, Supplement by C.F.
Gauß, Society for Industrial and Applied Mathematics, (1995).
[206] Stoer, J., Numerische Mathematik 1. 5. Auflage, Springer-Verlag, Berlin, Heidel-
berg, New York, (1989).
[207] Stokes, G.G.: On the Variation of Gravity on the Surface of the Earth. Trans. Camb.
Phil. Soc., 8:672–695, (1849).
[208] Svensson, S.L.: Pseudodifferential Operators. A New Approach to the Boundary
Value Problems of Physical Geodesy. Manuscr. Geod., 8:1–40, (1983).
[209] Szegö G.: Orthogonal Polynomials. American Mathematical Society Colloquium
Publications, 4th edition, Volume 23, American Mathematical Society, Providence,
Rhode Island, (1975).
[210] Taylor, A.E., Lay, D.: Functional Analysis. Wiley, New York, (1979).
[211] Tikhonov, A.N.: On the Stability of Inverse Problems. Dokl. Akad. Nauk. SSSR,
39:195–198, (1943).
[212] Tikhonov, A.N.: On the Solution of Incorrectly Formulated Problems and the Reg-
ularization Method, Dokl. Akad Nauk SSSR, 151:501–504, (1963).
[213] Tikhonov, A.N.: On Methods of Solving Incorrect Problems, Amer. Math. Soc.
Transl., 2, 70: 222–224, (1968).
[214] Tikhonov, A.N., Arsenin V.Ya.: Solutions of Ill-Posed Problems. Wiley, New York,
(1977).
[215] Tikhonov, A.N., Goncharsky, A.V., Stepanov, V.V., Yagola, A.G.: Numerical Meth-
ods for the Solution of Ill-Posed Problems. Kluwer, Dordrecht, (1995).
[216] Todhunter, I., A History of the Mathematical Theories of Attraction and of the
Figure on the Earth from the Time of Newton to that of Laplace, Macmillan,
London (reprint 1962 by Dover Publications, New York), (1878).
[217] Vening Meinesz, F.A.: A Formula Expressing the Deflection of the Plumb-Line in
the Gravity Anomalies and some Formulae for the Gravity Field and the Gravity
Potential Outside the Geoid. Proceedings of the Koninklijke Nederlandse Akademie
van Wetenschappen, 31(3):315–331, (1928).
[218] Vekua, I.N.: Über die Vollständigkeit des Systems harmonischer Polynome im
Raum. Dokl. Akad. Nauk 90, 495–498, (1953).
[219] Vogler, C.A.: Lehrbuch der praktischen Geometrie. Bd. II 1. Halbband, Braun-
schweig, (1894).
[220] Voronoi, G.: Sur un problème du calcul des fonctions asymptotiques. JRAM,
126:241–282, (1903).
[221] Wahba, G.: Spline Models for Observational Data. SIAM Philadelphia, (1990).
[222] Walsh, J.L.: The Approximation of Harmonic Functions by Harmonic Polynomials
and by Harmonic Rational Functions, Bull. Amer. Math. Soc. 35, 499–544, (1929).
[223] Wangerin, A.: Theorie des Potentials und der Kugelfunktionen, Walter de Gruyter
& Co, Berlin, Leipzig, (1921).
[224] Weizel, R.: C.F. Gauß, Angewandte Mathematik 200 Jahre nach seiner Geburt.
Zeitschrift für Vermessungswesen, 520–525, (1977).
[225] Wermer, J.: Potential Theory. Lecture Notes in Mathematics, 408, Springer, Berlin,
(1974).
[226] Werner, J., Numerische Mathematik 1. Vieweg Studium, Braunschweig, (1991).
[227] Werner, D.: Funktionalanalysis. 4. Auflage; Springer, Berlin, (2002).
[228] Weyl, H.: Über die Gleichverteilung von Zahlen mod. Eins. Math. Ann., 77:313–352,
(1916).
[229] Witte, B.: Die Berechnung von Schwerestörungen im Außenraum mit Hilfe des
Greenschen Satzes. Zeitschrift für Vermessungswesen (ZfV), 135–141, Stuttgart,
(1969).
[230] Witte, B.: Vergleich verschiedener Methoden für die Berechnung von Schwere-
störungen im Außenraum. Schweizerische Zeitschrift für Vermessung, Photogram-
metrie und Kulturtechnik, 161–168, (1969).
[231] Witte, B.: Die Bestimmung von Horizontalableitungen der Schwere im Außenraum
aus einer Weiterentwicklung der Stokesschen Funktion. Gerlands Beiträge zur Geo-
physik, 87–94, Leipzig, (1970).
[232] Witte, B.: Computational Procedures for the Determination of a Simple Layer
Model of the Geopotential From Doppler Observations. NOAA Technical Report
NOS-42, National Oceanic and Atmospheric Administration, National Ocean Ser-
vice, Rockville, Md., (1971).
[233] Witte, B., Sparla, P.: Vermessungskunde und Grundlagen der Statistik für das
Bauwesen. 8. Auflage, Wichmann – VDE-Verlag, Berlin, (2015).
[234] Wolf, H.: Ausgleichungsrechnung nach der Methode der kleinsten Quadrate. Dümm-
ler Verlag, Bonn, (1968).
[235] Wolf, H.: Ausgleichungsrechnung. Formeln zur praktischen Anwendung. Dümmler
Verlag, Bonn, (1975).
[236] Wolf, H.: C.F. Gauß und die Methode der kleinsten Quadrate. Allgemeine Vermes-
sungs-Nachrichten, 125–133, (1977).
[237] Wolf, H.: Die deutsche Landesvermessung in ihrer historischen Entwicklung.
Festschrift 100 Jahre Geodätische Lehre und Forschung in Hannover. Wissen-
schaftliche Arbeiten der Fachrichtung Vermessungswesen der Universität Hannover,
Nr. 100, 47–53. Hannover, (1981).
[238] Wolf, H.: Neues Altes in der Ausgleichungsrechnung. Vermessung, Photogramme-
trie, Kulturtechnik, 233–240, (1983).
[239] Wolfe, H.E.: Introduction to Non-Euclidean Geometry. Reprint of the first edition
1945. Dover Publications, Mineola, New York, (2012).
[240] Wußing, H.: Carl Friedrich Gauss. B.G. Teubner Verlag, Leipzig, (1989).
[241] Yoshida, K.: Functional Analysis. 5th Edition, Springer, Berlin, (1980).
Willi Freeden
Thomas Sonar
Institute for Computational Mathematics
University of Brunswick
Pockelsstraße 14
D-38106 Braunschweig, Germany
e-mail: t.sonar@tu-bs.de
Bertold Witte
Institute for Geodesy and Geoinformation
University of Bonn
Nussallee 17
D-53115 Bonn, Germany
e-mail: bertold.witte@uni-bonn.de

An Overview on Tools from Functional Analysis

Matthias Augustin, Sarah Eberle, and Martin Grothaus
Abstract. Many modern mathematical methods treat geodetic problems in

terms of functions from certain spaces, proving convergence properties of such
functions and regard the evaluation of such functions or their derivatives at
given points as operators. In doing so, knowingly or unknowingly, they use
the language of functional analysis.
This contribution aims at summarizing some fundamental concepts from
functional analysis which are used throughout this book. In this way, it tries
to add a layer of self-sufficiency and to act as supplement to other contribu-
tions for those readers who are not familiar with functional analytic tools. For
this purpose, we introduce, among others, the general ideas of vector spaces,
norms, metrics, inner products, orthogonality, completeness, Banach spaces,
Hilbert spaces, functionals, linear operators, different notions of convergence.
Then we show how functions can be interpreted as vectors in different kind
of function spaces, e.g., spaces of continuous functions, Lebesgue spaces, or
Sobolev spaces and how the more general concepts come into play here. More-
over, we have a brief glimpse at differential equations and how functional
analytic tools provide the necessary background to discuss them, and at the
idea of reproducing kernels and the corresponding reproducing kernel Hilbert
spaces.
Keywords. Functional analysis, metric spaces, normed spaces, function spaces,

Sobolev spaces, reproducing kernel Hilbert spaces, basis systems, operators,
convergence, weak derivatives, distributions, partial differential equations.
1. Introduction
Readers of this book who do not have a strong mathematical background might
wonder why it includes this chapter. Let us give a short motivation.
Most readers are probably familiar with the Fourier expansion and its spher-
ical relative, the expansion of a function in terms of spherical harmonics as it
is used, for example, to determine the gravitational potential of the Earth. This
166 M. Augustin, S. Eberle, and M. Grothaus
leads to the question which functions can be expanded in a Fourier series. Func-
tional analysis provides the context to discuss this question. As the determination
of Fourier coefficients requires the computation of integrals, the resulting answer
introduces the space of rapidly decreasing functions. However, this is probably not
the space a practitioner prefers. Dealing with signals – and in this context, the
gravitational field of the Earth might be considered as a signal – it is natural to
demand that a solution has finite energy, as it is otherwise physically impossible.
The energy density of a signal is proportional to the square of its amplitude. Thus,
the total energy is given by the integral of the square of the amplitude and has
to be finite. This yields the concept of square-integrable functions and the more
general Lebesgue spaces. It would be desirable to extend the concept of Fourier
expansions from the rapidly decreasing to square integrable functions.
The tools to discuss all of the above are given by functional analysis.
Let us consider another, more ambitious task. One of the main topics of
geodesy is the determination of the gravity field in the exterior of the Earth. As
the density distribution inside the Earth is not accessible, we cannot use Newton’s
volume potential for this purpose. Instead, early geodesists could only use data
from measurements at the Earth’s surface, yielding such results as Stokes’ famous
integral formula to compute the disturbing potential from gravity anomalies ([20]).
However, with the advancement of satellite technology, the situation changed dra-
matically in several ways. One of them is that, with modern GPS, gravity distur-
bances become available in more and more areas, leading to the conclusion that
instead of Stokes’ formula, a similar integral formula due to von Neumann be-
comes applicable. Nowadays, geodesists are confronted with a plethora of different
quantities, as not only gravity anomalies or gravity disturbances, but also, among
others, gravitational gradients, gravitational tensors, deflections of the vertical,
or height anomalies can be measured. Additionally, some of these quantities are
available from terrestrial measurements while others are collected by satellites.
This yields two questions:
1. Is it possible to find a framework which allows a common interpretation for
all of these different measurements?
2. Can we combine different types of data to get more accurate results?
If, for a moment, we consider only gravity anomalies and gravity disturbances,
the first of these two questions might be reformulated: what are the common
properties of the integral formulas by Stokes and by Neumann? Both formulas take
a certain function, assumed to be given on a sphere, as input and give another
function, the disturbing potential in the exterior of the Earth, as output. As we
try to deduce which kinds of inputs are allowed and how properties of the input
and output are linked to each other, we arrive again at the concepts of function
spaces, (integral) operators, limits and convergence. Once more, we have reached
the mathematical domain of functional analysis.
It is only a small step from opening the rich toolbox of functional analysis to
interpreting the above mentioned measurements as the application of operators.
An Overview on Tools from Functional Analysis 167
Some of these operators are linear, some non-linear and some derived by linearizing
non-linear operators. But it would be shortsighted to assume that functional anal-
ysis is only useful to describe the setting of the problem. It also provides tools to
solve those problems efficiently. However, the aim of this chapter is not to present
solutions to geodetic problems, as this is done in other chapters, but to provide
the reader with the necessary background to understand modern approaches in
mathematical geodesy.
For this purpose, we start by recalling definitions and properties of metric
spaces, normed spaces, Banach spaces, and Hilbert spaces as well as linear opera-
tors and sesquilinear or bilinear forms. This leads us to the consideration of differ-
ent kinds of convergence, e.g., with respect to a given norm, weak convergence and
weak convergence. Next, we specifically consider function spaces, starting with
classical spaces of continuous or continuously differentiable functions, but also tak-
ing into account spaces of distributions, Lebesgue spaces, and Sobolev spaces. The
latter ones are particularly useful when dealing with differential equations in weak
form, yielding results on existence, uniqueness and regularity of solutions. The
chapter is completed by a short discussion of reproducing kernel Hilbert spaces.
2. Basic concepts
This section summarizes some basic concepts from functional analysis taken from
[2, 4, 10].
2.1. Metric spaces, normed spaces, Banach spaces, and linear operators
Let N, N0 , Z, R, R+ , R+ 0 , and C denote the set of positive integers, non-negative
integers, integers, real numbers, positive real numbers, non-negative real numbers,
and complex numbers, respectively.
In the following, V , W , and Z are K-vector spaces over the field K ∈ {R, C}.
As a particular space, we have for any n ∈ N, the n-dimensional vector spaces
K × · · · × K = Kn .

n-times
For the elements x = (x1 , x2 , . . . , xn )T , y = (y1 , y2 , . . . , yn )T ∈ Kn we define
the (Euclidean) inner product x · y and its induced (Euclidean) norm x by

n
x·y = xi y i , (1)
i=1
6
7 n
√ 7
x = x·x =8 |xi |2 , (2)
i=1
where z denotes the complex conjugate of z for any z ∈ C and |·| is the absolute
value in K, i.e., for x ∈ R,
⎧
⎪
⎨ x, x > 0,
|x| = 0, x = 0, (3)
⎪
⎩ −x, x < 0,
and for z = x + iy ∈ C, x, y ∈ R, i the imaginary unit,
#
|z| = x2 + y 2 . (4)
The canonical orthonormal basis vectors are denoted by e1 , e2 ,. . . , en . In R3
we have
e1 = (1, 0, 0)T , e2 = (0, 1, 0)T , e3 = (0, 0, 1)T . (5)
Furthermore, let δn,m given by

1, n=m
δn,m = (6)
0, n = m
for n, m ∈ N0 , denote the Kronecker delta.
For any subset G ⊂ Rn , ∂G denotes the boundary and its closure is given by
G = G ∪ ∂G.
In addition, an open ball of radius r ∈ R+ centered at x ∈ Kn is defined by
Br (x) = {y ∈ Kn : x − y < r}. (7)
The corresponding sphere is given by
∂Br (x) = {y ∈ Kn : x − y = r}. (8)
Definition 2.1 (Metric Space). A function ρ : V × V → R+ 0 , u, v → ρ(u, v) is

called a metric if it has the following properties:
(a) for all u, v ∈ V , ρ(u, v) ≥ 0 and ρ(u, v) = 0 ⇐⇒ u = v,
(b) for all u, v ∈ V , ρ(u, v) = ρ(v, u),
(c) for all u, v, w ∈ V we have the triangle inequality ρ(u, v) ≤ ρ(u, w) + ρ(v, w).
Definition 2.2 (Convergence in Metric Spaces). A sequence (vn )n∈N in V is called
convergent with limit v0 ∈ V if lim ρ(vn , v0 ) = 0.
n→∞
Definition 2.3 (Cauchy Sequence). A sequence (vn )n∈N in V is called Cauchy

sequence if for any ε ∈ R+ exists a n0 ∈ N such that for all n, m ∈ N, n ≥ m ≥ n0 ,
we have ρ(vn , vm ) ≤ ε.
Definition 2.4 (Completeness). A metric space V is called complete if every Cauchy
sequence is convergent.
Definition 2.5 (Density). Let W, Z be two sets in a metric space V such that
W ⊂ Z ⊂ V . W is said to be dense in Z if for every z in Z and every ε ∈ R+
exists a w ∈ W such that ρ(w, z) ≤ ε.
Definition 2.6 (Separability). A metric space V is called separable if it contains a

countable, dense subset.
Definition 2.7 (Normed Space). A function · : V → R+ 0 , v → v is called a
norm if it has the following properties for all u, v ∈ V , α ∈ K:
(a) v ≥ 0 and v = 0 ⇐⇒ v = 0,
(b) αv = |α|v,
(c) u + v ≤ u + v.
From every norm · on V , a metric can be constructed by
ρ(u, v) = u − v = v − u, u, v ∈ V. (9)
Additionally to the defining properties of a metric, a metric defined from a norm
also satisfies for all u, v ∈ V , α ∈ K:
ρ(u + w, v + w) = ρ(u, v), (10)
ρ(αu, αv) = |α|ρ(u, v). (11)
Vice versa, a metric that also has properties (10) and (11) can be used to define a
norm via
v = ρ(v, 0), v ∈ V. (12)
Definition 2.8 (Complete System of Elements). Let V be a normed space and I a
suitable set to index a system of elements {vi : i ∈ I} ⊂ V . The system {vi : i ∈ I}
is called complete if span {vi : i ∈ I}, i.e., the set of all finite linear combinations
of the elements {vi : i ∈ I}, is dense in V .
Theorem 2.9. Any finite-dimensional normed space is complete.
Theorem 2.10. Any finite-dimensional subspace of a normed space V is closed.
Definition 2.11 (Banach Space). A normed space (V, ·V ) is called Banach space
if V is complete with respect to ·V .
If there is no risk of confusion, we usually do not state the norm explicitly
and also omit to index the norm by the corresponding space.
Definition 2.12 ((Schauder) Bases in Banach Spaces). Let V be a Banach space.
A sequence (vn )n∈N in V is called a basis, if every element v ∈ V can be written in

∞ N
the form v = an vn = lim an vn with coefficients an ∈ K and the elements
n=1 N →∞ n=1
of every finite subset of (vn )n∈N are linearly independent.
Whereas every Banach space with a countable basis is separable, not every
separable Banach space has a countable basis (see [5]).
Definition 2.13 (Operators and Functionals). Let V , W be two spaces and U an
arbitrary subset of V . A function L : U → W , U v → Lv ∈ W that maps every
element v ∈ U on a unique element Lv ∈ W is called an operator. The set U is

called the domain of the operator L and denoted by D(L). The set
R(L) = {w ∈ W : w = Lv, v ∈ D(L)} (13)
is called the range of L. The set
N (L) = {v ∈ D(L) : Lv = 0} (14)
is called the kernel or null space of L.
If R(L) ⊂ K, the operator is called a functional.
Definition 2.14 (Continuous Operator). Let (V, ρV ) and (W, ρW ) be two metric
spaces. An operator L : V ⊃ D(L) → R(L) ⊂ W is called continuous in v0 ∈
D(L) if for every sequence (vn )n∈N ⊂ D(L) with lim ρV (vn , v0 ) = 0 we have
n→∞
lim ρW (Lv0 , Lvn ) = 0.
n→∞
L is said to be continuous on D(L) if L is continuous for all v0 ∈ D(L).
Definition 2.15 (Linear Operator). Let D(L) ⊂ V be K-vector spaces and L :

D(L) → R(L) ⊂ W . L is called a linear operator if for arbitrary α ∈ K, and
v1 , v2 ∈ D(L) we have
L(αv1 + v2 ) = αLv1 + Lv2 . (15)
It follows from Definition 2.15 that N (L) is a linear subspace of D(L) and
R(L) is a linear subspace of W .
Definition 2.16 (Isometric Linear Operator). Let (V, ·V ), (W, ·W ) be two
normed spaces and L : V ⊃ D(L) → R(L) ⊂ W a linear operator. L is called
isometric if for every v ∈ D(L) we have LvW = vV .
In the following, we assume every operator to be linear.
Definition 2.17 (Inverse Operator). Let L : V ⊃ D(L) → R(L) ⊂ W such that for
every w ∈ R(L) exists a unique v ∈ D(L) with Lv = w. The operator that assigns
to every w ∈ R(L) its inverse image v ∈ D(L) is called the inverse operator to L,
denoted by L−1 , and is a linear operator from R(L) to D(L). We have
L−1 (Lv) = v for all v ∈ D(L), (16)
−1
L(L w) = w for all w ∈ R(L). (17)
As the existence of the inverse operator requires the uniqueness of the inverse
image, a linear operator L has an inverse operator L−1 if and only if N (L) = {0}.
Definition 2.18 (Bounded Operator and Operator Norm). Let (V, ·V ), (W, ·W )
be two normed spaces and L : V ⊃ D(L) → R(L) ⊂ W a linear operator. L is
bounded if it exists a C ∈ R+
0 such that
LvW ≤ C vV for all v ∈ D(L). (18)

The space of all linear bounded operators with domain D(L) and range R(L) is
denoted by L (D(L), R(L)). This space can be equipped with the operator norm,
given as
LvW
LL (D(L),R(L)) = sup = sup LvW = sup LvW . (19)
v∈D(L) vV v∈D(L) v∈D(L)
v =0 vV =1 vV ≤1
We have LvW ≤ LL (D(L),R(L)) vV .

The space V = L (V, K) is called the (topological) dual space to V and consists of
all continuous linear functionals on V . Elements of V are denoted by v . A norm
on V is defined in the same way as the general operator norm.
The dual space of V is denoted by V and called the bidual space of V .
A space is called reflexive, if V and V are isomorph.
Theorem 2.19 (Continuity of Linear Operators). Let (V, ·V ), (W, ·W ) be two
normed spaces and L : V ⊃ D(L) → R(L) ⊂ W a linear operator. L is continuous
if and only if L is bounded.
Definition 2.20 (Dual Space Adjoint Operator). Let (V, ·V ), (W, ·W ) be two
normed spaces and L ∈ L (V, W ). The dual space adjoint operator L : W → V
is defined by
(L w )(v) = w (Lv), v ∈ V, w ∈ W . (20)
As
|(L w )v| ≤ w W LL (V,W ) vV (21)
we have L ∈ L (W , V ) and L L (W ,V ) ≤ LL (V,W ) .
2.2. Sesquilinear forms, inner products, and Hilbert spaces
Definition 2.21 (Sesquilinear Form, Bilinear Form and Inner Product Space).
A function a(·, ·) : V × V → K is called a sesquilinear form if for every u, u1 , u2 ,
v, v1 , v2 ∈ V and ever α ∈ K we have
a(α u, v) = α a(u, v), (22)
a(u, α v) = α a(u, v), (23)
a(u1 + u2 , v) = a(u1 , v) + a(u2 , v), (24)
a(u, v1 + v2 ) = a(u, v1 ) + a(u, v2 ). (25)
If instead of (23), we have
a(u, α v) = α a(u, v), (26)
then a(·, ·) is called a bilinear form.
A sesquilinear form is called Hermitian if
a(u, v) = a(v, u). (27)
A bilinear form is called symmetric if
a(u, v) = a(v, u). (28)
A Hermitian sesquilinear form is called positive definite if

a(u, u) ≥ 0 and a(u, u) = 0 if and only if u = 0. (29)
A function (·, ·)V : V × V → K is called an inner product if it is a positive definite,
Hermitian sesquilinear form. The space V equipped with (·, ·)V is called an inner
product space.
Lemma 2.22 (Cauchy–Schwarz Inequality). Let (V, (·, ·)V ) be an inner product
space. For ever u, v ∈ V , we have
|(u, v)V |2 ≤ (u, u)V (v, v)V . (30)
Definition 2.23 (Pre-Hilbert Space and Hilbert Space). An inner product space
(V, (·, ·)V ) is called a pre-Hilbert space if it is equipped with the norm defined by

uV = (u, u)V . (31)
If V is complete with respect to the norm ·V defined by its inner product, it is
called a Hilbert space.
Definition 2.24 (Properties of Sesquilinear Forms). Let (V, (·, ·)V ) be a pre-Hilbert
space. A sesquilinear form a(·, ·) : V × V → K is called
(a) bounded or continuous, if there is a C ∈ R+ such that |a(u, v)| ≤ C uV vV
for all u, v ∈ V ,
2
(b) coercive, if there is a c ∈ R+ such that a(u, u) ≥ c uV for all u ∈ V .
The norm of a bounded sesquilinear form is defined as
aL (V ×V,K) = sup |a(u, v)| . (32)
uV ≤1, vV ≤1
Lemma 2.25. Let (V, (·, ·)V ) be a pre-Hilbert space. Then the inner product is con-
tinuous in both components and the norm satisfies the parallelogram identity

2 2 2 2
u + vV + u − vV = 2 uV + vV (33)
for all u, v ∈ V .
Definition 2.26 (Orthogonality, Orthonormal System). Let (V, (·, ·)V ) be a pre-
Hilbert space. Two elements u and v of V are orthogonal, if (u, v)V = 0.
A finite or countable set of elements, {un : n ∈ N } ⊂ V , N ⊂ N, is called an
orthonormal system if (ui , uk )V = δi,k for all i, k ∈ N.
Theorem 2.27 (Approximation in Pre-Hilbert Spaces, Bessel’s Inequality). Let
(V, (·, ·)V ) be a pre-Hilbert space and {ek : k = 1, . . . , n}
⊂ V , n ∈ N an or-
thonormal system. For every u ∈ V , the element a = nk=1 (u, ek )V ek satisfies
u − aV ≤ u − vV for all v ∈ span {ek : k = 1, . . . , n}.
The coefficients (u, ek )V are called Fourier coefficients of u with respect to
{uk : k = 1, . . . , n}.
Moreover, from u − aV ≥ 0 follows Bessel’s inequality

n
2 2
|(u, ek )V | ≤ uV . (34)
k=1
Theorem 2.28 (Orthonormal Hilbert Basis). Let (V, (·, ·)V ) be a pre-Hilbert space
and {ek : k ∈ N} ⊂ V an orthonormal system. This orthonormal system is a
Hilbert basis if it satisfies one of the following equivalent properties:
(a) span ({ek }k∈N ) is dense in V .
(b) The system {ek : k ∈ N} is closed, i.e., for all u ∈ V , there is a unique

n
sequence (ak )k∈N in K, such that u = lim ak e k .
n→∞ k=1

∞
(c) {ek : k ∈ N} is complete, i.e., u = (u, ek )V ek for all u ∈ V .
k=1
(d) The Parseval identity holds for {ek : k ∈ N}, i.e., for all u ∈ V we have
2
∞
2
uV = |(u, ek )V |
k=1
(e) The extended Parseval identity holds for {ek : k ∈ N}, i.e., for all u, v ∈ V

∞
we have (u, v)V = (u, ek )V (v, ek )V .
k=1
If (V, (·, ·)V ) is a Hilbert space, the above properties are equivalent to the property
that there exists no 0 = u ∈ V such that (u, ek )V = 0 for all k ∈ N.
Theorem 2.29 (Existence of an Orthonormal Basis). Every separable Hilbert space
contains an orthonormal basis.
Remark 2.30. If {uk : k ∈ N } ⊂ V , N ⊂ N, is an orthonormal system in a separa-
ble Hilbert space V which is not complete, we can find a system {vk : k ∈ N \ N } ⊂
V such that both systems together form an orthonormal basis.
Theorem 2.31 (Riesz Representation Theorem). Let (V, (·, ·)V ) be a Hilbert space.
For every bounded linear functional f ∈ V exists a unique element w ∈ V such
that f (v) = (v, w)V for every v ∈ V .
The function J : V → V , w → J(w) = (·, w)V is an isometric, conjugate
linear isomorphism. Consequently, f V = wV .
Remark 2.32. With the extended Parseval identity from Theorem 2.28, we get in
separable Hilbert spaces V
∞
∞
∞

f (v) = (v, w)V = (v, ek )V (w, ek )V = (v, ek )V (ek , w)V = (v, ek )V f (ek )
k=1 k=1 k=1
(35)
and thus
∞

w= f (ek ) ek . (36)
k=1
Definition 2.33 (Hilbert Space Adjoint Operator). Let (V, (·, ·)V ), (W, (·, ·)W ) be
Hilbert spaces, JV and JW the corresponding Riesz isomorphisms according to
Theorem 2.31 and L ∈ L (V, W ). The Hilbert space adjoint operator L∗ of L is
defined by
L∗ = JV−1 L JW , (37)
whereas L is the dual space adjoint operator to L according to Definition 2.20.
L is called self-adjoint, if L∗ = L.
Lemma 2.34. Let (V, (·, ·)V ), (W, (·, ·)W ) be Hilbert spaces and L ∈ L (V, W ). Then
we have for the Hilbert space adjoint operator L∗ ∈ L (W, V ) and
(v, L∗ w)V = (Lv, w)W . (38)
Moreover, we have L∗∗ = L, L∗ L (W,V ) = LL (V,W ) and L∗ LL (V,V ) =
2
LL (V,W ) .
Theorem 2.35 (Lax–Milgram Theorem). Let (V, (·, ·)V ) be a Hilbert space and
a(·, ·) : V × V → K a sesquilinear form. If a(·, ·) is continuous, i.e., bounded
with the constant C ∈ R+ , then there exists a unique bounded linear operator
A ∈ L (V, V ) such that
a(u, v) = (u, Av)V for all u, v ∈ V. (39)
We have AL (V,V ) ≤ C or, more precisely, AL (V,V ) = a(·, ·)L (V ×V,K) .
If a(·, ·) is also coercive with coercivity constant c ∈ R+ , then A is bijective,
i.e., there exists a -bounded- linear operator A−1 ∈ L (V, V ) which is the inverse
operator to A and A - −1 - 1
≤ c.
L (V,V )
Corollary 2.36. Let (V, (·, ·)V ) be a Hilbert space and a(·, ·) : V × V → K a
continuous, coercive sesquilinear form, A the corresponding operator according to
Theorem 2.35 and J : V → V the Riesz isomorphism according to Theorem 2.31.
For f ∈ V let u = A−1 J −1 f .
The so-defined u is the unique solution to the problem
a(u, v) = f (v) for all v ∈ V (40)
and satisfies
1
uV ≤ f V (41)
c
with c ∈ R+ being the coercivity constant to the sesquilinear form.
If, additionally, a(·, ·) is Hermitian, i.e., a(·, ·) is an inner product, then u is
also the unique minimizer of the functional
1
V v → F (v) = a(v, v) − Re(f (v)) ∈ R. (42)
2
2.3. Weak and weak convergence

Definition 2.37 (Duality). Let V, W be normed spaces. A bilinear form ·, · :
V × W → K is called duality, if
(a) for all 0 = v ∈ V exists a w ∈ W such that v, w = 0 and
(b) for all 0 = w ∈ W exists a v ∈ V such that v, w = 0.
The most common duality is the duality between a linear space V and its
dual space V . For a functional v ∈ V and a v ∈ V this duality is defined by
v , v = v (v).
Definition 2.38 (Weak and Weak Convergence and Compactness). Let V be a
Banach space.
(a) A sequence (vn )n∈N in V converges weakly to v ∈ V , also written
n→∞
vn −−−− v,
if and only if
lim vn , v = v, v for all v ∈ V . (43)
n→∞
(b) A sequence (vn )n∈N in V converges weakly to v ∈ V , also written

n→∞
vn −−−− v ,
if and only if
lim v, vn = v, v for all v ∈ V. (44)
n→∞

(c) Weak and weak Cauchy sequences are defined correspondingly.
(d) A subset U ⊂ V is called weak sequentially compact if and only if each
sequence in U possesses a weak convergent subsequence whose weak limit is
also in U .
(e) Weak compactness for subsets of V is defined analogously.
(f) Convergence with respect to the norm is subsequently called strong conver-
gence.
Lemma 2.39. Let V be a Banach space.
(a) Weak and weak limits are unique and weakly or weakly convergent sequences
are bounded in the norms of the corresponding spaces.
(b) 2
There exists
3 an isometric mapping J4 ∈ L (V, V ) that can be defined by
4 = v, v for any v ∈ V and every v ∈ V . V is called reflexive if
v , Jv
and only if J4 is surjective.
4n
(c) A sequence (vn )n∈N in V converges weakly to v ∈ V if and only if Jv
n∈N
4 ∈ V .
converges weakly to Jv
n→∞ n→∞
(d) From vn −−−− v follows vV ≤ lim inf vn V and from vn −−−− v follows
n→∞
v V ≤ lim inf vn V .
n→∞
Remark 2.40.
(i) If V is a Hilbert space, the Riesz representation Theorem 2.31 implies that
n→∞
vn −−−− v means lim (vn , u) = (v, u) for all u ∈ V .
n→∞
(ii) Strong convergence implies weak or weak convergence, respectively.
(iii) In the dual space V of a Banach space V , we now have three different
concepts of convergence:
n→∞
(a) strong convergence: vn −−−−→ v , i.e., lim vn − v V = 0,
n→∞
n→∞
(b) weak convergence: vn −−−− v , i.e., limn→∞ v, vn = v, v for all
v ∈ V , and
n→∞
(c) weak convergence: vn −−−− v , i.e., limn→∞ vn , v = v , v for all

v ∈V .
However, as V is always isomorph to at least a subset of V due to the
mapping constructed in Lemma 2.39, weak convergence in V always implies
weak convergence in V . If V is reflexive, both concepts are identical on V .
(iv) In a finite-dimensional normed space, strong and weak convergence coincide.
Lemma 2.41. Let V be a Banach space, W ⊂ V dense, Z ⊂ V dense, v ∈ V ,
v ∈ V , (vn )n∈N in V and (vn )n∈N in V .
n→∞
(a) vn −−−− v if and only if there is a C ∈ R+ such that vn V ≤ C for all
n ∈ N and limn→∞ vn , v = v, v for all v ∈ Z .
n→∞
(b) vn −−−− v if and only if there is a C ∈ R+ such that vn V ≤ C for all
n ∈ N and limn→∞ v, vn = v, v for all v ∈ W .
Lemma 2.42 (Mazur). Let V be a normed space and (vn )n∈N a sequence in V with
n→∞
N
vn −−−− v. For every ε ∈ R+ exists a linear combination u = ak vk , N ∈ N,
k=1

N
{ak }N
k=1 ⊂ R, ak ≥ 0 for all k, ak = 1 such that u − vV ≤ ε.
k=1
3. Function spaces
In order to deal with differential equations, we have to introduce some notation for
differentiation and integration. As it turns out, the classical strong concept of dif-
ferentiability is too restrictive. This leads to the definition of weak differentiability.
Different kinds of requirements on the differentiability of functions yield different
sets of functions which can be shown to be normed vector spaces. There are some
more remarkable properties of these spaces and the functions they contain as well
as interesting and useful relations between them.
Definition 3.1 ((Strong) Derivative). Let G be a bounded open subset of Rn , n ∈ N,
u : G → R, γ ∈ Nn0 , and k ∈ N0 . Let x be a point in Rn with coordinates xi ,
i ∈ N, i ≤ n. Throughout this section, these are cartesian coordinates.
The partial derivative Dγ u is defined by

∂ |γ| u
(Dγ u)(x) = ∂xγ11 . . . ∂xγnn u (x) = (x), x ∈ G , (45)
∂xγ11 . . . ∂xγnn
n
with |γ| = i=1 γi being the order of the derivative. The set of all derivatives of
u of order k at point x is denoted by (Dk u)(x) = {(Dγ u)(x) : |γ| = k}.
Definition 3.2 (Region). A subset G ⊂ Rn , n ∈ N is called a region, if it is open
and connected. Here, connected means that for any two points x, y ∈ G, there
exists a continuous function f : [0, 1] → G such that f (0) = x and f (1) = y.
Remark 3.3.
(i) If u is defined as a function on Rn , n ∈ N, such that u : Rn → R, the
restriction of u to G ⊂ Rn is denoted by u|G .
(ii) The gradient of a differentiable function u : Rn → R with respect to the
variable x is defined as the vector of all first derivatives and denoted by
∇x u = (∂x1 u, . . . , ∂xn u)T . (46)
We omit the index x if it is clear with respect to which variable the differen-
tiation has to be carried out.
(iii) The directional derivative of u with respect to a unit vector e is given by
(∇x u) · e. The directional derivative with respect to the outer unit normal
vector of a bounded region G is denoted by ∂n u.
(iv) The divergence ∇x · of a differentiable vector field h : G → Rn , n ∈ N, with
respect to the variable x is the scalar value
n
∂hk
∇x · h = . (47)
∂xk
k=1
(v) If ∇x f is differentiable, the Laplace operator Δx f with respect to the variable
x is given by
n
∂2f
Δx f = ∇x · (∇x f ) = . (48)
∂x2k
k=1
For vector-valued functions, the Laplace operator is defined component-wise.
As we now have introduced the notation for strong derivatives, we can define
function spaces of continuously differentiable functions (see, e.g., [1]).
Definition 3.4 (Spaces of Continuously Differentiable Functions). Let G ⊂ Rn ,
n ∈ N, be a region. For k ∈ N0 , we denote by Ck (G) the vector space of all
functions u : G → R which together with all their derivatives Dγ u of order |γ| ≤ k
are continuous on G. We write C(G) for C0 (G).9The space of infinitely continuously
∞
differentiable functions is given by C∞ (G) = k=0 Ck (G).
The subspace of functions in C (G) that have compact support in G is denoted
k
by Ckc (G). A function u has compact support in G if there is a compact set K ⊂ G

such that
supp(u) = {x ∈ G : u(x) = 0} ⊂ K . (49)
The spaces Ck (G) contain all functions u ∈ Ck (G) for which Dγ u is bounded and
uniformly continuous for all γ ∈ Nk0 with 0 ≤ |γ| ≤ k, i.e., it possesses a unique,
bounded, continuous extension to the closure G. These spaces are Banach spaces
when equipped with the norm
uCk (G) = max sup |(Dγ u)(x)| . (50)
0≤|γ|≤k x∈G
In some cases, functions are required to be more regular than just being
continuous, but requiring them to be continuously differentiable would be too
much. Thus, we introduce the spaces of Hölder-continuous functions ([1]).
Definition 3.5 (Hölder-Continuous Functions). Let G ⊂ Rn , n ∈ N, be a region,
γ ∈ Nk0 , and k ∈ N0 . The space Ck,s (G), 0 < s ≤ 1, is the subspace of functions
u ∈ Ck (G) whose derivatives Dγ u of order k satisfy
|(Dγ u)(x) − (Dγ u)(y)| ≤ C x − ys for all x, y ∈ G (51)
+
with a constant C ∈ R . We say u has Hölder-continuous derivatives of order k
with Hölder exponent s or, in the special case s = 1, u has Lipschitz-continuous
derivatives of order k. Ck,s (G) is a Banach space if equipped with the norm
|(Dγ u)(x) − (Dγ u)(y)|
uCk,s (G) = uCk (G) + max sup s . (52)
0≤|γ|≤k x,y∈G x − y
x =y
For r ≥ s > 0, the inclusion Ck,r (G) ⊂ Ck,s (G) is valid.

If a function u satisfies
|(Dγ u)(x) − (Dγ u)(y)|
lim sup s <∞ (53)
δ→0 x,y∈G x − y
0<x−y<δ
for each |γ| = k, the derivatives of order k of u are called uniformly Hölder-
continuous and u is an element of the space Ck,s
u (G).
When looking for functions whose values on the boundary of a region are
prescribed, it is often useful, if not even necessary, to restrict the kind of region
under consideration to answer questions of existence and uniqueness. The region
is required to have some kind of regularity. In order to give different kinds of
regularity properties, we need another definition ([1]).
Definition 3.6 (m-smooth Transformation). Let Φ be a one-to-one transformation
of a region G ⊂ Rn , n ∈ N, onto a region G ⊂ Rn with Ψ = Φ−1 . We call Φ
m-smooth if, writing y = Φ(x) and
y1 = φ1 (x1 , . . . , xn ), x1 = ψ1 (y1 , . . . , yn ),
y2 = φ2 (x1 , . . . , xn ), x2 = ψ2 (y1 , . . . , yn ),
.. ..
. .
yn = φn (x1 , . . . , xn ), xn = ψn (y1 , . . . , yn ), (54)
the functions φ1 , . . . , φn belong to Cm (G) and the functions ψ1 , . . . , ψn belong to

Cm (G).
The following summary of regularity conditions is taken from the book by
Adams [1] and is just slightly adapted. Here, a finite cone Cx with vertex x ∈ Rn
is defined as the set
Cx = Br1 (x) ∩ {x + C(y − x) : y ∈ Br2 (z)} (55)
with Br2 (z) being a ball around z ∈ Rn such that x ∈ / Br2 (z) and r1 , r2 , C ∈ R+ .
Two cones Cx and Cx are called congruent if there exists an isometry f : Rn → Rn
with f (Cx ) = f (Cx).
Given an index set J, a collection of sets {Uj : j ∈ J} is called a cover of a
set G if :
G⊂ Uj . (56)
j∈J
A cover is called open cover if all sets {Uj : j ∈ J} are open. An open cover is said
to be locally finite if any compact set in Rn can intersect at most finitely many
elements of {Uj : j ∈ J} ([1]). Locally finite collections of sets are countable. Thus,
we can assume J ⊂ N.
Definition 3.7 (Regularity of Domains). Let G ⊂ Rn , n ∈ N, be a region. G has
(i) the segment property if there exists a locally finite open cover {Uj : j ∈ J}
of ∂G and a corresponding set {yj : j ∈ J} of non-zero vectors such that if
x ∈ G ∩ Uj for some j, then x + εyj ∈ G for 0 < ε < 1;
(ii) the cone property if there exists a finite cone C such that each point x ∈ G
is the vertex of a finite cone Cx contained in G and congruent to C ;
(iii) the uniform cone property if there exists a locally finite open cover {Uj : j ∈ J}
of ∂G and a corresponding set {Cj : j ∈ J} of finite cones, each congruent to
some fixed finite cone C , such that
(a) for some finite M;∞ ∈ R+ , every Uj has a diameter less than M ,
(b) for some δ > 0, j=1 Uj ⊃ {x ∈ G : dist(x, ∂G) < δ},
;
(c) for every j, Qj = x∈G∩Uj (x + Cj ) ⊂ G,
(d) for some finite N ∈ N, every collection of N + 1 of the sets Qj has an
empty intersection;
(iv) the strong local Lipschitz property if there exist positive numbers δ and M ,
a locally finite open cover {Uj : j ∈ J} of ∂G, and for each Uj a real-valued
function fj of n − 1 real variables, such that
(a) for some finite N ∈ N, every collection of N + 1 of the sets Uj has an
empty intersection,
(b) for every pair of points x, y ∈ {z ∈ G : dist(z, ∂G) < δ} such that
x − y < δ, there exists j such that x, y ∈ {z ∈ Uj : dist(z, ∂Uj ) > δ},
(c) each function fj satisfies a Lipschitz condition with constant M ,
(d) for some cartesian coordinate system (xj,l )nl=1 in Uj , the set G ∩ Uj is
represented by the inequality xj,n < fj (xj,1 , . . . , xj,n−1 );
(v) the uniform Cm -regularity property if there exists a locally finite open cover
{Uj : j ∈ J} of ∂G and a corresponding set {Φj : j ∈ J} of m-smooth
one-to-one transformations ; with Φj taking Uj onto B1 (0) ⊂ Rn , such that
(a) for some δ > 0, ∞ j=1 j (B0.5 (0)) ⊃ {x ∈ G : dist(x, ∂G) < δ}, where
Ψ
−1
Ψ=Φ ,
(b) for some finite N ∈ N, every collection of N + 1 of the sets Uj has an
empty intersection,
(c) for each j, Φj (Uj ∩ G) = {y ∈ B1 (0) : yn > 0},
(d) if (φj,1 , . . . , φj,n ) and (ψj,1 , . . . , ψj,n ) denote the components of Φj and
Ψj , respectively, then there exists a finite M such that for all γ ∈ Nn0 ,
|γ| ≤ m, for every 1 ≤ i ≤ n, and for every j, we have |Dγ φj,i (x)| ≤ M ,
x ∈ Uj , and |Dγ ψj,i (y)| ≤ M , y ∈ B1 (0).
m≥1
For the different kinds of regularity, we have (v) =⇒ (iv) =⇒ (iii) =⇒ (i).
These regularity properties require G to lie on only one side of its boundary,
whereas the cone property does not impose this condition.
Remark 3.8.
(i) If G is bounded, the requirements for G being strong local Lipschitz reduce
to the condition that for each point x ∈ ∂G, there exists a neighborhood U
of x such that U ∩ ∂G is the graph of a Lipschitz-continuous function.
(ii) In some cases it is necessary to require that the parts of the one-to-one
transformation mentioned in the definition of the Cm -regularity property
have not only bounded derivatives, but Hölder-continuous ones. This yields
the Cm,s -regularity property.
As already mentioned, the above introduced definition of strong differentia-
bility with continuous or even Hölder-continuous derivatives is often too restrictive.
Therefore, we need some other, weaker definition of derivatives. To define these
weak derivatives, we need a definition of convergence in C∞ c (G) first (see, e.g., [1]).
Definition 3.9 (Convergence in C∞c (G)). Let G ⊂ R , n ∈ N, be a bounded region,

n
∞ ∞
(φl )l∈N in Cc (G), and φ0 ∈ Cc (G). The sequence (φl )l∈N is said to converge
towards φ0 in C∞c (G) for l → ∞ if there is a compact subset K ⊂ G such that
supp(φl ) ⊂ K for all l ∈ N , (57)
supp(φ0 ) ⊂ K , (58)
as well as all partial derivatives of φl of arbitrary order converge uniformly to those
of φ0 , i.e.,
l→∞
sup |(Dγ φl )(x) − Dγ φ0 (x)| −−−→ 0 for all γ ∈ Nn0 . (59)
x∈G
Remark 3.10.
(i) C∞c (G) is often denoted by D(G) and called the space of test functions, al-
though the latter identification is not unique. It is a topological vector space,
but not normable ([1, 17]).
(ii) A function φ defined on Rn is called finite if it vanishes outside a bounded

set. The space Cc∞ (Rn ) consist of all finite, arbitrary often continuously dif-
ferentiable functions on Rn .
The above definition allows us to define distributions (see [16]).
Definition 3.11 (Distribution). Let G ⊂ Rn , n ∈ N, be a bounded region. A dis-
tribution (or generalized function) is a linear functional f : C∞ c (G) → R which is
continuous in the following sense: If a sequence (φl )l∈N0 in C∞ c (G) converges for
l → ∞ towards φ ∈ C∞ c (G), then f (φl ) = f, φ l converges for l → ∞ towards f (φ).

The set of all distributions is denoted by (C∞ c (G)) .
The space of vector-valued distributions is defined accordingly.
Remark 3.12. The space of distributions is often denoted by D (G). If we consider

C∞ ∞
c (G) as a topological vector space, (Cc (G)) is its topological dual ([1, 17]).
There is another way to characterize functions that is useful to present here

in anticipation of a more general concept that we introduce later on. For this, we
have to explain our interpretation of the integral of a function.
Within this chapter, all integrals are understood in the sense of Lebesgue
integrals. In the following, we denote by Vn the Lebesgue measure on a given
measurable (e.g., open or closed) subset of Rn . If there is no confusion, we omit
the index n in dVn . A function f : Rn → R ∪ {−∞, ∞} is called Lebesgue-
measurable if the set {x : x ∈ Rn , f (x) > c} is measurable for arbitrary c ∈ R
(see [4]). We also have to define when a function is integrable. Following
[4], we
consider a measurable set A ∈ Rn and introduce step functions s = m k=1 ak χAk
with coefficients ak ∈ R and χAk the indicator function of Ak ⊂ A, given by

1, x ∈ Ak ,
χAk (x) = (60)
0, x ∈
/ Ak .
For step functions, we can define

m
s(x) dVn (x) = ak Vn (Ak ). (61)
A k=1
This allows us to define the integral for measurable positive functions f : Rn →

R+ ∪ {∞} as

f (x) dVn (x) = sup s(x) dVn (x) (62)
A A
whereas the supremum is taken over all step functions s which vanish outside A
and satisfy 0 ≤ s ≤ f inside A. For measurable functions f : Rn → R ∪ {−∞, ∞},
we introduce
f + (x) = max(f (x), 0), (63)
−
f (x) = max(−f (x), 0). (64)

If at least one of the integrals A f + (x) dVn (x), A f − (x) dVn (x) is finite, the
integral of f exists and is given by

f (x) dVn (x) = f + (x) dVn (x) − f − (x) dVn (x) (65)
A A A
with values in R ∪ {−∞, ∞}. A measurable function is called integrable if its

integral exists and takes values in R.
Now, we can define ([1])
Definition 3.13 (Locally Integrable Functions). Let G ⊂ Rn , n ∈ N, be a region. A
function u is called locally integrable on G if for every compact subset K ⊂ G we
have
|f (x)| dVn (x) < ∞. (66)
K
For every locally integrable function u, we can define a corresponding distri-

bution Tu ∈ (C∞c (G)) simply by

Tu (φ) = u(x)φ(x) dV (x), φ ∈ C∞ c (G) . (67)
G
Usually, notation is a little bit abused by also using u instead of Tu to denote the
corresponding distribution. Distributions that correspond in that way to a locally
integrable function are called regular.
There are many distributions for which no corresponding locally integrable
function can be found. The most prominent example is the evaluation of a function
φ at a certain point x, known as Dirac’s delta distribution. If 0 ∈ G, the evaluation
of a function φ ∈ C∞c (G) is given by
δ(φ) = φ(0). (68)

It is easy to prove that there is no locally integrable function for which

δ(x)φ(x) dV (x) = φ(0), φ ∈ C∞ (G) . (69)
G
However, δ satisfies Definition 3.11.

It is obvious how addition of two distributions and multiplication with a
constant should be defined on distributions. Distributions may even be multiplied

by smooth functions ([1]). For T ∈ (C∞ ∞
c (G)) and u ∈ C (G), the product uT ∈
∞
(Cc (G)) is defined by
(uT )(φ) = T (uφ), φ ∈ C∞
c (G) . (70)
The support of a distribution is defined as follows ([4]).

Definition 3.14 (Support of a Distribution). Suppose T ∈ (C∞ c (G)) for an open
bounded region G ⊂ R , n ∈ N. The support of T is defined as
n
! "
supp(T ) := x ∈ G : ∀δ > 0 : T |G∩Bδ (x) = 0 . (71)
The support of a distribution as defined in Definition 3.14 is closed (see [4])

and the concept of compact support is directly transfered to distributions.
Another operation which we need not only on distributions is convolution ([17]).
Definition 3.15 (Convolution). Let u be a function defined on Rn , n ∈ N, and
x, y ∈ Rn . We define
(x u)(y) = u(y − x) , (72)
ŭ(y) = u(−y) , (73)
(x ŭ)(y) = u(x − y) . (74)
Let v be another function on R . The convolution u ∗ v is defined as
n

(u ∗ v)(x) = u(y)v(x − y) dV (y) = u(y)(x v̆)(y) dV (y) (75)
Rn Rn
if the integral exists for V-almost all x ∈ Rn , i.e., the set N of all points x for which
the integral does not exist has vanishing measure, i.e., V (N ) = 0.
n
For a distribution u ∈ (C∞ ∞
c (R )) and φ ∈ Cc (R ), the function u ∗ φ is
n
defined by
(u ∗ φ)(x) := u(x φ̆). (76)

Theorem 3.16 (Properties of Convolutions). Let u ∈ (C∞ ∞
c (R )) , φ, ψ ∈ Cc (R ).
n n
Then the following holds:

(i) x (u ∗ φ) = (x u) ∗ φ = u ∗ (x φ) for all x ∈ Rn ;
(ii) u ∗ φ ∈ C∞ (Rn ) and u ∗ (φ ∗ ψ) = (u ∗ φ) ∗ ψ;
(iii) the operator L, defined by
Lφ = u ∗ φ, φ ∈ C∞ n
c (R ) , (77)
is a linear mapping of C∞ n
c (R )
∞
into C (R ) which satisfies x L = Lx ,
n
x ∈ Rn .
Until now, convolutions for distributions are only declared if a distribution
is convolved with an element of C∞ n
c (R ). The next lemma extends convolution to
∞ n
elements of C (R ) ([17]).

Lemma 3.17. Let u ∈ (C∞ ∞
c (R )) have compact support, φ ∈ C (R ), ψ ∈
n n
∞ ∞
Cc (R ). The convolution u ∗ φ ∈ C (R ) is well defined. Moreover,
n n
(i) x (u ∗ φ) = (x u) ∗ φ = u ∗ (x φ) for all x ∈ Rn ,

(ii) u ∗ ψ ∈ C∞ n
c (R ),
(iii) u ∗ (φ ∗ ψ) = (u ∗ φ) ∗ ψ = (u ∗ ψ) ∗ φ.
Convolutions may also be defined between distributions ([17]).
n
Lemma 3.18 (Convolutions between Distributions). Let u, v, w ∈ (C∞ c (R )) ,
n ∈ N.
(i) If at least one of u, v has compact support, the convolution u ∗ v is defined by
(u ∗ v) ∗ φ = u ∗ (v ∗ φ) for all φ ∈ C∞
c (R ) and u ∗ v = v ∗ u.
n
(ii) If at least one of the supports supp(u), supp(v) is compact, we have supp(u ∗
v) ⊂ supp(u) + supp(v).
(iii) If at least two of the supports supp(u), supp(v), supp(w) are compact, we
have (u ∗ v) ∗ w = u ∗ (v ∗ w).
We can now define weak derivatives by defining derivatives of distributions
([1, 17]).
Definition 3.19 (Weak Derivative). Let G be a bounded region in Rn , n ∈ N, and

u ∈ (C∞c (G)) . The weak derivative of u with respect to xi , i ∈ {1, . . . , n}, is
defined by
∂xi u, φ = − u, ∂xi φ , φ ∈ C∞c (G) . (78)
For a multi-index γ ∈ Nn0 , we have the generalization
Dγ u, φ = (−1)|γ| u, Dγ φ , φ ∈ C∞
c (G) . (79)
Remark 3.20. We use the same notation for weak derivatives and classical (strong)
partial derivatives (based on the limit of difference quotients). If a continuous
strong derivative exists, it coincides with the weak derivative as can be seen by
integration by parts ([17]).
Theorem 3.21 (Weak Derivatives and Convolution).
n n
(i) Suppose u ∈ (C∞ ∞ ∞
c (R )) and φ ∈ Cc (R ), n ∈ N, or u ∈ (Cc (R )) with
n
∞
compact support and φ ∈ C (R ), then D (u ∗ φ) = (D u) ∗ φ = u ∗ (Dγ φ)
n γ γ
for all γ ∈ Nn0 .

n
(ii) Suppose u ∈ (C∞ c (R )) and δ is the delta distribution, then D u = (D δ) ∗ u
γ γ
for all γ ∈ N0 . In particular, u = δ ∗ u.

n
n
(iii) Suppose u, v ∈ (C∞ c (R )) and at least one of them has compact support, then
D (u ∗ v) = (D u) ∗ v = u ∗ (Dγ v) for all γ ∈ Nn0 .
γ γ
A consequence of Theorem 3.21 is that it allows to give an informal integral

expression for Dirac’s delta distribution and its derivatives, with a slight abuse of
notation, by
Dγ δ(x − y)φ(y) dV (y) = (−1)|γ| (Dγ φ)(x) . (80)
Rn
It is easy to prove that the weak derivative of a distribution is also a distribu-
tion. Thus, for every distribution, there exist weak derivatives of arbitrary order.
Nevertheless, classes of distributions and their derivatives can be distinguished
if we introduce a new concept of regularity based on integrability. We begin by
defining Lebesgue spaces.
Definition 3.22 (Lebesgue Spaces). Let G be a bounded region in Rn , n ∈ N, and
p ∈ R+ . The Lebesgue space Lp (G) consists of all equivalence classes with respect
to the Lebesgue measure V of V-almost everywhere identical functions on G, whose
representatives u satisfy
|u(x)|p dV (x) < ∞ . (81)
G
Moreover, the space L∞ (G) contains all such equivalence classes whose rep-
resentatives are measurable, essentially bounded functions u : G → R, i.e.,
ess sup |u(x)| := inf sup |u(x)| < ∞ . (82)
x∈G V (N )=0 x∈G\N
Remark 3.23. It is convenient to identify a function with its respective equivalence

class. However, we remind the reader that it is, in general, not possible to evaluate
a function from Lp (G) at a point x ∈ G.
We summarize a few properties of the Lebesgue spaces ([1]).
Lemma 3.24 (Properties of Lebesgue Spaces). Let G be a bounded region in Rn ,

n ∈ N, and 1 ≤ p < ∞.
(i) The Lebesgue space Lp (G) is a Banach space with respect to the norm
p1
p
uLp (G) = |u(x)| dV (x) . (83)
G
L∞ (G) is a Banach space with respect to the norm

uL∞ (G) = ess sup |u(x)| . (84)
x∈G
(ii) The space L2 (G) is a Hilbert space if equipped with the scalar product

(u, v)L2 (G) = u(x)v(x) dV (x) . (85)
G
(iii) For arbitrary 1 ≤ p1 , p2 < ∞ with p1 ≥ p2 , we have Lp1 (G) ⊂ Lp2 (G) and
Lp1 (G) ⊂ L1loc (G), with L1loc (G) being the space of locally integrable functions.
(iv) Let 1 < p1 < ∞ and p2 such that p11 + p12 = 1. For u ∈ Lp1 (G), v ∈ Lp2 (G),
we have uv ∈ L1 (G) and
uvL1 (G) ≤ uLp1 (G) vLp2 (G) . (86)
This is known as Hölder’s inequality. It also holds for u ∈ L1 (G) and v ∈

L∞ (G). Then we have uv ∈ L1 (G).

(v) Let 1 < p1 < ∞ and p2 such that p11 + p12 = 1. Then the dual space (Lp1 (G))

of (Lp1 (G)) is isometrically isomorph to Lp2 (G). Moreover, L1 (G) is iso-

metrically isomorph to L∞ (G), but (L∞ (G)) is not isometrically isomorph to
1
L (G).
(vi) Cc (G) and C∞ c (G) are dense subspaces of L (G) for all 1 ≤ p ≤ ∞.
p
The definition of Lebesgue spaces allows us to evaluate the regularity of a

distribution by asking if it is also an element of some Lebesgue space. It comes nat-
urally to extend this to a distribution’s derivative. This gives rise to the definition
of Sobolev spaces ([1, 14, 21]).
Definition 3.25 (Sobolev Spaces). Let G be a bounded region in Rn , n ∈ N, and

1 ≤ p ≤ ∞. The Sobolev space Wk,p (G), k ∈ N0 , is defined as the subspace of
Lp (G) with
Wk,p (G) = {u ∈ Lp (G) : Dγ u ∈ Lp (G) for all γ ∈ Nn0 , |γ| ≤ k} . (87)
k,p
W (G) is a separable Banach space with respect to the norm
⎛ ⎞ p1

uWk,p (G) = ⎝ |Dγ u(x)|p dV (x)⎠ . (88)
|γ|≤k G
For p = 2, we denote Hk (G) = Wk,2 (G). These spaces are separable Hilbert spaces
with inner product

(u, v)Hk (G) = (Dγ u, Dγ v)L2 (G) . (89)
|γ|≤k
Moreover, the concept of Hölder-continuity can also be transferred to weak

derivatives in the following sense ([21]).
Definition 3.26 (Sobolev–Slobodeckij Spaces). Let G be a bounded region in Rn ,
n ∈ N, and 1 ≤ p ≤ ∞. The Sobolev–Slobodeckij space Wk,p (G) of fractional order
k = r + s with r ∈ N0 and 0 < s < 1 is defined as the subspace of Wr,p (G) with

Wk,p (G) = u ∈ Wr,p (G) : |Dγ u|s,p,G < ∞ for all γ ∈ Nn0 , |γ| = k , (90)
where the semi-norm |u|s,p,G is given by
p p1
|u(x) − u(y)|
|u|s,p,G = n+ps dV (x) dV (y) . (91)
G G x − y
Wk,p (G) is a Banach space if equipped with the norm

⎛ ⎞ p1

uWk,p (G) = ⎝uWr,p (G) + |Dγ u|s,p,G ⎠ .
p p
(92)
|γ|=k
As before, for p = 2, we denote Hk (G) = Wk,2 (G).

In what follows, we formulate results only for scalar-valued functions, al-
though similar results are valid for vector-valued functions.
The following relations of Sobolev spaces and spaces of continuous differen-
tiable functions is particularly useful when considering numerical solution schemes
(see [1]).
Lemma 3.27. Let G be a bounded region in Rn , n ∈ N, and 1 ≤ p < ∞. The
Sobolev space Wk,p (G), k ∈ N0 , is the completion of C∞ (G) with respect to the
norm ·Wk,p (G) .
If G has the segment property, then the set of restrictions to G of functions
in C∞ n
c (R ) is dense in W
k,p
(G).
Lemma 3.27 suggests the definition of another class of Sobolev spaces.

Definition 3.28. Let G be a bounded region in Rn , n ∈ N, and 1 ≤ p < ∞. The
Sobolev space W0k,p (G), k ∈ N0 , is defined as the completion of C∞
c (G) with respect
to the norm ·Wk,p (G) .
The above definition is motivated by the theory of partial differential equa-
tions. In general, if we are looking for a solution to a given partial differential
equation on G, it is necessary to prescribe boundary conditions to achieve unique-
ness. The simplest way to do this is by assuming that the solution shall vanish
at the boundary (so-called homogeneous Dirichlet boundary condition). We add
more details to these concepts later.
Remark 3.29. In general, W0k,p (G) = Wk,p (G). For conditions on G under which
those spaces are equal, the reader is referred to, e.g., [1].
With this definition, we can characterize Sobolev spaces with negative in-
dex ([1]).
Definition 3.30 (Sobolev Spaces with Negative Index). Let G be a bounded region
in Rn , n ∈ N, and 1 < p1 < ∞, p2 such that p11 + p12 = 1. The Sobolev space
W−k,p2 (G), k ∈ R+ , is defined as

W−k,p2 (G) = f ∈ (C∞ c (G)) : f W −k,p 2 (G) < ∞ , (93)
with
|f (u)|
f W−k,p2 (G) = sup . (94)
0 =u∈C∞
c (G)
uWk,p1 (G)
W−k,p2 (G) is the dual space of W0k,p1 (G).
An essential property of Sobolev spaces is the existence of the following em-
beddings ([1]).
Theorem 3.31 (Sobolev Embedding Theorem). Let G be a bounded region in Rn ,
n ∈ N, j, k ∈ N0 , 1 ≤ p1 , p2 < ∞.
(i) If G has the cone property, the following embeddings, marked by →, exist:
(a) Suppose kp1 < n and p1 ≤ p2 ≤ n−kp np1
1
. Then
Wj+k,p1 (G) → Wj,p2 (G) . (95)
(b) Suppose kp1 = n, p1 ≤ p2 < ∞. Then
Wj+k,p1 (G) → Wj,p2 (G) . (96)
Moreover, if p1 = 1 and, thus, k = n, this also holds for p2 = ∞.
(ii) If G has the strong local Lipschitz property, additional embeddings hold:
(a) Suppose kp1 > n > (k − 1)p1 . Then
n
Wj+k,p1 (G) → Cj,s (G), 0 < s < k − . (97)
p1
(b) Suppose n = (k − 1)p1 . Then
Wj+k,p1 (G) → Cj,s (G), 0<s<1. (98)

The last embedding holds for s = 1 if n = k − 1 and p1 = 1.
As mentioned above, the discussion of partial differential equations involves

the necessity to specify in some sense the values an element of a Sobolev space
takes on the boundary of a region G. This is not a trivial problem as elements
of Sobolev spaces are equivalence classes like the elements of Lebesgue spaces on
which the definition of Sobolev spaces is based. Lemma 3.27 allows us to find a
solution to this dilemma by introducing the trace operator ([13, 21]). However, as
this involves Sobolev spaces defined on the boundary ∂G of a bounded region, we
first have to define proper measures on the boundary, which are given by so-called
Hausdorff measures ([2]).
Definition 3.32 (Hausdorff Measure). Let S be a smooth surface in Rn , n ∈ N

given by
S = {(x, g(x)) ∈ Rn : x ∈ D} (99)
1
with D ⊂ R n−1
open and bounded and g ∈ C (D).
For any subset E ⊂ S for which the set E 4 := {x ∈ D : (x, g(x)) ∈ E} is
measurable with respect to the (n − 1)-dimensional Lebesgue measure Vn−1 , we
define the Hausdorff measure Sn−1 of E by

2
Sn−1 (E) := 1 + |∇g(x)| dVn−1 (x). (100)

E
The norms of all Sobolev spaces defined on the boundary ∂G of a region

G ⊂ Rn , n ∈ R are understood with respect to the corresponding Hausdorff
measure Sn−1 .
Remark 3.33. It is possible to extend the above definition in several ways. For
example, we can combine several surfaces which satisfy Definition 3.32 by gluing
them. This is necessary for closed surfaces, which do not directly satisfy Definition
3.32, e.g, the unit sphere which can be seen as the combination of two hemispheres.
For an even more general definition see, e.g., [19].
Theorem 3.34 (Trace Operator). Let G ⊂ Rn , n ∈ N, be a bounded region with the

uniform Cm,s -regularity property.
(i) Let 12 < k ≤ m + s, whereas for k ∈ N, k = m − 1, s = 1 is allowed. There is
1
a continuous linear operator T0 : Hk (G) → Hk− 2 (∂G), called trace operator,
such that
T0 u = u|∂G for all u ∈ Ck +1 (G) . (101)
If k ∈ N, we have u ∈ C (G).
k
(ii) Let k + 1 ≤ m + s, whereas for k ∈ N, m = k and s = 1 is allowed and

l ∈ N such that k − l > 12 . There is another continuous linear trace operator
<l 1
Tl : Hk (G) → Hk−i− 2 (∂G) such that
i=0

l
Tl u = u|∂G , ∂−n(x) u|∂G , . . . , ∂−n(x) u|∂G for all u ∈ Ck +l+1 (G) . (102)
If k ∈ N, we have u ∈ Ck+l (G). Here, ∂−n(x) u is the directional derivative

with respect to the inner normal on ∂G.
Remark 3.35. Definition 3.28 and Theorem 3.34 are compatible, as we have
T0 u = 0 on ∂G for all u ∈ W0k,p (G) , (103)
i.e., the elements of W0k,p (G) satisfy a homogeneous Dirichlet boundary condition.
For k ≥ 2, it is possible to show that
T0 D γ u = 0 on ∂G for all u ∈ W0k,p (G) with |γ| ≤ k − 1 . (104)
The definition of all the above spaces can be generalized to functions which
take values in a separable Banach space V ([16]). Let I ⊂ R be a bounded open
interval and V be a separable Banach space with (topological) dual V . We start
by defining C(I; V ) to be the space of all bounded continuous functions u : I → V ,
t → u(t) and equip it with the norm
uC(I;V ) = sup u(t)V . (105)
t∈I
Analogously, Ck (I; V ), k ∈ N, is defined as the space of all functions u : I → V

whose derivatives in I, i.e., with respect to t, up to order k are of class C(I; V ).
Moreover, the Lebesgue spaces Lp (G), 1 ≤ p < ∞, can be generalized to
Lp (I; V ) by substituting the absolute value in their definition and the definition
of their norms by the norm on V , thus yielding the norm
p1
p
uLp (I;V ) = u(t)V dV1 (t) . (106)
I
2
If V is a separable Hilbert space, L (I; V ) is also a Hilbert space. The space
L∞ (I; V ) consists of all measurable, essentially bounded functions u : I → V . It
is a Banach space with respect to the norm
uL∞ (I;V ) = ess sup u(t)V . (107)
t∈I
The spaces C(I; V ) and Lp (I; V ), 1 ≤ p ≤ ∞, are defined accordingly.

The generalization of Sobolev spaces to V -valued functions is straightforward.
We show how this is done for H1 (I; L2 (G)), where G ⊂ Rn , n ∈ N, is an open
bounded region. The corresponding norm is given by
12
2 2
uH1 (I;L2 (G)) = |u(x, t)| + |∂t u(x, t)| dV (x) dV1 (t) . (108)
I G
For separable Hilbert spaces V , the following embedding theorem can be

established:
Lemma 3.36 (Sobolev Lemma for Hilbert Space-valued Functions). Let I ⊂ R be
a bounded open interval and V be a separable Hilbert space. Then any function
u ∈ H1 (I; V ) has a continuous representative in C(I; V ).
Remark 3.37. See, e.g., [18] for a proof in a more general setting with Banach
spaces instead of Hilbert spaces.
A well-known tool in geodesy is Fourier transformation, which can also be
used to define Sobolev spaces. However, it is not possible to define Fourier trans-
formation for all distributions. Thus, we need the concept of rapidly decreasing
functions ([4]).
Definition 3.38 (Schwartz Space of Rapidly Decreasing Functions).
For φ ∈ C∞ (Rn ), k, l, n ∈ N0 , n = 0, α ∈ Nn0 , we define
α
pk,l (φ) := sup |x|k + 1 |D φ(x)| . (109)
x∈Rn
|α|≤l
φ is called rapidly decreasing if pk,l (φ) < ∞ for all k, l ∈ N0 . The space S (Rn ) of
all rapidly decreasing functions is called Schwartz space.
Convergence of a sequence (φj )j∈N in S (Rn ) is defined by
S
φj −→ φ ⇐⇒ lim pk,l (φj − φ) = 0 ∀ k, l ∈ N0 . (110)
j→∞
From the definition, it is clear that C∞

c (R ) ⊂ S (R ).
n n
We can now define Fourier transformation for rapidly decreasing functions ([4]):
Definition 3.39 (Fourier Transformation in Schwartz Space). Let φ ∈ S (Rn ). The
Fourier transform F φ of φ is defined by

n
F φ(ξ) := (2π)− 2 e−ix·ξ φ(x)dV (x), ξ ∈ Rn . (111)
Rn
We summarize some properties of the Fourier transformation F for rapidly

decreasing functions ([4]):
Theorem 3.40 (Properties of Fourier Transformation on S (Rn )). Let φ ∈ S (Rn ),
α ∈ Nn0 :
(i) xα φ, Dα φ, F φ, Dα F φ, F Dα φ ∈ S (Rn );
(ii) DαF φ =(−i)|α| F (xα φ), ξ α F φ = (−i)|α| F (Dα φ);
|·|2 |ξ|2
(iii) F e− 2 (ξ) = e− 2 ;
(iv) F : S (Rn ) → S (Rn ) is bijective, periodic with period 4, and bicontinuous
with inverse

−1 −n
F φ (x) = (2π) 2 eix·ξ φ(ξ)dV (ξ) . (112)
Rn
As mentioned earlier, Fourier transformation cannot be extended to the space

of distributions. However, we can consider the dual of the Schwartz space, S (Rn ).
n
Since C∞ ∞
c (R ) ⊂ S (R ), we have S (R ) ⊂ (Cc (R )) .
n n n
The elements of S (Rn ) are called tempered distributions. Any distribution

with compact support can be extended to be a tempered distribution. Regular
distributions are tempered distributions if their corresponding locally integrable
function satisfies u ∈ Lp (Rn ) for some 1 ≤ p ≤ ∞. A counter-example would be
2
the distribution corresponding to u : x → u(x) = e|x| .
We can now extend the Fourier transformation to S (Rn ) ([4]):
Definition 3.41 (Fourier Transformation of Tempered Distributions). Let T ∈

S (Rn ). The Fourier transform F T of T is defined by
F T (φ) := T (F φ) , φ ∈ S (Rn ) . (113)
Again, we summarize some properties of the Fourier transformation for tem-

pered distributions ([4]):
Theorem 3.42 (Properties of Fourier Transformation on S (Rn )).

(i) T ∈ S (Rn ) ⇒ F T ∈ S (Rn );
(ii) for regular distributions Tu ∈ S (Rn ), we have F Tu = TF u ;
(iii) F , F −1 : S (Rn ) → S (Rn ) are bijective, periodic with period 4, and bicon-
tinuous with
F F −1 T = F −1 F T = T . (114)
(iv) T ∈ S (Rn ), φ ∈ S (Rn ) ⇒ F −1 T (φ(ξ)) = F T (φ(−ξ)).
An important feature of the Fourier transformation is that it can be extended

to L2 (Rn ) ([4]). As an operator from L2 (Rn ) to L2 (Rn ), the Fourier transformation
is an isometric isomorphism ([4]):
Theorem 3.43 (Fourier Transformation on L2 (Rn )). The Fourier transformation

has a unique extension to L2 (Rn ) and for all φ, ψ ∈ L2 (Rn ), we have
(u, v)L2 (Rn ) = (F u, F v)L2 (Rn ) , (115)
i.e., F : L2 (Rn ) → L2 (Rn ) is an isometric isomorphism.
Fourier transformation can also be used to characterize Sobolev spaces and

is especially useful when discussing Sobolev spaces with non-integer index ([4])
Theorem 3.44 (Characterization of Hk (Rn ) and Hk (G) via Fourier Transforma-

tion). Let G ⊂ Rn , n ∈ N be a bounded region, k ∈ R+
0.
(i) For u ∈ Hk (Rn ) exist c1 , c2 ∈ R+ such that

2 2k 2 2
c1 uHk (Rn ) ≤ (1 + |ξ|) |F u| ≤ c2 uHk (Rn ) .
Rn
(ii) If there exists a continuous operator E : Hk (G) → Hk (Rn ) with Eu|G = u for
all u ∈ Hk (G), the space Hk (G) coincides with the restriction of functions in
Hk (Rn ) on G and the norm
- -
- k -
uHk (G) = inf -(1 + |·|) F u 4- (116)
k n 2 n L (R )
∈H (R )
u
|G
u
is equivalent to the norm in Hk (G).

Remark 3.45.
(a) For a domain G to satisfy an extension condition as in Theorem 3.44, it is
sufficient that G is bounded and has the uniform cone property. This result
is due to Calderón and Zygmund (see, e.g., [21] and the references therein).
(b) It is also possible to define spaces of Sobolev-type via Fourier transformation
based on other Lebesgue spaces Lp (Rn ). Those are called Bessel potential
spaces. However, they are in general not identical to the Sobolev or Sobolev–
Slobodeckij spaces as given in this chapter. For details, the reader is referred
to [1] and the references therein.
4. Differential equations
Assume we have an open bounded region G ⊂ Rn , n ∈ N, n > 1, and a map
k k−1
F : Rn × R n × · · · × Rn × R × G → R, k∈N. (117)
Then for u ∈ C (G)
k
k
F D u (x), . . . , u(x), x = 0 for all x ∈ G (118)
is a partial differential equation (PDE) of order k if at least one derivative of order
k is actually a part of the equation and no derivative of higher order than k is
present. This can be done analogously for systems of differential equations. We
only deal with linear PDEs here that can be written as

aγ (x) (Dγ u) (x) = f (x), γ ∈ Nn0 for all x ∈ G , (119)
|γ|≤k
with given coefficient functions aγ and right-hand side f . If f = 0, the PDE is

called homogeneous.
Many important systems of linear PDEs consist of second-order PDEs. There
are three main classes of these PDEs. We start by defining an elliptic differential
operator (cf., e.g., [6]).
Definition 4.1 (Elliptic PDEs). Let G ⊂ Rn , n ∈ N, u ∈ Ck (G). Let L be a linear
differential operator of second order such that
n
n
u → Lu = aij ∂xi ∂xj u + bi ∂xi u + cu (120)
i,j=1 i=1
with given functions aij : G → R, bi : G → R, and c : G → R sufficiently smooth.

L is called uniformly elliptic if there exists a constant C ∈ R+ such that

n
2
aij (x)yi yj ≥ C y (121)
i,j=1
for almost every x ∈ G and all y ∈ Rn .

With this definition, we can now define the two other main classes (cf., e.g.,
[6, 9]). For both, there is one distinguished variable, denoted by t rather than as
a component of a vector x, which is usually the time being distinct from spatial
variables summarized in x.
Definition 4.2 (Parabolic PDE). Let G ⊂ Rn , n ∈ N, (0, tend ) ⊂ R, tend ∈ R+ ,
u ∈ Ck (G × (0, tend)). Let L be a linear differential operator of second order such
that
n
n
u → Lu = aij ∂xi ∂xj u + bi ∂xi u + cu − ∂t u (122)
i,j=1 i=1
with given functions aij : G × (0, tend ) → R, bi : G × (0, tend) → R, and c :
G × (0, tend) → R sufficiently smooth.
L is called uniformly parabolic if there exist constants C0 , C1 > 0 such that
n
C0 y2 ≤ aij (x, t)yi yj ≤ C1 y2 (123)
i,j=1
for all (x, t) ∈ G × (0, tend) and all y ∈ Rn .

Definition 4.3 (Hyperbolic PDE). Let G ⊂ Rn , n ∈ N, (0, tend) ⊂ R, tend ∈ R+ ,
u ∈ Ck (G × (0, tend)). Let L be a linear differential operator of second order such
that
n
n
u → Lu = aij ∂xi ∂xj u + bi ∂xi u + cu − ∂t2 u (124)
i,j=1 i=1
with given functions aij : G × (0, tend ) → R, bi : G × (0, tend) → R, and c :
G × (0, tend) → R sufficiently smooth.
L is called uniformly hyperbolic if there exists a constant C ∈ R+ such that
n
2
aij (x, t)yi yj ≥ C y (125)
i,j=1
for all (x, t) ∈ G × (0, tend) and all y ∈ Rn .

Remark 4.4.
(i) It is possible that the character of a differential operator changes with x, e.g.,
when there is a function of x as coefficient of the time derivative term. Such
equations can be locally elliptic, parabolic, or hyperbolic instead of uniformly,
i.e., they are of one of these types on certain subregions of G.
(ii) Not all linear second-order PDEs are of one of the above classes for n > 2.
(iii) Conditions on the actual smoothness of coefficients depends on the specific
problem under considerations. For example, the so-called Cauchy–Kovalev-
skaya Theorem (see, e.g., [15]) needs analytic coefficients to guarantee the
existence and uniqueness of analytic solutions. If we want the operator L
in (119) to be continuous
from H l+k (G) to H l (G), this can be achieved by
requiring aγ ∈ C l G (see, e.g., [21]).
As already pointed out, the formulations of differential equations as given
above with strong partial derivatives is often not suited to find answers to the
questions of solvability, uniqueness of solutions, or their regularity. Instead, we
would like to have a formulation based on weak derivatives.
Let us assume that a linear second-order differential equation is given in its
strong form by
Lu(x) = f (x) for all x ∈ G . (126)
For simplicity, we equip this equation with the homogeneous Dirichlet boundary
condition
u(x) = 0 for all x ∈ ∂G . (127)
In the context of differential equations, we often use the abbreviation Γ = ∂G.
A classical strong solution of this PDE has to be in C2 (G) which is a rather
restrictive requirement. We can relax this requirement in two points. First, we can
change over to a weakly differentiable solution. For this purpose, suppose v is an
arbitrary function belonging to C∞ c (G), multiply the differential equation by v and
integrate over G. We obtain
(Lu, v)L2 (G) = (f, v)L2 (G) for all v ∈ C∞
c (G) . (128)
Moreover, we can relax the requirements on differentiability of u by performing an
integration by parts on the left-hand side, which gives us a bilinear form a(u, v).
Thus, a useful assumption on u is u ∈ H10 (G). Additionally, as C∞ c (G) is a dense
subspace of H10 (G), we can extend the space of functions with which we multiply
to H10 (G). This yields
a(u, v) = f (v) for all v ∈ H10 (G) . (129)
Here, we interpreted the right-hand side as a linear functional on H10 (G).
It is easy
to see that every solution of the strong formulation is also a solution of the weak
formulation. However, the opposite may not be true.
For other kinds of boundary conditions, the above procedure is changed in
two points. On the one hand, if the values of u on the boundary are given and
different from zero, u has to belong to another subspace of H1 (G). On the other
hand, if normal derivatives of u are specified in a Neumann boundary condition,
integration by parts yields some integrals over (parts of) the boundary Γ of G.
These are usually incorporated into the linear form f on the right-hand side.
Other modifications may be necessary for other boundary conditions.
To answer the question whether a unique solution to a PDE in its weak form
(129) exists, we can use Corollary 2.36 or directly the Theorems by Riesz (2.31) or
Lax–Milgram (2.35) for elliptic PDEs. There are also general results for parabolic
or hyperbolic systems, but those are out of the scope of this chapter.
5. Reproducing kernel functions

Let S ⊂ Kn , n ∈ N, x, y ∈ S and (V, (·, ·)V ) a Hilbert space of functions defined
on S or a superset of S. Following [3], we define
Definition 5.1 (Reproducing Kernel). A function K : S × S → K, is called a
reproducing kernel if
(a) for each fixed y ∈ S, we have K(·, y) ∈ V and
(b) for every f ∈ V and for every point y ∈ S, the reproducing property
f (y) = (f, K(·, y))V (130)
holds.
The question whether a given Hilbert space V has a reproducing kernel is
answered by Davis [3].
Theorem 5.2 (Aronszajn, Existence of a Reproducing Kernel). A necessary and
sufficient condition that V has a reproducing kernel function is that for each fixed
y ∈ S, the linear functional given by δ(y), i.e.,
V f → δ(y)f = f (y) ∈ K (131)
is bounded.
Reproducing kernel functions have the following properties ([3]):
Theorem 5.3 (Uniqueness and Symmetry of a Reproducing Kernel). If a Hilbert
space V possesses a reproducing kernel K, the kernel is unique and for all x, y ∈ S,
we have
K(x, y) = K(y, x), (132)
whereas the overline marks complex conjugation.
Theorem 5.4. Let V have a reproducing kernel and let limn→∞ f − fn V = 0.
Then, for each x ∈ S, we have
lim fn (x) = f (x). (133)
n→∞
The convergence holds uniformly in every subset U of S for which K : S → K,
y → K(y, y) is bounded.
As a consequence, we obtain the following corollary ([3]):
Corollary 5.5. If V has a reproducing kernel, then the expansion of a function with
respect to an orthonormal basis converges pointwise to the function and uniformly
in every subset U of S for which K : S → K, y → K(y, y) is bounded.
Reproducing kernels can also be used to find representatives for linear func-
tionals on V . Let L be a linear functional on the reproducing kernel Hilbert space
V . According to the Riesz representation Theorem 2.31, L has a representative h.
As h is an element of V , we have
h(y) = (h, K(·, y))V = (K(·, y), h)V . (134)
On the other hand, K(·, y) is an element of V and we can apply the functional L
to K(·, y) to get
LK(·, y) = (K(·, y), h)V . (135)
By comparing Eqs. (134) and (135), we get
Theorem 5.6. Let V have a reproducing kernel K(·, ·), and let L be a bounded
linear functional defined on V . Then the function
h(y) = LK(·, y) (136)
is in V and for all f ∈ V ,
L(f ) = (f, h)V . (137)
Moreover, we have
L2V = Ly Lx K(x, y). (138)
Here, the indices x and y on L are used to clarify that we first regard K(x, y) as
a function of x with parameter y and apply L with respect to x and then regard
Lx K(x, y) as a function of y with parameter x and apply L with respect to y.
An important example for a reproducing kernel Hilbert space is given in the
following definition.
Definition 5.7. Let G ⊂ C be a bounded region. The space of all functions f ∈
L2 (G) which are analytic is denoted by L2A (G).
Theorem 5.8. Let {hn : n ∈ N} be an orthonormal basis in L2A (G).
∞
Then K(x, y) = n=1 hn (x)hn (y) is a (and hence the) reproducing kernel for
L2A (G); that is, for all f ∈ L2A (G) we have

f (y) = (f (·), K(·, y))V = f (x)K(x, y) dx. (139)
G
K(·, ·) is known as the Bergman kernel.

Combining Theorems 5.6 and 5.8 yields
Corollary 5.9. Let {hn : n ∈ N} be an orthonormal basis in L2A (G) and L a bounded
linear functional defined on L2A (G). Then
∞

f (y) = hn (y)L hn (x) (140)
n=1
is the representative of L, and

∞

L2 = |L hn |2 = Ly Lx K(x, y). (141)
n=1
The above theorems show how an orthonormal basis can be used to find
a reproducing kernel. However, we can also use a reproducing kernel to find an
orthonormal basis using Theorem 5.6 ([3]).
Theorem 5.10. Let V be a Hilbert space of functions that has the reproducing kernel
K(·, ·). If (Ln )n∈N is a sequence of bounded linear functionals on V such that from
Ln f = 0, for all n ∈ N, follows that f = 0, then the functions
hn (y) = Ln K(·, y), n∈N (142)
form a basis for V .
Moreover, Theorem 5.8 can be used the other way around to examine whether
an orthonormal system is complete ([3]).
Theorem 5.11. Let K(·, ·) be the reproducing kernel of L2A (G) and {hn : n ∈ N} be
an orthonormal system. This system is complete if and only if
∞

K(x, x) = |hn (x)|2 for all x ∈ G. (143)
n=1
Another example with special relevance in geodesy is the space of spherical

harmonics of a fixed degree.
Let Ω = ∂B1 (0) ⊂ R3 be the unit sphere in R3 . Let Hn : R3 → R be a
homogeneous harmonic polynomial of degree n ∈ N, i.e., Hn (αx) = α3 Hn (x) for
all α ∈ R for all x ∈ R3 and Δx Hn vanishes on R3 . Then the restriction Yn = Hn |Ω
is a spherical harmonic of degree n and the following theorem holds ([8]):
Theorem 5.12 (Reproducing Kernel for Spherical Harmonics).
The space Harmn (Ω) of all spherical harmonics of degree n ∈ N is a reproducing
kernel Hilbert space with the kernel given by
2n + 1
K(ζ, η) = Pn (ζ · η), ζ, η ∈ Ω (144)
4π
with Pn : [−1, +1] → R the Legendre polynomial of degree n.
As the space of spherical harmonics in the above example is only finite di-
mensional, let us consider another example with an infinite-dimensional space.
Consider the operator of the Newton potential
A : L2 (B1 (0)) → R(A) (145)

f (y)
f → Af = dV (y) . (146)
Ω 4π |· − y| R3 \B1 (0)
The null space of A consists of all functions in L2 (B1 (0)) that are orthogonal to
harmonic functions in B1 (0). We denote the space of functions which are harmonic
in B1 (0) by Pot(B1 (0)). It can be shown (see, e.g., [7]) that A|Pot(B1 (0)) is a linear
bijective operator. We can define an inner product on R(A) by
(h1 , h2 )R(A) := (A−1 h1 , A−1 h2 )L2 (B1 (0)) . (147)
Moreover, R(A) equipped with this inner product is a Hilbert space which satisfies
Theorem 5.2.
To find the reproducing kernel, we observe that for any x ∈ R3 \ B1 (0),

f ∈ Pot(B1 (0)) and h = Af , we have
) *
1 1
h(x) = f, = Af, A
4π |x − ·| L2 (B1 (0)) 4π |x − ·| R(A)
) *
1
= h, A . (148)
4π |x − ·| R(A)
Thus, the reproducing kernel is given by
) *
1 1 1
KR(A) (x, y) = A = 2
dV (z). (149)
4π |x − ·| (4π) Ω |x − z| |y − z|
6. Summary
As we have seen in this chapter, functional analysis provides many useful concepts
to tackle geodetic problems. The structures and results which we discussed here are
the very foundations for the solution of such problems as the (stochastic) oblique
derivative problem ([11]), the inverse problem of determining the density distri-
bution in the Earth’s crust from gravity measurements ([7]) or the very successful
and still expanding applications of wavelets to deal with local data concentration
and data refinement ([7, 8]). They are also used in [12] to derive limit formulae
and jump relations of potential theory in Sobolev spaces.
As can be deduced from their success so far, the application of functional
analytic results and concepts has a key role in facing future challenges in geodesy.
Among those are the incorporation of heterogeneous data, i.e., measurements of
different quantities related to the gravity potential of the Earth to determine said
potential. Moreover, the unequal distribution of those measurements, which may
provide a high data density in some regions but show gaps in others, calls for
locally oriented methods as opposed to the classical, globally oriented methods.
References
[1] R. Adams. Sobolev Spaces. Academic Press, New York, 1975.
[2] H.W. Alt. Lineare Funktionalanalysis. Springer, Berlin, Heidelberg, 2012.
[3] P. Davis. Interpolation and Approximation. Blaisdell Publishing Company, Waltham,
MA, 1963.
[4] M. Dobrowolski. Angewandte Funktionalanalysis. Springer, Berlin, Heidelberg, 2010.
[5] P. Enflo. A counterexample to the approximation property in Banach spaces. Acta
Math., 130:309–317, 1973.
[6] L. Evans. Partial Differential Equations, volume 19 of Graduate Studies in Mathe-
matics. American Mathematical Society, Providence, 1998.
[7] W. Freeden and C. Gerhards. Geomathematically Oriented Potential Theory. Chap-
man & Hall / CRC Press, Boca Raton, 2012.
[8] W. Freeden and V. Michel. Multiscale Potential Theory with Applications to Geo-
science. Applied and Numerical Harmonic Analysis. Birkhäuser, Boston, 2004.
[9] A. Friedman. Partial Differential Equations of Parabolic Type. Prentice-Hall, Inc.,
Englewood Cliffs, 1964.
[10] G. Grosche, V. Ziegler, D. Ziegler, and E. Zeidler, editors. Teubner Taschenbuch der
Mathematik – Teil II. B.G. Teubner, Stuttgart, Leipzig, 1995.
[11] M. Grothaus and T. Raskop. The Outer Oblique Boundary Problem of Potential
Theory. Numer Funct Anal Optim, 30:711–750, 2009.
[12] M. Grothaus and T. Raskop. Limit Formulae and Jump Relations of Potential The-
ory in Sobolev Spaces. Int. J. Geomath., 1:51–100, 2010.
[13] J.-L. Lions. Equations Différentielles Operationelles et Problèmes aux Limites, vol-
ume 111 of Die Grundlehren der mathematischen Wissenschaften in Einzeldarstel-
lungen. Springer, Berlin, 1961.
[14] W. McLean. Strongly Elliptic Systems and Boundary Integral Equations. Cambridge
University Press, Cambridge, 2000.
[15] V.P. Mikhailov. Partial Differential Equations. MIR Publishers, Moscow, 1978.
[16] M. Renardy and R.C. Rogers. An Introduction to Partial Differential Equations.
Texts in Applied Mathematics. Springer, New York, 1993.
[17] W. Rudin. Functional Analysis. McGraw-Hill, New York, 1973.
[18] R.E. Showalter. Monotone Operators in Banach Space and Nonlinear Partial Dif-
ferential Equations, volume 49 of Mathematical Surveys and Monographs. American
Mathematical Society, Providence, 1996.
[19] L. Simon. Lectures on Geometric Measure Theory, volume 3 of Proceedings of the
Centre for Mathematical Analysis. Centre for Mathematical Analysis, The Australian
National University, Canberra, 1984.
[20] G.G. Stokes. On the variation of gravity on the surface of the Earth. Trans Camb
Phil Soc, 8:672–695, 1849.
[21] J. Wloka. Partial Differential Equations. Cambridge University Press, Cambridge,
1992.
Matthias Augustin
Mathematical Image Analysis Group
Fakultät 6
Saarland University
D-66123 Saarbrücken, Germany
Sarah Eberle
Numerical Analysis Group
Mathematical Institute
University of Tübingen
D-72076 Tübingen, Germany
Martin Grothaus
Functional Analysis and Stochastic Analysis Group
Fachbereich Mathematik

Ill-Posed Problems: Operator Methodologies

of Resolution and Regularization
Willi Freeden and M. Zuhair Nashed
Abstract. A general framework of regularization and approximation methods

for ill-posed problems is developed. Three levels in the resolution processes
are distinguished and emphasized: philosophy of resolution, regularization-
approximation schema, and regularization algorithms. Dilemmas and method-
ologies of resolution of ill-posed problems and their numerical implementations
are examined with particular reference to the problem of finding numerically
minimum weighted-norm least squares solutions of first kind integral equa-
tions (and more generally of linear operator equations with non-closed range).
An emphasis is placed on the role of constraints, function space methods, the
role of generalized inverses, and reproducing kernels in the regularization and
stable computational resolution of these problems. The thrust of the contri-
bution is devoted to the interdisciplinary character of operator-theoretic and
regularization methods for ill-posed problems, in particular in mathematical
geoscience.
Keywords. Ill-posed problems, inverse problems, regularization.
1. Introduction
This contribution on operator-theoretic approaches to ill-posed problems (IPP’s)
develops a general framework for regularization and approximation methods for
ill-posed problems. Most inverse problems are ill-posed. For example, gravimet-
ric and downward continuation problems of geodesy are ill-posed. Three levels
in the resolution processes are distinguished and analyzed in this expository re-
search paper: philosophy of resolution, regularization-approximation schema, and
regularization algorithms. Our essential objective is to provide an outlook within
Ill-posed and inverse problems represent classical topics in the research of mathematical geodesy.
Hence, the necessity of such a contribution comes naturally. This is the reason why the editors and
the publisher have decided to include this chapter here despite the fact that its content has been
extracted from W. Freeden, M.Z. Nashed, Operator-Theoretic and Regularization Approaches
to Ill-Posed Problems, GEM Int. J. Geomath., Springer, 2017 (https://doi.org/10.1007/s13137-
017-0100-0).
202 W. Freeden and M.Z. Nashed
which discretization and other approximation methods can be better motivated,

interpreted, and understood. Our development will be fairly general in scope and
theory, and it is applicable to a wide range of ill-posed problems. Each method for
resolution, whether regularized in the classical sense or non-regularized, involves
a critical “parameter” whose “optimal value” is crucial to the amenability and
numerical implementation of the method. For example, in Tikhonov-type regu-
larization it is the regularization parameter, or more generally the choice of the
regularization operator. In projection and other discrete methods, it is the opti-
mal dimension of the approximating subspaces. In discretization methods it is the
choice of the mesh size beyond which a further refinement will lead to instability.
In multiscale methods it is the scale parameter to determine the scale space in
which multiresolution is realizable relative to the data width. In iterative methods
it is the level at which one should terminate the iteration (i.e., it is the stopping
rule for the iterates). In filteration-truncation methods it is the number of terms to
be included, etc. This paper delineates unifying principles that quantify the choice
of the parameter, the type of estimates, and a priori information that are needed
to arrive at an “optimal” value for this parameter.
Methodologies and dilemmas of the resolution of ill-posed problems and their
numerical implementations are examined with particular reference to the problem
of finding minimum weighted-norm least squares solutions of linear operator equa-
tions with non-closed range. A common problem in all these methods is delineated:
Each method reduces the problem of resolution to a “non-standard” minimiza-
tion problem involving an unknown critical “parameter” whose “optimal” value
is crucial to the numerical realization and amenability of the method. The “non-
standardness” results from the fact that one does not have explicitly, or a priori,
the function to be minimized: It has to built up using additional information,
convergence rate estimates, noise characteristics and robustness conditions, etc.
Several results are discussed that represent and complement advances in regular-
ization of inverse and ill-posed problems. An emphasis is placed on the role of
constraints, function space methods, the role of generalized inverses, and repro-
ducing kernels in the regularization and stable computational resolution of these
problems. The thrust of the work is devoted to the interdisciplinary character of
operator-theoretic methods for ill-posed problems. It is hoped that the viewpoints
and approaches developed in this work for geodetically relevant obligations would
be found useful in connection with other ill-posed problems in diverse areas of
application. In fact, our purpose is to provide an outlook within which technical
results can be better motivated and understood. Within this framework, criteria
can be given relative to which the scope and limitations of the various methods
can be assessed. This is important both in theory and practice since there is no
cure-all method for ill-posed problems; therefore it is imperative to be able to
clarify why a certain method works in some context as well as when not to use
that method. The work discusses at length the intuitive principles that underlie
the various methods and establishes some results within this framework, thereby
omitting technicalities of the proofs.
Ill-Posed Problems: Operator Methodologies of Resolution 203
Within the framework of the paper we are mainly interested in pointing

out those aspects that are related to generalized inverses in (reproducing kernel)
Hilbert spaces and those which are not. Only standard notation from functional
analysis is used; reference may be made to any introductory book on functional
analysis (e.g., [140, 143]). In a number of ill-posed problems (for example, in math-
ematical geodesy, the gravimetry problem, the gradiometry problem, etc.), the
operator A is an integral operator, and the problem Ax = y is essentially one of
“solving” a Fredholm integral equation of the first kind. If y belongs to the range
of the operator A, we may ask for an exact solution, while in the case in which y
fails to belong to the range of A (the case more typically met in applied problems
of the type described above), we must confront the fundamental issue of deciding
what should be meant by a “solution” and, only then, seek appropriate techniques
for the resolution of the problem.
During the past three decades a substantial amount of machinery from func-
tional analysis, theory of special functions, optimization as well as approximation
theory and numerical analysis has been brought to bear on the resolution and
understanding of IPPs, and the interdisciplinary character of many inverse and ill-
posed problems has emerged very clearly. The interdisciplinary character of IPP’s
in Applied Sciences is also stressed in many survey papers, which also give excel-
lent account of the state of the art for various problems in practice and contain
extensive bibliographies. Three problems are essentially treated in the literature:
(1) the identification problem,
(2) the synthesis (or controllability) problem, and
(3) best approximate synthesis.
These problems are all subsumed in the general problem of studying an operator
equation of the form Ax = y, where A usually is assumed to be an operator with
non-continuous inverse.
The vivid research activity in the field of ill-posed and inverse problems has
led to a vast literature on inverse and ill-posed problems. We list only a selection
of contributions, where the reader is also referred to the literature therein.
textbooks: [14, 22, 60, 63, 105, 106, 124, 132, 137, 142, 148, 169, 170, 173, 177, 178,
248, 250, 254],
conference reports, handbooks: [7, 12, 25, 65, 82, 83, 107, 128, 141, 173, 177, 182,
183, 185, 187–189, 192, 223, 233].
In addition to the books we mention the journals:
Inverse Problems,
Inverse Problems in Science and Engineering,
Inverse Problems and Imaging,
Journal on Inverse and Ill-Posed Problems,
Journal of Mathematical Imaging and Vision,
Mathematical Inverse Problems.
Many authors contributed results in different areas of ill-posed and inverse

problems (note that he following list is rather incomplete, for more details the
reader is referred to the references in the aforementioned textbooks or the below
listed journal and handbook publications):
generalized inverse and least squares problems: [11, 39, 47, 50, 51, 57, 58, 60, 66,
95, 99, 114, 125, 151, 153, 157, 160, 164–166, 173, 175, 178, 182–186, 194,
197, 208, 209, 219, 220, 226, 240, 261],
truncated singular value, Tikhonov regularization, and discrepancy principles: [24–
28, 38, 52, 53, 60, 64, 68, 69, 92, 93, 105, 107, 118, 126, 146, 147, 152, 154,
155, 168, 176, 178, 179, 181, 184, 201, 202, 210, 216, 218, 225, 235, 243, 247,
249, 250, 256],
inversion in (reproducing kernel) Hilbert spaces: [67, 122, 144, 184, 195, 196, 232,
236, 258],
projection methods, moment problems: [6, 7, 34, 35, 52, 53, 61, 137, 180, 184, 199,
241, 262],
iterative methods, finite element methods, other computational methods: [3, 6, 11,
20, 32, 33, 38, 40, 54, 80, 92, 102, 108, 112, 113, 115–117, 133, 134, 138, 172,
184, 200, 204, 214, 242, 246, 251, 253, 255, 263, 266–268],
mollifier methods: [63, 148, 149],
variational methods, implicit function theorems, ill-posed problems in differential
equations: [1, 5, 9, 10, 14, 36, 42, 43, 48, 50, 56, 61, 62, 145, 189–191, 206,
215, 224, 230, 252, 257],
multiscale methods: [41, 44, 45, 74–77, 81, 91, 94, 141, 150, 152, 156, 161–163].
The GEM–International Journal on Geomathematics is a forum in which
geoscientifically relevant ill-posed problems gain appropriate recognition. Many
further references will be given in due course.
It is also hoped that the viewpoints and approaches developed in this paper
would be found useful in connection with other inverse problems of various (not
necessarily geoscientific) research areas.
2. Solvability of ill-posed operator equations

Schematically, a direct (forward) problem can be formulated as follows:
object −→ data information of the object.
The inverse problem is considered the “inverse” to the forward problem which
relates the object (sub)information to the object:
data information of the object −→ object.
An object may be understood to be the systematic relationship of all data
subinformation, object parameters, and other auxiliary information. It may be
linear or non-linear, deterministic or random, etc.
In standard functional analytic nomenclature (see, e.g., [120, 135, 229, 245,
265]) we are usually confronted with the following operator equation: Given spaces
X, Y equipped with the settings of norm and inner product, respectively. Consider
a mapping A from X to Y , i.e., (A; X, Y ) with
A : X → Y. (2.1)
The Direct Problem (DP) is as follows: Given x ∈ X, find y = Ax ∈ Y . The
Inverse Problem (IP) is as follows: Given an observed output y, find an input x
that produces it, i.e., Ax = y ∈ Y , or given a desired output z, find an input x
that produces an output y = Ax ∈ Y that is as “close” to z as possible.
A Well-Posed (Properly-Posed) Problem in the sense of Hadamard is as fol-
lows: For each “data” y ∈ Y , the operator equation X x → Ax = y ∈ Y has one
and only one solution, and the solution depends continuously on y. In more detail,
a mathematical problem is well posed in the sense of Hadamard (cf. [109, 110]), if
it satisfies the following properties:
(H1) Existence: For all (suitable) data, there exists a solution of the problem (in
an appropriate sense).
(H2) Uniqueness: For all (suitable) data, the solution is unique.
(H3) Stability: The solution depends continuously on the data.
According to this definition, a problem is ill posed or improperly posed in the sense
of Hadamard if one of these three conditions is violated.
As already mentioned, ill-posed problems arise in many branches of science,
engineering, and mathematics, including computer vision, natural language pro-
cessing, machine learning, statistics, statistical inference, medical imaging, remote
sensing, non-destructive testing, astronomy, geodesy and geophysics, exploration
and prospection, and many other fields. It should be noted that J. Hardamard
(1865–1963) dismissed ill-posed problems as irrelevant to physics or real world ap-
plications, but he was proven wrong four decades after his declaration. In fact, it
turned out that Hadamard’s classification had a tremendous influence on the de-
velopment of mathematics. Some years ago, starting from Hadamard’s properties
a more relevant understanding of ill-posedness was provided by a more detailed
functional analytical background (cf. [185]) that will be explained later on.
2.1. Finite-dimensional matrix equations and generalized inverse

Since any numerical approximation procedure usually leads to finite-dimensional
problems involving a singular functional analytic context, we begin with the re-
capitulation of finite systems of linear equations (see, e.g., [23, 99, 178, 260] and
the list of references therein for more details). After having treated the finite-
dimensional situation, we turn to the analogous theory in infinite-dimensional
operator framework.
Spectral matrix representation. We start with a linear matrix equation of the form
Ax = y, (2.2)
with A ∈ Kn×n (K = R or K = C) being a Hermitian positive definite matrix

with n rows and n columns. From the spectral theory it is well known that there
exist eigenvalues 0 < λ1 ≤ · · · ≤ λn and a corresponding unitary matrix U =
(u1 , . . . , un ) of eigenvectors ui ∈ Cn \{0} (i.e., uH
i uj = δi,j ) such that A has a
representation of the form
⎛ ⎞
λ1 0
⎜ .. ⎟
⎜ . ⎟
U H AU = ⎜ ⎜
⎟ = diag(λ1 , . . . , λn ).
⎟ (2.3)
⎝ .. ⎠
.
0 λn

n×n
The condition number of A is given by the quotient of the largest and smallest
eigenvalue, i.e., κ = λλn1 (note that λ1 > 0). For the sake of simplicity and coherence
with the analysis for the infinite-dimensional case below, we shall assume here that
the scaling is such that λn = 1, so that κ = λ−11 . The condition number is a measure
for the stable solvability of the problem (2.2).
Ill-conditioned matrix equations. Assume that we have noisy data y ε instead of y,
which satisfy
y ε − y ≤ ε (2.4)
in the Euclidean norm on K . Let x denote the solution with right-hand side y ε .
n ε
Then it follows from the spectral representation that

1 1
xε − x = U diag ,..., U H (y ε − y). (2.5)
λ1 λn
Hence, observing the orthogonality of eigenvectors we are led to the estimate
n
xε − x2 = λ−2 2 −2
i |ui (y − y)| ≤ λ1 y − y .
H ε ε 2
(2.6)
i=1
xε − x ≤ κy ε − y ≤ κ ε. (2.7)
The sharpness of this estimate can be immediately seen for y − y = εu1 . It is clear
ε
that with increasing condition number of the matrix A, the noise amplification
increases in the worst case. For large κ one therefore speaks of an “ill-conditioned
problem” (ICP). But it should be remarked that a finite-dimensional linear problem
is never ill posed (in the sense that the third condition in Hadamard’s classification
is violated), but for κ large one certainly comes close to this case.
We also observe that errors in low frequencies (i.e., corresponding to eigen-
vectors with large eigenvalues) are amplified less. Following our nomenclature we
see that an error in the lowest frequency, i.e., y ε − y = ε un is not amplified at
all. In fact, we just obtain xε − x = ε from the spectral representation. This
is a typical effect for inverse problems. It means that not all possible versions of
noise of the same size are equally bad, high-frequency noise corresponding to low
eigenvalues is always worse than low-frequency noise. However, in practice, we

are able to make any assumption on the noise only in rare exceptions, so that a
regularization method has to deal with arbitrary noise.
Until now, we have assumed that the matrix A is Hermitian positive definite,
i.e., the minimal eigenvalue is positive. If this is not the case, the matrix has a
non-trivial null space. If λr denotes
the minimal non-zero eigenvalue, then the
solution formula becomes x = ni=r λ−1 H
i ui ui y, and the problem is solvable if and
only if ui y = 0 for i < r. If the data set is noisy, i.e., instead of y we have y ε , we
H
are led to use the projection P y ε onto the range of A. In doing

n so we obtain for the
corresponding solution xε with data P y ε that xε − x = i=r λ−1 i ui ui (P y − y).
H ε
Since ui P y = ui y for i ≥ r we thus can estimate similarly as described above

H ε H ε
xε − x ≤ λr ε. Consequently, there is no error propagation in the null space

components and the noise amplification is actually determined by the minimal
non-zero eigenvalue.
Matricial generalized inverse (pseudoinverse, Moore–Penrose inverse). Let A ∈

Kn×m (K = R or K = C) be a matrix with n rows and m columns, y ∈ Km . Note
that A is not required to be square, no rank assumptions are made at this stage.
Then the linear system
Ax = y, x ∈ Kn , (2.8)
need not have a (unique) solution. If the system (2.8) is unsolvable, a reasonable
generalized notion of a solution is a “least squares solution”, which minimizes the
residual Ax − y in the Euclidean norm (note that stands for the Euclidean
norm in this subsection): A vector x ∈ Kn is a
(1) least squares solution of (2.8) if and only if
Ax − y = inf{Az − y : z ∈ Kn }, (2.9)
(2) best-approximate solution (or minimal norm solution) of (2.8) if and only if
x is a least squares solution and satisfies
x = inf{z : z is a least squares solution}. (2.10)
The following results are well known from classical linear algebra: A vector
x∗ is a solution of (2.9) if and only if the normal equations
AH Ax∗ = AH y (2.11)
are satisfied. The problem (2.9) possesses a unique solution if and only if A has
full rank. If A† ∈ Km×n may be understood as the matrix which assigns to each
y ∈ Kn the best-approximate solution of (2.8), then it is called the generalized
inverse (also designated, Moore–Penrose inverse or pseudoinverse) of A.
In order to construct A† and, hence, best-approximate solutions via the so-
called singular value decomposition (SVD) of A we recall the definition of singular
values of a matrix A:
Let σ1 , . . . , σr > 0 be such that σ12 ≥ σ22 ≥ · · · ≥ σr2 > 0 are the positive
eigenvalues of the matrix AH A (each one written down as often as its multiplicity
is). Then σ1 , . . . , σr are denoted the (non-zero) singular values of A.
This setting makes sense, since AH A is positive semidefinite. Obviously, r ≤
min{n, m}, where r is the rank of A. We know that a Hermitian matrix can
be diagonalized, where the diagonal elements are its eigenvalues. The following
theorem generalizes this result to the non-Hermitian case.
Let σ1 ≥ σ2 ≥ · · · ≥ σr > 0 be the singular values of A. Then there exist
unitary matrices U ∈ Km×m and V ∈ Kn×n such that
⎛ ⎞
σ1 0
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ σr ⎟
H
V AU = ⎜ ⎜ ⎟ (2.12)
0 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
0 0

m×n
holds true. The columns of U and V are eigenvectors of AH A and AAH , respec-
tively. The expression (2.12) is the so-called singular value decomposition (SVD)
of A.
The singular value decomposition (SVD) is not unique, since the unitary
matrices U and V are not. Obviously, from (2.12), we obtain
⎛ ⎞
σ1 0
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ σ ⎟ H
A=V ⎜ ⎜ r ⎟U , (2.13)
0 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
0 0
since V and U are invertible, the rank of A is r, the number of non-zero singular val-
ues (counted with multiplicity). Note that with U = (u1 , . . . , um ), V = (v1 , . . . vn ),
we have for i ∈ {1, . . . , r}
Aui = σi vi (2.14)
and
AH vi = σi ui , (2.15)
which follows from the singular value decomposition (2.12) via multiplication by
V and U H , respectively. The system {(σi ; ui , vi ) : i ∈ {1, . . . , r}} is the so-called
singular system for A. The system {v1 , . . . , vr } is an orthonormal basis for the
range R(A) = {y : Ax = y}, for any x ∈ Kn ,

r
r
Ax = Ax, vi vi = x, AH vi vi , (2.16)
i=1 i=1
which implies in connection with (2.15) that

r
Ax = σi x, ui vi (2.17)
i=1
holds true. Analogously, for all y ∈ K , m

r
AH y = σi y, vi ui . (2.18)
i=1
Note that if A has real entries, so U and V have. The notion of a singular
system and the expansions (2.17) and (2.18) generalize to compact operators on
infinite-dimensional spaces, e.g., integral operators, as we will see later on.
Let A have the SVD (2.12). Then
⎛ 1 ⎞
σ1 0
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ 1 ⎟ H
†
A =U⎜ ⎜ σ r ⎟V . (2.19)
0 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
0 0

n×m
This also implies the existence and uniqueness of a best-approximate solution.

Since A† b is the least squares solution of minimal norm, we obtain that A† b is a
solution of the normal equations AH Ax = AH b with minimal norm, i.e.,
A† y = (AH A)† AH y. (2.20)
†
This means that in order to approximate A b we may as well compute an ap-
proximation to the minimal-norm solution in the normal equations, a fact we will
heavily use in the construction of regularization methods, later on.
Historical remarks. It should be mentioned that during the last century, the con-
cept of a pseudoinverse (generalized inverse) has rated considerable attention in
the mathematical as well as geodetic literature (a bibliography, for example, listing
over 1700 references on the subject is due to [178]). One of the most significant
applications of generalized inverses is to problems of best fit. Therefore one might
seek such evidence in the writings of those who laid the foundations of the method
of least squares. C.F. Gauss developed the method of least squares in 1794, but
he did not publish his results until several years later (see, e.g., [55, 97, 217] for a
review of the subject). Gauss’s interest in the subject may be dated back to his
considerations of problems in geodesy. One should point out that Gauss [87, 88]
did not formally display A† . However, following [226], the ingredients for the con-
struction of a generalized inverse were essentially available to him, but he did not
use them. Indeed, there appears to be no evidence that he was inclined to proceed
in that direction. On the other hand, his approach to the problem of determining
best estimates is certainly in the spirit of generalized inverses. Early interest in
the first half of the last century in the subject of generalized inverses was initi-
ated by a paper on matrices by R. Penrose [208]. Indeed, basic elements of this
concept had been considered somewhat earlier. For example, E.H. Moore [165] pre-
sented a development of the notion (see also R. Baer [17], A. Bjerhammar [29, 30],
K. Friedrichs [84], F.Helmert [119], E.H. Moore [166], C.R. Rao, S.K. Mitra [220],
C.L. Siegel [238], and H. Wolf [261]). Moreover, in the setting of integral and dif-
ferential operators the concept was considered even earlier by I. Fredholm [70] and
W.A. Hurwitz [127], and by D. Hilbert [121] (see [223] for a discussion of gener-
alized inverses in classical analysis, and see also [23, 31, 178] for brief historical
sketches of the subject).
Truncated singular value regularization. The decomposition (2.19), more con-

cretely, the identity

r
y, vi
A† y = ui (2.21)
i=1
σi
also shows how errors in y affect the result A† y: Errors in components of y corre-
sponding to small singular values are amplified by the large factor of the singular
value, so that such data errors are dangerous. This explains the numerical insta-
bility of (2.21), if A has small singular values. Although the problem of computing
the best-approximate solution is well posed, it is then numerically unstable. The
first idea to reduce this instability is to replace (2.21) by
r
y, vi
xα = ui (2.22)
i=1
σi
σi2 ≥a
with an appropriately chosen value α > 0; this truncation is the first example of a
regularization, where the original problem is replaced by a neighboring one, which
is more stable. However, the choice of the “regularization parameter” α is quite
crucial.
If we use (2.22) with perturbed data y ε (with y − y ε ≤ ε), we obtain as the
“regularized solution”
r
y ε , vi
xεα = ui . (2.23)
i=1
σi
σi2 ≥α
We estimate the total error between xεα and the sought-for quantity A† y:
- -
- -
- r -
- y ε , vi r
y, v -
†
xα − A y = -
ε - ui −
i
ui - (2.24)
σi σi -
- i=1 -
-σ2 ≥α i=1
-
i
- - - -
- - - -
- - - -
- r y, vi r
y, v - - r
y, v − y ε
, v -
=-- ui −
i - -
ui - + -
i i
ui -
- .
- i=1 σi σi - - i=1 σi -
- i=1
σ2 ≥α - -σ2 ≥α -
i i
Since the elements ui are orthonormal, we have

- - - -2
- - - -
- - - -
- r
y, v
r
y, v - - r
y, v - r
|y, vi |2
- i
u −
i
u -=- i
u - = →0
- i i - - i - σi2
- i=1 σi σi - - i=1 σi -
- i=1
σ2 ≥α - -σ2 <α - i=1
σ2 <α
i i i
(2.25)
for α → 0. Hence, for sufficiently small α, the last sum is empty such that
- -2
- -
- r -
- y, vi − y ε , vi - r
|y − y ε , vi |2 1
r
ε2
- u - = ≤ |y − y ε , vi |2 ≤ .
- σi
i-
σi2 α i=1 α
- i=1 -
-σ2 ≥α - i=1
σ2 ≥α
i i
(2.26)
The second error term does not blow up as α → 0, since the sum has always at
most r terms, hence, it can be estimated by ε2 (min{σi2 : i ∈ {1, . . . , r}})−1 . The
sum (2.23) is called truncated singular value expansion. It can be interpreted as
applying a low-pass filter to the data.
Tikhonov regularization. Another way of making (2.21) → (2.22) more stable

would be to replace it by the sum

r
σi
xεα = y ε , vi ui . (2.27)
σ2
i=1 i
+α
This is a classical variant of the famous Tikhonov regularization method for matrix
equations, which we shall consider in more detail in infinite dimensions. It is helpful
to characterize it in a different way: Let xεα be defined by (2.27). Then, by the
orthonormality of the ui , we have, for all j ∈ {1, . . . , r},
σj
xεα , uj = y ε , vj (2.28)
σj2 + α
and
σj2 xεα , uj + αxεα , uj = σj y ε , vj . (2.29)
Now, because of (2.17) and (2.18), it follows that

r
r
r
σj y ε , vj uj = y ε , Auj uj = AH y ε , uj uj = AH y ε (2.30)
j=1 j=1 j=1
and

r
2 ε
r
ε H
σj xα , uj + αxεα , uj uj = xα , A Auj + αxεα , uj uj
j=1 j=1
= AH Axεα + αxεα , (2.31)
note that xεα is in the linear span of {u1 , . . . , ur } which follows from (2.27). Now,
the identity (2.29) implies in vector nomenclature that
H
A A + αI xεα = AH y ε , (2.32)
which is an alternative characterization of the Tikhonov regularization. From this

“regularized normal equation”, we can obtain still another characterization of xεα ,
namely as the unique minimizer of the so-called Tikhonov functional
x → Ax − y ε 2 + αx2 , (2.33)
which can be seen by putting the first derivative of the functional in (2.33) to 0,
resulting exactly in the linear equation (2.32). The minimization of (2.33) can be
seen as a combination of the two minimizations that appear in the definition of a
best-approximate solution. It has also interpretations as a penalty method, e.g.,
via Lagrange multipliers.
The computation explained above can be formally performed for α = 0, too.
In this case it shows that x is the solution of minimal norm of the normal equation
AH Ax = AH y which was already attacked by C.F. Gauss [87] (see also the contri-
butions by R.L. Plackett [217] and D.W. Robinson [226], and for a deeper insight
[178]). The monograph [178] also contains a brief historical sketch of this subject.
The book [260] (see also the references therein) presents an overview about numer-
ical methods and procedures. If AH A is invertible (and hence positive definite),
the normal equation AH Ax = AH y can be solved by standard Cholesky decom-
position, which leads to an alternative method for computing best-approximation
solutions, for which no SVD is needed.
2.2. Least squares problems and generalized inverses

As usual, the domain, range, and null space (kernel) of any operator A : D(A) →
R(A) are denoted by D(A), R(A), and N (A), respectively.
We start with solvability conditions of an operator equation in Hilbert spaces,
where the concepts of “distance” and “angle” are at the disposal for an applicant.
More concretely, let X and Y be Hilbert spaces and let A : X → Y be a bounded
linear operator whose range R(A) is not necessarily closed. Then we have the
orthogonal decompositions
X = N (A) ⊕ N (A)⊥ , (2.34)
Y = R(A) ⊕ R(A)⊥ , (2.35)
and
N (A∗ ) = R(A)⊥ , (2.36)
⊥
where N (A) is the orthogonal complement of N (A), R(A) is the closure of the
range of A, and A∗ is the adjoint operator of A, i.e., Ax, y = x, A∗ y for all
x ∈ X and y ∈ Y .
We consider the operator equation
Ax = y. (2.37)
Four (mutually exclusive) situations arise (cf. [184]):
1) R(A) is dense in Y , (hence, N (A∗ ) = {0}), and y ∈ R(A);
2) R(A) is dense in Y , and y ∈ / R(A);
3) R(A) is a proper subspace of Y , and y ∈ R(A) + R(A)⊥ ;
4) R(A) = Y , and y ∈ / R(A) + R(A)⊥ .
In case 1) one has, of course, a solution in the classical sense; in case 2) and 4) a
classical solution does not exist, while in case 3) a solution need not exist.
We say x is a “least squares solution” of (2.37) if
inf{Au − y : u ∈ X} = Ax − y. (2.38)
Since
Au − y2 = Au − Qy2 + y − Qy2 , (2.39)
where Q is the orthogonal projector of Y onto R(A), it is clear that a least squares
solution exists if and only if
y ∈ R(A) + R(A)⊥ , (2.40)
where R(A) + R(A)⊥ is a dense set in Y . For such y the set of all least squares
solutions of (2.37), denoted by L(y), is a non-empty closed convex set (indeed
L(y) is the translate of N (A) by a fixed element of N (y)), hence, it has a unique
element of minimal norm, denoted by A† y.
The generalized inverse (or pseudoinverse) A† is the linear operator which
assigns to each y ∈ D(A† ) := R(A)+ R(A)⊥ , the unique element in L(y)∩N (A)⊥ ,
so that L(y) = A† y + N (A). It is easy to show that A† y is the minimal norm
solution (equivalently the unique solution in N (A)⊥ ) of the normal equation
A∗ Ax = A∗ y (2.41)
2
(the equation obtained by setting the first variation of Ax − y equal to zero).
It also follows that A† = (A/N (A)⊥ )−1 Q so that A† can be characterized as the
linear operator with the function-theoretic properties:
D(A† ) = R(A) + R(A)⊥ , N (A† ) = R(A)⊥ = N (A∗ ) (2.42)
and
R(A† ) = N (A)⊥ . (2.43)
The equivalence of these characterizations of A† is established in [173] (see also
[104, 185] for a lucid exposition and [185, 194] for generalization to unbounded
operators).
In case 1) above, A† gives the minimal-norm solution of 3). In case 3), Equa-
tion (2.37) has a least squares solution (which is unique if and only if N (A) = {0}).
In both cases the infimum in (2.38) is attained and is equal to zero and y − Qy,
respectively. Case 2) and 4) are pathological and usually are not under discussion
in generalized inverse theory, since in both cases y ∈ / D(A† ), and the infimum in
(2.38) is not attained.
As canonical evolution of Hadamard’s classification, M.Z. Nashed [178, 184]
called the operator equation (2.37) well posed in the least squares (relative to
X and Y ) if for each y ∈ Y the equation has a unique least squares solution
(of minimal norm), which depends continuously on y; otherwise the problem is ill
posed. The advantage of adopting this notion of well-posedness is that it focuses on
infinite-dimensional problems (e.g., an inconsistent finite system of linear algebraic
equations will not be ill posed in above sense, while it is ill-posed in the sense of
Hadamard). It follows immediately from the open mapping theorem in functional
analysis (see, e.g., [245]) that the following statements are equivalent:
a) the problem (2.37) is well posed;
b) R(A) is closed;
c) A† is bounded.
Summarizing we are led to the following conclusion (see [16, 184, 185]): The
problem (A; X, Y ) is called well posed in the sense of Nashed, if R (A) is closed
in Y . If R (A) is not closed in Y , the problem (A; X, Y ) is called ill posed in the
sense of Nashed.
2.3. Weighted least squares problems

Very often we are interested in weighted minimal-norm least squares solutions. Let
LA (y) be the set of all least squares solutions of Ax = y, where A is a bounded
linear operator from X into Y . Let Z be a Hilbert space and L : DL ⊂ X → Z be
a closed linear operator with dense domain and closed range. For y ∈ D(A† ), we
consider the following problem: find w ∈ LA (y) such that
Lw ≤ Lu for all u ∈ LA (y). (2.44)
If L(N (A)) is closed and N (A) ∩ N (L) = {0}, then (2.44) has a unique solution
w(y). We denote by A†L the linear map induced by y → w(y) and call it the
weighted generalized inverse of A We define a new inner product and norm on
D(L) by
[u, v]L := (Au, Av) + (Lu, Lv) (2.45)
#
with uL := [u, u]L . We denote the space D(L) with this new inner product by
XL . It is easy to show that A†L y is the least squares solution of XL -minimal norm
of Ax = y. Let
M := {x ∈ X : L∗ Lx ∈ N (A)⊥ }. (2.46)
Then M is the orthogonal complement of N (A) with respect to [·, ·], and
R(A†L ) = M.
2.4. Singular value decomposition for compact operators

Next we discuss a certain set of operators, viz. compact operators, that turn out
to be specific prototypes for generating a large class of ill-posed problems.
Let X, Y be normed spaces. An operator A : X → Y is called compact, if
one of the following equivalent conditions is fulfilled:
(1) Every bounded subset U ⊂ X possesses an image in Y , which is relatively
compact, i.e., A(U ) is a compact set.
(2) For every bounded sequence {xn }n ⊂ X the sequence {Axn }n possesses a
subsequence that converges in Y .
By convention, we introduce the following settings:
K (X, Y ) = {A : X → Y : A is linear and compact} (2.47)
and
K (X) = K (X, X) . (2.48)
Let X, Y, Z be normed spaces. Then the following statements hold true:
(1) K (X, Y ) ⊂ L (X, Y ) .
(2) If A ∈ L(X, Y ) with dim R (A) < ∞, then A is compact.
(3) If A : X → Y is compact and B : Y → Z is continuous or A : X → Y is
continuous and B : Y → Z is compact, then AB : X → Z is compact.
(4) The identity operator I : X → X is compact if and only if X is finite-
dimensional.
(5) Let Y be a Banach space. Then K (X, Y ) is closed, i.e., a sequence of compact
operators {An }n ⊂ K (X, Y ) limn→∞ An − AX→Y = 0 has a compact
limit, i.e., the limit operator A is compact.
(4) If A is compact and invertible and X is not finite-dimensional, then A−1 is
not continuous.
Example 1. Let G be a regular region in Rq , i.e., a bounded region G dividing Rq
uniquely into the inner space G and the outer space G c = R3 \G, G = G ∪ ∂G, such
that the boundary ∂G is an orientable smooth Lipschitzian
manifold of dimension
q − 1, and suppose that K is of class C (0) G × G . We introduce the integral

operator A : C (0) G → C (0) G by letting

(AF ) (x) = K (x, y) F (y) dy, F ∈ C (0) G . (2.49)
G
The proof of the compactness of A can be based on a well-known theorem of

classical analysis, namely the Theorem of Arzelà–Ascoli. This theorem provides
two equivalent properties to the relative compactness of a subset U ⊂ C (0) G :

Let ∅ = G be regular. A subset U ⊂ C (0) G is relatively compact if and only
if the following two statements are valid:
(1) U is equicontinuous, i.e., for every ε > 0 there exists δ (ε) > 0, such that for
all F ∈ U
|F (x) − F (y)| < ε (2.50)
for all x, y ∈ G with x − y < δ (ε) .
(2) U is bounded, i.e., there exists an M > 0 with F C (0) (G ) < M for all F ∈ U.
In accordance with the theorem of Arzelà–Ascoli we are now interested in ap-
plying the conditions (1) and (2) to the context of the integral operator introduced
in (2.49):
(1) K is uniformly continuous in G × G. Therefore, AU is equicontinuous.
(2) Suppose that U ⊂ C (0) (G) is bounded. Assume that F is of class U . Then we
have

|AF (x)| = K (x, y) F (y) dy ≤ M sup |K (x, y)| G < ∞. (2.51)
G x,y∈G
In other words, AU is bounded.

As a consequence, by virtue of the theorem of Arzelà–Ascoli, we are able to con-
clude that AU is relatively compact, so that the operator A defined by (2.49) is
compact.

Example 2. Let G be a regular region in Rq , and K be of class L2 G × G , then

A : L2 G → L2 G given by

AF = K (·, y) F (y) dy, F ∈ L2 G (2.52)
G
is compact (the proof can be found in, e.g., [120, 139]).
G be a regular region in R . We introduce the operator A :
q
3.2 Let
Example
2
L G → L G given by

AF (x) = K (x, y) F (y) dy, F ∈ L2 (G). (2.53)
G
If K is continuous for x = y and weakly singular, i.e., there exist a value α ∈ (0, q)
and a constant C > 0 such that
1
|K (x, y)| ≤ C , (2.54)
|x − y|q−α
then A is compact (for the proof see, e.g., [135]). As a consequence, the Newton
volume integral

1 1
V (x) = AF (x) = F (y) dy, F ∈ L2 (G), (2.55)
4π G |x − y|
occurring in the so-called inverse gravimetry problem of determining the geologic

density distributions F inside the Earth from the Earth’s gravitational potential
V in G c ⊂ R3 forms a compact operator A.
Singular value decomposition. Next we are concerned with the introduction of

eigenvalues and eigenfunctions corresponding to an operator A ∈ L(X). Let X be
a normed space. Suppose that A is of class L (X).
(1) λ ∈ C is called a regular value of A if and only if λI −A is continuously invert-

ible. ρ (A) = {λ ∈ C : λI − A continuously invertible} is called the resolvent
set.
(2) σ (A) = C \ ρ (A) is called the spectrum of A.
(3) λ ∈ σ (A) is called an eigenvalue of A if N (λI − A) = {0}. The elements
of N (λI − A) \ {0} are called eigenvectors of A corresponding to the eigen-
value λ.
The following results are standard for a Banach space X (see, e.g., [120]):
(1) If λ ∈ σ (A), then |λ| ≤ A, i.e., the spectrum is bounded.
(2) σ (A) ⊂ C is compact.
Let X be a normed space. Suppose that A is a compact operator on X (i.e.,
A ∈ K (X)).
(1) If λ ∈ σ (A) \ {0} , then λ is an eigenvalue of A.
(2) If λ is an eigenvalue of A, then dim N (λI − A) < ∞.
(3) σ (A) is at most countable. Furthermore, 0 ∈ σ (A).
(4) 0 is the only accumulation point of σ (A).
Central in our considerations about compact operator is the following spectral

theorem for compact self-adjoint operators that can be seen in parallel to the finite-
dimensional case of matrix operators:
Let X be a Hilbert space. Assume that A is of class K (X) and that A is
self-adjoint, i.e., A∗ = A. Then there exists an orthonormal system {xi }i∈N ⊂ X
and a sequence {μi }i∈N ⊂ R (finite or countable) with |μ1 | ≥ |μ2 | ≥ · · · > 0, such
that
∞
Ax = μi x, xi xi (2.56)
i=1

Suppose that X and Y are Hilbert spaces. Furthermore, let A be of class
K(X, Y ). Then A∗ A is also compact and obviously self-adjoint. Due to the spectral
theorem there exist a sequence {λi }i∈N ⊂ R and an orthonormal system {xi }i∈N ⊂
X such that
∞

A∗ Ax = λi x, xi xi , x ∈ X. (2.57)
i=1
Suppose that λi ∈ σ (A∗ A)\ {0} and denote, as usual, by xi its corresponding
eigenvector. It follows that
2 2
λi xi = λi xi , xi = λi xi , xi X = A∗ Axi , xi X = Axi , Axi Y = Axi Y .
(2.58)
Therefore we are able to conclude that λi > 0.
Singular values. In the sequel, we assume that the eigenvalues are listed in the
chronological order as follows:
λ1 ≥ λ2 ≥ · · · ≥ λi ≥ λi+1 ≥ · · · ≥ 0. (2.59)
#
Set σj = λj . Moreover, let yi = σ1i Axi i.e., Axi = σi yi , i ∈ N, and

1 1 1
A∗ yi = A∗ Axi = A∗ Axi = λi xi = σi xi . (2.60)
σi σi σi
It is not hard to see that
1 1 1
yi , yk Y = Axi , Axk Y = A∗ Axi , xk X = λi xi , xk X
σi σk σi σk σi σk
(2.61)
λi σi
= xi , xk X = δi,k = δi,k .
σi σk σk
Thus, {yi }i∈N ⊂ Y forms a complete orthonormal system (ONS) in R(A), so that
⊥
the system {xi }i∈N is a complete ONS in N (A) . Now, assume that x is a member
⊥
of N (A) . Then it follows that
∞

x= x, xi X xi (2.62)
i=1
and
∞
∞

Ax = x, xi X Axi = σi x, xi X yi (2.63)
i=1 i=1
⊥
for all x ∈ N (A) .
Let X, Y be Hilbert spaces. The set {σi ; xi , yi }i∈N ⊂ (0, ∞) × X × Y is called
the singular system of an operator A ∈ K (X, Y ). The values σi are called the sin-
gular values of A. The elements xi , yi are called the singular vectors. Furthermore,
the series
∞
Ax = σi x, xi X yi , x ∈ X (2.64)
i=1
is called the singular value decomposition (SVD) of A.
Picard condition. The following condition plays an essential role in the solvability
of inverse problems.
If A : X → Y is compact with singular value decomposition (SVD)
{σi ; xi , yi }i∈N ,
then y ∈ R (A) is an element of R (A) if and only if

∞
2
|y, yi |
(2.65)
i=1
σi2
is convergent.
Obviously, from the Picard condition, it follows that σi−2 |y, yi |2 → 0 if
i → ∞ so that information about the decay of the Fourier coefficients of an element
y becomes available.
Let A be a compact operator (i.e., A ∈ K (X, Y )) with SVD {σi ; xi , yi }i∈N .
Then the generalized inverse (or pseudoinverse) of a compact operator is repre-
sentable in the form
∞
† 1
A y= y, yi xi (2.66)
i=1 i
σ

for all y ∈ D A† .
If R (A)
is finite dimensional (i.e., there exist only finitely many element yi ),
then R A† < ∞. Therefore, A† is compact. In particular, A† is continuous.
The representation of the generalized inverse in terms of the singular value
decomposition (2.66) opens the perspective to classify ill-posed problems. Indeed,
the summands σi−1 y, yi xi occurring in the series (2.66) depend closely on the
singular values. If the values σi are small, then the contribution by the series
(2.66) becomes large. The existence of SVD can be guaranteed for all compact
operators. Nevertheless, the concrete knowledge of SVD is critical. Only in rare
exceptions, SVD is explicitly known in practice.
Regularization methods. Obviously, the first two criteria (H1) and (H2) deter-
mining a well-posed problem in the sense of Hadamard, can always be enforced
by considering the generalized inverse A† . A violating of the third criterion, i.e.,
instability arises if the spectrum of the operator A is not bounded away from
zero. Thus, it seems to be natural to construct regularizing approximations via
modifying the smallest singular values. In accordance with the singular value de-
composition of the generalized inverse, it follows that such a modification of small
values and, hence, a construction of regularization operators can be obtained in
the form
∞

xα = Rα y = σi Fα (σi2 )y, yi xi y ∈ Y, (2.67)
i=1
with some function Fα : R+ → R+ such that
1
, λ > 0, α → 0.
Fα (λ) → (2.68)
λ
Such an operator Rα as defined by (2.67) and (2.68) may be understood as a
regularization operator if
λ|Fα (λ)| ≤ CFα < ∞, λ > 0. (2.69)
If (2.69) is satisfied, then we are able to see that

∞
∞

Rα y2 = σi2 (Fα (σi ))2 |y, yi |2 ≤ CF2 α |y, yi |2 ≤ CF2 α y2 , (2.70)
i=1 i=1
where CFα is a bound for the norm of Rα . Note that the pointwise convergence of
Fα immediately implies the pointwise convergence of Rα to A† .
Truncated singular value regularization. Let A be a compact operator (i.e., A ∈
K (X, Y )) with SVD {σi ; xi , yi }i∈N . The main idea of truncated singular value
decomposition is to ignore all singular values below a certain threshold value, which
we can identify with the regularization parameter α, hence, the representation of
the regularized solution is given by
1
, λ≥α
Fα (λ) = λ (2.71)
0, λ<α
i.e.,
1
xα = Rα y = y, yi xi , y ∈ Y, (2.72)
σi
σi ≥α
which explains the name truncated singular value decomposition, since all terms
in the sum corresponding to small singular values are truncated. Since 0 is the
only accumulation point of the singular values of a compact operator, the sum in
(2.72) is always finite for α > 0. In particular, only a finite number of singular
values and singular vectors has to be computed in order to realize this method.
On the other hand it should be mentioned that, for α being sufficiently small,
the number of singular values that need to be computed can increase strongly.
Obviously, CFα = α−1 .
Lavrentiev regularization. The main idea of this regularization method is to shift
all singular values by α, i.e.,
1
Fα (λ) = √ √ , λ > 0, α > 0, (2.73)
λ( λ + α)
so that
∞
1
xα = Rα y = y, yi xi , y ∈ Y. (2.74)
i=1
σi + α
In this case, the sum is really infinite and the full singular system is needed in
order to compute the solution. However, if A is a positive semidefinite operator
(and, thus, xi = yi ), we obtain
∞

(A + αI) xα = y, yi xi = y. (2.75)
i=1
Hence, the regularized solution can also be obtained in this case without any knowl-
edge of the singular system as the solution of the linear equation (A + αI) xα = y.
Clearly, (σ + α)−1 ≤ α−1 , hence, CFα = α−1 .
Tikhonov regularization. The regularized solution is given by

1
Fα (λ) = , λ > 0, α > 0, (2.76)
λ+α
so that
∞
σi
xα = Rα y = 2 y, yi xi , y ∈ Y. (2.77)
σ +α
i=1 i
As in the case of Lavrentiev regularization, we can compute xα defined by (2.77)
without any knowledge of the singular system. In fact, it is easy to see that
(A∗ A + αI) xα = A∗ y (2.78)
and, hence, we can solve a well-posed linear system to obtain xα . From this rep-
resentation it also follows that Tikhonov regularization is just Lavrentiev regular-
√
ization applied to the normal equation. It is not hard to see that λ2 + α ≥ 2λ α,
hence, CFα can be chosen as 2α−1/2 .
Asymptotic regularization. Asymptotic regularization is usually constructed from
the solution x of the initial value problem
x (t) = −A∗ (Ax(t) − y), t > 0, (2.79)
x(0) = 0, (2.80)
as xα = x( α1 ·). By representing x in terms of the singular vectors xi in the form
∞

x(t) = αi (t) xi (2.81)
i=1
with αi (0) = 0, we obtain from the singular value decomposition
αi (t) = −σi2 αi (t) + σi yi , y. (2.82)
This ordinary differential equation can be solved analytically by
1
αi (t) = 1 − exp(−σi2 t) yi , y. (2.83)
σi
Hence, the regularized solution is given by

λ 1
Fα (λ) = 1 − exp − (2.84)
α λ
i.e.,
∞
σi2 1
xα = 1 − exp − y, yi xi , y ∈ Y. (2.85)
i=1
α σ i
Error estimates. For the error between x† and xεα in the case of noisy data y ε , we
are able to write (with xεα = Rα y ε )
x† − xεα = (x† − xα ) + (xα − xεα ). (2.86)
†
The first term x − xα is the approximation error of the regularization method,
which is independent of the noise. The second term xα − xεα corresponding to the
propagation of data noise in the regularized case. By aid of the triangle inequality
it follows that
x† − xεα ≤ x† − xα + xα − xεα . (2.87)
Thus, the estimate of the error between the regularized solution and the exact
solution, can be handled by two error terms separately. It is clear that such an
estimation provides a guideline for the parameter choice, namely by choosing α
such that the terms on the right-hand side are balanced.
Next we deal with an estimate of the approximation error, which is indepen-
dent of the noise level ε:
Let Fα : R+ → R+ be a piecewise continuous function satisfying the assump-
tions
1
Fα (λ) → , λ > 0, α → 0,
λ
|Fα (λ)| ≤ Mα < ∞, λ > 0,
and
sup(λFα (λ)) ≤ CF < ∞ (2.88)
α,λ
for some constant CF > 0. Moreover, let the regularization operator be defined by
(2.67). Then, for all y ∈ D(A† ), we have
Rα y → A† y, α → 0. (2.89)
+
The function t → Fα (t), t ∈ R , converges pointwise to the function

0 , t>0
F (t) = (2.90)
1 , t = 0.
Due to the discontinuity at zero, the convergence of tFα (t) − 1 to zero is becoming
slower and slower as t decreases to zero. Since it is allowed to specify an arbitrarily
small singular value σi and the minimal norm solution x† = xi , the convergence of
regularized solutions is arbitrarily slow. On the other hand, we observe that there
is a possibly faster convergence if the components x† , xi decay sufficiently fast
compared to the eigenvalues. For example, if we have |x† , xi | ≤ cσiμ for some
constant c > 0 and μ > 0, then it follows
∞
lim sup Rα y − A† y2 ≤ lim sup c2 (σi Fα (σi ) − 1)2 σi2μ
α→0 α→0
n=1
∞

≤ c2 lim(σi1+μ Fα (σi ) − σiμ )2 . (2.91)
α
n=1
In other words, one has to consider the limit of the function t → |t1+μ Fα (t)−tμ | as
t → ∞ instead, which is usually much faster. For example, in case of the truncated
singular value decomposition, we obtain

1+μ 0 , t≥α
|t Fα (t) − t | =
μ
(2.92)
tμ , t < α.
If the singular values of the operatordecay sufficiently fast (which is the typi-
∞
cal case for ill-posed problems), e.g., n=1 σiμ < ∞, we are confronted with the
situation
2μ ∞
Rα y − A† y2 ≤ c2 σi ≤ c2 αμ σiμ (2.93)
σi <α i=1
so that the term Rα y − A†y μ/2

is of order α . Consequently, we somehow need
smoothness of the solution (in terms of the smoothing properties of the operator)
in order to obtain a convergence rate in terms of α. We shall pursue this idea
by introducing spaces of smoothness involving the absolute value of a compact
operator.
Next we are concerned with the propagation of the data error through the
regularization. Let Fα and CF be as given above, and let xα = Rα y, xεα = Rα y ε .
From the singular value decomposition it follows directly
∞
2 2
Axα − Axεα 2 ≤ σi Fα (σi2 ) |y − y ε , yi |2
i=1
∞

≤ CF2 |y − y ε , yi |2 = CF2 y − y ε 2 ≤ (CF ε)2 , (2.94)
n=1
so that
Axα − Axεα ≤ CF ε (2.95)
is valid. In the same way we obtain
∞

xα − xεα 2 ≤ (Fα (σi ))2 |y − y ε , yi |2
i=1
∞

≤ Mα2 |y − y ε , yi |2 = Mα2 y − y ε 2 ≤ (Mα ε)2 , (2.96)
i=1
so that
xα − xαα ≤ Mα ε (2.97)
is implied (note that (2.97) estimates the norm of Rα by CF ).
3. Operator methodologies of resolution

Numerous methods have been proposed for treating and regularizing various types
of ill-posed problems (IPP’s). The rationale in most methods for resolution (ap-
proximate solvability) of IPP is to construct a “solution” that is acceptable phys-
ically as a meaningful approximation and is sufficiently stable from the computa-
tional standpoint, hence, an emphasis is put on the distinction between “solution”
and “resolution”. As already mentioned, the main dilemma of modeling of ill-
posed problems is that the closer the mathematical model describes the IPP, the
worse is the “condition number” of the associated computational problem (i.e.,
the more sensitive to errors). For ill-posed problems, the difficulty is to bring ad-
ditional information about the desired solution, compromises, or new outlooks as
aids to the resolution of IPP. It is conventional to use the phrase “regularization
of an ill-posed problem” to refer to various approaches to circumvent the lack of
continuous dependence (as well as to bring about existence and uniqueness if nec-
essary). Roughly speaking, this entails an analysis of an IPP via an analysis of
an associated well-posed problems, i.e., a system (usually a sequence or a family)
of well-posed problems, yielding meaningful answers to the IPP. We distinguish
three aspects of regularization:
(a) strategy of resolution and reconstruction,
(b) regularization-approximation schema,
(c) regularization algorithms.
One of the purposes of our work is to dramatize this delineation with reference
to specific methods and results.
The strategy of resolution and reconstruction of ill-posed problems involves
one or more of the following intuitive ideas (cf. [184]):
(α) change the notion of what is meant by a solution (e.g., ε-approximate solu-
tion: Au − y ≤ ε, where ε > 0 is prescribed; quasi-solution: Au − y ≤
Ax − y for all x ∈ M, a prescribed subset of the domain of A; least squares
solution of minimal norm, etc.),
(β) modify the operator equation or the problem itself,
(γ) change the spaces and/or topologies,
(δ) specify the type of involved noise (“strong” or “weak” noise).
The philosophy of resolution leads to the use of algebraic methods versus func-
tion space methods, statistical versus deterministic approaches, strong versus weak
noise (see [50, 51, 53], where the concept of weakly bounded noise was first intro-
duced), etc.
By a regularization-approximation scheme we refer to a variety of methods
such as Tikhonov’s regularization, projection methods, multiscale methods, itera-
tive approximation, etc., that can be applied to ill-posed problems. These schemes
turn into algorithms once a resolution strategy can be effectively implemented. Un-
fortunately, this requires a determination of a suitable value of a certain parameter
associated with the scheme (e.g., regularization parameter, mesh size, dimension
of subspace in the projection scheme, specification of the level of a scale space,
classification of noise, etc.). This is not a trivial problem since it involves a trade-off
between accuracy and numerical stability, a situation that does not usually arise
in well-posed problems.
From the standpoint of mathematical and numerical analysis one can roughly
group “regularization methods” into three categories (cf. [184]):
(a) Regularization methods in function spaces is one category. This includes
Tikhonov-type regularization, the method of quasi-reversibility, the use for
certain function spaces such as scale spaces in multi-resolutions, the method
of generalized inverses (pseudoinverses) in reproducing kernel Hilbert spaces,

and multiscale wavelet regularization.
(b) Resolution of ill-posed problems by “control of dimensionality” is another cat-
egory. This includes projection methods, discretization and moment-discreti-
zation schemes. The success of these methods hinges on the possibility of ob-
taining approximate solutions while keeping the dimensionality of the finite-
dimensional problem within the “range of numerical stability”. It also hinges
on deriving error estimates for the approximate solutions that is crucial to
the control of the dimensionality.
(c) A third category are iterative and filtration methods which can be applied
either to the problem in function spaces or to a discrete version of it. The
crucial ingredient in iterative methods is to stop the iteration before insta-
bility creeps into the process. Thus iterative methods have to be modified
or accelerated so as to provide a desirable accuracy by the time a stopping
rule is applied. Filtration methods refer to procedures where, for example,
singular functions and values producing highly oscillatory solutions are elim-
inated. Various “low pass” filters can, of course, be used. The last sentence
in (b) is also crucial for the determination of a stopping rule.
3.1. Concept of regularization revisited
The concept of a regularizer plays an important role in obtaining an approximate
solution of an IPP in the presence of contamination in the data. We shall explain
show that it is, indeed, a generic concept that can be used to unify some of the
principles occurring in various regularization-approximation schemes (Tikhonov’s
regularization, truncated or filtered singular-value expansions, projection methods,
multiscale techniques, iterative methods, etc).
More explicitly, let X, Y be normed spaces and let A : X → Y be a one-to-one
mapping (not necessarily linear). Note that the assumption that A is one-to-one
is imposed for the convenience of the linear case. In fact, it can be dropped if one
uses generalized inverses, as is done in several papers of Nashed [176, 178]. This
assumption will also be dropped in this contribution after we will have motivated
and explained the auxiliary procedure for the case when A is assumed to be one-
to-one.
A regularizer to the operator equation Ax = y is a one-parameter family of
operators {Rt : t ∈ Γ} where Γ is an index set of real numbers with 0 ∈ Γ (the
closure of Γ) satisfying the following conditions:
(C1) For each t ∈ Γ, Rt is a continuous operator on all of Y into X.
(C2) For each x ∈ X, limt→0 Rt Ax − x = 0.
Strong noise. Ill-posed problems of mathematical practice are usually regular-
izable in following sense: For y ∈ R(A), let y ε be known with “noise level”,
y ε − y ≤ ε (“strong noise condition”). Consider the operator equation between
normed spaces X, Y
Ax = y ε . (3.1)
The existence of a regularizer enables us to calculate an “approximate solution”

xε − A−1 y → 0 and Axε − y → 0. We explain this fact in more detail. Let xε
be given by xε := Rt y ε . Then
xε − A−1 y ≤ Rt y − A−1 y + Rt y ε − Rt y, (3.2)
where we assume that both norms on the right side of (3.2) are known. Note that
Rt y − A−1 y gives a rate of convergence of Rt Ax to x and Rt y ε − Rt y is the
modulus of continuity of the operator Rt at y. Observe that Rt y ε − Rt y → 0 as
ε → 0 for any fixed t. For given ε > 0 we choose t = t(ε) to minimize the right side
of (3.2). Then xε = Rt(ε) y ε has the claimed property since Rt(ε) y ε − A−1 y → 0
as ε → 0.
At this stage we shall be primarily interested in the case when A will be
assumed to be linear. For simplicity, Rt will be required to be linear, too. We then
have
xε − A−1 y ≤ Rt y − A−1 y + M (t) ε (3.3)
where Rt ≤ M (t) (note that the operators Rt are not uniformly bounded since
A−1 is unbounded, so M (t) → ∞ as t → 0).
The notion of a regularizer can be easily extended to weighted least squares

problems (see Subsection 2.3): We say that the problem (3.1), or equivalently A†L , is
regularizable if there exists a one-parameter family of linear operators {Rt : t ∈ Γ}
with R(Rt ) ⊂ M such that limt→0 Rt Ax − x = 0 for x ∈ M and for each t > 0,
Rt is bounded. Again, the family {Rt } is not uniformly bounded in t since R(A)
is non-closed. As before, the existence of a regularizer provides us with a family
of approximate solutions determined by a well-posed problem. In the presence of
contamination in y, say y ε − y ≤ ε, the error Rt y ε − A†L yX → 0 as t → 0 (in
fact, it blows up). The criterion then is to choose t to minimize the error:
Rt y ε − A†L y ≤ Rt y − A†L y + Rt (y ε − y)

≤ Rt y − A†L y + Rt ε. (3.4)
The first term (regularization error ) tends to zero as t → 0, while the second term
(magnification of contamination error due to ill-posedness) tends to ∞:
- -
- † -
-AL y − Rt y - −→ 0

t→0 (3.5)

Rt (y − y ε ) ≤ Rt ε −→ ∞
- -
- † -
-AL y − Rt y - −→ ∞

t→∞ (3.6)

Rt (y − y ε ) ≤ Rt ε −→ 0
If we know an error estimate for the first term and a growth estimate for
Ct , a suitable t can be determined. Such estimates can be obtained for particular
regularizers Rt using additional information on the solution A†L y, e.g., smoothness,
and some robustness condition on Ct , e.g., one might estimate that Ct ≤ d(t) is
a known function which tends to ∞ as t → 0, and Rt y − A†L y ≤ b(t), where
b(t) → 0 as t → 0. Then, an optimal t(ε) can be easily calculated, and for this
t(ε), Rt(z) y ε → A†L y, as ε → 0. Concrete realizations of regularizers abound in
regularization methods, projection and iterative methods, etc., as we shall see
in the remaining work. In the case of an iterative scheme, t = 1/n, xn = Rt y
represents the nth iterate of a process which converges to A†L (or A−1 ) in the
absence of contamination in y. In projection methods t = 1/n, where n is the
dimension of the approximating subspace. In finite differences, t represents the
mesh size h. The preceding results then show how to obtain stable approximate
solutions in the presence of error in y it the needed estimates in (3.4) are available.
Weak noise. Let K : X → Y be a linear compact operator between the Hilbert
spaces X and Y . The inner products and norms of X and Y are denoted by ·, ·X ,
·, ·Y and ·X , ·Y (note that we do not use subscripts if they are clear from the
context, here and elsewhere). Consider the data y ∈ Y according to the equation
y = Kx0 + η , (3.7)
where η ∈ Y is the unknown noise and x0 ∈ X is an unknown element one wishes
to recover from the data y. The following model is imposed on the noise. Let
A : Y → Y be linear, compact, Hermitian, and positive-definite (i.e., y, Ay > 0
for all y ∈ Y , y = 0), and let
ε2 := η, Aη. (3.8)
We assume that ε is “small” and investigate what happens when ε → 0. The
operator A introduced above is not meant to be arbitrary. In fact, it must be
connected with K in the sense that, for some m ≥ 1 (not necessarily integer), the
range of K is continuously embedded into the range of Am , so that
A−m K : X → Y is continuous . (3.9)
If η satisfies (3.8), (3.9), it is referred to as weakly bounded noise.
Some comments should be made: In a deterministic setting, a reasonable
model for the noise is that it is “high-frequency”, and we would like to investigate
what happens when the frequency tends to ∞, but without the noise tending
to 0 strongly, that is without assuming that ηY → 0. Thus, η → 0 weakly
begins capturing the essence of “noise”. Then, for any linear compact operator
S : Y → Y , we would have SηY → 0. So, in this sense, there is nothing unusual

about (3.8) and (3.9). Moreover, we would like (3.8) to capture the whole truth,
i.e., the relations
η, Ap η = o(ε2 ) and η, Aq η = O(ε2 ) (3.10)
fail for p > 1 and q < 1 as ε → 0. This may be a tall order, although examples of
operators A and noises η satisfying (3.8)–(3.10) are easily constructed (for more
details see [50]). At the same time A is supposed to capture the smoothing effect
of K in the sense of (3.9). Ideally, one would like A−m K to be continuous with
a continuous inverse. The natural choice A = (KK ∗ )1/2m would achieve this, but
would have to be reconciled with (3.8) and possibly (3.10). The condition (3.9) is
not unreasonable.
Eggermont et al. [53] show how the weak noise model leads to simple bounds
on expressions like η, yY for y ∈ Am (Y ), the range of Am . For β > 0, they
introduce the inner product on Am (Y ), by letting
y, zm,β = y, zY + β 2m A−m , y, A−m zY , z ∈ Am (Y ), (3.11)
and denote the associated norm by · m,β . The following result is of interest in
itself, but it also later on plays a crucial role in the context of Tikhonov regular-
ization with weakly bounded noise:
Suppose that m ≥ 1. Under the assumptions (3.8), (3.9) on the weakly
bounded noise, for all y ∈ Am (Y ) and all β > 0
|η, yY | ≤ β −1/2 ε ym,β (3.12)
(note that the factor β −1/2 stays the same, regardless of m).
3.2. Use of compactness and a priori bounds

The use of a priori bounds (more generally, a priori information) about the so-
lution of an ill-posed problem has long been recognized to play a significant role
in bringing about continuous dependence (i.e., of providing a regularizing effect).
Early in the study of ill-posed problems, a fundamental observation was made by
Tikhonov (cf. [246, 249] for original references) that the restriction to a compact
set insures well-posedness. More precisely, suppose X and Y are metric spaces
and F : X → Y is a continuous injection, and let C ⊂ X be compact. Then
F −1 : F (C) → C is continuous: To show this, let W ⊂ C be open in the rela-
tive topology, then the complement of W relative to C, denoted by W c , is closed
and, hence, compact since C is compact. Continuity of F implies that F (W c ) is
compact and, therefore, closed. From this it follows that F (W ) is open, for F is
injective (and hence F (W c ) ∩ F (W ) is empty).
Remark. The use of differential operators as smoothing conditions often leads to

a setting in which the restriction of the domain of the operator to a compact
set automatically prevails. For example, consider the simple situation treated by
Tikhonov (see [250]), where
Jα (f ) = Af − g2 + α Ω(f ), α > 0, (3.13)
with
1
Ω(f ) = {p(x)[f (x)]2 + q(x)[f (x)]2 } dx, (3.14)
0
when p and q are positive, q is continuous, and q has a continuous derivative. Then
the set
Cr := {f ∈ L2 [0, 1] : Ω(f ) ≤ r} (3.15)
is compact in X for each r > 0. The smoothing function Ω(f ) can be con-
sidered to be induced by a differential operator L, i.e., Ω(f ) = Lf 2, where
L∗ Lf = −(pf ) + qf on [0, 1] with the boundary conditions f (0) = f (1) = 0.
In particular, taking p = q = 1, it follows that Cr is a ball in the Sobolev space
W21 [0, 1]. Thus by restricting solutions to lie in a ball in W21 [0, 1] the problem is
no longer ill-posed. Similar results can be obtained using more general differential
operators and related spaces that are compactly embedded in X. The success of
the aforementioned approach hinges on the fact that the unit ball of W21 [0, 1] is a
compact set in the topology of L2 [0, 1] (note that it is, of course, not compact in the
topology of W12 [0, 1], for more details the reader is referred, e.g., to [2, 184, 185]).
3.3. Tikhonov’s regularization

Let L : D(L) ⊂ X → Z, where Z is a Hilbert space, be a closed linear operator
with dense domain and closed range. We first assume (see also [176, 247, 249, 250])
that N (L) is finite dimensional and that N (L) ∩ N (A) = {0}. We endow D(L)
with the topology induced by the graph norm |u| := (u2 + Lu2)1/2 . Then L
becomes a bounded operator on D(L). We define a new inner product on D(L) by
[u, v]L = (Au, Av)Y + (Lu, Lv)Z (3.16)
#
Then the induced norm uL := [u, u] is equivalent to the graph norm of u. Thus,
both A and L are bounded operators on D(L) equipped with the inner product
(3.16); we denote this Hilbert space by XL . The discussion shows that, under the
hypotheses listed above, without loss of generality we may restrict ourselves to the
case in which A : X → Y and L : X → Z are both bounded. For each y ∈ D(A† ),
there is a unique element x ∈ L(y) which minimizes Lu. Let A†L y := x, and
define
M := {x ∈ X : L∗ Lu ∈ N (A)⊥ }. (3.17)
Then it is not difficult to show that M is the orthogonal complement of N (A) with
respect to the inner product (3.16), so that A†L is the generalized inverse relative
to the decompositions:
X : N (A) ⊕ M, Y = R(A) ⊕ R(A)⊥ . (3.18)
Under the above assumptions for each α > 0, there exists a unique xα which
minimizes
Jα (x) := Ax − y2Y + αLx2Z . (3.19)
Furthermore,
xα = (A∗ A + αL∗ L)−1 A∗ y (3.20)
and xα converges to A†L y for y ∈ D(A† ) as α → 0, and diverges otherwise. Thus
it follows that in the presence of contamination, say y is replaced by yε , where
yε − y ≤ ε for some ε > 0, the norm of the corresponding xεα may well diverge.
It should be noted that (A∗ A + αL∗ L)−1 A∗ does not converge in the uniform
operator topology as α → 0; in fact, (A∗ A + αL∗ L)−1 A∗ → ∞. Furthermore,
the equation system (A∗ A + αL∗ L)x = A∗ y is poorly conditioned for small α,
and hence numerically unstable. Thus, both contamination and numerical approx-
imation dictate that a choice for α has to be made which would be a suitable
compromise between accuracy and stability. Several procedures for the choice of
“optimal” α are available, some of which take into consideration various a priori
information. A simple choice that works would be α = ε2 .
Strongly bounded noise in Tikhonov’s regularization. We come back to the situ-
ation where K : X → Y is a linear compact operator between the Hilbert spaces
X and Y . Consider the data y ∈ Y according to the equation y = Kx0 + η, where
η ∈ Y is the unknown noise and x0 ∈ X is an unknown element one wishes to re-
cover from the data y. We study Tikhonov’s regularization as a scheme to recover
x0 from the data y in the strong noise model
y = Kx0 + η with ηY ≤ ε . (3.21)
The interest is in what happens when ε → 0. It should be noted that, in the
Tikhonov regularization scheme, the unknown x0 is estimated by specifying the
solution x = xα,ε of the problem
minimize Kx − y2Y + αx2X over x ∈ X (3.22)
for some regularization parameter α, α > 0, yet to be specified. This procedure
dates back to [216, 249]. Its minimizer exists and is unique. Moreover, it is well-
known (see, e.g., [106]) that convergence rates on the error xα,δ − x0 X can be
obtained from a source condition. For simplicity, it is assumed there that there
exists a z0 ∈ X such that the “source condition”
x0 = (K ∗ K) ν/2 z0 for some 0 < ν ≤ 2 (3.23)
holds true. Precise necessary and sufficient conditions are given in [202]. In the
study of convergence rates under the source condition (3.23), it is assumed here
that ν is known and that α is chosen accordingly. Clearly, one wants to obtain
bounds on the error xα,ε − x0 X . As usual, this is broken up into two parts
xα,ε − x0 X ≤ xα,ε − xα,0 X + xα,0 − x0 X , (3.24)
where xα,ε is the “noiseless” estimator, i.e., the minimizer of Kx − y2Y + αx2X .
Thus, xα,ε − xα,0 is the noise part of the error and xα,0 − x0 is the error introduced
by the regularization.
The following results (see, e.g., [50, 106]) are well known:
(1) There exists a constant c such that for all α, 0 < α ≤ 1,
1
xα,ε − xα,0 X ≤ c α− 2 ηY . (3.25)
(2) Under the source condition (3.23), there exists a constant such that for all α,
0 < α ≤ 1,
xα,ε − x0 X ≤ c α ν/2 . (3.26)
As a consequence, the two results (3.25), (3.26) above then provide the fol-
lowing convergence rates:
Assuming the source condition (3.23) and the condition (3.21) on the noise
for α → 0 we have

xα,ε − x0 X = O α− 1/2 ε + α ν/2 . (3.27)
Moreover, if α ε 2/(ν+1) then

xα,ε − x0 X = O ε ν/(ν+1) . (3.28)
Weak noise in Tikhonov’s regularization. Tikhonov’s regularization may also be

considered (cf. [49, 52, 53]) as the scheme to recover x0 from the data y in the
weak noise model
y = Kx0 + η . (3.29)
Thus, we assume that there is a smoothing operator A such that the noise η and
A satisfy (3.8) and (3.9). In particular, η, AηY = ε2 , and the discussion (cf. [53])
focusses on what happens when ε → 0. Formally, Tikhonov regularization does not
depend on the noise being strongly or weakly bounded. Thus x0 is estimated by the
solution x = xα,ε of the problem (3.22) for some positive regularization parameter
α yet to be specified. Again we want to obtain bounds on the error xα,ε − x0 X ,
and it is broken up as xα,ε − x0 X ≤ xα,ε − xα,0 X + xα,0 − x0 X , where
xα,0 is the “noiseless” estimator, i.e., the minimizer of Kx − y2Y + αx2X . Thus,
xα,ε − xα,0 is the noise part of the error and xα,0 − x0 is the error caused by the
regularization. It is useful to introduce a new norm on X by way of
x2α,X = Kx2Y + α x2X . (3.30)
Assuming again the source condition (3.23) we see that the noiseless part xα,0 − x0
can be covered as before, but the treatment of the noise part is markedly different
from the case of strong noise (see [51]):
(1) Under the conditions (4.216), (3.9) on the noise η, there exists a constant C
depending on A only such that for α → 0
1
xα,ε − xα,0 2α,X ≤ C α− 4m ε . (3.31)
This leads to the convergence rates (3.32) and (3.33) specified below, where
it is shown in [50] that they are optimal, following arguments from [201], but
assuming, in addition, that A−m K has a continuous inverse.
(2) Assuming the source condition (3.23) and the conditions (4.216) and (3.9)
on the noise for α → 0, we have
1 1
xα,ε − x0 X = O α− 2 − 4m ε + α ν/2 . (3.32)
Moreover, if α ε4m/(2mν+2m+1) then

xα,ε − x0 X = O ε2mν/(2mν+2m+1) . (3.33)
3.4. Characterization of regularizers

Let A be of class L(X, Y ) with non-closed range. In what follows we introduce con-
cepts of regularizing families for the ill-posed problem Ax = y based on bounded
outer inverses of the operator A:
A linear operator B : Y → X is called an inner inverse of A if ABA = A. A
(nonzero) linear operator B : Y → X is called an outer inverse of A if BAB = B.
In what follows, B is always taken to be a non-zero operator. In the case of
Hilbert spaces, the regularizers will approximate least squares solutions of Ax = y.
In the case of Banach spaces, we assume that A is injective, and R(A) is dense in
Y (otherwise the regularizers would apply to the equation Ax = Qy, where Q is a
continuous projector of Y onto R(A), whose existence has to be assumed).
We will classify ill-posed in Banach spaces according to the type of regular-
izing families that they admit (cf. [185] [186]). In the case of Hilbert spaces this
classification will particularly distinguish the set of all compact operators with
infinite-dimensional range within the set of all bounded operators with non-closed
range.
Many of the operator-theoretic aspects of ill-posed linear equation (including
regularization and stabilization methods) are really problems in operator ranges
and operator factorizations. In particular, ranges of outer inverses play several
roles. First, it should be noted that very often the constructed approximation so-
lution is in the range of some outer inverse (or an approximate outer inverse).
Second, the problem of finding an “optimal” outer inverse with a prescribed rank
can be solved for several classes of operators. Third, outer inverses with a pre-
scribed finite-dimensional range can be easily constructed; this cannot always be
done in case the prescribed range is infinite dimensional and the outer inverse is re-
quired to be bounded. Fourth, outer inverses have desirable “stability/continuous
dependence” properties which inner inverses or the generalized inverse lack.
Our purpose is to introduce notions of regularizers in form of bounded outer
inverses with infinite-dimensional range (within this class, convergent regularizers
can be selected to provide “optimal” resolution), approximate outer inverses and
approximate right inverses in scales of norms.
Remark. The set of all operators in L(X, Y ) that have bounded outer inverses with
infinite-dimensional range and the set of full-rank m × n matrices share several
common properties: Each of them is both open and dense, and all elements of each
of the sets have outer inverses with the maximal possible rank (namely, the same
as the rank of A). These properties and other results to be analyzed next indicate
that, in Hilbert space, an equation involving a bounded non-compact operator with
non-closed range is “less” ill-posed than an equation with a compact operator with
infinite-dimensional range. In comparison with least squares or generalized inverse
problems for m×n matrices, one may say that for operators with non-closed range,
the case of a non-compact operator corresponds to the full-rank case for matrices,
while the case of a (nondegenerate) compact operator is the infinite-dimensional
analog of the rank-deficient case for matrices.
Outer inverses in “solvability” and “regularization” of ill-posed problems. If X is

of dimension m and Y is of dimension n, it follows from the property ABA = A
that the rank of any inner inverse of A cannot be less than the rank of A. Similarly
it follows from the property BAB = B that the rank of any outer inverse of A
cannot exceed the rank of A. Moreover, if r := rank A, then one can construct
outer inverses of rank s for any s ≤ r and inner inverses of rank t for any r ≤
t ≤ min(m, n) (see, e.g., [178]). Even in the finite-dimensional case the possibility
of approximating the generalized inverse (or the least squares solution of minimal
norm in Hilbert space settings) by an outer inverse of rank s < r is an attractive
feature of outer inverses. A similar approximation by inner inverses is of course
not possible.
The situation is more drastic in the infinite-dimensional case as can be seen

from the following known result: If the range of A ∈ L(X, Y ) is non-closed, then
A has no bounded inner inverse. In fact, if B is any inner inverse of A, then AB
and BA are linear idempotents, with N (BA) = N (A) and R(A) = R(AB). Thus
the following algebraic decompositions hold:
X = N (A)+̇R(BA), (3.34)
Y = R(A)+̇N (AB), (3.35)
where denotes the algebraic direct sum. Now, suppose B is bounded, then
the projectors BA and AB are continuous (equivalently, the decomposition in
(3.34) and (3.35) are topological) and so R(A) is closed, which contradicts the
assumption. Thus, we are led to the statement:
No regularizer can be an inner inverse.
The non-boundedness of the inner inverse of A in case of a non-closed range

of A ∈ L(X, Y ) is actually a part of the following known statement (see [194]):
Let A ∈ L(X, Y ), where X and Y are Banach spaces. Then A has a bounded
inner inverse B if and only if the decompositions (3.34) and (3.35) are topological.
Equivalently, A has a bounded inner inverse if and only if N (A) and R(A) have
topological complements in X and Y , respectively, (i.e., the projectors on N (A)

and R(A) are continuous).
If X and Y are Hilbert spaces, then A ∈ L(X, Y ) has a bounded inner inverse
if and only if R(A) is closed. It should be noted in all these cases that it does not
mean that all inner inverses are bounded, unless A is invertible.
Truncated singular value expansions as outer inverses (cf. [185]): Let H1 and H2
be Hilbert spaces and let K : H1 → H2 be a (nonzero) compact linear operator.
Let K ∗ denote the adjoint of K. Since K ∗ K is a non-negative symmetric compact
linear operator on H1 we have in terms of the singular
∞ system {σk ; xk , yk } for K
the following spectral representation K ∗ Kx = k=1 σk2 x, xk xk , where {xk } is
an orthonormal set of eigenvectors of K ∗ K with K ∗ Kxk = σk2 xk with σ1 ≥ σ2 ≥
· · · > 0. Set yk := σk−1 Kxk . Then the yk ’s form an orthonormal set in H2 , and it
is easy to show that
∞

Kx = σk x, xk yk . (3.36)
k=1
Obviously, the series (3.36) is the singular value expansion (SVD) of K (note
that, if rank(K) = r, then the number of non-zero singular values is r and the
summation in (3.36) extends from k = 1 to r).
From here on, we assume, unless stated otherwise, that the range K is infinite
dimensional. Then we have an infinite number of non-zero singular values with
σn → 0 as n → ∞.
Let m be a fixed positive integer and define the operator Bm : H2 → H1 by

m
Bm y := σk−1 y, yk xk . (3.37)
k=1
It then follows that

m
m
KBm y = σk−1 y, yk Kxk = y, yk yk (3.38)
k=1 k=1
and =m >

m
Bm K Bm y = σk−1 y, yi yi , yk xk
k=1 i=1
m
= σk−1 y, yk xk
k=1
= Bm y. (3.39)
†
Thus, for each m ∈ N, Bm is an outer inverse of rank m. For each
∞ y ∈−1D(K ) :=
⊥ † †
R(K) + R(K) , Bm y − K y → 0 as m → ∞, where K y = i=1 σi y, yi xi ,
but the convergence is not uniform and the operators Bm are not uniformly
bounded.
The SVD is particularly useful because it permits a quantification of the

notion of near rank deficiency. It is well known from linear algebra that for any
m × n real or complex matrix A of rank r and any k < r,
inf {A − BFrob : rank(B) = k} = A − Ak Frob , (3.40)
where Ak is the truncated singular value decomposition (with k terms) of A and
· Frob is the Frobenius norm (see, e.g., [99]). In view of the above observation
that the TSVD is an outer inverse and by use of the well-known relation between
the singular values and singular vectors of A and those of the Moore–Penrose of
A, it follows that
! "
inf A† − BFrob : BAB = B, rank(B) = k = A† − Bk Frob, (3.41)
where

k
Bk y = σi−1 y, yi xi . (3.42)
i=1
The same analysis and properties of outer inverses can be easily extended to
linear inverse problems with discrete data (such as those that arise from moment
discretization of ill-posed linear integral and operator equations in Hilbert space
(see, e.g., [24, 178]) or when projection methods on finite-dimensional subspaces
are used (see, e.g., [105, 181]).
Outer inverses are not “equation solvers” (cf. [185]), i.e., if B is an outer
inverse for A, which is not also an inner inverse, then for y ∈ R(A), x := By
is not a solution to Ax = y, and for y ∈ / R(A), x := By is not a least squares
solution (in the case of a Hilbert space). One finds in some books statements like:
“Since almost every application of various generalized inverses involves subsets of
1-inverses (inner inverses), we will mainly consider inner inverses that satisfy addi-
tional conditions. . . ”. For ill-posed problems (see [185]), the situation is precisely
the opposite. We are not interested in a generalized or inner inverse that would be
an “equations solver”, since such an “inverse” will be unbounded. Rather we seek
a bounded operator that has some “inverse-like” properties, and that can serve
simultaneously as an “approximate inverse” and stabilizer to the inverse problem.
Indeed, outer inverses possess these qualities:
(a) If B is an outer inverse of A, then B is also an inner inverse of Ã := A|R(B).
(b) For all y ∈ R(AB), x := By is the unique solution in M = R(B) of the
equation Ax = y.
From these properties (see [185] for more details), it follows that the unique
solution of Ax = y in R(B) is a “regularized” solution and can be constructed in a
stable way. If R(BA) or R(B) is infinite-dimensional, then we have in a sense the
possibility of “infinite resolution”, and the equation with bounded outer inverses
of infinite rank are not as ill posed as those for which an outer inverse with infinite
rank does not exist.
Approximate outer and approximate right inverses. The concept of a regular-

izer plays an important role in obtaining an approximate solution of an ill-posed
problem. Let A : X → Y be a one-to-one mapping (not necessarily linear). In
the already known definition by Tikhonov (see e.g. [170, 249]) a regularizer is a
bounded “approximate” inverse or “approximate” generalized inverse. Any regu-
larizer must also satisfy the defining equations of an inner as well as outer inverse
approximately, but it cannot satisfy the defining equation of an inner inverse ex-
actly. This motivates considering regularizers (cf. [185]) that satisfy the defining
equation of an outer inverse and, in addition, have “maximum” rank.
A family F of regularizers by outer inverses for the ill-posed problem Ax = y
is said to be of type I if each B ∈ F satisfies the following conditions:
(i) B ∈ L(Y, X)
(ii) BAB = B,
(iii) the range of B is infinite-dimensional.
An ill-posed problem that does not admit a family of regularizers of type I is
said to be of type II.
Any B ∈ F is called a regularizer. Of course, u := By is not necessarily a good
approximation to the “solution” x of Ax = y for each B ∈ F , just like, say, (A∗ A+
αI)−1 A∗ y is not necessarily a good approximation for each α > 0. As already
mentioned, every “regularization method” involves a critical “parameter”, whose
optimal value – or at least a suitable choice of it – is crucial to the approximation
of the solution. Thus criteria or strategies have to be developed for selecting a
suitable or “optimal” outer inverse from our class of regularizers. We will not
address this problem here. Instead we focus only on the operator-theoretic aspects
of outer inverses as a class of regularizers.
As an example, we note that Tikhonov’s regularization provides an approxi-
mate outer inverse, but not an outer inverse. Let
Bα := (K ∗ K + αI)−1 K ∗ , α > 0. (3.43)
Then
Bα − Bα KBα = α(K ∗ K + αI)−1 Bα , α > 0. (3.44)
Note that
Bα − Bα KBα → 0, α → 0. (3.45)
Regularization operators obtained via spectral families (such as those in [19, 59,
104, 179]) are usually not outer inverses. The same is true for the “regularized”
truncated SVD.
Tikhonov’s regularization operator and other regularizers obtained by spec-
tral families are approximate outer inverses in the following sense:
A ∈ L(X, Y ) is approximately outer-invertible if, for each μ ∈ (0, 1), there
exists a Bμ ∈ L(Y, X) with the following properties:
(Bμ ABμ − Bμ )y ≤ C(μ)Bμ y (3.46)
and
Bμ y ≤ Γμ y for all y ∈ Y. (3.47)
Each such Bμ is called an approximate outer inverse of A. Properties of C(μ), Γμ
and Bμ are to be prescribed for convergence analysis.
For a compact linear operator K : H1 → H2 , the filtered truncated SVD

r
σi
Br y = y, yi xi , α>0 (3.48)
σ2
i=1 i
+α
is an approximate outer inverse, but not an outer inverse (choose μ = r−1 ).

It is also useful to introduce a notion of “approximate right-invertibility”,
which abstracts some characteristics of regularization methods:
A ∈ L(X, Y ) is called approximately right-invertible if, for each μ ∈ (0, 1),
there exists a norm · μ on X and a Bμ : Y → X such that for all y ∈ Y and all
x ∈ X,
ABμ y − y ≤ μ y (3.49)
Bμ yμ ≤ Γ(μ) y, (3.50)
and
xμ → x as μ → 0+ . (3.51)
Each such Bμ is called an approximate right inverse of A. Properties of C(μ) and
Γμ and Vμ are to be prescribed for convergence analysis.
Each such Bμ is called an approximate right inverse of A with a bound Γ(μ)
(note that Bμ need not be linear). For regularization one requires Γ(μ) = O(μ−γ )
or a similar behaviour. Again μ y in (3.49) may be replaced by C(μ) y.
For regularizers of type I we have bounded outer inverses with infinite-
dimensional range. For ill-posed problems for which such regularizers do not exist,
we may use approximate outer inverses as regularizers. Again, these can only be
approximate inner inverses or approximate right inverses.
Characterizations of ill-posed problems. Next we deal with characterizations of
ill-posed problems of so-called type I and II: Let Out(L) denote that set of all
(nonzero) outer inverses to a (nonzero) linear transformation L : V → W , where
V and W are vector spaces over the same field. The following proposition is im-
mediate:
Let L : V → W be a (nonzero) linear transformation. Then the following
statements are equivalent for any (nonzero) linear transformation M : W → V
(where again denotes algebraic direct sum):
(a) M ∈ Out(L).
(b) M L is idempotent and V = R(M ) N (M L).
(c) LM is idempotent and W = N (M ) R(LM ).
(d) LM is idempotent and R(M ) ∩ N (L) = {0}.
The next proposition establishes the existence of (algebraic) outer inverses

with a prescribed range:
Let V1 be a subspace of V such that V1 ∩ N (L) = {0} and W2 be an algebraic
complement of LV1 . Then there exists an outer inverse M to L such that R(M ) =
V1 and N (M ) = W2 . Under these conditions M |LV1 = (L|V1 )−1 .
We now consider bounded outer inverses of A ∈ L(X, Y ), A = 0, and their
connections with certain topological complements. A topological direct sum will
be denoted by ⊕. We are again interested in conditions under which there exists
a bounded outer inverse B with a range and a null space prescribed.
Let A : X → Y be a bounded linear operator with non-closed range. Then the
ill-posed problem Ax = y is of type I if and only if the following two conditions
hold:
(a) R(A) contains a closed infinite-dimensional subspace, say M .
(b) N (A) is (topologically) complemented in the subspace A−1 (M ), the inverse
image of M under A.
In the case of Hilbert spaces, these characterizing conditions take an explicit
and simple form.
Let A be a bounded linear operator on a Hilbert space H1 into a Hilbert
space H2 , and let the range of A be non-closed. Then the following statements are
equivalent:
(i) The ill-posed problem Ax = y is of type I,
(ii) R(A) contains a closed infinite-dimensional subspace,
(iii) A is not compact.
An operator A ∈ L(X, Y ) is strictly singular if the subspaces Z ⊂ X for with
the restriction A|Z has a bounded inverse on AZ, the image of Z under A, are
necessarily finite dimensional (see, e.g., [136]). This notion of a strictly singular
operator is clearly the right generalization of a compact operator in Hilbert space.
It is natural to ask if it is possible to characterize ill-posed problems of type I
in Banach spaces by the condition that “the operator is not strictly singular”.
However, this is not possible (cf. [185, 186]).
Remarks. Finally some comments should be made (following Nashed [185]):
(i) For various aspects of operator factorizations and operator ranges, [184] has
initiated the study of bounded (or densely defined closed) linear operators
which have bounded outer inverses of infinite rank within the framework of
operator factorization and operator ranges.
(ii) The classification of ill-posed linear problems as proposed here provides also
a classification of ill-posed non-linear problems based on properties of outer
inverses of the Fréchet or Hadamard derivative of the non-linear operator (cf.
[174]). Approximate outer inverses have been used by B.D. Craven and M.Z.
Nashed [36] in the context of inverse function theorems when the derivative
does not have a bounded inverse of a bounded generalized inverse.
(iii) Examples of operator equations with a non-compact bounded linear operator

with non-closed range arise from various integral and convolution operators
on the half-line and for certain generalized Wiener–Hopf operators. They
also arise if zero belongs to the continuous spectrum of a linear, bounded
selfadjoint and non-negative operator on a Hilbert space.
(iv) For various aspects of construction and computation of outer inverses of
a prescribed rank s ≤ r for a matrix of rank r, see [23] and several pa-
pers and the annotated bibliography in [177]. For constructions of outer in-
verses of bounded linear operators in Banach spaces the reader is referred to
[165, 166, 179]. For stability properties for outer inverses, see [184–186]. For
perturbation analysis of outer inverses, see [166, 179, 185, 186]. Convergence
analysis of regularization operators hinges upon perturbation and stability
properties of outer inverses (see [185, 186]).
4. Reconstruction methods and regularizing filters

Assume that A is of class L (X, Y ), X, Y Hilbert spaces. Let {Rt }t>0 be a family
of continuous operators (regularizers) from Y to X with Rt 0 = 0. If there exists
a mapping α : (0, ∞) × Y → (0, ∞), such that, for all elements y ∈ R (A) and
regularization parameter t = α(ε, y ε ),
!- - " ε→0
sup -A† y − Rα(ε,yε ) y ε - : y ε ∈ Y with y − y ε ≤ ε −→ 0, (4.1)

then the pair {Rt }t>0 , α is called a regularization method (or simply a regular-
ization) of A† .
If the sequence {Rt }t>0 is a subset of L (Y, X), then the regularization
{Rt }t>0 , α is called linear. The mapping α is known as the parameter choice
provided that
ε→0
sup {α (ε, y ε ) : y ε ∈ Y with y − y ε ≤ ε} −→ 0. (4.2)
If α is only dependent on ε, i.e., α (ε, y ε ) = α (ε), we speak of an a priori
parameter choice. If α is dependent on ε and y ε , i.e., α = α (ε, y ε ) , it is called an
a posteriori parameter choice.
The limit relation (4.1) can be equivalently written in the form
!- - " ε→0
sup -x − Rα(ε,yε ) y ε - : y ε ∈ Y with Ax − y ε ≤ ε −→ 0 (4.3)
⊥
for all x ∈ N (A) . A direct consequence of a regularization method is that the
limit relation - -
lim -A† y − Rα(ε,y) y - = 0 (4.4)
ε→0
holds true for all y ∈ R(A) (note that, in Equation (4.4), the regularization is
applied to y instead of y ε ). It is usual to collect all regularization parameters that
are relevant in the limit relation (4.4) in a set Γ = {α (ε, y) : ε > 0, y ∈ R (A)} .
Because of the fact that limε→0 α (ε, y) = 0 the set Γ possesses an accumulation
- -
point at 0. Moreover, limΓλ→0 -A† y − Rλ y - = 0 for all y ∈ R (A). All in all, we
are led to the following result:

Assume that A is of class L (X, Y ) . Let {Rt }t>0 , α be a regularization
method of A† . Then, the subfamily {Rλ }λ∈Γ converges pointwise in R (A) to A†
for λ → 0.
As a direct consequence we obtain the statement (see, e.g., Louis [148], Rieder
[227]):

Assume that A is of class L (X, Y ) . Let {Rt }t>0 , α be a regularization of
A† . If R (A) = R (A), then {Rt }t>0 is unbounded.
As already known, for a linear regularization, the reconstruction error
A† y − Rt y ε can be split by use of a family {Rt }t>0 in the following way:
- † - - † -
-A y − Rt y ε - ≤ -A y − Rt y - + Rt y − Rt y ε

approximation error data error
≤ Rt y − y ≤ Rt · ε.
ε
(4.5)
Once again, it should be pointed out, that both parts of the reconstruction
error, i.e., the approximation error and the data error, exhibit an opposed behavior
in limit considerations for t → 0 and t → ∞.
Hence, an essential task is to search for a value topt that balances both er-
rors. In fact, the regularization parameter should be chosen in such a way that
α (ε, y ε ) ≈ topt .
Classification of regularization methods. The question (cf. [148]) arises how to
classify regularization methods. Since all methods need to converge as ε → 0, an
obvious criterion of specifying their characteristics is the speed of the convergence,
i.e., we make the attempt to introduce a classification with respect to the speed
of convergence of the total error:
!- - " ε→0
sup -A† y − Rα(ε,yε ) y ε - : y ∈ R(A), y ε ∈ Y with y − y ε ≤ ε −→ 0. (4.6)
Unfortunately, it turns out that the speed of the convergence is arbitrarily slow
for all regularization methods.
Let A be of class L(X, Y ). Assume that R(A) = R(A). Let ({Rt }t>0 , α) be
a regularization of A† . Then there exists no function h : [0, ∞) → [0, ∞) with
limε→0 h(ε) = 0 such that
!- - "
sup -A† y − Rα(ε,yε ) y ε - : y ∈ R(A), y ≤ 1, y ε ∈ Y with y − y ε ≤ ε ≤ h(ε).
(4.7)
Powers of absolute values and smooth Hilbert spaces. Next our interest is to show
that the concept of smoothness can be used for the classification of regularization
methods. An auxiliary tool is the absolute value |A| of the operator A. In order
to motivate the setting |A| we start with the explanation of a functional calcu-
lus for compact operators which also helps us to introduce filters for purposes of
regularization in the next subsection. Even better, the functional calculus for com-
pact operators enables us to reduce spectral features to a study in terms of real
functions.
Let A be of class K (X, Y ) with the singular system {σj ; xj , yj }j∈N . Let
0 1
Φ : [0, ∞) → R be a piecewise continuous function defined on the interval 0, A2 .
Then we understand the operator Φ (A∗ A) to be given as
∞

Φ (A∗ A) x = Φ σj2 x, xj xj + Φ (0) PN (A) x. (4.8)
j=1
0 Note
1 that the series on the right side of (4.8) is convergent on the interval
0, A2 .
Let A ∈ K(X, Y ) be a compact operator with the singular system
{σj ; xj , yj }j∈N .
Suppose that Φ : [0, ∞) → R is piecewise continuous. Then the following properties
hold true:
(a) A = σ1 , i.e., the norm coincides with the largest singular value of A,
(b) Φ(A∗ A)A∗ = A∗ Φ(AA∗ ), (4.9)
(c) Φ(A A) = sup |Φ(σj2 )| ≤ sup |Φ(λ)|,
∗
(4.10)
j∈N 0≤λ≤A2
√
(d) Φ(A∗ A)A∗ = A∗ Φ(AA∗ ) = sup σj |Φ(σj2 )| ≤ sup λ |Φ(λ)| .
j∈N 0≤λ≤A2
(4.11)
Example. If Φ = 1, then it is clear that
∞

Φ (A∗ A) x = x, xj xj + PN (A) x = PR(A∗ ) x + PN (A) x = PX x = x. (4.12)
j=1
This explains the occurrence of the term Φ (0) PN (A) x in (4.8).

Example. If Φ (t) = t1/2 , t ≥ 0, then
∞
∞
1/2
(A∗ A) x= σj x, xj xj + Φ(0) PN (A) x = σj x, xj xj , (4.13)

j=1 j=1
=0
The operator |A| = (A∗ A)1/2 is called the absolute value of A. Analogously,
∗
|A | is given by
∞

1/2
|A∗ | y = (AA∗ ) y= σj y, yj yj , y ∈ Y. (4.14)
j=1
It is easy to see that

|A| x2 = |A|x, |A|x = A∗ Ax, x = Ax, Ax = Ax2 . (4.15)
Remark. Our notation (4.13) can be used to introduce fractional powers (A∗ A)μ ,
μ ≥ 0, of A∗ A:
∞

|A|2μ x = (A∗ A)μ x = σj2μ x, xj xj . (4.16)
j=1
Let X, Y be Hilbert spaces. Suppose that A is of class K(X, Y ). Then

(1) R(A∗ ) = R(|A|) = R (A∗ A)1/2 ,

(2) R(A) = R(|A∗ |) = R (AA∗ )1/2 .
Next we introduce subspaces of the Hilbert space X involving on the concept
of powers of the absolute value |A| of the operator A: For ν ≥ 0, let Xν be
defined by
ν ν ⊥
Xν = R (|A| ) = |A| x : x ∈ N (A) . (4.17)
Then the following properties can be verified by straightforward arguments:
⊥
(1) Xν ⊂ N (A) ⊂ X for all ν ≥ 0,
(2) Xν ⊂ Xμ for ν > μ ≥ 0,
⊥
(3) X0 = N (A) .
In connection with ∞

x = |A|ν z = σkν z, xk xk (4.18)
k=1
we are able to impose the following norm on Xν :
∞
∞
|z, xk | 2ν |x, xk |
2 2
2 2
xν = z = σ = . (4.19)
σk2ν k
σk2ν
k=1 k=1
Alternatively, the spaces Xν can be characterized by the norms (4.19), i.e.,

⊥
Xν = x ∈ N (A) : xν < ∞ . (4.20)
The spaces Xν impose conditions on the smoothness of the elements x ∈ X.
After these preliminaries about powers of absolute values we come back to the
discussion of the speed of the convergence of a regularization method involving the
concept of smoothness, i.e., the solution of an operator equation will be assumed
to be a member of a subspace Xν , ν > 0, of X.
Indeed, in a large number of ill-posed problems (A; X, Y ) the operator A
shows the property that the image Ax is smoother than x. In concrete situations
this leads us to functions xk in the singular system with strongly growing oscilla-
tions for increasing k. Looking at the higher frequency parts of an element x ∈ X,
i.e., the inner products x, xk for large k, we notice that there is a damping effect
on Ax by the factor σk . Hence, the norm xν can be interpreted in the sense
that x, xk /σkν → 0 for k → ∞ is demanded. In addition, the larger the value of
ν is chosen, the faster x, xk has to converge to 0. Therefore it can be concluded
that the element x does not contain relevant high frequency components. In other
words, x can be regarded as “smooth”.
Optimality of reconstruction methods. In the following, a stable reconstruction

method for the solution of the operator equation involving A ∈ L(X, Y ) is un-
derstood to be a continuous (not-necessarily linear) mapping T : Y → X with
T 0 = 0. The question is which reconstruction error does occur in the best worst
case, if the data are noisy.
The worst case error of a reconstruction method T for A corresponding to
the noise level ε and the additional information A† yν ≤ ρ is defined as
! "
Eν (ε, ρ, T ) = sup T y ε − A† y : y ∈ R(A), y ε ∈ Y, y − y ε ≤ ε, A† yν ≤ ρ .
(4.21)
Another expression (avoiding the occurrence of the operator A† ) can be formu-
lated by
Eν (ε, ρ, T ) = sup {T y ε − x : x ∈ Xν , y ε ∈ Y, Ax − y ε ≤ ε, xν ≤ ρ} .
(4.22)
Clearly, the smaller the worst case error, the better the reconstruction method.
The best worst case error for A corresponding to the noise level ε and the
additional restriction A† yν ≤ ρ is understood to be
Eν (ε, ρ) = inf {Eν (ε, ρ, T ) : T : Y → X continuous, T 0 = 0} . (4.23)
Note that the best worst case error, i.e., Eν (ε, ρ), depends on the problem
(i.e., on the operator A), but not on the reconstruction method.
The next result tells us about the quantity of the best worst case error.
Let A be of class L(X, Y ). Then we have
Eν (ε, ρ) = eν (ε, ρ), (4.24)
where
eν (ε, ρ) = sup {x : x ∈ Xν , Ax ≤ ε, xν ≤ ρ} . (4.25)
Clearly, our results characterize the best worst case error independently of
the knowledge of a specific reconstruction method.
Let A be of class L(X, Y ). Then, for ν > 0,
1 ν
eν (ε, ρ) ≤ ρ ν+1 ε ν+1 . (4.26)
Furthermore, there exists a sequence {εk }k∈N with εk → 0 for k → ∞ such that
1 ν
eν (ε, ρ) = ρ ν+1 ε ν+1 . (4.27)
In other words, the estimate (4.26) is sharp, i.e., it cannot be improved.
Regularizing filters. If A is an injective operator of class K(X, Y ), then A† can
be expressed in the form (A∗ A)−1 A∗ . The non-continuity is caused by the term
(A∗ A)−1 , that has to be stabilized. In connection with the functional calculus for
compact operators we are therefore led to filters as appropriate tools for regular-
ization.
. /
2
Let {Ft }t>0 , Ft : 0, A → R be a family of piecewise continuous functions
satisfying the conditions
/
2
(F1) limt→0 Ft (λ) = λ1 for all λ ∈ 0, A ,
. /
2
(F2) λ |Ft (λ)| ≤ CF for all λ ∈ 0, A and t > 0.
Then the family {Ft }t>0 is called a filter relative to A.
By virtue of Condition (F1), Ft (A∗ A) becomes a continuous operator, which
converges in pointwise sense to (A∗ A)−1 as t → 0. This is the reason why we let
Rt y = Ft (A∗ A) A∗ y, y ∈ Y. (4.28)
As a consequence, {Ft }t>0 filters the influence of small singular values of A on the
operator Rt in (4.28).
In terms of the singular system {σk ; xk , yk }k∈N we are able to write
∞

Ft (A∗ A) A∗ y = Ft σk2 σk y, yk xk + Ft (0) PN (A) A∗ y
k=1
∞

= Ft σk2 σk y, yk xk (4.29)
k=1
⊥
due to fact that PN (A) A∗ y = 0 (note that A∗ y ∈ R (A∗ ) = N (A) ). Considering
the approximation error we obtain, for y ∈ R(A), that
A† y − Rt y = A† y − Ft (A∗ A) A∗ y
= A† y − Ft (A∗ A) A∗ AA† y
= (I − Ft (A∗ A) A∗ A) A† y
= pt (A∗ A) A† y, (4.30)
. /
2
where the function pt : λ → pt (λ) , λ ∈ 0, A is given by
. /
2
pt (λ) = 1 − λFt (λ) , λ ∈ 0, A . (4.31)
The identity (4.30) leads us to the formulation of the following result:
Assume that A ∈ K (X, Y ). Let {Ft }t>0 be a filter. Then

A† y y ∈ D A† ,
lim Rt y = (4.32)
t→0 ∞ y∈/ D A† ,
where
Rt y = Ft (A∗ A) A∗ y
for y ∈ Y .
The next result concretizes the stability of Rt y ε under the noise level ε in
more detail:
Let {Ft }t>0 be a filter so that (F1), (F2) hold true. For y, y ε ∈ Y with
y − y ε ≤ ε, set xt = Rt y and xεt = Rt y ε . Then, for the residual term, we
have
Axt − Axεt ≤ CF ε, (4.33)
while, for the error term, it follows that
#
xt − xεt ≤ ε CF M (t)
M (t) = sup |Ft (λ)| . (4.34)
0≤λ≤A2
† †
Let
us continue with the estimate of the total error, thereby using x = A y,
†
y ∈ D A . We base our considerations on the usual splitting into the approxi-
mation error and the data error in the form
- † - - -
-A y − Rt y ε - ≤ -A† y − Rt y - + Rt y − Rt y ε
- -
= -x† − xt - + xt − xεt
t→0 - - #
≤ -x† − xt - + ε CF M (t). (4.35)

→0
†
Obviously, the approximation error x − Rt y tends to 0. The discussion of the
data error is much more problematic than the approximation error: From the
2
limit relation limt→0 Ft (λ) = λ1 imposed on a filter within the interval [0, A ] it
t→0
follows that M (t) −→ ∞. Therefore, for the total error, we are confronted with
a divergent behavior as t → 0. Nevertheless, convergence properties can be forced
by a suitable coupling of t and ε. In fact, it can be deduced that, under the a
priori parameter choice indicated above, the approximation error as well as the
data error converge to 0, if the noise level ε tends to zero (cf. [227]):
Let {Ft }t>0 be a filter. If we choose α : (0, ∞) → (0, ∞) , ε → α(ε), such
ε→0 # ε→0
that α (ε) −→ 0 as well as ε M (α (ε)) −→ 0, then {Rt }t>0 , α with Rt =
Ft (A∗ A) A∗ is a regularization of A† (by convention, {Ft }t>0 is called a regular-
izing filter).
Next we mention which additional requirement imposed on filters Rt =
Ft (A∗ A)A∗ will be necessary to guarantee the order optimality.
Let {Ft }t>0 be a regularizing filter for A ∈ L(X, Y ). Assume there are t0 > 0,
μ > 0, and a function ωμ : (0, t0 ] → R such that
sup λμ/2 |pt (λ)| ≤ ωμ (t) (4.36)
0≤λ≤A2
for all t ∈ (0, t0 ] (with pt (λ) = 1 − λFt (λ)). Let y ∈ R(A) and let x† = A† y in Xμ
with x† μ ≤ . Then the following estimates
(a) x† − xt X ≤ ωμ (t),
(b) Ax† − Axt Y ≤ ωμ+1 (t)
hold true for xt = Rt y = Ft (A∗ A)A∗ y and 0 < t < t0 . Moreover, let
μ
(1) ωμ (t) ≤ Cp t 2 for t → 0,
(2) M (t) = sup |Ft (λ)| ≤ CM t−1 for t → 0,
0≤λ≤A2
where μ > 0, Cp , CM > 0 are constants. Let the a priori parameter choice α :
(0, ∞) → (0, ∞) fulfill
μ+1
2 μ+1
2
ε ε
C1 ≤ α(ε) ≤ C2 , ε → 0, (4.37)
ρ ρ
where C1 , C2 are positive constants. Then, ({Rt }t>0 , α), Rt = Ft (A∗ A)A∗ , is an
order optimal regularization for A† with respect to Xμ .
Of course, we need to know the values ρ and μ to guarantee the order opti-
mality. Without the availability of ρ but based on the knowledge of μ, we are led
2
to α(ε) = Cε μ+1 with C being a positive constant to find an order optimal regu-
larization. Without any information of both parameters ρ as well as μ we have to
deal with a posteriori parameter choices.
An asymptotic behavior of ωμ determines the speed of convergence for the
reconstruction error. An important feature is the so-called qualification.
Let {Ft }t≥0 be a regularizing filter for A† , where A ∈ L (X, Y ) satisfying the
asymptotic relation
CM
M (t) ≤ , t → 0. (4.38)
t
where CM is a positive constant. The maximal value μ0 , such that there exists,
for all values μ ∈ (0, μ0 ] , a constant Cp > 0 satisfying
sup λ μ/2 |pt (λ)| ≤ Cp t μ/2 , t → 0, (4.39)
is called the qualification of the filter (remember pt (λ) = 1 − λFt (λ)).
In other words, the qualification can be regarded as the maximal rate of
decay. If the qualification is finite, there exists a parameter choice α (ε) such that
- † -
-A y − Rα(ε) y ε - = O ε μ0 /(μ0 +1) , ε → 0, (4.40)
holds true for A† y ∈ Xμ0 . If the qualification is infinite, there exists a parame-
ter choice α (ε) for which the error decay comes arbitrarily close to O (ε). As a
consequence, filters with infinite qualification are more advantageous than others.
Order optimality of special regularizing filters. In the following we recover impor-
tant examples of regularization methods constituted by filters, i.e., the truncated
singular value decomposition SVD and the Tikhonov–Phillips regularization.
Truncated singular value decomposition revisited: The SVD of A† is
∞
1
A† y = y, yk xk .
σk
k=1
We choose the filter 1

λ, λ ≥ t,
Ft (λ) = (4.41)
0, λ < t.
Then
∞

1
Rt y = Ft (A∗ A) A∗ y = Ft σk2 σk y, yk xk = y, yk xk (4.42)
√ σk
k=1 σk ≥ t
is the truncated SVD of A† (with finitely many summands).

Clearly we have
(1) limt→0 Ft (λ) = λ1 for all λ > 0,
(2) CF = sup0≤λ≤A2 λ |Ft (λ)| = 1 for all t ≤ A2 ,
(3) M (t) = sup0≤λ≤A2 |Ft (λ)| = 1t for all t > 0.
The total error can be described as follows:
- † - - -
-A y − Rt y ε - ≤ -A† y − Rt y - + Rt y − Rt y ε
- -
- -
- 1 - #
= - - y, yk xk - + ε CF M (t)
√ σk
-
- σk < t
-
⎛ ⎞ 1/2
Parseval
1 2 ε
= ⎝ |y, yk | ⎠ +√ . (4.43)
√ σk2 t
σk < t
2
Moreover, for all μ > 0 and 0 ≤ t ≤ A , we have
sup λ μ/2 |pt (λ)| = sup λ μ/2 |1 − λFt (λ)| = sup λ μ/2 = t μ/2 .
0≤λ≤A 2
0≤λ≤A 2 0≤λ≤t
Together with M (t) = 1t we find that this filter possesses an infinite qualification,
and it is order optimal for all μ > 0. However, one can show that the TSVD is not
optimal for any μ > 0.
Asymptotic regularization revisited. We choose the following filter
⎧
⎨ 1−exp(− λt ) λ>0
Ft (λ) = λ (4.44)
⎩ 1
λ = 0.
t
Then we obtain
2
∞
∞ 1 − exp − σk

t
Rt y = Ft σk2 σk y, yk xk = y, yk xk .
σk
k=1 k=1
It is not difficult to prove the following properties:
(1) limt→0 Ft (λ)= λ1 for λ > 0

1 − exp − λt , λ>0
(2) λ |Ft (λ)| = λ .≤1 , t > 0,
t, λ=0
(3) sup0≤λ≤A2 λ |Ft (λ)| = 1 = CF ,

(4) M (t) = sup0≤λ≤A2 |Ft (λ)| = 1t , t > 0 (note that Ft (λ) is monotonously
decreasing in λ and limλ→0 Ft (λ) = 1t ).
The qualification of the filter again is infinite.
Tikhonov’s regularization revisited. Using the filter
1
Ft (λ) = , t > 0, (4.45)
λ+t
we find that
∞
σk
Rt y = Ft (A∗ A) A∗ y = y, yk xk (4.46)
σ2 +t
k=1 k
and
∞
∞
σk
(A∗ A + tI) Rt y = 2 y, yk σk2 + t xk = σk y, yk xk (4.47)
σk + t
k=1 k=1
∞ ∞
∞

= y, σk yk xk = y, Axk xk = A∗ y, xk xk = A∗ y,
k=1 k=1 k=1
i.e., xt = Rt y is the unique solution of the equations (A∗ A + tI) xt = A∗ y. Such

equations are called regularized normal equations. It can be easily seen that
(1) limt→0 Ft (λ) = λ1 , λ > 0,
(2) λ |Ft (λ)| = λ+t
λ
≤ 1 = CF for all t > 0,
1
(3) M (t) = sup0≤λ≤A2 |Ft (λ)| = sup0≤λ≤A2 λ+t = 1t , t > 0.
In order to compute the qualification for Tikhonov’s regularization we take a look
at the term
λ μ2
μ μ
sup λ 2 |pt (λ)| = sup t2 λ
t
. (4.48)
0≤λ≤A 2
0≤λ≤A 2 +1

t
=hμ (λ,t)
In fact, we are led to distinguish two cases:

Case 1: For μ > 2: hμ (λ, t) is strictly monotonously increasing in λ.
Case 2: For μ ≤ 2: we find the estimate
μ
z2
μ
sup hμ (λ, t) ≤ t sup 2 . (4.49)
0≤λ≤A2 0≤z<∞ + 1
z

=Cp <∞
Summarizing our considerations we obtain

μ
μ Cp t 2 , : 0 < μ ≤ 2,
sup λ 2 |pt (λ)| ≤ (4.50)
0≤λ≤A2 Aμ−2 t, : μ > 2.
In other words, the qualification of the Tikhonov filter is μ0 = 2, i.e., we arrive

at the order optimality and even at the optimality for 0 < μ ≤ 2, if we use the
μ+1
2
parameter choice α(ε) = μ1 ρε .

Morozov’s discrepancy principle. Let y ∈ D A† and y ε ∈ Y with y − y ε < ε
and xεt = Ft (A∗ A) A∗ y ε . The idea of the discrepancy principle can be explained
as follows: Choose the parameter α = α (ε, y ε ) such that
Axεt − y ε ≈ ε. (4.51)
ε
In other words, the residual or discrepancy of y is assumed to have the order of
the data error.
In order to realize the assumption (4.51) we consider the defect function
d : t → d(t) = Axεt − y ε . (4.52)
It is not hard to see that
- -
d (t) = -AFt (A∗ A) A∗ y ε − y ε -
- -
= -pt (AA∗ ) y ε -
∞
2 ε - - 1/2
ε -2
= 2 2 -
pt σk |y , yk | + pt (0) PN (A ) y
∗ , (4.53)

k=1
=1
so that - - - -
limd (t) = -PN (A∗ ) y ε - = -PR(A)⊥ y ε -. (4.54)
t→0,t>0
- -
If y ∈/ R (A), then -PR(A)⊥ y ε -Y can be arbitrarily large. However, if y ∈
R (A), then we have
- - - -
-P ε-
= -PR(A)⊥ (y − y ε ) - ≤ y − y ε ≤ ε
R(A)⊥ y (4.55)
and, therefore,
lim d (t) ≤ ε (4.56)
t→0,t>0
In other words, for all τ > 1 exists a t0 with d (t) < τ ε for all t ≤ t0 .
Let τ > 1 be chosen (fixed) and {tk } be a strictly monotonously decreasing
sequence with limit zero. Determine k ∗ , such that
d (tk∗ ) ≤ τ ε ≤ d (ti ) , i = 1, . . . , k ∗ − 1 (4.57)
ε
Set α (ε, y ) = t .
k∗
Let A be of class L (X, Y ) . Suppose that {Ft }t>0 is a regularizing filter with
qualification μ0 > 1. Moreover, assume that M (t) ≤ CtM for t → 0. Let the param-
eter choice α : (0, ∞) × Y → (0, ∞) be taken in accordance with the discrepancy
principle, such that the sequence {tk }k satisfies tk = θk tk−1 , where 0 < ϑ ≤ θk < 1
for all k. Furthermore, suppose that τ > sup{|pt (λ)| : t > 0, 0 ≤ λ ≤ A2 } ≥
pt (0) = 1.
Then {Rt }t>0 , γ with Rt = Ft (A∗ A) A∗ is an order optimal regularization
of A† with respect to Xμ for all (0, μ0 − 1].
Note that, for the values μ ∈ (μ0 − 1, μ0 ] we obtain nothing, whereas, the
1 2
Tikhonov case leads to the order of convergence O(ε 2 ) instead of O(ε 3 ). This is
the reason why the discrepancy principle may produce suboptimal rates of con-
vergence in case of a finite qualification. However, in connection with an infinite
qualification, no such problems arise.
Generalized discrepancy principle: Next we are interested in a generalization of
the discrepancy principle. To this end we make a reformulation: Let {tk } be a
strictly monotonously decreasing sequence with limit zero. Then, tk∗ is chosen as
follows:
- -2
tk∗ = sup tk : -Axεtk − y ε - ≤ τ ε2
- -2
= sup tk : -ptk (AA∗ ) y ε - ≤ τ ε2
! ? @ "
= sup tk : y ε , p2tk (AA∗ ) y ε ≤ τ ε2 . (4.58)
In doing so we have used the function st = p2t .
For a generalization we allow arbitrary functions st in the following sense:
Let τ > 1 be chosen (fixed) and {tk } as before. Determine k ∗ such that
! "
tk∗ = sup tk : y ε , stk (AA∗ ) y ε ≤ τ ε2 . (4.59)
Set
α (ε, y ε ) = tk∗ . (4.60)
We have to look for functions st which yield order optimal methods for the whole
parameter domain (0, μ0 ]. An answer is given by the example:
3
2+ 2 t
st (λ) = pt μ0 (λ) = p3t (λ) = . (4.61)
t+λ
The generalized discrepancy principle with this family of functions st together
with the Tikhonov regularization is an order optimal method with respect to Xμ
for μ ∈ (0, 2].
In the previous considerations we have studied a number of a posteriori pa-
rameter choice rules which all depend in one way or the other on the computed
approximation – and on the given data error level ε. A perfect example to illus-
trate this general reasoning is the discrepancy principle where reconstructions are
discarded unless their data fit has the order to the noise level ε.
In practical examples such noise level information is not always available
(or reliable). For instance, a given discrete data vector may consist of a finite
number of measurements, for each of which we may or may not know the standard
deviation and/or a worst-case error bound. Typically, the worst-case bound will
be a severe overestimation, while the standard deviation might underestimate the
true error. both estimates may therefore lead to a significant loss of accuracy when
used in these parameter choice rules. Another uncertainty problem arises if we are
going to embed the discrete data into a continuous model by some interpolation
or approximation process. Then we have to estimate the L2 -norm of the difference
between the constructed function and the true data function from the discrete
noise information, and from a priori assumed smoothness properties of the data.
Often it is necessary to consider alternative (a posteriori) parameter choice
rules that avoid knowledge of the noise level, and to determine some realistic reg-
ularization parameter on the basis of the actual performance of the regularization
method under consideration. Such heuristic parameter choice rules will be called
error free. A good reference to these strategies is [63]. It must be emphasized,
however, that error free parameter choice rules cannot provide a convergent regu-
larization method in the strict sense. Still, there are examples where an error free
rule leads to better reconstructions than some sophisticated order-optimal rule,
cf., e.g., [113] for some numerical comparisons.
Another heuristic parameter choice rule which can be interpreted via some
kind of error estimation is the method of generalized cross-validation introduced
by Wahba (cf. [256] for the history of this method and a more detailed exposition).
It applies to problems where A is an operator into a finite-dimensional data space,
e.g., a generalized moment problem.
Another very popular error-free parameter choice rule has been advocated
by Hansen [116]. This method is based on an inspection of the residual norms
of the computed approximations, this time by relating them to the norms of the
approximations themselves (cf. [113]). In spite of its use in several applications,
for example, in satellite to satellite tracking, satellite gravity gradiometry (see,
e.g., [72, 79, 79, 94] [231], there still lacks a sound mathematical foundation of the
L-curve method.
Tikhonov–Phillips regularization. The Tikhonov–Phillips filter is given by
0 1
Ft (λ) = 1 /(λ + t) , t > 0, λ ∈ 0, A2 . (4.62)
Let A be of class L (X, Y ) . Without loss of generality, suppose that A is
⊥
injective (otherwise we have to replace X by N (A) ). Let Z be a Hilbert space and
B ∈ L (X, Z) be continuously invertible, i.e., there exists a value β > 0 such that
β xX ≤ BxZ for all x ∈ X. (4.63)
Before we deal with the generalization of the Tikhonov–Phillips regularization we
mention some preparatory results:
(1) Lax–Milgram Lemma Suppose that L ∈ L (X) . Assume there exists a
value λ > 0 such that
Lx, x ≥ λ x2 (4.64)
holds for all x ∈ X. Then L is continuously invertible and
- −1 -
-L - ≤ 1/ λ. (4.65)
(2) The stabilized normal equation
(A∗ A + tB ∗ B) x = A∗ y (4.66)
with y ∈ Y possesses a unique solution for all t > 0 which continuously de-
pends on y.
Our aim is to show that the solution of (4.66) can be equivalently obtained
by minimizing the argument of the Tikhonov–Phillips functional given by
2 2
Jt,y (x) = Ax − y + t Bx . (4.67)
Note that the second term on the right-hand side of (4.67) is called the penalty
term of the Tikhonov–Phillips functional:
Let A and B be given as indicated above. Moreover, assume that y ∈ Y and
t > 0. Then the following statements are equivalent:
(1) (A∗ A + tB ∗ B) xt = A∗ y,
(2) xt minimizes the functional
2 2
Jt,y (x) = Ax − y + t Bx , (4.68)
i.e., xt = arg min {Jt,y (x) : x ∈ X}.
Our purpose is to comment on this result in more detail: For y ∈ Y and the
family of generalized Tikhonov–Phillips regularizations Rt given by
xt = Rt y = (A∗ A + tB ∗ B)−1 A∗ y
= arg min {Jt,y (x) : x ∈ X} , (4.69)
the penalty term satisfies the estimate
1
Bxt ≤ √ y . (4.70)
t
In fact, the property (4.70) explains the role of the penalty term. If t is large,
then Bxt is small compared to Axt − yY . If t 1, then Bxt becomes
large in comparison with the residual term Axt − y. All in all, the choice of the
operator influences the character of xt . Some features can be strengthened, where
others can be weakened.

Let A and B be given as before. Furthermore, suppose that y ∈ D A† and
r > 0. Set - -
- 1 ∗ -
- † ∗ -
δr = inf - 2 B BA y − A y - : y ∈ Y, y ≤ r . (4.71)
β
Then the following statements hold true for xt = (A∗ A + tB ∗ B)−1 A∗ y:
- -2
(1) -xt − A† y - ≤ δr2 + tβ 2 r2 , r, t > 0,
(2) limt→0 xt = A† y.
Some additional effort is needed to formulate regularizations under a priori

parameter choice.
Let A, B given as before. If we choose γ : (0, ∞) → (0, ∞) such that
ε
lim γ (ε) = 0 and lim # = 0, (4.72)
ε→0 ε→0 γ (ε)
−1
then {Rt }t>0 , γ with Rt = (A∗ A + tB ∗ B) A∗ is a regularization of A† .
In order to derive further convergence results we impose further assumptions

on B: If {(σn ; xn , yn )} is the singular system of A and if, for B ∈ L (X, Z), we set
∞ . /
B ∗ Bx = βk2 x, xk xk , βk ∈ β, B2 (4.73)
k=1
(this is a particular specialization), then we get

−1
Rt y = (A∗ A + tB ∗ B) A∗ y
∞
σk
= y, yk xk . (4.74)
σ2 + tβk2
k=1 k
Note that the classical Tikhonov–Phillips regularization uses B = I, i.e., βk = 1
for all k. By βk we control which singular value gets damped and how strongly it
is regularized.
In general, Rt cannot be written as a filter, i.e., in the form Ft (A∗ A)A∗ .
However, the following helpful estimates can be easily verified (see [227]):
- - - - - -
-F 2 (A∗ A)A∗ y - ≤ -Rt y - ≤ -Fβ 2 t (A∗ A)A∗ y - for all y ∈ Y, (4.75)
B t
and
- - - - - -
-pβ 2 t (A∗ A)x- ≤ -(I − Rt A)x- ≤ -p 2 (A∗ A)x- for all x ∈ X, (4.76)
B t
1
where pt (λ) = 1 − λFt (λ) = t
λ+t and Ft (λ) = λ+t (as for the classical Tikhonov
filter).
The family {Rt }t>0 possesses the same asymptotic behavior for t → 0 as the
classical Tikhonov filter, the behavior for ε → 0 is independent of B.
Suppose that A ∈ K(X, Y ) and B ∈ L(X, Z). Assume that the representation
(4.73) holds true. Let Rt be given in the form Rt = (A∗ A + tB ∗ B)−1 A∗ , t > 0.
(a) If the a priori parameter γ is chosen such that
μ+1
2 μ+1
2
ε ε
Cγ ≤ γ(ε) ≤ CΓ for ε → 0, (4.77)
ρ ρ
(as in (4.37)), where Cγ and CΓ are positive constants, then the method
({Rt }t>0 , γ) is an order optimal regularization of A† with respect to Xμ ,
μ ∈ (0, 2].
(b) If we choose γ according to the discrepancy principle, then ({Rt }t>0 , γ) is an
order optimal regularization of A† with respect to Xμ , μ ∈ (0, 1]. The order
1
of decay of the error O(ε 2 ) is maximal.
(c) If we choose γ in accordance with the generalized discrepancy principle with
t∗k = sup{tk | η(tk ) ≤ τ ε2 }, where
η(t) = Axεt − y ε 2 − A∗ (Axεt − y ε ), (A∗ A + tB ∗ B)−1 A∗ (Axεt − y ε ), (4.78)
with y ∈ Y and
ε
= Rt y , then ({Rt }t>0 , γ) is an order optimal regular-
xεt ε
ization of A† with respect to Xμ , μ ∈ (0, 2].

2
Once again, it is possible to prove that O(ε 3 ) is the maximal order of decay
for the error. In other words, the qualification does not increase based on this
generalization. However, it should be remarked that the reconstructions using B =
I may lead to strongly different results. In particular, it is possible to adapt βk to
the spectrum of the noise, if this information is known or it can be estimated by
other methods.
4.1. Generalized inverses in reproducing kernel Hilbert spaces
Within the L2 -context the range of a compact linear operator K with infinite-
dimensional range is always non-closed. In [195–197] we are confronted with the
question: Can one endow R(K) with a new inner product that would make R(K)
a Hilbert space and that would have additional useful properties?
Reproducing kernel Hilbert space (RKHS) framework. A Hilbert space H of
complex-valued functions on a (bounded) set G (e.g., a regular region) is called a
reproducing kernel Hilbert space (RKHS) if all the evaluation functional H x →
x(t) ∈ C are continuous (bounded) for each fixed t ∈ G, i.e., there exists a positive
constant Ct for each t ∈ G such that |x(t)| ≤ Ct xH for all x ∈ H. By the Riesz
Representation Theorem, for each t ∈ G, there exists a unique element Qt such
that x(t) = x, Qt H for all x ∈ H. The reproducing kernel Q(·, ·) : G × G → C of
a RKHS H is defined by Q(s, t) = Qs , Qt H , s, t ∈ G.
We list some basic properties of RKHS’s that are particularly relevant in
approximation and estimation theory:
• Q(s, t) = Q(t, s) for all t, s ∈ G.
• Q(s, s) ≥ 0# ∈ G.
for all s #
• |Q(s, t)| ≤ Q(s, s) Q(t, t) for all s, t ∈ G.
• The reproducing kernel Q(s, t) on G × G is a non-negative definite Hermitian
kernel. Conversely by the Aronszajn–Moore Theorem, every non-negative def-
inite Hermitian function Q(·, ·) on G × G determines a unique Hilbert space
HQ for which Q(·, ·) is a reproducing kernel ([15]) (note that a complex-
valued kernel F on G × G is said to be positive definite if, for any n points
t1 , . . . , tn ∈ G, the matrix A = (F (ti , tj ))1≤i,j≤n is non-negative definite, i.e.,

n
uH Au = ui F (ti , tj ) uj ≥ 0 (4.79)
i,j=1
for all u = (u1 , . . . , un ) ∈ Cn ).

• A closed subspace H̃ of a RKHS H is also a RKHS. Moreover, the orthogonal
projector P of H onto H̃ and the reproducing kernel Q̃(s, t) of the RKHS H̃
are related by P f (s) = f, Q̃s , s ∈ G for all f ∈ H where Q̃k = P Q.
• In a RKHS, the element representing a given bounded linear functional L
can be expressed by means of the reproducing kernel: L(f ) = f, hH , where
h = L(Q).
Similarly, for a bounded linear operator L on H to H, we have that
Lf (t) = Lf, h = f, L∗ h.
• If G is a bounded domain or if G is an unbounded domain but

|Q(t, s)|2 dt ds < ∞, (4.80)
G×G
and Q(s, t) is continuous on G ×G, then HQ is a space of continuous functions.

• Every finite-dimensional function space is a RKHS H with reproducing kernel
n
Q(s, t) = ui (s) ui (t), (4.81)
i=1
where {ui }ni=1is an orthonormal basis for H (notice that the sum in the above
definition of the kernel Q is invariant under the choice of an orthonormal
basis).
• If the integral relation

|Q(s, t)|2 ds dt < ∞, (4.82)
G×G
holds true, then Q(·, ·) has a countable sequence of eigenvalues and eigen-
functions (Theorem of Mercer).
• Let {ϕn }n∈N be a sequence of complex functions defined on G such that, for
every t ∈ G,
∞
|ϕn (t)|2 < ∞. (4.83)
n=1
∞ ∞
For every sequence {cn }n∈N with n=1 |cn |2 < ∞, the series n=1 cn ϕn (t)
is then convergent in C for every t ∈ G. The functions which are the sums
of such series form a linear subspace H, on which we are able to define the
structure of a separable Hilbert space by taking as scalar product, for
∞
∞

f= cn ϕn , g = dn ϕn , (4.84)
n=1 n=1
the number
∞

f, gH = cn dn . (4.85)
n=1
This space has a reproducing kernel, namely
∞

Q(x, y) = ϕn (t)ϕn (s), t, s ∈ G × G. (4.86)
n=1
• Let H be a separable RKHS, then its reproducing kernel Q(·, ·) has the
expansion
∞

Q(s, t) = ϕn (t) ϕn (s), (4.87)
n=1
where {ϕn }∞
n=1 is an orthonormal basis for H (we remark that for a general
∞
separable Hilbert space H, n=1 ϕn (t) ϕn (s) is not a reproducing kernel
(note that L2 (G) is not an RKHS) and also that φn ’s do not generally cor-
respond to sampling expansions. If they do,∞ i.e., if ϕn (t) = Q(tn , t) for some
sequence {tn }, then we have that f (t) = n=1 f (tn ) ϕn (t), this constitutes
a sampling theorem.)
• If the reproducing kernel Q(s, t) of a RKHS H is continuous on G ×G, then H
is a space of continuous functions (being uniformly continuous on a bounded
G). This follows from
|x(t) − x(s)| = |x, Qt − Qs H | ≤ xH Qt − Qs H (4.88)
and
Qt − Qs 2 = Q(t, t) − 2Q(t, s) + Q(s, s) (4.89)
for all s, t ∈ G.
• Strong convergence in a RKHS H implies pointwise convergence and uniform
convergence on compact sets, because of the fact
#
|x(t) − xn (t)| = |x − xn , Qt H | ≤ Q(t, t) x − xn H . (4.90)
• Let HQ denote the RKHS with reproducing kernel Q, and denote the inner
product and norm in HQ by ·, ·HQ and · HQ , respectively. Note that
Q(s, s )(= Qs (s )) is a non-negative definite Hermitian kernel on G × G, and
that {Qs , s ∈ G} spans HQ since Qs , xHQ = 0, s ∈ G, implies x(s) = 0. For
more properties of reproducing kernel spaces the reader is referred to, e.g.,
[15, 37, 131] and the references therein.
• For every positive definite kernel Q(·, ·) on G × G, there exist a zero mean
Gaussian process with Q(·, ·) as its covariance, giving rise to the relation
between Bayes estimates, Gaussian processes, and optimization processes in
RHKS (for more details the reader is referred to the geodetic literature, see,
e.g., [100, 159, 160, 167], and the monographs [148, 256]).
Interest in reproducing kernel Hilbert spaces have increased in recent years,
as the computer capacity has made solutions of ever larger and more complex
problems practicable. Indeed, new reproducing kernel representations and new ap-
plications (in particular in physical geodesy and geophysics) are being contributed
at a rapid rate. For example, a certain RHKS in terms of outer harmonics al-
lows the adequate determination of the Earth’s gravitational potential (see, e.g.,
[71, 237] for early spline approaches) in consistency with gravitational observables
of heterogeneous type (that are interpretable as (bounded) linear functionals on
the RKHS under consideration).
Hilbert–Schmidt operator theory. An RKHS HQ with RK Q determines a self-
adjoint Hilbert–Schmidt operator (also denoted by Q) on L2 (G) to L2 (G) by letting

(Qx)(s) = Q(s, s ) x(s ) ds , x ∈ L2 (G). (4.91)
G
Since Q is assumed to be continuous, then by the Theorem of Mercer (see, e.g.,
[139]), the operator Q has an L2 (G)-complete orthonormal system ofeigenfunctions
∞
{φi }∞
i=1 and corresponding eigenvalues {λ i } ∞
i=1 with λi ≥ 0 and i=1 λi < ∞.
Thus Q is a trace-class operator (see, e.g., [46]) so that Q(·, ·) has the uniformly
convergent Fourier expansions
∞

Q(s, s ) = λi φi (s) φi (s ) (4.92)
i=1
and
∞

Qx = λi x, φi L2 (G) φi , (4.93)
i=1
where ·, ·L2 (G) is the inner product in L2 (G). It is well known (see, for example,
[249]) that the space HQ consists of all x ∈ L2 (G) satisfying the condition
∞
2
λi −1 x, φi L2 (G) < ∞, (4.94)
i=1
(note that the notational convention 0/0 = 0 is being adopted) with inner product
·, ·HQ given by
∞

x1 , x2 HQ = λ−1
i x1 , φi L2 (G) x2 , φi L2 (G) . (4.95)
i=1
The operator Q has a well-defined symmetric square root Q1/2 which is a Hilbert–
Schmidt operator (see, e.g., [195–197]):
∞ #

Q1/2 x = λi x, φi L2 (G) φi . (4.96)
i=1
Thus, since N (Q) = N (Q1/2 ), we have

HQ = Q1/2 (L2 (G)) = Q1/2 L2 (G) N (Q) . (4.97)
1/2 †
(Q ) has the representation
∞ #

(Q1/2 )† x = ( λi )† x, φi L2 (G) φi (4.98)
i=1
⊥
on HQ ⊕HQ (⊥ in L2 (G)), where, for Θ a real number, Θ† = Θ−1 for Θ = 0; Θ† = 0
for Θ = 0. Similarly, Q† has the representation
∞

Q† x = λ†i x, φi L2 (G) φi . (4.99)
i=1
For any operator Q on L2 (G), induced by an RK Q(s, s ) as defined in (4.91), we

shall adopt the notational conventions (cf. [195–197])
Q−1 := Q† , (4.100)
−1/2 1/2 †
Q := (Q ) . (4.101)
This leads to the relations

xHQ = inf pL2 (G) : p ∈ L2 (G), x = Q1/2 p , x ∈ HQ . (4.102)
and
x1 , x2 HQ = Q1/2 x1 , Q1/2 x2 L2 (G) , x1 , x2 ∈ HQ , (4.103)
and, if x1 ∈ HQ and x2 ∈ HQ with x2 = Qρ for some ρ ∈ L2 (G), then
x1 , x2 HQ = x1 , ρL2 (G) . (4.104)
Relationship between generalized inverses and L2 -topology. We are now ready to

explore properties of the generalized inverse of a linear operator between two RK
spaces (cf. [195–197]). To this end we let X = L2 (G) and Y = L2 (H) denote the
Hilbert spaces of square-integrable real-valued functions on the closed, bounded
domains G and H, respectively. Let A be a linear operator from X into Y . Let ⊂
denote point set inclusion only, and suppose that A has the following properties:
HQ ⊂ D(A) ⊂ X, (4.105)
where HQ is an RKHS with continuous RK on G × G;
A(HQ ) = HR ⊂ HR ⊂ Y, (4.106)
where HR and HR are RKHS’s with continuous RK’s on H × H; so that
N (A) in HQ is closed in HQ . (4.107)
We emphasize, in particular, that the space HR is not necessarily closed in the
topology of HR .
Let A†(X,Y ) denote the generalized inverse of A, when A is considered as a
mapping from X into Y , and let A†(Q,R) denote the generalized inverse of A when
A is considered as a mapping from HQ into HR . Now, the topologies in (X, Y ) are
not the same as the topologies in (HQ , HR ). Thus, the generalized inverses A†(X,Y )
and A†(Q,R) show distinct continuity properties, in general. We shall develop the
relation between A†(Q,R) and certain (X, Y ) and (Y, Y ) generalized inverses. In the
sequel, the operators R : Y → Y and R1/2 : Y → Y are defined from the RK of
HR analogous to Q and Q1/2 (see (4.93) and (4.96)). We continue the notational
†
convention of (4.100), i.e., R−1 = R† = R(Y,Y ) and R
−1/2
= (R1/2 )†(Y,Y ) .
From [195–197] we are able to deduce the following result: Under assumptions
(4.105)–(4.107), let y ∈ D(A†(Q,R) ), i.e., y ∈ HR ⊕ HR
⊥
(⊥ in HR ). Then
y ∈ D(Q1/2 (R−1/2 AQ1/2 )†(X,Y ) R−1/2 ) (4.108)

and
A†(Q,R) y = Q1/2 (R−1/2 AQ1/2 )†(X,Y ) R−1/2 y. (4.109)
Moreover, it follows that A†(Q,R) is bounded, provided that A(HQ ) = HR .

It should be noted that an operator A may satisfy the assumption A(HQ ) =

HR while failing to have a closed range in the space Y . This is, for example,
the case if A is a Hilbert–Schmidt linear integral operator (with non-degenerate
kernel) on X. It is this observation which makes RKHS useful in the context of
regularization and approximation of ill-posed linear operator equations.
Explicit representation of minimal norm solutions. We assume that HQ is chosen

so that the linear functionals {Et : t ∈ H} defined by
Et x = (Ax)(t) (4.110)
are continuous in HQ . Then, by the Riesz representation theorem, there exists

{ηt , t ∈ H} ∈ HQ such that
(Ax)(t) = ηt , xHQ , t ∈ H, x ∈ HQ , (4.111)
where ηt is explicitly given by
ηt (s) = yr , Qs HQ = (AQs )(t) (4.112)
(ηt (s) is readily obtained in a more explicit form from (4.112) if A is a differential
or integral operator).
Let R(t, t ) be the non-negative definite kernel on H × H given by
R(t, t ) = ηt , ηi HQ , t, t ∈ H. (4.113)
Assume that HR is the RKHS with RK R given by (4.113). Let Rt be the element
of HR defined by Rt (t ) = R(t, t ), and let ·, ·HR be the inner product in HR .
Suppose that V is the closure of the span of {ηt , t ∈ H} in HQ . Now, {Rt , t ∈ H}
spans HR , and by the properties of RKHS, we have
ηt , ηt HQ = R(t, t ) = Rt , Rt HR . (4.114)
Thus there is an isometric isomorphism between the subspace V and HR , generated

by the correspondence
ηt ∈ V ∼ Rt ∈ HR . (4.115)
Then, x ∈ V ∼ y ∈ HR if and only if ηt , xHQ = y(t) = Rt , gHR , t ∈ H, i.e., if
and only if y(t) = (Ax)(t), t ∈ H. Thus, A(HQ ) = A(V ) = HR . The null space of
A in HQ is {x : x ∈ HQ , AxHR = 0}. Since
ηt , xHQ = 0, t∈H and x ∈ HQ ⇒ x ∈ V ⊥ , (4.116)
and x ∈ V implies xQ = AxV , it follows that the null space of A in HQ is V ⊥

( ⊥ in HQ ). Hence, (4.110) entails that the null space of A : HQ → HR in HQ is
always closed, irrespective of the topological properties of A : X → Y .
We list the following table of corresponding sets and elements, under the
correspondence “∼” of (4.115), where the entries on the left are in HQ :
V ∼ HR , (4.117)
x ∼ y, (4.118)
ηt ∼ Rt , (4.119)
PV Qs ∼ ηs∗ . (4.120)
Here, PV is the projector from HQ onto the (closed) subspace V , y(t) = ηt , xHQ ,
t ∈ H, and ηs∗ = AQs = A(PV Qs ), i.e.,
ηs∗ (t) = ηi , PV Qs HQ = ηt (s). (4.121)
This leads to the following result (the proof is given in [196, 197]):
Let A and HQ satisfy (4.110), and let R be given by (4.114), where ηi is
defined by (4.111). Let ηs∗ = AQs . Then, for y ∈ HR ,
(A†(Q,R) y)(x) = ηs∗ , yHR , s ∈ G. (4.122)
We also obtain another operator representation of A†(Q,R) :

Under the assumptions
(i) D(A∗ ) is dense in Y , where A∗ is the adjoint of A considered as an operator
from X to Y ,
(ii) A and HQ satisfy (4.111),
(iii) HQ and HR = A(HQ ) possess continuous reproducing kernels,
we have, for y ∈ HR ,
(A†(Q,R) y)(s) = (QA∗ (AQA∗ )†(Y,Y ) y)(s), s ∈ G. (4.123)
Poorly conditioned operator equations. For A : X → Y , the pseudocondition

number of A (relative to the norms of X and Y ) is given by
AxY A† yX
γ(A; X, Y ) = sup sup . (4.124)
x =0 xX y =0 yY
x∈D(A) y∈D(A† )
The equation Ax = y is said to be poorly conditioned in the spaces X, Y if the

number γ(A; X, Y ) is much greater than 1 (note that 1 ≤ γ(A; X, Y ); for ill-posed
problems, γ is not finite).
Suppose HQ is an RKHS with HQ ⊂ D(A), and A and HQ satisfy (4.110)
with A(HQ ) = HR , R given by (4.114). Then γ(A; HQ , HR ) = 1. To see this, write
x ∈ HQ in the form x = x1 + x2 , where x2 ∈ V ⊥ . Then Ax = Ax1 = y1 and
y1 HR = x1 HQ . Thus
y1 HR x1 HQ
γ(A; HQ , HR ) = sup sup = 1. (4.125)
x =0 xHQ y1 =0 y1 HR
On the other hand, the number γ(A; X, Y ) may be large. Thus, the casting of the
operator equation Ax = y in the reproducing kernel spaces HQ , HR always leads
to a well-conditioned (indeed, optimally-conditioned) problem.
Regularization of pseudosolutions in reproducing kernel spaces. We study proper-
ties of regularized pseudosolutions (in RKHS) xα of the operator equation Ax = y,
where y is not necessarily in the range of the operator A. By a regularized pseudoso-
lution we mean a solution to the variational problem: Find xα in HQ to minimize
Jy (x) = y − Ax2HP + αx2HQ , α > 0, (4.126)
where HQ is an RKHS in the domain of A, · HP denotes the norm in an RKHS
HP with RK P , HP ⊂ Y , Jy ⊂ Y , φy (x) is assigned the value +∞ if y − Ax ∈
/ HP ,
and α > 0. We suppose A and HQ satisfy (4.110), hence, A(HQ ) = HR possesses
an RK. As before, A may be unbounded, invertible, or compact considered as an
operator from X(= L2 (G)) to Y (= L2 (H)). It is assumed that y possesses a (not
necessarily unique) representation y = y0 + ξ, for some y0 ∈ A(HQ ) and ξ ∈ HP ,
where ξ may be thought of as a “disturbance”.
For α > 0, let HαP be the RKHS with RK αP (t, t ), where P (t, t ) is the RK
on H × H associated with HP . We have HP = HαP and
· 2HP = α · 2HαP . (4.127)
Let R(α) = R + αP , and let HR(α) be the RKHS with RK R(α) = R(α, t, t ).
According to [15], HR(α) is the Hilbert space of functions of the form
y = y0 + ξ, (4.128)
where y0 ∈ HR and ξ ∈ HP . Following [15], we note that this decomposition is not
unique unless HR and HP have no element in common except the zero element.
The norm in HR(α) is given by
y2R(α) = min{y0 2HR + ξ2HαR : y0 ∈ HR , ξ ∈ HP , y0 + ξ = y}, (4.129)
where, however, the y0 and ξ attaining the minimum in (4.129) are easily shown
to be unique by the strict convexity of the norm.
Consider the problem of finding xα ∈ HQ to minimize Jy (x) in (4.126) for
y ∈ HR(α) . Then y − Axα must be in HP and it is obvious that xα ∈ V , the
orthogonal complement of the null space of A in HQ . For any x ∈ V , xHQ =
AxHR by the isometric isomorphism between V and HR , and (4.126) may be
written in the equivalent form: Find xα ∈ V to minimize
αAx2HR + y − Ax2HP . (4.130)
Comparing (4.129) and (4.130) with the aid of (4.127), we see that y0 and ξ
attaining the minimum on the right-hand side of (4.129) are related to the solution
xα of the minimization problem (4.130), by
y0 = Axα and ξ = y − Axα . (4.131)
A representation of the solution xα is given (see [195, 196]) as follows:
Suppose D(A∗ ) is dense in Y, HQ ⊂ D(A) and A and HQ satisfy (4.110).

Suppose HQ , HR (= A(HQ )) and HP ⊂ Y all have continuous RK’s. Then, for
y ∈ HR(α) , the unique minimizing element xα ∈ HQ of the functional Jy (x) is
given by
ηs∗ , yR(α) = xα (s) = (QA∗ (AQA∗ + αP )†(Y,Y ) y)(s) s ∈ G, (4.132)
where ηs∗ = AQs . We call the (linear) mapping which assigns to each y ∈ HR(α) the
unique minimizing element xα the regularization operator of the equation Ax = y.
The most useful situations occur, of course, when HR is strictly contained
in HR(α) . For example, HR may be a dense subset of Y in the Y -topology and
⊥
HR(α) a bigger dense subset. We deal with this situation later. If HR (in Y ) is
not empty, then P may be chosen so that the closure of HP in the Y -topology
⊥
equals HR in Y . Then HP ∩ HR = {0}, HαP and HR are orthogonal subspaces
of HR(α) (see, e.g., [15]), and the decomposition (4.128) is unique. In this case
we have the following theorem which shows that the regularization operator is
indeed a generalized inverse in an appropriate RKHS: If HP ∩ HR = {0}, then the
minimizing element xα of (4.126) is the solution to the problem: Find x ∈ L(y)
to minimize
xHQ , (4.133)
where
L(y) = {x : x ∈ HQ , y − AxHR(α) = inf y − AzHR(α) }. (4.134)
z∈HQ
It should be remarked that, in our approach, we have

A(HQ ) = HR ⊂ HR(α) ⊂ Y. (4.135)
Replacing HR and HR in (4.106) by HR and HR(α) , respectively, we get from
(4.109)
A†(Q,R(α)) y = Q1/2 [(R + αP )−1/2 AQ1/2 ]†(X,Y ) (R + αP )−1/2 y (4.136)
for y ∈ D(A†(Q,R(α)) ) .
It is helpful to remember that the topology on HR is not, in general, the
restriction of the topology of HR(α) , with the notable exception of the case HR ∩
HP = {0}. In [129] a concrete example is provided arising in the approximate
solution of boundary value problems, where HR is not a closed subspace of HR(α) .
If HR ∩ HP = {0}, then HR is a closed subspace of HR(α) , and we have
A†(Q,R(α)) = QA∗ (R + αP )−1 . (4.137)
Note that in this case, the generalized inverse and the regularization operator
coincide. If HR = A(HQ ) is not closed in HR(α) , then the regularization operator
and the generalized inverse are different. Also, the right-hand side of (4.136) and
(4.137) are not the same: (4.137) has maximal domain HR(α) , while (4.136) has
⊥
maximal domain HR ⊕ HR (⊥ in HR(α) ).
Rates of convergence to the generalized inverse. We note some properties of xα
as α → 0 when HR ⊂ HP . If y ∈ HR = A(HQ ), then we have xα → A†(Q,R) y
as α → 0; here we may say something about the rate of convergence if certain
additional conditions are satisfied (compare also with [129]). However, y may not
be in the domain of A†(Q,R) . This situation can occur if, for example, HR is dense in
HR(1) . In this case, limα→0 xα Q = ∞: Assume that y = Ax0 + ξ0 , where x0 ∈ V ,
ξ0 ∈ HP and suppose that HR ⊂ HP . Then the following properties hold true:
(i) B = P −1/2 R1/2 is a bounded operator on Y = L2 (H).
(ii) If ξ0 = 0 and (B ∗ B)−1 R−1/2 (Ax0 )L2 (H) < ∞, then
A†(Q,R) y − xα 2HQ = O(α2 ). (4.138)
(iii) If ξ0 = 0 and (B ∗ B)−1/2 R−1/2 (Ax0 )L2 (H) < ∞, then
A†(Q,R) y − xα 2HQ = O(α). (4.139)
(iv) If ξ0 ∈
/ HR , then limα→0 xα HQ = ∞.
Here inverses indicated by “– ” are the generalized inverses in the topology
of L2 -spaces.
4.2. Projection methods

First, we recall the definition of a projection operator known from functional anal-
ysis: Let X be a normed space over the field K (K is equal to R or C). Let U ⊂ X
be a closed subspace. A linear bounded operator P : X → X is called a projection
operator on U if P x ∈ U for all x ∈ X and P x = x for all x ∈ U . Every non-trivial
projection operator satisfies P 2 = P and P ≥ 1.
The following two examples provide some important projection operators:
(a) (Orthogonal Projection.) Let X be a pre-Hilbert space. Suppose that U ⊂ X
is a complete subspace. Let P x ∈ U be the best approximation to x in U , i.e.,
P x satisfies the relation
P x − x ≤ u − x for all u ∈ U. (4.140)
P : X → U is linear and P x ∈ U is characterized by x − P x, u = 0 for all
u ∈ U , i.e., x − P x ∈ U ⊥ . Therefore,
x2 = P x + (x − P x)2
= P x2 + x − P x2 + 2 Rex − P x, P x ≥ P x2 ,

=0
i.e., P = 1.
(b) (Interpolation Operator.) Let X = C (0) ([a, b]) be the space of real-valued con-
tinuous functions on [a, b] supplied with the supremum norm ·C (0) [a,b] . Then
X is a normed space over R. Let U = span{u1 , . . . , un } be an n-dimensional

subspace and t1 , . . . , tn ∈ [a, b] such that the interpolation problem in U is
uniquely solvable, i.e., det(uj (tk )) = 0. We define P x ∈ U by the interpolant
of x ∈ C (0) ([a, b]) in U , i.e., u = P x ∈ U satisfies u(ti ) = x(ti ) for all
i = 1, . . . , n. Then P : X → U is a projection operator.
Examples for U in (b) are spaces of algebraic or trigonometric polynomials.
As a drawback of these choices, we note that from the results of Faber (see, for
example, [198]) the interpolating polynomials of continuous functions x do not, in
general, converge to x as the degree of the polynomials tends to infinity. Nonethe-
less, trigonometric interpolation at equidistant points converges with optimal order
of convergence.
Next we are concerned with a certain class of projection methods in Banach
spaces, essentially following the monograph of A. Kirsch [137]: Let X and Y be
Banach spaces and A : X → Y be bounded and one-to-one. Furthermore, let
Xn ⊂ X and Yn ⊂ Y be finite-dimensional subspaces of dimension n and Qn :
Y → Yn be a projection operator. For y ∈ Y , the projection method for solving
the equations Ax = y is to solve the equations
Qn Axn = Qn y for xn ∈ Xn . (4.141)
Assume that {x̂1 , . . . , x̂n } and {ŷ1 , . . . , ŷn } are bases of Xn and Yn , respectively.
Then Qn y and every Qn Ax̂j , j = 1, . . . , n, are representable in the forms

n
n
Qn y = βi ŷi and Qn Ax̂j = Bij ŷi , j = 1, . . . , n, (4.142)
i=1 i=1
n
with βi , Bij ∈ K. The linear combination xn = j=1 αj x̂j solves (4.141) if and
only if α = (α1 , . . . , αn )T ∈ Kn solves the finite system of linear equations

n
Bij αj = βi , i = 1, . . . , n. (4.143)
i=1
We are led to the following important classes of projection methods for A :

X → Y being a bounded and one-to-one operator.
Let X and Y be pre-Hilbert spaces and Xn ⊂ X and Yn ⊂ Y be finite-
dimensional subspaces with dimXn = dimYn = n. Let Qn : Y → Yn be the
orthogonal projection. Then the projected equation Qn Axn = Qn y is equivalent to
Axn , zn = y, zn for all zn ∈ Yn . (4.144)
We let Xn = span{x̂1 , . . . , x̂n } and Yn = span{ŷ1 , . .
. , ŷn }. Looking for a solution
n
of (4.144) in the form of a linear combination xn = j=1 αj x̂j we arrive at

n
αj Ax̂j , ŷi = y, ŷi for i = 1, . . . , n, (4.145)
j=1
or in matrix-vector Aα = β, where Aij = Ax̂j , ŷi and βi = ŷ, ŷi , i = 1, . . . , n.

A priori assumption. For the remaining part of this subsection about projection
methods, it is helpful to make the following a priori assumption (APA):
(i) Let A : X → Y be a linear, bounded, and injective operator between Banach
spaces, Xn ⊂ X and Yn ⊂ Y be finite-dimensional subspaces of;dimension
n and Qn : Y → Yn be a projection operator. We assume that n∈N Xn is
dense in X and the Qn A|Xn : Xn → Yn is one-to-one and, thus, invertible.
(ii) Let x ∈ X be the solution of
Ax = y. (4.146)
By xn ∈ Xn , we understand the unique solutions of the equations
Qn Axn = Qn y, n ∈ N. (4.147)
As a consequence of (APA) we are allowed to represent the solutions xn of
(4.147) in the form xn = Rn y, where Rn : Y → Xn ⊂ X is defined by
Rn := (Qn A|Xn )−1 Qn : Y → Xn ⊂ X. (4.148)
Suppose that (APA) is valid throughout this section. The projection method
is called convergent if the approximate solutions xn ∈ Xn of (4.147) converge to
the exact solution x ∈ X of (4.146) for every y ∈ A(X), i.e., if the limit relation
Rn Ax = (Qn A|Xn )−1 Qn Ax → x, n → ∞, (4.149)
holds true for every x ∈ X.
Obviously, this definition of convergence coincides with the definition of a
regularization strategy for the equation Ax = y. Therefore, the projection method
converges if and only if Rn is a regularization strategy for the equation Ax = y.
;
Convergence can only be expected if we require that n∈N Xn is dense in X
and Qn y → y for all y ∈ A(X). For a compact operator A, however, this property
is not sufficient for the convergence. In fact we have to assume an additional
boundedness condition:
The solution xn = Rn y ∈ Xn of (4.147) converges to x for every y = Ax if
and only if there exists c > 0 such that
Rn A ≤ c for all n ∈ N. (4.150)
If (4.150) is satisfied, the following error estimate can be shown to be valid
xn − x ≤ (1 + c) min zn − x (4.151)
zn ∈Xn
with the same constant c as in (4.150).

The estimates (4.150) and (4.151) can be verified in straightforward way.
Suppose that Rn A is bounded. The operator Rn A is a projection operator
onto Xn since for zn ∈ Xn we have Rn Azn = (Qn A|Xn )−1 Qn Azn = zn . Thus we
are able to deduce that
xn − x = (Rn A − I)x = (Rn A − I)(x − zn ) for all z n ∈ Xn . (4.152)
It follows that
xn − x ≤ (c + 1)x − zn for all z n ∈ Xn (4.153)
;
such that (4.151) is valid. Convergence xn → x follows from the fact that n∈N Xn
is dense in X.
Thus far, we were concerned with the case, where the right-hand side y is
exactly known. Next we consider the case where the right-hand side is known only
approximately, i.e., we start from an element y ε ∈ Y with y ε − y ≤ ε. To this
end we understand the operator Rn from (4.148) as a regularization operator in
the usual sense so that we are led to distinguish two kinds of errors for the right-
hand side. A straightforward application of the triangle inequality yields with
xεn := Rn y ε the inequality
xεn − x ≤ xεn − Rn y + Rn y − x
≤ Rn y ε − y + Rn Ax − x. (4.154)
As usual, we are confronted with the dilemma of IP: The error ε of the right-hand
side is multiplied by the norm of Rn . The second term describes the discretization
error against the exact data.
In practice one solves the discrete system (4.143) where the vector β is re-
placed by a perturbed vector β ε ∈ Kn with

n
|β ε − β|2 = |βjε − βj |2 ≤ ε2 . (4.155)
j=1
We will call this fact the discrete perturbation of the right-hand

n side. Instead of
(4.143) one solves Bαε = β and defines xεn ∈ Xn by xεn = j=1 αεj x̂j . Note that
the choice of the basis functions x̂j ∈ Xn and ŷj ∈ Yn are essential rather than the
norm of Y . Unfortunately, it turns out, that the condition number of B reflects
the ill-conditioning of the equation Ax = y. In this respect it should be mentioned
that it suffices to study the question of convergence for the “principal part” of the
operator A under discussion. More concretely, if the projection method converges
for an operator A, then convergence and error estimates also hold true for A + C,
where C is compact relative to A (i.e., A−1 C is compact).
The proof of the following result is, e.g., given in the monograph [137]:
Let C : X → Y be a linear operator with C(X) ⊂ A(X) such that A + C is
one-to-one and A−1 C is compact in X. Assume, furthermore, that the projection
method converges for A, i.,e., that Rn Ax → x, n → ∞, for every x ∈ X, where
Rn = (Qn A|Xn )−1 Qn .
Then it also converges for A + C i.e.,
(Qn (A + C)|Xn )−1 Qn (A + C)x → x, n → ∞, for all x ∈ X.
Let x ∈ X be the solution of (A + C)x = y and xε ∈ Xn be the solution of the

corresponding projected equation Qn (A + C)xεn = ynε for some ynε ∈ Yn . Then there
exists a constant c > 0 such that
0 1
x − xεn ≤ c A−1 Cx − Rn Cx + A−1 y − Rn ynε (4.156)
for all sufficiently large n and ε > 0.
It should be mentioned that the first term on the right-hand side of (4.156)
is just the error of the projection methods for the equations Ax = Cx without
perturbation of the right-hand side. This allows to assure the estimate
A−1 Cx − Rn Cx ≤ (1 + c) min A−1 Cx − zn .
zn ∈XN
The second term on the right-hand side of (4.156) is the error for the equation Ax =
y. Hence, our results include both the continuous and the discrete perturbations
of the right-hand side. Forthe continuous case we set ynε := Qn y ε , while in the
n
discrete case, we set ynε = i=1 βiε yi .
All in all, a framework for reduction of inverse and identification problems
to finite-dimensional problems exists within the concept of projection methods
(cf. [184]). Projection methods (e.g., spline or finite-element functions) can be
either applied directly to IPP or to the regularized problem, i.e., to the problem
of minimizing a Tikhonov type functional (see, e.g., [158, 199, 241]). The moment-
discretization (or semidiscretization) method for integral equations of the first
kind and for IPP in reproducing kernel Hilbert spaces (see [178, 195]), viewed
as a projection method, is both quasi-optimal and robust. It is also particularly
suited when values of the data function y(s) are known only at a finite number
of points. These properties, together with the convergence and commutativity
properties established in [178] may account for the favorable behavior of computer
implementations of this method [16], which the authors [16] call Nashed’s method.
Galerkin methods. We deal with the situation that X and Y are (real or complex)
Hilbert spaces. Moreover, A : X → Y is assumed to be linear, bounded, and one-
to-one; Xn ⊂ X and Yn ⊂ Y are assumed to be finite-dimensional subspaces with
dim Xn = dim Yn = n; and Qn : Y → Yn is the orthogonal projection operator
onto Yn . Then, Qn Axn = Qn y reduces to the so-called Galerkin equations
Axn , zn = y, zn for all zn ∈ Yn . (4.157)
Choosing bases {x̂1 , . . . , x̂n } and {ŷ1 , . . . , ŷn } of
Xn and Yn , respectively, we are
n
led to a finite system in the coefficients of xn = j=1 αj x̂j (see (4.145)):

n
Bij αj = βi , i = 1, . . . , n, (4.158)
i=1
where we have used the abbreviations Bij = Ax̂j , ŷi Y and βi = y, ŷi Y . We
observe that Bij and βi coincide with the settings in (4.142) only if the set {ŷj :
j = 1, . . . , n} forms an orthonormal basis of Yn .
It should be remarked that the Galerkin method is also known as the Petrov–
Galerkin method (see [215]) since Petrov was the first to consider the general
situation of (4.157). The special case X = Y and Xn = Yn was studied by Bubnov
in 1914 and later by Galerkin in 1915 (see [86]). For this reason, this special case
is also known as the Bubnov–Galerkin method. In the case when the operator A
is self-adjoint and positive definite, we will see that the Bubnow–Galerkin method
coincides with the Rayleigh–Ritz method (see [221, 228]).
Error estimates. The following error estimates for the Galerkin method of
the form (4.154) (see, e.g., [137]) differ only in the first term, which corresponds
to the perturbation of the right-hand side. The second term bounds the error for
the exact right-hand side and tends to zero, provided the boundedness assumption
(4.150) is satisfied.
Assume that the Galerkin equations (4.157) are uniquely solvable for every
right-hand side of the equation Ax = y.
(a) Let y ε ∈ Y with y − y ε ≤ ε be given and xεn ∈ Xn be the solution of
Axεn , zn = y ε , zn for all zn ∈ Yn . (4.159)
Then the following error estimate holds true:
xεn − x ≤ εRn + Rn Ax − x. (4.160)
(b) Let B and β be given by (4.158) and β ε ∈ Kn with |β − β ε | ≤ ε, where | · |
denotes
the Euclidean norm in Kn . Let αε ∈ Kn be the solution of Bαε := β ε . Set
n
xn := j=1 αεj x̂j ∈ Xn . Then the following error estimate holds true:
ε
an
xεn − x ≤ ε + Rn Ax − x, (4.161)
λn
xεn − x ≤ bn Rn ε + Rn Ax − x, (4.162)
where
⎧- - ⎫
⎨-- n -
- n ⎬
an = max - ρ x̂ - : |ρ | 2
= 1 , (4.163)
⎩- -
j j j
- j=1 - ⎭
j=1
⎧6 X
- - ⎫
⎨7 - - ⎬
7 n
- n -
bn = max 8 |ρj |2 : - - ρ j ŷ j
-=1 ,
- (4.164)
⎩ - j=1 - ⎭
j=1
and λn > 0 denotes the smallest singular value of the matrix B.

Next we are interested in deriving error estimates for three particularly in-
teresting choices for the finite-dimensional subspaces Xn and Yn (see, e.g., [137]).
The cases, where Xn and Yn are coupled by Yn = A(Xn ) or Xn = A∗ (Yn ) will
lead to the least squares method or the dual least squares method, respectively.
In addition we will study the Bubnov–Galerkin method for the case where A ad-
ditionally satisfies the so-called Garding inequality. In all cases, we formulate the
Galerkin equations for the perturbed cases first without using particular bases and
then with respect to given bases in Xn and Yn .
Least squares method. For a finite-dimensional subspace Xn ⊂ X, determine
xn ∈ Xn such that
Axn − y ≤ Azn − y for all zn ∈ Xn . (4.165)
Clearly, existence and uniqueness of xn ∈ Xn can be guaranteed easily since Xn
is finite-dimensional and A is assumed to be one-to-one. The solution xn ∈ Xn of
the least squares problem is characterized by
Axn , Azn = y, Azn for all zn ∈ Xn . (4.166)
We notice that this method is a special case of the Galerkin method, where we
have Yn = A(Xn ). Choosing a basis {x̂j , j = 1, . . . , n} of Xn leads to the finite
linear system

n
αj Ax̂j , Axi = βi = y, Ax̂i for all i = 1, . . . , n, (4.167)
j=1
i.e., in matrix-vector nomenclature Bα = β. The corresponding matrix B ∈ Kn×n

with Bij = Ax̂j , Ax̂j Y is Hermitian and positive definite, since A is assumed to
be one-to-one.
Of practical interest is the case where the right-hand side is perturbed by an
error. Let xεn ∈ XN solve the equations
Axεn , Azn = y ε , Azn for all zn ∈ Xn , (4.168)
where y ε ∈ Y is the perturbed right-hand side satisfying y ε − yY ≤ ε. For
the discrete perturbation, we assume that β ∈ Kn is replaced by β ε ∈ Kn with
|β ε −β| ≤ ε, where |·| denotes the Euclidean norm in Kn . This leads
n to the following
finite system of equations in the unknown coefficients of xεn = j=1 aεj x̂j :

n
αεj Ax̂j , Ax̂i = βiε for all i = 1, . . . , n. (4.169)
j=1
The system (4.169) is uniquely solvable, since the matrix B is positive definite.
Obviously, for least squares methods, the boundedness condition (4.150) is
not satisfied without imposing additional assumptions (for more details we refer,
e.g., to [139, 234]):
Let A : X → Y be a linear, bounded, and injective operator between
; Hilbert
spaces. Suppose that Xn ⊂ X form finite-dimensional subspaces such that n∈N Xn
is dense in X. Let x ∈ X be the solution of Ax = y and xεn ∈ Xn be the least
squares solution from (4.168) or (4.169). Denote by σn the quantity
σn = max {zn : zn ∈ Xn , Azn = 1} . (4.170)
Suppose that there exists c > 0 independent of n, such that

min {x − zn + σn A(x − zn )} ≤ cx for all x ∈ X. (4.171)
zn ∈Xn
Then, the least squares method is convergent, and we have Rn ≤ σn . Moreover,
we have the error estimate
x − xεn ≤ rn σn ε + c min {x − zn : zn ∈ Xn } (4.172)
for some c > 0. Here, rn = 1, if xεn ∈ Xn solves (4.168), i.e., ε measures the
continuous perturbation y ε− yY . If ε measures the discrete error |β ε − β| in the
n
Euclidean norm and xn = j=1 αεj x̂j ∈ Xn , where the vector αε = (αε1 , . . . , αεn )T
ε
solves (4.169), then rn is given by

⎧6 - ⎛ ⎞- ⎫
⎨7 - - ⎬
7 n
- n
-
rn = max 8 |ρj |2 : -
-A⎝ ρj x̂j ⎠-
- =1 . (4.173)
⎩ - - ⎭
j=1 j=1
For further numerical aspects of least squares method, we refer, e.g., to [57,
58, 134, 157, 173, 178].
Dual least squares method. As another variant of the Galerkin method, we come to
the dual least squares method. In this case the boundedness condition (4.150) is in-
deed always satisfied: Given some finite-dimensional subspaces Yn ⊂ Y , determine
un ∈ Yn such that
A∗ un , zn = y, zn for all zn ∈ Yn , (4.174)
where, as always, A∗ : Y → X denotes the adjoint of A. Then xn = A∗ zn is called
the dual least squares solution. It is a special case of the Galerkin method, where
Xn = A∗ (Yn ). Writing (4.174) for y = Ax in the form
A∗ un , A∗ zn = x, A∗ zn for all zn ∈ Yn , (4.175)
we see that the dual least squares method is just the least squares method for the
equation A∗ u = x. This explains the standard terminology in the literature.
Suppose now that the right-hand side is perturbed. Let y ε ∈ Y be given
such that y ε − y ≤ ε. Instead of the linear equation (4.174), we determine
xεn := A∗ uεn ∈ Xn via
A∗ uεn , A∗ zn = y ε , zn for all zn ∈ Yn . (4.176)
For discrete perturbations, we specify a basis {ŷj , j = 1, . . . , n} of Yn and assume
that the right-hand side βi = y, ŷi Y , i = 1, . . . , n, of the resulting linear equations
are perturbed by a vector β ε ∈ Kn with |β ε −β| ≤ ε, where |·| denotes the Euclidean
norm in Kn . Instead of (4.174) we are then led to

n
xεn = A∗ uεn = αεj A∗ ŷj , (4.177)
j=1
where αε ∈ Kn solves the linear equation

n
αεj A∗ ŷj , A∗ ŷi = βiε , i = 1, . . . , n. (4.178)
j=1
Results on convergence and error estimates are listed, e.g., in the textbook [137].
Let X and Y be Hilbert spaces. Suppose that A : X → Y is linear, bounded,
an one-to-one such that the range; A(X) is dense in Y . Let Yn ⊂ Y form finite-
dimensional subspaces such that n∈N Yn is dense in Y . Assume that x ∈ X is the
solution of Ax = y. Then the linear equations (4.176) and (4.178) are uniquely
solvable for every right-hand side and every n ∈ N. Furthermore, the dual least
squares method is convergent, and we have
Rn ≤ σn = max{zn : zn ∈ Yn , A∗ zn = 1}. (4.179)
Moreover, we have the error estimate
x − xεn ≤ rn σn ε + c min{x − zn : zn ∈ A∗ (Yn )} (4.180)
for some c > 0. Here, rn = 1 if ∈ Xn solves (4.176), i.e., ε measures
rnε n the norm
y ε − y in Y . If ε measures the discrete error |β ε − β| and xδn = j=1 αεj A∗ ŷj ∈
Xn , where αε solves (4.178), then rn is given by
⎧6 - - ⎫
⎨7 - - ⎬
7 n
- n -
rn = max 8 |ρj |2 : -
- ρ j ŷ j
-=1
- (4.181)
⎩ - j=1 - ⎭
j=1
(note that rn = 1 if {ŷj , j = 1, . . . , n} constitutes an orthonormal system in Y ).

Bubnov–Galerkin method. We assume that A : X → X is a linear and bounded
operator and Xn , n ∈ N, are finite-dimensional subspaces. The Galerkin method
amounts to the problem of determining xn ∈ Xn such that
Axn , zn = y, zn for all zn ∈ Xn . (4.182)
This special case is called the Bubnov–Galerkin method. Again, we consider the
perturbation of the right-hand side. If y ε ∈ Y is chosen such that y ε − y ≤
ε represents a perturbed right-hand side, then instead of (4.182) we study the
equation system
Axεn , zn = y ε , zn for all zn ∈ Xn . (4.183)
An alternative is to choose a basis {x̂j , j = 1, . . . , n} of Xn . We assume
that the right-hand side βi = (y, ŷi ), i = 1, . . . , n of the Galerkin equations are
perturbed by a vector β ε ∈ Kn with |β ε − β| ≤ ε, where | · | denotes again the
Euclidean norm in Kn . In this case, instead of (4.182), we have to solve

n
aεj Ax̂j , x̂i = βiε for i = 1, . . . , n. (4.184)
j=1
n
For αε ∈ Kn we set xεn = j=1 αεj x̂j .
Next we show that the Rayleigh–Ritz method, in fact, is a special case of the
Bubnov–Galerkin method.
Rayleigh–Ritz method. Let A : X → X be also self-adjoint and positive
definite, so that Ax, yX = x, AyX and Ax, xX > 0 for all x, y ∈ X with
x = 0. We introduce the functional
J(z) = Az, z − 2Rey, z for z ∈ X. (4.185)
The identity
J(z) − J(x) = 2ReAx − y, z − x + A(z − x), z − x (4.186)
and the positivity of A tells us that x ∈ X is the unique minimum of J if and
only if x solves Ax = y. The Rayleigh–Ritz method is to minimize J over the
finite-dimensional subspace Xn . From (4.186), we see that if xn ∈ Xn minimizes
J on Xn , then , for zn = xn ± εun with un ∈ Xn and ε > 0, it follows that
0 ≤ J(zn ) − J(xn ) = ±ε 2ReAxn − y, un + ε2 Aun , un
for all un ∈ Xn . By dividing ε > 0 and afterwards letting ε → 0 we find that
xn ∈ Xn satisfies the equation system (4.182). If, on the other hand, xn ∈ Xn
solves (4.182), then we get from (4.186),
J(zn ) − J(xn ) = A(zn − xn ), zn − xn ≥ 0
for all zn ∈ Xn . Therefore, the Rayleigh–Ritz method coincides with the Bubnov–
Galerkin method.
Finally we are interested in the Bubnov–Galerkin method for the important
class of coercive operators. As preparatory material we briefly recapitulate some
settings (see, e.g., [137]):
(i) A Gelfand triple (V, X, V ∗ ) consists of a reflexive Banach space V , an Hilbert
space X, and the dual space V ∗ of V such that
(a) V is a dense subspace of X,
(b) the imbedding J : V → X is bounded.
It is conventional to write (see, e.g., [137]) V ⊂ X ⊂ V ∗ since we can identify
X with a dense subspace of V ∗ . This identifications given by the dual operator
J ∗ : X → V ∗ of J, where we identify the dual of the Hilbert space X by
itself. From (x, y) = J ∗ x, y, for all x ∈ X and y ∈ V we see that with this
identification the dual pairing ·, · in (V ∗ , V ) is an extension of the inner
product (·, ·) in X, i.e., we write
x, y = (x, y) for all x ∈ Y and y ∈ V.
Furthermore, we have the estimates
|x, y| ≤ xV ∗ yV for all x ∈ V ∗ , y ∈ V,
thus,
|x, y| ≤ xV ∗ yV for all x ∈ X, y ∈ V.
∗
It is well known that J is one-to-one and has a dense range.
(ii) Let V be a reflexive Banach space with dual space V ∗ . We denote the norms
in V and V ∗ by · V and · V ∗ , respectively. A linear bounded operator
A : V ∗ → V is called coercive if there exists γ > 0 with
Rex, Ax ≥ γ x2V for all x ∈ V ∗ , (4.187)
where ·, · denotes the dual pairing in (V ∗ , V ).
(iii) The operator A satisfies Garding’s inequality if there exists a linear compact
operator C : V ∗ → V such that A + C is coercive, i.e.,
Re(x, Ax) ≥ γ x2V ∗ − Rex, Cx for all x ∈ V ∗ .
Note that, by the same argument as in the Lax–Milgram theorem, it can
be shown that every coercive operator is an isomorphism from V ∗ onto V .
Coercive operators play an important role in the study of partial differential
equations and integral equations by variational methods. In the conventional
definition, the roles of V and V ∗ are interchanged. For integral operators that
are “smoothing”, our definition seems to be more appropriate. However, both
definitions are equivalent in the sense that the inverse operator A−1 : V → V ∗
is coercive in the usual sense with γ replaced by γ/A2 .
Convergence of the Bubnov–Galerkin method. After these preparations we are in
the position to formulate convergence of the Bubnov–Galerkin method for coercive
operators (see [137]).
Let (V, X, V;∗ ) be a Gelfand triple, and Xn ⊂ V be finite-dimensional sub-
spaces such that n∈N Xn is dense in X. Let K : V ∗ → V be coercive with constant
γ > 0. Let x ∈ X be the solution of Ax = y. Then we have the following results:
(a) There exist unique solutions of the Galerkin equations (4.182)–(4.184), and
the Bubnov–Galerkin method converges in V ∗ with
x − xn V ∗ ≤ c min{x − zn V ∗ : zn ∈ Xn } (4.188)
for some c > 0.
(b) Define the quantity ρn > 0 by
ρn = max{u : u ∈ Xn , uV ∗ = 1} (4.189)
and the orthogonal projection operator Pn from X onto Xn . The Bubnov–
Galerkin method converges in X if there exists c > 0 with
c
u − Pn uV ∗ ≤ u for all u ∈ X. (4.190)
ρn
In this case, we have the estimates
1 2
Rn ≤ ρ (4.191)
γ n
and
x − xεn ≤ c[rn ρ2n + min{x − zn : zn ∈ Xn }] (4.192)
for some c > 0. Here rn = 1 if xεn ∈ Xn solves (4.183), i.e., ε measures the norm
y ε −
y in X. If ε measures the discrete error |β ε − β| in the Euclidean norm and
n
xεn = j=1 αεj x̂j ∈ Xn , where αε solves (4.184), then rn is given by
⎧6 ⎫
⎨77 n
n ⎬
rn = max 8 |ρj |2 : ρj x̂j = 1 . (4.193)
⎩ ⎭
j=1 j=1
Again, we note that rn = 1 if {x̂j , j = 1, . . . , n} forms an orthonormal system in

X. For further details, we refer to [200] and the monographs [22, 137, 139, 148].
4.3. Multiscale methods as regularization schemes

Next a compact operator equation is dealt within regularization methods, based
on filtering techniques by means of wavelets. In a general setup a singular integral
approach to regularization is established, decomposition/reconstruction regulariza-
tion wavelets are introduced in the frequency space which allow the regularization
in form of a multiresolution analysis. Two different types of regularization wavelets
are discussed in more detail, namely (non-locally supported) Tikhonov–Phillips
regularization wavelets and (bandlimited) truncated singular value decomposition
wavelets. Our considerations closely follow Freeden, Schneider [77] about regular-
ization and multiresolution.
Let (X, ·, ·) and (Y, ·, ·) be separable real functional Hilbert spaces over
domains GX and GY , respectively, i.e., X, respectively, Y consists of functions
x : GX → R, respectively, y : GY → R. We consider a linear, compact operator
A : X → Y satisfying N (A) = {0}, R(A) Y , R(A) = Y, so that A∗ A : X → X
is self-adjoint.
In the sequel, we denote by {xn }n=0,1,... a complete orthonormal system in
(X, ·, ·X ) and by {yn }n=0,1,... a complete orthonormal system in (Y, ·, ·Y ) such
that the singular values {σn }n=0,1,... of A satisfy Axn = σn yn , A∗ yn = σn xn , n ∈
N0 . Since A is supposed to be injective it follows that σn > 0 for all n ∈ N0 .
Any p ∈ P , P ∈ {X, Y }, can be represented in terms of an orthonormal (Fourier)
expansion with respect to {pn }n=0,1,...
∞

p= p∧
P (n)pn (4.194)
n=0
with
p∧
P (n) = p, pn P , (4.195)
where the equality in (4.194) is understood in the · P -sense. In conclusion, any
element of R(A) admits an expansion of the form
∞
Ax = σn x∧
X (n)yn . (4.196)
n=0
As is well known, the Picard condition tells us that the problem
Ax = y, x ∈ X, y ∈ Y (4.197)
has a solution if and only if y ∈ Y satisfies

∞

(σn−1 yY∧ (n))2 < ∞. (4.198)
n=0
In this case it is known that the solution of (4.197) is representable in the form
∞

x= σn−1 yY∧ (n)xn = A† y. (4.199)
n=0
Since the right-hand side y is error affected (due to the inaccuracy of the
measuring instrument) in any practical application, the series (4.199) will not
converge in general. For that purpose we are interested in regularized solutions.
The idea (cf. [77]) we follow is to represent the J-level regularization of the
problem
Ax = y, x ∈ X, y ∈ Y (4.200)
by means of a wavelet analysis. The overall advantage of such a method is that
we obtain a J + 1-level regularization by starting with the J-level regulariza-
tion and adding so-called detail information. It becomes clear that any classical
regularization method based on a filtered singular value decomposition can be re-
formulated in terms of our wavelet method. Thus, any known parameter choice
strategy depending on the special method is also applicable and, moreover, any of
the corresponding error estimates holds true, too. For that reason we omit these
discussions here.
As is well known, a family {RJ }J∈Z of linear operators RJ : Y → X, J ∈ Z,
is a regularization of A† if it satisfies the following properties:
(i) RJ is bounded on Y for all J ∈ Z,
(ii) for any member y ∈ R(A), the limit relation limJ→∞ RJ y = A† y holds in
the · X -sense.
The kernel xJ = RJ y is called the J-level regularization of the problem (4.197).
Product kernels. A function ΓP,Q ·, · : GP × GQ → R, P, Q ∈ {X, Y }, of the
form
∞
P,Q
Γ (x, y) = Γ∧ (n) pn (x) qn (y), x ∈ GP , y ∈ GQ , Γ∧ (n) ∈ R, n ∈ N0 ,
n=0
(4.201)
is called a (P, Q)-(product) kernel. Note that the indices P and Q in ΓP,Q are
associated to the variables x and y, respectively, such that ΓP,Q (x, y) = ΓQ,P (y, x).
The sequence {Γ∧ (n)}n=0,1,... is called the symbol of the (P, Q)-kernel. For brevity,
a (P, P )-kernel is simply said to be a P -kernel.
An important question for our investigations is as follows: Fix one variable
of a product kernel, what are the conditions for the product kernel (as a function
of the remaining variable) to be of class P ∈ {X, Y }? The answer is provided by
the concept of symbol admissibility.
A symbol {Γ∧ (n)}n=0,1,... is called P -admissible, P ∈ {X, Y }, if it satisfies

the following conditions:
∞
2
(Ai) (Γ∧ (n)) < ∞, (4.202)
n=0
∞

∧ 2
(Aii) sup (Γ (n)pn (x)) < ∞. (4.203)
x∈GP n=0
A symbol {Γ∧ (n)}n=0,1,... is called ((P, Q)-admissible, P, Q ∈ {X, Y }, or simply)

admissible, if it is P -admissible as well as Q-admissible.
From the definition of admissibility we immediately obtain the result:
Let {Γ∧ (n)}n=0,1,... be the symbol of an (P, Q)-kernel, P, Q ∈ {X, Y }.

(α) If {Γ∧ (n)}n=0,1,... is P -admissible, then ΓP,Q (x, ·) ∈ Q for every (fixed)
x ∈ GP ,
(β) If {Γ∧ (n)}n=0,1,... is Q-admissible, then ΓP,Q (·, y) ∈ P for every (fixed)
y ∈ GQ .
Convolutions. A fundamental tool for our wavelet theory is the concept of
a convolution we introduce below: Let ΓX,X be an X-kernel with X-admissible
symbol. Suppose that F is of class X. Then we understand the convolution of
ΓX,X and F to be the function given by
∞

? @
(ΓX,X ∗ x)(t) = ΓX,X (t, ·), x X = Γ∧ (n)x∧
X (n)xn (x), x ∈ GX . (4.204)
n=0
We immediately see that (ΓX,X ∗ x)∧ ∧ ∧

X (n) = Γ (n)xX (n) and Γ
X,X
∗ x ∈ X. In
X,Y
analogous way we define the convolution of an (X, Y )-kernel Γ having an X-
admissible symbol with a function y ∈ Y to be the expression
∞

? @
(ΓX,Y ∗ y)(t) = ΓX,Y (t, ·), y Y = Γ∧ (n)yY∧ (n)xn (t), t ∈ GX , (4.205)
n=0
and it follows that ΓX,Y ∗ y ∈ X. Lastly, the convolution of an (X, Y )-kernel ΓX,Y
having an Y -admissible symbol with a function F ∈ X is given by
∞

? @
(ΓX,Y ∗ x)(s) = ΓX,Y (·, s), x X = Γ∧ (n)x∧
X (n)yn (s), s ∈ GY , (4.206)
n=0
and we have ΓX,Y ∗ x ∈ Y . Next we proceed with the convolution of two product
kernels leading to the following result: Let ΓX,X be an X-kernel with X-admissible
symbol and let ΥX,Y be an (X, Y )-kernel with Y -admissible symbol. Then
(Γ ∗ Υ)X,Y (t, s) = (ΓX,X ∗ ΥX,Y (·, s))(t)
? @
= ΓX,X (t, ·), ΥX,Y (·, s) X
∞

= Γ∧ (n)Υ∧ (n)xn (t)yn (s), t ∈ GX , s ∈ GY (4.207)
n=0
represents an (X, Y )-kernel with admissible symbol

((Γ ∗ Υ)X,Y )(n) = Γ∧ (n)Υ∧ (n). (4.208)
Dilation and shifting. In order to prepare the fundamentals of the forthcoming

wavelet theory we are now interested in countable families {ΓP,Q J }, J ∈ Z, of
product kernels ΓP,Q
J , P, Q ∈ {X, Y }. Observing our notations we are able to
define a dilation operator acting on these families in the following way: let ΓP,Q J
be a member of the family of product kernels.
Then the dilation operator DK , K ∈ Z is defined by DK ΓP,Q J = ΓP,Q
J+K .
P,Q P,Q P,Q
Especially, we obtain ΓJ = DJ Γ0 , J ∈ Z. Thus we refer Γ0 to as a “mother
kernel”.
Moreover, we define a shifting operator StP , t ∈ GP , P ∈ {X, Y }, by StP ΓP,Q J =
ΓP,Q
J (t, ·), t ∈ GP , J ∈ Z, resp. S Q P,Q
Γ
s J = Γ P,Q
J (·, s), s ∈ G Q , J ∈ Z. In doing
so we consequently get ΓP,Q P,Q P,Q
J (t, ·) = St DJ Γ0 , t ∈ GP , J ∈ Z, resp. ΓJ (·, s) =
P
P,Q
SsQ DJ Γ0 , s ∈ GQ , J ∈ Z.
Regularization scaling functions. Next we are concerned with a wavelet based
regularization technique of problem (4.200): Let {(ΦJ )∧ (n)}n=0,1,... , J ∈ Z, be an
(X, Y )-admissible symbol of a family of product kernels which additionally satisfies
the following properties:
i) limJ→∞ σn ((ΦJ )∧ (n))2 = 1, n ∈ N,
ii) ((ΦJ+1 )∧ (n))2 ≥ ((ΦJ )∧ (n))2 , J ∈ Z, n ∈ N,
iii) limJ→−∞ ((ΦJ )∧ (n))2 = 0, n ∈ N,
iv) σ0 ((ΦJ )∧ (0))2 = 1, J ∈ Z.
Then {(ΦJ )∧ (n)}n=0,1,... is said to be the generating symbol of a regularization
scaling function (with respect to (4.197)). The (X, Y )-kernel
∞

d
ΦX,Y
J = (ΦJ )∧ (n)xn yn (4.209)
n=0
is called a decomposition regularization scaling function, the (X, X)-kernel

∞

r X,X
ΦJ = (ΦJ )∧ (n)xn xn (4.210)
n=0
is called a reconstruction regularization scaling function.

From the results of the previous subsection it is clear that d ΦX,Y

J (t, ·) ∈ Y, t ∈
r X,X
GX , J ∈ Z, ΦJ (t, ·) ∈ X, t ∈ GX , J ∈ Z, and ( ΦJ ∗ ΦJ )
r d X,Y
is an (X, Y )-kernel
with (X, Y )-admissible symbol {((ΦJ )∧ (n))2 }n=0,1,... . Observing these properties
we are able to verify the following result which is central for our considerations:
Let {(ΦJ )∧ (n)}n=0,1,..., J ∈ Z, be the generating symbol of a regularization
scaling function. Then, for any y ∈ Y ,
xJ = rΦX,X
J ∗ (d ΦX,Y
J ∗ y) (4.211)
represents the J-level regularization of problem (4.200). If, in addition, y ∈ R(A),
then
lim xJ − A† y = 0. (4.212)
J→∞
For RJ : Y → X defined by
X,Y
RJ = rΦJ ∗d ΦJ ∗ y, y ∈ Y
we have (see [77])
∞

RJ y = ((ΦJ )∧ (n))2 yY∧ (n) xn
n=0
and
∞

2
RJ ≤ ((ΦJ )∧ (n))4 , J ∈ Z. (4.213)
n=0
As an immediate consequence of our results we obtain the result: Let y be a
member of R(A). Suppose that y ε ∈ Y denotes the right-hand side of problem
(4.200) with noise level y − y ε ≤ ε. Moreover, assume that {(ΦJ )∧ (n)}n=0,1,... ,
J ∈ Z, is the generating symbol of a regularization scaling function, where the
parameter J = J(ε) is assumed to satisfy
(1) limε→0 J(ε) = ∞,
(2) limε→0 ε((ΦJ )∧ (n))2 = 0.
Then we have
lim (rΦJ ∗ d ΦJ )X,Y ∗ y ε − A† y = 0. (4.214)
ε→0
Condition iii) seems to be unnecessary for the proof of (4.214) and, in fact, it
is. Nevertheless, in what follows we need this assumption for our multiresolution
analysis and the (spectral) introduction of wavelets.
Multiresolution analysis. For any y ∈ R(A) each (rΦJ ∗ d ΦJ )X,Y ∗ y provides a
regularization of the solution A† y at scale J by ”smoothing” the Fourier coefficients
of A† y with the symbol {((ΦJ )∧ (n))2 }n=0,1,... . In terms of filtering, (rΦJ ∗ d ΦJ )X,Y
may be interpreted as a low-pass filter. Accordingly we understand the scale spaces
VJ to be the image of R(A) under the operator RJ :
VJ = RJ (R(A)) = {(rΦJ ∗ d ΦJ )X,Y ∗ y : y ∈ R(A)}. (4.215)
This leads us to the properties formulated in the following statement:
The scale spaces satisfy the following properties:

i) {X0 } ⊂ VJ ⊂ VJ ⊂ X, J ≤ J , i.e., for any right-hand side y ∈ R(A)
of problem (4.200), all J-level regularizations with fixed parameter J are
sampled
9∞ in a scale space VJ with the above property,
ii) J=−∞ VJ = {x0 },
;∞ ·X
iii) J=−∞ VJ = X,
iv) if xJ ∈ VJ , then D−1 xJ ∈ VJ−1 , J ∈ Z.
If a collection of subspaces of X satisfies the above conditions we call them
a regularization multiresolution analysis (RMRA).
Regularization wavelet functions. The definition of the regularization scaling func-

tion allows us to introduce regularization wavelets. An essential point is the def-
inition of a decomposition and a reconstruction regularization wavelet associated
to regularization mother wavelets. This definition, of course, has to be formulated
in close relation to a prescribed regularization scaling function.
Let {(Φj )∧ (n)}n=0,1,... , j ∈ Z, be the generating symbol of a regularization

scaling function. Then the (X, Y )-admissible generating symbol {(Ψj )∧ (n)}n=0,1,... ,
j ∈ Z, and the (X, X)-admissible generating symbol {(Ψ̃j )∧ (n)}n=0,1,... , j ∈ Z, re-
spectively, are defined by the “scaling equation”
(Ψ̃j )∧ (n)(Ψj )∧ (n) = ((Φj+1 )∧ (n))2 − ((Φj )∧ (n))2 . (4.216)
Correspondingly, the (X, Y )-kernel
∞

d
ΨX,Y
J = (ΨJ )∧ (n) xn yn (4.217)
n=0
is called the decomposition regularization wavelet, while the (X, X)-kernel

∞

r
Ψ̃X,X
J = (Ψ̃J )∧ (n) xn xn (4.218)
n=0
is called the reconstruction regularization wavelet. The corresponding regulariza-

tion mother wavelets are denoted by d ΨX,Y0 and r Ψ̃X,X
0 , respectively.
Using this notation, any decomposition regularization wavelet, respectively,

any reconstruction regularization wavelet can be interpreted as a dilated and
shifted copy of the corresponding mother wavelet.
d
ΨX,Y
J (t, ·) = StX DJ d ΨX,Y
0 , (4.219)
r
ΨJ Ψ̃X,X
J (t, ·) = StX DJ rΨX,X
0 . (4.220)
At this stage it becomes obvious why we required Condition iv). A conclusion of

(4.216) is a vanishing 0th moment of the regularization wavelets. Moreover, from
(4.216) it can be readily deduced that
2

J
2

J
((Φ0 )∧ (n + 1)) = (Ψ̃j )∧ (n)(Ψj )∧ (n) = ((Φ0 )∧ (n)) + (Ψ̃j )∧ (n)(Ψj )∧ (n).
j=−∞ j=0
(4.221)
Thus, we easily see in connection with 4.216 that
X,Y J
J
r
Φ̃J+1 ∗ d ΦJ+1 = (r Ψ̃j ∗ d Ψj )X,Y = (rΦ0 ∗ d Φ0 )X,Y + (r Ψ̃j ∗ r Ψj )X,Y .
j=−∞ j=0
(4.222)
In analogy to the definition of the operator RJ we consider now convolution op-
erators SJ : Y → X, J ∈ Z, defined by
SJ y = (r Ψ̃J ∗ d ΨJ )X,Y ∗ y. (4.223)
It describes the “detail information” of the right-hand side y at scale J. From
Equation (4.222) it follows that the operator RJ+1 can be decomposed in the
following way:
J
RJ+1 = R0 + Sj . (4.224)
j=0
But this gives rise to introduce the detail spaces as follows:
WJ = SJ (R(A)) = {(r Ψ̃J ∗ d ΨJ )X,Y ∗ y : y ∈ R(A)}. (4.225)
The space WJ contains the “detail information” needed to go from a regularization
at level J to a regularization at level J + 1. Note that
VJ = VJ−1 + WJ−1 , (4.226)

J
J
Wj = V 0 + Wj = VJ+1 . (4.227)
j=−∞ j=0
It is worth mentioning that, in general, the sum decomposition is neither direct

nor orthogonal. Two examples leading to orthogonal and non-orthogonal multires-
olution are introduced in the next subsection.
Any right-hand side y ∈ Y can now be decomposed as follows. Starting with
R0 y we find
J
RJ+1 y = R0 y + Sj y (4.228)
j=0
for any J ∈ Z. In other words, the partial reconstruction RJ y is nothing else than
the “difference of two smoothings” at two consecutive scales,
SJ y = RJ+1 y − RJ y. (4.229)
For what follows we define the regularization wavelet transform at scale J ∈ Z
and position t ∈ GX by letting
RW T (y)(J; t) = d ΨX,Y
J (t, ·), yY , y ∈ Y. (4.230)
From (4.222) it is not hard to verify the main result in this context:
Let {(ΦJ )∧ (n)}n=0,1,... , J ∈ Z, be the generating symbol of a regularization
scaling function. Suppose that
{(Ψj )∧ (n)}n=0,1,... , {(Ψ̃j )∧ (n)}n=0,1,... , j∈Z
are the generating symbols of the corresponding regularization wavelets. Further-
more, let y be of class Y . Then,

J
xJ = (rΦ0 ∗ d Φ0 )X,Y ∗ y + r
Ψ̃j ∗ RW T (y)(j; ·) (4.231)
j=0
denotes the J-level regularization of the problem (4.200) satisfying

lim xJ − A† y = 0 (4.232)
J→∞
provided that y ∈ R(A).
Equation (4.231) shows the essential characteristic of regularization wavelets.
By adding the so-called detail information of level J as the difference of two
smoothings of two consecutive scales J + 1 and J, we change the regularized
solution from xJ to xJ+1 thereby satisfying limJ→∞ xJ = A† y in the case that
y ∈ R(A). Of course, this can be understood as a kind of a permanence principle.
Table: The multiscale method as presented here can be illustrated by the scheme
shown in Figure 4.1.
R0 (y) R1 (y) R2 (y) ... −→ A† y

j→∞
V0 ⊂ V1 ⊂ V2 ... = A† X
V0 + W0 + W1 + W2 ... = A† X
R0 (y)+ S0 (y) + S1 (y) + S2 (y)+ ... = A† y
Figure 4.1. Multiresolution regularization scheme.
Some generating symbols. The singular values {σn }n=0,1,... of A satisfy Axn =
σn yn , A∗ yn = σn xn , n ∈ N0 . Keeping these facts in mind we are led to introduce
the following examples of generating symbols of a regularization scaling function:
i) Truncated singular value decomposition (bandlimited regularization).
a) orthogonal:
−1/2
∧ σn for n = 0, . . . , NJ
(ΦJ ) (n) = , (4.233)
0 for n ≥ NJ + 1

0 for J ∈ Z, J < 0
NJ = , (4.234)
2J − 1 for J ∈ Z, J ≥ 0
b) non-orthogonal:
⎧
⎪ −1/2
⎨ σn for n = 0, . . . , MJ
(ΦJ )∧ (n) = σ
−1/2
(τ (n)) 1/2
for n = MJ + 1, . . . , NJ , (4.235)
⎪
⎩ 0
n J
for n ≥ NJ + 1

0 for J ∈ Z, J < 0
NJ = ,
2J+1 − 1 for J ∈ Z, J ≥ 0
(4.236)
0 for J ∈ Z, J < 0
MJ =
2J − 1 for J ∈ Z, J ≥ 0
and
τJ (n) = 2 − 2−J (n + 1), n ∈ [2J − 1, 2J+1 − 1], J ∈ N0 . (4.237)
It is easy to see that case a) leads to an orthogonal RMRA, i.e., the detail
and the scale spaces satisfy the orthogonality conditions
VJ+1 = VJ ⊕ WJ , WJ ⊥ WK , K = J, K, J ≥ 0. (4.238)
In case b) the scale and detail spaces are still finite dimensional, but the
detail spaces are no longer orthogonal.
ii) Tikhonov’s regularization (non-bandlimited regularization).
a) classical
12
σn
(ΦJ )∧ (n) = , n ∈ N, J ∈ Z, (4.239)
σn2 + γJ2
b) Tikhonov–Phillips
12
∧ σn
(ΦJ ) (n) = , n ∈ N, J ∈ Z (4.240)
σn + γJ (n + 14 )4
2 2
with {γJ }, J ∈ Z, being a sequence of real numbers satisfying limJ→∞ γJ = 0

and limJ→−∞ γJ = ∞. Case a) leads to the minimization of the Tikhonov
functional, where the penalty term is given by the norm of the regularized so-
lution of problem (4.200). Case b) leads to the minimization of the Tikhonov–
Phillips functional, where the penalty term is given by the linearized spherical
bending energy of the regularized solution of problem (4.200). Both cases lead
to infinite-dimensional scale and detail spaces and, furthermore, the RMRA
is neither direct nor orthogonal.
4.4. Iterative methods as regularization schemes

From linear algebra we borrow the idea to use iterative solvers of the equation
Ax = y, y ∈ Y , in the following way: Suppose that the matrix A ∈ Kn×n is
written in the form A = B − C with an invertible B. Then we are able to write
Ax = y, in equivalent form
Ax = y ⇔ (B − C) x = y ⇔ Bx = Cx + y ⇔ x = B −1 Cx + B −1 y. (4.241)
Let us base the determination of a (uniquely determined) fixed point on the fol-
lowing iteration:
x(0) ∈ Kn , arbitrary,
Bx(n+1) = Cx(n) + y, n ∈ N0 .
! (n) "
Clearly, if x converges to x∗ ∈ Rn , then it follows that Ax∗ = y.
In numerical linear algebra the following procedures are convenient which
should be recapitulated here: Let A ∈ Rn×n be given. Let us decompose the
matrix A in the form A = L + D + U such that
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0 a11 0 0 ∗
⎜ .. ⎟ ⎜ .. ⎟ ⎜ .. ⎟
L=⎝ . ⎠ D=⎝ . ⎠ U =⎝ . ⎠ (4.242)
∗ 0 0 ann 0 0
We distinguish the following concepts:
(1) (Jakobi Method) We let
B = D, C = −L − U. (4.243)
Then we are led to
x(n+1) = D−1 (L + U ) x(n) + D−1 y, n ∈ N. (4.244)
(2) (Gauss–Seidel Method) We choose
B = D + L, C = −U. (4.245)
Then we are led to
−1 −1
x(n+1) = (D + L) U x(n) + (D + L) y, n ∈ N. (4.246)
(3) (Richardson Method) For ω > 0, let
1 1
B = I, C = I − A. (4.247)
ω ω
Then we are led to the recursion
−1 −1
1 1 1
x(n+1) = I I − A x(n) + I y
ω ω ω

1
=ω I − A x(n) + ωy
ω
= (I − ωA) x(n) + ωy, n ∈ N. (4.248)
Since iteration methods known from linear algebra that use the upper or
lower triangular part of a matrix cannot be adequately transferred to operator
equations, we are not able to use the first two of the aforementioned iteration
procedures. However, the third method (i.e., Richardson method or successive
relaxation method) can be applied leading to a variant called Landweber iteration.
Landweber iteration. In order to solve the normal equation
A∗ Ax = A∗ y, y ∈ Y, x ∈ X, (4.249)
the Landweber iteration starting from the initial value x(0) ∈ X is defined by
x(m+1) = (I − ωA∗ A) x(m) + ωA∗ y

= x(m) + ωA∗ y − Ax(m) , m = 0, 1, . . . (4.250)
For simplicity, we introduce the family of operators {Rm }m∈N ⊂ L (Y, X) by

Rm y = x(m) :
2
Let 0 < ω < and A ∈ L (X, Y ). Then, for x(0) ∈ X, we have
A2
†
A y + PN (A) x(0) y ∈ D A†
lim Rm y = (4.251)
m→∞ ∞ y∈/ D A†
The Landweber iteration is characterized by
x(m) = Rm y = Fm (A∗ A) A∗ y + (I − ωA∗ A)m x(0) , (4.252)
where

m−1
1 − (1 − ωλ)m
Fm (λ) = ω (1 − ωλ)j = . (4.253)
j=0
λ
Thus, the Landweber iteration represents a regularization with a filter that cannot
be used as such if we do not know the singular values of our problem.
m
The polynomial λ → pm (λ) = 1 − λFm (λ) = (1 − ωλ) is called residual
polynomial.
• For x(0) = 0 it follows that A† y is the solution of the normal equation
A∗ Ax = A∗ y (4.254)
showing minimal norm (as usual). This result is also obtained for x(0) ∈
N (A)⊥ , since then PN (A) x(0) = 0.
• For x(0) = 0 we see that A† y + PN (A) x(0) is the solution of the normal
equation
A∗ Ax = A∗ y (4.255)
with minimal distance to x(0) , i.e., the x(0) -minimum norm solution.
- - - -
- † - - -
-A y + PN (A) x(0) − x(0) - = min -x − x(0) - : A∗ Ax = A∗ y (4.256)
Obviously, the choice of the initial value of the Landweber iteration for inverse
problems is of great importance for the solution (see, e.g., [227]). We investigate
the same choice for a noiseless and noisy right-hand side.
! " ! "
Let A ∈ L (X, Y ), y, y ε ∈ Y with y − y ε Y < ε. Let x(m) and x(m),ε
be the Landweber iteration sequences with respect to y and y ε corresponding
to the
(0) (0),ε 2
same initial value x = x ∈ X and the same parameter ω ∈ 0, A2 . Then
- - √
- (m) (m),ε - 2ω , m = 1
-x −x -≤ε √ (4.257)
mω , m ≥ 2.
The estimate (4.257) suggests that the parameter ω should be chosen as small
as possible. However, we have seen that the convergence of the approximation
error depends on the estimate |1 − ωλ| < 1. As a consequence, for small ω the
approximation error converges very slowly.
In more details, we separate the total error in the usual way
- † - - -
-A y − Rm y ε - = -A† y − Rm y - + Rm y − Rm y ε (4.258)

approximation data
error error
- - m→∞ √
with -A† y − Rm y - −→ 0 if x(0) = 0 and Rm y − Rm y ε ≤ mω ε, i.e., the
Landweber iteration is a regularization with a suitably chosen m. Our aim now is
to find a stopping rule for m.
In fact, we obtain for the residual term

Ax(m),ε − y ε = A (I − ωA∗ A) x(m−1),ε + ωA∗ y ε − y ε
= Ax(m−1),ε − ωAA∗ Ax(m−1),ε + ωAA∗ y ε − y ε
= (I − ωAA∗ ) Ax(m−1),ε − (I − ωAA∗ ) y ε

= (I − ωAA∗ ) Ax(m−1),ε − y ε . (4.259)
2
Under the choice 0 < ω < A 2 we have
- - - -
- (m),ε - - -
-Ax − y ε - ≤ I − ωAA∗ · -Ax(m−1),ε − y ε -
- -
- -
≤ -Ax(m−1),ε − y ε - , (4.260)
i.e., the norms

of the residuals are monotonously decreasing in m. Furthermore,
/ D A† , it follows that Ax(m−1),ε − y ε ∈
if y ε ∈ / N (A∗ ) (due to the fact that
y ε ∈ D(A† )).
2
If 0 < ω < A2
/ N (A∗ ), then
and y ∈
(I − ωAA∗ ) y < y . (4.261)

- - - -
/ D A† , then -Ax(m),ε − y ε - < -Ax(m−1),ε − y ε -, i.e., the residual
If y ε ∈
is strictly monotonously decreasing. Hence, we are confronted with the typical
dilemma of ill-posed problems that a small residual terms does not imply a small
error. The monotonicity of the residual term suggests to use a discrepancy principle
as a kind of “stopping rule”. This observation goes back to [38]. In more detail,
let τ > 1 be fixed. We are interested in determining m∗ ∈ N0 , such that
- (m∗ ),ε - - -
-Ax − y ε - ≤ τ ε < -Ax(m),ε − y ε - m = 0, 1, . . . , m∗ − 1. (4.262)
The discrepancy principle (4.262) seems to be particularly suitable for Land-
weber iteration, since a residual term larger than 2ε implies the monotonicity of
the error:
Suppose -that A ∈ L (X, Y ), y ∈ R (A) and y ε ∈ Y with y − y ε < ε. If

- (m),ε
-Ax − y ε - > 2ε and 0 < ω < A1
2 , then
- † - - -
-A y − x(m+1),ε - < -A† y − x(m),ε -, m ∈ N0 , (4.263)
i.e., the error is also strictly monotonously decreasing.
In other words, as already announced the error decreases monotonously just
like the residual as long as the residual stays larger than 2ε.
An upper bound for the number of iterations when using the discrepancy
principle is as follows (cf. [227]):
Assume that A ∈ L (X, Y ), y ∈ R (A), and y ε ∈ Y with y − y ε < ε.
1
Suppose that 0 < ω < A2 . The discrepancy principle (4.262) with τ > 1 yields
the stopping index m∗ = m∗ (ε, y ε ) ≤ CL ε−2 with the constant CL > 0 for the
Landweber iteration.
Note that this result does not require any assumptions on the smoothness of
the solution. With such information the discrepancy principle enables us to stop
the iteration much earlier.
2
Suppose that A ∈ L(X, Y ), 0 < ω < 2/ A , y ∈ R(A), y ∈ Y with
y − y ε < ε and x(0) = 0. Then the Landweber iteration together with the discrep-
ancy principle (4.262) is an order optimal regularization of A† , i.e., the Landweber
iteration possesses infinite qualification. The stopping index can be estimated as
follows
−2
m∗ = m∗ (ε, y ε ) ≤ Cμ ε μ+1 (4.264)
with Cμ > 0.
If we choose as an initial value 0 = x(0) with x(0) ∈ N (A)⊥ , it suffices to
study the convergence of the sequence {x̃(m),ε } with x̃(0),ε = 0 which results from
the Landweber method applied to the equation Ax = y ε − Ax(0) . The minimum
norm solution of this equation is A† y − x(0) and if A† y ∈ Xμ , we also need to have
that x(0) ∈ Xμ to obtain the optimal order of decay for the error, i.e., O(ε μ/(μ+1) ).
Unfortunately, for unknown μ we have to choose 0 as starting value.
The disadvantage of the Landweber iteration is that its convergence is rather
slow, i.e., the stopping index m∗ is often large. This is the reason why semi-iterative
methods (see, e.g., [227] and the references therein) come into play to accelerate
the convergence.
Semi-iterative methods. The characteristics of Landweber iteration are as follows;
x(0) = 0 (4.265)
and
x(m) = Fm (A∗ A) A∗ y (4.266)
with Fm (λ) given by (4.253), i.e.,

m−1
1 − (1 − ωλ)m
Fm (λ) = ω (1 − ωλ)j = . (4.267)
j=0
λ
For λ ∈ [0, A 2 ] we have

1
Fm (λ) → , m → ∞. (4.268)
λ
Moreover, we are able to show that
sup λ|Fm (λ)| = sup |1 − (1 − ωλ)m | ≤ 2, (4.269)
λ∈[0,A2 ] λ∈[0,A2 ]
so that {Fm }m∈N is a regularizing filter. The residual polynomials corresponding

to the filter polynomials are of degree m:
m
pm (λ) = 1 − λFm (λ) = (1 − ωλ) . (4.270)
All in all, the Landweber iteration procedure is as follows:
y − Ax(m) = pm (AA∗ ) y. (4.271)
In order to accelerate the Landweber iteration we are led to the idea (see, e.g., the
monograph [227]), to replace the polynomial filter by another filter family, that
shows a faster convergence to 1/λ.
To this end we consider an alternative polynomial Fm of degree m − 1, so
that its residual polynomial λ → pm (λ) = 1 − λFm (λ) is of degree m. Letting
x(m) := Fm (A∗ A) A∗ y we obtain as residuum y − Ax(m) = pm (AA∗ ) y. Now,
if {Fm }m∈N is a regularizing filter, the corresponding family {pm }m∈N has the
If {Fm } is a regularizing filter, the residual polynomials {pm } satisfy the
(i) lim pm (λ) = 0 for λ ∈ [0, A 2 ] .
m→∞
0 21
(ii) {pm } is uniformly bounded on the interval 0, A (by 1 + CF ).
(iii) pm (0) = 1 for all m.
Conversely, if {pm } is a family of polynomials satisfying the properties (i),
(ii) (iii), then
1 − pm (λ)
Fm (λ) = , m ∈ N0 (4.272)
λ
constitutes a regularizing filter.
From the theory on special functions of mathematical physics” (see, e.g., [73])
we borrow the following result:
Let the family {pm } fulfill the following conditions:
0 1
1. pm is a polynomial of degree m on the interval 0, A 2 .
2
A
0 21
2. pm (λ) pn (λ) w (λ) dλ = 0 if n = m, where w (λ) > 0 for λ ∈ 0, A
0
and w is piecewise continuous.
3. pm (0) = 1, m ∈ N0 .
Then there exist Am , Bm ∈ R satisfying
pm (λ) = pm−1 (λ) + Am (pm−1 (λ) − pm−2 (λ)) − Bm λpm−1 (λ) . (4.273)
0 21
The polynomials {pm } are orthogonal polynomials on the interval 0, A
with
0 respect to the weight function w (λ). Note that w (λ) induces a measure on
21
0, A . In transition to iteration we are led to
x(0) = 0, (4.274)
1 − p1 (λ)
x(1) = F1 (A∗ A) A∗ y with F1 (λ) = , (4.275)
λ

x(m) = x(m−1) + Am x(m−1) − x(m−2) + Bm A∗ y − Ax(m−1) for m ≥ 2.
(4.276)
This is the reason why methods of type (4.274), (4.275), (4.276) are called semi-
iterative (note that, for each iteration step, two previous iteration values are re-
quired).
For simplicity, assume now that A ≤ 1 which can be achieved by scaling
the operator. As a consequence, we only need to consider polynomials on the
interval [0, 1]. Then the following results are known for semi-iterative methods
(see, e.g., [227]):
(1) Let {pm } be a sequence of residual polynomials, so that it is uniformly
bounded on [0, 1], normalized by pm (0) = 1, and it converges pointwise to 0 on
(0, 1]. If {Fm } is the corresponding sequence of filters, then for A ∈ L(X, Y ) and
x(0) ∈ X it follows that

lim x(m) = lim x(0) + Fm (A∗ A)A∗ (y − Ax(0) )
m→∞ m→∞
†
A y + PN (A) x(0) : y ∈ D(A† ),
=
∞ : y∈ / D(A† ).
(2) Suppose that A is of class L(X, Y ). Let {x(m) }, {x(m),ε } be the iterates
of a semi-iterative method with respect to y ∈ R(A) and y ε ∈ Y using the same
initial value. Let the residual polynomials of the semi-iterative method be uniformly
bounded by Cp > 0. Then, for the data error, we have
- (m) -
-x − x(m),ε - ≤ 2Cp mε. (4.277)
Together with a stopping rule that fulfills
m∗ (ε) → ∞ , εm∗ (ε) → 0 for ε → 0 (4.278)
†
the semi-iterative method is a regularization of A .
In order to determine the speed of convergence we investigate

ωμ (m) = sup λμ/2 |pm (λ)| (4.279)
0≤λ≤1
(3) Each sequence of polynomials {pm }m∈N0 , pm (0) = 1, which satisfies the
best asymptotic behavior
ωμ (m) = O(m−μ ) for m → ∞ (4.280)
for some μ > 0, is uniformly bounded on [0, 1] and converges pointwise to 0 on
(0, 1]. In other words {pm }m∈N0 given in such a way is a sequence of residual
polynomials, for which, in addition, ωα (m) = O(m−α ) for 0 < α ≤ μ.
The discrepancy principle (cf. (4.262)) can also be used as stopping rule.
(4) Let A ∈ L(X, Y ) and y ∈ R(A). Let the normalized polynomials
{pm }m∈N0 , pm (0) = 1, satisfy (4.280) for some μ > 1. Then the corresponding
semi-iterative method with starting value x(0) = 0 is an order optimal regulariza-
tion of A† with respect to Xα for 0 < α ≤ μ−1 if it is combined with the discrepancy
principle (4.262) as stopping rule with τ > sup{ pm C[0,1] |m ∈ N0 } ≥ 1.
The stopping index satisfies

m∗ = m∗ (ε, y ε ) = O ε−1/(α+1) (4.281)
for ε → 0.
Normalized polynomials that fulfill (4.280) automatically lead to semi-iter-
ative order optimal regularization methods. The reason for this is that (4.280)
implies the uniform boundedness of the polynomials on [0, 1] as well as
sup{|Fm (λ)| : λ ∈ [0, 1]} ≤ 2τ m2
for the corresponding filters.
In the general case that pt (λ) = 1 − λFt (λ) we are confronted with the
situation (see, e.g., [227]) that
ωμ (t) ≤ Cp t μ/2 for t → 0 (4.282)
which is the analogue of (4.280). However, this neither implies the uniform bound-
edness of {pt }t>0 nor an estimate like |Ft (λ)| ≤ CF t−α .
Gradient method. The method of successive approximation gradient, and related
iterative methods can be used for finding approximate solutions of ill-posed prob-
lems (see, e.g., [108, 133, 134, 172, 179] and the references therein for more details).
We let X and Y be two Hilbert spaces, both over K, and let A be a bounded
linear operator on X into Y . As already known, the linear equation
Ax = y, y∈Y (4.283)
may or may not have a solution depending on whether or not y is in R(A), the
range of A, and even if y ∈ R(A) the solution of (4.283) need not be unique. For
any bounded linear operator A : X → Y, R(A) and R(A∗ ) are closed subspaces of
X and Y , respectively, hence, X = R(A) ⊕ R(A)⊥ and Y = R(A∗ ) ⊕ R(A∗ )⊥ . The
relations R(A) = R(A∗ )⊥ , R(A∗ ) = R(A)⊥ , R(A)⊥ = R(A∗ ), R(A) = R(AA∗ )

are also valid (see, for instance, [265]). Let P denote the orthogonal projection of
X onto R(A∗ ) and let Q denote the orthogonal projection of A on R(A). Then
Ax = AP x for all x ∈ X, A∗ y = A∗ Qy for each y ∈ Y , and the restriction of A to
R(A)⊥ has an inverse, which is not necessary.
In either case, i.e., y ∈ R(A) unique or non-unique solution of (4.283), one can
seek a best approximate solution, i.e., a solution which minimizes the quadratic
function J(x) = Ax − y 2 . Such a solution always exist for all y ∈ Y if R(A)
is closed. If R(A) is arbitrary, a best approximation does not exist for all y ∈ Y ,
however, it does exist for all y ∈ R(A)⊥ .
We consider the conjugate gradient (CG) method (cf. [133, 134]) that min-
imizes J(x) = Ax − y 2 at each step. That is, choose an initial vector x0 ∈ X,
then compute r0 = p0 = A∗ (Ax0 − y), where A∗ is the adjoint of A. If p0 = 0,
compute x1 = x0 − α0 p0 , where α0 = r0 2 / Ap0 2 . For i = 1, 2, . . ., compute
ri = A∗ (Axi − y) = ri−1 − αi−1 A∗ Api−1 , (4.284)
where
ri−1 , pi−1
αi−1 = , (4.285)
Api−1 2
and if ri = 0, then compute
pi = ri + βi−1 pi−1 , (4.286)
where
ri , A∗ Api−1
βi−1 = − . (4.287)
Api−1 2
Set
xi+1 = xi − αi pi . (4.288)
We examine some properties of the CG algorithm, for the case of an arbitrary
bounded linear operator. To be more concrete, the domain of the generalized in-
verse of A is D(A† ) = R(A) + R(A)† . If y ∈ D(A† ), then Qy = ȳ is in the range of
T and v = A† y = A† ȳ and y = Qy = Av = AA† y. Since Q is an orthogonal projec-
tion, the functional J can be written as J(x) = Ax − y 2 = Ax − ȳ 2 + ȳ − y 2 .
Thus, minimizing J is equivalent to minimizing the functional Ax − ȳ 2 which
we denote by g(x). Setting
u = v + (I − P )x0 = A† y + (I − P )x0 (4.289)
one can define the error vector e = x − u and the vector r = A∗ (Ax − y) =
A∗ (Ax − ȳ). Then
(A∗ A)e = r (4.290)
and
r, e = Ax − ȳ 2 = g(x). (4.291)
The sequence of iterates {xi } generated by the CG method (4.284)–(4.288) is
contained in the flat x0 + R(A∗ ) with both ri and pi , for i = 0, 1, 2, . . ., in R(A∗ ).
Moreover, p0 , p1 , . . . , pi−1 form an A∗ A-orthogonal set of vectors and their span is

an i-dimensional subspace of R(A∗ ). If at the ith step, ri = 0, then both Axi − y
and Axi − ȳ are vectors in R(A∗ ) = R(A)⊥ . However, Axi and ȳ are also in
R(A), and therefore txi = ȳ, implying that g(xi ) = 0. In this case, the iteration
terminates at the ith step and we have

i−1
xi = x0 − αk pk = u, (4.292)
k=0
as well as

i−1
A† y = P x0 − αi pk . (4.293)
k=0
Therefore, unless explicitly mentioned otherwise, we shall assume that the

CG method does not terminate in a finite number of steps, that is ri = 0 for
i = 0, 1, . . . . We list some known identities for the CG method.
For indices satisfying k = 0, 1, 2, . . ., i and i = 0, 1, 2, . . ., we have
ri , rk = pi , rk , (4.294)
Api Ari , (4.295)
2
ri+1
βi = , (4.296)
ri 2
ri 2 pk 2
pi , pk = , (4.297)
rk 2

i
rj
pi = ri 2 (4.298)
j=0
rj 2
and
ri 2 pi . (4.299)
Setting g(xi ) = ri , ei = Axi − ȳ 2 , where ei = xi − u, one finds that
g(xi ) − g(xi+1 ) = αi ri 2 . (4.300)
Finally, xi minimizes the functionals J(x) and y(x) on the i-dimensional flat
x0 + span{p0 , p1 , . . . , pi−1 }. (4.301)
It is worth mentioning the following three properties (see [134]):
(i) For k = 0, 1, 2, . . . , i
g(xi ) = ei , rk = ri , ek . (4.302)
For i = 0, 1, 2, . . .,
pi , ei ri 2 = g(xi ) pi 2 . (4.303)
(ii) The inequality

ei+1 2 ei 2 − αi g(xi ) (4.304)
holds for i = 0, 1, 2, . . ..
(iii) For any non-negative integers i and j, both pi , ei and ei , ej are non-
negative numbers.
The main result on the CG method for bounded linear operators with closed
range depends heavily upon the following observation:
Let X and Y be two Hilbert spaces over the same field and let A be a bounded
linear transformation mapping X into Y . If R(A) is closed, then S = A|R(A∗ ),
the restriction of A to R(A∗ ), has a bounded inverse and U = A∗ A|R(A∗ ) is a
positive definite operator onto R(A∗ ).
Let μ(x) := U x, x/x, x, x = 0, be the Rayleigh quotient of U . Since U is a
bounded symmetric positive definite linear operator on the Hilbert space R(A∗ ),
the spectral bounds
m = inf{μ(x) : x ∈ R(A∗ )} (4.305)
and
M = sup{μ(x) : x ∈ R(A∗ )} (4.306)
are positive and finite.
Let X and Y be two Hilbert spaces over the real field and let A be a bounded
linear transformation mapping X into Y . If the range of A is closed then the
conjugate gradient method (4.284)–(4.288) converges monotonously to the least
squares solution u = A† y + (I − P )x0 of Ax = y. Moreover, if m and M are the
spectral bounds of U = {A∗ A|R(A∗ )}, then
2i
g(x0 ) M − m
xi − u 2 , i = 0, 1, 2, . . . . (4.307)
m M +m
Altogether, let X and Y be two Hilbert space over the same field. If A is a
linear transformation mapping X into Y of rank r, then the conjugate gradient
method associated with the system Ax = y converges in at most r steps to the
least squares solution u = A† y + (I − P )x0 .
When R(A∗ ) is not closed, then the operator U = A|R(A∗ ) need not be
positive definite, and therefore the generalization of an inequality cannot longer
be utilized. However, it is still possible to establish convergence of the CG method
under mild restrictions.
Let X and Y be two Hilbert spaces over the real field, and let A be a bounded
linear operator mapping X to Y . If Qy ∈ R(AA∗ A), then the conjugate gradient
method (4.284)–(4.288), with initial value x0 ∈ R(A∗ A), converges monotonously
to the least squares solution of minimal norm u = A† y. In fact,
A 2 x0 − A† y 2 A∗† x0 − (AA∗ )† 2
xi − u 2 , i = 1, 2, . . . .
A 2 A∗† x0 − (AA∗ )† y 2 + i x0 − (AA∗ )† y 2
(4.308)
4.5. Stochastic regularization methods

Let X be a Hilbert space. Our statistical approach starts from an equation of type
(see, e.g., [101, 148])
Ax = y + η (4.309)
where x, y, η are considered as values of jointly distributed random variables.
Random variables. On the probability space Ω equipped with the probability mea-
sure P we understand ξ : Ω → X as Hilbert space-valued random variable. If X is
a function space, then we denote by ξ a stochastic process. For x ∈ X we obtain by
ξx = x, ξ a real-valued random variable. Hence, for a complete orthogonal system
{xi } in X, we are able to introduce by ξ (i) = ξxi = xi , ξ an infinite number of
jointly distributed random variables.
In what follows we suppose that ξ has a vanishing expectation value, i.e., we
have
E[x, ξ] = 0 (4.310)
for all x ∈ X (note that (4.310) does not mean any restriction, if (4.310) is violated
we are allowed to go over to the random variable ξ˜ = ξ − E[ξ]). Furthermore, we
assume that the random variable has a finite second moment, so that
(i) E[|x, ξ|2 ] < ∞ for all x ∈ X,
(ii) the expectation value is continuous at x.
Then it follows that E[x, ξξ, z] is a continuous, symmetric, non-negative bilin-
ear form on X, hence, there exists a linear, continuous, selfadjoint, non-negative
operator Rξξ : X → X satisfying
Rξξ x, z = Cov(x, z) = E[x, ξξ, z], (4.311)
Rξξ is called covariance operator.
Example. Let G be a regular region. Assume that X = L2 (G) and ξ(α) is defined
for α ∈ G. Then we are able to identify the covariance operator with the covariance
function resulting in the autocovariance function given by
Rξξ (α, β) = Cov(ξ(α), ξ(β)) = E[ξ(α) · ξ(β)]. (4.312)
Application to an L2 (G)-function x yields the identity

Rξ,ξ x(α) = Rξξ (α, β)x(β) dβ. (4.313)
G
As white noise we denote the Gaussian process characterized by
Rξξ (α, β) = η 2 δ(α − β), (4.314)
i.e.,
Rξξ = η 2 I. (4.315)
Let ξ, η be jointly distributed random variables with ξ : Ω → X, ζ : Ω → Y .
Then we define the cross-covariance operator by Rξζ : Y → X as follows:
Rξζ y, x = E [y, ζY x, ξX ] . (4.316)
Best linear estimator and Tikhonov–Phillips regularization. Next we have a look

at the equation
Aξ = β + ζ, ξ : Ω → X, β, ζ : Ω → Y. (4.317)
We assume that the inverse A−1 exists. The problem is to estimate ξ under the
knowledge of β and ζ. To this end, we suppose that E[ξ] = E[β] = 0 such that
−1
Rξζ = 0 (i.e., ξ and ζ are uncorrelated) and Rζζ exists. It is not difficult to show
that (see, e.g., [101, 148])
Rββ = ARξξ A∗ + Rζζ , (4.318)
Rξβ = Rξξ A∗ . (4.319)
In order to realize a least squares estimation we have to consider a linear estimator
of ξ, i.e., a random variable ξL = Lβ, where L : Y → X is a linear and continuous
operator such that xL = Ly is a solution of the operator equation. Central in our
considerations is the following statement, that is standard in statistical geodesy
(see, e.g., [101] and the references therein):
If Rζζ is assumed to have a continuous inverse, then
−1
L = Rξβ Rββ = Rξξ A∗ (ARξξ A∗ + Rζζ )−1 (4.320)
0 1
minimizes the functional E |x, ξ − Lβ|2 for all x.
Under the simplifying assumptions
Rξξ = I, Rζζ = η 2 I
we obtain for (4.320)
L = A∗ (AA∗ + η 2 I)−1 = (A∗ A + η 2 I)−1 A∗ . (4.321)
In other words, in similarity to the Tikhonov–Philipps regularization, we are led
to normal equations in order to determine the best linear estimator.
For arbitrary covariance operators we obtain x = Ly by minimizing
−1 −1
Rζζ (Ax − y), Ax − y + Rξξ x, x = |||Ax − y|||2 + ||x|| (4.322)
with
−1
|||y||| = Rζζ y, y (4.323)
and
−1
x = Rξξ x, x . (4.324)
Observing these facts we are finally able to come to the following conclusion:
The best linear estimator is a special Tikhonov–Phillips regularization method,
or, the Tikhonov–Phillips regularization method is a special linear estimator.
The stochastic approach also allows an interpretation of Bayes estimation as
Tikhonov–Phillips regularization (for more details the reader is referred, e.g., to
[101, 148] and the references therein).
4.6. Mollifier methods

The original idea of the mollifier method can be characterized as follows: We are
interested in the solution x† of Ax = y, but we realize that the problem is “too
ill-posed” for being able to determine x† accurately. Thus, we compromise by
changing the problem into a more well-posed one, namely that of trying to deter-
mine a mollified version Eρ x† of the solution, where Eρ is a suitable “mollification
operator” depending on a parameter ρ. The heuristic motivation is that the trou-
ble usually comes from high frequency components of the data and of the solution,
which are damped out by mollification (which, in fact, defines mollification).
In abstract nomenclature, early mollifier methods for ill-posed problems were
studied in [148]. Our approach is based on [63]. First we recapitulate the essential
ingredients. Again we start from the operator equation
Ax = y, x ∈ X, y ∈ Y (4.325)
with X, Y Hilbert spaces and R(A) non-closed. Our aim is to introduce operators
Eρ : X → X such that
Eρ x → x, x ∈ X, ρ → 0. (4.326)
IF X is a suitable function space, we are able to represent Eρ by a mollifier eρ via
the equation
Eρ (x)(s) = eρ (s, ·), xX . (4.327)
Instead of x† we now look for Eρ x† for some ρ > 0, thereby assuming that eρ has
a representation
A∗ vsρ = eρ (s, ·) (4.328)
with vsρ ∈ Y . Then, if Ax† = y, we can compute Eρ x† as follows:
(Eρ x† )(x) = eρ (s, ·), x† X = A∗ vsρ , x† X = vsρ , Ax† Y = vsρ , yY , (4.329)
i.e.,
(Eρ x† )(s) = vsρ , yY . (4.330)
Hence, the problem of solving (4.325) reduces to that of solving (4.328), which is
also ill posed as soon as (4.325) is. However, the right-hand side of (4.328) (which
is actually a family of equations depending on the parameter s) is usually given
analytically, since the mollifier eρ is chosen. Hence, there is no (or much less) error
in the data of (4.328), and these equations can be solved (by regularization) much
better than (4.325). As soon as an approximation for vsρ has been computed, it
can be used to solve (4.325) for any right-hand side y via (4.330). If we define the
operator Sρ : Y → X via the estimate
(Sρ y)(s) = vsρ , y, (4.331)
then, by (4.330), this operator maps the right-hand side of (4.325) to mollified
solutions. This motivates the term approximate inverse of A used for Sρ , also for
the more general case that (4.328) is not solvable. In this case, (4.328) is replaced by
A∗ vsρ − eρ (s, ·) X → min, (4.332)
which, is equivalent to
AA∗ vsρ = Aeρ (s, ·). (4.333)
Note that one needs the requirement that (4.333) is solvable. The function vsρ
is called reconstruction kernel ; uniqueness can be enforced by solving (4.333) in
the best-approximate sense, i.e., by selecting the solution of (4.333) with minimal
norm: vsρ = (A∗ )† eρ (s, ·). Assume, for simplicity, that R(A) is dense in Y , so that
(AA∗ )−1 exist. Then we have with vsγ defined by (4.332):
Sρ y = (AA∗ )−1 Aeρ (s, ·), yY = eρ (s, ·), A∗ (AA∗ )−1 yX
= eρ (s, ·), (A∗ A)† A∗ yX = (Eρ x† )(s), (4.334)
i.e., Sρ y is the mollified version of the best-approximate solution of (4.330). This
justifies (4.332).
Let A be compact with singular system (σn ; xn , yn ). Let a regularization
method realized in standard way, i.e.,
∞

xα = σn Fα (σn2 )y, yn Y yn . (4.335)
n=1
If we assume that X and Y are suitable function spaces, then (4.335) can be
written as
xα (s) = vsρ , yY (4.336)
with
∞

vsρ (t) = σn Fα (σn2 )yn (t)yn (s). (4.337)
n=1
Now, vsρ can be written in the form (4.333) with
∞

eρ (s, t) = σn2 Fα (σn2 )yn (s)yn (t). (4.338)
n=1
Hence, xα can be considered as a mollified solution Eρ x† with Eρ given (in the

sense of (4.326)) by the mollifier (4.338), so that linear regularization methods can
also be viewed as mollifier methods.
The underlying “suitable function space” have to be such that point evalu-
ation is continuous wherever used. The choice of the mollifier, of course, depends
on what one wants to achieve; frequently used choices are singular integral kernels
(such as Haar kernel, sinc kernel, etc.). Note again that each of these mollifiers
can be applied only to equations, so that (4.333) admits a solution.
4.7. Backus–Gilbert method

The Backus–Gilbert method (cf. [18]) treats moment problems of the type
x, ki X = βi , i ∈ {1, . . . , n}, (4.339)
with given elements ki ∈ X, for example X = L2 (G), G regular region, so that

(4.339) takes the form

ki (t)x(t) dt = βi , i ∈ {1, . . . , n}, (4.340)
G
which can be thought of as resulting from discretizing an integral equation of the
first kind
k(s, t)x(t) dt = β(s) (4.341)
G
by collocating at points s1 , . . . , sn , so that ki (t) = k(si , t), βi = β(si ).
With Ax = (x, k1 X , . . . , x, kn X )T , β = (β1 , . . . , βn ), (4.339) can be writ-
ten in the form Ax = β with X = L2 (G), Y = Rn . In the Backus–Gilbert method,
one looks for an approximate inverse S : Rn → L2 (G) for A by defining
n
Sy = yi vi , (4.342)
i=1
with functions vi ∈ L2 (G) to be determined as follows: since

= n >
n
SAx = x, ki X vi = x, ki vi , (4.343)
i=1 i=1
i.e., for the concrete case (4.340)
D E

n
(SAx)(s) = x(t) ki (t)vi (s) dt, (4.344)
G i=1
one should aim at determining the functions vi such that
n
ki (t)vi (s) ∼ δ(|s − t|). (4.345)
i=1
The question is how to formalize the requirement (4.345). In the classical approach
to the Backus–Gilbert method [18], this is done by minimizing, for any fixed s ∈ G
and some chosen τ , the functional
n 2

2τ
(v1 , . . . , vn ) → |s − t| ki (t)vi dt. (4.346)
G
k=1
Under the normalization constraint
n
ki (t)vi dt = 1, (4.347)
G i=1
we then take vi (s) := vi . The constraint (4.347) just says that for x ≡ 1, (SAx)(s) =
x(s) holds. The parameter τ (in [18] τ = 1) determines the concrete method.
The common feature between mollification and the Backus–Gilbert method
is the following: In both cases, an approximate inverse (determined by vsρ or by
the vi (s)) is determined independently from the right-hand side of the equation,
which can then be used to explicitly represent an approximate solution via (4.336)
or via (4.342). By use of Lagrange multipliers, the Backus–Gilbert basis functions
v1 , . . . , vn can be determined pointwise from the linear system

G(s) w v 0
= , s ∈ Ω, (4.348)
T
w 0 λ 1
with

G(s)ij = |s − t|2τ ki (t)kj (t) dt, i, j ∈ {1, . . . , n}, (4.349)
G

wi = ki (t) dt, i ∈ {1, . . . , n}. (4.350)
G
Note that the matrix of this system depends on s while in the corresponding system
(4.333) for mollifier methods, s enters only in the right-hand side.
4.8. Numerical dilemmas and methodologies

The numerical analysis of all ill-posed problem ultimately involves solutions of
finite-dimensional problems in order to obtain numerical approximations. This
often entails a two-stage regularization. One first may “regularize” the problem in
function spaces and then apply numerical methods to approximate the solution
of a well-posed problem (or a family of such problems). On the other hand, one
may “discretize” or approximate the IPP by finite-dimensional problems and then
resolve numerical instability of these problems by methods of numerical linear
algebra that are suitable for discrete ill-posed problems (see, e.g., [177, 184, 253]).
Two alternate routes are represented by the following diagram:
c r
P → Pt
p p p
d r
Fn → Pn → Pn,t Pt,n
Here, P is a given ill-posed problem and Pt is a “regularized” version of
P by use of a certain regularization scheme in a certain function space. Fn is
a problem in a finite-dimensional function space and Pn is a finite-dimensional
algebraic problem, both obtained by approximation of P ; whereas Pn,t denotes a
“regularization” of the (numerically unstable) problem Pn and Pt,n is the numerical
approximation of the problem Pt , which is numerically stable for t not too small. In
the diagram “c r” denotes continuous regularization, “d r” discrete regularization
and “p” denotes a generic “projection” (discretization, projection method, etc.).
At this stage, after having some knowledge of most of the regularization
methods, we mention some procedures which provide concrete realizations of these
schemes (cf. [184]):
P → Pt Tikhonov’s regularization, (multiscale) regularizer operators in func-
tion spaces, quasi-reversibility methods, replacement of the IPP by
a stable minimization problem depending on a parameter, iterative

methods in function spaces;
P → Fn truncated singular value decompositions (TSVD), truncated series
expansion, moment discretization, projection methods;
P → Pn finite difference method with collocation, reduction of Fn to algebraic
equations;
Pn → Pn,t decomposition methods or regularization for linear algebraic equa-
tions, TSVD for matrices;
Pγ → Pt,n various numerical methods for solving well-posed problems, e.g., dis-
cretization, projection methods, multiscale procedures, etc.
Returning to the general scheme, we let x, xt , xn , xt,n , xn,t denote, respec-
tively, the “solution” (classical or least square of minimal norm) of the problems
P, Pt , Pn , Pt,n , Pn,t in the absence of contamination in the data, and let xε , xεt , etc.,
denote the corresponding solution when the data are contaminated (y is replaced
by y ε , where y ε −y ≤ ε for some ε > 0). If we assume that the various regularizer
schemes are convergent (e.g., xt → x as t → 0, xn,t → xn as t → 0 for each fixed
n, xn → x as n → ∞, etc.) in the absence of error, then estimates similar to (3.3)
and (3.4) can be used to calculate an “approximate solution” xε as before. For
example,
xεt,n − x ≤ xεt,n − xεt + xεt − xt + xt − x . (4.351)
Here, xεt,n − xεt X is an approximation error estimate that provides a rate of
convergence of the approximation scheme for the well-posed problem Pt for a fixed
t. xεt − xt is an estimate for the contamination error which can be estimated if
the robustness of Pt is known and xt − x is a regularization error. Similarly,
xεn,t − x ≤ xεn,t − xtn + xεn − xn + xn − x . (4.352)
Note, however, that now xεn,t −xtn
is an error in the regularization of the problem
Pn , xεn − xn is an error due to the propagation of contamination into the discrete
system, etc.
The dilemmas and methodologies of mathematical and numerical analysis of
IPP involve the following facts and observations:
(i) For most regularization and approximation schemes
lim xt = x (4.353)
t→0
and
lim xn = x (4.354)
n→∞
in the absence of contamination. For some approximation (projection)

schemes, non-convergence can take place even without contamination.
(ii) Neither the double limit
lim xεt,n , (4.355)
t→0,
n→∞
nor the iterated limits [limn→∞ limt→0 and limt→0 limn→∞ ] of xεt,n and xεn,t
exist. In fact, xεt,n X and xεn,t X blow up as n → ∞ and t → 0. It should be
emphasized that this blow-up is intrinsically inherent in all IPP regardless of
any regularization-approximation scheme. Thus, the best one can achieve for
a numerical resolution of IPP is to minimize the error: x−xεt,n or x−xεn,t ,
and to find “paths” along which xt,n and xn,t converge to x as ε → 0.
(iii) The alternative routes diagram is non-commutative, in general. It is not al-
ways clear which path along the diagram is more effective. One has to com-
pare the minimum errors of x − xεt,n X and x − xεn,t X and to take into
consideration the computational complexity of the two paths. For some sim-
ple schemes, the diagram is commutative (e.g., Tikhonov regularization and
TSVD for a compact operator commute).
(iv) Sharp resolutions of an ill-posed problem (i.e., an optimal compromise be-
tween accuracy and numerical stability) hung upon “optimal parameter choice
criteria” (for t and n). Analytic criteria for this choice are often not avail-
able due to the lack of sharp rates of convergence in the preceding estimates,
except for Tikhonov regularization or related methods based on simple vari-
ational principles. Often the parameter is chosen by an interactive computa-
tional scheme, based on rough analytic estimates.
References
[1] Acar, R., Vogel, C.R.: Analysis of Bounded Variation Penalty Methods for Ill-Posed
Problems. Inverse Problems, 10:1217–1229, 1994.
[2] Adams, R.A.: Sobolev Spaces. Academic Press, New York, San Francisco, 1975.
[3] Alber, Y.I.: The Solution of Nonlinear Equations with Monotone Operators in Ba-
nach Spaces, Siberian Math. J., 16:1–8, 1975.
[4] Alber, Y.I.: Iterative Regularization in Banach Spaces, Soviet Math. (Iz. VUZ),
30:1–8, 1986
[5] Alber, Y.I.: The Regularization Method for Variational Inequalities with Non-
smooth Unbounded Operators in Banach Spaces. Appl. Math. Lett., 6:63–68, 1993.
[6] Alber, Y.I.: Generalized Projection Operators in Banach Spaces: Properties and
Applications, Functional Differential Equations, Proceedings of the Israel Seminar,
1:1–21, 1994.
[7] Alber, Y.I.: Metric and Generalized Projection Operators in Banach Spaces: Prop-
erties and Applications, In: Theory and Applications of Nonlinear Operators of
Monotone and Assertive Type (A. Kartsatos, ed.). Marcel Dekker, New York, 15–
50, 1996.
[8] Alber, Y.I., Notik, A.: Perturbed Unstable Variational Inequalities with Unbounded
Operator on Approximately Given Sets. Set-Valued Analysis, Vol. 1, 4:393–402,
1993.
[9] Alber, Y.I., Reich, S.: An Iterative Method for Solving a Class of Nonlinear Operator
Equations in Banach Spaces, Panamerican Math. J., 4:39–54, 1994.
[10] Alber, Y.I., Rjazanceva, I.: Variational Inequalities with Discontinuous Monotone
Mappings, Soviet Math. Dokl., 25: 206–210, 1982.
[11] Albert, A.: Regressions and the Moore–Penrose Pseudoinverse. Academic Press,
1972.
[12] Angell, T.S., Nashed, M.Z.: Operator-Theoretic and Computational Aspects of Ill-
Posed Problems in Antenna Theory. In: Proc. Int. Symp. Math. Theory of Networks
and Syst., Delft Univ. of Technology, The Netherlands, 499–511, 1979.
[13] Anger, G.: A Characterization of Inverse Gravimetric Source Problem through Ex-
tremal Measures. Rev. Geophys. Space Phys., 19:299–306, 1981.
[14] Anger, G.: Inverse Problems in Differential Equations. Akademie-Verlag, Berlin,
1990.
[15] Aronszajn, N.: Theory of Reproducing Kernels. Trans. Am. Math. Soc., 68:337–404,
1950.
[16] Baart, M.L.: Methods for Solution of Fredholm Integral Equations of the First Kind.
CSIR, Pretoria, Techn. Rep., 104, 1979.
[17] Baer, R.: Linear Algebra and Projective Geometry. Academic Press, New York,
1952.
[18] Backus, G.E., Gilbert, F.: Numerical Applications of a Formalism for Geophysical
Inverse Problems. Geophys. J.R. Astron. Soc., 13:247–276, 1967.
[19] Bakusinksii, A.B.: A general method for constructing regularizing algorithms for
a linear incorrect equation in Hilbert space. U.S.S.R. Comp. Math. Meth. Phys.
(7):279–284, 1967.
[20] Bakusinskii, A.B.: On the Principle of Iterative Regularization, USSR Comput.
Math. and Math. Phys., 19:256–260, 1979.
[21] Barzaghi, R., Sansò, F.: Remarks on the Inverse Gravimetric Problem. Boll. Geod.
Scienze Affini, 45:203–216, 1986.
[22] Baumeister, J.: Stable Solution of Inverse Problems. Vieweg, Braunschweig, 1987.
[23] Ben-Israel, A., Greville, T.N.E., Generalized Inverses: Theory and Applications.
New York: Wiley-Interscience, 1974.
[24] Bertero, M., de Mol, C., Viano, G.A.: Linear Inverse Problems With Discrete Data.
Inverse Problems, 4:573–594, 1985/88.
[25] Bertero, M., Brianzi, P., Pike, E.R., Rebolia, L.: Linear Regularizing Algorithms for
Positive Solutions of Linear Inverse Problems. Proc. R. Soc. Lond. A, 415:257–275,
1988.
[26] Bertero, M. DeMol, C., Viano, G.A.: The Stability of Inverse Problems. In: In-
verse Scattering Problems in Optics, H.P. Baltes, ed., Berlin, Heidelberg, New York,
Springer, 1979.
[27] Binder, A., Engl, H.W., Groetsch, C.W., Neubauer, A., Scherzer, O.: Weakly
Closed Nonlinear Operators and parameter identification in Parabolic Equations
by Tiknonov Regularization. Appl. Anal., 55:215–234, 1994.
[28] Bissantz, H., Hohage, T., Munk, A.: Ruymgaart, F.: Convergence Rates of General
Regularization Methods for Statistical Inverse Problems and Applications. SIAM
J. Numer. Anal., 45:2610–2626, 2007.
[29] Bjerhammar, A.: Rectangular Reciprocal Matrices, with Special Reference to Geo-
detic Calculations. Bulletin Géodésique, 25, 188–220, 1951.
[30] Bjerhammar, A.: Theory of Errors and Generalized Matrix Inverses. Elsevier Sci-
entific Publ. Co., Amsterdam, 1973.
[31] Boullion, T.L., Odell, P.L., Generalized Inverse Matrices. Wiley-Interscience, New
York, (1971).
[32] Bruck, R.E.: A Strongly Convergent Iterative Solution of 0 ∈ U x for a Maximal
Monotone Operator U in Hilbert Space, J. Math. Anal. Appl., 48:114–126, 1974.
[33] Burger, M., Kaltenbacher, B.: Regularizing Newton–Kaczmarz Methods for Non-
linear Ill-Posed Problems. SIAM J. Numer. Anal., 44:1775–1797, 2006.
[34] Cavalier, L., Golubev, G.K.: Risk Hull Method and Regularization by Projections
of Ill-posed Inverse Problems. Ann. Stat., 34:1653–1677, 2006.
[35] Cavalier, L., Golubev, G.K., Picard, D., Tsybakov, A.B.: Oracle Inequalities for
Inverse Problems. Ann. Stat., 30:843–874, 2002.
[36] Craven, B.D., Nashed, M.Z.: Generalized Implicit Function Theorems when the
Derivative has no Bounded Inverse: Theory, Methods, and Applications. Nonlinear
Analysis, 6:375–387, 1982.
[37] Davis, P.J.: Interpolation and Approximation. Blaisdell, New York, 1963.
[38] M. Defrise, C. de Mol.: A Note on Stopping Rules for Iterative Regulaization Meth-
ods and Filtered SVD. In: Inverse Problems: An Interdisciplinary Study. Academic
Press, San Diego, 261–268, 1987.
[39] Desbat, L., Girard, D.: The “Minimum Reconstruction Error” Choice of Regular-
ization Parameters: Some More Efficient Methods and Their Application of Decon-
volution Problems. SIAM, J. Sci. Compt., 16:187–1403, 1995.
[40] Diaz, J.B. and Metcalf, F.T.: On Integration Procedures for Equations of the First
Kind, Ax = y, and Picard’s Criterion for the Existence of a Solution. Math. Comp.
24:923–935, 1970.
[41] Dicken V. and Maass P.: Wavelet-Galerkin Methods for Ill-Posed Problems. J. Inv.
and Ill-posed Probl., 4:203–222, 1996.
[42] Dimitiev,V.I., Il’inski, A.S., Svenshnikov, A.G.: The Developments of Mathematical
Methods for the Study of Direct and Inverse Problems in Electrodynamics. Russian
Math. Surveys, 31:133–152, 1976.
[43] Dobson D.C., Scherzer, O.: Analysis of Regularized Total Variation Penalty Meth-
ods for Denoising. Inverse Problems, 12:601–617, 1996.
[44] Donoho D.L.: Nonlinear Solution of Linear Inverse Problems by Wavelet-Vaguelette
Decomposition. Appl. Comp. Harm. Anal., 2:101–126, 1995.
[45] Donoho D.L. and Johnstone, I.M.: Minimax Estimation Via Wavelet Shrinkage.
Ann. Statistics, 26:879–921, 1998.
[46] Dunford N., Schwarz, J.: Linear Operators, Vol. II, Wiley-Interscience, New York,
1963.
[47] Duris, C.S.: Optimal Quadrature Formulas Using Generalized Inverses. Part I. Gen-
eral Theory and Minimum Variance Formulas, Math. Comp. 25:495–504, 1971.
[48] Eggermont, P.P.B.: Maximum Entropy Regularization for Fredholm Integral Equa-
tions of the First Kind. SIAM J. Math. Anal., 24:1557–1576, 1993.
[49] Eggermont P.P.B., LaRicca, V.N.: Maximum Penalized Likelihood Estimation. Vol-
ume II, Regression, Springer, New York, 2009. 10
[50] Eggermont P.P.B., LaRicca, V.N. Nashed, M.Z.: On Weakly Bounded Noise in Ill-
Posed Problems. Inverse Problems. 25:115018–115032, 2009.
[51] Eggermont, P.P.B.; LaRiccia, V.N.; Nashed, M.Z.: Noise Models for Ill-Posed Prob-
lems. In: Handbook of Geomathematics, Vol. 2, (Freeden, W., Nashed, M.Z., Sonar,
T., Eds.), 741–762, Springer, Heidelberg, 2010.
[52] Eggermont P.P.B., LaRicca, V.N. Nashed, M.Z.: Moment Discretization of Ill-Posed
Problems with Discrete Weakly Bounded Noise. Int. J. Geomath., 3:155–178, 2012.
[53] Eggermont, P.N., LaRiccia, V., Nashed, MZ: Noise Models for Ill-Posed Problems.
In: Handbook of Geomathematics, Vol. 2, 2nd edition, (Freeden, W., Nashed, M.Z.,
Sonar, T., Eds.), 1633–1658, Springer, New York, 2015.
[54] Eicke, B.: Iteration Methods for Convexly Constrained Ill-posed Problems in Hilbert
Space, Numer. Funct. Anal. Optim., 13:413–429, 1992.
[55] Eisenhart, C., Carl Friedrich Gauss. International Encyclopedia of Social Sciences,
New York, Vol. VI, 74–81, 1986.
[56] Ekeland, I.; Temam, R. Convex Analysis and Variational Problems. North-Holland,
Amsterdam, 1976.
[57] Elden, L.: Algorithms for the Regularization of Ill-Conditioned Least Squares Prob-
lems. BIT, 17:134–145, 1977.
[58] Elden, L.: An Algorithm for the Regularization of Ill-Conditioned Banded Least
Squares Problems. SIAM J. Sci. Stat. Comput.,5:237–254,1984.
[59] Engl, H.: Discrepancy Principles for Tikhonov Regularization of Ill-Posed Problems,
Leading to Optimal Convergence Rates, J. Optim. Theory Appl., 52:209–215, 1987.
[60] Engl, H.: Integralgleichungen. Springer Lehrbuch Mathematik, Wien, 1997.
[61] Engl, H.W., Nashed, M.Z.: Stochastic Projectional Schemes for Random Linear Op-
erator Equations of the First and Second Kinds, Numer. Funct. Anal. and Optimiz.,
1:451–473, 1979.
[62] Engl, H.W., Nashed, M.Z.: New Extremal Characterizations of Generalized Inverses
of Linear Operators. J. Math. Anal. Appl., 82:566–586, 1981.
[63] Engl, H.W., Hanke, M., Neubauer, A.: Regularization of Inverse Problems. Kluwer,
Dordrecht, 1996.
[64] Engl H.W., Kunisch K. and Neubauer, A.: Convergence Rates for Tikhonov Regu-
larisation of Non-Linear Ill-Posed Problems. Inverse Problems, 5:523–540, 1989.
[65] Engl, H., Louis, A.K., Rundell, W. (Eds.): Inverse Problems in Geophysical Appli-
cations. SIAM, Philadelphia, 1997.
[66] Flemming, J.; Hofmann, B.: A New Approach to Source Conditions in Regular-
ization with General Residual Term. Numer. Funct. Anal. Optimiz., 31:254–284,
2010.
[67] Frankenberger H, Hanke M.: Kernel Polynomials for the Solution of Indefinite and
Ill-Posed Problems. Numer. Alg., 25:197–212, 2002.
[68] Franklin J.N.: Well-posed Stochastic Extensions to Ill-Posed Linear Problems. J.
Math. Anal. Appl., 31:682–716, 1970.
[69] Franklin, J.N.: On Thikhonov’s Method for Ill-Posed Problems. Math. Comp.,
28:889–907, 1974.
[70] Fredholm I.: Sur une classe d’équations functionelles. Acta Mathematica 27:365–
390, 1903.
1981.
[72] Freeden, W.: Multiscale Modelling of Spaceborne Geodata, B.G. Teubner, Stuttgart,
Leipzig, 1999.
[73] Freeden, W., Gutting, M.: Special Functions of Mathematical (Geo)Physics. Birk-
häuser, Basel, (2013).
[74] Freeden, W., Maier, T.: Spectral and Multiscale Signal-to-Noise Thresholding of
Spherical Vector Fields. Computational Geosciences, 7(3):215–250, 2003.
[75] Freeden, W., Michel, V.: Multiscale Potential Theory (with Applications to Geo-
science). Birkhäuser, Boston, 2004.
[76] Freeden, W., Nutz, H.: Satellite Gravity Gradiometry as Tensorial Inverse Problem.,
Int. J. Geomath., 2:177–218, 2012.
[77] Freeden, W., Schneider, F.: Regularization Wavelets and Multiresolution, Inverse
Problems, 14:493–515, 1998.
[78] Freeden, W., Schreiner, M.: Satellite Gravity Gradiometry (SGG): From Scalar to
Tensorial Solution. In: Handbook of Geomathematics, Vol. 1 (Freeden, W., Nashed,
M.Z., Sonar,T., Eds.), 269–302, Springer, Heidelberg, 2010.
Tensorial Solution. In: Handbook of Geomathematics, Freeden, W., Nashed, M.Z.,
and Sonar, T., Eds., Springer, 2nd edition, New York, Heidelberg, 2nd edition,
2015.
[80] Freeden, W., Witte, B.: A Combined (Spline-)Interpolation and Smoothing Method
for the Determination of the Gravitational Potential from Heterogeneous Data.
Bull. Geod., 56:53–62, 1982.
[81] Freeden, W., Michel, V., Nutz, H.: Satellite-to-Satellite Tracking and Satellite Grav-
ity Gradiometry (Advanced Techniques for High-Resolution Geopotential Field De-
termination). J. Eng. Math., 43:19–56, 2002.
[82] Freeden, W., Nashed, Z., Sonar, T. (Eds.): Handbook of Geomathematics. Springer,
New York, 2nd edition, Vol. 1, 2, and 3, 2015.
[83] Freeden, W., Schneider, F., Schreiner, M.: Gradiometry – An Inverse Problem in
Modern Satellite Geodesy, in: Engl, H.W., Louis, A., Rundell, W. (Eds.), GAMM-
SIAM Symposium on Inverse Problems: Geophysical Applications, 179–239, 1997.
[84] Friedrich, K.: Allgemeine für die Rechenpraxis geeignete Lösung für die Aufgaben
der kleinsten Absolutsumme und der günstigsten Gewichtsverteilung, Zeitschrift
für Vermessungswesen, 337–358, 1937.
[85] Gajewski, H., Gröger, K., Zacharias, K.: Nichtlineare Operatorgleichungen und Op-
eratordifferentialgleichungen. Akademie-Verlag, Berlin, 1974.
[86] Galerkin, B.G.: Expansions in Stability Problems for Elastic Rods and Plates (in
Russian). Vestnik Inzkenorov, 19:897–908, 1915.
[87] Gauss, C.F.: Theoria motus corporum coelestium in sectionibus conicis solem am-
bientium, Hamburg, 1809, Werke 7. Translated into English by C.H. Davis, 1963.
[88] Gauss, C.F.: Theoria combinationis observationum erroribus minimis obnoxiae, 1.
[89] Gebbauer, B., Scherzer, O.: Impedance-Acoustic Tomography, SIAM J. Appl.
Math., 69:565–576, 2008.
[90] Gelf́and I.M., Shilov, G.E.: Generalized Functions. Vol. 1 and 2. Academic Press,
New York, 1964.
[91] Gerhards, C.: Spherical Multiscale Methods in Terms of Locally Supported Wavelets:
Theory and Application to Geomagnetic Modeling. Ph.D.-thesis, Geomathematics
Group, University of Kaiserslautern, 2011.
[92] Gfrerer, H.: An a Posteriori Parameter Choice for Ordinary and Iterated Tikhonov
Regularization of Ill-Posed Problems Leading to Optimal Convergence Rates. Math.
Comp., 49:523–542, 1987a.
[93] Gfrerer, H.: Supplement to: An a Posteriori Parameter Choice for Ordinary and
Iterated Tikhonov Regularization of Ill-Posed Problems Leading to Optimal Con-
vergence Rates. Math. Comp., 49:S5–S12, 1987b.
[94] Glockner, O.: On Numerical Aspects of Gravitational Field Modelling from SST
and SGG by Harmonic Splines and Wavelets (with Application to CHAMP Data),
Ph.D.-thesis, Geomathematics Group, University of Kaiserslautern, 2001.
[95] Gholami, A.; Siahkoohi, H.R.: Regularization of Linear and Nonlinear Geophysical
Ill-Posed Problems with Joint Sparsity Constraints. Geophys. J. Int., 180:871–882,
2010.
[96] Gohberg, I., Goldberg, S., Kaashoek, M.A.: Classes of Linear Operators, Vol. II.
Birkhäuser OT 63, 1993.
[97] Goldstine, H.H.: A History of Numerical Analysis from the 16th Through the 19th
Century. Springer-Verlag, New York, 1977.
[98] Golomb M. and Weinberger, H.F.: Optimal Approximation and Error Bounds, In
One Numerical Approximation,Proc. Sympos. Madison, Wis., 1958, (R.E. Langer,
ed.), University of Wisconsin Press, 117–190, 1959.
[99] Golub, G.H., Van Loan, C.F.: Matrix Computations. 3rd ed., The John Hopkins
Univ. Press, Baltimore, M.D., 1996.
[100] Grafarend, E.W.: Six Lectures on Geodesy and Global Geodynamics. In: Pro-
ceedings of the Third International Summer School in the Mountains (Moritz, H.,
Sünkel, H. (Eds.)), 531–685, 1982.
[101] Grafarend, E.W., Awange, J.L.: Applications of Linear and Nonlinear Models.
Springer, Berlin, Heidelberg, 2012.
[102] Graves, J., Prenter, P.M.: On Generalized Iterative Filters for Ill-Posed Problems.
Numer. Math., 30:281–299, 1978.
[103] Grenander U.: Abstract Inference. Wiley, New York, 1981.
[104] Groetsch, C.W.: Generalized Inverses of Linear Operators. Marcel Dekker. Inc.,
New York, 1977.
[105] Groetsch, C.W.: The Theory of Tikhonov Regularization for Fredholm Equations of
the First Kind. Pitman, London-Boston, 1984.
schweig, 1993.
[107] Groetsch, C.W.: Comments on Morozov’s Discrepancy Principle. In Hämmerlin, G.;
Hoffmann, K.H., Eds., Improperly Posed Problems and Their Numerical Treatment,
97–104, Birkhäuser, Basel, 1983.
[108] Groetsch C.W., Scherzer, O.: Iterative Stabilization and Edge Detection. In: Con-
temp. Math., 313:129–141. Providence, RI: Amer. Math. Soc., 2002.
[109] Hadamard, J.: Sur les problèmes aux dérivés partielles et leur signification physique.
Princeton Univ. Bull., 13:49–52, 1902.
[110] Hadamard, J.: Lectures on the Cauchy Problem in Linear Partial Differential Equa-
tions, Yale University Press, New Haven, 1923.
[111] Haddad R.A. Parsons, T.W.: Digital Signal Processing: Theory, Applications and
Hardware. Computer Science Press, 1991.
[112] Hanke, M.: Conjugate Gradient Type Methods for Ill-Posed Problems. Pitman Re-
search Notes in Mathematics, Longman House, Harlow, Essex, 1995.
[113] Hanke, M., Hansen, P.C.: Regularization Methods for Large Scale Problems. Sur-
veys Math. Indust., 3:253–315, 1993.
[114] Hanke, M., Scherzer, O.: Inverse Problems Light: Numerical Differentiation. Amer.
Math. Monthly, 108:512–521, 2001.
[115] Hanke, M.; Vogel, C.R.: Two-Level Preconditioners for Regularized Inverse Prob-
lems. Numer. Math., 83:385–402, 1999.
[116] Hansen, P.C.: Analysis of Discrete Ill-Posed Problems by Means of the L-Curve.
SIAM Rev., 34:561–580, 1992.
[117] Hanson, R.J.: A Numerical Methods for Solving Fredholm Integral Equations of
the First Kind. SIAM J. Numer. Anal. 8:616–662, 1971.
[118] Hegland, M.: Variable Hilbert Scales and their Interpolation Inequalities with Ap-
plications to Tikhonov Regularization. Appl. Anal., 59:207–223, 1995
[119] Helmert, F.: Die Ausgleichungsrechnung nach der Methode der kleinsten Quadrate.
Teubner, Berlin, 1907.
[120] Heuser, H.: Funktionalanalysis. 4. Auflage, Teubner, 1975.
[121] Hilbert, D.: Grundzüge einer allgemeinen Theorie der linearen Integralgleichungen.
Teubner, Leipzig, 1912.
[122] Hille, E.: Introduction to the General Theory of Reproducing Kernels, Rocky Moun-
tain, J. Math., 2:321–368, 1972.
[123] Hirzebruch, Scharlau: Funktional Analysis I, BI 296a, 1971.
[124] Hofmann, B.: Inverse Probleme. Teubner 1999.
[125] Hofmann, B.; Mathé, P.; Weiszäcker, H. von: Regularisation in Hilbert Space under
Unbounded Operators and General Source Conditions. Inverse Problems, 25:115–
130, 2009.
[126] Hohage, T.; Pricop, M.: Nonlinear Tikhonov Regularization in Hilbert Scales for
Inverse Boundary Value Problems with Random Noise. Inverse Problems and Imag-
ing, 2:271–290, 2008.
[127] Hurwitz, W.A.: On the Pseudo-Resolvent to the Kernel of an Integral Equation.
Transactions of the American Mathematical Society, 13:405–418, 1912.
[128] Ismail, M., Nashed, M.Z., Zayed, A., Ghaleb, A.: Mathematical Analysis, Wavelets
and Signal Processing. Contemporary Mathematics, Vol. 190, American Mathemat-
ical Society. Providence, RI, 1995.
[129] Ivanov V.K., Kudrinskii, V.Yu.: Approximate Solution of Linear Operator Equa-
tions in Hilbert Space by the Method of Least Squares. I. Z. Vycisl. Mat. i Mat.
Fiz., 6:831–944, 1966.
[130] Jacobsen, M., Hansen, P.C.; Saunders, M.A.: Subspace preconditioned LSQR for
Discrete Ill-Posed Problems. BIT Numer. Math., 43:975–989, 2003.
[131] Jorgensen P., Tian, F.: Graph Laplacians and Discrete Reproducing Kernel Hilbert
Spaces from Restrictions. Stochastic Analysis and Applications, 34:722–747, 2016.
[132] Kaipio J., Somersalo, E.: Statistical and Computational Inverse Problems. Springer,
Berlin, 2005.
[133] Kammerer, W.J., Nashed, M.Z.: The Convergence of the Conjugate Gradient
Method for Singular Linear Operator Equations. SIAM J. Numer. Anal., 9:165–
181, 1972.
[134] W.J. Kammerer, M.Z. Nashed: Iterative Methods for Best Approximate Solutions
of Linear Integral Equation of the First and Second Kind. J. Math. Anal. Appl.,
40:547–573, 1972.
[135] Kantorowitsch, L.W., Akilow, G.P.: Funktionalanalysis in normierten Räumen.
Akademie-Verlag, Berlin, 1964.
[136] Kato, T.: Perturbation Theory for Nullity Deficiency and Other Quantities of Linear
Operators. J. Analyse Math., 6:271–322, 1958.
[137] Kirsch, A.: An Introduction to the Mathematical Theory of Inverse Problems. 2nd
ed., Springer, Heidelberg, 1996.
[138] Kowar, R., Scherzer, O.: Convergence analysis of a Landweber–Kaczmarz method
for solving nonlinear ill-posed problems, in: S. Romanov, S.I. Kabanikhin, Y.E.
Anikonov, A.L. Bukhgeim, (Eds.), Ill-Posed and Inverse Problems, VSP Publishers,
Zeist, 2002.
[139] Kress, R.: Linear Integral Equations. 2nd ed., Springer, Berlin, 1989.
[140] Kreyszig, E.: Introductory Functional Analysis with Applications. John Wiley and
Sons, New York, 1978.
[141] Larson, D., Massopust, P., Nashed, M.Z., Nguyen, M.C., Papadakis, M.,
Zayed, A., (Eds.): Frames and Operator Theory in Analysis and Signal Processing.
Contemporary Mathematics. Vol. 451, American Mathematical Society. Providence,
RI, 2008.
[142] Lavrentiev, M.M.: Some Improperly Posed Problems of Mathematical Physics, Iz-
dat. Sibirsk. Otdel, Akad. Nauk. SSSR, Novosibirsk, 1962, English Transl., Springer
Tracts in Natural Philosophy, Vol. 11, Springer-Verlag, Berlin, 1967.
[143] Lieusternik, L.A., Sobolev, V.J.: Elements of Functional Analysis. New York: Un-
gar, 1961.
[144] Lin, Y., Brown, L.D.: Statistical Properties of the Method of Regularization with
Periodic Gaussian Reproducing Kernel. Ann. Statist., 32(4):1723–1743, 2004.
[145] Liskovets, O.A.: Regularization of Variational Inequalities with Pseudo-Monotone
Operators on Approximately Given Sets, Differ. Equations, 11:1970–1977, 1989.
[146] Liu, F.; Nashed, M.Z.: Tikhonov Regularization of Nonlinear Ill-Posed Problems
with Closed Operators in Hilbert Scales. J. Inverse Ill-Posed Problems, 5:363–376,
1997.
[147] Locker J., Prenter, P.M.: Regularization with Differential Operators. J. Math. Anal.
Appl., 74:504–529, 1980.
[148] Louis, A.K.: Inverse und schlecht gestellte Probleme. Teubner, Stuttgart, 1989.
[149] Louis, A.K., Maass, P.: A Mollifier Method for Linear Equations of the First Kind.
Inverse Problems, 6:427–440, 1989.
[150] Louis, A.K., Maass, P., Rieder, A.: Wavelets: Theorie und Anwendungen. B.G.
Teubner Studienbücher, 1998.
[151] Mair, B.A.; Ruymgaart, F.H.: Statistical Inverse Estimation in Hilbert Scales.
SIAM J. Appl. Math., 56:1424–1444, 1996
[152] Mathé, P.: The Lepskii Principle Revisited, Inverse Problems, 22:111–115, 2006.
[153] Mathé, P.; Hofmann, B.: How General are General Source Conditions? Inverse
Problems, 24: 2008.
[154] Mathé, P., Pereverzev, S.V: The Discretized Discrepancy Principle Under General
Source Conditions, J. Complexity, 22:371–381, 2006a.
[155] Mathé, P., Pereverzev, S.V: Regularization of Some Linear Ill-Posed Problems with
Discretized Random Noisy Data. Math. Comp., 75:1913–1929, 2006b.
[156] Mathé, P.; Pereverzev, S.V.: Geometry of Linear Ill-Posed Problems in Variable
Hilbert Scales. Inverse Problems, 19:789–803, 2003.
[157] Marti, J.T.: An Algorithm For Computing Minimum Norm Solutions of Fredholm
Integral Equations of the First Kind. SIAM J. Numer. Anal., 15:1071–1076, 1978.
[158] Marti, J.T.: On the Convergence of an Algorithm Computing Minimum-Norm So-
lutions of Ill-Posed Problems. Math. Comp., 34:521–527, 1980.
[159] Meissl, P.A.: A Study of Covariance Functions Related to the Earth’s Disturbing
Potential. Department of Geodetic Science, No. 151, The Ohio State University,
Columbus, OH, (1971).
[160] Meissl, P.A.: Hilbert Spaces and Their Applications to Geodetic Least Squares
Problems. Boll. Geod. Sci. Aff., N. 1, 181–210, 1976.
[161] Michel, V.: A Multiscale Method for the Gravimetry Problem: Theoretical and Nu-
merical Aspects of Harmonic and Anharmonic Modelling. Ph.D.-thesis, Geomath-
ematics Group, University of Kaiserslautern, Shaker, Aachen, 1999.
[162] Michel, V.: Scale Continuous, Scale Discretized and Scale Discrete Harmonic
Wavelets for the Outer and the Inner Space of a Sphere and Their Application
to an Inverse Problem in Geomathematics. Appl. Comp. Harm. Anal. (ACHA),
12:77–99, 2002.
[163] Michel, V.: A Multiscale Approximation for Operator Equations in Separable Hilbert
Spaces – Case Study: Reconstruction and Description of the Earth’s Interior. Ha-
bilitation Thesis, University of Kaiserslautern, Geomathematics Group, Shaker,
Aachen, 2002a.
[164] Miller, K.: Least Squares Methods for Ill-Posed Problems with a Prescribed
Bounded. SIAM J. Math. Anal., 1:52–74, 1970.
[165] Moore, E.H.: On the Reciprocal of the General Algebraic Matrix. Bulletin of the
American Mathematical Society, 26:394–395, 1920.
[166] Moore, E.H.: General Analysis. Memoirs of the American Mathematical Society,
1:197–209, 1935.
[167] Moritz, H.: Advanced Physical Geodesy. Herbert Wichmann Verlag, Karlsruhe, Aba-
cus Press, Tunbridge, 1980.
[168] Morozov, V.A.: On the Solution of Functional Equations by the Method of Regu-
larization. Soviet Math. Doklady, 7:414–41, 1966.
[169] Morozov, V.A.: Methods for Solving Incorrectly Posed Problems. (In Russian),
Moscow Univ., 1975 (English transl., editor M.Z. Nashed), Springer, New York,
1984.
[170] Morozov, V.A.: Methods for Solving Incorrectly Posed Problems. Springer, New
York, 1984.
[171] Nagel, H.-H., On the Estimation of Optical Flow: Relations Between New Ap-
proaches and Some New Results. Artif. Intell. 33,299–324 1987.
[172] Nashed, M.Z.: Steepest Descent for Singular Linear Operator Equations. SIAM J.
Numer. Anal., 7:358–362, 1970.
ed., Academic, New York, 311–359, 1971.
[174] Nashed, M.Z.: Differentiability and Related Properties of Nonlinear Operators:
Some Aspects of the Role of Differentials in Nonlinear Functional Analysis, in:
Nonlinear Functional Analysis and Applications, L.B. Rall, ed., 103–309, Academic
Press, New York, 1971.
[175] Nashed, M.Z.: Some Aspects of Regularization and Approximations of Solutions
of Ill-Posed Operator Equations. In: Proc. 1972 Army Numerical Analysis Conf.,
Edgewood, MD, 163–181, 1972.
[176] Nashed, M.Z.: Approximate Regularized Solutions to Improperly Posed Linear In-
tegral and Operator Equations. In: Constructive and Computational Methods for
Differential and Integral Equations. D.L. Colton and R.P. Gilbert, Eds., Springer,
New York, LNM 430, 289–322, 1974.
[177] Nashed, M.Z. (ed.): Generalized Inverses and Applications. Academic Press, New
York, 1976.
Generalized Inverses and Applications. Academic Press, New York, 193–244, 1976.
[179] Nashed, M.Z.: Perturbations and Approximation for Generalized Inverses and Lin-
ear Operators. In: Generalized Inverses and Applications. Academic Press, New
York, 325–396, 1976.
[180] Nashed, M.Z.: On Moment-Discretization and Least-Squares Solutions of Linear
Integration Equations of the First Kind. J. Math. Anal. Appl., 53:359–366, 1976.
[181] Nashed, M.Z.: Regularization and Approximation of Ill-Posed Problems in System
Theory. In: Proc. 1979 Conf. on Information Sciences and Systems, G.G.L. Meyer
and C.R. Westgate, Eds., The Johns Hopkins Univ., New York, 568–575, 1979.
[182] Nashed, M.Z.: New Applications of Generalized Inverses in System and Control
Theory. In: Proc. 1980 Conf. on Information Sciences and Systems, J.B. Thomas.
ed. Princeton. NJ, Princeton, 353–358, 1980
[183] Nashed, M.Z.: Continuous and Semicontinuous Analogous of Iterative Method of
Cimmino and Kaczmarz with Applications to the Inverse Radon Transform. In:
Mathematical Aspects of Computerized Tomography. G.T. Herman and F. Natterer,
Eds., Springer, New York, 160–178, 1981.
Problems with Applications to Antenna Theory. IEEE Trans. Antennas Propa-
gation, 29: 220–231, 1981.
[185] Nashed, M.Z.: A New Approach to Classification and Regularization of Ill-Posed
Operator Equations. In: H. Engl and C.W. Groetsch, Eds., Inverse and Ill-Posed
Problems, Band 4, Notes and Reports in Mathematics and Science and Engineering,
Academic Press, Boston, 1987.
[186] Nashed, M.Z.: Inner, Outer, and Generalized Inverses in Banach and Hilbert Spaces,
Numer. Funct. Anal. Optim., 9:261–326, 1987
[187] Nashed, M.Z.: Inverse Problems, Moment Problems and Signal Processing: Un
Menage a Trois, Mathematics in Science and Technology. Siddiqi, A.H., Singh,
R.C. Manchanda, P. (Eds.), World Scientific, New Jersey, 1–19, 2010.
[188] Nashed, M.Z., Engl, H.W.: Random Generalized Inverses and Approximate Solution
of Random Operator Equations. In: Approximate Solution of Random Equations,
A.T. Bharucha-Reid, ed., North-Holland, New York, 149–210, 1979.
[189] Nashed, M.Z., Lin, F.: On Nonlinear Ill-Posed Problems II: Monotone Operator
Equations and Monotone variational Inequalities. In: Theory and Applications of
Nonlinear Operators of Monotone and Assertive Type, A. Kartsatos, ed., Marcel
Dekker, New York, 223–240, 1996.
[190] Nashed, M.Z., Scherzer, O.: Stable Approximation of Nondifferentiable Optimiza-
tion Problems with Variational Inequalities, Contemp. Math., 204:155–170, 1997.
[191] Nashed, M.Z., Scherzer, O.: Stable Approximation of a Minimal Surface Problem
with Variational Inequalities. Abst. and Appl. Anal., 2:137–161, 1997.
[192] Nashed, M.Z., Scherzer, O. (Eds.): Inverse Problems, Image Analysis and Medical
Imaging. Contemporary Mathematics, Vol. 313, American Mathematical Society,
Providence, R.I., 2002.
[193] Nashed, Z.M., Sun, Q.: Function Spaces for Sampling Expansions, in: X. Shen, A.I.
Zayed (Eds.), Multiscale Signal Analysis and Modeling, Springer Science + Business
Media, 81–104, 2013.
[194] Nashed, M.Z., Votruba, F.G.: A Unified Operator Theory of Generalized Inverses.
In: Generalized Inverses and Applications. M.Z. Nashed ed., Academic Press, New
York, 1–109, 1976.
[195] Nashed, M.Z., Wahba, G.: Generalized Inverses in Reproducing Kernel Spaces: An
Approach to Regularization of Linear Operator Equations. SIAM J. Math. Anal.,
5:974–987, 1974.
[196] Nashed, M.Z., Wahba, G.: Approximate Regularized Pseudosolution of Liner Op-
erator Equations when the Data-Vector is not in the Range of the Operator. Bull.
Amer. Math. Soc., 80:1213–1218, 1974a.
[197] Nashed, M.Z., Wahba, G.: Convergence Rates of Approximate Least Squares Solu-
tions of Linear Integral and Operator Equations of the First Kind. Math. Comput.,
28:69–80, 1974b.
[198] Natanson, I.P.: Constructive Function Theory. Frederick Ungar Publ. Co., New
York, 1965.
[199] Natterer F.: The Finite Element Method for Ill-Posed Problems. RAIRO Analyse
Numérique, 11:271–278, 1977.
[200] Natterer F.: Regularisierung schlecht gestellter Probleme durch Projektionsver-
fahren. Numer. Math., 28:329–341, 1977.
[201] Natterer F.: Error Bounds for Tikhonov Regularization in Hilbert Scales. Appl Anal.
18:29–37, 1984.
[202] Neubauer, A.: On Converse and Saturation Results for Tikhonov Regularization of
Linear Ill-Posed Problems. SIAM J. Numer. Anal., 34:517–527, 1997.
[203] Novikoff, P.: Sur le problème inverse du potentiel. Comptes Rendus de l’Académie
des Sciences de l’URSS 18:165–168, 1938.
[204] Ortega, J.M. and Rheinboldt, W.C.: On Discretization and Differentiation of Op-
erators with Applications to Newton’s Method, SIAM J. Numer. Anal., 3, 1966.
[205] Parker, R.L.: The Theory of Ideal Bodies for Gravity Interpretation. Geophys. J.
R. Astr. Soc., 42:315–334, 1975.
[206] Payne, L.E.: Improperly Posed Problems in Partial Differential Equations. SIAM
Publications, Philadelphia, 1975.
[207] Pedlosky, J.: Geophysical Fluid Dynamics. Springer, New York, Heidelberg, Berlin,
1979.
[208] Penrose, R.: A Generalized Inverse for Matrices. Proceedings of the Cambridge
Philosophical Society, 51:406–413, 1955.
[209] Penrose, R.: On Best Approximate Solutions of Linear Matrix Equations. Proceed-
ings of the Cambridge Philosophical Society, 25:17–19, 1956.
[210] Pereverzev, S.V., Schock E.: On the Adaptive Selection of the Parameter in Regular-
ization of Ill-Posed Problems, SIAM Journal on Numerical Analysis, 43:2060–2076,
2005.
[211] Perry, W.L.: On the Bojarski–Lewis Inverse Scattering Method. IEEE Trans. An-
tennas Propagat., 6:826–829, 1974.
[212] Perry, W.L.: Approximate Solution of Inverse Problems with Piecewise Continuous
Solution. Radio Science, 12:634–642, 1977.
130:233–235, 1900.
[214] Petryshyn, W.V.: On generalized inverses and uniform convergence of (I − βK)n
with applications to iterative methods. J. Math. Anal. Appl., (18):417–439, 1967,
MR 34 , 8191.
[215] Petrov, G.I.: Application of Galerkin’s Method to a Problem of the Stability of the
Flow of a Viscous Fluid (in Russian). Priklad. Mate. Mekh., 4:3–12, 1940.
[216] Phillips, B.L.: A Technique for the Numerical Solution of Certain Integral Equations
of the First Kind. J. Assoc. Comput. Math. 9:84–97, 1962.
[217] Plackett, R.L.: A Historical Note on the Method of Least Squares. Biometrika,
36:458–460, 1949.
[218] Plato, R.: Optimal Algorithms for Linear Ill-Posed Problems Yielding Regulariza-
tion Methods. Numer. Funct. Anal. Optimization, 11:111–118, 1990.
[219] Rado, R.: Note on Generalized Inverses of Matrices. Proceedings of the Cambridge
Philosophical Society, 52:600–601, 1956.
[220] Rao, C.R., Mitra S.K.: Generalized Inverse of Matrices and Its Applications. Wiley,
New York, 1971.
[221] Rayleigh, L.: The Theory of Sound. Oxford Univ. Press, 1896.
[222] Reed, M., Simon, B.: Functional Analysis. I, Academic Press, New York, 1972.
[223] Reid, W.T.: Generalized Inverses of Differential and Integral Operators. Theory
and Applications of Generalized Inverses of Matrices (T.L. Boullion and P.L. Odell,
Eds.). Symposium Proceedings, Texas Tech University Mathematics Series, Vol. 4.,
Lubbock; Texas.
[224] Ribière G.: Régularisation d’opérateurs. Rev. Informat. Recherche Opérationnelle,
1:57–79, 1967.
[225] Richter G.R.: Numerical Solution of Integral Equations of the First Kind with
Non-Smooth Kernels. SIAM J. Numer. Anal., 15:511–522, 1978.
[226] Robinson, D.W.: Gauss and Generalized Inverses. Historia Mathematics, 7:118–125,
1980.
[227] Rieder, A.: Keine Probleme mit Inversen Problemen. Vieweg 2003.
[228] Ritz, W.: Über lineare Funktionalgleichungen. Acta Math., 41:71–98, 1918.
[229] Rudin, L.I.: Functional Analysis. Mc Graw-Hill, New York, 1973.
[230] Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear Total Variation Based Noise Removal
Algorithms. Physica D, 60: 259, 1992.
[231] Rummel R., Balmino, G., Johannessen, J., Visser, P., Woodworth P.: Dedicated
Gravity Field Missions – Principles and Aims. J. Geodyn., 33:3–20, 2002.
[232] Saitoh, S.: Theory of Reproducing Kernels and its Applications. Longman, New
York, 1988.
[233] Scherzer, O. (Ed.): Handbook of Mathematical Methods in Imaging. Springer, New
York, 2015.
[234] Seidmann, T.I.: Non-Convergence Results for the Application of Least-Squares Es-
timations onto Ill-Posed Problems. J. Optim. Theory Appl., 30:535–547, 1980.
[235] Shinozaki, Sibuya M., and Ranabe, K.: Numerical algorithms for the Moore–
Penrose Inverse of a Matrix: Direct Methods. Ann. Inst. Statist. Math. 24:193–203,
1972.
[236] Showalter D.W., Ben-Israel, B.: Representation and Computation of the General-
ized Inverse of a Bounded Linear Operator between two Hilbert Spaces. Atti Accad.
Naz. Kincei Rend. Cl. Sci, Fis. Mat. Natur., (8), v. 48:184–194, MR 42, 8310, 1970.
Phys. Earth Planet. Inter., 28:215–229, 1982.
[238] Siegel, C.L.: Über die analytische Theorie der quadratischen Formen, III. Annals
of Mathematics 38:212–291, 1937.
[239] Shin C.E., Sun, Q.; Stability of localized operators. J. Funct. Anal., 256:2417–2439,
2009.
[240] Sjöberg, L.: Station Adjustment of Directions Using Generalized Inverses. In: Borre,
K., Welsch, W. (Eds.), International Federation of Surveyors FIG – Proceedings
Survey Control Networks Meeting of Study Group 5B, 7th–9th July, 1982, Aal-
borg University Centre, Denmark. Schriftenreihe des Wissenschaftlichen Studien-
gangs Vermessungswesen der Hochschule der Bundeswehr München, Heft 7, 381–
399, 1982.
[241] Song, M.: Regularization-Projection Methods and Finite Element Approximations
for Ill-Posed Linear Operator Equations. Ph.-D. Thesis, Univ. Michigan, 1978.
[242] Strand, O.N.: Theory and Methods Related to the Singular Function Expansion and
Landweber’s Iteration for Integral Equations of the First Kind. SIAM J. Numer.
Anal., 11:798–825, 1974.
[243] Sudakov V.N., Khalfin, L.A.: A Statistical Approach to the Correctness of the
Problems of Mathematical Physics. Dokl. Akad. Nauk SSSR., 157-1058-1060, 1964.
[244] Tadmor, E., Nezzar, S., and Vese, L.: A Multiscale Image Representation Using
Hierarchical (BV, L2 ) Decompositions. Multiscale Model. Simul., 2:554–579, 2004.
[245] Taylor, A.E., Lay, D.: Functional Analysis. Wiley, New York, 1979.
[246] Tikhonov, A.N.: On the Stability of Inverse Problems. Dokl. Akad. Nauk SSSR,
39:195–198, 1943.
ularization Method, Dokl. Akad. Nauk SSSR, 151:501–504, 1963.
[248] Tikhonov, A.N.: On Methods of Solving Incorrect Problems, Amer. Math. Soc.
Transl., 2:222–224, 1968.
[249] Tikhonov, A.N., Arsenin V.Y.: Solutions of Ill-Posed Problems. Wiley, New York,
1977.
[250] Tikhonov, A.N., Goncharsky, A.V., Stepanov, V.V., Yagola, A.G.: Numerical Meth-
ods for the Solution of Ill-Posed Problems. Kluwer, Dordrecht. 1995.
[251] Twomey, S.: On the Numerical Solution of Fredholm Integral Equations of the First
Kind by the Inversion of the Linear System Produced by Quadrature. J. Assoc.
Comp. Mach., 10:97–101, 1963.
[252] Vainberg, M.M.: Variational Method and Method of Monotone Operators, Wiley,
New York, 1973.
[253] Varah, J.: On the Numerical Solution of Ill-Conditioned Linear Systems with Ap-
plications to Ill-Posed Problems. SIAM J. Numer. Anal., 10:257–267, 1973.
[254] Vogel, C.R.: Computational Methods for Inverse Problems. SIAM, Philadelphia,
2002.
[255] Vogel, C.R. Oman, M.E.: Iterative Methods for Total Variation Denoising. SIAM
J.Sci. Comput., 17:227–238, 1996.
[256] Wahba, G.: Convergence Rates of Certain Approximate Solutions of Fredholm In-
tegral Equations of the First Kind. J. Approx. Theo., 7:167–185. 1973.
[257] Weck, N.: Zwei inverse Probleme in der Potentialtheorie. In: Mitt. Inst. Theor.
Geodäsie, Universität Bonn, 4:27–36, 1972.
[258] Weiner, H.W. (Eds.): Reproducing Kernel Hilbert Spaces. Applications in Statistical
Signal Processing. Hutchinson Ross, Stroudsburg, PA, 1982.
[259] Werner, D.: Funktionalanalysis. 4. Auflage; Springer, Berlin, 2002.
[260] Werner, J., Numerische Mathematik 1. Vieweg Studium, Braunschweig, 1991.
[261] Wolf, H.: Ausgleichungsrechnung. Formeln zur praktischen Anwendung. Dümmler
Verlag, Bonn, 197).
[262] Xia X.G., Nashed M.Z.: The Backus–Gilbert Method for Signals in Reproducing
Hilbert Spaces and Wavelet Subspaces. Inverse Problems, 10:785–804, 1994.
[263] Xia X.G., Nashed M.Z.: A Modified Minimum Norm Solution Method for Band-
limited Signal Extrapolation with Inaccurate Data. Inverse Problems, 13:1641–1661,
1997.
[264] Yao, K.: Applications of Reproducing Kernel Hilbert Spaces – Bandlimited Signal
Models. Inform. and Control, 11:429–444, 1967.
[265] Yosida, K.: Functional Analysis. 5th ed., Springer, Berlin, 1965.
[266] Zhou, L., Li, X, Pan, F.: Gradient-Based Iterative Identification for Wiener Non-
linear Systems with Non-Uniform Sampling. Nonlinear Dynam., 76:627–634, 2014.
[267] Zwaan, M.: Approximation of the Solution to the Moment Problem in a Hilbert
Space, Numer. Functional Anal. Optimization, 11:601–612, 1990.
[268] Zwaan, M.: MRI Reconstruction as a Moment Problem, Math. Methods Appl. Sci,
15:661–675, 1992.
Willi Freeden
M. Zuhair Nashed
Mathematics Department
University of Central Florida
Orlando, USA

Geodetic Observables and Their Mathematical

Treatment in Multiscale Framework
Willi Freeden and Helga Nutz
Abstract. For the determination of the Earth’s gravitational field various

types of observations are available nowadays, e.g., from terrestrial gravimetry,
airborne gravimetry, satellite-to-satellite tracking, satellite gravity gradiome-
try, etc. The mathematical relation between these observables on the one hand
and the gravitational field and the shape of the Earth on the other hand is
called the integrated concept of physical geodesy. In this paper, an integrated
concept of physical geodesy in terms of harmonic wavelets is presented. Es-
sential tools for approximation are Runge–Walsh type integration formulas
relating an integral over an internal sphere to suitable linear combinations of
observational functionals, i.e., linear functionals representing the geodetic ob-
servables in terms of gravitational quantities on and outside the Earth. A scale
discrete version of multiresolution is described for approximating the gravi-
tational potential on and outside the Earth’s surface. Furthermore, an exact
fully discrete wavelet approximation is developed for the case of bandlimited
wavelets. A method for combined global outer harmonic and local harmonic
wavelet modeling is proposed corresponding to realistic Earth’s models.
Keywords. Integrated wavelet concept, scaling function, Runge–Walsh approx-

imation, geodetic observables, Meissl schemata.
1. Introduction
Gravity as observed on the Earth’s surface is the combined effect of the gravita-
tional mass attraction and the centrifugal force due to the Earth’s rotation. Under
the assumption that the centrifugal force is explicitly known, the determination
of the gravity mainly reduces to getting knowledge of the gravitation. According
to the classical Newton Law of Gravitation (1687), knowing the density distribu-
tion of a body such as the Earth, the gravitational potential can be computed
everywhere in the Euclidean space R3 .
316 W. Freeden and H. Nutz
Although Earth’s gravitational field modeling is always governed by the same

physical laws, it changes its nature when it is seen from different spatial and time
scales. To be more concrete, if one looks at gravitational field determination on
the basis of an increasing spatial magnification and accuracy, we have to go from
something that is suitably characterized by a simple mass point, on astronomical
scale, to what is described by a global truncated multipole (i.e., outer harmonic)
model, at scales corresponding to satellite altimetry, down to wavelengths of about
100 km. By further zooming in we can reach a spatial resolution of about 1 km
showing a very complicated pattern, strongly related to the shape of the Earth
and to irregular masses inside the Earth’s crust. Simultaneously, the error in the
knowledge of the gravitational field models goes from 5 Gal, the flattening effect,
down to 10 mGal in a today’s global model, down to about 10−1 mGal at the
regional 1 km resolution or even better. There is also a change of the gravitational
field in the time scale depending on the time interval under consideration, for in-
stance, gravitational changes due to geotectonic displacements of masses inside the
Earth on very long time scales. It changes because of motions of the rotational axis
inside the Earth’s body and it shows a periodic change because of the continent
and ocean reactions to the torques generated by the moon and the sun. Finally,
gravitation shows a change because of human activities, for instance, because of
the presence of artificial lakes, height’s variations in the water-bearing stratum un-
der cities, etc. It is also worth mentioning that there are certain relations between
different scales in the time-like behaviour and in the space-like behaviour of the
gravitational field. In any way, it may be assumed for global up to regional mod-
eling purposes that the time-like variations of the field are either well predictable
(like tides etc.) or so slow as to be neglected, e.g., on the scale of a decade, or so
small and local as to be beyond the scope of interest. Thus, global gravitational
field modeling as scientific issue is by definition based on the assumption of a sta-
tionary gravitational field with a spatial resolution ranging from a worldwide scale
down to about 1 km and from about 1000 Gal of the full field down to, at least,
10−1 mGal, or even better in some regional areas.
What we would like to present in this contribution are mathematical struc-
tures in straightforward continuation to the monograph [19] by which the grav-
itational part of the gravity field can be approximated progressively better and
better, reflecting an increasing flow of observations of terrestrial, airborne and/or
satellite type, e.g., terrestrial gravimetry, airborne gravimetry, satellite altime-
try, satellite-to-satellite tracking (SST), satellite gravity gradiometry (SGG), etc.
More precisely, we shall try to outline the canonical bridge of gravitational field
determination from the well-established global outer harmonic approximation cor-
responding to a spherical Earth to modern multiscale methods involving the actual
geometry of the Earth’s surface (thereby neglecting, e.g., the small effect of the
atmosphere in the outer space).
The so-called disturbing potential is probably the most crucial quantity in
gravity field modeling. The disturbing potential is a scalar quantity which is ob-
tained as the difference between the gravity potential of the Earth and the normal
Geodetic Observables and Their Mathematical Treatment 317
gravity potential of a reference surface, usually an ellipsoid. The deviations of the

gravity potential from the normal potential are relatively small. Note that both
the gravity potential and the normal gravity potential contain the same centrifugal
potential. Thus, the disturbing potential is harmonic in the outer space.
At this stage some remarks should be made in order to clarify our approach
in more detail:
1. The mathematical connection between the observables, the gravity field and
the shape of the Earth is called the integrated concept of physical geodesy.
2. The foundation of the integrated geodesy approach is the fact that every
geodetic measurement is a functional which may assumed to be suitably
linearizable by introducing, e.g., normal potentials associated to a reference
surface such as an ellipsoid. In other words, the relation between the object
function, i.e., the geopotential and the data, may be supposed to be linear.
3. More and more measurements refer to satellites and cannot be modeled as
functionals of the gravitational potential on the boundary. Although these
observations show a denser observational distribution, they are much more
difficult to handle, since they show an exponentially spectral smoothing while
moving to the outer space. As a consequence, essential knowledge of the grav-
itational potential should be based on ground observations, but gravitational
field modeling cannot be treated only within a boundary-value formulation
because of spaceborne observations. This fact is the reason why we do not
speak of the “geodetic boundary-value problem (GBVP)” but of the “inte-
grated concept”.
4. Concerning the layout of this contribution a particular interest is focussed
on the satellite methods SST und SGG, which are introduced within the
framework of pseudodifferential operators assuming non-spherical (orbital)
geometry.
5. An important feature of our contribution are the so-called Meissl schemata
which are graphical illustrations for the conversion of data both on different
heights (terrestrial level, satellite orbit) and of different degrees of derivative
of the gravitational potential. The comparison between data on the (spheri-
cal) Earth’s surface and the orbital sphere was primarily carried out by Meissl
(1971) and has been transformed by Rummel [60, 61, 63] and by Rummel and
van Gelderen [64, 65] into a more general framework concerning relations be-
tween different gravity quantities in the framework of outer harmonics. One
of our objectives is the extension of the Meissl schemata to the concept of
multiscale decomposition of scalar functions, vector, and tensor fields. In
principle, we follow the ideas of mathematical classification first presented in
[19, 29, 32–34] for the scalar case and extended in the Ph.D.-thesis [58] to
the vector and tensor approach.
2. Current state of gravity field determination

Positioning systems are ideally located as far as possible from the Earth, whereas
gravity field sensors are ideally located as close as possible to the Earth. Following
these basic principles, various positioning and gravity field determination tech-
niques have been designed. Sensors may be sensitive to local or global features of
the gravity field. Considering the spatial location of the data, we may distinguish
between terrestrial (surface), airborne, and spaceborne methods. Regarding the
data type we have various measurement principles of the gravity field (see, for
example, [9–11, 51] and the references therein for more details) leading to different
types of data.
2.1. Important geodetic observables

(a) Gravity Measurements: The force of gravity provides a directional structure
to the space above the Earth’s surface. It is tangential to the vertical plumb
lines and perpendicular to all (level) equipotential surfaces. Any water sur-
face at rest is part of a level surface. Level (equipotential) surfaces are ideal
reference surfaces, for example, for heights. The geoid is defined as that level
surface of the gravity field which best fits the mean sea level. Gravity vectors
can be measured by absolute or relative gravimeters. The highest available ac-
curacy relative gravity measurements are conducted at the Earth’s surface.
Measurements on ships and in aircrafts deliver reasonably good data only
after the removal of inertial noise. Gravity data are converted into gravity
anomalies by subtracting a corresponding reference potential derived from
a simple gravity field model associated to an, e.g., ellipsoidal surface (see
also Appendix A). Gravity anomalies are furthermore converted into mean
gravity anomalies by a proper averaging process over well defined areas. It
should be pointed out that the distribution of Earth’s gravity data on a
global scale is far from being homogeneous with large gaps, in particular over
oceans but also over land. In addition, the quality of the data is very distinct.
Thus, terrestrial gravity data coverage now and in the foreseeable future is
far from being satisfactory for the global purpose of geoidal determination
(at an accuracy of essentially less than one centimeter).
(b) Vertical Deflections. The direction of the gravity vector can be obtained by
astronomical positioning. Measurements are only possible on the Earth’s sur-
face. Observations of the gravity vector are converted into so-called vertical
deflections by subtracting a corresponding reference direction derived from a
simple gravity field model associated to an ellipsoidal surface. Vertical deflec-
tions are tangential fields of the anomalous potential in a spherical Earth’s
model. Due to the high measurement effort required to acquire these types
of data compared to a gravity measurement, the data density of vertical
deflections is much less than that of gravity anomalies. Gravitational field
determination based on the observation of vertical deflections and combined
with gravity is feasible in smaller areas with good data coverage.
(c) Satellite Radar Altimetry. Satellite radar altimetry has demonstrated an im-
pressive capability of mapping the surface of the oceans. The ocean surface
is a good approximation of an equipotential surface and, as such, its offset
from the geoid at mean sea level (mean in terms of time) is called sea surface
topography. This offset, which can be as large as two meters, reflects many
effects including the variables salinity, ocean temperature, ocean currents,
variable atmospheric conditions such as wind and air pressure perturbations,
tides, etc. Since the sea surface topography refers to the geoid, the precise
and sufficiently detailed knowledge of the geoid is mandatory.
(d) Global Gravitational Field Models. On the basis of all satellite data, collected
over the last decades in orbits at different altitudes and inclinations, only long
wavelength components of the global gravity field can be recovered. There
are two reasons for this fact: First, an orbit as such is rather insensitive to
local features of the gravitational field, and this insensitivity increases with
increasing orbit altitude. Second, the satellites which can and are being used
are flying at altitudes which are too high for a better purpose such as local
gravimetry. Therefore, satellite-only global gravity field models are reliable
to a moderate maximum degree expressed in a potential representation in
terms of spherical harmonics. Considering the shortcomings of satellite-only
gravity field models and of the information content of surface data, several in-
stitutions have been working for many years on the combination of both data
sets. This work in geodesy has resulted in various gravitational field models in
terms of spherical harmonics. All gravity field data available worldwide have
entered into the production of this model. Therefore, such models represent
the latest state of the art in global gravitational field knowledge.
2.2. Satellite concepts and airborne data

The three satellite concepts which are of importance for gravity field determination
are satellite-to-satellite tracking in the high-low mode (SST hi-lo), satellite-to-
satellite tracking in the low-low mode (SST lo-lo), and satellite gravity gradiometry
(SGG). Common to all three concepts is that the determination of the Earth’s
gravitational field is based on the measurement of the relative motion (in the
Earth’s gravity field) of test masses.
1. Satellite-to-Satellite Tracking. In the case of SST hi-lo the low flying test mass
is a low earth orbiter (LEO) and the high flying test masses are the satellites
of the GNSS-system (i.e., GPS, GLONASS, Galileo, and Beidou). As, for
example, the GNSS-receiver mounted on the LEO always “contacts” four or
even more of the GNSS satellites the relative motion of the LEO can be mon-
itored three-dimensionally, i.e., in all three coordinate directions. The lower
the orbit of the LEO the higher is its sensitivity with respect to the spatial
variations of the gravitational forces but to skin forces as well (atmospheric
drag, solar radiation, albedo, etc.). The latter have either to be compensated
for by a drag-free mechanism or be measured by a three axis accelerometer.
Also the high orbiters, the GNSS satellites, are affected by non-gravitational
forces. However the latter can be modeled quite well. They affect mainly
the very long spatial scales, and to a large extent their effect averages out.
In addition, the ephemerides of the GNSS satellites are determined very ac-
curately by the large network of ground stations. In the case of SST lo-lo
the relative motion between two LEOs, chasing each other, is measured with
highest precision. The quantity of interest is the relative motion of the centre
of mass of the two satellites. Again, the effect of non-gravitational forces on
the two spacecraft either has to be compensated actively or be measured.
2. Satellite Gravity Gradiometry. The satellite gravity gradiometry technique
is the measurement of the relative acceleration, not between free falling test
masses like satellites, but of test masses at different locations inside one satel-
lite. Each test mass is enclosed in a housing and kept levitated (floating, with-
out ever touching the walls) by a capacitive or inductive feedback mechanism.
The difference in feedback signals between two test masses is proportional to
their relative acceleration and exerted purely by the differential gravitational
field. Non-gravitational acceleration of the spacecraft affects all accelerome-
ters inside the satellite in the same manner and so ideally drops out during
differencing. The rotational motion of the satellite affects the measured dif-
ferences. However, the rotational signal (angular velocities and accelerations)
can be separated from the gravitational signal, if acceleration differences are
taken in all possible (spatial) combinations (= full tensor gradiometer). In
order to achieve a higher sensitity, an orbit as low as possible is of great
importance.
In a unified view on spaceborne missions (see, e.g., [9–11, 51]), one can argue
that the basic observable in all three cases is gravitational acceleration. In the case
of SST hi-lo, with the motion of the high orbiting GNSS satellites assumed to be
perfectly known, this corresponds to an in situ 3-D acceleration measurement in
the LEO. For SST lo-lo it is the measurement of acceleration difference over the
intersatellite distance and in the line-of-sight (LOS) of the LEOs. Finally, in the
case of gradiometry, it is the measurement of acceleration differences in 3-D over
the tiny baseline of the gradiometer. In short we are confronted with the following
situation:
SST hi-lo: 3-D acceleration = gravitational gradient,
SST lo-lo: acceleration difference = difference in gradient,
SGG: differential = gradient of gradient (“tensor”).
As explained in more detail by W. Freeden [19], in mathematical sense, it is a
transition from the first derivative of the gravitational potential via a difference in
the first derivative to the second derivative. The guiding parameter that determines
sensitivity with respect to the spatial scales of the Earth’s gravitational potential
is the distance between the test masses, being almost infinite for SST hi-lo and
almost zero for gradiometry.
3. Airborne Gravimetry. Airborne gravimetry is a highly sensitive detection

method of the gravitational potential of the Earth by a gravity accelero-
meter mostly for regional and/or local purposes. Proposals to implement
airborne gravimetry go back to the late fifties of the last century, and first
flight experiments were already done in the early sixties. A major obstacle
of such techniques at that time was the inaccuracy of navigational informa-
tion (e.g., velocity and acceleration of the space vehicle) which is needed to
obtain the desired precision. Although at an appropriate level of accuracy
airborne gravimetry is vastly superior in economy and efficiency to pointwise
terrestrial methods, there were serious doubts in the seventies and eighties of
ever achieving useful results. In the early nineties, however, great advances
in GNSS technology opened new ways to resolve the navigational problems.
More explicitly, altitude, position, and velocity of the airborne gravity sys-
tem become sufficiently computable from the inertial measurements updated
by GNSS carrier phase and Doppler observations. Vehicle accelerations are
derivable from GNSS data only, so that in a third step the airborne gravity
disturbance is determinable from the difference between the force vector and
the GNSS-derived acceleration vector. Nowadays, some industrial companies
are perfecting their system concepts by paying careful attention to the op-
erational conditions under which an airborne gravimeter works best, also for
progress in gravimetric exploration.
All in all, over the last decades, geoscientists have realized the great complex-
ity of the Earth and its environment. In particular, the knowledge of the gravity
potential and its level (equipotential) surfaces have become an important issue. It
was realized that dedicated highly accurate gravity field sensors, when operating
in an isolated manner, have their shortcomings, and combining data from differ-
ent sensors is therefore the way forward. At this stage of development, the global
determination of the Earth’s gravitational field is a mathematical challenge which
should include the numerical progress obtainable by modern multiscale approxi-
mation.
2.3. Gravity field applications

The knowledge of the gravitational field of the Earth is of great importance
for many applications from which we only mention some significant examples
(cf. [19, 61]):
(i) Geodesy and Civil Engineering. Accurate heights are needed for civil con-
structions, mapping, etc. They are obtained by leveling, a very time consum-
ing and expensive procedure. Nowadays, geometric heights can be obtained
fast and efficiently from space positioning (GNSS). The geometric heights
are convertible to leveled heights by subtracting the precise geoid, which is
achieved by a high resolution gravitational potential. To be more specific, in
those areas where good gravity information is available already, the future
data information will eliminate all medium and long wavelength distortions in
unsurveyed areas. For example, GNSS (GPS, GLONASS, Galileo, or Beidou)

together with today’s satellite missions provide high quality height informa-
tion at global scale.
(ii) Satellite Orbits. For any positioning from space, the uncertainty in the orbit
of the spacecraft is the limiting factor. The spaceborne techniques eliminate
basically all gravitational uncertainties in satellite orbits.
(iii) Solid Earth Physics. The gravity anomaly field derivable from future satellite
observations has its origin mainly in mass inhomogeneities of the continen-
tal and oceanic lithosphere. Together with height information and regional
tomography, a much deeper understanding of tectonic processes is obtainable.
(iv) Physical Oceanography. Altimeter satellites in combination with a precise
geoid deliver global dynamic ocean topography. From ocean topography,
global surface circulation and its variations in time can be computed re-
sulting in efficient ocean modeling. Circulation allows the determination of
transport processes of, e.g., polluted material. Moreover, ocean modeling is
an important indicator of climate change.
(v) Earth System. There is a growing awareness of global environmental problems
(for example, the CO2 -question, the rapid decrease of rain forests, global sea
level changes, etc.). What is the role of the airborne methods and satellite
missions in this context? They do not tell us the reasons for physical pro-
cesses, but it is essential to bring the phenomena into one system (e.g., to
make sea level records comparable in different parts of the world). In other
words, equipotential surfaces such as the geoid may be viewed as an almost
static reference for many rapidly changing processes and at the same time as
a “frozen picture” of tectonic processes that evolve in geological time spans.
(vi) Exploration Geophysics and Prospecting. Knowledge of local geologic struc-
tures can easily be gained by means of terrestrial and airborne data so grav-
ity prospecting can be done over land or sea areas using different techniques
and equipment. Terrestrial gravimetry was first applied to prospect for salt
domes (e.g., in the Gulf of Mexico), and later for looking for anticlines in
continental areas. In future, embedded in (regional) airborne and (global)
spaceborne gravity information such as satellite-to-satellite tracking (SST)
and/or satellite gravity gradiometry (SGG) (see, e.g., [19, 27, 32] and the
references therein), new promising components in gravimetrically oriented
modeling can be expected, for example, based on multiscale modeling pro-
viding reconstruction and decomposition of geological signatures.
2.4. Principles of multiscale approximation

Spaceborne observation combined with terrestrial and airborne activities provide
huge datasets of the order of millions of data (see [9–11, 51, 63]). Standard math-
ematical theory and numerical methods are not at all adequate for the solution
of data systems with such a structure, because these methods are not adapted
to the specific properties of the data set. They quickly reach their capacity limit
even on very powerful computers. An adequate reconstruction of the gravitational
field from the huge and heterogeneous data material requires a careful multiscale
analysis of the gravitational potential, fast solution techniques, and a proper sta-
bilization of the inverse character of satellite problems by regularization. In order
to achieve these objectives various strategies and structures must be introduced
reflecting the different aspects of geopotential determination. While global long-
wavelength modeling can be adequately done by use of spherical harmonic expan-
sions, it becomes more and more obvious that harmonic splines and/or wavelets
are most likely the candidates for medium and short-wavelength approximation.
The concept of harmonic wavelets, however, demands its own nature which only on
exploration areas of small size may be developed to some extend from the theory
in Euclidean spaces. Fundamental results known from the Euclidean wavelet ap-
proach have to be recovered. Nevertheless, the stage is set for working out and im-
proving essential ideas and results involving harmonic wavelets. Why are harmonic
wavelets important in future gravitational potential determination? Following [19],
the answer is summarized in the following sentence:
Harmonic wavelets are “building blocks” that enable fast decorrelation of gravi-
tational data. Thus three features are incorporated in this way of thinking about
georelevant harmonic wavelets, namely basis property, decorrelation, and efficient
algorithms. These aspects should be discussed in more detail:
(i) Basis property
Wavelets are building blocks for the approximation of arbitrary functions
(signals). In mathematical understanding this formulation expresses that the
set of wavelets forms a “frame” (see, e.g., [6] for details in classical one-
dimensional theory).
(ii) Decorrelation
Wavelets possess the ability to decorrelate the signal. This means, that the
representation of the signal via wavelet coefficients occurs in a “more con-
stituting” form as in the original form reflecting a certain amount of space
and frequency (more accurately, momentum) information. The decorrelation
enables the extraction of specific information contained in a signal through
a particular number of coefficients. Signals usually show a correlation in the
frequency (momentum) domain as well as in the space domain. Obviously,
since data points in a local neighborhood are stronger correlated as those data
points far-off from each other, signal characteristics often appear in certain
frequency bands. In order to analyze and reconstruct such signals, we need
“auxiliary functions” providing localized information in the space as well as
in the frequency domain. In applications, different approaches have been re-
alized in the field of signal analysis before the occurrence of wavelets: on the
one hand, the Fourier theory allows a trendsetting bandlimited decomposi-
tion, on the other hand, the Haar theory offers short-wavelets spacelimited
decomposition. The (Heisenberg) uncertainty principle (see, e.g., [21]) tells
us that a simultaneous sharp localization in frequency as well as space do-
main is exclusive. Even more within a “zooming-in process”, the amount of
frequency as well as space contribution can be specified in quantitative way.

A so-called scaling function forms a compromise in which a certain balanced
amount of frequency and space localization in the sense of the uncertainty
principle is realized. In consequence, each scaling function depends on two
variables, namely a “shifting” and a scaling parameter, which control the
amount of the space localization to be available at the price of the frequency
localization, and vice versa. Associated to each scaling function is a wavelet
function, which here is simply understood to be the difference of two succes-
sive scaling functions. All in all, filtering (convolution) with a scaling function
takes the part of a lowpass filter, while convolution with the corresponding
wavelet function provides a bandpass filtering. A multiscale approximation of
a signal is the successive execution of an efficient evaluation process by use of
scaling and wavelet functions which show more and more space localization at
the cost of frequency localization. The wavelet transform within a multiscale
approximation lays the foundation for the decorrelation of a signal.
(iii) Efficient algorithms
Wavelet transformation provides efficient algorithms because of the space-
localizing character. The successive decomposition of the signal by use of
wavelets at different scales offers the advantage for efficient and economic
numerical calculation (e.g., tree algorithm). The detail information stored
in the wavelet coefficients leads to a reconstruction from a rough to a fine
resolution and to a decomposition from fine to rough resolution in form of tree
algorithms. In particular, the decomposition algorithm is an excellent tool for
the post-processing of a signal into “constituting blocks” by decorrelation,
e.g., the specification of signature bands corresponding to certain geological
formations.
3. Geodetically relevant Sobolev spaces

We start our mathematical foundation of Meissl schemata by introducing some
basic information related to the theory of geodetic observables within the frame-
work of Sobolev spaces. We adopt the following general scheme of notation which
is non-standard in geodesy, but extremely helpful in establishing Meissl schemata
especially for the vectorial and tensorial framework. Capital letters (F , G, . . . )
are used for scalar functions, small letters (f , g, . . . ) represent vector fields and
small boldface letters (f , g, . . . ) represent tensor fields of second rank. As usual, a
scalar function having k continuous derivatives is said to be of class C (k) whereas
L2 denotes the Hilbert space of square integrable functions. A vector field having
k continuous derivatives is said to be of class c(k) and l2 denotes the Hilbert space
of square-integrable vector fields. Finally, the space of all tensor fields having k
continuous derivatives is denoted by c(k) and l2 denotes the Hilbert space of all
square-integrable tensor fields.
Σ ⊂ R3 is called a regular surface if Σ is the boundary of a regular region

Σint ⊂ R3 , i.e., Σ = ∂Σint , with the following properties (cf. [20]):
(i) Σ constitutes an orientable piecewise smooth Lipschitzian manifold
of dimension 2.
(ii) The origin is contained in Σint .
(iii) Σ divides R3 into the “inner space” Σint and the “outer space”
Σext = R3 \Σint , Σint = Σint ∪ Σ.
Georelevant regular surfaces Σ are, for example, the sphere, the ellipsoid, the
telluroid, the geoid, and the regular Earth’s surface.
The geometric concept to be discussed in our approach is as follows (see
Figure 3.1): Σ denotes the Earth’s surface which we assume to be known and

Figure 3.1. Geometric concept characterizing the surface of the

Earth Σ and the orbit of a satellite Γ.
regular. Γ is the orbit of a satellite which is not necessarily a closed surface. σ is the
radius of a so-called Runge (in the jargon of geodesy, Bjerhammar) sphere inside
the Earth, that is σ < α = inf x∈Σ |x|. The value γ is a lower bound of the lowest
possible altitude of the satellite, i.e., γ < inf x∈Γ |x|. Ωext
σ = {x ∈ R
3
: |x| > σ}
denotes the outer space of the sphere Ωσ with radius σ around the origin 0, whereas
Σext denotes the outer space of the (actual) Earth.
Let V : Ωext ext 3 ext 3 3
σ → R, v : Ωσ → R , and v : Ωσ → R ⊗ R , respectively, be a
ext
scalar, vector, and tensor field on the set Ωσ . We say that V , v, v, respectively,
are harmonic on Ωext σ if V , v, v are twice continuously differentiable on Ωσ and
ext
ext
ΔV = 0, Δv = 0, Δv = 0 on Ωσ .
Without proof we mention some well-known theorems concerning harmonic

fields on Ωext
σ (for the proofs see, for example, [20, 38, 47]):
(1) Every harmonic field in Ωext ext

σ is analytic in Ωσ , i.e., every harmonic field is
determined by its local properties.
(2) Harnack’s convergence theorem: Let Vδ : Ωext σ → R, vδ : Ωext
σ → R3 , and
ext 3 3 ext
vδ : Ωσ → R ⊗ R , respectively, be harmonic on Ωσ for each value δ (0 <
δ < δ0 ), and regular at infinity. Moreover, let
Vδ → V, δ → 0, δ > 0, (3.1)
vδ → v, δ → 0, δ > 0, (3.2)
vδ → v, δ → 0, δ > 0, (3.3)
uniformly on each subset K of Ωext ext ext

σ with dist(K, ∂Ωσ ) > 0. Then V : Ωσ →
R, v : Ωσ → R , and v : Ωσ → R ⊗ R , respectively, is harmonic on Ωext
ext 3 ext 3 3
σ
and regular at infinity.
ext
(3) Let V : Ωext
σ → R be twice continuously differentiable on Ωσ and continuous
ext (0) ext (2) ext ext
on Ωσ , i.e., V ∈ C (Ωσ ) ∩ C (Ωσ ), harmonic on Ωσ , and regular at
infinity. Then the maximum/minimum principle tells us that
sup |V (x)| ≤ sup |V (x)| . (3.4)
x∈Ωext x∈Ωσ
σ
(4) There is a so-called fundamental solution (singularity function) S : x →

|x − y|−1 , x = y, with respect to the Laplace operator Δ such that the
fundamental theorem of potential theory

1 ∂V ∂ 1
(y) − V (y) dω(y)
∂Ωext |x − y| ∂ν ∂νy |x − y|
σ
⎧
⎨ −4πV (x), x ∈ Ωext σ ,
= −2πV (x), x ∈ ∂Ωextσ ,
⎩
0, x∈ / Ωext
σ ,
holds true.
3.1. Scalar outer harmonic and Sobolev theory

As already explained, we let Ωσ ⊂ R3 be the sphere around the origin with radius
σ > 0, Ωint ext
σ is the inner space of Ωσ , and Ωσ is the outer space. We let Ω = Ω1 .
By virtue of the isomorphism Ω ξ → σξ ∈ Ωσ we assume functions F : Ωσ → R
to be defined on Ω. It is clear that the function spaces defined on Ω admit their
natural generalizations as spaces of functions defined on Ωσ . We have, for example,
C (∞) (Ωσ ), Lp (Ωσ ), etc.
Let {Yn,m }n∈N0 ;m=1,...,2n+1 be an L2 -orthonormal system of (surface) spheri-
cal harmonics. Obviously, such an L2 (Ω)-orthonormal system of spherical harmon-
ics forms an orthogonal system on Ωσ (with respect to (·, ·)L2 (Ωσ ) ). More explicitly,
we have

x x
(Yn,k , Yp,q )L2 (Ωσ ) = Yn,k Yp,q dω(x) = σ 2 δn,p δk,q , (3.5)
Ωσ |x| |x|
where δn,p is the Kronecker symbol and dω is the surface element. With the re-
lationship ξ ↔ σξ, the surface gradient ∇∗;σ and the Beltrami operator Δ∗;σ on
Ωσ , respectively, have the representation ∇∗;σ = (1/σ)∇∗;1 = (1/σ)∇∗ , Δ∗;σ =
(1/σ 2 )Δ∗;1 = (1/σ 2 )Δ∗ , where ∇∗ , Δ∗ are the surface gradient and the Beltrami
operator of the unit sphere Ω.
We now introduce the system {Yn,k σ
}n=0,1,...; k=1,...,2n+1 by letting

1 x
σ
Yn,k (x) = Yn,k , x ∈ Ωσ . (3.6)
σ |x|
Due to (3.5) the system {Yn,k
σ
}n=0,1,...;k=1,...,2n+1 is an orthonormal basis in L2 (Ωσ ):
·L2 (Ωσ )
L2 (Ωσ ) = span σ
n=0,1,...; (Yn,k ) . (3.7)
k=1,...,2n+1
The system {Hn,m (σ; ·)}n∈N0 ;m=1,...,2n+1 , of scalar outer harmonics defined by
n+1
1 σ x
Hn,m (σ; x) = Yn,m , x ∈ Ωext
σ ,
σ |x| |x|
satisfies the following properties:
• Hn,m (σ; ·) is of class C (∞) (Ωext
σ ),
• Hn,m (σ; ·) is harmonic in Ωext ext
σ , i.e., Δx Hn,m (σ; x) = 0 for x ∈ Ωσ ,
−1
• Hn,m is regular at infinity, i.e., |Hn,m (σ; x)| = O(|x| ), x| → ∞,
1
• n,m (σ; ·)|Ωσ = σ Yn,m ,
H
• Ωσ
Hn,m (σ; x)Hk,l (σ; x)dω(x) = δn,k δm,l .
As it is well known (cf., e.g., [32, 57]), the addition theorem of outer harmonics
reads as follows:
2n+1
2 n+1
2n + 1 σ x y
Hn,m (σ; x)Hn,m (σ; y) = Pn · , (3.8)
m=1
4πσ 2 |x| |y| |x| |y|
for all (x, y) ∈ Ωext ext

σ × Ωσ and n ∈ N0 , where Pn denotes the Legendre polynomial
of degree n. Harm n (Ωext
σ ) denotes the space of all outer harmonics of order n,
n ∈ N0 :
Harm n (Ωext
σ ) = spanm=1,...,2n+1 (Hn,m (σ; ·)).
It is well known that dim(Harm n (Ωext ext

σ )) = 2n + 1. We let Harm p,...,q (Ωσ ) be the
space of all linear combinations of the functions Hn,m (σ; ·) on Ωσ , n = p, . . . , q,
ext
m = 1, . . . , 2n + 1, i.e.,
5
q
Harm p,...,q (Ωext
σ ) = Harm n (Ωext
σ ).
n=p
The space Harm p,...,q (Ωext

σ ) has the reproducing kernel KHarm p,...,q (Ωext (·, ·)
σ )
given by
q 2n+1

KHarm p,...,q (Ωext ) (x, y) = Hn,m (σ; x)Hn,m (σ; y)
σ
n=p m=1
q 2
2n + 1 σ2 x y
= Pn · . (3.9)
n=p
4πσ 2 |x| |y| |x| |y|
Pot (Σext ) denotes the space of all functions (potentials) U : Σext → R with
• U ∈ C (2) (Σext ),
• U satisfies the Laplace equation in the outer space, i.e., Δx U (x) = 0, x ∈ Σext ,
• U is regular at infinity, i.e., |U (x)| = O(|x|−1 ), |x| → ∞.
As usual, for k = 0, 1, . . . , we let Pot (k) (Σext ) be the space of functions F : Σext →
R such that F |Σext ∈ Pot (Σext ) and F ∈ C (k) (Σext ), in brief,
Pot (k) (Σext ) = Pot (Σext ) ∩ C (k) (Σext ). (3.10)
It is known from [13] and [17] that
·L2 (Σ)
L2 (Σ) = span n=0,1,...; (Hn,m (σ; ·))|Σ , (3.11)
m=1,...,2n+1
·C (0) (Σ)
C (0) (Σ) = span n=0,1,...; (Hn,m (σ; ·))|Σ . (3.12)
m=1,...,2n+1
Furthermore (cf. [13]),

·C (0) (Σext )
Pot (0) (Σext ) = span n=0,1,...; (Hn,m (σ; ·))|Σext . (3.13)
m=1,...,2n+1
Next we introduce Sobolev spaces H(Ωext σ ) (cf. [14]). We start with a general
definition based on the concept of summable sequences, give some examples for
spaces with a reproducing kernel structure, and, finally, introduce the well-known
Hs (Ωext
σ )-spaces.
The introduction of the Sobolev spaces may be based on a linear space A
consisting of all sequences {An } of real numbers An , n = 0, 1, . . ., i.e.,
A = {{An } : An ∈ R, n = 0, 1, . . .} .
For given sequences {An }, {Bn } ∈ A we denote by N (Bn−1 An ) the set of all non-
negative integers n for which Bn A−1 −1
n exists and is different from 0. Let N0 (Bn An )
−1
denote the complement of N (Bn An ) in N0 . Consequently, it follows that N0 =
N (Bn−1 An )∪N0 (Bn−1 An ) and N (Bn−1 An )∩N0 (Bn−1 An ) = ∅. In particular, if {Bn }
is chosen such that Bn = 1 for all n ∈ N0 , N (An ) is the set of all integers n ∈ N0 for
which An = 0, and N0 (An ) is the set of all integers n ∈ N0 with An = 0. Further
on N (An ) is always assumed to be non-void. Moreover, we write N instead of
N (An ) if no confusion is likely to arise.

(∞)
Consider the set E(Ωext ext
σ ) = E({An }; Ωσ ) of all functions F ∈ Pot (Ωext
σ )
of the form
2n+1

F = F ∧ (n, m)Hn,m (σ; ·) (3.14)
n∈N m=1
with
F ∧ (n, m) = F ∧L2 (Ωσ ) (n, m) = F (y)Hn,m (σ; y) dω(y)
Ωσ
satisfying
2n+1

A2n (F ∧ (n, m))2 < ∞ (3.15)
n∈N m=1
(note that Σn∈N means that the sum is extended over all non-negative integers n
with n ∈ N ). From the Cauchy–Schwarz inequality it follows that

2n+1

2 ∧ ∧
An F (n, m)G (n, m) (3.16)

n∈N m=1
1/2 1/2
2n+1
2n+1

≤ A2n (F ∧ (n, m))2 A2n (G∧ (n, m))2
n∈N m=1 n∈N m=1
for all F, G ∈ E(Ωext

σ ), hence, the left-hand side of (3.16) is finite whenever each
member of the right-hand side is finite. This is the reason why we are able to
impose on E(Ωext
σ ) an inner product (·, ·)H({An };Ωext ) by letting
σ
2n+1

(F, G)H({An };Ωext ) = A2n F ∧ (n, m)G∧ (n, m). (3.17)
σ
n∈N m=1
The associated norm is given by

1/2
2n+1

F H({An };Ωext ) = A2n ∧
(F (n, m)) 2
. (3.18)
σ
n∈N m=1
Summarizing our results we therefore obtain the following definition.

Definition 3.1. The Sobolev space H(Ωext ext
σ ) (more accurately: H({An }; Ωσ )) is the
ext ext
completion of E(Ωσ )(= E({An }; Ωσ )) under the norm · H({An };Ωext ) :
σ
·H({A
n };Ωext
σ )
H({An }; Ωext ext
σ ) = E({An }; Ωσ ) .
H(Ωext
σ ) equipped with the inner product corresponding to the norm (3.18) is a
∗{A }
Hilbert space. The system {Hn,m n (σ; ·)} given by
∗{An }
Hn,m (σ; x) = A−1
n Hn,m (σ; x), x ∈ Ωext
σ , (3.19)
∗ ∗{A }
is a Hilbert basis. We simply write Hn,m (σ; ·) instead of Hn,m n (σ; ·) if no confusion
is likely to arise.
Consider the Beltrami operator Δ∗;σ on the sphere Ωσ . We know that

1 ∗ 1
Δ∗;σ Yn,m = 2
Δ Yn,m = − 2 n(n + 1)Yn,m
σ σ
for n ∈ N0 ; m = 1, . . . , 2n + 1 (note that Δ∗;1 = Δ∗ ). Thus we formally have
s/2 s
1 n + 12
−Δ∗;σ + 2 Yn,k = Yn,m
4σ σ
and
s/2 ∧ s
∗;σ 1 n + 12
−Δ + 2 F (n, m) = F ∧ (n, m)
4σ σ
for all n ∈ N0 ; m = 1, . . . , 2n + 1.
Definition 3.2. For any given value s ∈ R, the Sobolev space Hs (Ωext
σ ) is the
completion of E(Ωext
σ ) under the norm · Hs (Ωext ) :
σ
·H
s (Ωext
σ )
Hs (Ωext ext
σ ) = E(Ωσ ) .
Hs (Ωext
σ ) equipped with the inner product (·, ·)Hs (Ωext is a Hilbert space. The
σ )
system {Hn,m
s
(σ; ·)} given by
s
σ
s
Hn,m (σ; x) = Hn,m (σ; x), x ∈ Ωext
σ , (3.20)
n + 12
is a Hilbert basis.
Hence, the norm in Hs (Ωext
σ ) reads as follows:
⎛ ⎞1/2
s/2 2
1
F Hs (Ωext ) = ⎝ −Δ∗;σ
x + 2
F (x) dω(x)⎠ . (3.21)
σ
Ωσ 4σ
ext
H0 (Ωext
σ ) may be understood as the space of all harmonic functions in Ωσ , regular
2
at infinity, corresponding to L -restrictions (note that the potentials in H0 (Ωext σ )
are uniquely determined by their L2 -(Dirichlet) boundary conditions on Ωσ ). Ac-
cording to our construction, Pot (∞) (Ωext ext
σ ) is a dense subspace of Hs (Ωσ ) for each
ext ext
s. If t < s, then F Ht (Ωext ) ≤ F Hs (Ωext ) and Hs (Ωσ ) ⊂ Ht (Ωσ ).
σ σ
If we associate to U the outer harmonic expansion (3.14) it is of fundamental
importance to know when the series (3.14) converges uniformly on the whole set
Ωext
σ . To this end we need the concept of summable sequences.
Definition 3.3. A sequence {An }n∈N0 ∈ A is called summable if

∞
2n + 1
< ∞. (3.22)
n=0
A2n
Lemma 3.4 (Sobolev Lemma). Assume that the sequences {An }

n∈N0 , {Bn }n∈N0 ∈

A are such that {Bn An }n∈N0 is summable. Then each F ∈ H {Bn−1 An }; Ωext
−1
σ
corresponds to a potential of class Pot (0) (Ωext

σ ).
The Sobolev Lemma which is proved in [19] states that in the case of summa-
bility of the sequence
{Bn−1 An }n∈N
0 , the Fourier series in terms of the basis func-
tions Hn,m ∈ H {Bn−1 An }; Ωext
σ is continuous on the boundary Ωσ . In particular,
we have the following statement (cf. [19]).
Lemma 3.5. If U ∈ Hs (Ωext

σ ), s > k + 1, then U corresponds to a potential of class
Pot (k) (Ωext
σ ).
3.2. Vectorial outer harmonic and Sobolev theory

We now extend the theory of scalar outer harmonics and scalar Sobolev spaces
to the vectorial case. We use a system of vector spherical harmonics (cf. [21]) in
order to generate the set of vector outer harmonics in such a way, that the Laplace
equation is fulfilled componentwise.
(i)
Let {ỹn,m }i=1,2,3;n∈N0i ;m=1,...,2n+1 be a set of vector spherical harmonics sat-
isfying the condition of being a set of eigenfunctions of the Beltrami operator,
with
0, i = 1,
0i = (3.23)
1, i = 2, 3.
(see, e.g., [21, 32, 58], for a detailed introduction and profound discussion of these
vector spherical harmonics). In the nomenclature of [32], the vector outer harmon-
(i)
ics hn,m (σ; ·) of degree n and kind i are defined by
n+2
(1) 1 σ (1) x
hn,m (σ; x) = ỹn,m , n = 0, 1, . . . ; m = 1, . . . , 2n + 1, (3.24)
σ |x| |x|
n
1 σ x
h(2) (σ; x) = ỹ (2)
, n = 1, 2, . . . ; m = 1, . . . , 2n + 1, (3.25)
n,m
σ |x| n,m
|x|
n+1
(3) 1 σ (3) x
hn,m (σ; x) = ỹn,m , n = 1, 2, . . . ; m = 1, . . . , 2n + 1, (3.26)
σ |x| |x|
for x ∈ Ωext
σ . The following properties are satisfied:
(i)
• hn,m (σ; ·) is of class c(∞) (Ωext σ ),
(i) (i)
• Δx hn,m (σ; x) = 0 for x ∈ Ωext σ , i.e., every component function hn,m · ε satis-
k
fies the Laplace equation,

(i) (i)
• hn,m is regular at infinity, i.e., |hn,m (σ; x)| = O(|x|−1 ),
(2)
|hn,m (σ · x)| = O(|x|−2 ), |x| → ∞
(i) (i)
• hn,m (σ; ·)|Ωσ = (1/σ)ỹn,m ,
(i) (j) (i) (j)
• (hn,m (σ; ·), hl,s (σ; ·))l2 (Ωσ ) = Ωσ hn,m (σ; x)hl,s (σ; x) dω(x) = δi,j δn,l δm,s .
We introduce
·c(0) (Ωext )
(i)
harm (i) (Ωext
σ ) = span n=0i ,...; hn,m (σ; ·) σ
, (3.27)
m=1,...,2n+1
·c(0) (Ωext )
(i)
harm(Ωext
σ ) = spani=1,2,3;n=0i ,...; hn,m (σ; ·)
σ
. (3.28)
m=1,...,2n+1
Some results concerning addition theorems for outer harmonics using Legendre
tensors and Legendre vectors can be found in the Ph.D.-thesis [58] and are not
discussed here.
Lemma 3.6. Let {Hn,m (σ; ·)}n∈N0 ;m=1,...,2n+1 be a system of scalar outer harmon-
ics. Then
·l2 (Σ)
span{Hn,m (σ; ·)εi |Σ }i=1,2,3 = l2 (Σ),
·c(0) (Σ)
span{Hn,m (σ; ·)εi |Σ }i=1,2,3 = c(0) (Σ).
(i)
Theorem 3.7. Let {hn,m (σ; ·)}i=1,2,3;n=0i ,...; be a system of vector outer harmonics
m=1,...,2n+1
as defined in (3.24)–(3.26). Then the following statements hold true:
·l2 (Σ)
(i)
l2 (Σ) = span (hn,m (σ; ·))|Σ ,
i=1,2,3;n=0i ,...;
m=1,...,2n+1
·c(0) (Σ)
(i)
c(0) (Σ) = span (hn,m (σ; ·))|Σ .
i=1,2,3;n=0i ,...;
m=1,...,2n+1
In order to define the vectorial potential space pot (Σext ) we need the diver-
gence and curl operator, which are defined by
3
3

∂Fi
div f (x) = (x), f= Fi εi , (3.29)
i=1
∂xi i=1
and
3
∂Fk
(curl f (x))i = εijk (x), (3.30)
∂xj
j,k=1
where εijk is the alternator defined by

⎧
⎪
⎨ +1, (i, j, k) is an even permutation of (1, 2, 3),
εijk = −1, (i, j, k) is an odd permutation of (1, 2, 3), (3.31)
⎪
⎩ 0, (i, j, k) is not a permutation of (1, 2, 3).
By pot (Σext ) we denote the space of all vector fields f : Σext → R3 satisfying
(i) f ∈ c(1) (Σext ),
(ii) f is a harmonic vector field: divf = 0, curlf = 0 in Σext ,
(iii) f is regular at infinity: |f (x)| = O(|x|−2 ), |x| → ∞.
Furthermore, we let
pot (k) (Σext ) = pot (Σext ) ∩ c(k) (Σext ), (3.32)
which is meant in the same sense as we explained in the scalar case. It is well
known (see, e.g., [38]), that every function f ∈ c(k) (Σext ) satisfying curlf = 0
is the gradient of a function V ∈ C (k+1) (Σext ): f = ∇V . As a consequence, we
get that every f ∈ pot (Σext ) can be represented as a gradient field f = ∇V ,
where V ∈ Pot (Σext ), and vice versa. Furthermore, it is obvious, that a function
3
f ∈ pot (Σext ) of the form f = i=1 Fi εi fulfills Fi ∈ Pot (Σext ).
For arbitrary ε > 0, we have an integer N = N (ε) and coefficients an,m ,
n = 0, . . . , N ; m = 1, . . . , 2n + 1, such that

N 2n+1

sup F (x) − an,m Hn,m (σ; x) < ε. (3.33)
x∈Σ
n=0 m=1
For the gradient of Hn,m (σ; ·) we obtain

∇x Hn,m (σ; x) = C h(1)
n,m (σ; x), (3.34)
with a constant factor C, which leads us to (cf. [25])
·c(0) (Σext )
(1)
pot (0) (Σext ) = span n∈N0 ; (hn,m (σ; ·))|Σext (3.35)
m=1,...,2n+1
(Runge–Walsh approximation property).

In analogy to the scalar case, we define Sobolev spaces for vector fields. We do
not restrict our considerations to pot (∞) (Ωext
σ ) as a reference space for the definition
(1)
of vectorial Sobolev spaces, because in this case only the hn,m -part would be taken
into account.
Consider the space a defined by
T
a = {{an } | an = A(1)
n , A(2)
n , A(3)
n ∈ R3 , A(i)
n = 0, n ∈ N0 }. (3.36)
(i)
Obviously, we have {An }n∈N0 ∈ A for i ∈ {1, 2, 3}.
For {an }n∈N0 ∈ a we define
∞ 2n+1

(i)
e(i) (Ωext
σ ) = f ∈ harm (Ω ext ) :
σ |A(i) 2
n | (f, h (i) 2
) 2
n,m l (Ωσ ) < ∞ , (3.37)
n=0i m=1
i ∈ {1, 2, 3}. Equipped with the inner product

∞ 2n+1
3
(f, g)h(Ωext ) = |A(i) 2 (i) (i)
n | (f, hn,m )l2 (Ωσ ) (g, hn,m )l2 (Ωσ ) , (3.38)
σ
i=1 n=0i m=1
f, g ∈ e(i) (Ωext (i) ext

σ ), the space e (Ωσ ) becomes a pre-Hilbert space. We define the
(i) (i) (i) (i)
Sobolev space h (Ωσ ) = h ({An }; Ωext
ext ext
σ ) to be the completion of e (Ωσ )
under the norm · h(Ωext ) , which denotes the norm associated to (·, ·)h(Ωext ) :
σ σ
·h(Ωext )
h(i) ({A(i) ext (i) ext
n }; Ωσ ) = e (Ωσ )
σ
. (3.39)
We use the following notation
3
5 3
5
h(Ωext ext
σ ) = h({an }; Ωσ ) = h(i) (Ωext
σ ) = h(i) ({A(i) ext
n }; Ωσ ). (3.40)
i=1 i=1
The space h(Ωext

σ ) equipped with the inner product (·, ·)h(Ωext ) is a Hilbert space σ
(i)∗{A(i)
n }
with Hilbert basis {hn,m (σ; ·)}i=1,2,3; n=0i ,...; m=1,...,2n+1 given by
(i)∗{A(i)
n }
hn,m (σ; x) = (A(i)
n )
−1 (i)
hn,m (σ; x), x ∈ Ωext
σ . (3.41)
We can, therefore, expand a function f ∈ h(Ωext
σ ) as a Fourier series in terms of
(i)∗{A(i)
n }
the basis hn,m :
∞ 2n+1
3 (i)∧h({a
n };Ωext
(i)∗{A(i)
n }
f= f σ ) (n, m)hn,m , (3.42)
i=1 n=0i m=1
where
(i)∧h({a (i)∗{A(i)
n }
f n };Ωext
σ ) (n, m) = f (i)∧ (n, m) = (f, hn,m )h(Ωext ) . (3.43)
σ
In analogy to the scalar spaces Hs (Ωext

σ ), we define the vectorial spaces
hs (Ωext
σ ) by

1 s
(i) (i) n + 2
hs (Ωext
σ )= h ; Ωext
σ ) , (3.44)
σ
3
5
hs (Ωext
σ )= h(i) ext
s (Ωσ ). (3.45)
i=1
The space hs (Ωext

σ ) equipped with the inner product (·, ·)hs (Ωext ) is a Hilbert space σ
(i)s
with Hilbert basis {hn,m (σ; ·)}i=1,2,3; n=0i ,...; m=1,...,2n+1 given by
s
(i)s σ
hn,m (σ; x) = h(i)
n,m (σ; x), x ∈ Ωext
σ . (3.46)
n + 12
In the case of the space h0 (Ωext
σ ) we understand the norm · h0 (Ωext to be the
σ )
· l (Ωσ ) - norm.
2
Next, the scalar Sobolev Lemma 3.4 will be extended to vector fields.
Definition 3.8. A sequence {an }n∈N0 ∈ a is called summable if
∞
2n + 1
2 < ∞, (3.47)
(i)
n=0i An
for i = 1, 2, 3.
In the sequel, {b−1

n }n∈N0 ∈ a means the sequence given by
−1 (2) −1 (3) −1 T
b−1
n = Bn(1) , Bn , Bn , (3.48)
and
(1) −1 (2) (2) −1 (3) (3) −1 T
b−1 (1)
n an = An Bn , An Bn , An Bn . (3.49)
Lemma 3.9 (Vectorial Sobolev Lemma). Assume, that {an }n∈N0 , {bn }n∈N0 ∈ a are
sequences such that {b−1 −1 ext
n an }n∈N0 ∈ a is summable. Then each f ∈ h({bn an }; Ωσ )
ext
corresponds to a function of class harm(Ωσ ).
3.3. Tensorial outer harmonic and Sobolev theory

The extension of vectorial to tensorial theory is straightforward (see [21, 32, 58]).
(i,k)
With the help of a system {ỹn,m } of tensor spherical harmonics we can derive a
(i,k)
set of tensor outer harmonics {hn,m (σ; ·)} satisfying the Laplace equation compo-
nentwise.
(i,k)
Let {ỹn,m }i,k=1,2,3;n∈N0 ;m=1,...,2n+1 with
⎧
⎪
⎨ 0, (i, k) ∈ {(1, 1), (2, 1), (3, 1)},
0ik = 1, (i, k) ∈ {(1, 2), (1, 3), (2, 3), (3, 3)}, (3.50)
⎪
⎩ 2, (i, k) ∈ {(2, 2), (3, 2)},
be a set of tensorial spherical harmonics satisfying the condition of being eigen-

functions of the Beltrami operator (see, e.g., the Ph.D.-thesis [58] for a detailed
introduction and profound discussion of these tensor spherical harmonics). The
(i,k)
tensor outer harmonics hn,m (σ; ·) of degree n and kind (i, k) are then defined by
n+3
(1,1) 1 σ (1,1) x
hn,m (σ; x) = ỹn,m , (3.51)
σ |x| |x|
n+1
(1,2) 1 σ (1,2) x
hn,m (σ; x) = ỹn,m , (3.52)
σ |x| |x|
n+1
(2,1) 1 σ (2,1) x
hn,m (σ; x) = ỹn,m , (3.53)
σ |x| |x|
n−1
(2,2) 1 σ (2,2) x
hn,m (R; x) = ỹn,m , (3.54)
σ |x| |x|
n+1
(3,3) 1 σ (3,3) x
hn,m (σ; x) = ỹn,m , (3.55)
σ |x| |x|
n+2
(1,3) 1 σ (1,3) x
hn,m (σ; x) = ỹn,m , (3.56)
σ |x| |x|
n
(2,3) 1 σ (2,3) x
hn,m (σ; x) = ỹun,m , (3.57)
σ |x| |x|
n+2
(3,1) 1 σ (3,1) x
hn,m (σ; x) = ỹn,m , (3.58)
σ |x| |x|
n
(3,2) 1 σ (3,2) x
hn,m (σ; x) = ỹn,m , (3.59)
σ |x| |x|
where x ∈ Ωext
σ , n = 0ik , . . . ; m = 1, . . . , 2n + 1. The following properties are
satisfied:
(i,k)
• hn,m (σ; ·) is of class c(∞) (Ωextσ ),
(i,k) (i,k)
• Δx hn,m (σ; x) = 0 for x ∈ Ωext σ , i.e., the component functions of hn,m (σ; ·)
fulfill the Laplace equation,
(i,k) (i,k)
• hn,m is regular at infinity, i.e., |hn,m (σ; x)| = O(|x|−3 ), |x| → ∞.
(i,k) (i,k)
• hn,m (σ; ·)|Ωσ = (1/σ)ỹn,m ,
(i,k) (p,q) (i,k)
• (hn,m (σ; ·), hl,s (σ; ·))l2 (Ωσ ) = Ωσ hn,m (σ; x)hp,q
l,s (σ; x)dω(x)
= δi,p δk,q δn,l δm,s .
Moreover, we define
·c(0) (Ωext )
(i,k)
harm(i,k) (Ωext
σ ) = span n=0ik ...; hn,m (σ; ·) σ
, (3.60)
m=1,...,2n+1
·c(0) (Ωext )
(i,k)
harm(Ωext
σ ) = spani,k∈{1,2,3};n=0ik ...; hn,m (σ; ·)
σ
. (3.61)
m=1,...,2n+1
Some results concerning addition theorems for outer harmonics can be for-
mulated both for the tensor product of two tensor outer harmonics and for the
product of a scalar and a tensor outer harmonic. They can be found in the Ph.D.-
thesis [58] and are not discussed in this contribution.
Lemma 3.10. Let {Hn,m (σ; ·)}n∈N0ik ;m=1,...,2n+1 be a system of scalar outer har-
monics. Then
·l2 (Σ)
span{Hn,m (σ; ·)εi ⊗ εk |Σ } = l2 (Σ), (3.62)
·c(0) (Σ)
span{Hn,m (σ; ·)εi ⊗ εk |Σ) } = c(0) (Σ). (3.63)
(i,k)
Theorem 3.11. Let {hn,m }i,k=1,2,3;n=0ik ,...; be a system of tensor outer harmonics.
m=1,...,2n+1
Then the following statements hold true:
·l2 (Σ)
(i,k)
l2 (Σ) = span (hn,m (σ; ·))|Σ , (3.64)
i,k=1,2,3;n=0ik ,...;
m=1,...,2n+1
·c(Σ)
(i,k)
c(Σ) = span (hn,m (σ; ·))|Σ . (3.65)
i,k=1,2,3;n=0ik ,...;
m=1,...,2n+1
In order to define a tensorial counterpart pot(Σext ) of the space pot (Σext ),

we need the divergence and the curl operator of tensor fields. Having (3.29) in
mind, we define div f by

3
3

∂Fi,j
(div f (x))i = (x), f= Fi,j εi ⊗ εj . (3.66)
j=1
∂xj i,j=1
Furthermore, based on (3.30) we have the following definition of curl f :

3
∂Fj,k
(curl f (x))i,j = εipk (x). (3.67)
∂xp
p,k=1
The space pot(Σext ) denotes the space of all tensor fields f : Σext → R3 ⊗ R3
satisfying the following properties:
(i) f ∈ c(1) (Σext ),
(ii) f is a harmonic tensor field: div f = 0, curl f = 0 in Σext ,
(iii) f is regular at infinity: |f (x)| = O(|x|−3 ), |x| → ∞.
Furthermore, we let
pot(k) (Σext ) = pot(Σext ) ∩ c(k) (Σext ), (3.68)
which we understand in the same sense as in the scalar and vectorial case. As
shown, e.g., in [38], every tensor function f ∈ c(k) (Σext ) with curl f = 0 is the
gradient of a vector field v ∈ c(k+1) (Σext ):
f = ∇v, (3.69)
where ∇v is the tensor of second rank defined by
∂vi
(∇x v)ij (x) = (x). (3.70)
∂xj
Therefore, every member v ∈ pot(Σext ) can be represented as a gradient field
v = ∇v, where v is of class pot (Σext ), and vice versa. As a consequence of this, in
connection with the fact that every v ∈ pot (Σext ) can be represented as a gradient
field v = ∇V with V ∈ Pot (Σext ), we finally get that a tensor field v ∈ pot(Σext )
can be represented as the Hesse tensor of a scalar field V ∈ Pot (Σext ):
v = ∇ ⊗ ∇V, (3.71)
and vice versa. 3
It is obvious, that f ∈ pot(Σext ) of the form f = i,k=1 Fi,k εi ⊗ εk fulfills
Fi,k ∈ Pot (Σext ). In addition, we are able to show that
·c(0) (Σext )
(1,1)
pot(0) (Σext ) = span n∈N0 ; (hn,m (σ; ·))|Σext (3.72)
m=1,...,2n+1
(Runge–Walsh approximation property).

Our purpose is now to define Sobolev spaces for tensor fields in analogy to
the vectorial Sobolev spaces. We introduce the linear space a in the following way:
a = {{an } | an ∈ R3 ⊗ R3 , A(i,k)
n = 0, n ∈ N0 ; m = 1, . . . , 2n + 1; i, k ∈ {1, 2, 3}},
(3.73)
where ⎛ ⎞
(1,1) (1,2) (1,3)
An An An
⎜ (2,3) ⎟
an = ⎝ A(2,1)
n
(2,2)
An An ⎠, (3.74)
(3,1) (3,2) (3,3)
An An An
(i,k)
with {An }n∈N0 ∈ A for i, k ∈ {1, 2, 3}.
Let us now consider a sequence {an }n∈N0 ∈ a. Then we define
∞ 2n+1

(i,k)
e(i,k) (Ωext
σ ) = f ∈ harm (Ω ext ) :
σ |A(i,k) 2
n | (f , h (i,k) 2
) 2
n,m l (Ωσ ) < ∞ ,
n=0ik m=1
(3.75)
i, k ∈ {1, 2, 3}. Equipped with the inner product
3
∞ 2n+1

(f , g)h(Ωext ) = |A(i,k)
n |2 (f , h(i,k) (i,k)
n,m )l2 (Ωσ ) (g, hn,m )l2 (Ωσ ) , (3.76)
σ
i,k=1 n=0ik m=1
f , g ∈ e(i,k) (Ωextσ ), the space e

(i,k)
(Ωext
σ ) becomes a pre-Hilbert space. We de-
(i,k) (i,k)
fine the Sobolev space h (Ωσ ) = h(i,k) ({An }; Ωext
ext
σ ) to be the completion
(i,k)
of e (Ωext ) under the norm · h(Ωext , which denotes the norm associated to
σ )
σ
(·, ·)h(Ωext ) :
σ
·h({Ωext )
h(i,k) ({A(i,k)
n }; Ωext
σ )= e
(i,k) (Ωext )
σ
σ
. (3.77)
We use the following notation
3
5
h(Ωext
σ ) = h(i,k) (Ωext
σ ). (3.78)
i,k=1
The space h(Ωext

σ ) equipped with the inner product (·, ·)h({Ωext ) is a Hilbert space. σ
(i,k)∗{A(i,k) }
The system {hn,m n
(σ; ·)}i,k∈{1,2,3};n∈N0ik ; , given by
m=1,...,2n+1
(i,k)∗{A(i,k) }
hn,m n
(σ; x) = (A(i,k)
n )−1 h(i,k)
σ , (3.79)
represents an h(Ωext ext

σ )-orthonormal Hilbert basis in h(Ωσ ).
As a consequence, we can expand a function f ∈ h(Ωext
σ ) as a Fourier series
(i,k)∗{A(i,k)
n }
in terms of the basis hn,m :
3
∞ 2n+1
(i,k)∧h({a
n };Ωext
(i,k)∗{A(i,k)
n }
f= f σ ) (n, m)hn,m , (3.80)
i,k=1 n=0ik m=1
where
(i,k)∧h({a (i,k)∗{A(i,k) }
f n };Ωext
σ ) (n, m) = f (i,k)∧ (n, m) = (f , hn,m n
)h(Ωext . (3.81)
σ
(i)
Finally, in analogy to the vectorial spaces hs (Ωext
σ ), we define
s
(i,k) ext (i,k) n + 12 ext
hs (Ωσ ) = h ; Ωσ , (3.82)
σ
3
5
hs (Ωext
σ )= h(i,k)
s (Ωext
σ ). (3.83)
i,k=1
The space hs (Ωext

σ ) equipped with the inner product (·, ·)hs (Ωext ) is a Hilbert
σ
(i,k)s
space. The system {hn,m (σ; ·)}i,k∈{1,2,3}; n∈N0ik ; m=1,...,2n+1 , given by
s
σ
h(i,k)s
n,m (σ; x) = 1 h(i,k)
σ , (3.84)
n+ 2

σ )-orthonormal Hilbert basis in h(Ωσ ).
Our next goal is to extend the Sobolev Lemma 3.4 to tensor fields.
Definition 3.12. A sequence {an }n∈N0 ∈ a is called summable if

∞
2n + 1
2 < ∞ (3.85)
(i,k)
n=0ik An
for i, k ∈ {1, 2, 3}.
In the sequel, {b−1

n }n∈N0 ∈ a represents the sequence given by
⎛ −1 −1 −1 ⎞
(1,1) (1,2) (1,3)
⎜ B B B
−1 ⎟
n n n
⎜ −1 −1 ⎟
−1 ⎜
bn = ⎜ B n (2,1)
Bn
(2,2)
Bn
(2,3) ⎟, (3.86)
⎟
⎝ −1 −1 −1 ⎠
(3,1) (3,2) (3,3)
Bn Bn Bn
and {a−1
n bn }n∈N0 ∈ a is given by
⎛ −1 −1 −1 ⎞
(1,1) (1,1) (1,2) (1,2) (1,3) (1,3)
⎜ An B n An B n An Bn ⎟
⎜ (2,1) (2,1) −1 −1 −1 ⎟
b−1 a = ⎜ An Bn
(2,2)
An Bn
(2,2) (2,3)
An
(2,3)
Bn ⎟ . (3.87)
n n ⎜ ⎟
⎝ −1 −1 −1 ⎠
(3,1) (3,1) (3,2) (3,2) (3,3) (3,3)
An Bn An Bn An Bn
Lemma 3.13 (Tensorial Sobolev Lemma). Assume, that the sequences {an }n∈N0 ,
−1
{bn }n∈N0 ∈ a are such that {bn an }n∈N0 ∈ a is summable. Then each f ∈
h {b−1 ext corresponds to a function of class harm(Ωext ).
n an }; Ωσ σ
4. Pseudodifferential operators and geodetic nomenclature

All gravitational information under discussion in physical geodesy leads to operator
equations relating the (disturbing) potential to geodetically relevant observables.
In physical geodesy, one can think of observables as operating on an “input signal”
F (e.g., the (disturbing) potential) to produce an (scalar, vectorial or tensorial)
output signal of the form
ΛF = G (4.1)
(for example, geoidal undulation, gravity anomaly, radial or tangential deriva-
tives), where Λ is a certain (scalar, vectorial or tensorial) operator. Note, that
later on we will differentiate in our notation weather we deal with scalar, vecto-
rial or tensorial observables, but in this introductory part of the text for reason of
readability we do not distinguish the geodetic quantities. Fortunately, it is the case
in geodetic applications involving the (disturbing) potential that large portions of
interest can be well approximated by operators that represent linear, rotation-
invariant pseudodifferential operators.
The standard pseudodifferential operators Λ occurring in physical geodesy
(cf. [69]) have to reflect the Pizzetti concept (cf. [36, 59]):
1. The mass within the reference ellipsoid for establishing the disturbing poten-
tial F is equal to the mass of the Earth.
2. The center of the reference ellipsoid coincides with the center of the Earth.
3. The value of the potential on the geoidal surface and the value of the normal
potential on the reference ellipsoidal surface are the same.
4. There are no masses outside the geoid (remove-restore-principle from masses
outside the geoid).
5. The constructive approximation is simplified for reasons of computational
economy from an ellipsoidal to a spherical framework by Runge–Walsh jus-
tification (see the contribution [4] in this volume).
The presentation of the classical quantities in gravitational potential determi-
nation can be formulated within the framework of pseudodifferential operators. To
be more concrete, in our approach we deal with radial, tangential and mixed (first-
and second-order) derivatives of the Earth gravitational potential. Two important
properties have to be taken into account specifying the operators which we study
in the sequel. On the one hand, the mathematical modeling should lead to a consis-
tent setup. It turns out that this requirement is, in fact, assured by the operators.
On the other hand, we demand the assigned operators to be isotropic for structural
∂
reasons. In consequence (see also [63]), the (scalar) tangential derivatives ∂ϕ and
∂
∂t are of no interest for us because they do not lead to isotropic operators in a
scalar framework. Instead of using scalar tangential operators we decide to go over
to the vectorial (and tensorial) tangential derivative using the surface gradient ∇∗ .
Indeed, we want to point out that we have the choice between two viable variants
namely either to develop a scalar anisotropic theory for component modeling, or to
turn over to vectorial/tensorial isotropic theory. In this contribution, we prefer the
second variant, expecting that the development of a vector/tensor theory provides

us with a versatile tool for modeling geodetically relevant vector and tensor fields
and solving the SST and SGG problem in a simply structured isotropic framework.
The observables we discuss are presented in Tables 1, 2 and 3.
Quantity Operator Symbol Order

n−1
gravity anomaly ΛA σ 1
geoid undulations ΛU σ2 0
Stokes operator ΛSt σ
n−1 −1
first radial derivative Λ ∂ − n+1
σ 1
∂r
(n+1)(n+2)
second radial derivative Λ ∂2 σ2 2
∂r2
n
upward continuation ΛUP C σ
γ −∞
n
scalar SST ΛSST σ
γ
n+1
γ −∞
n
(n+1)(n+2)
scalar SGG ΛSGG σ
γ γ2 −∞
Table 1. Scalar geodetic observables leading to isotropic pseudodiffer-

ential operators (note that the symbol is given with respect to Hn,m ).
4.1. Scalar theory

We start with the scalar definition and give some examples.
Definition 4.1. Let Hs (Ωext ext
τ ) and Hs (Ωρ ) be Sobolev spaces, τ, ρ > 0. Further-
∧
more, let {Λ (n)}n∈N0 be a sequence of real numbers. The operator Λ : Hs (Ωext
τ ) →
Hs (Ωext
ρ ) defined by
∞ 2n+1

ΛF = Λ∧ (n)F ∧ (n, m)Hn,m
s
(ρ; ·) (4.2)
n=0 m=1
is called a scalar pseudodifferential operator of order t, if
|Λ∧ (n)|
lim
n→∞ n + 1 t
= const = 0 (4.3)
2
for some t ∈ R. The sequence {Λ∧ (n)}n∈N0 is called the symbol of Λ. Moreover, if
the limit relation
|Λ∧ (n)|
lim =0 (4.4)
n→∞ n + 1 t
2
holds for all t ∈ R, then the operator is called a pseudodifferential operator of order
−∞.
Quantity Operator Symbol Order

first tangential derivative ∇∗ n n+1
, 1
σ
2n+1
n+1 n
σ 2n+1 ,
0

second mixed derivative ∇∗ ∂V
∂r ,
n(n+1)
σ2
n+1
, 2
2
2n+1
(n+1) n
σ2 2n+1 ,
0
n
vectorial SST λSST σ n n+1
, −∞,
γn γ
2n+1
σ n+1 n
γ γ 2n+1 ,
0
n
n(n+1)
vectorial SGG λSGG σ
2
n+1
−∞
n γ 2 2n+1
γ
σ (n+1) n
γ γ 2n+1 ,
0
Table 2. Vectorial geodetic observables leading to isotropic pseudodif-

(i)
ferential operators (note that the symbol is given with respect to hn,m ,
i = 1, 2, 3 from top to down for each operator).
Note that the convergence of the series in (4.2) is understood in Hs (Ωext

ρ )-
topology. As an immediate consequence (cf. [69]), we have the important relation
s
ΛHn,m (τ ; ·) = Λ∧ (n)Hn,m
s
(ρ; ·). (4.5)
In other words, we have the requirement that the outer harmonics are the eigen-
functions of the operator Λ, and the invertibility has to be controlled by the in-
vertibility of the values Λ∧ (n), n ∈ N0 . The symbol has many appealing properties
(cf. [69]): It is easily seen that
(Λ
+ Λ
)∧ (n) = (Λ
)∧ (n) + (Λ
)∧ (n), (4.6)

∧
∧
∧
(Λ Λ ) (n) = (Λ ) (n)(Λ ) (n), (4.7)
for all n ∈ N0 .
As any “output function” (output signal) can be expanded into an orthogonal
series of outer harmonics
∞ 2n+1
∞ 2n+1

G = ΛF = Λ∧ (n)F ∧ (n, m)Hn,m
s
(ρ; ·) = G∧ (n, m)Hn,m
s
(ρ; ·)
n=0 m=1 n=0 m=1
(4.8)
Swcpvkv{ Qrgtcvqt U{odqn Qtfgt
o)o,3*
q
ugeqpf s s 4 )4o,3*)4o,5* )o , 4*)o , 3*)4o , 3*)4o , 5*- 4
vcpigpvkcn )o,3*)o3* q 4
fgtkxcvkxg 4 )4o3*)4o,3* 5o -
2-
o)o,4*
q
4 )4o,5*)4o,3* )o , 3* )4o , 3*)4o , 5*-
o)o,3*)o,4*
q
4 )4o3*)4o,3* o)o 3*)4o 3*)4o , 3*-
2- hqt )j< l* 3 g)4< 5*< )5< 3*< )5< 4*< )5< 5*h
o q
o)o,3*
vgpuqtkcn THH 4 )4o,3*)4o,5* )o , 4*)o , 3*)4o , 3*)4o , 5*- 2
UII o
)o,3*)o3* q
4
4 )4o3*)4o,3* 5o -
2-
o q
o)o,4*
4 )4o,5*)4o,3* )o , 3* )4o , 3*)4o , 5*-
o q
o)o,3*)o,4*
4 )4o3*)4o,3* o)o 3*)4o 3*)4o , 3*-
2- hqt )j< l* 3 g)4< 5*< )5< 3*< )5< 4*< )5< 5*h
Table 3. Tensorial geodetic observables leading to isotropic pseudodif-

(i,k)
ferential operators (note that the symbol is given with respect to hn,m ,
i, k = 1, 2, 3, from top to down ((1, 1), (1, 2), . . . , (3, 2), (3, 3)) for each
operator).
in the sense of · Hs(Ωext ) , we are confronted with a spectral representation of the

ρ
form
G∧ (n, m) = (ΛF )∧ (n, m) = Λ∧ (n) F ∧ (n, m), n ∈ N0 , k = 1, . . . , 2n + 1. (4.9)
∧
This means that the “amplitude spectrum” {G (n, m)} of the response of Λ is
described in terms of the amplitude spectrum of functions (signals) F by a simple
multiplication by the “transfer” Λ∧ (n).
The following list contains (scalar) pseudodifferential operators which are of
importance for geodetic applications.
Consider a potential F of the class Hs (Ωext
σ ), that is
∞ 2n+1

F = F ∧ (n, m)Hn,m
s
(σ; ·), (4.10)
n=0 m=1
where we use the geometric concept as explained in Section 3 and shown in Fig-
ure 3.1.
(i) Gravity Anomalies. The problem of determining the disturbing potential U
with Λ(U ) = F from prescribed gravity anomalies F is the “fundamental
problem of classical physical geodesy” (see, e.g., [37, 43, 53, 69]). The operator
related to gravity anomalies Λ : Hs (Ωext ext
σ ) → Hs (Ωσ ) has the symbol
n−1
Λ∧ (n) = . (4.11)
σ
(ii) Geoid Undulations. The operator related to geoid undulations Λ : Hs (Ωext σ ) →
Hs (Ωext
σ ) has the symbol
Λ∧ (n) = σ 2 . (4.12)
(iii) Stokes Operator . This operator is defined by

σ
Λ(F )(x) = St(x, y)F (y), dω(y), x ∈ Ωσ (4.13)
4π Ωσ
where St(·, ·) is the Stokes kernel (cf. [32, 68, 69]).
The Stokes operator Λ : Hs (Ωext ext
σ ) → Hs (Ωσ ) has the symbol

0, for n = 1
(Λ)∧ (n) = σ (4.14)
n−1 , for n = 0, 2, 3, 4, . . . .
(iv) Upward Continuation Operator . The upward continuation operator associates
(0)
to F ∈ Hs (Ωext σ ) the solution ΛF of the Dirichlet problem ΛF ∈ Pot (Ωext
γ )
corresponding to the boundary values (ΛF )|Ωγ = F |Ωγ . The upward contin-
uation operator Λ : Hs (Ωext ext
σ ) → Hs (Ωγ ) has the symbol
n
σ
Λ∧ (n) = , n ∈ N0 . (4.15)
γ
The upward continuation operator indeed plays an important role in the
mathematical treatment of spaceborne problems, since it relates potential
values at height σ to potential values at height γ(> σ).
(v) Operator of the (Negative) First-order Radial Derivative on Ωσ . This operator
associates to F ∈ Hs (Ωext σ ) the solution ΛF of the Dirichlet problem ΛF ∈
Pot (0) (Ωext
σ ) corresponding to the boundary values (ΛF )|Ωσ = − ∂r ∂
F |Ωσ . Λ
∧
is a pseudodifferential operator of order 1 with symbol {Λ (n)}n∈N0 given by
n+1
Λ∧ (n) = , n ∈ N0 . (4.16)
σ
In fact, Λ is the “harmonic continuation” of the radial derivative on Ωσ into
the outer space Ωext σ and is important in case of the SST problem.
(vi) Operator of the Second-order Radial Derivative on Ωσ . This operator as-
sociates to F ∈ Hs (Ωext σ ) the solution ΛF of the Dirichlet problem ΛF ∈
∂2
Pot (0) (Ωext
σ ) corresponding to the boundary values (ΛF )|Ωσ = ∂r 2 F |Ωσ . Λ is
∧
a pseudodifferential operator of order 2 with symbol {Λ (n)}n∈N0 given by
(n + 1)(n + 2)
Λ∧ (n) = , n ∈ N0 . (4.17)
σ2
Λ is the “harmonic continuation” of the second radial derivative on Ωσ into
the outer space Ωext
σ and is important in case of the SGG problem.
4.2. Vectorial theory

We now introduce vectorial pseudodifferential operators and give two examples.
(i)
Definition 4.2. Let Hs (Ωext ext
τ ) be a scalar Sobolev space and hs (Ωρ ) a vectorial
(i)∧
Sobolev space, τ, ρ > 0, i ∈ {1, 2, 3}. Furthermore, let {λ (n)}n∈N0i be a se-
(i)
quence of real numbers for i = 1, 2, 3. The operator λ(i) : Hs (Ωext ext
τ ) → hs (Ωρ )
defined by
∞ 2n+1

λ(i) F = λ(i)∧ (n)F ∧ (n, m)h(i)s
n,m (ρ; ·) (4.18)
n=0i m=1
is called a vectorial pseudodifferential operator of kind i and order t, if

|λ(i)∧ (n)|
lim = const = 0 (4.19)
n→∞ (n + 1 )t
2
for some t ∈ R. Moreover, if the limit relation

|λ(i)∧ (n)|
lim =0 (4.20)
n→∞ (n + 1 )t
2
holds for all t ∈ R, then the operator λ(i) is called a vectorial pseudodifferential
operator of kind i and order −∞. The sequence {λ(i)∧ (n)} is called the symbol of
λ(i) . Further on, the operator λ : Hs (Ωext ext
τ ) → hs (Ωρ ) defined by
3

λ= λ(i) , (4.21)
i=1
is called a vectorial pseudodifferential operator of order t, where t = max3i=1 (order

of λ(i) ). Moreover, if the limit relation
|λ(i)∧ (n)|
lim =0 (4.22)
n→∞ (n + 1 )t
2
holds for all t ∈ R, and all i ∈ {1, 2, 3}, then the operator λ is called a vectorial
pseudodifferential operator of order −∞.
We now give two examples of vectorial pseudodifferential operators which are

important for geodetic applications. We use the surface gradient on the sphere Ωσ
defined by
1
∇∗;σ = ∇∗ . (4.23)
σ
(iv) The Operator of the First-order Tangential Derivatives on Ωσ . This operator
associates to F ∈ Hs (Ωext
σ ) the solution λF of the Dirichlet problem λF ∈
hs (Ωext
σ ) corresponding to the boundary value (λF )|Ωσ = ∇∗,σ F |Ωσ . λ is a
pseudodifferential operator of order 1 with symbol {λ(i)∧ (n)}n∈N0i given by

⎧
⎪
⎪ nσ 2n+1
n+1
, i = 1,
⎨
(i)∧
λ (n) = n+1 n (4.24)
⎪
⎪ σ 2n+1 , i = 2,
⎩
0, i = 3.
In fact, Λ is the “harmonic continuation” of the tangential derivative on Ωσ
into the outer space Ωext
σ and is important in case of the SST problem.
(v) The Operator of the (Negative) Second-order Mixed Derivatives on Ωσ . This
operator associates to F ∈ Hs (Ωext
σ ) the solution λF of the Dirichlet problem
ext
λF ∈ hs (Ωσ ) corresponding to the boundary values (λF )|Ωσ = − ∂r∂
∇∗,σ
ξ F |Ωσ .
λ is a pseudodifferential operator of second order with symbol {λ(i)∧ (n)}n∈N0i
given by
⎧
⎪
⎪
n(n+1) n+1
⎨ σ2 2n+1 , i = 1,
λ(i)∧ (n) = (n+1)2 n (4.25)
⎪
⎪ σ2 2n+1 , i = 2,
⎩
0, i = 3.
Λ is the “harmonic continuation” of the second-order mixed derivatives on
Ωσ into the outer space Ωext
4.3. Tensorial theory

The introduction of tensorial pseudodifferential operators is straightforward.
(i,k)
Definition 4.3. Let Hs (Ωext
τ ) be a scalar Sobolev space and hs (Ωext
ρ ) a ten-
sorial Sobolev space, τ, ρ > 0, i, k ∈ {1, 2, 3}. Furthermore, for i, k ∈ {1, 2, 3},
let λ(i,k)∧ (n)n∈N0ik be a sequence of real numbers. The operator λ(i,k)∧ (n) :
(i,k)
Hs (Ωext
τ ) → hs (Ωext
ρ ) defined by
∞ 2n+1

λ(i,k) F = λ(i,k)∧ (n)}F ∧ (n, m)h(i,k)s
n,m (ρ; ·) (4.26)
n=0ik m=1
is called a tensorial pseudodifferential operator of kind (i, k) and order t, if the

limit relation
| λ(i,k)∧ (n)|
lim = const = 0 (4.27)
n→∞ (n + 1 )t
2
is satisfied for some t ∈ R. Moreover, if the limit relation
| λ(i,k)∧ (n)|
lim =0 (4.28)
n→∞ (n + 12 )t
holds for all t ∈ R, then the operator λ is called a pseudodifferential operator
of kind (i, k) and order −∞. The sequence { λ(i,k)∧ (n)} is called the (spherical)
symbol of λ(i,k) . Further on, the operator λ : Hs (Ωext ext

τ ) → hs (Ωρ ) defined by
3
3
λ= λ(i,k) , (4.29)
i=1 k=1
is called a tensorial pseudodifferential operator of order t, where t = max3i,k=1

(order of λ(i,k) ). Moreover, if the limit relation
| λ(i,k)∧ (n)|
lim =0 (4.30)
n→∞ (n + 12 )t
holds for all t ∈ R, and all i, k ∈ {1, 2, 3}, then the operator λ is called a pseudo-
differential operator of order −∞.
Finally, we mention one important example.

(iv) The Operator of the Second-order Tangential Derivatives on Ωσ . This op-
erator associates to F ∈ Hs (Ωext σ ) the solution λF of the Dirichlet prob-
lem λF ∈ hs (Ωext σ ) corresponding to the boundary values ( λF )|Ωσ =
∇∗,σ ⊗∇∗,σ F |Ωσ . It is a pseudodifferential operator of order 2 with the symbol
{ λ(i,k)∧ (n)}n∈N0ik given by
λ(i,k)∧ (n) (4.31)

⎧ #
⎪ n(n+1)
⎪ σ2 (2n+1)(2n+3) √ (n + 2)(n + 1)(2n + 1)(2n + 3), (i, k) = (1, 1),
⎪
⎪
⎪
⎪ −(n+1)(n−1) 2
⎨ σ2 (2n−1)(2n+1) 3n , # (i, k) = (1, 2),
= −n(n+2)
⎪ σ2 (2n+3)(2n+1) (n
#
+ 1) (2n + 1)(2n + 3), (i, k) = (2, 1),
⎪
⎪
⎪
⎪
n(n+1)(n+2)
n(n − 1)(2n − 1)(2n + 1), (i, k) = (2, 2),
⎪ 2
⎩ σ (2n−1)(2n+1)
0, else.
Λ is the “harmonic continuation” of the second-order tangential derivatives

on Ωσ into the outer space Ωext
5. Reproducing kernel structure and observational functionals

Of great importance for our considerations are Sobolev spaces equipped with a
reproducing kernel structure. The importance of the reproducing kernel lies in the
fact that it determines the norm of the dual space. Furthermore, no computational
work must be done to evaluate inner products involving reproducing kernel expres-
sions. Within this section, we focus on scalar theory and essentially follow [19].
The extension to vectorial and tensorial reproducing kernel Sobolev spaces is not
hard to perform.
5.1. Reproducing Hilbert spaces

Theorem 5.1. Let the sequence {An } be summable in the sense of Definition 3.3.
Then H(Ωext ext
σ ) (more explicitly, H({An }; Ωσ )) is a Hilbert subspace of the space
(0) ext ext
Pot (Ωσ ). The space H(Ωσ ) has the reproducing kernel function
KH(Ωext ) (·, ·) : Ωext ext
σ × Ωσ → R
σ
given by
2n+1

∗{An } ∗{An }
KH(Ωext ) (x, y) = Hn,m (σ; x)Hn,m (σ; y),
σ
n∈N (An ) m=1
x, y ∈ Ωext
σ .
If H(Ωext
σ ) has a reproducing kernel, then the Fourier (orthogonal) expan-
∗
sion of a potential in terms of the Hilbert basis {Hn,k (σ; ·)} in H(Ωextσ ) converges
uniformly on the domain Ωextσ (cf. [3, 7]). To be more specific, the relation
- -
- 2n+1
-
- ∧ ∗{An } -
lim -F − F (n, m)Hn,m (σ; ·)- =0
N →∞- -
n∈N m=1 ext H(Ωσ )
n≤N
implies

2n+1

lim sup F (x) − F ∧ (n, m)Hn,m
∗
(σ; x) = 0.
N →∞
x∈Ωext

σ n∈N m=1
n≤N
The representer of a bounded linear functional L on H(Ωext σ ) has a sim-

ple expression. More explicitly, L(x) = LKH(Ωext ) (·, x), x ∈ Ωext
σ , is in H(Ωext
σ ),
σ
and for all F ∈ H(Ωext

σ ) we have LF = (F, L)H(Ωext (note that x is held fixed
σ )
and L is applied to KH(Ωext ) (·, x) as a function of the first variable). Obviously,
σ
∗
(L, L)H(Ωext ) = LLKH(Ωext ) (·, ·) = (L, L)H(Ωext )∗ . The dual space H(Ωext
σ ) of
σ σ σ
H(Ωext ext
σ ) (i.e., the space of all linear bounded functionals on H(Ωσ )) is a Hilbert
1
space with respect to · H(Ωext )∗ = (·, ·)H(Ω ; the spaces H(Ωext ext ∗
σ ) and H(Ωσ )
2
σ ext )∗
σ
are known as isomorphic and isometric (see, e.g., [7]).
Reproducing kernel representations may be used to act as basis system in
reproducing Sobolev spaces.
Theorem 5.2. Let {An } be summable in the sense of Definition 3.3. Assume that
X is a countable dense set of points on a regular surface Ξ ⊂ Ωext
σ (for example,
Runge sphere Ωσ , real Earth’s surface Σ). Then
·H(Ωext )
spanx∈X KH(Ωext ) (x, ·) σ = H(Ωext
σ ).
σ
Theorem 5.2 allows an obvious generalization by means of bounded linear

functionals on H(Ωext
σ ).
Theorem 5.3. Let {An } be summable. Assume that X is a countable dense set of
∗
linear functionals in H(Ωext
σ ) . Then
·H(Ωext )
spanL∈X LKH(Ωext ) (·, ·) σ = H(Ωext
σ ).
σ
The set of all finite linear combinations of outer harmonics is dense in the
space Pot (0) (Ωext ext
σ ) in the sense of · C (0) (Ωext ) . Hence, H(Ωσ ) is a dense subset
σ
of Pot (0) (Ωext

σ ), too. This leads us to the following corollary.
Corollary 5.4. Under the assumption of Theorem 5.3

·C (0) (Ωext )
spanL∈X LKH(Ωext ) (·, ·) σ = Pot (0) (Ωext
σ ).
σ
Next we come to the problem of specifying certain types of sequences {An }

such that H(Ωext ext
σ )(= H({An }; Ωσ )) is a reproducing kernel Hilbert space. We
restrict ourselves to those kernel functions which are usable later on in multiscale
approximation. Other types of kernel functions which are known from spline inter-
polation or smoothing procedures (see, for example, [14–16, 18, 20, 49, 55, 56, 72])
are not discussed here.
Our list of (reproducing) kernel functions is divided into two parts, namely
bandlimited kernel functions such as Shannon’s kernel, smoothed Shannon ker-
nels, etc., and non-bandlimited kernel functions such as rational kernel functions,
exponential kernel functions, (smoothed) Haar kernel functions, etc.
5.2. Bandlimited kernel functions
These kernel functions are characterized by the property that only a finite number
of coefficients An does not vanish. Consequently, the reproducing kernel Hilbert
space is of finite dimension.
At this stage two important cases of bandlimited kernels should be mentioned:
(a) The Shannon Kernel (see Figure 5.1). For a non-negative integer N we let

1, n ∈ [0, N + 1),
An =
0, n ∈ [N + 1, ∞),
i.e., N (An ) = {0, . . . , N }. Obviously, the reproducing kernel Hilbert space
H(Ωext ext
σ ) is equal to the space Harm 0,...,N (Ωσ ) of outer harmonics of degree
≤ N . The reproducing kernel function KH(Ωext ) (·, ·) : Ωext ext
σ × Ωσ → R, i.e.,
σ
the Shannon kernel, reads as follows:
2n+1

∗ ∗
KH(Ωext ) (x, y) = Hn,m (σ; x)Hn,m (σ; y)
σ
0≤n≤N m=1
2n + 1 σ 2 n+1
x y

= Pn · . (5.1)
4πσ 2 |x| |y| |x| |y|
0≤n≤N
Observing the well-known recursion relation for Legendre polynomials
(n+1)(Pn+1 (t)−Pn (t))−n(Pn (t)−Pn−1 (t)) = (2n+1)(t−1)Pn (t), n ≥ 1, (5.2)
we obtain for (x, y) ∈ Ωσ × Ωσ

x y N +1 x y x y
· − 1 KH(Ωext ) (x, y) = PN +1 · − PN · .
|x| |y| σ 4πσ 2 |x| |y| |x| |y|
(5.3)
(a) K(x, y) for N = 25 − 1 (b) K ∧ (n) for N = 25 − 1
(c) K(x, y) for N = 27 − 1 (d) K ∧ (n) for N = 27 − 1
Figure 5.1. Shannon kernel with N = 25 − 1 (above) and N = 27 − 1

(below): space domain, i.e., K(x, y) for (x, y) ∈ Ωσ × Ωσ in sectional
representation (left) and frequency domain, i.e., K ∧ (n) = An (right).
(b) Smoothed Shannon Kernels (see Figure 5.2). For (fixed) non-negative integers
N, M with N > M + 1 we let
⎧
⎪
⎨ 1, n ∈ [0, M + 1),
N −m
An = N −M , n ∈ [M + 1, N + 1),
⎪
⎩ 0, n ∈ [N + 1, ∞).
Of course, many other suitable choices can be found for practical purposes.
(a) K(x, y) (b) K ∧ (n)
Figure 5.2. Smoothed Shannon kernel with M = 26 and N = 27 − 1:

space domain, i.e., K(x, y) for (x, y) ∈ Ωσ × Ωσ in sectional representa-
tion (left) and frequency domain, i.e., K ∧ (n) = An (right).
5.3. Non-bandlimited kernel functions

All non-bandlimited kernels share the property that an infinite number of coef-
ficients An is different from zero. The corresponding reproducing Hilbert kernel
spaces are infinite-dimensional. We mention rational kernels, exponential kernels,
and “locally supported” kernels, i.e., (smoothed) Haar kernels.
(a) Rational Kernels (see Figure 5.3). Let {An } be a sequence of real numbers
An satisfying the following conditions:
(i) n → A2n , n ∈ N0 , is a (real) rational function (in the integer variable n).
(ii) There exist two positive constants C, C
with
2+ε α
n + ( 12 ) n + ( 12 )
C ≤ A2n ≤ C
(5.4)
σ σ
for some ε > 0, α ≥ 2 + ε.
Then the norm reads
∞ 2n+1
2
F 2H(Ωext ) = A2n (F ∧ (n, m)) .
σ
n=0 m=1
For the reproducing kernel in H(Ωextσ ) we find the representation

∞ 2 n+1
1 2n + 1 σ x y
KH(Ωext ) (x, y) = Pn · .
σ A2 4πσ 2
n=0 n
|x| |y| |x| |y|

n+ 12 −β
An σ is summable for all β < ε/2.
(b) Exponential Kernels. An alternative to come to candidates of reproducing
kernel sum representations with an exponential rate of convergence is to use
a sequence {An } of the form
σ n
An = Bn , n ∈ N , (5.5)
σ
(a) K(x, y) (b) K ∧ (n)
Figure 5.3. Rational kernel with A2n = (1 + n)−s , s = 6.5 : space

domain, i.e., K(x, y) for (x, y) ∈ Ωσ × Ωσ in sectional representation
(left) and frequency domain, i.e., K ∧ (n) = An (right).
with σ
< σ and Bn satisfying
α
n + ( 12 )
0< Bn2 ≤C
(5.6)
σ
for all n ∈ N , some value α and a positive constant C
. The radius σ
(< σ)
should be taken close to the value σ (i.e., σ
is assumed to be the radius of a
Runge sphere so that σ/σ
is close to 1). It is evident that an “inner radius” σ
gives additional flexibility in choosing the norm of the Hilbert space and also
results in more general sequences {An } being possible. On the other hand,
the radius σ
appears as an artificial value in the infinite sum of the kernel
to force an exponential
rate of sum convergence. In conclusion, the sequence
1 −β
n+ 2
An σ is summable for every β.
Kernel representations of type (5.5) for (x, y) ∈ Ωext

σ × Ωσ
ext
1 2n + 1 n+1
σ
2 x y
KH(Ωext ) (x, y) = Pn ·
σ Bn2 4πσ 2 |x| |y| |x| |y|
n∈N
are well known from geophysical applications (see, for example, [14, 32, 55,
72]).
Far- and Near-Field Methods as well as Multipole Methods are explained
in the Ph.D.-thesis [39] and can also be found in [24, 40, 41] and in the
contribution [42] in this volume.
Of particular importance for purposes of minimum norm (spline) in-
terpolation and smoothing (cf., e.g., [14–16, 18, 72]) are kernels, which are
available in terms of elementary functions. We only mention here (cf. [52]):
(i) Abel–Poisson kernel (see Figure 5.4):
Bn2 = 1, n ∈ N0 . (5.7)
σ σ
(a) K(x, y) for σ
= 0.7 (b) K ∧ (n) for σ
= 0.7
σ σ
(c) K(x, y) for σ
= 0.9 (d) K ∧ (n) for σ
= 0.9

Figure 5.4. Abel–Poisson kernel with σσ = 0.7 (above) and σσ = 0.9
(below): space domain, i.e., K(x, y) for (x, y) ∈ Ωσ × Ωσ in sectional
The kernel reads as follows:

|x| |y| |x|2 |y|2 − σ
4
KH(Ωext ) (x, y) = , x, y ∈ Ωext
σ ,
σ 4πσ
2 (L(x, y))3/2
L(x, y) = |x|2 |y|2 − 2σ
2 x · y + σ
4 .
(ii) “Singularity kernel” (see Figure 5.5)
Bn2 = (2n + 1)/2, n ∈ N0 . (5.8)
The kernel is given by
1 1
KH(Ωext ) (x, y) = , x, y ∈ Ωext
σ .
σ 4π (L(x, y)) 12
(iii) “Logarithmic kernel” (see Figure 5.6)
Bn2 = (2n + 1)(n + 1), n ∈ N0 . (5.9)
(a) K(x, y) (b) K ∧ (n)

Figure 5.5. Singularity kernel with σσ = 0.7: space domain, i.e.,
K(x, y) for (x, y) ∈ Ωσ × Ωσ in sectional representation (left) and fre-
quency domain, i.e., K ∧ (n) = An (right).
(a) K(x, y) (b) K ∧ (n)

Figure 5.6. Logarithmic kernel with σσ = 0.7: space domain, i.e.,
K(x, y) for (x, y) ∈ Ωσ × Ωσ in sectional representation (left) and fre-
quency domain, i.e., K ∧ (n) = An (right).
Now we have

1 2σ
2
KH(Ωext ) (x, y) = ln 1 + , x, y ∈ Ωext
σ ,
σ 4πσ
2 M (x, y)
with
1
M (x, y) = (L(x, y)) 2 + |x| |y| − σ
2 .
(c)“Locally Supported” Kernels (Smoothed Haar Kernels, see Figure 5.7):

(k)
Consider the piecewise polynomial function Bh : [−1, +1] → R, k = 0, 1, . . .
and h ∈ (0, 1) given by

(k)
0, t ∈ [−1, h),
Bh (t) = (t−h)k (5.10)
(1−h)k , t ∈ [h, 1],
(cf. [5, 20, 21, 26, 35, 67]). Let ξ ∈ Ω = Ω1 be fixed. Then the ξ-zonal function
(k) (k)
Bh (ξ ·) : Ω → R has a local support. More explicitly, the support of Bh (ξ ·)
is the cap with centre ξ characterized by
(k)
suppBh (ξ ·) = {η ∈ Ω : h ≤ ξ · η ≤ 1}.
(0)
The ξ-zonal function Bh (ξ ·) : Ω → R given by

(0) 0 for ξ · η ∈ [−1, h),
Bh (ξ · η) =
1 for ξ · η ∈ [h, 1].
(k)
is called the Haar kernel at position ξ ∈ Ω, while Bh (ξ ·), k > 0, are called
“smoothed” Haar kernels at position ξ ∈ Ω.
(a) Haar kernel K(x, y) (b) Symbol K ∧ (n) of the Haar kernel
(c) Smoothed Haar kernel K(x, y) (d) Symbol K ∧ (n) of the smoothed
Haar kernel
Figure 5.7. Haar kernel (above) and smoothed Haar kernel (below)
with h = 0.7: space domain, i.e., K(x, y) for (x, y) ∈ Ωσ ×Ωσ in sectional
An easy calculation shows that the iterated “Haar kernel”

(k) (k) (k)
(Bh )(2) (ξ ·) = (Bh ∗L2 (Ω) Bh )(ξ ·)
also has a cap with centre ξ as a local support:
(k)
supp(Bh )(2) (ξ ·) = {η ∈ Ω : 2h2 − 1 ≤ ξ · η ≤ 1}.
(k)
Expanding Bh in terms of Legendre polynomials we obtain
∞
(k) 2n + 1 (k) ∧
Bh = (Bh ) (n)Pn , (5.11)
n=0
4π
where
+1 k
(k) t−h
(Bh )∧ (n) = 2π Pn (t) dt, n = 0, 1, . . . .
−1 1−h
The recurrence formulae for Legendre polynomials give us
(k)
(k + 1)(Bh )∧ (0) = 2π(1 − h), (5.12)
(k) (k)
(k + 2)(Bh )∧ (1) = (k + 1 + h)(Bh )∧ (0), (5.13)
(k) (k)
(n + k + 2)(Bh )∧ (n + 1) = (2n + 1)h(Bh )∧ (n)
(k)
+ (k + 1 − n)(Bh )∧ (n − 1) (5.14)
(for more details the reader is referred to [26]).

(0)
For k = 0 it is easy to see that (Bh )∧ (n) = O(n−3/2 ), n → ∞.
Moreover, from the recurrence relations Eqs. (5.12)–(5.14) it follows that

(k) ∧
(Bh ) (n) = O(n−(3/2)−k ), n → ∞.
Furthermore, [67] has shown the following statements:
(k)
(i) (Bh )∧ (n) = 0 for n = 0, 1, . . ., k + 2.
(k) k+ 3
(ii) For n ≥ k + 2, (Bh )∧ (n) = 0 if and only if Cn−k−1
2
(h) = 0 (where
k+ 32
Cm is the Gegenbauer polynomial of order m with respect to k + 32 ).
This leads us to the following result: For k ≥ 0, h ∈ (0, 1), the sequence

(k)
((Bh )∧ (n))−1 , n ∈ N,
An = (5.15)
0, n ∈ N0
is summable.
In case of locally supported kernels we have the following lemma:
(0)
Lemma 5.5. H(Ωext σ ) ⊂ Pot (Ωext
σ ), as defined by (5.15), is a reproducing kernel
Hilbert space with the reproducing kernel
(k) (2) ∧ 2n + 1
2 n+1
σ

x y

KH(Ωext ) (x, y) = Bh (n) P n · .
σ 4πσ 2 |x| |y| |x| |y|
n∈N
(5.16)
Moreover, for x = σξ, y = ση, we have
(2) x (2)
y
σ 2 KH(Ωext ) (x, y)
(k) (k)
= Bh · = Bh (ξ · η),
σ |x|=σ, |x| |y|
|y|=σ
where
(2) x x y

(k)
supp Bh · = y ∈ Ωσ : 2h2 − 1 ≤ · ≤1 .
|x| |x| |y|
In other words, reproducing kernel Hilbert spaces of potentials defined
on and outside the sphere Ωσ are found such that the “restriction” (x, y) →
KH(Ωext ) (x, y), (x, y) ∈ Ωσ × Ωσ , is a locally supported (zonal) function on Ωσ
σ
(k)
(note that (Bh )(2) (ξ · η) is a zonal function, i.e., depends only on the scalar
product of the unit vectors ξ and η).
6. Ill-posedness of the satellite problems

The question of subsets X ⊂ Ωext γ on which observations are required in order
to uniquely determine the potential F |Σext , is answered in this section. In order
to handle existence and stability of the solution we give a reformulation of the
pseudodifferential operators as convolution operators.
6.1. Scalar SST and SGG problem
Throughout the remaining part of this contribution, the sequence {An } ∈ A gener-
ating the reference space H(Ωextσ ) for gravitational field determination is assumed
to satisfy the so-called ‘consistency conditions’:
Definition 6.1. A sequence {An } ∈ A is said to satisfy the consistency conditions
(CC1) and (CC2) relative to [σ, σ inf ), if the following conditions are satisfied:
(CC1) An is different from 0 for all n ∈ N0 , i.e.,
An = 0, n = 0, 1, . . . , (6.1)
and
(CC2) there exists a value τ with σ ≤ τ < σ inf such that
∞ σ n 1
(2n + 1) < ∞. (6.2)
n=0
τ A2n
The “downward continuation problem” of determining the potential F ∈
Hs (Ωext ext
σ ) from “satellite data” G ∈ Hs (Ωγ ) reads as follows.
(i) (Scalar) SST Problem (Corresponding to the First-order Radial Derivative).

Let the values G(x), x ∈ X, for some subset X ⊂ Ωext γ be known from a
ext
function G of the class Hs (Ωγ ). We search for a potential F |Σext with F
being from Hs (Ωext
σ ) that fulfills the (scalar) SST operator equation with the
SST operator Λ : Hs (Ωext ext
σ ) → Hs (Ωγ ) given by
ΛF (x) = G(x), x ∈ X, (6.3)

where

x
(ΛF )(x) = − · ∇x F (x)||x|=γ = G(x), x ∈ X. (6.4)
|x|
Equation (6.4) means that the SST operator is the composition of the radial
derivative and the upward continuation operator. Having in mind that the
symbol of a pseudodifferential operator Λ : Hs (Ωext ext
σ ) → Hs (Ωγ ) satisfies
∧
ΛHn,m (σ; ·) = Λ (n)Hn,m (γ; ·), we have
s s
n
∧ n+1 σ
Λ (n) = , n = 0, 1, . . . , (6.5)
γ γ
and the SST operator is given by
∞ 2n+1

ΛF (x) = Λ∧ (n)F ∧ (n, m)Hn,m
s
(γ; x). (6.6)
n=0 m=1
(ii) (Scalar) SGG problem (Corresponding to the Second-order Radial Derivative).

Let the values G(x), x ∈ X, for some subset X ⊂ Ωext
γ be known from a func-
ext
tion G of the class Hs (Ωγ ). We search for a potential F |Σext with F being
from Hs (Ωext
σ ) such that
ΛF (x) = G(x), x ∈ X, (6.7)

where the SGG operator Λ : Hs (Ωext
σ ) → Hs (Ωext
γ ) with the symbol
n
(n + 1)(n + 2) σ
Λ∧ (n) = , n = 0, 1, . . . , (6.8)
γ2 γ
is given by

x x
ΛF (x) = − · ∇x − · ∇x F (x)||x|=γ
|x| |x|
∞ 2n+1

= Λ∧ (n)F ∧ (n, m)Hn,m
s
(γ; x). (6.9)
n=0 m=1
In the case of combined SST/SGG data we have the following formulation in terms
of pseudodifferential operators.
(iii) Combined (scalar) SST/SGG problem. Let the values G1 (x), x ∈ X1 ⊂ Ωext γ
and G2 (x), x ∈ X2 ⊂ Ωext
γ be known from a function of class Hs (Ω ext ). Let
γ
the symbols of the two corresponding pseudodifferential operators Λ1 and Λ2
be given by
n
∧ σ n+1
Λ1 (n) = , n = 0, 1, . . . for SST, (6.10)
γ γ
n
σ (n + 1)(n + 2)
Λ∧2 (n) = , n = 0, 1, . . . for SGG. (6.11)
γ γ2
Find a potential F ∈ Hs (Ωext
σ )|Σext such that
(Λ1 F ) (x) = G1 (x), x ∈ X1 , (6.12)

(Λ2 F ) (x) = G2 (x), x ∈ X2 . (6.13)
In order to give an answer to the question of subsets X ⊂ Ωext

γ on which
data are necessary to assure uniqueness of the solution F , we define Hs (Ωext
σ )-
fundamental systems.
Definition 6.2. A system X = {xn }n=0,1,... of points xn ∈ Ωext γ is called an
Hs (Ωext
σ )-fundamental system in Ω ext , if the conditions F ∈ H (Ωext ) and
γ s σ
F (xn ) = 0 for n = 0, 1, . . . imply F = 0.
For fundamental systems we get the following uniqueness theorems which are
proved in the Ph.D.-thesis [58].
Theorem 6.3. Let X = {xn }n=0,1,... be an Hs (Ωext ext
σ )-fundamental system in Ωγ .
Then the potential F |Σext solving the (scalar) SST or SGG problem is uniquely
defined.
Theorem 6.4. Let X1 ⊂ Ωext
σ , X 2 ⊂ Ωσ
ext such that X = X ∪ X = {x }
1 2 n n=0,1,...
is an Hs (Ωσ )-fundamental system in Ωext
ext
γ . Then the potential F |Σext solving the
combined (scalar) SST/SGG problem is uniquely defined.
In order to present the results concerning the ill-posedness of the satellite
problems, we essentially follow [19]. We reformulate the SST and SGG problem as
a convolution equation using kernel functions.
Definition 6.5. Let α, β ∈ R, α ≥ σ and β ≥ σ. Then any kernel K α,β (·, ·) :
Ωext ext
α × Ωβ → R of the form
∞
2n+1

K α,β (x, y) = K ∧ (n) s
Hn,m s
(α; x)Hn,m (β; y) (6.14)
n=0 m=1
(x, y) ∈ Ωext ext

α × Ωβ , is called an Hα,β -kernel.
n
The sequence {(K α,β )∧ (n)}n∈N0 with (K α,β )∧ (n) = αβ
σ2 K ∧ (n), n =
0, 1, . . ., is called the (α, β)-symbol of the Hα,β -kernel K α,β (·, ·). The (σ, σ)-symbol
of the Hα,β -kernel K α,β (·, ·) is simply called the symbol of the Hα,β -kernel.
Definition 6.6. An Hα,β -kernel K α,β (·, ·) with symbol {K ∧(n)}n=0,1,... is called
admissible, if the following conditions are satisfied:
∞ ∧ 2
(i) n=0 (K (n)) < ∞, 2s
∞ ∧ 2
(ii) n=0 (2n + 1) (K (n)) σ
n+ 1 < ∞.
2
The first property in Definition 6.6 ensures that K ∧ (n) → 0 as n → ∞,

whereas the second condition implies the following lemma.
Lemma 6.7. Let α, β ∈ R, α ≥ σ, β ≥ σ.
(i) If K α,β (·, ·) is an admissible Hα,β -kernel with the symbol {K ∧(n)}n=0,1,... ,
then K α,β (x, ·) is an element of Hs (Ωext ext
β ) for every (fixed) x ∈ Ωα .
(ii) If K α,β (·, ·) is an admissible Hα,β -kernel with the symbol {K ∧(n)}n=0,1,... ,
then K α,β (·, y) is an element of Hs (Ωext ext
α ) for every (fixed) x ∈ Ωβ .
Suppose now that F, G are elements of class Hs (Ωext σ ). Then we understand

the Hs (Ωext
σ )-convolution of F and G simply to be the inner product in Hs (Ωext
σ ),
i.e.:
F ∗ G = (F, G)Hs (Ωext ) . (6.15)
σ
(More precisely, we had to write F ∗ G = F ∗Hs (Ωext ) G.) By definition, we let

σ
∧
F (n, k) = F ∗ s
Hn,k (σ; ·) (6.16)
for n ∈ N (An ); k = 1, . . . , 2n + 1. It follows from (6.15) via the Parseval identity
that
2n+1

F ∗G= F ∧ (n, k)G∧ (n, k),
n∈N k=1
for F, G ∈ Hs (Ωext
σ ).
We now define the convolution of an admissible Hα,β -kernel against a function
F ∈ Hs (Ωext
β ) as follows:
(K α,β ∗ F )(x) = K α,β (x, ·) ∗ F

∞ 2n+1

= K ∧ (n)F ∧ (n, m)Hn,m
s
(α; x), x ∈ Ωext
α . (6.17)
n=0 m=1
It directly follows that (K α,β ∗ F )∧ (n, m) = K ∧ (n)F ∧ (n, m) and K α,β ∗ F ∈

Hs (Ωext
α ). In analogous way we define the convolution of an Hα,β -kernel K
α,β
(·, ·)
against a function F ∈ Hs (Ωext
α ) by
(K α,β ∗ F )(y) = K α,β (·, y) ∗ F
∞ 2n+1

= K ∧ (n)F ∧ (n, m)Hn,m
s
(β; y), y ∈ Ωext
β , (6.18)
n=0 m=1
and K α,β ∗ F is an element of Hs (Ωext

β ).
If L, K are admissible Hσ,σ -kernels, then the Hs (Ωext
σ )-convolution L ∗ K is
defined by
(L ∗ K)(x, y) = (L(x, ·), K(·, y))Hs (Ωext ) , (x, y) ∈ Ωext ext
σ × Ωσ .
σ
Obviously, (L ∗ K)(·, ·) is an admissible Hσ,σ -kernel, and it is not difficult to see

that

(L ∗ K)∧ (n) = L∧ (n)K ∧ (n), n ∈ N (K ∧ (n)L∧ (n))−1 An .
We usually write K (2) (·, ·) = (K ∗ K)(·, ·) to indicate the convolution of an
Hσ,σ -kernel with itself. K (2) (·, ·) = (K ∗ K)(·, ·) is said to be the iterated ker-
nel of K(·, ·). More generally, K (p) (·, ·) = (K (p−1) ∗ K)(·, ·) for p = 2, 3, . . ., and
K (1) (·, ·) = K(·, ·) for p = 1. Obviously, we have

(K (2) )∧ (n) = (K ∧ (n))2 .
In order to give an answer to the question of ill-posedness of the (scalar)
SST or SGG problem, the continuity of the inverse additionally has to be inves-
tigated. The answer to this question requires the reformulation of the problem as
convolution equation. Starting from a pseudodifferential operator Λ : Hs (Ωext
σ ) →
Hs (Ωext
γ ) given by
∞ 2n+1

ΛF = Λ∧ (n)F ∧ (n, m)Hn,m
s
(γ; ·), (6.19)
n=0 m=1
we can interpret the symbol of the pseudodifferential operator as the symbol of an

Hσ,γ -kernel (K Λ )σ,γ presuming that the symbol satisfies the admissibility condi-
tions. The pseudodifferential operator is then given by the convolution identity
ΛF (x) = (K Λ )σ,γ (·, x) ∗ F, x ∈ Ωext
γ , (6.20)
Λ ∧ ∧
for F ∈ Hs (Ωext
σ ), where (K ) (n) = Λ (n), n = 0, 1, . . .. Obviously, we have
(K Λ )σ,γ (·, x) ∗ Hn,m

s
(σ; ·) = (K Λ )γ,σ (x, ·) ∗ Hn,m
s
(σ; ·)
= Λ∧ (n)Hn,m
s
(γ; x), (6.21)
for all n ∈ N; m = 1, . . . , 2n + 1, or, equivalently,
s
ΛHn,m (σ; ·) = Λ∧ (n)Hn,m
s
(γ; ·). (6.22)
Having a look at the (scalar) SST and SGG operator, we get the following result.
Theorem 6.8. The Hσ,γ -kernel (K Λ )σ,γ defined by the symbol
⎧ n
⎨ σγ n+1
γ , n = 0, 1, . . . for SST,
Λ∧ (n) = n (6.23)
⎩ σ (n+1)(n+2)
, n = 0, 1, . . . for SGG,
γ γ2
1 s
n+
is admissible, if { σ 2 } is summable in the sense of Eq. (3.3).
Theorem 6.9. Let Λ : Hs (Ωext ext

σ ) → Hs (Ωγ ) be a pseudodifferential operator with
Λ σ,γ
(K ) satisfying the admissibility conditions. Then the pseudodifferential opera-
tor Λ is bounded and Λ = max |Λ∧ (n)|. Further on, Λ is an injective operator.
n∈N0
From functional analysis (see, e.g., [70, 77]), we know that the SST and SGG
operators are compact as being so-called Hilbert–Schmidt operators. Summing up
the preceding considerations we finally get the following result.
Theorem 6.10. Let
ΛF = G, F ∈ Hs (Ωext
σ ), G ∈ Hs (Ωext
γ ), (6.24)
be the (scalar) SST or SGG problem. Then Λ is a compact operator with infinite-
dimensional range. Furthermore, Λ−1 is not bounded on Hs (Ωext
γ ). The (scalar)
SST or SGG problem is solvable if and only if
G∧ (n, m) 2
∞ 2n+1
< ∞. (6.25)
n=0 m=1
Λ∧ (n)
Remembering Hadamard’s definition of a well-posed problem (existence,

uniqueness and continuity of the inverse), we consequently see that the (scalar)
SST or SGG problem is ill posed, as it violates the first and third condition.
6.2. Vectorial SST and SGG problem

Following [58], we additionally formulate uniqueness results for the (vectorial) 1 s
n+ 2
SST and SGG problems. Let Hs (Ωext σ ) be a (scalar) Sobolev space with σ
satisfying the consistency condition (CC2) relative to [σ, τ ) (see Eq. (6.2)). Further
(i)
on, let hs (Ωext γ ), i = 1, 2, be (vectorial) Sobolev spaces. Then the “downward
continuation problem” of determining the potential F ∈ Hs (Ωext σ ) from “satellite
(1) (2)
data” g ∈ hs (Ωext γ ) ⊕ h s (Ω ext ) reads as follows.
γ
(i) (Vectorial) SST problem (Corresponding to the First-order Tangential Deriv-
ative). Let the values g(x), x ∈ X, for some subset X ⊂ Ωext
γ be known from
(1) (2)
a function g of the class hs (Ωext ext
γ ) ⊕ hs (Ωγ ). We search for a potential
ext
F |Σext with F being of the class Hs (Ωσ ) such that
λF (x) = g(x), x ∈ X, (6.26)
(1) (2)
where the SST Operator λ : Hs (Ωext ext ext
σ ) → hs (Ωγ ) ⊕ hs (Ωγ ) is given by
(λF )(x) = ∇∗,σ

ξ F (x)||x|=γ , (6.27)
with x = |x|ξ. Observing the symbol
⎧ n
⎨ σγ n
γ
n+1
2n+1 , i = 1; n = 0, 1, . . . ,
(i)∧ n
λ (n) = (6.28)
⎩ σ n+1 n
γ γ 2n+1 , i = 2; n = 1, 2, . . . ,
the (vectorial) SST operator can be written as

∞ 2n+1
2
λF (x) = λ(i)∧ (n)F ∧ (n, m)h(i)s
n,m (γ; x). (6.29)
i=1 n=1 m=1
In the case of SGG-data the mixed derivatives can be handled within vectorial
framework.
(ii) (Vectorial) SGG problem (Corresponding to the Second-order Mixed Deriva-
tives). Let the values g(x), x ∈ X, for some subset X ⊂ Ωext
γ be known from
(1) (2)
a function g of the class hs (Ωext ext
γ ) ⊕ hs (Ωγ ). We search for a potential
ext
F |Σext with F being of the class Hs (Ωσ ) such that
λF (x) = g(x), x ∈ X, (6.30)
(1) (2)
where the SGG operator λ : Hs (Ωext σ ) → hs (Ωext ext
γ ) ⊕ hs (Ωγ ) with symbol
⎧ n
⎨ σγ n(n+1)
γ2
n+1
2n+1 , i = 1; n = 0, 1, . . . ,
(i)∧
λ (n) = n (6.31)
⎩ σ (n+1) 2
n
γ γ2 2n+1 , i = 2; n = 1, 2, . . . ,
is given by
∞ 2n+1
2
λ(i)∧ (n)F ∧ (n, m)h(i)s
n,m (γ; x). (6.32)
i=1 n=1 m=1
γ on which data
(i)
are necessary to get uniqueness of the solution F , we define hs (Ωext
σ )-fundamental
systems.
Definition 6.11. A system X = {xn }n=0,1,... of points xn ∈ Ωext σ is called an
(i) (i)
hs (Ωσ )-fundamental system in Ωσ , if the conditions g ∈ hs (Ωext
ext ext
σ ) and
g(xn ) = 0 for n ∈ N0 imply g = 0, i ∈ {1, 2, 3}. Further on, X is called an
(i) (j) (i) (j)
hs (Ωext ext ext
σ ) ⊕ hs (Ωσ )-fundamental system, if g ∈ hs (Ωσ ) ⊕ hs (Ωσ ) and
ext
g(xn ) = 0 for n ∈ N0 imply g = 0 for i, j ∈ {1, 2, 3} with i = j.

We now obtain the following uniqueness theorem.
(1) (2)
Theorem 6.12. Let X = {xn }n=0,1,... be an hs (Ωext ext
γ ) ⊕ hs (Ωγ )-fundamental
ext
system in Ωγ . Then the potential F |Σext solving the (vectorial) SST or SGG
problem is uniquely defined up to an additive constant C.
Definition 6.13. Let α, β ∈ R, α ≥ σ and β ≥ σ. Then any kernel k (i),α,β (·, ·) :
ext 3
Ωext
α × Ωβ → R of the form
∞
2n+1

k (i),α,β (x, y) = k (i)∧ (n) s
Hn,m (α; x)h(i)s
n,m (β; y), (6.33)
n=0i m=1
ext (i)
(x, y) ∈ Ωext
α × Ωβ , is called an hα,β -kernel. Furthermore,
3

k α,β (x, y) = k (i),α,β (x, y), (6.34)
i=1

α × Ωβ , is called an hα,β -kernel.
∧
The sequence { k (i),α,β (n)}n∈N0i with
∧ n
(i),α,β αβ
k (n) = k (i)∧ (n), n = 0i , . . . , (6.35)
σ2
(i)
is called the (α, β)-symbol of the hα,β -kernel k (i),α,β (·, ·). The (σ, σ)-symbol of the
(i) (i)
hα,β -kernel k (i),α,β (·, ·) is simply called the symbol of the hα,β -kernel.
(i)
Definition 6.14. An hα,β -kernel k (i),α,β (·, ·) with symbol {k (i)∧ (n)}n=0i ,... is called
∞ (i)∧
(i) n=0i (k (n))2 < ∞,
∞ (i)∧ 2 σ 2s
(ii) n=0i (2n + 1) k (n) n+ 12
< ∞,
∞ (1)∧ 2 σ 2s
(iii) (a) n=0 (2n + 1)(2n + 3) k (n) n+ 12
< ∞,
∞ (2)∧ 2 σ 2s
(b) n=1 (2n + 1)(2n − 1) k (n) n+ 12
< ∞,
∞ (3)∧ 2 2s
(c) n=1 (2n + 1)(2n + 1) k (n) σ
n+ 1
< ∞.
2
(i)
Furthermore, the hα,β -kernel is called admissible, if the hα,β -kernels, i ∈ {1, 2, 3},
are admissible.
The second and the third condition imply the following lemma.
(i)
(i) If k (i),α,β (·, ·) is an admissible hα,β -kernel with the symbol {k (i)∧ (n)}n=0i ,... ,
(i)
then k (i),α,β (x, ·) is an element of hs (Ωext ext
β ) for every (fixed) x ∈ Ωα .
(i)
(ii) If k (i),α,β (·, ·) is an admissible hα,β -kernel with the symbol {k (i)∧ (n)}n=0i ,... ,
(i),α,β
then the component functions k (·, y) ε are elements of Hs (Ωext
T l
α ) for
every (fixed) x ∈ Ωext β , l ∈ {1, 2, 3}.
(i)
Our next step is the definition of the convolution of an admissible hα,β -kernel
against a function f ∈ hs (Ωext
β ) as follows:
(k (i),α,β ∗ f )(x) = k (i),α,β (x, ·) ∗ f

∞ 2n+1

= k (i)∧ (n)f (i)∧ (n, m)Hn,m
s
α . (6.36)
n=0i m=1
It directly follows that (k (i),α,β ∗ f )∧ (n, m) = k (i)∧ (n)f (i)∧ (n, m), n = 0i , i ∈
{1, 2, 3}, and k (i),α,β ∗f ∈ Hs (Ωext
α ). In an analogous way we define the convolution
of an hα,β -kernel k α,β (·, ·) against a function F ∈ Hs (Ωext
α ) by
(k α,β F )(y) = k α,β (·, y) F

∞ 2n+1
3
= k (i)∧ (n)F ∧ (n, m)h(i)s
n,m (β; y), y ∈ Ωext
β , (6.37)
i=1 n=0i m=1
and k α,β F is an element of hs (Ωext

β ).
Our next purpose is to present the formulation of the vectorial SST respec-
tively SGG operators with the help of convolutions. This enables us to give an
answer to the question of continuity of the inverse. We start from a pseudodiffer-

ential operator λ : Hs (Ωext ext
σ ) → hs (Ωγ ) given by
∞ 2n+1
3
λF = λ(i)∧ (n)F ∧ (n, m)h(i)s
n,m (γ; ·), (6.38)
i=1 n=0i m=1
and interpret the symbol of the pseudodifferential operator as the symbol of an

hσ,γ -kernel (k λ )σ,γ presuming that the symbol satisfies the admissibility condi-
λF (x) = (k λ )σ,γ (·, x) F, x ∈ Ωext
γ , (6.39)
λ (i)∧
for F ∈ Hs (Ωext
σ ), where (k ) (n) = λ(i)∧ (n), i = 1, 2, 3; n = 0i , . . .. Obviously,
we have
(i)
(k λ )σ,γ (·, x) Hn,m
s
(σ; ·) = λ(i)∧ (n)h(i)s
n,m (γ; x), (6.40)
for all i = 1, 2, 3; n = 0i , . . .; m = 1, . . . , 2n + 1, or, equivalently
λ(i) Hn,m
s
(σ; ·) = λ(i)∧ (n)h(i)s
n,m (γ; ·). (6.41)
Having a look at the (vectorial) SST and SGG operator, we get the following
result.
Theorem 6.16. The hσ,γ -kernel (k λ )σ,γ defined by the symbol

⎧ n
⎨ σγ n
γ
n+1
2n+1 , n = 0, 1, . . . for SST,
λ (1)∧ (1)∧
(k ) (n) = λ (n) = n
⎩ σ n(n+1) n+1
, n = 0, 1, . . . for SGG,
γ γ2 2n+1
(6.42)
and
⎧ n
⎨ σγ n+1
γ
n
2n+1 , n = 1, 2, . . . for SST,
λ (2)∧ (2)∧ n
(k ) (n) = λ (n) =
⎩ σ (n+1) 2
n
γ γ2 2n+1 , n = 1, 2, . . . for SGG,
s (6.43)
n+ 12
is admissible, if { σ } is summable and satisfies, in addition, condition (iii)
in Definition 6.14.
Theorem 6.17. Let λ : Hs (Ωext ext

σ ) → hs (Ωγ ) be a pseudodifferential operator with
(k λ )σ,γ satisfying the admissibility conditions, and λ(i)∧ (n) = 0, i ∈ {1, 2, 3},
n = 0i , . . .. Then the pseudodifferential operator λ is bounded and

3 (i)∧
λ = max λ (n), (6.44)
n∈N0
i=1
where we let λ(2)∧ (0) = λ(3)∧ (0) = 0. Further on, λ is an injective operator.
Finally, we get the following result.

Theorem 6.18. Let
(1) (2)
λF = g, F ∈ Hs (Ωext ext ext
σ ), g ∈ hs (Ωγ ) ⊕ hs (Ωγ ) (6.45)
be the (vectorial) SST or SGG problem. Then λ is a compact operator with infinite-
(1) (2)
dimensional range. Furthermore, λ−1 is not bounded on hs (Ωext ext
γ ) ⊕ hs (Ωγ ).
The SST/SGG problem is solvable if and only if
2 g (i)∧ (n, m) 2
∞ 2n+1
< ∞. (6.46)
i=1 n=0 m=1
λ(i)∧ (n)
i
We consequently get that the (vectorial) SST/SGG problem is ill posed be-
cause existence and continuity of the inverse are violated.
6.3. Tensorial SGG problem
The formulation of the definitions and theorems for the tensorial case is straight-
forward. Let Hs (Ωext σ ) be a (scalar) Sobolev space satisfying the consistency con-
(i,k)
dition (CC2) relative to [σ, τ ) (see Eq. (6.2)). Further on, let hs (Ωextγ ), (i, k) ∈
{(1, 1), (1, 2), (2, 1), (2, 2)}, be (tensorial) Sobolev spaces. Then the “downward
continuation problem” of determining the potential F ∈ Hs (Ωext σ ) from “satel-
ext
lite data” g ∈ hs (Ωγ ), where we use the abbreviation
SGG
(1,1) (1,2) (2,1) (2,2)

hSGG
s (Ωext
γ )= h (Ωext
γ ) ⊕ hs (Ωext
γ ) ⊕ hs (Ωext
γ ) ⊕ hs (Ωext
γ ), (6.47)
reads as follows.
(i) (Tensorial) SGG problem (Corresponding to the Second-order Tangential Der-
ivative). Let the values g(x), x ∈ X, for some subset X ⊂ Ωext γ be known
from a function g of the class hSGG (Ω ext ). We search for a potential F |
s γ Σext
with F being from Hs (Ωext
σ ) such that
λF (x) = g(x), x ∈ X, (6.48)
where the SGG operator λ : Hs (Ωext
σ ) → hs
SGG
(Ωext
γ ) is given by
∗,σ ∗,σ
( λF )(x) = ∇ ⊗ ∇ F (x)||x|=γ , (6.49)
with x = |x|ξ. With the help of the symbol
λ(i,k)∧ (n)
⎧
⎪
⎪
n
n(n+1) (1,1)
⎪
⎪ γ
σ
γ 2 (2n+1)(2n+3) νn , (i, k) = (1, 1), n = 0, 1, . . . ,
⎪
⎪
⎪ σ n (n+1)(n−1)
⎪
⎪ (1,2)
⎨ − γ γ 2 ((2n−1)(2n+1) νn , (i, k) = (1, 2), n = 1, 2, . . . ,
= n
⎪
⎪ (2,1)
⎪
⎪ − σ n(n+2)
νn , (i, k) = (2, 1), n = 0, 1, . . . ,
⎪
⎪ γ γ 2 (2n+3)(2n+1)

⎪
⎪
⎪ σ n n(n+1)(n+2)
⎩ (2,2)
γ 2
γ (2n−1)(2n+1) νn , (i, k) = (2, 2), n = 2, 3, . . . ,
(6.50)
with
νn(1,1) = (n + 1)(n + 2)(2n + 1)(2n + 3), (6.51)
νn(1,2) = 3n4 , (6.52)
νn(2,1) 2
= (n + 1) (2n + 1)(2n + 3), (6.53)
νn(2,2) = n(n − 1)(2n − 1)(2n + 1), (6.54)
the SGG operator can be written as
∞ 2n+1

λF (x) = λ(i,k)∧ (n)F ∧ (n, m)h(i,k)s
n,m (γ; x), (6.55)
(i,k)∈I SGG n=0ik m=1
where I SGG = {(1, 1), (1, 2), (2, 1), (2, 2)} is the index set for the tensorial
SGG problem.

γ on which
(i,k)
data are necessary to get uniqueness of the solution F , we define hs (Ωext
σ )-
fundamental systems.
Definition 6.19. A system X = {xn }n=0,1,... of points xn ∈ Ωext σ is called an

(i,k) (i,k)
hs (Ωσ )-fundamental system in Ωσ , if the conditions g ∈ hs (Ωext
ext ext
σ ) and
g(xn ) = 0 for n ∈ N0 imply g = 0, i, k ∈ {1, 2, 3}. In analogy the fundamental
(i,k)
systems are defined for spaces which are direct sums of the spaces hs (Ωext σ ),
i, k ∈ {1, 2, 3}.
As in the scalar and vectorial case we have the following theorem.
Theorem 6.20. Let X = {xn }n=0,1,... be an hSGG s (Ωext

γ )-fundamental system in
Ωext
γ . Then the potential F | Σext solving the (tensorial) SGG problem is uniquely
defined up to a term of the form
1 2n+1
n+1
σ 1 x
V (x) = cnm Yn,m , x ∈ Ωext
σ , (6.56)
n=0 m=1
|x| σ |x|
for constants c01 , c11 , c12 , c13 ∈ R.
We finally shortly present the results using the reformulation as convolution

equation.
Definition 6.21. Let α, β ∈ R, α ≥ σ and β ≥ σ. Then any kernel k(i,k),α,β (·, ·) :

ext 3 3
Ωext
α × Ωβ → R ⊗ R of the form
∞
2n+1

k(i,k),α,β (x, y) = k(i,k)∧ (n) s
Hn,m (α; x)h(i,k)s
n,m (β; y), (6.57)
n=0ik m=1
ext (i,k)
(x, y) ∈ Ωext
α × Ωβ , is called an hα,β -kernel. Furthermore,
3

kα,β (x, y) = k(i,k),α,β (x, y), (6.58)
i,k=1

α × Ωβ , is called an hα,β -kernel.
∧
The sequence { k(i,k),α,β (n)}n∈N0ik with
∧ n
αβ
k(i,k),α,β (n) = k(i,k)∧ (n), n = 0ik , . . . , (6.59)
σ2
(i,k)
is called the (α, β)-symbol of the hα,β -kernel k(i,k),α,β (·, ·). The (σ, σ)-symbol of
(i,k) (i,k)
the hα,β -kernel k(i,k),α,β (·, ·) is simply called the symbol of the hα,β -kernel.
(i,k)
Definition 6.22. An hα,β -kernel k(i,k),α,β (·, ·) with the symbol {k(i,k)∧ (n)}n=0ik ,...
is called admissible, if the following conditions are satisfied:
∞ (i,k)∧
(i) n=0ik (k (n))2 < ∞,
∞ 2s
(i,k)∧
(ii) n=0ik (2n + 1)(k (n))2 n+ σ
1 < ∞,
∞
2
2s
(1,1)∧
(iii) (a) n=0 (2n + 1)(2n + 5)(k (n))2 n+ σ
1 < ∞,
2
∞ 2s
(i,k)∧
(b) n=0ik (2n + 1)(2n + 3)(k (n))2 n+ σ
1 < ∞,
2
(i, k) ∈ {(1, 3), (3, 1)},
∞ 2s
(i,k)∧
(c) n=0ik (2n + 1)(2n + 1)(k (n))2 n+ σ
1 < ∞,
2
(i, k) ∈ {(1, 2), (2, 1), (3, 3)},
∞ 2s
(i,k)∧ 2
n=0ik (2n + 1)(2n − 1)(k < ∞,
σ
(d) (n)) n+ 1
2
(i, k) ∈ {(2, 3), (3, 2)},
∞ 2s
(2,2)∧
(e) n=2 (2n + 1)(2n − 3)(k (n))2 n+ σ
1 < ∞.
2
(i,k)
Furthermore, the hα,β -kernel is called admissible, if all hα,β -kernels, i, k ∈
{1, 2, 3}, are admissible.
The second and the third condition imply the following lemma.
(i,k)
1. If the kernel k(i,k),α,β (·, ·) is an admissible hα,β -kernel with the symbol given
(i,k)
by {k(i,k)∧ (n)}n=0ik ,... , then k(i,k),α,β (x, ·) is an element of hs (Ωext β ) for
every (fixed) x ∈ Ωα . ext
(i,k)
2. If the kernel k(i,k),α,β (·, ·) is an admissible hα,β -kernel with the symbol
{k(i,k)∧ (n)}n=0ik ,... , then the component functions k(i,k),α,β (·, y) · εj ⊗ εl are
elements of Hs (Ωext ext
α ) for every (fixed) x ∈ Ωβ , j, l ∈ {1, 2, 3}.
(i,k)
We now define the convolution of an admissible hα,β -kernel against a func-
tion f ∈ hs (Ωext
β ) as follows:
(k(i,k),α,β ∗ f )(x) = k(i,k),α,β (x, ·) ∗ f (6.60)

∞ 2n+1

= k(i,k)∧ (n)f (i,k)∧ (n, m)Hn,m
s
α .
n=0ik m=1
It follows directly that (k(i,k),α,β ∗ f )∧ (n, m) = k(i,k)∧ (n)f (i,k)∧ (n, m), n = 0ik ,
i, k ∈ {1, 2, 3}, and k(i,k),α,β ∗ f ∈ Hs (Ωext
α ). In an analogous way we define the
convolution of an hα,β -kernel kα,β (·, ·) against a function F ∈ Hs (Ωext α ) by
(kα,β F )(y) = kα,β (·, y) F (6.61)

3
∞ 2n+1

= k(i,k)∧ (n)F ∧ (n, m)h(i,k)s
n,m (β; y), y ∈ Ωext
β ,
i,k=1 n=0ik m=1
and kα,β F is an element of hs (Ωext

β ). Our next purpose is to present the formu-
lation of the tensorial SGG operator with the help of convolutions. This enables
us to give an answer to the question of continuity of the inverse. We start from a
pseudodifferential operator λ : Hs (Ωext ext
σ ) → hs (Ωγ ) given by
3
∞ 2n+1

λF = λ(i,k)∧ (n)F ∧ (n, m)h(i,k)s
n,m (γ; ·), (6.62)
i,k=1 n=0ik m=1
and interpret the symbol of the pseudodifferential operator as the symbol of an

hσ,γ -kernel (k λ )σ,γ presuming that the symbol satisfies the admissibility condi-
λF (x) = (k λ )σ,γ (·, x) F, x ∈ Ωext

γ , (6.63)
for F ∈ Hs (Ωext λ )(i,k)∧ (n) = λ(i,k)∧ (n), i, k = 1, 2, 3; n = 0 , . . ..

σ ), where (k ik
Obviously, we have
(k λ )σ,γ (·, x) Hn,m

s
(σ; ·) = λ(i,k)∧ (n)h(i,k)s
n,m (γ; x), (6.64)
for all i, k = 1, 2, 3; n = 0ik , . . .; m = 1, . . . , 2n + 1, or, equivalently,
λ(i,k) Hn,m
s
(σ; ·) = λ(i,k)∧ (n)h(i,k)s
n,m (γ; ·). (6.65)
Having a look at the (tensorial) SGG operator, we get the following result.
Theorem 6.24. The hSGG λ )σ,γ defined by the symbol

σ,γ -kernel (k
⎧ n
⎪
⎪ n(n+1) (1,1)
⎪
⎪
σ
γ 2 (2n+1)(2n+3) νn , (i, k) = (1, 1),
⎪
⎪
γ

⎪ n
⎪
⎪
⎨ − σγ (n+1)(n−1) (1,2)
νn , (i, k) = (1, 2),
(i,k)∧ γ 2 ((2n−1)(2n+1)
λ (n) = n (6.66)
⎪
⎪ (2,1)
⎪
⎪ − σ n(n+2)
ν , (i, k) = (2, 1),
⎪
⎪ γ 2
γ (2n+3)(2n+1)

n
⎪
⎪ n
⎪
⎩ σ n(n+1)(n+2) (2,2)
γ γ 2 (2n−1)(2n+1) νn , (i, k) = (2, 2),
1 s
n+ 2
is admissible, if σ is summable and satisfies, in addition, condition (iii) in
Definition 6.22.
We finally get the following results.
Theorem 6.25. Let λ : H(Ωext ext
σ ) → h(Ωγ ) be a pseudodifferential operator with
λ
(k ) σ,γ
satisfying the admissibility conditions, and λ(i,k)∧ (n) = 0, i ∈ {1, 2, 3},
n = 0ik , . . .. Then the pseudodifferential operator λ is bounded and

3
λ = max λ (i,k)∧
(n), (6.67)
n∈N0
i,k=1
where the sum has to be understood in the same sense as in the vectorial case.
Further on, λ is an injective operator.
Theorem 6.26. Let
λF = g, F ∈ Hs (Ωext
σ ), g ∈ hSGG
s (Ωext
γ ), (6.68)
be the (tensorial) SGG problem. Then λ is a compact operator with infinite-
dimensional range. Furthermore, λ−1 is not bounded on hSGG s (Ωext
γ ). The SGG
problem is solvable if and only if
∞ 2n+1
2
g(i,k)∧ (n, m)
< ∞. (6.69)
(i,k)∈I SGG n=0ik m=1
λ(i,k)∧ (n)
We consequently have that the (tensorial) SGG problem is ill posed because
existence and continuity of the inverse are violated.
7. Geodetically oriented wavelet approximation

In this section we present a multiscale approach based on wavelet approximation.
Note that all modern multiscale approaches have a conception of wavelets as con-
stituting multiscale building blocks in common, which provide a fast and efficient
way to decorrelate a given signal data set. As already mentioned in Section 2.4,
this characterization contains three basic attributes (basis property, decorrelation
and efficient algorithms), which are common features of all classical wavelets and
form the key for a variety of applications, particularly for signal reconstruction
and decomposition, thresholding, data compression, denoising, etc.
7.1. Scalar wavelet theory
We start with the presentation of the scalar theory, where we follow the approach
given in [19]. First, we define an Hσ,σ -multiresolution analysis. We use the abbre-
viation Φ(2) (·, ·) = (Φ ∗ Φ)(·, ·), where Φ is an Hσ,σ -kernel.
Definition 7.1. Let {Φj (·, ·)}j∈N0 be a family of admissible Hσ,σ -kernels as de-
fined in Definition 6.6. Then the family {Vj (Ωext σ )}j∈N0 of scale spaces Vj (Ωσ )
ext
defined by
(2)
Vj (Ωext
σ ) = {Φj ∗ F : F ∈ Hs (Ωσ )},
ext (7.1)
is called an Hσ,σ -multiresolution analysis, if the following properties are satisfied:
(i) V0 (Ωext ext ext ext
σ ) ⊂ · · · ⊂ Vj (Ωσ ) ⊂ Vj+1 (Ωσ ) ⊂ · · · ⊂ Hs (Ωσ ),
; ·H (Ωext )
(ii) Vj (Ωext
σ )
s σ
= Hs (Ωext
σ ).
j∈N0
Wavelet analysis is based on the idea of splitting the function into a lowpass
part and several bandpass parts. The so-called scaling function corresponds to the
lowpass filter, whereas the bandbass filters are the shifted and dilated versions of
the wavelet, which are defined as differences between successive scaling functions
with the help of a so-called refinement equation.
Definition 7.2. A family {{ϕj (n)}n∈N0 }j∈N0 of sequences {ϕj (n)}n∈N0 is called a
generator of a scaling function, if it satisfies the following requirements:
(i) (ϕj (0))2 = 1, for all j ∈ N0 ,
2
(ii) (ϕj (n))2 ≤ (ϕj (n)) , for all j, j
∈ N0 with j ≤ j
and all n ∈ N,
2
(iii) lim (ϕj (n)) = 1, for all n ∈ N.
j→∞
Based on the definition of a generator of a scaling function, we now introduce

Hσ,σ -scaling functions.
Definition 7.3. A family {Φj (·, ·)}j∈N0 of Hσ,σ -kernels Φj (·, ·) defined by Φ∧
j (n) =
ϕj (n), n, j ∈ N0 , i.e.,
∞
2n+1

Φj (x, y) = ϕj (n) s
Hn,m s
(σ; x)Hn,m (σ; y), x, y ∈ Ωext
σ , (7.2)
n=0 m=1
is called an Hσ,σ -scaling function, if it satisfies the following properties:
(i) Φj (·, ·) is an admissible Hσ,σ -kernel for every j ∈ N0 (in the sense of Definition
6.6),
(ii) {Φ∧j (n)n∈N0 }j∈N0 constitutes a generator of a scaling function (in the sense
of Definition 7.2).
The following theorem shows the approximation property of an Hσ,σ -scaling
function.
Theorem 7.4. Let {Φj (·, ·)}j∈N0 be an Hσ,σ -scaling function. Then
(2)
lim F − Φj ∗ F Hs (Ωext ) = 0 (7.3)
j→∞ σ
holds for all F ∈ Hs (Ωext

σ ).
We now introduce the dilation and the shifting operator in order to define
an Hσ,σ -approximate identity. Let J, J1 , J2 ∈ N0 and x ∈ Ωext
σ . Then we define the
dilation operator DJ1 and the shifting operator Sx by
DJ1 : ΦJ2 (·, ·) → DJ1 ΦJ2 (·, ·) = ΦJ1 +J2 (·, ·), (7.4)
Sx : ΦJ (·, ·) → Sx ΦJ (·, ·) = ΦJ (x, ·). (7.5)
The shifting operator Sy acting on the second variable is defined in an analogous
way. Note that by definition ΦJ (·, ·) = DJ Φ0 (·, ·) for any J ∈ N0 .
Definition 7.5. Let {Φj (·, ·)}j∈N0 be an Hσ,σ -scaling function. Then {Pj }j∈N0 with
Pj : Hs (Ωext ext
σ ) → Hs (Ωσ ) defined by

(2)
Pj (F )(x) = Sx Dj Φ0 (·, ·), F
Hs (Ωext
σ )

(2)
= Φj (x, ·), F
Hs (Ωext
σ )
(2)
= (Φj ∗ F )(x), (7.6)
for F ∈ Hs (Ωext ext
σ ), x ∈ Ωσ , is called an Hσ,σ -approximate identity.
The kernel Φ0 is called mother kernel of the Hσ,σ -scaling function. Theorem
7.4 leads to
lim F − Pj (F ) Hs (Ωext ) = 0. (7.7)
j→∞ σ
The following theorem clarifies the connection between the concept of multireso-
lution analysis and the scaling functions.
Theorem 7.6. Let {Φj (·, ·)}j∈N0 be an Hσ,σ -scaling function. Then {Vj (Ωext
σ )}j∈N0
forms an Hσ,σ -multiresolution analysis.
We now turn to the definition of the primal and dual wavelet.
Definition 7.7. Let {Φj (·, ·)}j∈N0 be an Hσ,σ -scaling function. Then the families of
Hσ,σ -kernels {Ψj (·, ·)}j∈N0 , {Ψ̃j (·, ·)}j∈N0 given by
Ψ∧
j (n) = ψj (n), n, j ∈ N0 , (7.8)
Ψ̃∧
j (n) = ψ̃j (n), n, j ∈ N0 , (7.9)
are called (primal) Hσ,σ -wavelet and dual Hσ,σ -wavelet, respectively, if all Hσ,σ -
kernels Ψj (·, ·), Ψ̃j (·, ·), j ∈ N0 , are admissible and the symbols {ψj (n)}, {ψ̃j (n)},
in addition, satisfy the (scalar) refinement equation
ψ̃j (n)ψj (n) = (ϕj+1 (n))2 − (ϕj (n))2 (7.10)
for all j, n ∈ N0 .
The following equation is a direct consequence of the refinement equation:

J
(ϕJ+1 (n))2 = (ϕ0 (n))2 + ψ̃j (n)ψj (n), J ∈ N0 , (7.11)
j=0
for all n ∈ N0 . This property finally leads to the reconstruction formula which
states how the original function F ∈ Hs (Ωext σ ) can be derived from a lowpass part
and the corresponding bandpass parts (see Theorem 7.9).
We now turn to the definition of the wavelet transform. To this end we define
N−1 = N0 ∪ {−1} and let ψ−1 (n) = ψ̃−1 (n) = ϕ0 (n), for n ∈ N0 , Ψ−1 (·, ·) =
Ψ̃−1 (·, ·) = Φ0 (·, ·). This abbreviation simplifies our notation. Then we define the
space
∞

Hs (N−1 × Ωext ext
σ ) = {H : N−1 × Ωσ → R : (H(j; ·), H(j; ·))Hs (Ωext ) < ∞}
σ
j=−1
(7.12)
with inner product
∞

(H1 , H2 )Hs (N−1 ×Ωext ) = (H1 (j; ·), H2 (j; ·))Hs (Ωext ) (7.13)
σ σ
j=−1
and corresponding norm

⎛ ⎞ 1/2
∞

H Hs(N−1 ×Ωext ) = ⎝ H(j; ·) 2H ext
⎠ . (7.14)
σ s (Ωσ )
j=−1
With the help of the dilation operator Dj and the shifting operator Sy we
introduce the following abbreviation:
Ψj;y (·) = Ψj (·, y) = Sy Ψj (·, ·) = Sy Dj Ψ0 (·, ·), (7.15)
Ψ̃j;y (·) = Ψ̃j (·, y) = Sy Ψ̃j (·, ·) = Sy Dj Ψ̃0 (·, ·). (7.16)
Definition 7.8. Let {Ψj (·, ·)}j∈N−1 be a (primal) Hσ,σ -wavelet. Then
W T : Hs (Ωext ext
σ ) → Hs (N−1 × Ωσ ),
defined by
(W T )(F )(j; y) = (Ψj;y , F )H(Ωext ) = (Ψj ∗ F )(y), (7.17)
σ
is called Hσ,σ -wavelet transform of F at position y ∈ Ωext

σ and scale j ∈ N−1 .
Having the definition of the scale spaces Vj (Ωext

σ ) in mind, we now define the
detail spaces Wj (Ωext
σ ) at scale j by

Wj (Ωext ext
σ ) = Ψ̃j ∗ Ψj ∗ F : F ∈ Hs (Ωσ ) , j ∈ N0 . (7.18)
Theorem 7.9 (Scalar Reconstruction Formula for the Outer Space). Let the fami-
lies {Ψj (·, ·)}j∈N0 and {Ψ̃j (·, ·)}j∈N0 , respectively, be a (primal) Hσ,σ -wavelet and
its dual corresponding to an Hσ,σ -scaling function {Φj (·, ·)}j∈N0 . Then
∞

F = Ψ̃j ∗ Ψj ∗ F (7.19)
j=−1
holds for all F ∈ Hs (Ωext

σ ) (in · Hs (Ωext ) -sense).
σ
We now solve the (scalar) SST or SGG problem using bandlimited harmonic
wavelets. First, we define Hα,α -scaling functions with the help of a generator of
a scaling function {{ϕj (n)}n∈N0 }j∈N0 . Since the generator does not depend on σ,
we can directly extend the theory to the case of Hα,α -scaling functions Φα,α
j with
α ≥ σ:
∞ 2n+1

Φα,α
j (x, y) = ϕj (n) s
Hn,m s
(α; x)Hn,m (α; y), (7.20)
n=0 m=1
where
∧
(Φα,α
j ) (n) = ϕj (n). (7.21)
As a consequence, Theorem 7.4 is valid substituting σ by α. Furthermore, the
definition of the scale spaces can be directly transferred in the following way:
(2)
Vj (Ωext
α ) = {(Φj )
α,α
∗ F : F ∈ Hs (Ωext
α )}, (7.22)
where
(2)
(Φj )α,α = Φα,α
j ∗ Φα,α
j . (7.23)
The system {Vj (Ωext
α )}of scale spaces forms a multiresolution analysis due to
Theorem 7.6. We now investigate the solution of the restriction of an operator
Λ : Hs (Ωext ext
σ ) → Hs (Ωγ ) to a scale space Vj :
Λ : Vj (Ωext ext
σ ) → Vj (Ωγ ). (7.24)
Note that Λ(Vj (Ωext ext
σ )) ⊂ Vj (Ωγ ) is automatically fulfilled, because every F ∈
ext
Vj (Ωσ ) of the form
(2)
F = Φj ∗ Q, Q ∈ Hs (Ωext
σ ) (7.25)
∧
with Fourier coefficients F (n, m) = (ϕ∧ 2 ∧
j (n)) Q (n, m) leads to
∞ 2n+1

ΛF (x) = Λ∧ (n)F ∧ (n, m)Hn,m
s
(γ; x)
n=0 m=1
∞ 2n+1

= Λ∧ (n)(ϕ∧ 2 ∧ s
j (n)) Q (n, m)Hn,m (γ; x)
n=0 m=1
(2) (2)
= (Φj )γ,γ ∗ (ΛQ) = (Φj )γ,γ ∗ G, (7.26)
where we let G = ΛQ ∈ Hs (Ωext
γ ). Thus, we get the following theorem.
Theorem 7.10. The restriction of the operator Λ : Hs (Ωext ext

σ ) → Hs (Ωγ ) to a scale
ext
space Vj (Ωσ ), j ∈ N0 , i.e.,
Λ|Vj (Ωext ) : Vj (Ωext ext
σ ) → Vj (Ωγ ) (7.27)
σ
is injective. Moreover, we have the following results:

(i) If the families {{ψj (n)}n∈N0 }j∈N0 and {{ψ̃j (n)}n∈N0 }j∈N0 are bandlimited
(for example, ψj (n) = ψ̃j (n) = 0 for all n ≥ 2j ), then the restricted operator
is even bijective. To be more specific, for G ∈ Hs (Ωext γ ) the unique solution
Fj ∈ Vj (Ωext
σ ), j ∈ N 0 , of the equation
(2)
ΛFj = (Φj )γ,γ ∗ G (7.28)
is given by
(2)
Fj = (Φj )σ,σ ∗ Q, (7.29)
where Q ∈ Hs (Ωext
σ ) is given by

G∧ (n,m)
Λ∧ (n) , n ∈ [0, 2j ),
Q∧ (n, m) = (7.30)
0, n ∈ [2j , ∞),
n = 0, 1, . . . ; m = 1, . . . , 2n + 1.
(ii) If the families {{ψj (n)}n∈N0 }j∈N0 and {{ψ̃j (n)}n∈N0 }j∈N0 are not bandlim-
ited, the equation
(2)
ΛFj = (Φj )γ,γ ∗ G (7.31)
Λ Λ
has a solution Fj ∈ Vj (Ωext ext
σ ) provided that G ∈ Hs (Ωσ ), where Hs (Ωσ ) is
ext
a suitable Sobolev space (see the Ph.D.-thesis [58] for a detailed introduction).
In this case, the unique solution of the equation is given by
(2)
Fj = (Φj )σ,σ ∗ Q, (7.32)
σ ) is obtainable by
G∧ (n, m)
Q∧ (n, m) = , (7.33)
Λ∧ (n)
n = 0, 1, . . .; m = 1, . . . , 2n + 1.
We now define the primal wavelets {Ψα,α
j (·, ·)}j∈N0 and the dual wavelets
{Ψ̃α,α
j (·, ·)}j∈N0
for α ≥ σ in the way as we did in the case of the scaling functions
and get
∞
2n+1

Ψα,α
j (x, y) = ψj (n) s
Hn,m s
(α; x)Hn,m (α; y), (7.34)
n=0 m=1
∞
2n+1

Ψ̃α,α
j (x, y) = ψ̃j (n) s
Hn,m s
(α; x)Hn,m (α; y), (7.35)
n=0 m=1
where
∧ ∧
(Ψα,α
j ) (n) = ψj (n), (Ψ̃α,α
j ) (n) = ψ̃j (n). (7.36)
The detail spaces are defined in canonical manner:
Wj (Ωext
α ) = {(Ψj ∗ Ψ̃j )
α,α
∗ F : F ∈ Hs (Ωext
α )}, (7.37)
where
(Ψj ∗ Ψ̃j )α,α = Ψα,α
j ∗ Ψ̃α,α
j . (7.38)
The reconstruction formula given in Theorem 7.9 is valid substituting Ψ̃j ∗ Ψ ∗ F
by (Ψ̃j ∗ Ψ)α,α ∗ F . Theorem 7.10 can now be transferred to the restriction on
detail spaces and we get the following theorem.
Theorem 7.11. The restriction of the operator Λ : Hs (Ωext ext
σ ) → Hs (Ωγ ) to a detail
ext
space Wj (Ωσ ), j ∈ N0 , i.e.,
Λ|Wj (Ωext ) : Wj (Ωext ext
σ ) → Wj (Ωγ ) (7.39)
σ
is injective. Moreover, we have the following results:

(i) If the family {{ϕj (n)}n∈N0 }j∈N0 is bandlimited (for example, ϕj (n) = 0 for
all n ≥ 2j ), then the restricted operator is even bijective. To be more specific,
for G ∈ Hs (Ωext ext
γ ) the unique solution Hj ∈ Wj (Ωσ ), j ∈ N0 , of the equation
ΛHj = (Ψ̃j ∗ Ψj )γ,γ ∗ G (7.40)

is given by
Hj = (Ψ̃j ∗ Ψj )σ,σ ∗ Q, (7.41)
∧
G (n,m)
Λ∧ (n) , n ∈ [0, 2j+1 ),
Q∧ (n, m) = (7.42)
0, n ∈ [2j+1 , ∞),
n ∈ N0 ; m = 1, . . . , 2n + 1.
(ii) If the family {{ϕj (n)}n∈N0 }j∈N0 is non-bandlimited, the equation
ΛHj = (Ψ̃j ∗ Ψj )γ,γ ∗ G (7.43)
has a solution Hj ∈ Wj (Ωextσ ) provided that G ∈ HsΛ (Ωext
σ ),
where HsΛ (Ωext
σ )
is a suitable Sobolev space (cf. the Ph.D.-thesis [58] for a detailed definition).
In this case, the unique solution of the equation is given by
Hj = (Ψ̃j ∗ Ψj )σ,σ ∗ Q, (7.44)
G∧ (n, m)
Q∧ (n, m) = , (7.45)
Λ∧ (n)
n ∈ N0 ; m = 1, . . . , 2n + 1.
Up to now, we have summarized some results about the filtered solution, i.e.,
the solution when we restrict the operator to scale or detail spaces. In the case of
the unfiltered solution, we have the following theorem.
Theorem 7.12. Let G ∈ Hs (Ωext γ ) satisfy the condition G ∈ im(Λ). Then the unique
solution F ∈ Hs (Ωext
σ ) of the equation ΛF = G is given by
G∧ (n, m)
F ∧ (n, m) = , (7.46)
Λ∧ (n)
n ∈ N0 ; m = 1, . . . , 2n + 1.
Examples for scaling functions

To make the preceding considerations more concrete, we would like to show that all
reproducing kernel functions introduced in Section 5 may be used as Hσ,σ -scaling
functions. We essentially follow [19] and distinguish in accordance with Definition
7.2 two cases, viz. (1) bandlimited Hσ,σ -scaling functions and (2) non-bandlimited
(1) Bandlimited Hσ,σ -scaling Functions. Suppose that H(Ωext σ ) is a Sobolev space
(satisfying the consistency conditions (CC1) and (CC2) relative to [σ, σ inf )).
Consider sequences {ϕj (n)}n∈N0 with “local support” (for example, ϕj (n) =
0 for all n ≥ 2j , j ∈ N0 ). Thus all members Φj (·, ·) of an associated Hσ,σ -
scaling function {Φj (·, ·)}j∈N0 with (Φj )∧ (n) = ϕj (n), n ∈ N0 , are band-
limited. This allows to deal with finite-dimensional scale spaces Vj (Ωext σ ).
Consequently, all spaces Vj (Ωext
σ ) have finite-dimensional basis systems.
(1a) Shannon Hσ,σ -scaling function (see Figure 7.1). Consider the family
{{ϕj (n)}n∈N0 }j∈N0
(a) Φ4 (x, y) (b) ϕ4 (n)
Figure 7.1. Shannon Hσ,σ -scaling function for j = 4 and An = 1:

space domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ in sectional represen-
tation (left) and frequency domain, i.e., ϕj (n) (right).
given by
1, n ∈ [0, 2j ),
ϕj (n) =
0, n ∈ [2j , ∞).
The family {{ϕj (n)}n∈N0 }j∈N0 forms a generator of a scaling function in
the sense of Definition 7.2. The Hσ,σ -scaling function {Φj (·, ·)}j∈N0 reads as
follows:
1 2n + 1 σ 2 n+1
x y

Φj (x, y) = Pn · ,
j
A2n 4πσ 2 |x| |y| |x| |y|
n≤2 −1

σ × Ωσ . A remarkable property is that Φj (·, ·) coincides with its
iterations:
(k) (k−1)
Φj (·, ·) = (Φj ∗H Φj )(·, ·), k = 2, 3, . . . .
The scale spaces
5
Vj (Ωext ext
σ ) = Pj (H(Ωσ )) = Harm n (Ωext
σ ), j ∈ N0 ,
n≤2j −1
satisfy the properties:

σ ) ⊂ · · · ⊂ Vj (Ωσ ) ⊂ Vj+1 (Ωσ ) ⊂ · · · ⊂ H(Ωσ ),
; ·H(Ωext )
(ii) Vj (Ωext
σ )
σ
= H(Ωextσ ),
j∈N0
9
(iii) Vj (Ωext
σ ) = Harm 0 (Ωσ ).
ext
j∈N0
The multiresolution analysis is orthogonal. As a matter of fact, the Shannon
“detail spaces” Wj (Ωext ext ext
σ ) = Vj+1 (Ωσ ) Vj (Ωσ ) of different scales j do
not have any common frequencies. Consequently, the orthogonality of the
outer harmonics immediately implies the orthogonality of the Shannon detail
spaces. The scale spaces Vj (Ωext ext
σ ), j ∈ N0 , form an H(Ωσ )-multiresolution
analysis. Apart from this, it can be even verified that the decomposition
of the scale space Vj+1 (Ωext ext
σ ) into the scale space Vj (Ωσ ) and the detail
ext
space Wj (Ωσ ) is orthogonal. This orthogonality of the decomposition easily
follows from the already known fact that
5
Vj+1 (Ωext
σ )= Harm n (Ωext
σ )
0≤n≤2j+1 −1
5 5
= Harm n (Ωext
σ )⊕ Harm n (Ωext
σ )
0≤n≤2j −1 2j ≤n≤2j+1 −1
= Vj (Ωext
σ ) ⊕ Wj (Ωext
σ ). (7.47)
On the one hand, the orthogonal structure of the Shannon multiresolution
analysis seems to be very profitable. On the other hand, it is not surprising
that the Shannon Hσ,σ -scaling function shows strong oscillations. This is the
price to be paid for the sharp separation “in momentum space”. For numerical
purposes it is often advisable to discuss “smoothed versions” of the Shannon
(a) Φ4 (x, y) (b) ϕ4 (n)
Figure 7.2. Smoothed Shannon Hσ,σ -scaling function for j = 4 and

An = 1, h = 12 : space domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ in
sectional representation (left) and frequency domain, i.e., ϕj (n) (right).
kernels. But this automatically implies the loss of the orthogonality in the
multiresolution analysis.
(1b) Smoothed Shannon Hσ,σ -scaling Function (see Figure 7.2). For fixed h ∈ [0, 1)
we now consider the family {{ϕj (n)}n∈N0 }j∈N0 given by
⎧
⎪
⎨ 1, n ∈ [0, 2j h),
−j
1−2 n
1−h , n ∈ [2 h, 2 ),
ϕj (n) = j j
⎪
⎩
0, n ∈ [2 , ∞).
j
The family {{ϕj (n)}n∈N0 }j∈N0 defines a generator of an Hσ,σ -scaling func-
tion. Obviously, {Φj (·, ·)}j∈N0 with (Φj )∧ (n) = ϕj (n) for n, j ∈ N0 is an
Hσ,σ -scaling function. Clearly, for each n ∈ N0 , {ϕj (n)}j∈N0 is monotonously
increasing. The kernels Φj (·, ·): Ωext ext
σ × Ωσ → R read as follows:
2n + 1 ϕj (n) σ 2 n+1
x y

Φj (x, y) = P n · .
j
4πσ 2 A2n |x| |y| |x| |y|
n≤2 −1
The value h ∈ [0, 1) represents a “control parameter” of the smoothing ef-

fect of the Hσ,σ -scaling function {Φj (·, ·)}j∈N0 . The scale spaces Vj (Ωext
σ ),
j ∈ N0 , form an Hσ,σ -multiresolution analysis. This multiresolution analysis,
however, is not orthogonal, since Vj+1 (Ωext ext ext
σ ) = Vj (Ωσ ) + Wj (Ωσ ), j ∈ N0 ,
cannot be understood as orthogonal sum decomposition.
(1c) Cubic Polynomial (CP) Hσ,σ -scaling Function (see Figure 7.3). In order to
gain a higher intensity of the smoothing effect than in the case of the Hσ,σ -
scaling function (1b), we introduce a function ϕ0 : [0, ∞) → R in such a
way that ϕ0 |[0,1] coincides with the uniquely determined cubic polynomial
p : [0, 1] → [0, 1] with the properties:
p(0) = 1, p(1) = 0, p
(0) = 0, p
(1) = 0.
(a) Φ4 (x, y) (b) ϕ4 (n)
Figure 7.3. CP Hσ,σ -scaling function for j = 4 and An = 1: space

domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ in sectional representation
(left) and frequency domain, i.e., ϕj (n) (right).
It is not difficult to see that these properties are fulfilled by

p(t) = (1 − t)2 (1 + 2t), t ∈ [0, 1].
This leads us to a function ϕ0 : [0, ∞) → R given by

(1 − t)2 (1 + 2t), t ∈ [0, 1),
ϕ0 (t) =
0, t ∈ [1, ∞).
It is obvious that ϕ0 is a monotonously decreasing function. In [31] a con-
struction principle of deriving scaling functions from a “mother function”
ϕ0 : [0, ∞) → R by letting ϕj (t) = ϕ0 (2−j t), t ∈ [0, ∞), is described and we
thus define the family {{ϕj }j∈N0 }n∈N0 with ϕj (t) = ϕ0 (2−j t), t ∈ [0, ∞), by

(1 − 2−j t)2 (1 + 2−j+1 t), t ∈ [0, 2j ),
ϕj (t) = ϕ0 (2−j t) =
0, t ∈ [2j , ∞).
{ϕj (n)}j∈N0 is a monotonously increasing sequence for each n ∈ N0 , hence,
{Φj (·, ·)}j∈N0 defines an Hσ,σ -scaling function. The finite-dimensional scale
spaces Vj (Ωext
σ ), j ∈ N0 , represent a non-orthogonal Hσ,σ -multiresolution
analysis.
Finally, it should be remarked that one can think of other ways to “smooth”
the Shannon generator but these are not discussed.
(2) Non-bandlimited Hσ,σ -scaling functions. Next we take a look at non-bandli-
mited generators of scaling functions. In other words, all Hσ,σ -scaling func-
tions {Φj (·, ·)}j∈N0 discussed in the following share the property that their
“generators” {{ϕj (n)}n∈N0 }j∈N0 have a “global support”. Since there are only
a few conditions for a family {{ϕj (n)}n∈N0 }j∈N0 to generate an Hσ,σ -scaling
function, there are various possibilities for its concrete realization. In our ap-
proach we concentrate on three types: Tikhonov, rational, and exponential
(a) Φ5 (x, y) (b) ϕ5 (n)
(c) Φ7 (x, y) (d) ϕ7 (n)
Figure 7.4. Tikhonov Hσ,σ -scaling function for j = 5 (above) and j =

7 (below) and An = 1: space domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ
in sectional representation (left) and frequency domain, i.e., ϕj (n)
(right).
(2a) Tikhonov Hσ,σ -scaling Function (see Figure 7.4). Consider the family
{{ϕj (n)}n∈N0 }j∈N0
given by

1, n = 0,
ϕj (n) = 2
1/2 (7.48)
τn
2 +(2−j )2
τn , n = 1, 2, . . . ,
where the sequence {τn }n∈N0 with τn = 0 for all n ∈ N0 is given in such a
way that
∞
∞ 2
2 τn
(i) τn < ∞ , (ii) (2n + 1) < ∞.
n=0 n=0
An
It is not hard to see that the family {Φj (·, ·)}j∈N0 constitutes an Hσ,σ -scaling
function. The Tikhonov Hσ,σ -scaling function plays an important role in the
theory of regularization wavelets.
(a) Φ4 (x, y) (b) ϕ4 (n)
Figure 7.5. Rational Hσ,σ -scaling function for j = 4 and An = 1,

τ = 5: space domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ in sectional
representation (left) and frequency domain, i.e., ϕj (n) (right).
(2b) Rational Hσ,σ -scaling Functions (see Figure 7.5). Consider ϕj : [0, ∞) → R
given by
ϕj (t) = (1 + 2−j t)−τ , t ∈ [0, ∞), τ > 1. (7.49)
Clearly, for all values τ > 1, the family {{ϕj (n)}n∈N0 }j∈N0 forms a generator
of a scaling function. All functions ϕj , j ∈ N0 , define admissible Hσ,σ -kernels
Φj (·, ·), j ∈ N0 , if, in addition, τ > 1 is chosen in such a way that
∞
(1 + 2−j n)−2τ
(2n + 1) <∞ (7.50)
n=0
A2n
1 s
n+ 2
for j ∈ N0 . For example, in the case of Hs (Ωext σ ), i.e., An = σ for
n = 0, 1, . . ., we find s + τ > 1 to satisfy the estimate (7.50). More generally,
(1 + n)−2τ A−2 n = O(n
−2−ε
) for n → ∞ with ε > 0 together with τ > 1 is a
sufficient condition to define an admissible Hσ,σ -kernel Φj (·, ·), j ∈ N0 . The
Hσ,σ -scaling function {Φj (·, ·)}j∈N0 consists of the kernels
∞ 2 n+1
(1 + 2−j n)−τ 2n + 1 σ x y
Φj (x, y) = Pn · ,
n=0
A2n 4πσ 2 |x| |y| |x| |y|

σ × Ωσ . The functions ϕj , j ∈ N0 , are monotonously decreasing
on the interval [0, ∞) for all values τ > 1 and all j ∈ N0 . Therefore, the scale
spaces Vj (Ωext
σ ) form an Hσ,σ -multiresolution analysis provided that both
τ > 1 and the summability condition (7.50) is valid.
(2c) Exponential Hσ,σ -scaling Functions (see Figures 7.6 and 7.7). Choose ϕj :
[0, ∞) → R, j ∈ N0 , to be defined by
−j
ϕj (t) = e−2 H(t)
, t ∈ [0, ∞), (7.51)
(a) Φ4 (x, y) (b) ϕ4 (n)
Figure 7.6. Abel–Poisson Hσ,σ -scaling function for j = 4 and An = 1,

τ = 1: space domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ in sectional
representation (left) and frequency domain, i.e., ϕj (n) (right).
(a) Φ4 (x, y) (b) ϕ4 (n)
Figure 7.7. Gauss–Weierstraß Hσ,σ -scaling function for j = 4 and

An = 1, τ = 1: space domain, i.e., Φj (x, y) for (x, y) ∈ Ωσ × Ωσ in
sectional representation (left) and frequency domain, i.e., ϕj (n) (right).
where H : [0, ∞) → [0, ∞) satisfies the properties:

– H ∈ C (∞) [0, ∞),
– H(0) = 0,
– H(t) > 0 for t > 0,
– H(t) < H(t
) whenever 0 < t < t
.
The sequence {ϕj (n)}j∈N0 is monotonously increasing for each n ∈ N0 . The
functions ϕj , j ∈ N0 , define an Hσ,σ -scaling function {Φj (·, ·)}j∈N0 by letting
(Φj )∧ (n) = ϕj (n), n ∈ N0 , provided that Φj (·, ·), j ∈ N0 , are admissible
Hσ,σ -kernel functions. It is not hard to see that
(Φj ∗ Φj )(x, y)
∞ −j 2 n+1
(e−2 H(n) )2 2n + 1 σ x y
= Pn ·
n=0
A2n 4πσ 2 |x| |y| |x| |y|
∞ −(j−1) 2 n+1
e−2 H(n)
2n + 1 σ x y
= Pn ·
n=0
A2n 4πσ 2 |x| |y| |x| |y|
= Φj−1 (x, y) (7.52)
holds for all j ∈ N and all (x, y) ∈ Ωext ext
σ × Ωσ . The scale spaces Vj (Ωσ )
ext
constitute an Hσ,σ -multiresolution analysis. Altogether we find the following

result for exponential Hσ,σ -scaling functions: The family {Pj }j∈N0 of opera-
(2)
tors Pj : H(Ωext ext
σ ) → H(Ωσ ) (defined by Pj (F ) = Φj ∗ F , F ∈ H(Ωσ ))
ext
forms an Hσ,σ -contracting approximate identity (called the exponential Hσ,σ -

contracting approximate identity), i.e., the following properties are satisfied:
(i) Pj is a bounded linear operator for every j ∈ N0 and P∞ = I (identity),
(ii) Pj−1 = Pj Pj for all j ∈ N0 ,
(iii) limj→∞ F − Pj (F ) H(Ωext ) = 0 for all F ∈ H(Ωext
σ ),
σ
(iv) Pj (F ) H(Ωext ) ≤ F H(Ωext ) for all j ∈ N0 , F ∈ H(Ωext

σ ).
σ σ
As examples we mention the Abel–Poisson Hσ,σ -contracting approxi-

mate identity given by H(t) = αt, α > 0, and the Gauss–Weierstraß Hσ,σ -
contracting approximate identity given by H(t) = αt(t + 1), α > 0.
Remark 7.13. Non-bandlimited scaling functions become bandlimited ones by suit-
able truncation in momentum space. To be more specific, if {Φj (·, ·)}j∈N0 is a
non-bandlimited Hσ,σ -scaling function, then {Γj (·, ·)}j∈N0 given by (Γj )∧ (n) =
(Φj )∧ (n) for n ∈ [0, 2j ) and (Γj )∧ (n) = 0 for n ∈ [2j , ∞) represents a bandlimited
Hσ,σ -scaling function.
We now explain the connection between the solution in the scale spaces and
the unfiltered solution.
Theorem 7.14. Suppose that G is of class HsΛ (Ωext ext
γ ). Let F ∈ Hs (Ωσ ) be the
unique solution of ΛF = G. Then
(2)
Fj = (Φj )σ,σ ∗ F (7.53)
is the unique solution in Vj (Ωext
σ ) of the equation
(2)
ΛFj = (Φj )γ,γ ∗ G (7.54)
for every j ∈ N0 . Furthermore, the limit relation
(2)
lim (ΦJ )σ,σ ∗ F = F (7.55)
J→∞
holds (in · Hs (Ωext ) -sense).

σ
In the case of bandlimited scaling functions, the preceding theorem shows

that the (scalar) SST or SGG problem is well posed: A unique solution always
exists and due to the finite dimension of the scale spaces the solution is also
stable. According to the multiscale approach the solution in the scale space is
given by adding the solution of the corresponding detail spaces to the solution of
the scale space of a lower scale. Because of the limit relation given in Theorem
7.14 the filtered solutions converge to the unfiltered solution in the Sobolev space
Hs (Ωext
σ ). If we now turn to non-bandlimited scaling functions, the stability of
the solution cannot be ensured, because the (scalar) SST or SGG problem is an
exponentially ill-posed problem with unbounded inverse operator Λ−1 . In order
to obtain a well-posed problem, we have to replace the inverse operator by an
appropriate bounded operator, that is we have to use a regularization of Λ−1 .
Definition 7.15. A family of linear operators Sj : Hs (Ωext ext
γ ) → Hs (Ωσ ), j ∈ N0 , is
−1
called a regularization of Λ , if it satisfies the following properties:
(i) Sj is bounded on Hs (Ωext
γ ) for all j ∈ N0 ,
(ii) for any member G ∈ im(Λ), the limit relation lim SJ G = Λ−1 G holds (in
J→∞
· Hs (Ωext ) -sense).
σ
The function FJ = SJ G is called the J-level regularization of Λ−1 G. In our

approach we want to represent the J-level regularization with the help of harmonic
wavelets which guarantees that we can calculate the J + 1-level regularization by
adding the corresponding detail information to the J-level regularization. In order
to formulate the multiscale regularization concept, we start with the definition of
a generator of a regularization scaling function by modifying Definition 7.2.
generator of a regularization scaling function with respect to Λ−1 , if it satisfies the
following requirements:
(i) (ϕj (0))2 = Λ∧1(0) , for all j ∈ N0 ,
(ii) (ϕj (n))2 ≤ (ϕj (n))2 , for all j, j
∈ N0 with j ≤ j
and all n ∈ N,
(iii) lim (ϕj (n))2 = Λ∧1(n) , for all n ∈ N.
j→∞
Now we are able to define the decomposition and reconstruction regulariza-

tion scaling functions in such a way that the corresponding convolutions lead to
the J-level approximation of Λ−1 G, G ∈ im(Λ).
Definition 7.17. Let {{ϕj (n)}n∈N0 }j∈N0 be a generator of a regularization scaling
σ,γ
function with respect to Λ−1 . Then a family {d Φj (·, ·)}j∈N0 of admissible Hσ,γ -
kernels given by
∞
2n+1

d σ,γ s s
Φj (x, z) = ϕj (n) Hn,m (σ; x)Hn,m (γ; z), (7.56)
n=0 m=1
(x, z) ∈ Ωext ext

σ × Ωγ , is called a decomposition regularization Hσ,γ -scaling function
with respect to Λ−1 , whereas a family {r Φσ,σ
j (·, ·)}j∈N0 of admissible Hσ,σ -kernels
given by
∞ 2n+1

r σ,σ s s
Φj (x, y) = ϕj (n) Hn,m (σ; x)Hn,m (σ; y), (7.57)
n=0 m=1

σ × Ωσ is called a reconstruction regularization Hσ,σ -scaling function
with respect to Λ−1 .
Obviously, the regularization scaling functions fulfill
σ,γ
d
Φj (x, ·) ∈ Hs (Ωext
γ ), x ∈ Ωext
σ , j ∈ N0 , (7.58)
r
Φσ,σ
j (x, ·) ∈ Hs (Ωext
σ ), x ∈ Ωext
σ , j ∈ N0 . (7.59)
As already stated, we obtain the following theorem:
Theorem 7.18. Let {{ϕj (n)}n∈N0 }j∈N0 be a generator of a regularization scal-
ing function with respect to Λ−1 . If we define the admissible Hσ,γ -kernel (r Φj ∗
d
Φj )σ,γ (·, ·) by
σ,γ
(r Φj ∗ d Φj )σ,γ (x, z) = r Φσ,σ
j (x, ·) ∗ Φj (·, z),
d
(7.60)

σ × Ωγ , then
FJ = (r ΦJ ∗ d ΦJ )σ,γ ∗ G, G ∈ Hs (Ωext
γ ), (7.61)
represents the J-level regularization of Λ−1 G.
If, in addition, G ∈ im(Λ) = HsΛ (Ωext
γ ), then
lim FJ − Λ−1 G Hs (Ωext ) = 0. (7.62)

J→∞ σ
If we define the convolution operators SJ : Hs (Ωext ext

γ ) → Hs (Ωσ ), J ∈ N0 , by
SJ (G) = (r ΦJ ∗ d ΦJ )σ,γ ∗ G, (7.63)

and introduce the scale spacesSJ (im(Λ)) as follows
SJ (im(Λ)) = {(r ΦJ ∗ d ΦJ )σ,γ ∗ G : G ∈ im(Λ)}, (7.64)
the following theorem holds.
Theorem 7.19. The scale spaces satisfy the following properties:

(i) S0 (im(Λ)) ⊂ · · · ⊂ SJ (im(Λ)) ⊂ SJ (im(Λ)) ⊂ Hs (Ωext

σ ), J ≤ J , i.e., for any
right-hand side G ∈ im(Λ) of the (scalar) SST or SGG problem, all J-level
regularizations with fixed parameter J are sampled in a scale space SJ (im(Λ))
with the above property,
;∞ · ext
(ii) J=0 SJ (im(Λ)) Hs (Ωσ ) = Hs (Ωext σ ).
A set of subspaces of Hs (Ωext

σ ) satisfying the conditions of Theorem 7.19 is
called regularization Hσ,γ -multiresolution analysis (RMRA) of the (scalar) SST or
SGG problem.
We now turn to the definition of regularization wavelets following the proce-
dure described in the case of regularization scaling functions. Obviously, we have
to define decomposition and reconstruction regularization wavelets.
Definition 7.20. Let {{ϕj (n)}n∈N0 }j∈N0 be a generator of a regularization scal-

ing function with respect to Λ−1 . Then the generating symbols {ψj (n)}n∈N0 and
{ψ̃j (n)}n∈N0 of the corresponding regularization wavelets are defined by the re-
σ,γ
finement equation (7.10). The admissible Hσ,γ -kernel {d Ψj (·, ·)}j∈N0 given by
∞
2n+1

d σ,γ s s
Ψj (x, z) = ψj (n) Hn,m (σ; x)Hn,m (γ; z), (7.65)
n=0 m=1
(x, z) ∈ Ωext
σ × Ωγ
ext is called the decomposition regularization H
σ,γ -wavelet, while
r σ,σ
the admissible Hσ,σ -kernel { Ψ̃j (·, ·)}j∈N0 given by
∞
2n+1

r σ,σ s s
Ψ̃j (x, y) = ψ̃j (n) Hn,m (σ; x)Hn,m (σ; y), (7.66)
n=0 m=1

σ × Ωσ is called the reconstruction regularization Hσ,σ -wavelet.
We now define the convolution operators Tj : Hs (Ωext ext

γ ) → Hs (Ωσ ), j ∈
N0 , by
Tj (G) = (r Ψ̃j ∗ d Ψj )σ,γ ∗ G, G ∈ Hs (Ωext
γ ). (7.67)
Obviously, due to the refinement equation, the operator SJ+1 can be represented
in the form
J
SJ+1 = S0 + Tj . (7.68)
j=0
Thus, we now introduce the detail spaces TJ (im(Λ)) by

TJ (im(Λ)) = (r Ψ̃J ∗ d ΨJ )σ,γ ∗ G : G ∈ im(Λ) . (7.69)
The space TJ (im(Λ)) contains the detail information which has to be added in
order to turn from the J-level regularization to the J + 1-level regularization:
SJ+1 (im(Λ)) = SJ (im(Λ)) + TJ (im(Λ)). (7.70)
In general, the sum is neither direct nor orthogonal.
Theorem 7.21. Let {{ϕj (n)}n∈N0 }j∈N0 be a generator of a regularization scaling
function with respect to Λ−1 . Suppose that {{ψj (n)}n∈N0 }j∈N0 , {{ψ̃j (n)}n∈N0 }j∈N0
more, let G be of class Hs (Ωext
γ ). Define the regularization Hσ,γ -wavelet transform
at scale j ∈ N0 and position x ∈ Ωext
σ by
σ,γ
(RWT)(G)(j; x) = d ΨJ (x, ·) ∗ G, G ∈ Hs (Ωext
γ ). (7.71)
Then
r σ,γ
J−1
σ,σ
FJ = Φ0 ∗ Φ0
d
∗G+ r
Ψ̃J ∗ (RW T )(G)(j; ·) (7.72)
j=0
is the J-level regularization of the (scalar) SST or SGG problem satisfying

lim FJ − Λ−1 G Hs (Ωext ) = 0 (7.73)
J→∞ σ
provided that G ∈ im(Λ) = HsΛ (Ωext

γ ).
Some examples of regularization wavelets and numerical calculations can be

found in [19], where, in addition, all the above-mentioned theorems are proved.
7.2. Vectorial wavelet theory

We now give the extension of the scalar wavelet theory to the vectorial case. First
we define vectorial scaling functions and wavelets. The reconstruction formula is
the main result stating how the function can be split into a lowpass part and an
infinite sum of bandpass parts. Then we solve the (vectorial) SST or SGG problem
(i) (i)
defining regularization wavelets. We use the notation Φ̂j Φ̂j ∗ f instead of
(i) (i) (i) (i)
Φ̂j (Φ̂j ∗ f ), and Φ̂j Φ̂j ∗ f = 3i=1 Φ̂j Φ̂j ∗ f (i) .
(i) (i)
Definition 7.22. Let {Φ̂j (·, ·)}j∈N0 be a family of admissible hσ,σ -kernels, i ∈
(i) (i)
{1, 2, 3}. Then the family {Vj (Ωext ext
σ )}j∈N0 of scale spaces Vj (Ωσ ) defined by
(i) (i) (i)

Vj (Ωext ext
σ ) = {Φ̂j Φ̂j ∗ f : f ∈ hs (Ωσ )}, (7.74)
(i)
is called an hσ,σ -multiresolution analysis, if the following properties are satisfied:
(i) (i) (i) (i)
σ ) ⊂ · · · ⊂ Vj (Ωσ ) ⊂ Vj+1 (Ωσ ) ⊂ · · · ⊂ hs (Ωσ ),
; (i) ext hs (Ωext
·
σ ) (i)
(ii) Vj (Ωσ ) = hs (Ωext
σ ).
j∈N0
Definition 7.23. Let {Φ̂j (·, ·)}j∈N0 be a family of admissible hσ,σ -kernels. The set
of scale spaces Vj (Ωext
σ ) defined by
Vj (Ωext ext
σ ) = {Φ̂j Φ̂j ∗ f : f ∈ hs (Ωσ )} (7.75)
(i) (i)
is called an hσ,σ -multiresolution analysis, if {Vj (Ωext ext
σ )}j∈N0 is an hs (Ωσ )-multi-
resolution analysis for every i ∈ {1, 2, 3}.
Our next purpose is to define scaling functions.
(i) (i)
Definition 7.24. A family {{ϕj (n)}n∈N0i }j∈N0 of sequences {ϕj (n)}n∈N0i is called
a generator of a scaling function of kind i, i ∈ {1, 2, 3}, if it satisfies the following
requirements:
(i)
(i) (ϕj (0i ))2 = 1 for all j ∈ N0 ,
2
(i) (i)
(ii) (ϕj (n))2 ≤ ϕj (n) for all j, j
∈ N0 with j ≤ j
and all n ∈ N0i +1 ,
(i)
(iii) lim (ϕj (n))2 = 1 for all n ∈ N0i +1 .
j→∞
Furthermore, the family {{{ϕ(i) (n)}i∈{1,2,3} }n∈N0i +1 }j∈N0 is called a generator of

a scaling function, if {{ϕ(i) (n)}n∈N0i }j∈N0 are generators of a scaling function of
kind i, i ∈ {1, 2, 3}.

hσ,σ -scaling functions.
(i) (i)
Definition 7.25. A family {Φ̂j (·, ·)}j∈N0 of h(i) -kernels Φ̂j (·, ·) defined by
(i)∧ (i)
Φ̂j (n) = ϕj (n), j ∈ N0 , n ∈ N0i , i ∈ {1, 2, 3},
i.e.,
∞
2n+1

(i) (i)
Φ̂j (x, y) = ϕj (n) s
Hn,m (σ; x)h(i)s
n,m (σ; y), x, y ∈ Ωext
σ , (7.76)
n=0i m=1
(i)
is called an hσ,σ -scaling function, if it satisfies the following properties:
(i) (i)
(i) Φ̂j (·, ·) is an admissible hσ,σ -kernel for every j ∈ N0 ,
(i)∧
(ii) {{Φ̂j (n)}n∈N0i }j∈N0 constitutes a generator of a scaling function of kind i.
Furthermore, the family {Φ̂j (·, ·)}j∈N0 of hσ,σ -kernels Φ̂j (·, ·) is called an hσ,σ -
(i) (i)
scaling function, if {Φ̂j }j∈N0 are hσ,σ -scaling functions for i ∈ {1, 2, 3}.
The following approximation property can be derived.
Theorem 7.26. Let {Φ̂j (·, ·)}j∈N0 be an hσ,σ -scaling function. Then
lim f − Φ̂j Φ̂j ∗ f hs (Ωext ) = 0 (7.77)

j→∞ σ
holds for all f ∈ hs (Ωext

σ ).
Definition 7.27. Let {Φ̂j (·, ·)}j∈N0 be an hσ,σ -scaling function. Then {Pj }j∈N0 with
Pj : hs (Ωext ext
σ ) → hs (Ωσ ) defined by
Pj (f )(x) = Φ̂j Φ̂j ∗ f, f ∈ hs (Ωext ext

σ ), x ∈ Ωσ , (7.78)
is called an hσ,σ -approximate identity.
The kernel Φ̂0 is called the mother kernel of the hσ,σ -scaling function.
Theorem 7.28. Let {Φ̂j (·, ·)}j∈N0 be an hσ,σ -scaling function. Then {Vj (Ωext
σ )}j∈N0
defined in (7.75) forms an hσ,σ -multiresolution analysis.
We are now at the point to define the (primal/dual) wavelet with the help of
the bilinear refinement equation.
Definition 7.29. Let {Φ̂j (·, ·)}j∈N0 be an hσ,σ -scaling function. Then the families
˜
of hσ,σ -kernels {Ψ̂j (·, ·)}j∈N0 , {Ψ̂j (·, ·)}j∈N0 given by
(i)
(Ψ̂j )(i)∧ (n) = ψj (n), j ∈ N0 , n ∈ N0i , i ∈ {1, 2, 3}, (7.79)
˜ (i)
(Ψ̂j )(i)∧ (n) = ψ̃j (n), j ∈ N0 , n ∈ N0i , i ∈ {1, 2, 3}, (7.80)
are called (primal) hσ,σ -wavelet and dual hσ,σ -wavelet, respectively, if all hσ,σ -ker-
˜ (i) (i)
nels Ψ̂j (·,·), Ψ̂j (·, ·), j ∈ N0 , are admissible and the symbols {ψj (n)}, {ψ̃j (n)},
in addition, satisfy the (vectorial) refinement equation
(i) (i) (i) (i)
ψ̃j (n)ψj (n) = (ϕj+1 (n))2 − (ϕj (n))2 (7.81)
for all j ∈ N0 , n ∈ N0i , i ∈ {1, 2, 3}.
The following equation can directly be seen:
(i) (i)

J
(i) (i)
(ϕJ+1 (n))2 = (ϕ0 (n))2 + ψ̃j (n)ψj (n), J ∈ N0 , (7.82)
j=0
(i)
for all n ∈ N0i . We now define the wavelet transform. To this end we let ψ−1 (n) =
(i) (i) ˜
ψ̃−1 (n) = ϕ0 (n) and Ψ̂−1 (·, ·) = Ψ̂−1 (·, ·) = Φ̂0 (·, ·) for n ∈ N0i , i ∈ {1, 2, 3}.
We remember that we have already defined the space Hs (N−1 × Ωext σ ) (see Eqs.
(7.12)–(7.14))
Definition 7.30. Let {Ψ̂j (·, ·)}j∈N−1 be a (primal) hσ,σ -wavelet. Then (W T )(i) :
hs (Ωext ext
σ ) → Hs (N−1 × Ωσ ) defined by
(i)
(W T )(i) (f )(j; y) = (Ψ̂j ∗ f )(y) (7.83)
is called hσ,σ -wavelet transform of kind i of f at position y ∈ Ωext

σ and scale
j ∈ N−1 .
(i)
As usual, we define the detail space Wj (Ωext
σ ) at scale j by
(i) ! ˜ (i) (i) "

Wj (Ωext ext
σ ) = Ψ̂j Ψ̂j ∗ f : f ∈ hs (Ωσ ) , (7.84)
and !˜ "
Wj (Ωext ext
σ ) = Ψ̂j Ψ̂j ∗ f : f ∈ hs (Ωσ ) . (7.85)
Theorem 7.31 (Vectorial Reconstruction Formula for the Outer Space). Let the
˜
families {Ψ̂j (·, ·)}j∈N0 and {Ψ̂j (·, ·)}j∈N0 , respectively, be a (primal) hσ,σ -wavelet
and its dual corresponding to an hσ,σ -scaling function {Φ̂j (·, ·)}j∈N0 . Then
∞
˜
f= Ψ̂j Ψ̂j ∗ f (7.86)
j=−1
σ ) (in · hs (Ωext ) -sense).

σ
Our next purpose is to solve the (vectorial) SST or SGG problem with the
(i)
help of bandlimited harmonic wavelets. First, we transfer the theory of hσ,σ -scaling
(i) (i),α,α
functions to the case of hα,α -scaling functions Φ̂j with α ≥ σ:
∞
2n+1

(i),α,α (i)
Φ̂j (x, y) = ϕj (n) s
Hn,m (α; x)h(i)s
n,m (α; y), (7.87)
n=0i m=1
where
(i),α,α (i)
(Φ̂j )∧ (n) = ϕj (n). (7.88)
Obviously, Theorem 7.26 can be directly transferred substituting σ by α. The scale
spaces are defined in the following way:
(i) (i),α,α (i),α,α
Vj (Ωext
α ) = {Φ̂j Φ̂j ∗ f : f ∈ hs (Ωext
α )}. (7.89)
(i)
α )} of scale spaces forms a multiresolution analysis.
(i)
Theorem 7.32. The restriction of the operator λ(i) : Hs (Ωext ext
σ ) → hs (Ωγ ) to a
scale space Vj (Ωext
σ ), j ∈ N0 , i.e.,
(i)
λ(i) |Vj (Ωext ) : Vj (Ωext ext
σ ) → Vj (Ωγ ), (7.90)
σ
is injective for i = 1, whereas in the case of i ∈ {2, 3} the Fourier coefficient

of degree 0 cannot be recovered and the Fourier coefficients of degree n ≥ 1 are
uniquely defined. Moreover, we have the following results:
(i)
(i) If the families {{ϕj (n)}n∈N0i }j∈N0 , i ∈ {1, 2, 3}, and {{ϕj (n)}n∈N0 }j∈N0
(i)
are bandlimited (for example, ϕj (n) = ϕj (n) = 0 for all n ≥ 2j ), then
the restricted operator is even bijective (in the sense described above). To be
(i)
more specific, for g (i) ∈ hs (Ωext
γ ) the (in the case of i = 2, 3 up to Fourier
coefficients of degree 0) unique solution Fj ∈ Vj (Ωext
σ ), j ∈ N0 , of the equation
(i),γ,γ (i),γ,γ
λ(i) Fj = Φ̂j Φ̂j ∗ g (i) (7.91)
is given by
Fj = Φσ,σ
j ∗ Φσ,σ
j ∗ Q, (7.92)
(i)∧
g (n,m)
, n ∈ [0i , 2j ),
Q∧ (n, m) = λ(i)∧ (n) (7.93)
0, n ∈ [2j , ∞).
(i)
(ii) If the families {{ϕj (n)}n∈N0i }j∈N0 , i ∈ {1, 2, 3}, and {{ϕj (n)}n∈N0 }j∈N0 are
non-bandlimited, the equation
(i),γ,γ (i),γ,γ
λ(i) Fj = Φ̂j Φ̂j ∗ g (i) (7.94)
(i) (i)Λ (i)Λ
has a solution Fj ∈ Vj (Ωext
σ ) provided that g ∈ hs (Ωext
γ ), where hs (Ωext
γ )
is a suitable Sobolev space (see the Ph.D.-thesis [58] for more details). In this
case, the (in the case of i = 2, 3 up to Fourier coefficients of degree 0) unique

solution is given by
Fj = Φσ,σ
j ∗ Φσ,σ
j ∗ Q, (7.95)
σ ) is obtainable in spectral language by
g (i)∧ (n, m)
Q∧ (n, m) = , (7.96)
λ(i)∧ (n)
n = 0i , . . . ; m = 1, . . . , 2n + 1.
3 The (i)
following corollary shows that in the case of general operators λ =
i=1 λ we have to claim an additional assumption onto the function g.
3 (i)
Corollary 7.33. The restriction of the operator λ = i=1 λ to a scale space
ext
Vj (Ωσ ), j ∈ N0 , i.e.,
3
5 (i)
λ|Vj (Ωext ) : Vj (Ωext
σ )→ Vj (Ωext
γ ) (7.97)
σ
i=1
(i)
has, in general, no solution. Under the assumption ϕj (n) = ϕj (n), i ∈ {1, 2, 3},
we have to claim, in addition, that
g (i)∧ (n, m) g (l)∧ (n, m)
= , (7.98)
λ(i)∧ (n) λ(l)∧ (n)
with i, l ∈ {1, 2, 3}; n = max (0i , 0l ), . . .; m = 1, . . . , 2n + 1.
i,l∈{1,2,3}
Then the results in Theorem 7.32 can directly be transferred.
Note that according to Theorem 7.32 the restriction of a pseudodifferential
operator of kind i to a scale space Vj (Ωext ) is injective. Therefore, in the case of
3 σ
a pseudodifferential operator λ = i=1 λ(i) each pseudodifferential operator λ(i)
leads to a unique solution. The additional assumption (7.98) is thus necessary, in
order to guarantee that the pseudodifferential operators of kind i do not lead to
different solutions.
With the help of the refinement equation (7.81) we now define the primal
(i),α,α ˜ (i),α,α
wavelets {Ψ̂j (·, ·)}j∈N0 and the dual wavelets {Ψ̂j (·, ·)}j∈N0 for α ≥ σ,
i ∈ {1, 2, 3}:
∞
2n+1

(i),α,α (i)
Ψ̂j (x, y) = ψj (n) s
Hn,m (α; x)h(i)s
n,m (α; y), (7.99)
n=0i m=1
∞
2n+1

˜ (i),α,α (i)
Ψ̂j (x, y) = ψ̃j (n) s
Hn,m (α; x)h(i)s
n,m (α; y), (7.100)
n=0i m=1
where
(i),α,α ∧ (i) ˜ (i),α,α ∧ (i)
(Ψ̂j ) (n) = ψj (n), (Ψ̂j ) (n) = ψ̃j (n). (7.101)

(i) (i),α,α ˜ (i),α,α
Wj (Ωext
α ) = {Ψ̂j Ψ̂j ∗ f : f ∈ hs (Ωext
α )}. (7.102)
Theorem 7.31 can be directly transferred by substituting the convolutions
with respect to the sphere Ωσ by the corresponding convolutions with respect to
the sphere Ωα . We now transfer Theorem 7.32 to the case of the detail spaces
and get the following theorem, where we use the terms injectivity, bijectivity, and
uniqueness in the same sense as before (i.e., up to Fourier coefficients of degree 0
in the case of i = 2, 3).
(i)
Theorem 7.34. The restriction of the operator λ(i) : Hs (Ωext ext
σ ) → hs (Ωγ ) to a
detail space Wj (Ωext
σ ), j ∈ N0 , i.e.,
(i)
λ(i) |Wj (Ωext ) : Wj (Ωext ext
σ ) → Wj (Ωγ ) (7.103)
σ
(i)
with ψj (n) = ψj (n) is injective. Moreover, we have the following results:
(i)
(i) If the families {{ϕj (n)}n∈N0i }j∈N0 , i ∈ {1, 2, 3}, and {{ϕj (n)}n∈N0 }j∈N0
(i)
are bandlimited (for example, ϕj (n) = ϕj (n) = 0 for all n ≥ 2j ), then the
(i)
restricted operator is even bijective. To be more specific, for g (i) ∈ hs (Ωext
γ )
the unique solution Hj ∈ Wj (Ωextσ ), j ∈ N 0 , of the equation
˜ (i),γ,γ (i),γ,γ
λ(i) Hj = Ψ̂j Ψ̂j ∗ g (i) (7.104)
is given by
Hj = Ψ̃σ,σ
j ∗ Ψσ,σ
j ∗ Q, (7.105)
(i)∧
g (n,m)
, n ∈ [0i , 2j+1 ),
Q∧ (n, m) = λ(i)∧ (n) (7.106)
0, n ∈ [2j+1 , ∞).
(i)
(ii) If the families {{ϕj (n)}n∈N0i }j∈N0 , i ∈ {1, 2, 3}, and {{ϕj (n)}n∈N0 }j∈N0 are
non-bandlimited, the equation
˜ (i),γ,γ (i),γ,γ
λ(i) Hj = Ψ̂j Ψ̂j ∗ g (i) (7.107)
has a solution Hj ∈ Wj (Ωext

σ ) provided that the condition
∞ 2n+1
g (i)∧ (n, m)
<∞ (7.108)
n=0 m=1
λ(i)∧ (n)
i
(i)
is satisfied for g (i) ∈ hs (Ωext
γ ). In this case, the unique solution of the equa-
tion is given by
Hj = Ψ̃σ,σ
j ∗ Ψσ,σ
j ∗ Q, (7.109)
g (i)∧ (n, m)
Q∧ (n, m) = , (7.110)
λ(i)∧ (n)
n = 0i , . . .; m = 1, . . . , 2n + 1.
3
Corollary 7.35. The restriction of the operator λ = i=1 λ(i) to a detail space
Wj (Ωext
σ ), j ∈ N0 , i.e.,
3
5 (i)
λ|Wj (Ωext ) : Wj (Ωext
σ )→ Wj (Ωext
γ ) (7.111)
σ
i=1
(i) (i)
has, in general, no solution. Under the assumption ψj (n) = ψj (n) and ψ̃j (n) =
ψ̃j (n), i ∈ {1, 2, 3}, we have to claim, in addition, that
g (i)∧ (n, m) g (l)∧ (n, m)
= , (7.112)
λ(i)∧ (n) λ(l)∧ (n)
with i, l ∈ {1, 2, 3}; n = max (0i , 0l ), . . .; m = 1, . . . , 2n + 1.
i,l∈{1,2,3}
Then the results in Theorem 7.34 can be directly transferred.
the solution when we restrict the operator to scale or detail spaces. In this case,
we have injectivity (in the case of i = 2, 3 up to Fourier coefficients of degree 0)
3
for the operators λ(i) , whereas in the case of general operators λ = i=1 λ(i) we
have to claim that (7.98) is valid. In the case of the unfiltered solution, we obtain
the following theorem.
(i)
Theorem 7.36. Let g (i) ∈ hs (Ωext
γ ) satisfy the condition g
(i)
∈ im(λ(i) ), i ∈ {1,2,3}.
Then the unique solution F ∈ Hs (Ωext σ ) (in the case of i = 2, 3 up to Fourier
coefficients of degree 0) of the equation λ(i) F = g (i) is given by
g (i)∧ (n, m)
F ∧ (n, m) = , (7.113)
λ(i)∧ (n)
3
n = 0i , . . .; m = 1, . . . , 2n + 1. In the case of the operator λ = i=1 λ(i) we have
to claim, in addition, that (7.112) holds in order to guarantee the solvability.
Last, we explain the connection between the solution in the scale spaces and
(i)Λ
Theorem 7.37. Suppose that g (i) is of the class hs (Ωext ext
γ ). Let F ∈ Hs (Ωσ ) be
the unique (in the case of i = 2, 3 up to Fourier coefficients of degree 0) solution
of λ(i) F = g (i) . Then
(2)
Fj = (Φj )σ,σ ∗ F (7.114)
(i),γ,γ (i),γ,γ
λ(i) Fj = Φ̂j Φ̂j ∗ g (i) (7.115)

(2)
lim (ΦJ )σ,σ ∗ F = F (7.116)
J→∞

σ
The preceding theorem shows that in the case of bandlimited scaling functions
the (vectorial) SST or SGG problem is well posed, because a unique solution always
exists and due to the finite dimension of the scale spaces the solution is also stable.
We now investigate the case of non-bandlimited scaling functions and it turns out
that the stability cannot be ensured. The reason is that the (vectorial) SST or SGG
problem is an exponentially ill-posed problem with unbounded inverse operator
λ−1 . Therefore, we have to turn to regularization methods and replace the inverse
operator by an appropriate bounded operator.
(i) (i)
Definition 7.38. A family of linear operators Sj : hs (Ωext ext
γ ) → Hs (Ωσ ), j ∈ N0 ,
is called a regularization of (λ(i) )−1 , i ∈ {1, 2, 3}, if it satisfies the following prop-
erties:
(i) (i)
(i) Sj is bounded on hs (Ωext γ ) for all j ∈ N0 ,
(ii) for any member g ∈ im(λ(i) ), the limit relation
(i)
(i)
lim SJ g (i) = (λ(i) )−1 g (i) (7.117)
J→∞

σ
(i)
The operator S : hs (Ωext ext
γ ) → Hs (Ωσ ) given by S|h(i) (Ωext ) = Sj is called a
s γ
−1
regularization of λ .
The function FJ = SJ g is called the J-level regularization of λ−1 g, whereas
(i) (i)
FJ = SJ g (i) is called the J-level regularization of (λ(i) )−1 g. Within our multi-
scale approach, we now represent the (J + 1)-level regularization using the J-level
regularization by adding the corresponding detail information. To this end, we
first introduce a multiscale regularization concept starting with the definition of a
generator of a regularization scaling function.
(i) (i)
Definition 7.39. A family {{ϕj (n)}n∈N0i }j∈N0 of sequences {ϕj (n)}n∈N0i ,
i ∈ {1, 2, 3}, is called a generator of a regularization scaling function with respect
to (λ(i) )−1 , if it satisfies the following requirements:
(i) 1
(i) (ϕj (0i ))2 = λ(i)∧ (0i )
for all j ∈ N0 ,
(i) (i)
(ii) (ϕj (n)) ≤ (ϕj (n)) for all j, j
∈ N0 with
2 2
j ≤ j
and all n ∈ N0i +1 ,
(i)
(iii) lim (ϕ (n))2 = λ(i)∧1 (n) for all n ∈ N0i +1 .
j→∞ j
(i)
Furthermore, {{{ϕj (n)}i∈{1,2,3} }n∈N0i }j∈N0 is called a generator of a regulariza-
tion scaling function with respect to λ−1 , if (λ(i) )−1 is a generator of a regulariza-
tion scaling function with respect to (λ(i) )−1 for every i = 1, 2, 3.
We now define decomposition and reconstruction regularization scaling func-

tions.
(i)
Definition 7.40. Let {{ϕj (n)}n∈N0i }j∈N0 be a generator of a regularization scaling
(i),σ,γ
function with respect to (λ(i) )−1 . Then a family {d Φ̂j (·, ·)}j∈N0 of admissible
(i)
hσ,γ -kernels given by
∞
2n+1

(i),σ,γ (i)
d
Φ̂j (x, z) = ϕj (n) s
Hn,m (σ; x)h(i)s
n,m (γ; z), (7.118)
n=0i m=1
(i)
σ × Ωγ , is called a decomposition regularization hσ,γ -scaling function
(i),σ,σ (i)
with respect to (λ(i) )−1 , whereas a family {r Φ̂j (·, ·)}j∈N0 of admissible hσ,σ -
kernels given by
∞
2n+1

(i),σ,σ (i)
r
Φ̂j (x, y) = ϕj (n) s
Hn,m (σ; x)h(i)s
n,m (σ; y), (7.119)
n=0i m=1
(i),σ,σ
σ ×Ωσ , is called a reconstruction regularization h -scaling function
(i) −1
with respect to (λ ) .
We obtain the following theorem:
(i)
Theorem 7.41. Let {{ϕj (n)}n∈N0i }j∈N0 be a generator of a regularization scaling
function with respect to (λ(i) )−1 , i ∈ {1, 2, 3}. If we formally define
(i) (i)
(r Φ̂j d Φ̂j )σ,γ (·, ·)
by
(i) (i) (i),σ,σ (i),σ,γ
(r Φ̂j d Φ̂j )σ,γ (x, z) = r Φ̂j (x, ·) d Φ̂j (·, z), (7.120)

σ × Ωγ , then
(i) (i) (i)

FJ = (r Φ̂J d Φ̂J )σ,γ ∗ g (i) , g (i) ∈ h(i) ext
s (Ωγ ), (7.121)
represents the J-level regularization of (λ(i) )−1 g (i) . If, in addition, g (i) ∈ im(λ(i) ),
then
(i)
lim FJ − (λ(i) )−1 g (i) Hs (Ωext ) = 0. (7.122)
J→∞ σ
Furthermore,
3
3

(i) (i)
FJ = (r Φ̂J d Φ̂J )σ,γ ∗ g (i) , g= g (i) ∈ hs (Ωext
γ ), (7.123)
i=1 i=1
represents the J-level regularization of λ−1 g. If, in addition, g ∈ im(λ), then

lim FJ − λ−1 g Hs (Ωext ) = 0. (7.124)
J→∞ σ
(i) (i)
We now define the convolution operators SJ : hs (Ωext ext
γ ) → Hs (Ωσ ),
J ∈ N0 , by
(i) (i) (i)
SJ (g (i) ) = (r Φ̂J d Φ̂J )σ,γ ∗ g (i) , (7.125)
whereas the convolution operator SJ : hs (Ωext
γ ) → Hs (Ωext
σ ), J ∈ N0 , is given by
3
(i)
SJ (g) = SJ (g (i) ). (7.126)
i=1
(i)
Furthermore, we introduce the corresponding scale spaces SJ (im(λ(i) )),
i ∈ {1, 2, 3}, and SJ (im(λ)) as follows
(i) (i)

(i)
SJ (im(λ(i) )) = (r Φ̂J d Φ̂J )σ,γ ∗ g (i) : g (i) ∈ im(λ(i) ) , (7.127)
3 3
(i) (i)
SJ (im(λ)) = (r Φ̂J d Φ̂J )σ,γ ∗ g (i) : g = g (i) ∈ im(λ) . (7.128)
i=1 i=1
(i) (i) (i)
(i) S0 (im(λ(i) )) ⊂ · · · ⊂ SJ (im(λ(i) )) ⊂ SJ (im(λ(i) )) ⊂ Hs (Ωext
σ ), J ≤ J ,
(i) (i)
i.e., for any right-hand side g ∈ im(λ ) of the (vectorial) SST or SGG
problem, all J-level regularizations with fixed parameter J are sampled in a
(i)
scale space SJ (im(λ(i) )) with the above property,
; (i)
·H (Ωext )
(ii) ∞
s
(i) σ
= Hs (Ωext
J=0 SJ (im(λ )) σ ).
(i)
Obviously, Theorem 7.42 is also valid substituting SJ by SJ which leads to
the following corollary.
Corollary 7.43. The scale spaces satisfy the following properties:

(i) S0 (im(λ)) ⊂ · · · ⊂ SJ (im(λ)) ⊂ SJ (im(λ)) ⊂ Hs (Ωext

σ ), J ≤ J , i.e., for any
right-hand side g ∈ im(λ) of the (vectorial) SST or SGG problem, all J-level
regularizations with fixed parameter J are sampled in a scale space SJ (im(λ))
;∞ · ext
(ii) J=0 SJ (im(λ)) Hs (Ωσ ) = Hs (Ωext σ ).

σ ) satisfying the conditions of Corollary 7.43 is
called regularization hσ,γ -multiresolution analysis (RMRA) of the (vectorial) SST
or SGG problem.
(i)
Definition 7.44. Let {{ϕj (n)}n∈N0i }j∈N0 be a generator of a regularization scal-
(i)
ing function with respect to (λ(i) )−1 . Then the generating symbols {ψj (n)}n∈N0i ,
(i)
{ψ̃j (n)}n∈N0i of the corresponding regularization wavelets are defined by the re-
(i) (i),σ,γ
finement equation (7.81). The admissible hσ,γ -kernels {d Ψ̂j (·, ·)}j∈N0 given by
∞
2n+1

(i),σ,γ (i)
d
Ψ̂j (x, z) = ψj (n) s
Hn,m (σ; x)h(i)s
n,m (γ; z), (7.129)
n=0i m=1
(i)
σ ×Ωγ , are called the decomposition regularization hσ,γ -wavelets, while
(i) ˜ (i),σ,σ
the admissible hσ,σ -kernels {r Ψ̂j (·, ·)}j∈N0 given by
(i),σ,σ ∞
2n+1

r˜ (i)
Ψ̂j (x, y) = ψ̃j (n) s
Hn,m (σ; x)h(i)s
n,m (σ; y), (7.130)
n=0i m=1
(i)
σ × Ωσ are called the reconstruction regularization hσ,σ -wavelets.
(i) (i)
We now define the convolution operators Tj : hs (Ωext ext
γ ) → Hs (Ωσ ), j ∈
N0 , i = 1, 2, 3, by
(i) ˜ (i) (i)
Tj (g (i) ) = (r Ψ̂j d Ψ̂j )σ,γ ∗ g (i) , g (i) ∈ h(i) ext
s (Ωγ ), (7.131)
and the convolution operator Tj : hs (Ωext ext

γ ) → Hs (Ωσ ), j ∈ N0 , by
3
(i)
TJ (g) = TJ (g (i) ). (7.132)
i=1
(i)
Obviously, due to the refinement equation, the operators SJ+1 and SJ+1 can be
represented in the form
(i) (i)

J
(i)
SJ+1 = S0 + Tj , (7.133)
j=0

J
SJ+1 = S0 + Tj . (7.134)
j=0
(i)
Thus, we now introduce the detail spaces TJ (im(λ(i) )) and TJ (im(λ)) by

(i) ˜ (i) (i)
TJ (im(λ(i) )) = (r Ψ̂J d Ψ̂J )σ,γ ∗ g (i) : g (i) ∈ im(λ(i) ) , (7.135)
3 3
˜ (i) (i)
TJ (im(λ)) = (r Ψ̂J d Ψ̂J )σ,γ ∗ g (i) : g = g (i) ∈ im(λ) . (7.136)
i=1 i=1
In terms of the multiscale concept, the space TJ (im(λ)) contains the detail
information which has to be added in order to turn from the J-level regularization
to the (J + 1)-level regularization:
SJ+1 (im(λ)) = SJ (im(λ)) + TJ (im(λ)). (7.137)
(i)
Theorem 7.45. Let {{ϕj (n)}n∈N0i }j∈N0 be a generator of a regularization scaling
(i)
function with respect to (λ(i) )−1 , i ∈ {1, 2, 3}. Suppose that {{ψj (n)}n∈N0i }j∈N0 ,
(i)
{{ψ̃j (n)}n∈N0i }j∈N0 are the generating symbols of the corresponding regulariza-
(i)
tion wavelets. Furthermore, let g (i) be of class hs (Ωext
γ ). Define the regularization
(i)
hσ,γ -wavelet transform at scale j ∈ N0 and position x ∈ Ωext
σ by
(i),σ,γ
(RW T )(g (i) )(j; x) = d Ψ̂j (x, ·) ∗ g (i) , g (i) ∈ h(i) ext
s (Ωγ ). (7.138)
Then
(i) (i)
J−1 (i),σ,σ
(i) r˜
FJ = ( r
Φ̂0 d
Φ̂0 )σ,γ ∗h + Ψ̂j (RW T )(g (i) )(j; ·)
j=0
is the J-level regularization of the (vectorial) SST or SGG problem satisfying

lim FJ − (λ(i) )−1 g (i) Hs (Ωext ) = 0 (7.139)
J→∞ σ
provided that g (i) ∈ im(λ(i) ).

7.3. Tensorial wavelet theory
The extension from vector to tensor theory is performed in this section. First,
we define tensorial scaling functions and wavelets and give the reconstruction for-
mula. The solution of the tensorial SGG problem is presented using regularization
wavelets.
(i,k)
Definition 7.46. Let {Φj (·, ·)}j∈N0 , i, k ∈ {1, 2, 3}, be a family of admissible
(i,k) (i,k) (i,k)
hσ,σ -kernels. Then the family {Vj (Ωext
σ )}j∈N0 of scale spaces Vj (Ωext
σ ) de-
fined by
(i,k) (i,k) (i,k)
Vj (Ωext
σ ) = {Φj Φj ∗ f : f ∈ hs (Ωext
σ )}, (7.140)
(i,k)
is called an hσ,σ -multiresolution analysis, if the following properties are satisfied:
(i,k) (i,k) (i,k) (i,k)
(i) V0 (Ωext
σ ) ⊂ · · · ⊂ Vj (Ωext ext
σ ) ⊂ Vj+1 (Ωσ ) ⊂ · · · ⊂ hs (Ωext
σ ),
; (i,k) ext hs (Ωext
·
σ ) (i,k)
(ii) Vj (Ωσ ) = hs (Ωextσ ).
j∈N0
Definition 7.47. Let {Φj (·, ·)}j∈N0 be a family of admissible hσ,σ -kernels. The set
of scale spaces Vj (Ωext
σ ) defined by
Vj (Ωext ext
σ ) = {Φj Φj ∗ f : f ∈ hs (Ωσ )} (7.141)
(i,k) (i,k)
is called an hσ,σ -multiresolution analysis, if {Vj (Ωext
σ )}j∈N0 is an hs (Ωext
σ )-
multiresolution analysis for every i, k ∈ {1, 2, 3}.
We now define the scaling functions.
(i,k) (i,k)
Definition 7.48. A family {{ϕj (n)}n∈N0ik }j∈N0 of sequences {ϕj (n)}n∈N0ik is
called a generator of a scaling function of kind (i, k), i, k ∈ {1, 2, 3}, if it satisfies
the following requirements:
(i,k)
(i) (ϕj (0ik ))2 = 1, for all j ∈ N0 ,
2
(i,k) (i,k)
(ii) (ϕj (n))2 ≤ ϕj (n) , for all j, j
∈ N0 with j ≤ j
and all n ∈ N0ik +1 ,
(i,k)
(iii) lim (ϕj (n))2 = 1, for all n ∈ N0ik +1 .
j→∞
Furthermore, the family {{{ϕ(i,k) (n)}i,k∈{1,2,3} }n∈N0ik }j∈N0 is called a generator

of a scaling function, if {{ϕ(i,k) (n)}n∈N0ik }j∈N0 are generators of a scaling function
of kind (i, k), i, k ∈ {1, 2, 3}.
hσ,σ -scaling functions.
(i) (i,k)
Definition 7.49. A family {Φj (·, ·)}j∈N0 of h(i,k) -kernels Φj (·, ·) defined by
(i,k)∧ (i,k)
Φj (n) = ϕj (n), j ∈ N0 , n ∈ N0ik , i.e.,
∞
2n+1

(i,k) (i,k)
Hn,m (σ; y)h(i,k)s
n,m (σ; x), x, y ∈ Ωext
σ , (7.142)
n=0ik m=1
(i,k)
is called an hσ,σ -scaling function, if it satisfies the following properties:
(i,k) (i,k)
(i) Φj (·, ·) is an admissible hσ,σ -kernel for every j ∈ N0 ,
(i,k)∧
(ii) {{Φj (n)n∈N0ik }j∈N0 constitutes a generator of a scaling function of kind
(i, k).
Furthermore, the family {Φj (·, ·)}j∈N0 of hσ,σ -kernels Φj (·, ·) is called an hσ,σ -
(i,k) (i,k)
scaling function, if {Φj }j∈N0 are hσ,σ -scaling functions for i, k ∈ {1, 2, 3}.
As in the scalar and vectorial theory, the following approximation theorem is
valid.
Theorem 7.50. Let {Φj (·, ·)}j∈N0 be an hσ,σ -scaling function. Then
lim f − Φj Φj ∗ f hs (Ωext ) = 0 (7.143)
j→∞ σ

σ ).
Definition 7.51. Let {Φj (·, ·)}j∈N0 be an hσ,σ -scaling function. Then {Pj }j∈N0 with
Pj : hs (Ωext ext
σ ) → hs (Ωσ ) defined by
Pj (f )(x) = Φj Φj ∗ f , f ∈ hs (Ωext
σ ), x ∈ Ωext
σ , (7.144)
is called an hσ,σ -approximate identity.
The kernel Φ0 is called the mother kernel of the hσ,σ -scaling function. We
obtain the following theorem.
Theorem 7.52. Let {Φj (·, ·)}j∈N0 be an hσ,σ -scaling function. Then {Vj (Ωext
σ )}j∈N0
given in (7.141) forms an hσ,σ -multiresolution analysis.
The next purpose is to define the primal and dual wavelet with the help of
the tensorial refinement equation.
Definition 7.53. Let {Φj (·, ·)}j∈N0 be an hσ,σ -scaling function. Then the families
of hσ,σ -kernels {Ψj (·, ·)}j∈N0 , {Ψ̃j (·, ·)}j∈N0 given by
(i,k)
(Ψj )(i,k)∧ (n) = ψj (n), j ∈ N0 , n ∈ N0ik , i, k ∈ {1, 2, 3}, (7.145)
(i,k)∧ (i,k)
(Ψ̃j ) (n) = ψ̃j (n), j ∈ N0 , n ∈ N0ik , i, k ∈ {1, 2, 3}, (7.146)
are called (primal) hσ,σ -wavelet and dual hσ,σ -wavelet, respectively, if all hσ,σ -
(i,k)
kernels Ψj (·, ·), Ψ̃j (·, ·), j ∈ N0 , are admissible and the symbols {ψj (n)},
(i,k)
{ψ̃j (n)}, in addition, satisfy the (tensorial) refinement equation
(i,k) (i,k) (i,k)
ψ̃j (n)ψj (n) = (ϕj+1 (n))2 − (ϕj (n)(i,k) )2 (7.147)
for all j ∈ N0 , n ∈ N0ik , i, k ∈ {1, 2, 3}.
As a direct consequence we get the following equation:
(i,k) (i,k)

J
(i,k) (i,k)
(ϕJ+1 (n))2 = (ϕ0 (n))2 + ψ̃j (n)ψj (n), J ∈ N0 , (7.148)
j=0
(i,k)
for all n ∈ N0ik . We now define the wavelet transform. To this end we let ψ−1 (n) =
(i,k) (i,k)
ψ̃−1 (n) = ϕ0 (n), for n ∈ N0ik , i, k ∈ {1, 2, 3}, Ψ−1 (·, ·) = Ψ̃−1 (·, ·) = Φ0 (·, ·).
We remember the space H(N−1 × Ωext σ ) (see Eqs. (7.12)–(7.14)).
Definition 7.54. Let {Ψj (·, ·)}j∈N−1 be a (primal) hσ,σ -wavelet. Then (W T )(i,k) :
hs (Ωext ext
σ ) → Hs (N−1 × Ωσ ) defined by
(i,k)
(W T )(i,k) (f )(j; y) = (Ψj ∗ f )(y) (7.149)
is called hσ,σ -wavelet transform if kind (i, k) of f at position y ∈ Ωext

σ and scale
j ∈ N−1 .
(i,k)
As usual, we define the detail space Wj (Ωext
σ ) at scale j by
(i,k) (i,k) (i,k)

Wj (Ωext
σ ) = {Ψ̃j Ψj ∗ f : f ∈ h(s Ωext
σ )}, (7.150)
and
Wj (Ωext ext
σ ) = {Ψ̃j Ψj ∗ f : f ∈ hs (Ωσ )}. (7.151)
Theorem 7.55 (Tensorial Reconstruction Formula for the Outer Space). Let the
families {Ψj (·, ·)}j∈N0 and {Ψ̃j (·, ·)}j∈N0 , respectively, be a (primal) hσ,σ -wavelet
and its dual corresponding to an hσ,σ -scaling function {Φj (·, ·)}j∈N0 . Then
∞

f= Ψ̃j Ψj ∗ f (7.152)
j=−1

σ ) (in · hs (Ωext ) -sense). σ
We now solve the (tensorial) SGG problem using regularization wavelets.

(i,k) (i,k)
First, we transfer the theory of hσ,σ -scaling functions to the general case of hα,α -
(i,k),α,α
scaling functions Φj with α ≥ σ:
∞
2n+1

(i,k),α,α (i,k)
Hn,m (α; x)h(i,k)s
n,m (α; y), (7.153)
n=0ik m=1
where
(i,k),α,α ∧ (i,k)
(Φj ) (n) = ϕj (n). (7.154)
Theorem 7.50 can be directly transferred substituting σ by α. The scale spaces
are defined in the following way:
(i,k) (i,k),α,α (i,k),α,α
Vj (Ωext
α ) = {Φj Φj ∗ f : f ∈ hs (Ωext
α )}. (7.155)
(i,k)
α )} of scale spaces forms a multiresolution analysis.
(i,k)
Theorem 7.56. The restriction of the operator λ(i,k) : Hs (Ωext
σ ) → hs (Ωext
γ ) to
a scale space Vj (Ωext
σ ), j ∈ N 0 , i.e.,
(i,k)
λ(i,k) |Vj (Ωext ) : Vj (Ωext
σ ) → Vj (Ωext
γ ), (7.156)
σ
is injective for (i, k) ∈ {(1, 1), (2, 1), (3, 1)}, whereas in the case of (i, k) ∈ {(1, 2),
(1, 3), (2, 3), (3, 3)} the Fourier coefficient of degree 0 cannot be recovered and the
Fourier coefficients of degree n ≥ 1 are uniquely defined. In the case of (i, k) ∈
{(2, 2), (3, 2)} the Fourier coefficient of degree 0 and 1 cannot be recovered and the
Fourier coefficients of degree n ≥ 2 are uniquely defined (in the following text,
injectivity, bijectivity and uniqueness is always used in this sense).
Moreover, we have the following results:
(i,k)
(i) If the families {{ϕj (n)}n∈N0ik }j∈N0 and {{ϕj (n)}n∈N0 }j∈N0 , i,k ∈ {1,2,3},
(i,k)
are bandlimited (for example, ϕj (n) = ϕj (n) = 0 for all n ≥ 2j ), then the
restricted operator is even bijective (in the sense described above). To be more
(i,k)
specific, for g(i,k) ∈ hs (Ωext ext
γ ) the unique solution Fj ∈ Vj (Ωσ ), j ∈ N0 ,
of the equation
(i,k),γ,γ (i,k),γ,γ
λ(i,k) Fj = Φj Φj ∗ g(i,k) (7.157)
is given by
Fj = Φσ,σ
j ∗ Φσ,σ
j ∗ Q, (7.158)
(i,k)∧
g (n,m)
, n ∈ [0ik , 2j ),
Q∧ (n, m) = λ(i,k)∧ (n) (7.159)
0, n ∈ [2j , ∞).
(i,k)
(ii) If the families {{ϕj (n)}n∈N0ik }j∈N0 , i,k ∈ {1,2,3}, and {{ϕj (n)}n∈N0 }j∈N0
are non-bandlimited, the equation
λ(i,k) Fj = Φj Φj ∗ g(i,k) (7.160)
(i,k) (i,k)Λ
has a solution Fj ∈ Vj (Ωext
σ ) provided that g ∈ hs (Ωext
γ ), where
(i,k)Λ
hs (Ωext
γ ) is an appropriate Sobolev space (see the Ph.D.-thesis [58] for
more details). In this case, the unique solution of the equation is given by
Fj = Φσ,σ
j ∗ Φσ,σ
j ∗ Q, (7.161)
σ ) is obtainable in spectral language by
g(i,k)∧ (n, m)
Q∧ (n, m) = , (7.162)
λ(i,k)∧ (n)
n = 0ik , . . .; m = 1, . . . , 2n + 1.
The following corollary shows that in the case of general operators λ =
3 (i,k)
i,k=1 λ we have to claim an additional assumption onto the function g.

Corollary 7.57. The restriction of the operator λ = 3i,k=1 λ(i,k) to a scale space
Vj (Ωext
σ ), j ∈ N0 , i.e.,
3
5 (i,k)
λ|Vj (Ωext ) : Vj (Ωext
σ )→ Vj (Ωext
γ ) (7.163)
σ
i,k=1
(i,k)
has, in general, no solution. Under the assumption ϕj (n) = ϕj (n), i, k ∈
{1, 2, 3}, we have to claim, in addition, that
g(i,k)∧ (n, m) g(l,r)∧ (n, m)
= , (7.164)
λ(i,k)∧ (n) λ(l,r)∧ (n)
with i, k, l, r ∈ {1, 2, 3}; n = max (0ik , 0lr ), . . .; m = 1, . . . , 2n + 1. Then the
i,k,l,r∈{1,2,3}
results in Theorem 7.56 can be directly transferred.
With the help of the refinement equation (7.147) we now define the primal
(i,k),α,α (i,k),α,α
wavelets {Ψj (·, ·)}j∈N0 and the dual wavelets {Ψ̃j (·, ·)}j∈N0 for α ≥ σ,
i, k ∈ {1, 2, 3}:
∞
2n+1

(i,k),α,α (i,k)
Ψj (x, y) = ψj (n) s
Hn,m (α; x)h(i,k)s
n,m (α; y), (7.165)
n=0ik m=1
∞
2n+1

(i,k),α,α (i,k)
Ψ̃j (x, y) = ψ̃j (n) s
Hn,m (α; x)h(i,k)s
n,m (α; y), (7.166)
n=0ik m=1
where
(i,k),α,α ∧ (i,k) (i,k),α,α ∧ (i,k)
(Ψj ) (n) = ψj (n), (Ψ̃j ) (n) = ψ̃j (n). (7.167)

(i,k) (i,k),α,α (i,k),α,α
Wj (Ωext
α ) = {Ψj Ψ̃j ∗ f : f ∈ hs (Ωext
α )}. (7.168)
Theorem 7.55 can be directly transferred by substituting the convolutions with
respect to the sphere Ωσ by the corresponding convolutions with respect to the
sphere Ωα . We now transfer Theorem 7.56 to the detail spaces and get the following
theorem, where we use the terms injectivity, bijectivity, and uniqueness in the same
sense as before.
(i,k)
Theorem 7.58. The restriction of the operator λ(i,k) : Hs (Ωext
σ ) → hs (Ωext
γ ) to
a detail space Wj (Ωext
σ ), j ∈ N 0 , i.e.,
(i,k)
λ(i,k) |Wj (Ωext ) Wj (Ωext
σ ) → Wj (Ωext
γ ) (7.169)
σ
(i,k)
with ψj (n) = ψj (n) is injective. Moreover, we have the following results:
(i,k)
(i) If the families {{ϕj (n)}n∈N0ik }j∈N0 and {{ϕj (n)}n∈N0 }j∈N0 , i,k ∈ {1,2,3},
(i,k)
are bandlimited (for example, ϕj (n) = ϕj (n) = 0 for all n ≥ 2j ), then
the restricted operator is even bijective. To be more specific, for g(i,k) ∈
(i,k)
hs (Ωext ext
γ ) the unique solution Hj ∈ Wj (Ωσ ), j ∈ N0 , of the equation
λ(i,k) Hj = Ψ̃j ∗ Ψj ∗ g(i,k) (7.170)
is given by
Hj = Ψ̃σ,σ
j ∗ Ψσ,σ
j ∗ Q, (7.171)
(i,k)∧
g (n,m)
, n ∈ [0ik , 2j+1 ),
Q∧ (n, m) = λ(i,k)∧ (n) (7.172)
0, n ∈ [2j+1 , ∞).
(i,k)
(ii) If the families {{ϕj (n)}n∈N0ik }j∈N0 , i,k ∈ {1,2,3}, and {{ϕj (n)}n∈N0 }j∈N0
are non-bandlimited, the equation
λ(i,,k) Hj = Ψ̃j Ψj ∗ g(i,k) (7.173)
has a solution Hj ∈ Wj (Ωext

σ ) provided that the condition
∞ 2n+1
g(i,k)∧ (n, m)
<∞ (7.174)
n=0ik m=1 λ(i,k)∧ (n)
is satisfied for g(i,k) ∈ hs (i,k) (Ωext

γ ). In this case, the unique solution of the
equation is given by
Hj = Ψ̃σ,σ
j ∗ Ψσ,σ
j ∗ Q, (7.175)
g(i,k)∧ (n, m)
Q∧ (n, m) = , (7.176)
λ(i,k)∧ (n)
n = 0ik , . . .; m = 1, . . . , 2n + 1.
Furthermore, we have the following corollary.
3
Corollary 7.59. The restriction of the operator λ = i,k=1 λ(i,k) to a detail space
Wj (Ωext
σ ), j ∈ N0 , i.e.,
3
5 (i,k)
λ|Wj (Ωext ) : Wj (Ωext
σ )→ Wj (Ωext
γ ) (7.177)
σ
i,k=1
has, in general, no solution. Under the assumption

(i,k) (i,k)
ψj (n) = ψj (n) and ψ̃j (n) = ψ̃j (n), i, k ∈ {1, 2, 3},
we have to claim, in addition, that
g(i,k)∧ (n, m) g(l,r)∧ (n, m)
= , (7.178)
λ(i,k)∧ (n) λ(l,r)∧ (n)
with i, k, l, r ∈ {1, 2, 3}; n = max (0ik , 0lr ), . . . ; m = 1, . . . , 2n+1. Then the results
i,k,l,r
in Theorem 7.58 can be directly transferred.
the solution when we restrict the operator to the scale or detail spaces. In this
case, the injectivity for the operators λ(i,k) could be proved, whereas in the case
3
of general operators λ = i,k=1 λ(i,k) we have to claim that (7.164) is valid. In
the case of the unfiltered solution, we obtain the following theorem.
(i,k) (i,k)
Theorem 7.60. Let g(i,k) ∈ hs (Ωext γ ) satisfy the condition g ∈ im(λ ), i, k ∈
(i,k)
ext
{1, 2, 3}. Then the unique solution F ∈ Hs (Ωσ ) of the equation λ F = g(i,k)
is given by
g(i,k)∧ (n, m)
F ∧ (n, m) = , (7.179)
λ(i,k)∧ (n)
3
n = 0ik , . . .; m = 1, . . . , 2n + 1. In the case of the operator λ = i,k=1 λ(i,k) we
have to claim, in addition, that (7.178) holds in order to guarantee the solvability.
Last, we explain the connection between the solution in the scale spaces and
(i,k)Λ
Theorem 7.61. Suppose that g(i,k) is of the class hs (Ωext ext
γ ). Let F ∈ Hs (Ωσ )
be the unique solution of λ(i,k) F = g(i,k) . Then
(2)
Fj = (Φj )σ,σ ∗ F (7.180)

λ(i,k) Fj = Φj Φj ∗ g(i,k) (7.181)
(2)
lim (ΦJ )σ,σ ∗ F = F (7.182)
J→∞
σ
The preceding theorem shows that in the case of bandlimited scaling functions
the (tensorial) SGG-problem is well posed, because a unique solution always exists
and due to the finite dimension of the scale spaces the solution is also stable. We
now investigate the case of non-bandlimited scaling functions, where the stability
cannot be ensured and we have to use regularization methods.
(i,k) (i,k)
Definition 7.62. A family of linear operators Sj : hs (Ωext γ ) → Hs (Ωσ ),
ext
j ∈ N0 , is called a regularization of (λ(i,k) )−1 , i, k ∈ {1, 2, 3}, if it satisfies the

(i,k) (i,k)
(i) Sj is bounded on hs (Ωext
γ ) for all j ∈ N0 ,
(ii) for any member g(i,k) ∈ im(λ(i,k) ), the limit relation
(i,k) (i,k)
lim SJ g = (λ(i,k) )−1 g(i,k) (7.183)
J→∞
σ
(i,k)
The operator S : hs (Ωext
γ ) → Hs (Ωext
σ ) given by S|h(i,k) (Ωext ) = Sj is called a
s γ
−1
regularization of λ .
The function FJ = SJ g is called the J-level regularization of λ−1 g, whereas
(i,k) (i,k)
FJ = SJ g(i,k) is called the J-level regularization of (λ(i,k) )−1 g. Within our
multiscale approach, we now represent the (J + 1)-level regularization using the
J-level regularization by adding the corresponding detail information. To this end
we first introduce a multiscale regularization concept starting with the definition
of a generator of a regularization scaling function.
(i,k) (i,k)
Definition 7.63. A family {{ϕj (n)}n∈N0ik }j∈N0 of sequences {ϕj (n)}n∈N0ik ,
i, k ∈ {1, 2, 3}, is called a generator of a regularization scaling function with respect
to (λ(i,k) )−1 , if it satisfies the following requirements:
(i,k) 1
(i) (ϕj (0ik ))2 = , for all j ∈ N0 ,
λ(i,k)∧ (0ik )
(i,k) (i,k)
(ii) (ϕj (n))2 ≤ (ϕj (n))2 , for all j, j
∈ N0 with j ≤ j
and all n ∈ N0ik +1 ,
(i,k) 1
(iii) lim (ϕj (n))2 = , for all n ∈ N0ik +1 .
(λ
(i,k)
j→∞ )∧ (n)
(i,k)
Furthermore, {{{ϕj (n)}i,k∈{1,2,3} }n∈N0ik }j∈N0 is called a generator of a regu-
larization scaling function with respect to λ−1 , if (λ(i,k) )−1 is a generator of a
regularization scaling function with respect to (λ(i,k) )−1 for every i, k = 1, 2, 3.
We now define decomposition and reconstruction regularization scaling func-

tions.
(i,k)
Definition 7.64. Let {{ϕj (n)}n∈N0ik }j∈N0 be a generator of a regularization
scaling function with respect to (λ(i,k) )−1 , i, k ∈ {1, 2, 3}.
(i,k),σ,γ (i,k)
Then a family {d Φj (·, ·)}j∈N0 of admissible hσ,γ -kernels given by
∞
2n+1

(i,k),σ,γ (i,k)
d
Φj (x, z) = ϕj (n) s
Hn,m (σ; x)h(i,k)s
n,m (γ; z), (7.184)
n=0ik m=1
(i,k)
σ × Ωγ , is called a decomposition regularization hσ,γ -scaling func-
(i,k),σ,σ
tion with respect to (λ(i,k) )−1 , whereas a family { r Φj (·, ·)}j∈N0 of admissible
(i,k)
hσ,σ -kernels given by
∞
2n+1

(i,k),σ,σ (i,k)
r
Hn,m (σ; x)h(i,k)s
n,m (σ; y), (7.185)
n=0ik m=1
(i,k)
σ × Ωσ is called a reconstruction regularization hσ,σ -scaling function
with respect to (λ(i,k) )−1 .
We obtain the following theorem:
(i,k)
Theorem 7.65. Let {{ϕj (n)}n∈N0ik }j∈N0 be a generator of a regularization scal-
ing function with respect to (λ(i,k) )−1 , i, k ∈ {1, 2, 3}. If we formally define
(i,k) (i,k) σ,γ
(r Φj d Φj ) (·, ·)
by
(i,k) (i,k) σ,γ (i,k),σ,σ (i,k),σ,γ
(r Φj d Φj ) (x, z) = r Φj (x, ·) d Φj (·, z), (7.186)
σ × Ωγ , then
(i,k) (i,k) (i,k) σ,γ
FJ = (r ΦJ d ΦJ ) ∗ g(i,k) , g(i,k) ∈ h(i,k)
s (Ωext
γ ), (7.187)
represents the J-level regularization of (λ(i,k) )−1 g(i,k) . If, in addition, g(i,k) ∈
im(λ(i,k) ), then
(i,k)
lim FJ − (λ(i,k) )−1 g(i,k) Hs (Ωext ) = 0. (7.188)
J→∞ σ
Furthermore,
3
3

(i,k) (i,k) (i,k)
FJ = ( r
ΦJ d
ΦJ )σ,γ ∗g , g= g(i,k) ∈ hs (Ωext
γ ), (7.189)
i,k=1 i,k=1
represents the J-level regularization of λ−1 g. If, in addition, g ∈ im(λ), then

lim FJ − λ−1 g Hs (Ωext ) = 0. (7.190)
J→∞ σ
(i,k) (i,k)
We define the convolution operators SJ : hs (Ωext ext
γ ) → Hs (Ωσ ),
J ∈ N0 , by
(i,k) (i,k) (i,k) σ,γ
SJ (g(i,k) ) = (r ΦJ d ΦJ ) ∗ g(i,k) , (7.191)
whereas the convolution operator SJ : hs (Ωext ext
γ ) → Hs (Ωσ ), J ∈ N0 , is given by
3
(i,k)
SJ (g) = SJ (g(i,k) ). (7.192)
i,k=1
(i,k)
Furthermore, we introduce the corresponding scale spaces SJ (im(λ(i,k) )), i, k ∈
{1, 2, 3}, and SJ (im(λ)) as follows
(i,k)

(i,k) (i,k)
SJ (im(λ(i,k) )) = (r ΦJ d ΦJ )σ,γ ∗ g(i,k) : g(i,k) ∈ im(λ(i,k) ) , (7.193)

3 3

(i,k) (i,k) σ,γ

SJ (im(λ)) = (r ΦJ d ΦJ ) ∗ g(i,k) : g = g(i,k) ∈ im(λ) .
i,k=1 i,k=1
(7.194)
(i,k) (i,k) (i,k)
(i) S0 (im(λ(i,k) )) ⊂ · · · ⊂ SJ (im(λ(i,k) )) ⊂ SJ (im(λ(i,k) )) ⊂ Hs (Ωext σ ),

(i,k) (i,k)
J ≤ J , i.e., for any right-hand side g ∈ im(λ ) of the (tensorial) SGG
problem, all J-level regularizations with fixed parameter J are sampled in a
(i,k)
scale space SJ (im(λ(i,k) )) with the above property,
;∞ (i,k)
·H (Ωext )
(ii) J=0 SJ (im(λ(i,k) ))
s σ
= Hs (Ωext
σ ).
(i,k)
Obviously, Theorem 7.66 is also valid substituting SJ by SJ which leads
to the following corollary.
Corollary 7.67. The scale spaces satisfy the following properties:

(i) S0 (im(λ)) ⊂ · · · ⊂ SJ (im(λ)) ⊂ SJ (im(λ)) ⊂ Hs (Ωext

σ ), J ≤ J , i.e., for
any right-hand side g ∈ im(λ) of the (tensorial) SGG problem, all J-level
regularizations with fixed parameter J are sampled in a scale space SJ (im(λ))
;∞ · ext
(ii) J=0 SJ (im(λ)) Hs (Ωσ ) = Hs (Ωextσ ).

σ ) satisfying the conditions of Corollary 7.67 is
called regularization hσ,γ -multiresolution analysis (RMRA) of the (tensorial) SGG
problem.
(i,k)
Definition 7.68. Let {{ϕj (n)}n∈N0ik }j∈N0 be a generator of a regularization
−1
scaling function with respect to λ(i,k) . Then the generating symbols
(i,k) (i,k)
{ψ̃j (n)}n∈N0ik , {ψj (n)}n∈N0ik
of the corresponding regularization wavelets are defined by the refinement equation
(i,k) (i,k),σ,γ
(7.147). The admissible hσ,γ -kernels {d Ψj (·, ·)}j∈N0 given by
∞
2n+1

(i,k),σ,γ (i,k)
d
Ψj (x, z) = ψj (n) s
Hn,m (σ; x)h(i,k)s
n,m (γ; z), (7.195)
n=0ik m=1
(i,k)
(x, z) ∈ Ωext ext are called the decomposition regularization h
σ × Ωγ σ,γ -wavelets,
(i,k) (i,k),σ,σ
while the admissible hσ,σ -kernels {r Ψ̃j (·, ·)}j∈N0 given by
∞
2n+1

(i,k),σ,σ (i,k)
r
Ψ̃j (x, y) = ψ̃j (n) s
Hn,m (σ; x)h(i,k)s
n,m (σ; y), (7.196)
n=0ik m=1
(i,k)
σ × Ωσ are called the reconstruction regularization hσ,σ -wavelets.
(i,k) (i,k)
We now define the convolution operators Tj : hs (Ωext ext
γ ) → Hs (Ωσ ),
j ∈ N0 , i, k = 1, 2, 3, by
(i,k) (i,k) (i,k) σ,γ
Tj (g(i,k) ) = (r Ψ̃j d Ψj ) ∗ g(i,k) , g(i,k) ∈ h(i,k)
s (Ωext
γ ), (7.197)
and the convolution operator Tj : hs (Ωext ext
γ ) → Hs (Ωσ ), j ∈ N0 , by
3
(i,k)
TJ (g) = TJ (g(i,k) ). (7.198)
i,k=1
(i,k)
Obviously, due to the refinement equation the operators SJ+1 and SJ+1 can be
(i,k) (i,k)

J
(i,k)
SJ+1 = S0 + Tj , (7.199)
j=0

J
SJ+1 = S0 + Tj . (7.200)
j=0
(i,k)
Thus, we now introduce the detail spaces TJ (im(λ(i,k) )) and TJ (im(λ)) by

(i,k) (i,k) (i,k)
TJ (im(λ(i,k) )) = (r Ψ̃J d ΨJ )σ,γ ∗ g(i,k) : g(i,k) ∈ im(λ(i,k) ) , (7.201)
3 3
r (i,k) d (i,k) σ,γ (i,k) (i,k)

TJ (im(λ)) = ( Ψ̃J ΨJ ) ∗ g :g= g ∈ im(λ) .
i,k=1 i=1
(7.202)
In terms of the multiscale concept, the space TJ (im(λ)) contains the detail infor-
mation which has to be added in order to turn from the J-level regularization to
the (J + 1)-level regularization:
SJ+1 (im(λ)) = SJ (im(λ)) + TJ (im(λ)). (7.203)
(i,k)
Theorem 7.69. Let {{ϕj (n)}n∈N0ik }j∈N0 be a generator of a regularization scal-
ing function with respect to (λ(i,k) )−1 , i, k ∈ {1, 2, 3}. Suppose that
(i,k) (i,k)
{{ψj (n)}n∈N0ik }j∈N0 , {{ψ̃j (n)}n∈N0ik }j∈N0
(i,k) (i,k)
more, let g(i,k) be of the class hs (Ωext
γ ). Define the regularization hσ,γ -wavelet
transform at scale j ∈ N0 and position x ∈ Ωextσ by
(i,k),σ,γ
(RW T )(g(i,k) )(j; x) = d Ψj (x, ·) ∗ g(i,k) , g(i,k) ∈ h(i,k)
s (Ωext
γ ). (7.204)
Then

(i,k) σ,γ
J−1
(i,k),σ,σ
(i,k)
FJ = r
Φ0 d Φ0 ∗ h(i,k) + r
Ψ̃j (RW T )(g(i,k) )(j; ·) (7.205)
j=0
is the J-level regularization of the (tensorial) SGG problem satisfying
lim FJ − (λ(i,k) )−1 g(i,k) Hs (Ωext ) = 0 (7.206)

J→∞ σ
provided that g(i,k) ∈ im(λ(i,k) ).
7.4. Combined outer harmonic and wavelet concept

In geodetic practice, there exists a variety of realizations of spherical harmonic
models of the Earth’s external gravitational potential. In [19] it is explained how
to combine an outer harmonic model of fixed order m with a harmonic wavelet
model. The justification for such a combined model is the fact that on the one
hand the appropriate candidate for the approximation of the low frequency parts
of the gravitational potential (i.e., global modeling) is a spherical harmonic (i.e., a
multipole) model of moderate order m and on the other hand for the representation
of the high frequency parts (i.e., local modeling) new wavelet techniques have to
come into play (see also the investigations in spherical continuous wavelet theory
[33, 34]).
Starting point of this model is the “refinement equation” (compare Eq. (7.10))
ψ̃j (n)ψj (n) = (ϕj+1 (n))2 − (ϕj (n))2 .
It is clear that ψ̃j (n)ψj (n) = 0 if and only if (ϕj+1 (n))2 = (ϕj (n))2 . Due to condi-
tion (i) in Definition 7.2, the wavelet (or its dual) satisfy the mean value condition
ψj (0) = 0, i.e., it has to oscillate. For purposes of combined approximation we need,
however, (ϕj+1 (n))2 = (ϕj (n))2 for all n ∈ [0, . . . , m]. Under these assumptions it
may be guaranteed that the wavelets constructed in this way have more vanishing
moments and we call them wavelets of order m. In [19] the reconstruction formula
for such wavelets is studied in more detail. The transition of the combined outer
harmonic and wavelet concept to the vectorial and tensorial case is also easy to
perform.
8. Bandlimited Runge–Walsh multiscale approximation

In the previous sections we developed several methods of wavelet approximation.
We briefly reformulate the main results: Let {Ψj (·, ·)}j∈N0 be an Hσ,σ -wavelet
corresponding to an Hσ,σ -scaling function {Φj (·, ·)}j∈N0 . Then any potential F ∈
Hs (Ωext
σ ) can be expressed by a multiscale approximation given by
(2)

J−1
Φ0 ∗ F + Ψ̃j ∗ Ψj ∗ F, F ∈ Hs (Ωext
σ ). (8.1)
j=0
For a numerical realization, the discretization of the Hs (Ωext σ )-convolutions

(i.e., the Hs (Ωext
σ )-inner products) occurring in the J-level wavelet approximation
is necessary. For that purpose we observe that any Hs (Ωext σ )-convolution is express-
ext
ible as a bounded linear functional on Hs (Ωσ ). Thus fully discretized wavelet ap-
proximation amounts to the problem of approximating a bounded linear functional
(i.e., an Hs (Ωext
σ )-inner product) by a linear combination of known bounded linear
functionals. In this context it should again be mentioned that following our nomen-
clature an H0 (Ωextσ )-inner product can be identified with an ordinary integral over
the sphere Ωσ . Therefore, fully discretized H0 (Ωext σ )-wavelet approximation can
be organized appropriately by numerical integration (cubature) over the sphere
Ωσ . Looking at the inner products in our general Hs (Ωext σ )-framework we are con-
fronted with convolutions involving a pseudodifferential operator Λ with symbol
Λ∧ (n) = An for n ∈ N0 . Their discretization requires the knowledge of linear
(observational) functionals for the potential F ∈ Hs (Ωext σ ) under consideration.
Usually, in gravitational field determination, these (observational) functionals are
heterogeneous in nature. In addition, the approximate formulae have to be for-
mulated in dependence on the scale parameter, since increasing space localization
demands increasing data material.
All these requirements, however, do not lead to a unique procedure for dis-
cretizing Hs (Ωext
σ )-convolutions. Many variants of approximate formulae are rea-
sonable and conceivable. In fact, the choice of a suitable method is essentially
dependent on the purpose for which scaling functions and wavelets are used. Un-
fortunately, it turns out that each of the discretization methods has its own draw-
back. Nevertheless, a lot of approximation schemata for Hs (Ωext σ )-convolutions can
be found so that at least some of the requests can be fulfilled. As most important
discretization rules we mention:
1. Fast Fourier techniques and multipole techniques (cf. [19, 39, 74]) are eco-
nomical in time, but they are based on evaluation functionals on equiangular
latitude-longitude grids. Thus the sample points are merely equidistributed
on the (ϑ, ϕ)-parameter interval [0, π] × [0, 2π] in Euclidean space R2 , but not
on a sphere.
2. Polynomial (i.e., outer harmonic) exact approximation of bandlimited func-
tions is a well-established tool for application to bandlimited potentials of
moderate degree (cf. [12, 28, 29, 54]). The problem is that the preliminary
work includes the solution of a linear system of equations (which is full-sized

and tends to be ill conditioned for an increasing number of nodal points).
However, it can be shown that (outer harmonics) exact approximation of
bandlimited potentials can be used very efficiently (without a priori solving
any linear system) on equiangular grids (cf. [23, 39]).
3. Another method for the approximate evaluation of Hs (Ωext σ )-convolutions,
which includes the exact approximation of bandlimited functions as a special
case, is harmonic spline exact best approximation (cf. [12, 14, 19]). It can be
applied appropriately for modeling the medium to short wavelength parts of
a signal.
4. The low discrepancy method (cf. [21, 48]) represents an adequate tool if a
great number of data is available, so that the solution of linear equations
should be avoided. Sufficient accuracy can be guaranteed only if a high num-
ber of equidistributed data points are available. Thus it is of advantage for
integrands of high complexity (e.g., short wavelength parts of a signal).
In what follows, it will be shown that both discretization techniques, i.e.,
outer harmonic and spline exact integration, lead to pyramid schemata adapted to
the space localization properties of the potential we are interested in. To be more
specific, the bandlimited variant of fast wavelet computation (based on the Shan-
non kernel and its modifications) can be based on outer harmonic exact formulae
for the evaluation of Hs (Ωextσ )-inner products. It is proposed for the application
to moderate phenomena of space localization (i.e., low-to-medium wavelength ap-
proximation) so that one can work with smaller data sets (cf. [31, 32]). In fast
computation by bandlimited wavelets the number of wavelet coefficients is re-
duced, since they contain information of a more extended area. In addition, a
certain spectral band is expressible exactly in terms of wavelets because of their
bandlimited character. The non-bandlimited variant of fast wavelet evaluation (us-
ing non-bandlimited kernels such as Tikhonov, rational, exponential, and “locally
supported” kernels (cf. [29, 31]) is meant for the application to seriously space
localizing potentials (i.e., short wavelength approximation). In consequence, huge
data sets can be handled since only a small subset of the data is needed for the
purpose of numerical evaluation. On the other hand, a large number of wavelet
coefficients is needed, since they only give local information related to a small
area. Again, we are confronted with the drawback that large linear systems must
be solved in an a priori step to obtain the weights in (spline exact) best approxi-
mation formulae. In the non-bandlimited case, however, panel clustering or sparse
matrix techniques (cf. [23]) are efficiently applicable because of the strong space
localization properties of the non-bandlimited kernel functions.
Next, the use of outer harmonic exact approximation will be discussed in
more detail following [19]. A constructive version of the Runge–Walsh theorem
will be developed in terms of bandlimited wavelets. The advantage is that when
using bandlimited wavelets, we do not need the wavelet transform at all positions.
It suffices to know a finite set of linear functionals for each scale J to evaluate
the wavelet transform exactly. In conclusion, each J-level wavelet approximation

(2)
ΦJ ∗ F can be expressed exactly as a finite sum.
Our concept using bandlimited wavelets is presented under the assumption
that the families {Φj (·, ·)}j∈N0 , {Ψj (·, ·)}j∈N0 , and {Ψ̃j (·, ·)}j∈N0 consist of band-
limited kernels such that
ϕj (n) = 0, n = 0, . . . , 2j − 1 (8.2)
and
ϕj (n) = 0, n = 2j , 2j + 1, . . . . (8.3)
In the following we use the notation
Hp,...,q (Ωext ext
σ ) = Harm p,...,q (Ωσ ). (8.4)
Consequently, we have
Φj (x, ·) ∈ H0,...,2j −1 (Ωext
σ ), (8.5)
and
Ψj (x, ·), Ψ̃j (x, ·) ∈ H0,...,2j+1 −1 (Ωext
σ ) (8.6)
for all x ∈ Ωext
σ . Thus the scale spaces and the detail spaces, respectively, fulfill
the relations
Vj = H0,...,2j −1 (Ωext
σ ), Wj ⊂ H0,...,2j+1 −1 (Ωext
σ ). (8.7)
Suppose now that there is known a set {v1 , . . . , vM } of M values vi ,
i = 1, . . . , M , from a potential V (for example, the gravitational potential or
the anomalous potential of the Earth) of class Pot (0) (Σext ) corresponding to lin-
ear (observational) functionals L1 , . . . , LM . Then an extended version of Helly’s
theorem (cf. [76]) tells us that, corresponding to the potential V ∈ Pot (0) (Σext ),
there exists a member F (i.e., a Runge–Walsh approximation of the (anomalous)
potential) of class Hs (Ωextσ ) such that F |Σext is in an (ε/2)-neighbourhood to V
(understood in uniform topology on Σext ) and Li F = vi , i = 1, . . . , M (note that
we may write more accurately F0,...,∞ instead of F to indicate that all Harm n -
spaces generally contribute to the “nature” of F when the Earth’s gravitational
potential is required). Moreover, there exists an element F0,...,m (i.e., a bandlim-
ited approximation to the Runge–Walsh approximation) of class H0,...,m (Ωext σ )
such that the restriction F0,...,m |Σext may be considered to be in (ε/2)-accuracy to
F |Σext uniformly on Σext and, in addition, Li F0,...,m = Li F = vi , i = 1, . . . , M . In
other words, corresponding to a potential V ∈ Pot (0) (Σext ) there exists on Σext
a bandlimited potential in H(Ωext ext
σ ), (namely, F0,...,m ∈ H0,...,m (Ωσ )) consistent
with the original data in ε-accuracy (i.e., vi = Li F = Li F0,...,m , i = 1, . . . , M ).
This is the reason why we are interested in wavelet approximations of potentials
F0,...,m of class H0,...,m (Ωext
σ ) uniformly on Σ
ext from a finite set of functional val-
ues (note that, for the Earth’s anomalous potential, the approximation consistent
with the original data may be found in the class H2,...,m (Ωextσ ) which is a subspace
of H0,...,m (Ωext
σ )).
Our strategy is to represent F0,...,m ∈ H0,...,m (Ωext

σ ) by a J-level approxima-
(2)
tion ΦJ ∗ F0,...,m with J chosen in such a way that 2J+1 − 1 ≥ m (note that
(2)
F0,...,m coincides with ΦJ+1 ∗ F0,...,m uniformly on Σext in the case of Shannon
(2)
wavelets). We want to express the J-level wavelet approximation ΦJ+1 ∗ F0,...,m
of the potential F0,...,m ( with 2J+1 − 1 ≥ m) exactly only by use of the M values
v1 , . . . , vM corresponding to the linear functionals L1 , . . . , LM .
First, our purpose is to apply outer harmonic based approximation formulae.
To this end, we introduce fundamental systems of bounded linear functionals and
derive some approximation formulae. Consider the matrix
⎛ ⎞
L1 H0,1 (σ; ·) ... LN H0,1 (σ; ·)
⎜ .. .. ⎟
m=⎝ . . ⎠ (8.8)
L1 Hm,2m+1 (σ; ·) . . . LN Hm,2m+1 (σ; ·)
m
associated to a system of N ≥ n=0 (2n + 1) = (m + 1)2 (linearly independent)
bounded linear functionals L1 , . . . , LN on H(Ωext σ ). According to well-known ar-
guments of approximation theory, the matrix (8.8) is not of maximal rank for all
systems {L1 , . . . , LN }, N ≥ (m + 1)2 . However, it is clear from a well-known con-
struction principle (see, for example, [19]) that there exist systems {L1 , . . . , LN }
possessing a non-degenerate matrix (8.8).
Definition 8.1. A system {L1 , . . . , LN } of N ≥ (m+1)2 bounded linear functionals
on H(Ωext ext
σ ) is called an H0,...,m (Ωσ )-fundamental system, if the conditions F ∈
ext
H0,...,m (Ωσ ) and Li F = 0, i = 1, . . . , N , imply F = 0.
From Definition 8.1 it is clear that the matrix (8.8) is of maximal rank (m+1)2
if and only if {L1 , . . . , LN } is an H0,...,m (Ωext
σ )-fundamental system. Moreover, it
should be noted that the addition theorem of outer harmonics gives us
⎛ ⎞
L1 L1 KH0,...,m (Ωext ) (·, ·) . . . L1 LN KH0,...,m (Ωext ) (·, ·)
σ σ
⎜ .. .. ⎟
mT m = ⎝ . . ⎠.
LN L1 KH0,...,m (Ωext ) (·, ·) . . . LN LN KH0,...,m (Ωext ) (·, ·)
σ σ
The Gram matrix m m is regular if and only if the system {L1 , . . . , LN } is

T
an H0,...,m (Ωext σ )-fundamental system. Moreover, it is clear that the property of

{L1 , . . . , LN } of being an H0,...,m (Ωext
σ )-fundamental system, is independent of the
choice of the H0,...,m (Ωext σ )-orthonormal basis.
For later use we introduce the following definition.
Definition 8.2. Let Ξ be a regular surface with Ξ ⊂ Ωext σ .
ext
Let {L1 , . . . , LN } be an H0,...,m (Ωσ )-fundamental system of Dirichlet func-
tionals L1 , . . . , LN on H(Ωext
σ ) (i.e., Li F = F (yi ) for yi ∈ Ξ, i = 1, . . . , N and all
F ∈ H(Ωext
σ )). Then the associated system {y1 , . . . , yN } is called an H0,...,m (Ωext
σ )-
Dirichlet-fundamental system on Ξ.
Let {L1 , . . . , LN } be an H0,...,m (Ωext

σ )-fundamental system of Neumann func-
tionals Li , i = 1, . . . , N (i.e., Li F = (λ · (∇F ))(yi ) for yi ∈ Ξ and all F ∈ H(Ωext
σ ))
with λ : Ξ → R3 being a unit vector field satisfying inf x∈Ξ ν(x) · λ(x) > 0
(where ν denotes the outer normal). Then the system {y1 , . . . , yN } is called an
H0,...,m (Ωext
σ )-Neumann-fundamental system on Ξ (relative to λ).
Let {L1 , . . . , LN } be an H0,...,m (Ωext

σ )-fundamental system in the sense of
Definition 8.1. Suppose that F is a potential of class H(Ωextσ ). Furthermore, let P
be an element of H0,...,m (Ωextσ ) with the representation
m 2n+1

P = P ∧ (n, l)Hn,l (σ; ·).
n=0 l=1
Then, for all solutions a ∈ RN , a = (a1 , . . . , aN )T , of the linear system

N
ak Lk Hn,l (σ; ·) = P ∧ (n, l), (8.9)
k=1
n = 0, . . . , m; l = 1, . . . , 2n + 1, we find

N m 2n+1

P = ak (Lk Hn,l (σ; ·)) Hn,l (σ; ·). (8.10)
k=1 n=0 l=1
Observing this fact we get the following theorem.
Theorem 8.3. Let {L1 , . . . , LN } be an H0,...,m (Ωext
σ )-fundamental system of bounded
linear functionals on H(Ωext σ ). Then the identity

N
N
F ∗P = ak Lk F − ak Lk KHm+1,...,∞ (Ωext ) ∗ F
σ
k=1 k=1
holds for all F ∈ H(Ωext

σ ) and all solutions a ∈ R , a = (a1 , . . . , aN ) , satisfying
N T
the linear system (8.9).

By virtue of the Cauchy–Schwarz inequality it follows from Theorem 8.3 that
the estimate

N

F ∗ P − ak Lk F

k=1
N N 1/2 (8.11)

≤ ak as Lk Ls KHm+1,...,∞ (Ωext ) (·, ·) F Hm+1,...,∞ (Ωext )
σ σ
k=1 s=1
holds for all F ∈ H(Ωext

σ ) and all solutions a ∈ R , a = (a1 , . . . , aN ) , satisfying
N T
ext
(8.9). In particular, we have for F ∈ H0,...,m (Ωσ )

N
F ∗P = ak Lk F, (8.12)
k=1
since F Hm+1,...,∞ (Ωext ) = 0. But this shows us that

σ

N
KH0,...,m ∗ P = ak Lk KH0,...,m (Ωext ) (·, ·) (8.13)
σ
k=1
holds for all a ∈ Rn , a = (a1 , . . . , aN )T , satisfying the linear equations (8.9).

Next we adopt a famous criterion due to [73] from Theorem 8.3.
Lemma 8.4. The following statements are equivalent:
N
(i) lim ak Lk Hn,l (σ; ·) = 0, n = m + 1, m + 2, . . .; l = 1, . . . , 2n + 1,
N →∞ k=1

N
(ii) F ∗ P = lim ak Lk F, F ∈ H(Ωext
σ ).
N →∞ k=1
As shown in [19], the definition of fundamental systems and approximation

formulae leads us to exact approximation rules on H0,...,2m (Ωext
σ )-spaces. To this
end we have to summarize shortly some results concerning interpolation by outer
harmonics (see [19]).
We start mentioning the Shannon sampling theorem for the finite-dimensional
space H0,...,m (Ωext
σ ).
Lemma 8.5. Let F be in H0,...,m (Ωext σ ). Assume that {L1 , . . . , LN } forms an

H0,...,m (Ωext
σ )-fundamental system. Then F can be reconstructed from its samples
at the bounded linear functionals L1 , . . . , LN by the following interpolation formula

N
F (x) = (Lk F )PkN (x), x ∈ Ωext
σ ,
k=1
where the “Lagrangians” PkN ∈ H0,...,m (Ωext

σ ), k = 1, . . . , N , are given by

N
PkN = N
wl,k Ll KH0,...,m (Ωext ) (·, ·)
σ
l=1
N
and the coefficients wl,k have to satisfy the linear equations

N
N
wl,k Li Ll KH0,...,m (Ωext ) (·, ·) = δi,k ,
σ
l=1
i, k = 1, . . . , N .
Next we come to some aspects on numerical integration on the sphere. The-
orem 8.3 allows as special cases the following variants.
Lemma 8.6 (Koksma–Hlawka formula of approximation order 0). Let F be of class
H(Ωext
σ ) with {An } being summable in the sense of Definition 3.3. Assume that
{y1N , . . . , yN
N
} is a subset of points on Ωσ . Then the integral formula

1
F (y) dω(y)
4πσ 2 Ωσ
N
N (8.14)
= wkN F (ykN ) − wkN KH1,...,∞ (Ωext ) (· , ykN ), F
σ H1,...,∞ (Ωext
σ )
k=1 k=1
N
holds for all wN = (w1N , . . . , wN
N T
) with k=1 wkN = 1 (e.g., wkN = 1/N ).
Lemma 8.7 (Koksma–Hlawka formula of approximation order m). Let F be a
member of class H(Ωext σ ) with {An } being summable in the sense of Definition
3.3. Assume that {y1N , . . . , yN N
} ⊂ Ωextσ is an H0,...,m (Ωext
σ )-Dirichlet-fundamental
system, i.e., a pointset on the sphere Ωσ such that
⎛ N ⎞
KH0,...,m (Ωext ) (y1N , y1N ) ... KH0,...,m (Ωext ) (y1N , yN )
σ σ
⎜ .. .. ⎟
⎝ . . ⎠
N
KH0,...,m (Ωext ) (yN , y1N ) ... KH0,...,m (Ωext ) (yNN N
, yN )
σ σ
is regular. Then the integral formula

1
F (y) dω(y) (8.15)
4πσ 2 Ωσ
N
N
= wkN F (ykN ) − wkN KHm+1,...,∞ (Ωext ) (· , ykN ), F
σ Hm+1,...,∞ (Ωext
σ )
k=1 k=1
holds for all wN = (w1N , . . . , wN

N T
) , satisfying

N
wlN = 1, (8.16)
l=1

N
wlN Hn,k (σ; ylN ) = 0, n = 1, . . . , m, k = 1, . . . , 2n + 1. (8.17)
l=1
Finally we are interested in an extension of the Koksma–Hlawka formula

for spherical integrals (see Lemma 8.7) to H0,...,m (Ωext σ )-inner products. To this
end we understand the summable sequence {An } generating the reference space
H(Ωext
σ ) to be the symbol of a pseudodifferential operator A with AHn,k (σ; ·) =
A∧ (n)Hn,k (σ; ·) = An Hn,k (σ; ·) for all n ∈ N0 ; k = 1, . . . , 2n + 1. Then the frame-
work of the space H0,...,m (Ωext
σ ) tells us that
m 2n+1

F ∗P = F ∧ (n, k)P ∧ (n, k)
n=0 k=1

= (AF )(y)(AP )(y) dω(y) (8.18)
Ωσ
holds for all F ∈ H(Ωext ext

σ ) and P ∈ H0,...,m (Ωσ ). Moreover, we see that

(AF )(y)(AP )(y) dω(y) = F (y)(A2 P )(y) dω(y).
Ωσ Ωσ
2
Clearly, A P is a member of class H0,...,m (Ωextσ ) (as defined in the foregoing). As-
2
suming F to be of class H0,...,m (Ωext
σ ), F (A P )|Ωext is the product of two elements
σ
2
of class H0,...,m (Ωext
σ ), hence, F (A P )|Ωext is a member of class H0,...,2m (Ωext
σ ). In
σ
connection with Lemma 8.7 this leads us to the following result.
Lemma 8.8. Let F and P be elements of class H0,...,m (Ωext σ ).
Assume that {y1N , . . . , yN
N
} ⊂ Ωσ is an H0,...,2m (Ωext
σ )-Dirichlet-fundamental
system on Ωσ (with N ≥ (2m + 1)2 ). Then the identity

N
F ∗P = wkN F (ykN )(A2 P )(ykN )
k=1
N
holds for all w = (w1N , . . . , wN
N T
) satisfying

N
wlN = 1, (8.19)
l=1

N
wlN Hn,k (σ; ylN ) = 0, n = 1, . . . , 2m; k = 1, . . . , 2n + 1. (8.20)
l=1
In particular, we have

N
KH0,...,m ∗ F = wkN F (ykN )KHarm 0,...,m (Ωext ) (·, ykN ).
σ
k=1
Lemma 8.8 is an essential tool for the development of “tree algorithms” (pyra-
mid schemata) in bandlimited harmonic wavelet theory.
Lemma 8.9. Let the system {y1M , . . . , yM M
} ⊂ Ωσ , M = (2m + 1)2 , define an
H0,...,2m (Ωext
σ )-Dirichlet-fundamental system. Furthermore, suppose that P0,...,m ,
Q0,...,m , respectively, are elements of class H0,...,m (Ωσ ). Then the identity

M
2
P0,...,m ∗ Q0,...,m = bM M M
n P0,...,m (yn )(A Q)0,...,m (yn ) (8.21)
n=1
holds for all weights bM M

1 , . . . , bM satisfying

M
bM M M
r KH0,...,2m (Ωext ) (yi , yr )
σ
r=1

= KH0,...,2m (Ωext ) (yiM , x) dω(x), i = 1, . . . , M. (8.22)
σ
Ωσ
Furthermore, we have the following results.

2 ext
Lemma 8.10. Let {LM 1 , . . . , LM }, M = (m + 1) , be an H0,...,m (Ωσ )-fundamental
M
system, and suppose that P0,...,m and Q0,...,m are members of H0,...,m (Ωext
σ ). Then
the identity
m 2n+1
M
P0,...,m ∗ Q0,...,m = r (Q0,...,m ∗ Hn,k (σ; ·)) Lr P0,...,m
dn,k M
(8.23)
n=0 k=1 r=1
holds for all weights dn,k n,k

1 , . . . , dM ; n = 0, . . . , m; k = 1, . . . , 2n + 1, satisfying the
linear equations
M
r Lr Hl,i (σ; ·) = δn,l δk,i ,

dn,k M
r=1
l = 0, . . . , m; i = 1, . . . , 2l + 1.
In order to reduce the number of weights in our approximation rules we
formulate the following lemma.
Lemma 8.11. Under the assumptions of Lemma 8.10, the formula

M
Q0,...,m ∗ P0,...,m = r Lr P0,...,m
dM M
(8.24)
r=1
holds for all weights dM M

1 , . . . , dM satisfying the linear equations

M
r Li Lr KH0,...,m (Ωext ) (·, ·)

dM M M
σ
r=1
m 2n+1

= LM
i Hn,k (σ; ·) Q0,...,m ∗ Hn,k (σ; ·) = Li Q0,...,m ,
M
(8.25)
n=0 k=1
i = 1, . . . , M.
It should be mentioned that on the one hand the number of integration
weights is reduced, but on the other hand the integration weights depend on
Q0,...,m . Other variants of discretization rules have been presented by W. Free-
den and W. Schneider [30] which allow different aspects of approximation. In this
work, however, we restrict ourselves to the above results (more explicitly, Lemma
8.9, Lemma 8.10, Lemma 8.11) based on linear systems of O(M )-dimension.
In what follows the Runge concept is of basic interest. Once again, it tells
us that to any potential V ∈ Pot (0) (Σext ) (for example, the Earth’s gravitational
potential) there exists a function F (namely, a Runge–Walsh approximation) har-
monic in Ωextσ and being regular at infinity in the sense that the absolute error
becomes arbitrarily small on the whole space Σext . In this formulation as we already
mentioned, the Runge–Walsh theorem is a pure existence theorem. It guarantees
only the existence of an approximating potential and does not provide a method
to find it. The theorem merely describes the theoretical background of approxi-
mating a potential by another one defined on a larger harmonicity domain. The
results developed now, however, enable us to derive a constructive version of the

Runge–Walsh theorem by means of a J-level wavelet approximation when the po-
tential F we are looking for is assumed to be a member of class H(Ωext
σ )|Σext (note
ext (0) ext
that H(Ωσ )|Σext is a uniformly dense subset of Pot (Σ )). Essential tools of
our considerations are the approximation formulae formulated above.
2 ext
1 , . . . , LM }, M = (m+1) , be an H0,...,m (Ωσ )-fundamental
Theorem 8.12. Let {LM M
system. Furthermore, suppose that {y1 , . . . , yMj } ⊂ Ωσ , Mj = (2mj + 1)2 , de-

Mj Mj
fine H0,...,2mj (Ωext

σ )-Dirichlet-fundamental systems for j = 0, . . . , J. Moreover,
assume that from a potential F0,...,m ∈ Harm 0,...,m (Ωext σ ) there are known the
data LMi F 0,...,m = vi , i = 1, . . . , M . Then, under our assumption of bandlim-
ited wavelets, the fully discrete J-level wavelet approximation of F0,...,m reads as
follows:
(2)
(α) ΦJ ∗ F0,...,m

M0 m 2k+1
M
= b0n dk,l 2
s Ak ϕ0 (k)vs Hk,l (σ; yn )Φ0 (yn , ·)
M0 M0
n=1 k=0 l=1 s=1

J−1 Mj m 2k+1
M
2
bjn s Ak ψj (k)vs Hk,l (σ; yn )Ψ̃j (yn , ·),
dk,l Mj Mj
+ (8.26)
j=0 n=1 k=0 l=1 s=1
where the weights dk,l k,l

1 , . . . , dM ; k = 0, . . . , m; l = 1, . . . , 2k + 1, satisfy the
linear equations

M
s Ls Hn,i (σ; ·) = δn,k δi,l ,

dk,l M
(8.27)
s=1
n = 0, . . . , m; i = 1, . . . , 2n + 1, and bj1 , . . . , bjMj ; j = 0, . . . , J, satisfy the

linear equations

Mj
M M
bjn KH0,...,2m (Ωext ) (yi j , ynMj ) = KH0,...,2m (Ωext (yi j , x) dω(x),
j σ j σ )
n=1 Ωσ
(8.28)
i = 1, . . . , Mj .
(2)

M0
M
J−1 Mj

M
(β) ΦJ ∗ F0,...,m = b0n d˜0,n
s vs Φ0 (yn , ·) +
M0
bjn s vs Ψ̃j (yn , ·),
dj,n Mj
n=1 s=1 j=0 n=1 s=1

(8.29)
where the weights d˜0,n ˜0,n
1 , . . . , dM ; n = 1, . . . , M0 , satisfy the linear equations

M
d˜0,n M M M 2
s Li Ls KHarm 0,...,m (Ωext ) (·, ·) = Li (A Φ0 )(yn , ·),
M0
(8.30)
σ
s=1
i = 1, . . . , M , and the weights dj,n j,n

1 , . . . , dM ; j = 0, . . . , J; n = 1, . . . , Mj ,
satisfy

M
2
s Li Ls KHarm 0,...,m(Ωext ) (·, ·) = Li (A Ψj )(yn , ·),
dj,n M M M Mj
(8.31)
σ
s=1
i = 1, . . . M , and the coefficients bj1 , . . . , bjMj ; j = 0, . . . , J satisfy the linear

system (8.28).
It should be remarked that a great number of linear systems must be solved
in an a priori step. But if we look carefully we realize that we are always confronted
with the same coefficient matrix. Having inverted the coefficient matrix once, all
weights for numerical integration can be obtained by a matrix-vector multiplication
and stored elsewhere (in an a priori step for computation). In addition, it should
be mentioned that the solution of the linear systems determining the weights of
the reconstruction step (8.28) can be avoided completely if we place the knots for
numerical integration of the wavelet coefficients for each detail step j = 0, . . . , J −1
on a special longitude-latitude grid on the sphere Ωσ . The corresponding set of
integration weights for reconstruction purposes are explicitly available without
solving any linear system (for more details concerning numerical integration the
reader is referred, e.g., to a paper due to Driscoll Healy [8]).
Until now the linear (observational) functionals have not been specified in
more detail in our bandlimited wavelet approach presented above. In fact, the
different types of linear functionals enable us to develop three important variants
of wavelet approximation in the reality of gravitational potential determination:
(1) Terrestrial-only Multiscale Approximation. The linear functionals are under-
stood to represent gravity observations (function values and/or derivatives)
related to locations on the Earth’s surface. If the data material is homoge-
neous, i.e., the linear functionals are all of the same type, terrestrial-only
approximation reduces to the wavelet solution of a boundary-value problem
of potential theory from discretely given data.
(2) Spaceborne-only Multiscale Approximation. In this case the linear functionals
are understood to represent data measured by spacecraft in locations of Ωext γ .
As result we get a spaceborne-only approximation.
In practice, however, we are confronted with the situation that terrestrial,
airborne as well as spaceborne data are available in gravitational potential deter-
mination (cf. [1, 2, 19, 22, 32, 45, 46, 50, 60, 62, 63, 66, 71, 75]). As a matter of
fact, there are some areas on the continents (for example, some parts of Australia,
Europe, and North-America), where the gravity field has been surveyed in much
detail. Thus it is reasonable that such areas may be used for the verification or
the calibration of the results obtained from spaceborne data.
(3) Combined Multiscale Approximation. Linear functionals representing terres-
trial, airborne, and spaceborne observations are taken into account, i.e., nu-
merical computation is required for a heterogeneous data set.
8.1. Runge–Walsh wavelet approximation of classical boundary value problems

corresponding to regular surfaces
The wavelet representations (Theorem 8.12) of a bandlimited potential from a
given finite set of linear functionals admit a variety of applications. The list includes
the following examples of classical boundary value problems:
(i) Dirichlet Problem. First we are interested in the wavelet approximation
(2)
ΦJ ∗ F0,...,m of the solution of the exterior Dirichlet problem
F0,...,m |Σext ∈ Harm 0,...,m (Σext ), F0,...,m |Σ = G0,...,m .
2
under the knowledge of the M = (m + 1) boundary data
vi = LM M M
i F0,...,m = F0,...,m (xi ) = G0,...,m (xi ), i = 1, . . . , M.
Theorem 8.13. Under the assumptions of Theorem 8.12 the fully discrete J-level
wavelet approximation of the solution of the exterior Dirichlet problem F0,...,m |Σext
∈ Harm 0,...,m (Σext ), (F0,...,m )|Σ = G0,...,m reads as follows:
(2)
(α) ΦJ ∗ F0,...,m (8.32)

M0 m 2k+1 M
= b0n 2
s Ak ϕ0 (k)G0,...,m (xs )Hk,l (σ; yn )Φ0 (yn , ·)
dk,l M M0 M0
n=1 k=0 l=1 s=1

J−1 Mj
m 2k+1
M
2
bjn s Ak ψj (k)G0,...,m (xs )Hk,l (σ; yn )Ψ̃j (yn ; ·)
dk,l M Mj Mj
+
j=0 n=1 k=0 l=1 s=1
(2)

M0 M
(β) ΦJ ∗ F0,...,m = b0n d˜0,n
s G0,...,m (xs )Φ0 (yn , ·)
M M0
(8.33)
m=1 s=1

J−1 Mj
M
bjn s G0,...,m (xs )Ψ̃j (yn , ·).
dj,n M Mj
+
j=0 n=1 s=1
The formulae (α), (β) of Theorem 8.13 are especially valid on the regular
(2)
(Earth’s) surface Σ, i.e., we automatically obtain by ΦJ ∗ F0,...,m |Σ a J-level
wavelet approximation of the “boundary function” F0,...,m |Σ = G0,...,m (by apply-
(2)
ing Shannon wavelets we even know that ΦJ ∗ F0,...,m = F0,...,m ). In other words,
a wavelet representation of a (bandlimited) function on regular surfaces has been
found from a discrete data set of function values.
By treating non-bandlimited potentials F ∈ Hs (Ωext
σ ), s > 1, the developed
integration formulae are only valid in approximate sense. To be more concrete,
(2)
if ΦJ ∗ F denotes the J-level wavelet approximation we actually calculate an
(2)
approximation ΦJ ∗ F0,...,m by performing the numerical integration methods
in (α), (β) of Theorem 8.13. Since this approximation also is harmonic in Σext the
(2) (2)
biggest absolute error between ΦJ ∗ F and its numerical approximation ΦJ ∗
F0,...,m is attained at the boundary Σ. Thus, the numerical error can be estimated
by the use of the following theorem (cf. [29, 30]).
Theorem 8.14. Let F satisfy F ∈ Hs (Ωext

σ ), F |Σ = G, s > 1. Furthermore, assume
that XM = {x1 , . . . , xM } ⊂ Σ, M = (m + 1)2 , is an H0,...,m (Ωext
Σ M M
σ )-Dirichlet-
fundamental system on Σ. Then, for any Q ∈ H0,...,m (Ωext
σ ), we have
M
M
M C
F (x)Q(x) dω(x) − dr G(xr ) ≤ s−1 |dr | F Hs (Ωext ) ,
M
Ωσ m σ
r=1 r=1
(8.34)
where C is a constant depending only on s and dM M

1 , . . . , dM are the weights of the
integration rule.
(2)
(ii) Neumann Problem. Now we are interested in the wavelet approximation ΦJ ∗
F0,...,m of the solution of the oblique Neumann problem
∂F0,...,m
F0,...,m |Σext ∈ Harm 0,...,m (Σext ), = G0,...,m ,
∂λ
under the knowledge of the M = (m + 1)2 boundary data
∂F0,...,m M
vi = LM
i F0,...,m = (xi ) = G0,...,m (xM
i ), i = 1, . . . , M,
∂λ
where λ : Σ → R3 is a C [1,ρ) -unit vector field (such that 0 < ρ < 1 for λ = ν
and ρ = 0 for λ = ν) forming an angle with the outer normal ν satisfying
inf ν(x) · λ(x) > 0 (8.35)

x∈Σ
at any point of Σ.
Note that the boundedness of the linear functionals of the oblique derivative on Σ
follows from well-known arguments (cf. [16, 18, 20]).
For the decomposition step we need in contrast to the Dirichlet problem an
integration method in terms of oblique derivatives on Σ. From our results we obtain
a fully discrete wavelet approximation for the solution of the exterior Neumann
problem.
Σ 2 ext
Theorem 8.15. Let XM = {xM
1 , . . . , xM } ⊂ Σ, M = (m+ 1) , be an H0,...,m (Ωσ )-
M
Mj Mj
Neumann-fundamental system on Σ. Furthermore, let XMj = {y1 , . . . , yMj },
Mj = (2mj + 1)2 , be H0,...,2mj (Ωext σ )-Neumann-fundamental systems on Ωσ for
j = 0, . . . , J. Moreover, assume that from a function F0,...,m ∈ H0,...,m (Ωext
σ ) there
Σ
are known the oblique derivatives G0,...,m = (∂F0,...,m /∂λ) at all points of XM .
Then, under our assumption of bandlimited wavelets, the fully discrete J-level
wavelet approximation of the solution of the exterior Neumann problem F0,...,m ∈
H0,...,m (Σext ), (∂F0,...,m )/∂λ = G0,...,m reads as follows:

(2)
(α) ΦJ ∗ F0,...,m

M0 m 2k+1
M
= b0n 2
s Ak ϕ0 (k)G0,...,m (xs )Hk,l (σ; yn )Φ0 (yn , ·)
dk,l M M0 M0
n=1 k=0 l=1 s=1

J−1 Mj m 2k+1
M
2
bjn s Ak ψj (k)G0,...,m (xs )Hk,l (σ; yn )Ψ̃j (yn , ·),
dk,l M Mj Mj
+
j=0 n=1 k=0 l=1 s=1
(8.36)
where the weights dk,l k,l

1 , . . . , dM ; k = 0, . . . , m; l = 1, . . . , 2k + 1 have to satisfy the
linear equations

M
∂Hn,i (σ; xM
s )
dk,l
s = δn,k δi,l , n = 0, . . . , m; i = 1, . . . , 2n + 1,
s=1
∂λ
and bj1 , . . . , bjM , j = 0, . . . , J must satisfy the linear equations (8.28).

(2)
(β) ΦJ ∗ F0,...,m

M0
M
= b0n d˜0,n
s G0,...,m (xs )Φ0 (yn , ·)
M M0
n=1 s=1

J−1 Mj

M
bjn s G0,...,m (xs )Ψ̃j (yn , ·),
dj,n M Mj
+ (8.37)
j=0 n=1 s=1
where the weights d˜0,n ˜0,n

1 , . . . , dM ; n = 1, . . . , M0 , have to satisfy the linear equations

M
∂ ∂ ∂
d˜0,n
s KH0,...,m (Ωext
σ )(·, ·) = (A2 Φ0 )(ynM0 , ·), (8.38)
s=1
∂λyiM ∂λysM ∂λyiM
i = 1, . . . , M , and the weights dj,n j,n

1 , . . . , dM ; j = 0, . . . , J; n = 1, . . . , Mj , must
satisfy

M
∂ ∂ ∂
dj,n K ext (·, ·) = (A2 Ψj )(ynMj , ·), (8.39)
s=1
s
∂λyiM ∂λysM H0,...,m (Ωσ ) ∂λyiM
i = 1, . . . , M , and bj1 , . . . , bjMj ; j = 0, . . . , J, satisfy the linear equations (8.28).
The formulae (α), (β) of Theorem 8.15 are especially valid on Σ. Thus, we
(2)
obtain by ∂(ΦJ ∗ F0,...,m )/∂λ a J-level wavelet approximation of G0,...,m =
∂F0,...,m /∂λ.
In order to examine the error in the integration formulae when we turn over
to non-bandlimited potentials we finally mention the following theorem.
Theorem 8.16. Let F satisfy F ∈ Hs (Ωext ∂F

σ ), ∂λ = G, s > 2. Furthermore, let
Σ 2 ext
XM = {xM1 , . . . , xM } ⊂ Σ, M = (m + 1) , be an H0,...,m (Ωσ )-Neumann-funda-
M
ext
mental system on Σ. Then, for any Q ∈ H0,...,m (Ωσ ), we have
M
M
M C
F (x)Q(x) dω(x) − dr G(xr ) ≤ s−2
M
|dr | F Hs (Ωext ) ,
M
Ωσ m σ
r=1 r=1
(8.40)
where C is a constant depending only on s and dM M

1 , . . . , dM are the weights of the
integration rule.
Hence, by treating non-bandlimited potentials F ∈ Hs (Ωext σ ), s > 2, we

obtain in similarity to the Dirichlet case a J-level wavelet approximation by per-
forming the numerical rules as indicated by (α), (β) of Theorem 8.15, and the
numerical errors can be estimated using Theorem 8.16.
Remark 8.17. The existence of all types of fundamental systems to be needed

in our preceding approximation rules is guaranteed by a well-known induction
procedure (as described, for example in [21, 24, 57]. Furthermore, more detailed
remainder estimates for the integration formulae can be found in [28]).
8.2. Pyramid schemata based on outer harmonic exact approximation

Our purpose now is to use two variants of exact (outer harmonic) approximation
to derive tree algorithms, i.e., pyramid schemata for fast evaluation of bandlimited
potentials. Without loss of generality, we assume that {Φj (·, ·)}j∈N0 , {Ψj (·, ·)}j∈N0 ,
and {Ψ̃j (·, ·)}j∈N0 are families of bandlimited kernels satisfying the conditions (8.2)
and (8.3). Variant 1 is based on the ideas of Lemma 8.8 using evaluation (i.e.
Dirichlet functionals) on a sphere, while Variant 2 is based on the Shannon sam-
pling Theorem 8.5 in terms of linear functionals. Both variants are particularly
suitable for application to medium wavelength parts of a signal (potential). As
shown in [19], Variant 2 can be extended to non-bandlimited potentials. This vari-
ant is therefore also suitable for the transition from medium to short wavelength
parts of a signal (potential).
Variant 1. The key ideas of our first discretization method using outer harmonic
exact approximation formulae are based on the following observations:
(1) For some suitably large J, the scale space VJ+1 (Ωext ext
σ ) = H0,...,2J+1 −1 (Ωσ )
is “sufficiently close” to H(Ωextσ ). Consequently, for each potential F ∈ H(Ωσ ),
ext
there exists a bandlimited potential of class VJ+1 (Ωext

σ ) such that the error between
(2)
F and ΦJ+1 ∗F (understood in · H(Ωext ) -topology) is negligible. This is the reason
σ
why the input data vlNJ , l = 1, . . . , NJ , are assumed to be given from a potential
of class VJ+1 (Ωext
σ ) (for the remainder of this subsection).
N
(2) For j = 0, . . . , J, the generating coefficients bl j and nodal points
N
yl j ∈ Ωσ of the exact outer harmonic formulae of order 2j+2 − 2(= 2 · (2j+1 − 1))
(cf. Lemma 8.8) are determined such that

Nj
N N N
KH (Ωext ∗P = bl j KHarm (Ωext (·, yl j )P (yl j )
0,...,2j −1 σ ) 0,...,2j −1 σ )
l=1
holds for all P ∈ H0,...,2j −1 (Ωext

σ ) with Nj ≥ ((2
j+2
− 2) + 1)2 = (2j+2 − 1)2 . The
Nj
coefficients bl may be calculated from the linear equations

Nj
N N N
bl j K H (Ωext (yi j , yl j )
0,...,2j+2 −2 σ )
l=1
1 N
= KH ext (x, yi j ) dω(x), (8.41)
4πσ 2 0,...,2j+2 −2 (Ωσ )
Ωσ
i = 1, . . . , Nj , in an a priori step and stored elsewhere.

Our goal is to show that all convolutions occurring in the J-level wavelet
approximation of a bandlimited potential (of order 2J+1 − 1) can be evaluated
exactly by means of outer harmonic approximation formulae. As a matter of fact,
what we realize is the following pyramid scheme: Starting from a sufficiently large
J, there exist vectors aNj ∈ RNj , j = 0, . . . , J (being, of course, dependent on the
potential F ∈ H(Ωext σ ) under consideration) such that the following statements
hold true:
(i) For j = 0, . . . , J, all wavelet coefficients can be calculated via the formulae
N j N N
(W T )(F )(j; ·) = ai j Ψj (·, yi j ).
i=1
(ii) The vectors aj ∈ RNj are obtainable from aj+1 ∈ RNj+1 by recursion:
Nj Nj
Nj+1 Nj+1 N Nj+1
ai = bi al KHarm ext (yi j , yl ),
l=1 0,...,2j+1 −1 (Ωσ )
i = 1, . . . , Nj .
(iii) The vectors satisfy, in addition, the identities
(2)

Nj
N (2) N
Φj+1 ∗ F = ai j Φj+1 (·, yi j )
i=1
and

Nj
N N
(Ψ̃j ∗ Ψj ) ∗ F = ai j (Ψ̃j ∗ Ψj )(·, yi j ).
i=1
Our considerations are divided into two parts, viz. the initial step concerning the
scale level J and the pyramid step establishing the recursion relation.
The Initial Step. Observing the exact (outer harmonic) formulae we obtain from
Lemma 8.8 for all potentials F ∈ VJ+1 (Ωext ext
σ ) = H0,...,2J+1 −1 (Ωσ )

NJ
KH (Ωext ∗F = bN J NJ
l F (yl )KHarm (Ωext (·, ylNJ ).
0,...,2J+1 −1 σ ) 0,...,2J+1 −1 σ )
l=1
It follows that a NJ
∈R NJ
, aNJ = (aN J NJ T
1 , . . . , aNJ ) , given by
aN
l
J
= bN J NJ NJ NJ
l F (yl ) = bl vl , l = 1, . . . , NJ , (8.42)
satisfies the equation

NJ
KH (Ωext ∗F = aN J
i KHarm (Ωext (·, yiNJ ).
0,...,2J+1 −1 σ ) 0,...,2J+1 −1 σ )
i=1
Note that the coefficients aN

i
J
are dependent on F . Again Lemma 8.8 now implies
the following result.
Lemma 8.18. Let F be of class VJ+1 (Ωext ext
σ ) = H0,...,2J+1 −1 (Ωσ ). Suppose that
∧
K(·, ·) is (an Hσ,σ -kernel) such that K (n) = 0 for all n > 2 J+1
− 1. Then the
coefficients (8.42) satisfy the equation

NJ
2
K ∗F = aN J NJ
i A K(·, yi ).
i=1
It should be noted that

2J+1
−1 2n+1

2
A K(x, y) = A2n K ∧ (n) ∗
Hn,k (σ; x)Hn,k (σ; y) (8.43)
n=0 k=1
for all (x, y) ∈ Ωext ext

σ × Ωσ . Furthermore, the vector a
NJ
is independent of the
choice of the Hσ,σ -kernel K(·, ·).
As special cases we obtain from Lemma 8.18 the following identities:

NJ
2
ΦJ+1 ∗ F = aN J NJ
i A ΦJ+1 (·, yi ), (8.44)
i=1

NJ
2
(ΦJ+1 ∗ ΦJ+1 ) ∗ F = i A (ΦJ+1 ∗ ΦJ+1 )(·, yi ),
aN J NJ
(8.45)
i=1
and

NJ
2
ΨJ ∗ F = aN J NJ
i A ΨJ (·, yi ), (8.46)
i=1

NJ
2
(Ψ̃J ∗ ΨJ ) ∗ F = i A (Ψ̃J ∗ ΨJ )(·, yi ).
aN J NJ
(8.47)
i=1
The Pyramid Step. An essential tool for the pyramid step is the following lemma.
Lemma 8.19. Let F be of class VJ+1 (Ωext

σ ). Suppose that K(·, ·) is an Hσ,σ -kernel
with K ∧ (n) = 0 for all n > 2J − 1. Then the vector aNJ−1 ∈ RNJ−1 , aNJ−1 =
N N
(a1 J−1 , . . . , aNJ−1
J−1
)T , given by
NJ−1 NJ−1 NJ−1
ai = bi (KH (Ωext ∗ F )(yi ), i = 1, . . . , NJ−1 ,
0,...,2J −1 σ )
satisfies the equation

NJ−1
A2 K(·, yi
NJ−1 NJ−1
K∗F = ai ).
i=1
Suppose that K(·, ·) satisfies the assumption of Lemma 8.19. Looking at our
foregoing results we notice that there are two ways of discretizing an H-convolution
K ∗ F . On the one hand we obtain from Lemma 8.18

NJ
2
K ∗F = aN J NJ
i A K(·, yi ) (8.48)
i=1
with coefficients aN J NJ
1 , . . . , aNJ given by
aN
i
J
= bN J NJ NJ NJ
i F (yi ) = bi vi , i = 1, . . . , NJ . (8.49)
It is remarkable that the coefficients are independent of the choice of the kernel
K(·, ·). As particularly important case we mention

NJ
KH ext ∗F = aN J
i KHarm ext (yiNJ , ·). (8.50)
0,...,2J −1 (Ωσ ) 0,...,2J −1 (Ωσ )
i=1
On the other hand, we are able to deduce from Lemma 8.19 that

NJ−1
A2 K(·, yi
NJ−1 NJ−1
K ∗F = ai ) (8.51)
i=1
N N
with coefficients a1 J−1 , . . . , aNJ−1
J−1
given by
NJ−1 NJ−1 NJ−1
ai = bi (KH ext ∗ F )(yi ), (8.52)
0,...,2J −1 (Ωσ )
i = 1, . . . , NJ−1 . Inserting (8.50) into (8.52) we find
NJ−1 NJ−1

NJ
NJ−1
ai = bi aN J
l KHarm (Ωext (yi , ylNJ ) (8.53)
0,...,2J −1 σ )
l=1
N
for i = 1, . . . , NJ−1 . In other words, the coefficients ai J−1 can be calculated recur-
sively. Moreover, the coefficients are independent of the special choice of the kernel
K(·, ·). This finally leads us to the following discretization of the H-convolutions
NJ−1

A2 ΦJ (·, yi
NJ−1 NJ−1
ΦJ ∗ F = ai ), (8.54)
i=1
NJ−1

A2 (ΦJ ∗ ΦJ )(·, yi
NJ−1 NJ−1
(ΦJ ∗ ΦJ ) ∗H F = ai ), (8.55)
i=1
and
NJ−1

A2 ΨJ−1 (·, yi
NJ−1 NJ−1
ΨJ−1 ∗ F = ai ), (8.56)
i=1
NJ−1

A2 (Ψ̃J−1 ∗ ΨJ−1 )(·, yi
NJ−1 NJ−1
(Ψ̃J−1 ∗ ΨJ−1 ) ∗ F = ai ). (8.57)
i=1
In conclusion, we end up with the following pyramid scheme for the decomposition
of a potential F :
F −→ aNJ −→ aNJ−1 −→ ··· −→ aN0
↓ ↓ ↓
(W T )(F )(J; ·) (W T )(F )(J − 1; ·) (W T )(F )(0; ·).
The reconstruction of the wavelet coefficients can be performed as described before
via the formula
Rj (F ) = Ψ̃j ∗ (W T )(F )(j; ·)

Nj
bi j (W T )(F )(j; yi j )A2 Ψ̃j (·, yi j ).
N N N
= (8.58)
i=1
This leads us to the following scheme:

(W T )(F )(0; yiN0 ) (W T )(F )(1; yiN1 )
↓ ↓
R0 (F ) R1 (F )

P0 (F ) → + P1 (F ) → + ··· .
According to our approach the wavelet transform (W T )(F )(j; ·) is given by the
N N
coefficients a1 j , . . . , aNjj . This also enables us to reconstruct the potential only by
N
use of the coefficients ai j , rather than calculating the wavelet coefficients of F :

Nj
ai j A2 (Ψ̃j ∗ Ψj )(·, yi j ).
N N
Rj (F ) =
i=1
Thus the decomposition and reconstruction, respectively, can be simplified as fol-

lows:
F → aNJ → aNJ−1 → · · · → aN0
and
aN 0 aN 1 aN 2
↓ ↓ ↓
R0 (F ) R1 (F ) R2 (F )

P0 (F ) → + → P1 (F ) → + → P2 (F ) → + → ··· .
That means the reconstruction of the potential is not performed with Ψ̃j . Instead
we have used the H(Ωextσ )-convolution Ψ̃j ∗ Ψj . Of particular significance is that
the vectors aNj do not depend on the special choice of the bandlimited scaling
function. As a matter of fact, we are able to reconstruct the potential with respect
to different types of wavelets just by use of the vectors aNj .
Remark 8.20. The critical point of our pyramid scheme is the determination of the
N
coefficients bl j , j = 0, . . . , J, from the linear system (8.41) which provides outer
harmonic exactness up to the order 2j+2 − 2. It should be mentioned that the
solution of this linear system can be avoided completely if we place the knots for
each detail step j = 0, . . . , J on a spherical longitude-latitude grid on the sphere
Ωσ . The corresponding set of weights is explicitly available without solving any
linear system from results due to [8].
Variant 2. In what follows we use outer harmonic exact approximation (Lemma

8.5) to develop a bandlimited variant of the pyramid scheme based on the Shannon
sampling theorem. Our approach consists of the following steps:
(i) According to our bandlimited wavelet approach the (reference) Sobolev space
2j
H(Ωext
σ ) is subdivided by a nested sequence of 2 -dimensional scale spaces
Vj (Ωext ext ext ext
σ ) as follows: · · · ⊂ Vj (Ωσ ) ⊂ Vj+1 (Ωσ ) ⊂ · · · ⊂ H(Ωσ ).
(ii) Vj (Ωext
σ ), j ∈ N 0 , can be identified with the set
2
H0,...,2j −1 (Ωext ext
σ ) = H({An /(ϕj (n)) }; Ωσ ),
(4)
and Φj (·, ·) is the uniquely determined reproducing kernel in (Vj (Ωext
σ ),
(·, ·)Vj (Ωext ) ) with (·, ·)Vj (Ωext ) given by
σ σ
(·, ·)Vj (Ωext ) = (·, ·)H({An /(ϕj (n))2 };Ωext ) .

σ σ
22j (linearly
N N
(iii) For each j ∈ N0 , consider sequences {L1 j , . . . , LNjj } of Nj ≥
independent) bounded linear functionals on H(Ωext σ ) such that

Nj (4) Nj (4)
Vj (Ωext
σ ) = span L1 Φ j (·, ·), . . . , L Φ
Nj j (·, ·) .
Then it also follows that

Nj (2) Nj (2)
Vj (Ωext
σ ) = span L1 Φj (·, ·), . . . , LNj Φj (·, ·) .
(iv) Vj (Ωext
σ ), j ∈ N0 , can be identified with the set H({An /ϕj (n)};Ωσ ),
ext
(2)
and Φj (·, ·) is the reproducing kernel in Vj (Ωext
σ ), (·, ·)V (1/2) (Ωext ) with
j σ
(·, ·)V (1/2) (Ωext ) defined by
j σ
(·, ·)V (1/2) (Ωext ) = (·, ·)H({An /ϕj (n)};Ωext ) .

j σ σ
The key idea of our fast evaluation method using the Shannon sampling
theorem in terms of linear functionals is based on the following observations:
(1) For some suitably large J, the scale space VJ (Ωext
σ ) is “sufficiently close”
ext ext
to H(Ωσ ). Consequently, for each F ∈ H(Ωσ ), there exists a function of class
(2)
VJ (Ωext
σ ) such that the error between F and ΦJ ∗ F (understood in · H(Ωext ) - σ
topology) is negligible. This is the reason why the input data vkNJ = LN J
k F,
ext
k = 1, . . . , NJ , are assumed to be of a potential F of class VJ (Ωσ ) for the re-
mainder of this subsection.
(2) For j = 0, . . . , J, consider sequences {L1 j , . . . , LNjj } of Nj ≥ 22j (linearly
N N
independent) bounded linear functionals on H(Ωext

σ ) such that

N (2) Nj (2)
Vj (Ωext ext
σ ) = H0,...,2j −1 (Ωσ ) = span L1 Φj (·, ·), . . . , LNj Φj (·, ·) .
j
N
In an a priori step the coefficients wl,kj have to be determined from the systems
of linear equations (see Lemma 8.5)

Nj
N N N (2)
wl,kj Li j Ll j Φj (·, ·) = δi,k , i, k = 1, . . . , Nj ,
l=1
j = 0, . . . , J, and can be stored elsewhere. Looking carefully at the linear systems,

N
it can be recognized that the coefficients wl,kj do not depend on the particular
function F under consideration, but only on the chosen linear functionals and
pointsets.
Next our considerations are divided into two parts, viz. the initial step con-
cerning the scale level J and the pyramid step establishing the recursion relation.
The Initial Step. The exact approximation

NJ
JNJ S = i Li S,
aN J NJ
S ∈ VJ (Ωext
σ ),
i=1
to the bounded linear functionals L on VJ (Ωext

σ ) defined by
LS = (S, F )V (1/2) (Ωext ) = S ∗V (1/2) F, S ∈ VJ (Ωext ext

σ ), F ∈ VJ (Ωσ ),
J σ J
is given by

NJ
(2)
aN
i
J
= NJ
wi,k LLN
k ΦJ (·, ·),
J
i = 1, . . . , NJ .
k=1
Note that in order to clarify the convolution we use a lower index at the sym-
bol “∗” in the following text if necessary. In accordance with our assumption
(2) (1/2)
F ∈ VJ (Ωext
σ ) and the reproducing property of ΦJ (·, ·) in VJ (Ωext
σ ) we see that
(2)
ΦJ ∗V (1/2) F = F . Thus we find
J

NJ
(2)

NJ
NJ
aN k ΦJ (·, ·) ∗V (1/2) F ) = Lk F =
NJ
i
J
= wi,k (LN J NJ NJ
wi,k NJ NJ
wi,k vk
J
k=1 k=1 k=1
for i = 1, . . . , NJ . This leads us to the following conclusion.
Lemma 8.21. If F is a member of class VJ (Ωext
σ ), then the identity

NJ
S ∗V (1/2) F = i Li S
aN J NJ
J
i=1
holds for all S ∈ VJ (Ωext

σ ).
Lemma 8.21 immediately enables us to formulate the following lemma.

Lemma 8.22. Let F be a member of class VJ (Ωext

NJ
K ∗F = i Li K(·, ·)
aN J NJ
i=1
holds for all Hσ,σ -kernels K(·, ·) with K ∧ (n) = 0 for n = 2J ,J +1, . . ..
The next theorem clarifies the remarkable consequences for our wavelet con-
cept.
Theorem 8.23. Under the assumptions of Lemma 8.22 we have

NJ
ΦJ ∗ F = i Li ΦJ (·, ·),
aN J NJ
(8.59)
i=1

NJ
(ΦJ ∗ ΦJ ) ∗ F = i Li (ΦJ ∗ ΦJ )(·, ·),
aN J NJ
(8.60)
i=1
and

NJ
ΨJ−1 ∗ F = i Li ΨJ−1 (·, ·),
aN J NJ
(8.61)
i=1

NJ
(Ψ̃J−1 ∗ ΨJ−1 ) ∗ F = i Li (Ψ̃J−1 ∗ ΨJ−1 )(·, ·).
aN J NJ
(8.62)
i=1
1 , . . . , aN J ) ∈ R
In conclusion, the vector aNJ = (aN J NJ T NJ
does not depend on
(2)
the special choice of the ΦJ (·, ·)-kernel in VJ (Ωext
σ ). Wavelet transform, lowpass,
and bandpass filter can be computed by use of the same set of coefficients.
The Pyramid Step. This step provides an algorithm such that aNJ ∈ RNJ serves
as starting vector for aNj ∈ RNj , j = 0, . . . , J − 1, which fulfill the following
properties:
(i) The vectors aNj satisfy
(2)

Nj
N N (2)
Φj ∗F = ai j Li j Φj (·, ·),
i=1
j = 0, . . . , J.
(ii) The wavelet transforms are given by

Nj
N N
Ψj−1 ∗ F = ai j Li j Ψj−1 (·, ·),
i=1
j = 1, . . . , J.
(iii) The vector aNj is obtainable from aNj+1 , j = 0, . . . , J − 1, by recursion.
In the remainder of this section the properties (i), (ii) and (iii) are described
in more detail. The exact approximations JNj , j = 0, . . . , J − 1,

Nj
N N
JN j S = ai j Li j S, S ∈ Vj (Ωext
σ )
i=1
to the bounded linear functional L on Vj (Ωext

σ ) defined by
(2)
LS = S ∗V (1/2) (Φj ∗H F ), S ∈ Vj (Ωext ext
σ ), F ∈ VJ (Ωσ ),
j
(2)
(note that Φj ∗H F ∈ Vj (Ωext
σ )) are given by the coefficients
Nj

Nj
N N (2)
al = wl,ij Li j Φj (·, ·), l = 1, . . . , Nj .
i=1
Consequently it is easily seen that for l = 1, . . . , Nj
Nj

Nj
N N (2)
al = wl,ij Li j (Φj (·, ·) ∗ F ).
i=1
Thus we obtain the following lemma.
Lemma 8.24. If F is a member of class Vj (Ωext
(2)

Nj
N N
S ∗V (1/2) (Φj ∗H F ) = ai j Li j S
j
i=1
holds for all S ∈ Vj (Ωext

σ ). In particular,
(2)

Nj
N N (2)
Φj ∗H F = ai j Li j Φj (·, ·).
i=1
By the same arguments as given in the last subsection we obtain the following
lemma.
Lemma 8.25. Let F be a function of class Vj (Ωext

Nj
N N
K ∗F = ai j Li j K(·, ·)
i=1
holds for all Hσ,σ -kernels K(·, ·) with K ∧ (n) = 0, n = 2j , 2j + 1, . . ..

Finally we get the following results.
Theorem 8.26. Under the assumptions of Lemma 8.25 we have

Nj
N N
Φj ∗ F = ai j Li j Φj (·, ·),
i=1

Nj
N N
(Φj ∗ Φj ) ∗ F = ai j Li j (Φj ∗ Φj )(·, ·),
i=1
and

Nj
N N
Ψj−1 ∗ F = ai j Li j Ψj−1 (·, ·),
i=1

Nj
N N
(Ψ̃j−1 ∗ Ψj−1 ) ∗ F = ai j Li j (Ψ̃j−1 ∗ Ψj−1 )(·, ·).
i=1
From Theorem 8.26 we are able to deduce that

(2)

NJ−1
NJ−1 NJ−1 (2)
ΦJ−1 ∗F = ai Li ΦJ−1 (·, ·), (8.63)
i=1
where
NJ−1

NJ−1
N NJ−1 (2)
al = wl,iJ−1 Li (ΦJ−1 (·, ·) ∗ F ). (8.64)
i=1
On the other hand, by virtue of Lemma 8.22, we have
(2)

NJ
(2)
ΦJ−1 ∗ F = i Li ΦJ−1 (·, ·).
aN J NJ
(8.65)
i=1
Combining (8.64) and (8.65) we obtain
N

NJ−1 NJ
N NJ−1 (2)
al J−1 k Li
wl,iJ−1 aN LN
k ΦJ−1 (·, ·)
J J
= (8.66)
i=1 k=1
N N
for l = 1, . . . , NJ−1 . Assuming the sets {L1 j , . . . , LNjj } to be hierarchical, i.e.,
Nj Nj+1
Li = Li , i = 1, . . . , Nj ; j = 0, . . . , J − 1, and observing the symmetry of the
N
matrix (wl,iJ−1 ) we gain a reduction of computational costs as follows:
N

NJ−1 NJ
N NJ−1 (2)
al J−1 k Li
wi,lJ−1 aN LN
k ΦJ−1 (·, ·)
J J
=
i=1 k=1

NJ−1 NJ−1
N NJ−1 N (2)
k Li
wi,lJ−1 aN Lk J−1 ΦJ−1 (·, ·)
J
=
i=1 k=1

NJ−1

NJ
N NJ−1 (2)
k Li
wi,lJ−1 aN LN
k ΦJ−1 (·, ·)
J J
+
i=1 k=NJ−1 +1
NJ−1

NJ
N NJ−1 (2)
= aN k Li
wi,lJ−1 aN LN
k ΦJ−1 (·, ·).
J J J
l +
i=1 k=NJ−1 +1
The recursion relation (8.66) leads us to the following decomposition scheme:

F → aN J → aNJ−1 → ··· aN 0
↓ ↓ ↓
(W T )(F )(J; ·) (W T )(F )(J − 1; ·) (W T )(F )(0; ·).
The bandpass filter Rj (F ) can be deduced from the formula

Nj
N N
Rj (F ) = Ψ̃j ∗ (W T )(F )(j; ·) = ai j Li j (Ψ̃j ∗ Ψj )(·, ·). (8.67)
i=1
This allows the following reconstruction scheme of F :
aN 0 aN 1 aN 2
↓ ↓ ↓
R0 (F ) R1 (F ) R2 (F )

P0 (F ) −→ + −→ P1 (F ) −→ + −→ P2 (F ) −→ + −→ · · · .
We have seen that the vectors aNj do not depend on the special choice of the scaling
function {Φj (·, ·)}j∈N0 . In other words, we are able to reconstruct a function with
respect to different wavelets just by the knowledge of the vectors aNj .
Let us finally make some comments concerning the pyramid schemata:
(1) In signal processing a variant of the pyramid scheme is known as subband
coding. This technique was originally studied before wavelet theory. The de-
composition step consists of applying a lowpass and a bandpass filter followed
by downsampling; the reconstruction consists of upsampling followed by fil-
tering.
(2) Any bandlimited potential can be reconstructed exactly via the pyramid
scheme by use of bandlimited wavelets (see also [67]). In this case spline
exact approximation coincides with polynomial (i.e., outer harmonic) exact
approximation. The scale and detail spaces are finite-dimensional so that the
detail information of a potential is only determined by a finite number of

wavelet coefficients for each scale.
(3) In case of evaluation functionals and (radial) derivatives at certain points on
a sphere Ωr , r ≥ σ, the numerical effort can be drastically reduced by three
integration procedures on the sphere. The first method is to use gridded
pointsystems and then to apply FFT-techniques (cf. the Ph.D.-thesis [74]).
The second technique is to use a suitable Gauss-quadrature rule in north-
south direction. The third method is to apply the idea of fast summation
and panel clustering (cf. [23, 39]). For more details concerning numerical
integration on the sphere the reader is referred to [21, 44].
(4) The pyramid scheme provides a powerful tool in interpreting and constructing
lowpass and bandpass filters. The wavelets localize in space and frequency.
This makes wavelets particularly useful for data compression. Compression
techniques aim at reducing storage requirements and speeding up read or
write operations to or from disks. For the compression scheme we are ready
to accept an error as long as the quality after compression is acceptable.
(5) Another application is, that for the evaluation of a potential or its derivatives
at a point, only wavelet coefficients close to the point have to be taken into
account. This enables us to observe local features of the geopotential in a
global model.
Example. In the foregoing we have seen that bandlimited harmonic wavelets pro-
vide “building blocks” that enable fast decorrelation of geopotential data. Next
we are interested in discussing the concept of multiresolution analysis from practi-
cal point of view. To be more specific, the multiresolution analysis “looks at” the
Earth’s gravitational potential through a microscope, whose resolution gets finer
and finer. Thus it associates to the gravitational potential a sequence of smoothed
versions, labelled by the scale parameter. This aspect is illustrated by the figures
below for the (bandlimited) EGM96 model. The computation has been performed
on the basis of the CP-wavelets following Variant 1.
-100.0 0.0 100.0 200.0 -200.0 0.0 200.0
[100 Gal m] [100 Gal m]
P3 (F ) R3 (F )
-200.0 0.0 200.0 400.0 -100.0 0.0 100.0
[100 Gal m] [100 Gal m]
P4 (F ) R4 (F )
-500.0 0.0 500.0 0.0 100.0
[100 Gal m] [100 Gal m]
P5 (F ) R5 (F )
Figure 8.1. EGM96 CP-wavelet representation at height 0 km.

-500.0 0.0 500.0 -50.0 0.0 50.0
[100 Gal m] [100 Gal m]
P6 (F ) R6 (F )
-500.0 0.0 500.0 -50.0 0.0 50.0
[100 Gal m] [100 Gal m]
P7 (F ) R7 (F )
-500.0 0.0 500.0
[100 Gal m]
P8 (F )
Figure 8.2. EGM96 CP-wavelet representation at height 0 km (cont.).

9. Illustrations of Meissl schemata

In this section we derive Meissl schemata for the SST and SGG operators (cf. [20,
32, 58]). In our contribution we focus on the gravitational potential, but obviously,
the results are also valid for the disturbing potential.
9.1. Meissl schemata based on outer harmonic framework

We start from the scalar Fourier expansion of the gravitational potential V in
terms of outer harmonics
∞ 2n+1

V (x) = V ∧ (n, m)Hn,m
s
(σ; ·). (9.1)
n=0 m=1
If the observables are given both at minimum satellites altitude γ and at

minimum Earth’s radius σ (see Figure 3.1), the symbols of the pseudodifferential
operators for the SST and SGG problem can be arranged in a Meissl scheme. The
symbols at the arrows indicate how the Fourier coefficients of degree n change
at the transition form one quantity to another. In order to avoid confusion the
corresponding basis functions are also given. In the case of radial derivatives we
remember that the basis system Hn,m fulfills
Hn,m (σ; ·)|Ωσ = (1/R)Yn,m . (9.2)
Therefore, we get the Meissl scheme for radial derivatives given in Figure 9.1.
Figure 9.1. Meissl scheme for radial derivatives.

Figure 9.2. Meissl scheme for first-order tangential derivatives and

second-order mixed derivatives.
If vectorial observables are investigated, we need that

(2) n + 1 (1) n
o Yn,m = −n ỹn,m + (n + 1) ỹ (2) , (9.3)
2n + 1 2n + 1 n,m
which yields the Meissl schemata in Figures 9.2 and 9.3.

Finally, in the case of second-order tangential derivatives (∇∗ ⊗ ∇∗ ) we cal-
culate
n + 1 (1,1) n + 2 (2,1) 2(n + 1)
∇∗ ⊗ ∇∗ ỹn,m
(1)
= ρ(1,1) ỹn,m + ρ(2,1) ỹn,m + ρ(2,2) ỹ(2,2)
n
2n + 3 n
2n + 3 n
(2n + 1)(2n − 1) n,m
(9.4)
and
2n(n + 1) n−1
∇∗ ⊗ ∇∗ ỹn,m
(2)
= τn(1,1) (−1) ỹ(1,1) + τn(1,2) ỹ(1,2)
(2n + 1)(2n + 3) n,m (2n − 1)(2n + 1) n,m
2n(n + 2) n
+ τn(2,1) ỹ(2,1) + τn(2,2) (−1) ỹ(2,2) ,
(2n + 3)(2n + 1) n,m (2n − 1)(2n + 1) n,m
(9.5)
Figure 9.3. Meissl scheme for first-order radial derivatives and

second-order mixed derivatives.
(i,k) (i,k)
where the constants ρn and τn are given by

(i,k)
νn
ρn(i,k) = , (9.6)
(2n + 1)(n + 1)

(i,k)
νn
τn(i,k) = . (9.7)
(2n + 1)n
In conclusion, we get the Meissl scheme for first- and second-order tangential
derivatives (see Figure 9.4).
9.2. Meissl schemata based on kernel function framework

In order to derive Meissl schemata based on kernel functions we want to recapit-
ulate the convolutions which are used in this section (see Table 4).
Our point of departure is the description of a function F ∈ Hs (Ωext
σ ) in terms
of outer harmonics
∞ 2n+1

F (x) = F ∧ (n, m)Hn,m
s
(σ; x), (9.8)
n=0 m=1
Figure 9.4. Meissl scheme for first- and second-order tangential derivatives.
x ∈ Ωext
σ , and we first derive the kernel functions corresponding to the SST and
SGG operators.
Scalar SST and SGG Operators
The SST and SGG operators are given by the convolution equation
ΛF (x) = (K Λ )σ,γ (·, x) ∗ F, x ∈ Ωext
γ , (9.9)
where the symbol of the kernel (K Λ )σ,γ is given by
⎧ n
⎨ σ n+1
, n = 0, 1, . . . for SST,
(K Λ )∧ (n) = Λ∧ (n) = γ n γ (9.10)
⎩ σ (n+1)(n+2)
2 , n = 0, 1, . . . for SGG.
γ γ
K ∗F

∞ 2n+1
F, K(·, y) ∈ Hs (Ωext
σ )
= K ∧ (n)F ∧ (n, m)Hn,m
s
(γ; ·)
n=0 m=1
(i)
k ∗f

∞ 2n+1 (i)
f, k (i) (·, y) ∈ hs (Ωext
σ )
= k (i)∧ (n)f (i)∧ (n, m)Hn,m
s
(γ; ·)
n=0i m=1
kF

3 (i)∧
∞ 2n+1 F ∈ Hs (Ωext
σ ),
(i)s
= k (n)F ∧ (n, m)hn,m (γ; ·) k(·, y) ∈ hs (Ωext
i=1 n=0i m=1 σ )
k(i,k) ∗ f
(i,k)∧
∞ 2n+1 f , k(i,k) (·, y)
= k (n)f (i,k)∧ (n, m)Hn,m
s
(γ; ·) ∈ h(i,k) (Ωext
s σ )
n=0̃ik m=1
kF
3 ∞ 2n+1
F ∈ Hs (Ωext
σ ),
= k(i,k)∧ (n)F ∧ (n, m)h(i,k)s
n,m (γ; ·) k(·, y) ∈ hs (Ωext
σ )
i,k=1 n=0̃ik m=1
Table 4. List of the convolutions.
Vectorial SST and SGG Operators

In the vectorial case we have
λF (x) = (k λ )σ,γ (·, x) F, x ∈ Ωext
γ , (9.11)
λ (i)∧
with the symbol (k ) (n) given by
⎧ n
⎨ − σ n n+1
γ γ 2n+1 , n = 1, 2, . . . for SST,
(k λ )(1)∧ (n) = λ(1)∧ (n) = n+1
⎩ − σ n(n+1) n+1
γ γ2 2n+1 , n = 1, 2, . . . for SGG,
(9.12)
and
⎧ n
⎪
⎨ σγ n+1 n
λ (2)∧ (2)∧ γ 2n+1 , n = 1, 2, . . . for SST,
(k ) (n) = λ (n) = n+1
⎪
⎩ σγ (n+1)2 n
γ2 2n+1 , n = 1, 2, . . . for SGG,
(9.13)
and (k λ )(3)∧ (n) = 0.
Tensorial SGG Operator
This operator is given by
λF (x) = (k λ )σ,γ (·, x) F, x ∈ Ωext
γ , (9.14)
where we have the symbol

⎧
⎪
⎪
n
1 n(n+1) (1,1)
⎪
⎪
σ
γ 2 (2n+1)(2n+3) νn , (i, k) = (1, 1),
⎪
⎪
γ
n
⎪
⎪
⎪
⎪ σ 1 −(n+1)(n−1) (1,2)
νn , (i, k) = (1, 2),
⎨ γ γ 2 ((2n−1)(2n+1)

λ (i,k)∧ (i,k)∧ n
(k ) (n) = λ (n) = 1 −n(n+2) (2,1)
⎪
⎪
σ
2
γ (2n+3)(2n+1) νn , (i, k) = (2, 1),
⎪
⎪
γ
n
⎪
⎪
⎪
⎪ σ 1 n(n+1)(n+2) (2,2)
νn , (i, k) = (2, 2),
⎪
⎪ γ γ 2 (2n−1)(2n+1)
⎩
0, else.
(9.15)
Upward Continuation Operators
The kernels of the (scalar) upward continuation operators KU , KU , and KU are
given by
σ n
∞ 2n+1
s s
KU (x, y) = Hn,m (γ; x)Hn,m (σ; y), (9.16)
n=0 m=1
γ
σ n+1
∞ 2n+1
s s
KU (x, y) =
Hn,m (γ; x)Hn,m (σ; y), (9.17)
n=0 m=1
γ
σ n+2
∞ 2n+1
s s
KU (x, y) = Hn,m (γ; x)Hn,m (σ; y). (9.18)
n=0 m=1
γ
The upward continuation operators for vector and tensor fields can be introduced
s(i)
in the same way by use of the vectorial and tensorial basis functions hn,m and
s(i,k)
hn,m , i, k ∈ {1, 2, 3}.
The Meissl schemata for the scalar/vectorial/tensorial wavelets can now be
derived as follows:
Scalar Meissl Scheme. From the reconstruction formula in the scalar case (7.9) we
get
∞
∞

F (x) = Ψ̃j ∗ (W T )(F )(j; x) = (Ψ̃j ∗ Ψj ∗ F )(x), (9.19)
j=−1 j=−1
x ∈ Ωext
σ , whereas
∞

∂F
(x) = Ψ̃j ∗ Ψj ∗ K σ∂ ∗ F (x), (9.20)
∂r j=−1
∂r
where the kernel of the first radial derivative K σ∂ on the sphere Ωσ is given by
∂r
n + 1
∞ 2n+1
σ
K ∂ (x, y) = − s
Hn,m s
(σ; x)Hn,m (σ; y). (9.21)
∂r
n=0 m=1
σ
2
∂
The same calculation for the second radial derivative ∂r 2 leads to
∞

∂2F
(x) = Ψ̃j ∗ Ψj ∗ K R∂ 2 ∗ F (x), (9.22)
∂r2 j=−1 ∂r2
where K σ∂ 2 is given by
∂r2
∞ 2n+1
(n + 1)(n + 2)
σ s s
K ∂2 (x, y) = Hn,m (σ; x)Hn,m (σ; y)
∂r2
n=0 m=1
σ2

= K σ∂ ∗ K̃ σ∂ (x, y), (9.23)
∂r ∂r
and the kernel K̃ σ∂ is given by

∂r
n + 2
∞ 2n+1
K̃ σ∂ (x, y) = − s
Hn,m s
(σ; x)Hn,m (σ; y). (9.24)
∂r
n=0 m=1
σ
Therefore, we get the Meissl scheme shown in Figure 9.5.
Figure 9.5. Meissl scheme for kernel functions (scalar case).
Scalar/Vectorial Meissl Scheme. The extension the the case of vectorial operators
is straightforward:
2
∞

(2),σ (i) (i) σ,(i)
o F (x) = Ψ̃j Ψj ∗ ko(2),σ F (x), (9.25)
j=−1 i=1
σ,(i)
where the kernel functions ko(2),σ are given by

σ,(1)
n
∞ 2n+1
n + 1 s(1)
ko(2),σ (x, y) = − s
hn,m (σ; x)Hn,m (σ; y), (9.26)
n=0 m=1
σ 2n + 1
σ,(2)
n + 1 n
∞ 2n+1
ko(2),σ (x, y) = hs(2) s
n,m (σ; x)Hn,m (σ; y). (9.27)
n=1 m=1
σ 2n + 1
In the SGG case we calculate

∞
2
∂F (i) (i)
o(2),σ
σ,(i)
(x) = Ψ̃j Ψj ∗ ko(2) ∂ F (x), (9.28)
∂r j=−1 i=1
∂r
σ,(i)
where the kernels ko(2),σ ∂ are given by
∂r

σ,(1)
n + 1 n n + 1
∞ 2n+1
ko(2),σ ∂ (x, y) = − hs(1) s
n,m (σ; x)Hn,m (σ; y)
∂r
n=0 m=1
σ σ 2n + 1

σ,(1)
= ko(2),σ K σ∂ (x, y), (9.29)
∂r
σ,(2)
n + 1 n + 1 n
∞ 2n+1
ko(2),σ ∂ (x, y) = hs(2),σ (σ; x)Hn,m
s
(σ; y)
∂r
n=1 m=1
σ σ 2n + 1 n,m

σ,(2)
= ko(2) K σ∂ (x, y). (9.30)
∂r
Summing up, we finally get the Meissl schemata given in Figures 9.6 and 9.7 for
the vector approach.
Scalar/Vectorial/Tensorial Meissl Scheme. We get
∇∗,σ ⊗ ∇∗,σ F (x)
∞

(i,k) (i,k) σ,(i,k)
= Ψ̃j Ψj ∗ k∇∗,σ ⊗∇∗,σ F (x), (9.31)
j=−1 (i,k)∈
{(1,1),(1,2),(2,1),(2,2)}
σ,(i,k)
where the kernel functions k∇∗,σ ⊗∇∗,σ are given by
σ,(1,1)
(1,1)
∞ 2n+1
n(n + 1) s(1,1) s
k∇s,σ ⊗∇∗,σ (x,y) = νn hn,m (σ;x)Hn,m (σ;y),
n=0 m=1
σ 2 (2n + 1)(2n + 3)
(9.32)
σ,(1,2)
(1,2)
∞ 2n+1
(n − 1)(n + 1)
k∇∗,σ ⊗∇∗,R (x,y) = − νn hs(1,2) (σ;x)Hn,m
s
(σ;y),
n=1 m=1
σ 2 (2n − 1)(2n + 1) n,m
(9.33)
Figure 9.6. Meissl scheme for kernel functions (scalar/vectorial case).
σ,(2,1)
(2,1)
∞ 2n+1
n(n + 2)
k∇∗,σ ⊗∇∗,σ (x,y) = − νn hs(2,1) (σ;x)Hn,m
2 (2n + 1)(2n + 3) n,m
s
(σ;y),
n=0 m=1
σ
(9.34)
∞ 2n+1

σ,(2,2) (2,2) n(n + 1)(n + 2)
k∇∗,σ ⊗∇∗,σ (x,y) = νn hs(2,2) (σ;x)Hn,m
2 (2n − 1)(2n + 1) n,m
s
(σ;y).
n=2 m=1
σ
(9.35)
σ,(i,k)
Note that the kernels k∇∗,σ ⊗∇∗,σ , (i, k) ∈ {(1, 1), (1, 2), (2, 1), (2, 2)} can be split
σ,(i,k) 2 σ,(i,k),(l) σ,(l) σ,(i,k)(l)
into k∇∗,σ ⊗∇∗,σ = l=1 k∇∗,σ ko(2),σ , where the kernels k∇∗,σ are given by
∞ 2n+1
n+1
ρn(1,1) hn,m
σ,(1,1),(1) s(1,1)
k∇∗,σ = (σ; x)hs(1)
n,m (σ; y), (9.36)
n=0 m=1
σ(2n + 3)

∞ 2n+1
2n(n + 1)

τn(1,1) hs(1,1)
σ,(1,1),(2)
k∇∗,σ = − s(2)
n,m (σ; x)hn,m (σ; y), (9.37)
n=1 m=1
σ(2n + 1)(2n + 3)
Figure 9.7. Meissl scheme for kernel functions (scalar/vectorial case).

σ,(1,2),(1)
k∇∗,σ = 0, (9.38)
∞ 2n+1
n−1
τ (1,2) hs(1,2)
σ,(1,2),(2) s(2)
k∇∗,σ = n,m (σ; x)hn,m (σ; y), (9.39)
n=1 m=1
σ(2n − 1)(2n + 1) n
∞ 2n+1
n+2
ρ(2,1) hn,m
σ,(2,1),(1) s(2,1)
k∇∗,σ = (σ; x)hs(1)
n,m (σ; y), (9.40)
n=0 m=1
σ(2n + 3) n
∞ 2n+1
2n(n + 2)
τ (2,1) hs(2,1)
σ,(2,1),(2) s(2)
k∇∗,σ = n,m (σ; x)hn,m (σ; y), (9.41)
n=1 m=1
σ(2n + 3)(2n + 1) n
∞ 2n+1
2(n + 2)
ρn(2,2) hn,m
σ,(2,2),(1) s(2,2)
k∇∗,σ = (σ; x)hs(1)
n,m (σ; y), (9.42)
n=0 m=1
σ(2n − 1)(2n + 1)

∞ 2n+1
n

τn(2,2) hs(2,2)
σ,(2,2),(2)
k∇∗,σ = − s(2)
n,m (σ; x)hn,m (σ; y). (9.43)
n=2 m=1
σ(2n − 1)(2n + 1)
∞ 2n+1 s(i,k) s(l)

The convolution of the kernel kσ,(i,k),(l) = n=0 m=1 kσ,(i,k),(l)∧ (n)hn,m hn,m
and the vector field f (l) ∈ h(Ωext
σ ) is given by
∞ 2n+1

kσ,(i,k),(l) f (l) = kσ , (i, k), (l) ∧ (n)f (l)∧h (n, m)hn,m
s(i,k)
(σ; ·). (9.44)
n=0̃ik m=1
Thus, we get the Meissl scheme given in Figure 9.8.
Figure 9.8. Meissl scheme for kernel functions (scalar/vectorial/ten-

sorial case). (Note that the tensor-2 wavelets could not be written in
bold letter for technical reasons.)
10. Conclusions
As already pointed out, accurate knowledge of the gravitational potential of the
Earth is required in order to solve, for example, problems in geodesy, navigation,
oceanography, solid Earth physics, and exploration geophysics. In physical geodesy
it is the essential pre-stage of geoid computation. Earlier it was envisaged that the
gravitational potential could be determinable as a solution of a boundary value
problem. The classical problem was the Stokes problem, the boundary values were
the gravity anomalies, for which the hitherto unrealistic assumption of global (ter-
restrial) coverage was required. But today we are confronted with the situation
where also other quantities give information about the Earth’s gravity potential,
for example, gravity disturbance vector or second-order gradients of the distur-
bance potential from air- and spacecraft. In recent years the geometric shape of
the Earth, continents and ocean surface, became measurable with unprecedented
precision, due to the enormous progress of space methods like GNSS, VLBI, SLR,
and satellite altimetry. The mathematical connection between the gravitational
data within a georelevant geometry is the integrated concept. Usually, this con-
cept is formulated in the framework of a reproducing kernel Hilbert space H(Ωext σ )
consisting of potentials harmonic down to an internal (Runge) sphere Ωσ . Math-
ematically, the gravitational (anomalous) potential of the Earth is assumed to be
an element of such a space H(Ωext ext
σ ). In the Hilbert space H(Ωσ ) any element may
be represented by its expansion with respect to a complete system of kernel ex-
pressions Li KH(Ωext ) (·, ·) related to (linear) observables Li on H(Ωext
σ ). Because of
σ
the reproducing kernel structure imposed on H(Ωext σ ), orthonormalization of a fi-

nite system {Li KH(Ωext ) (·, ·)}i=1,...,N is equivalent to the spline problem of finding
σ
the minimum norm interpolant in the associated · H(Ωext ) -metric. When using
σ
minimum norm interpolation (or smoothing), however, the normal equation ma-
trix (Li Lk KH(Ωext ) (·, ·))i,k=1,...,N is in general a full matrix, reflecting the certain
σ
status of decorrelation guaranteed by the reproducing kernel (covariance function)
under consideration. This problem causes numerical difficulties which may to a
certain extend be overcome by several techniques (for example, fast summation,
panel clustering, etc.). But the numerical obstacles are the main reasons why ap-
proximation methods of the Earth’s gravitational field determination based on
spline procedures could not keep pace with the increasing flow of observational
information. In other words, the serious drawback of spline approximation is that
there is no efficient transition from global to local modeling by only using one
kernel (covariance) function with (fixed) space/momentum localization property.
The power of harmonic wavelets lies in the fact that kernel functions with
variable space/momentum localization come into use according to a suitable dila-
tion process. By using a sequence of more and more kernels reflecting the various
levels of space/momentum localization the reference Sobolev space H(Ωext σ ) is de-
composed into a nested sequence of approximating subspaces
· · · Vj (Ωext ext
σ ) ⊂ Vj+1 (Ωσ ) ⊂ · · ·
reflecting the different stages of decorrelation. In doing so, harmonic wavelets may
be used as mathematical means for breaking up a complicated function (such as
the Earth’s gravitational potential) into many simple pieces at different scales and
positions. This allows multiresolution analysis and compression of data. The par-
ticular efficiency of wavelets is caused by the property that only a few wavelet
coefficients in the wavelet table are needed in areas where the gravitational poten-
tial is “smooth”, whereas stronger resolution of a complicated pattern is settled by
a zooming-in capability. Wavelets offer canonical tools for combined terrestrial, air-
borne, and spaceborne data management under realistic assumptions imposed on
the geometry of the Earth’s surface and the “orbital configuration”. Fast computa-
tion becomes available in form of tree algorithms. This enables gravitational poten-
tial determination with millions of data. Thus harmonic (regularization) wavelets
are particularly important for inverse multiscale modeling of spaceborne data. In
a subsequent step geoid computation can be based on a highly accurate gravi-
tational potential derived from a homogeneous set of spaceborne data combined
with terrestrial and/or airborne data.
For inverse multiscale modeling of spaceborne data two different ways of
wavelet regularization are available, namely bandlimited truncated singular value
decomposition and non-bandlimited regularization using, e.g., Tikhonov, rational,
exponential, and “locally supported” kernels. In accordance with the uncertainty
principle the different constituting elements of regularization may be explained
as follows: Non-bandlimited regularization wavelets tend to be extremely space
localizing. Thus huge data sets of irregular distribution can be handled since only
data in a small neighborhood, whose size is determined by the particular choice of
the wavelet type, is needed for the purpose of evaluating the wavelet coefficients.
On the other hand, a large number of wavelet coefficients depending on the choice
of the wavelet for the regularization is needed, since the wavelet coefficients only
give local information of a small neighborhood. It appears that non-bandlimited
regularization is an appropriate tool of local gravity surveys for oil and mineral
exploration. However, little practical work has been done yet in this application
area for non-synthetic data sets, although the use of linear functionals allows a very
promising combination of terrestrial and/or airborne data within a unified setup
in terms of wavelets. Moreover, fast summation techniques and panel clustering is
adequately applicable in pyramid schemata.
Bandlimited regularization wavelets show more moderate phenomena of space
localization so that one can work with smaller data sets in numerical evaluation.
In consequence, the number of wavelet coefficients can be reduced, since they
contain information of a more extended area. Moreover, a certain spectral band can
be expressed exactly in terms of wavelets because of their bandlimited character
even when the airborne data are combined with terrestrial information. Pyramid
schemata can be based on exact (outer harmonic) approximation. In conclusion,
dependent on the space/momentum character of the bandlimited wavelets inverse
multiscale gravity modeling of spaceborne data can be handled successfully by
multiresolution analysis.
Finally, it should be pointed out that our approach is given within a spher-
ical context. Geodesists sometimes believe that ellipsoidal reference surfaces in
combination with ellipsoidal harmonics might be the better choice. No doubt,
an ellipsoidally reflected multiscale formulation is mathematically interesting and
geodetically relevant. However, its numerical realization is by far more compli-
cated than the spherical oriented variant chosen for our study here. As a matter of
fact, Meissl schemata are involved with gravitational quantities not including the
centrifugal influence. In this case, however, Runge–Walsh methods corresponding
to Runge–Walsh (Bjerhammar) spheres form an adequate alternative which, in
the opinion of the authors, is superior when numerical purposes come into play
because of the much more efficient and economical structure inherent in spherical
framework. Even better, Runge–Walsh procedures are not only applicable for ellip-
soidal reference surfaces, but also for geometrically complicated reference surfaces
such as telluroid, or (co)geoid.
Acknowledgment. The authors thank the “Federal Ministry for Economic Affairs
and Energy, Berlin” and the “Project Management Jülich” for funding the project
“SPE” (funding reference number 0324016).
11. Appendix A: List of basic gravity field quantities

The list of this appendix essentially follows [ESA1]. It provides an introductory col-
lection of quantities used in classical geodesy that could not be explained through-
out the paper:
Definition Observation method
Gravity potential, W: Differences between values in two

Sum of the gravitational and the points observed by levelling.
centrifugal potential.
Equipotential surface: Points on one surface determined

Surface where the gravity potential is regionally by tide-gauges, which define
equal to a constant. the regional mean sea level.
Height datum: Mean sea-level calculated from

The equipotential surface best tide-gauges for a specific time period.
agreeing with local mean sea level.
Geoid: The equipotential surface which agrees

The equipotential surface which agrees with a global set of tide-gauges and
with global mean sea level. leveling bench-marks.
Gravity: Observed by absolute (e.g., free fall

Magnitude of gradient of the gravity experiment) or relative (as a
potential at Earth’s surface and of the difference) spring gravimeter.
gravitational potential in the outer
space.
Gravity gradient: Certain linear combinations
Derivatives of the gravity vector, i.e., measured by torsion-balance at
second-order derivatives of W . Earth’s surface, by difference between
accelerometers in space (gradiometry).
Mean Earth Ellipsoid: Surface which gives best fit to mean
Ellipse rotated around the ε3 -axis, sea-level, and which has centrum in
with center at the Earth’s gravity the gravity centre.
center.
Height above ellipsoid: Observed indirectly by GPS from
Height above mean Earth ellipsoid cartesian coordinates.
measured along the normal to
ellipsoid.
Geoid height: Observed by GPS at tide-gauge or at
Height of a point on the geoid above leveling point.
the reference ellipsoid.
Orthometric height: Observed by leveling and converted to
Height from geoid measured along a metric units by dividing with gravity.
plumb-line (often height above mean
sea-level).
GNSS: GNSS: GPS, GLONASS, Galileo or
A satellite navigation system with Beidou.
global coverage.
Gravity anomaly: It is a value derived by subtracting
A model gravity potential with a measured and normal gravity. The
reference ellipsoid as an equipotential normal gravity is calculated in a point
surface is used to calculate normal with the ellipsoidal height put equal to
gravity (needed is latitude and the orthometric height.
orthometric height).
12. Appendix B: List of basic units in gravitational field theory

Units and orders in gravity field theory are the following: The gravity is expressed
in m/s2 or in milligal (1 mgal= 10−5 m/s2 ); the mean Earth gravity is about
981 000 mgal, and varies from 978 100 mgal to 983 200 mgal from equator to
pole due to the Earth’s flattening and rotation. Deviations due to density inho-
mogenities, mountain ridges, etc. range from tens to hundreds of milligals. On the
other hand, the excursions of the geoid, measured from the mean Earth ellipsoid,
amount to about −105 and +90 meters. Gravity gradients are expressed in Eőtvős
(1E = 10−9 s2 ). The largest component is the vertical gravity gradient, being
on Earth’s surface of about 3000E (gravity changes by 3 ·10−6 m/s2 per meter
of elevation). The horizontal components are approximately half this size, mixed
gradients are below 100E for the normal field. Gravity gradient anomalies can be
much larger and reach about 1000E in mountainous areas (for more details see,
for example, [R4]).
SI units traditional
gravity
10−2 ms−2 1 Gal

10−5 ms−2 1 mGal
−8 −2
10 ms 1μ Gal
gravity potential
10m2 s−2 1 kGal ·m

gravity gradients
10−9 s−2 1E
References
[1] Arabelos, D., Tscherning, C.C. (1995) Regional Recovery of the Gravity Field from
SGG and Gravity Vector Data Using Collocation. J. Geophys. Res., 100, B 11, 22009–
22015.
[2] Arabelos, D., Tscherning, C.C. (1998) Calibration of Satellite Gradiometer Data
Aided by Ground Gravity Data. J. of Geodesy, 72, 617–625.
[3] Aronszajn, N. (1950) Theory of Reproducing Kernels. Trans. Am. Math. Soc., 68,
337–404.
[4] Augustin, M., Freeden, W., Nutz, H. (2018) About the Importance of the Runge–
Walsh Concept for Physical Geodesy. In: Freeden, W., Nashed , M.Z. (Eds.), Hand-
book of Mathematical Geodesy, this volume, Springer International Publishing, 517–
560.
[5] Cui, J., Freeden, W., Witte, B. (1992) Gleichmäßige Approximation mittels sphäri-
scher Finite-Elemente und ihre Anwendung auf die Geodäsie. Z. f. Vermessungswes.,
ZfV, 117, 266–278.
[6] Daubechies, I. (1992) Ten Lectures on Wavelets. SIAM.
[7] Davis, P.J. (1963) Interpolation and Approximation. Blaisdell Publishing Company.
[8] Driscoll, J.R., Healy, R.M. (1994) Computing Fourier Transforms and Convolutions
on the 2-Sphere. Adv. Appl. Math., 15, 202–250.
[9] ESA (1996) The Nine Candidate Earth Explorer Missions. Publications Division
ESTEC, Nordwijk, SP-1196 (1).
[10] ESA (1998) European Views on Dedicated Gravity Field Missions: GRACE and
GOCE. ESD–MAG–REP–CON–001.
[11] ESA (1999) Gravity Field and Steady-State Ocean Circulation Mission. ESTEC,
Nordwijk, ESA SP-1233 (1).
[12] Freeden, W. (1979) Über eine Klasse von Integralformeln der Mathematischen
Geodäsie. Veröff. Geod. Inst. RWTH Aachen, Heft 27.
[13] Freeden, W. (1980) On the Approximation of External Gravitational Potential with
Closed Systems of (Trial) Functions, Bull. Geod., 54, 1–20.
[14] Freeden, W. (1981) On Approximation by Harmonic Splines. Manuscr. Geod., 6,
193–244.
[15] Freeden, W. (1982) Interpolation and Best Approximation by Harmonic Spline Func-
tions. Boll. Geod. Sci. Aff., 1, 105–120.
[16] Freeden, W. (1982) On Spline Methods in Geodetic Approximation Problems. Math.
Meth. in the Appl. Sci., 4, 382–396.
[17] Freeden, W. (1983) Least Squares Approximation by Linear Combinations of
(Multi-) Poles, Dept. Geod. Science, 344, The Ohio State University, Columbus.
[18] Freeden, W. (1987) Harmonic Splines for Solving Boundary Value Problems of Po-
tential Theory. In: Algorithms for Approximation (J.C. Mason, M.G. Cox, Eds.), The
Institute of Mathematics and its Applications, Conference Series, vol. 10, Clarendon
Press, Oxford, 507–529.
[19] Freeden, W. (1999) Multiscale Modelling of Spaceborne Geodata, B.G. Teubner,
Stuttgart, Leipzig.
[20] Freeden, W., Gerhards, C. (2013) Geomathematically Oriented Potential Theory.
CRC Press, Taylor & Francis, Boca Raton.
[21] Freeden, W., Gervens, T., Schreiner, M. (1998) Constructive Approximation on
the Sphere (With Applications to Geomathematics). Oxford Science Publications,
Clarendon.
[22] Freeden, W., Glockner, O., Thalhammer, M. (1999) Multiscale Gravitational Field
Recovery from GPS-Satellite-to-Satellite Tracking. Studia Geoph. Geod, 43, 229–264.
[23] Freeden, W., Glockner, O., Schreiner, M. (1998) Spherical Panel Clustering and Its
Numerical Aspects. J. of Geodesy, 72, 586–599.
[24] Freeden, W., Gutting, M. (2013) Special Functions of Mathematical (Geo-)Physics.
Birkhäuser, Basel.
[25] Freeden, W., Kersten, H. (1981) The Geodetic Boundary Value Problem Using the
Known Surface of the Earth. Veröff. Geod. Inst. RWTH Aachen, Heft 29.
[26] Freeden, W., Mason, J.C. (1990) Uniform Piecewise Approximation on the Sphere.
In: Algorithms for Approximation II (J.C. Mason, M.G. Cox, Eds.), Chapman and
Hall Mathematics, 320–333.
[27] Freeden, W., Nutz, H.. (2011) Satellite Gravity Gradiometry as Tensorial Inverse
Problem. Int. J. Geomath, 2, 177–218.
[28] Freeden, W., Reuter, R. (1982) Remainder Terms in Numerical Integration Formulas
of the Sphere. Internat. Series Numeric. Math., 61, 151–170.
[29] Freeden, W., Schneider, F. (1998a) Wavelet Approximation on Closed Surfaces and
Their Application to Boundary Value Problems of Potential Theory, Math. Meth.
Appl. Sci., 21: 129–165.
[30] Freeden, W., Schneider, F. (1998b) An Integrated Wavelet Concept of Physical
Geodesy. J. of Geodesy, 72, 259–281.
[31] Freeden, W., Schreiner, M. (1997) Orthogonal and Non-orthogonal Multiresolution
Analysis, Scale Discrete and Exact Fully Discrete Wavelet Transform on the Sphere.
Constr. Approx. 14, 493–515.
[32] Freeden, W., Schreiner, M. (2009) Spherical Functions of Mathematical Geosciences
– A Scalar, Vectorial, and Tensorial Setup. Springer, Heidelberg.
[33] Freeden, W., Windheuser, U. (1996) Spherical Wavelet Transform and Its Discretiza-
tion, Adv. Comput. Math., 5, 51–94.
[34] Freeden, W., Windheuser, U. (1997) Combined Spherical Harmonic and Wavelet Ex-
pansion – A Future Concept in Earth’s Gravitational Determination, Appl. Comput.
Harm. Anal., 4, 1–37
[35] Gerhards, C. (2011) Spherical Multiscale Methods in Terms of Locally Supported
Wavelets: Theory and Application to Geomagnetic Modeling. Ph.D.-thesis, Univer-
sity of Kaiserslautern.
[36] Grafarend, E.W., Klapp, M., Martinec, Z. (2015) Spacetime Modelling of the Earth’s
Gravity Field by Ellipsoidal Harmonics. In: Handbook of Geomathematics, 2nd. edi-
tion (W. Freeden, M.Z. Nashed, T. Sonar, Eds.), Springer, 381–496.
[37] Groten, E. (1979) Geodesy and the Earth’s Gravity Field I, II. Dümmler.
[38] Gurtin, ME (1971) Theory of Elasticity, Handbuch der Physik, 6.
[39] Gutting, M. (2008) Fast Multipole Methods for Oblique Derivative Problems. Ph.D.-
thesis, University of Kaiserslautern, Geomathematics Group, Shaker, Aachen.
[40] Gutting, M. (2012) Fast Multipole Accelerated Solution of the Oblique Derivative
Boundary Value Problem. Int. J. Geomath. 3(2), 233–252.
[41] Gutting, M. (2015) Fast Spherical/Harmonic Spline Modeling. In: Handbook of Geo-
mathematics, 2nd. edition (W. Freeden, M.Z. Nashed, T. Sonar, Eds.), Springer,
2711–2746.
[42] Gutting, M. (2018) Parameter Choices for Fast Harmonic Spline Approximation.
In: Freeden, W., Nashed , M.Z. (Eds.), Handbook of Mathematical Geodesy, this
volume, Springer International Publishing, 605–639
[43] Heiskanen, W.A., Moritz, H. (1967) Physical Geodesy. W.H. Freeman and Company.
[44] Hesse, K., Sloan, I.H., Womersly, R.S. (2015) Numerical Integration on the Sphere.
In: Handbook of Geomathematics, 2nd. edition (W. Freeden, M.Z. Nashed, T. Sonar,
Eds.), Springer, 2711–2746
[45] Ilk, K.H., Reigber, C., Rummel, R. (1978) The Use of Satellite-to-Satellite Track-
ing for Gravity Parameter Recovery. Proc. of the European Workshop on Space
Oceanography, Navigation and Geodynamics (SONG), ESA SP–137.
[46] Ilk, K.H., Sigl, R., Thalhammer, M. (1990) Regional Gravity Field Recovery from
SST (GPS-Aristoteles), from Gradiometer Measurements and Their Combination.
In CIGAR II, WP’s 4.30, 4.40, 4.50.
[47] Kellogg, O.D. (1929) Foundations of Potential Theory. Frederick Ungar Publishing
Company.
[48] Klug, M. (2014) Integral Formulas and Discrepancy Estimates Using the Fundamen-
tal Solution to the Beltrami Operator on Regular Surfaces. Ph.D.-thesis, University
of Kaiserslautern, Geomathematics Group.
[49] Krarup, T. (1969) A Contribution to the Mathematical Foundation of Physical
Geodesy. Publ. of the Danish Geodetic Institute, No. 44, Copenhagen.
[50] Kusche, J. (1998): Regional adaptive Schwerefeldmodellierung für SST-Analysen. In:
Progress in Geodetic Science at GW98, (W. Freeden ed.), Shaker, 266–273.
[51] Laur, H., Liebig, V. (2015): Earth Observation Satellite Missions and Data Access.
In: Handbook of Geomathematics, 2nd. edition (W. Freeden, M.Z. Nashed, T. Sonar,
Eds.), Springer, 147–170.
[52] Magnus, W., Oberhettinger, F., Soni, R.P. (1966) Formulas and Theorems for the
Special Functions of Mathematical Physics. Die Grundlehren der mathematischen
Wissenschaften in Einzeldarstellungen, Bd. 52, Springer.
[53] Meissl, P.A. (1971) A Study of Covariance Functions Related to the Earth’s Dis-
turbing Potential. The Ohio State University, Department of Geodetic Science and
Surveying, Columbus, OSU Report No. 152.
[54] Michel, V. (1999) A Multiscale Method for the Gravimetry Problem: Theoretical and
Numerical Aspects of Harmonic and Anharmonic Modelling. Ph.D.-thesis, University
of Kaiserslautern, Geomathematics Group.
[55] Moritz, H. (1980) Advanced Physical Geodesy. Herbert Wichmann Verlag, Karl-
sruhe, Abacus Press, Tunbridge, Wells, Kent.
[56] Moritz, H., Sünkel, H. (1978) Approximation Methods in Geodesy. H. Wichman
Verlag, Karlsruhe.
[57] Müeller, C. (1966) Spherical Harmonics. Lecture Notes in Mathematics, 17, Springer-
Verlag
[58] Nutz, H. (2002) A Unified Setup of Gravitational Field Observables. Ph.D.-thesis,
University of Kaiserslautern, Geomathematics Group, Shaker, Aachen.
[59] Pizetti, P. (1910) Sopra il calcoba tesrico delle deviazioni del geoide dall’ellisoide.
Att. A. Acad. Sci. Torino, 46, 331–350.
[60] Rummel, R. (1975) Downward Continuation of Gravity Information from Satellite to
Satellite Tracking or Satellite Gradiometry in Local Areas. Dept. Geodetic Science,
221, The Ohio State University, Columbus.
[61] Rummel, R. (1979) Determination of the Short-wavelength Components of the Grav-
ity Field from Satellite-to-Satellite Tracking or Satellite Gradiometry. Manuscr.
Geod., 4, 107–148.
[62] Rummel, R. (1986) Satellite Gradiometry. In: Lecture Notes in Earth Sciences 7,
Mathematical and Numerical Techniques in Physical Geodesy (H. Sünkel, ed.), 318–
363, Springer.
[63] Rummel, R. (1997) Spherical spectral properties of the earth’s gravitational potential
and its first and second derivatives. In Sansò, S. and Rummel, R., Eds., Geodetic
Boundary Value Problems in View of the One Centimeter Geoid. Volume 65 of
Lecture Notes in Earth Science, pages 359–404. Springer, Berlin, Heidelberg.
[64] Rummel, R., van Gelderen, M. (1992) Spectral Analysis of the Full Gravity Tensor,
Geophys. J. Int., 111, 159–169
[65] Rummel, R, van Gelderen, M (1995) Meissl Scheme – Spectral Characteristics of
Physical Geodesy. Manuscr. Geod., 20, 379–385.
[66] Rummel, R., van Gelderen, M., Koop, R., Schrama, E., Sanso, F., Brovelli, M.,
Miggliaccio, F., Sacerdote, F. (1993) Spherical Harmonic Analysis of Satellite Gra-
diometry. Netherlands Geodetic Commission, New Series, No. 39.
[67] Schreiner, M. (1997) Locally Supported Kernels for Spherical Spline Interpolation. J.
of Approx. Theory, 89, 172–194.
[68] Stokes, G.G. (1849) On the Variation of Gravity on the Surface of the Earth. Trans.
Cambridge Phil. Soc., 8, 672–695.
[69] Svensson, S.L. (1983) Pseudodifferential Operators – a New Approach to the Bound-
ary Value Problems of Physical Geodesy. Manuscr. Geod., 8, 1–40.
[70] Taylor, A.E., Lay, D.C. (1980) Introduction to Functional Analysis, 2nd edition. New
York, Chichester, Brisbane, Toronto.
[71] Thalhammer, M. (1995) Regionale Gravitationsfeldbestimmung mit zukünftigen
Satellitenmissionen (SST und Gradiometrie). Deutsche Geodätische Kommission bei
der Bayerischen Akademie der Wissenschaften, Reihe C, Dissertation. Heft Nr. 437.
[72] Tscherning, C.C., Rapp, R.H. (1974) Closed Covariance Expressions for Gravity
Anomalies, Geoid Undulations, and Deflections of the Vertical Implies by Anomaly
Degree-Variance Models. Dept. Geod. Sci., 208, The Ohio State University, Colum-
bus.
[73] Weyl, H. (1916) Über die Gleichverteilung von Zahlen mod. Eins. Math. Ann., 77,
313–352.
[74] Windheuser, U. (1995) Sphärische Wavelets: Theorie und Anwendung in der Phy-
sikalischen Geodäsie. Ph.D.-thesis, University of Kaiserslautern, Geomathematics
Group, Shaker, Aachen.
[75] Xu, P.L., Rummel, R. (1994) A Simulation Study of Smoothness Methods in Recovery
of Regional Gravity Fields. Geophys. J. Int., 117, 472–486.
[76] Yamabe, H. (1959) On an Extension of Helly’s Theorem. Osaka Math. J., 2, 15–22.
[77] Yosida, K (1978) Functional Analysis, Springer-Verlag, Berlin, Heidelberg, New
York.


The Analysis of the Geodetic Boundary

Value Problem: State and Perspectives
Fernando Sansò
Abstract. The geodetic boundary value problem is mathematically a free-

boundary, oblique derivative boundary value problem for the Laplace oper-
ator. The solution of the problem is the determination of the shape of the
Earth and of its gravity field. The analysis of such a problem, specially for
its non-linear formulation, is hard, so it started only in 1976 with a paper by
L. Hörmander [13].
Since then the research has continued for both the non-linear and the
linearized version, till recent years. In this article the author tries to give
an overview on the subject, including a new result for the so-called Simple
Molodensky Problem.
Keywords. Geodetic boundary value problem,linearization, scalar and vector

variants, Molodensky problem.
Overview
In §1 the formulation of the vector GBVP is recalled, especially because it is
historically its first formulation. However the section introduces the reader to the
characteristic quasi non-uniqueness of the solution of the GBVP, due to a quasi-
invariance of the data under translation of the solution.
In §2 the scalar formulation of the GBVP is introduced and its version under a
partial Legendre transform is presented, what is called the GBVP in Marussi space.
The new formulation is a fixed boundary oblique-derivative BVP for a certain non-
linear partial differential equation only recently published. A theorem of existence
and uniqueness in Hölder spaces, derived from intermediate Schauder estimates,
is recalled. The result is nice in that on data we put regularity requirements which
are very close to a realistic model of the Earth.
460 F. Sansò
In §3 a careful linearization of the problem is performed, showing that the

resulting boundary operator can be written in the form of a simple spherical oper-
ator plus a perturbation. The problem defined using only the spherical part of the
boundary operator is known as the Simple Molodensky Problem (SMP) [15, 20].
In §4 the SMP is analyzed in suitable Sobolev spaces, obtaining a new un-
conditional theorem of existence and uniqueness, even for Lipschitz domains.
In §5 the above theorem is carried over to the linearized GBVP, by a sim-
ple perturbation technique. This is achieved at the cost of putting geometrical
constraints on the boundary.
In §6 final considerations and some items that would be interesting to in-
vestigate, are reported. Finally the author would like to warn the reader that
throughout the text many times the same letter C is used to define constants that
can actually have different values. This notwithstanding, the context shall be clear.
1. Introduction to the Vector Geodetic Boundary Value Problem

The Geodetic Boundary Value Problem (GBVP) is basically to determine the
figure of the Earth from as many as possible (and realistic) measurements on the
gravity field and the least possible knowledge of the geometry.
After one century of evolution, across XIX and XX centuries, from the early
G. Stokes research to the publication of the book by S. Molodensky, Eremiev and
Yurkina [19], the problem has got a more rigorous formulation in the sixties of the
last century, thanks to the work of A. Marussi and T. Krarup, the first interested in
finding a suitable system of coordinates, (called by him Intrinsic Coordinates [16])
that would simplify from the roots the free-boundary characteristic of the GBVP,
the second interested in a rigorous formulation of the GBVP in its non-linear form,
in order to perform at least a correct linearization process [15, 20].
In this way we got the first definition of the vector, non-linear GBVP, based
on the following hypotheses: assume that the Earth is a stationary body B, with
a figure (the boundary S of B), constant in time and in uniform rotation (with
constant angular velocity ω) around an axis having a fixed direction in an inertial
reference system, as well as with respect to the body of the Earth itself; assume
further that B is the support of a stationary mass distribution, that produces
a stationary Newtonian field with potential V (x) called gravitational potential,
that is also in uniform rotation, with angular velocity ω, as seen from an inertial
reference system; therefore we can define a gravity potential
1
W (x) = V (x) + ω 2 ρ2
2
(ρ = distance of x from the rotation axis, that we
take as Z coordinate axis, i.e., ρ2 = X 2 + Y 2 ) (1.1)
The Analysis of the Geodetic Boundary Value Problem 461
the gradient of which

g(x) = ∇W (x) = ∇V + ω 2 xe
(xe = projection on the (X, Y ) plane of x) (1.2)
is the gravity field, namely the force field per unit mass felt by a point mass at
rest in a reference system rotating with the Earth (a so-called Terrestrial Reference
System).
At every point of S we assume to be able to perform a gravimetric measure-
ment providing
g(x) = |∇W | , x ∈ S ; (1.3)
moreover we assume that the direction of the vertical
g(x)
n(x) = − , (1.4)
g(x)
is also observable by astrogeodetic measurements. The unit vector n is usually
parameterized by 2 angles Σ = (Λ, Φ), the astrogeodetic longitude and latitude,
that identify the direction of n in the terrestrial reference system. In addition we
assume that, by combining leveling networks with gravimetry, we come to know
the potential W (x), ∀x ∈ S.
If we accept the above hypotheses, we arrive at the following definition of the
Vector GBVP.
Definition 1.1 (VGBVP). Given
g(x) = −g(x)n(x) , W (x) x∈S (1.5)
to find a surface
S ≡ {x = x(Σ)} (1.6)
and a regular harmonic potential V (x)
ΔV (x) = 0 , x ∈ Ω ≡ (B ∪ S)c , (1.7)
such that
1
W (Σ) = V (x) + ω 2 (X 2 + Y 2 )x=x(Σ) . (1.8)
2
g(Σ) = g(Σ)n(Σ) ≡ ∇V + ω 2 xe x=x(Σ) (1.9)
As one can see the problem is a free boundary, non-linear BVP for the Laplace
operator.
A first remark on this formulation is that the centrifugal terms in (1.9), (1.8)
are small compared to the main gravitational part. For instance max |ω 2 xe | ∼
3 Gal (1 Gal = 1 cm sec−2 ), as opposed to |∇V | ∼ 103 Gal. Even more, if we
compute centrifugal potential and acceleration by using a reference figure, like the
Earth ellipsoid, the residual unknown part goes down to 10−5 ÷ 10−6 times the
main terms.
462 F. Sansò
It is then only natural to think of analyzing first a rotation-free VGBVP, and

to send the solution of (1.3), (1.9)(1.8) to a subsequent iteration process. So we
take as a reference the problem of the following Definition 1.2.
Definition 1.2 (rotation-free VGBVP). To find a surface S ≡ {x = X(Σ)} and a
potential V (x) such that
ΔV (x) = 0 x ∈ Ω (the exterior of S) (1.10)
μ
V (x) = + O r12 (regularity at infinity) (1.11)
r
V (x)|S = V [x(Σ)] = W (Σ) (1.12)
∇V (x)|S = (∇V )[x(Σ)] = g(Σ) ; (1.13)
here μ stems for μ = GM, where G is Newton’s universal constant and M is the
mass of the Earth.
Let us underline that solving (1.10), (1.11), (1.12), (1.13) in a suitably lin-
earized form, is known to provide solutions good at the limit of present day errors
in the observations and even beyond. So the conceptual importance of the non-
linear GBVP is to have theorems that guarantee not only the existence and the
uniqueness of the solution but also its continuous dependence on the data, in such a
way that we do not introduce unwanted simplifications in passing to the linearized
version.
One feature that emerges clearly in the rotation-free VGBVP is its invariance
under translation.
It is in fact easy to verify that given a solution S ≡ {x = X(Σ)}, V (x)(x ∈ Ω)
of (1.10) to (1.13), the following family
S
≡ {x = X(Σ) + c}, V
(x) = V (x − c) (∀x ∈ Ω
≡ Ω + c) (1.14)
provides a solution too, for every constant vector c.
Indeed the original formulation was hiding this fact, because the small cen-
trifugal terms are breaking the invariance in the X and Y directions, leaving
unaltered only that along Z. Nevertheless it is clear that centrifugal terms give
second-order variations to data, under an equatorial shift, when this last is small.
Therefore the problem is naturally weakly sensitive to such parameters that will
then be determined at most with large errors; a feature this that we want to avoid.
So we want first of all to find conditions that assure uniqueness of the solution.
We claim that such conditions can be to fix the origin of the coordinates at the
barycenter of the masses; as it is known this corresponds to the asymptotic formula
(see [28])
μ 1
V (x) = + O , (1.15)
r r3
meaning that V (x) has not a first-order harmonic component, when r runs far
away from the masses. This is equivalent to say that we will search for a solu-
tion in a space of harmonic functions of co-dimension 3. This indeed puts three
constraints on the spaces of data W (Σ), ∇W (Σ). Following Hörmander instead
of formulating this requirement on the data we rather augment the space of un-
knowns by introducing 3 unknown constants into the boundary relation, namely
we substitute (1.12) with the relation
V (x)|S = W (Σ) + a · v(x)|S (1.16)
where a is a 3D vector and V (x) a triple of functions that guarantee the satisfaction
of (1.15). The simplest choice can be
Yik (x) xk
vk (x) = ÷ 3 (1.17)
r2 r
although other choices can be done.
So the final formulation of the VGBVP, under a non-rotating condition, is:
⎧ ΔV = 0 in Ω
⎪
⎪
⎪
⎨ V |S = V (Σ) + a · v(x)
(1.18)
⎪
⎪ ∇V |S = g(Σ)
⎪
⎩
V = μr + O r13 (r → ∞)
A first result on the analysis of (1.18) is found in the seminal paper by L.
Hörmander [13]. Basically the result is that a solution V (x) exists in Hölder space
H2,λ (Ω), X(Σ) ∈ H2,λ (σ), (σ the unit sphere), if W (Σ), g(Σ) are close enough in
H2,λ (Ω) to spherical counterparts, i.e., W0 (Σ), g0 (Σ), satisfying the relation
μ|g0 (Σ)| = V (Σ)2 . (1.19)
This is obtained with the application of hard implicit function theorems (see
[21]). A simpler but effective approach, the so-called Gravity Space approach [25,
26], makes use of the Legendre transformation, where
g = ∇V (x) (1.20)
becomes a new coordinate system and a new adjoint potential ψ, defined by
ψ =g·x−V (1.21)
is introduced.
It is then easy to see that S is mapped to a fixed surface Sg , Ω is mapped to
the interior Ωg of Sg and the Laplace equation for V becomes
T rΨ−1 = 0 ,
) 2 *
∂ ψ (1.22)
Ψ= .
∂gi ∂gk
One then verifies that, symmetrically to (1.20), the relation x = ∇ψ(g) holds, so
that the boundary condition becomes
g · ∇ψ − ψ|Sg = V (Σ) . (1.23)
Furthermore the uniqueness constraint (1.15) becomes
ψ = μ1/2 g 1/2 + O(g 3/2 ) (g → 0) ; (1.24)
464 F. Sansò
suitably modifying (1.23), with the addition of 3 unknown constants, as

∂ψ
g − ψ|Sg = V (Σ) + c · g(Σ) (1.25)
∂ψ
one gets a perfect balance and existence, and uniqueness of ψ(g) in H2,λ follows
in a suitable neighborhood of the spherical solution ψ0 = μ1/2 g 1/2 . Once ψ is
retrieved, one gets S from the already mentioned relation
x(Σ) = ∇ψ|Sg . (1.26)
Better results can even be obtained by using the so-called intermediate Schauder
estimates [8, 23].
But it is time now to go to a different BVP, which is more adherent to the
physical situation of observable quantities.
2. From the vector to the scalar GBVP (SGBVP)

The formulation of the VGBVP, as for any mathematical model, requires a good
deal of abstraction. First of all the gravity field in continental areas has been
surveyed in a rather inhomogeneous way, concerning both the spatial resolution
and the accuracy of data.
For instance Antarctica and Africa are only marginally covered by gravimet-
ric measurements, while South America has a very uneven gravity data set. On
the other hand the vast area of the ocean has only a very poor coverage of di-
rect measurements, although a complex elaboration of satellite observations and
oceanographic models can provide a significant data set in terms of both resolution
and accuracy (see O. Andersen in [28], Chapt. 9).
But it is the data set of astrogeodetic observation, i.e., those defining the
direction of the vertical n (see (1.4)), that is so quantitatively and qualitatively
poor, to make the formulation of the VGBVP too far away from reality.
Fortunately, the physical structure of the Earth gravity field helps us in find-
ing a more realistic model. As a matter of fact, the variations of g in vertical
direction are roughly two orders of magnitude larger than in any horizontal di-
rection. This reflects the prevailing horizontal layering of the bulky density of the
body of the Earth. This means also that even an imprecise knowledge of the hor-
izontal position of the measurement point is acceptable without introducing too
large errors in the determination of the gravity potential.
This circumstance has suggested geodesists (see [23, 24]) to give a differ-
ent formulation of the GBVP where the boundary is “free” only in the vertical
direction.
Introducing ellipsoidal coordinates (σ, h) ≡ (λ, ϕ, h) (longitude, latitude, el-
lipsoidal height) (see [28], §1.11) we can formulate the scalar GBVP as follows:
to find S ≡ {h = H(σ)} and V (x) = V (σ, h) such that

⎧
⎪ ΔV = 0 in Ω ≡ {h ≥ H(σ)}
⎪
⎪
⎪
⎨ V |S = W (σ) − 1 ω 2 (X 2 (σ) + Y 2 (σ))
2
(2.1)
⎪
⎪
⎪ |∇V + ω 2 (X(σ)eX + Y (σ)eY )||S = g(σ)
⎪
⎩
V (x) = O 1r .
This problem is scalar precisely in the sense that of the unknown point P
on S we know the projection on the ellipsoid, i.e., σ, but we do not know the
scalar H(σ). Such lack of information is compensated by giving both W (σ) and
g(σ) at P .
Following the same reasonings as in §1, we can eliminate, with a very minor
error, the dependence on the rotational potential too, so arriving at the somewhat
simpler formulation
⎧
⎪
⎪ ΔV = 0 in Ω ≡ {h ≥ H(σ)}
⎪
⎪
⎨ V |S = W (σ)
(2.2)
⎪
⎪ |∇V ||S = g(σ)
⎪
⎪
⎩ V = O 1r .
An interesting historical remark is that, although (2.2) and (1.10)–(1.13) are
two different problems, they have been considered as one problem only, in geodetic
literature, until the difference was clearly highlighted in [24]. The reason why this
could happen is that both problems, linearized and posed under the so-called
spherical approximation (see [28] §2.6) are reduced to the same problem, known
as the Simple Molodensky Problem, that we shall treat later in §4.
This remark is useful to us, to understand that it is convenient, though not
strictly necessary, to reformulate (2.2) in the same fashion as (1.18), namely adding
3 scalar unknowns (the vector a) and putting a stricter constraint on the asymp-
totic behaviour of V (x), i.e.,
⎧ ΔV = in Ω
⎪
⎪
⎪
⎪
⎨ V |S = W (σ) + a · v(x)
|∇V |S = g(σ) (2.3)
⎪
⎪
⎪
⎪ μ
⎩ v = + O r13 r→∞.
r
As in (1.16), (1.17) the simplest choice of v(x) is
x
c(x) = 3 . (2.4)
r S
The SGBVP has been first formulated by F. Sacerdote and F. Sansò [24],
where the idea of a partial Legendre transform has been proposed. A significant
step forward in the analysis of the problem has been done in [23]. Also recently
in [27] a variant of the aforementioned paper has been introduced and we shall
shortly summarize it hereafter.
466 F. Sansò
In this context one has to mention the work of Seitz and Heck [30], who
gave a numerical evidence of the superior characteristics of the SGBVP on the
VGBVP. Basically the concept is to swap the ordinary spherical coordinate r with
the potential V , considered, together with σ = (λ, ϕ), as a radial coordinate in
a Marussi space. The name is because Marussi has proposed long ago to use the
potential, in this case the gravity potential, as intrinsic coordinate [16]. This is
best done, first by imagining to put r(V, σ) as the solution of the implicit equation
V = V (r, σ) , (2.5)
and then considering the field variable
1
Y (V, σ) = . (2.6)
r(V, σ)
Putting
∂2 Y
Y
= ∂Y
∂V , Y
= ∂V 2
(2.7)
Yσ = ∇σ Y (V, σ), Yσσ = Δσ Y (V, σ) ,
by exploiting the implicit function theorem one arrives at the following field equa-
tion (see [27])
(Y 2 + |Yσ |2 )Y
− 2Y
Yσ · Yσ
+ Y
2 Yσσ = 0 . (2.8)
Considering in Marussi’s space the known surface
SM ≡ {V ≡ W (σ)} (2.9)
and its internal domain
ΩM ≡ {(V, σ); V ≤ W (σ)} (2.10)
we have that (2.8), translating the first of (2.3), has to hold in ΩM . Let us for the
moment disregard the additional term in the second of (2.3), a·v, that we shall add
again when the formulation of the SGBVP in Marussi space will be completed. So
the second of (2.3) is just used to define SM . The third of (2.3), further considering
the relations
2
2 ∂V 1
g = + 2 |∇σ V |2 (2.11)
∂r r
Y
Vσ + Yσ ≡ 0 , (2.12)
can be written as
Y
2 g 2 (σ) = Y 2 (Y 2 + |Yσ |2 ) (2.13)
to hold on SM .
Finally the fourth of (2.3) becomes
Y = μ−1 V + O(V 3 ) (V → 0) . (2.14)
A first remark is that the field
Y = μ−1 V , (2.15)
corresponding to the spherical field

μ μ
V = , g= 2 , (2.16)
r r
is indeed a solution of (2.8), it satisfies the boundary relation (2.13), when
g 2 (σ) = μ−2 W 4 (σ) , (2.17)
as well as the asymptotic condition (2.14). Therefore it makes sense to search for
a form of the SGBVP, which highlights that the actual field Y (V, σ) is the sum of
a spherical field plus a perturbation. This is achieved posing
Y = μ−1 V (1 + y) , (2.18)
where y will be considered a quantity small of the first order. Physical estimations
show that y, which is an a-dimensional quantity, is between 10−3 and 10−2 , so
confirming our guess. Substitution of (2.18) into (2.8) and (2.13) leads to the BVP

ΔM y = F2 (y) + F3 (y)
(2.19)
V y
− y = h + b2 (y) + b3 (y) + b4 (y)
where
∂2 2 ∂ 1
ΔM = + + 2 Δσσ , (2.20)
∂V 2 V ∂V r
k = 1, 2 Fk (y) = non-linear differential operators
homogenous of degree k in y (2.21)
k = 1, 2, 3 bk (y) = non-linear differential boundary
operators homogeneous of degree k in y (2.22)
) *
1 μ2 g 2 (σ)
h(σ) = 1− . (2.23)
2 W 4 (σ)
The explicit form of Fk (y) and bk (y) can be found in [27].
Note has to be taken that since (2.17) holds, up to quantities small of the
first order, h(σ) has to be a first-order small quantity too.
Since y has to be small, it is natural that if we want to obtain a local invert-
ibility theorem for (2.19), around y = 0, one has to look at its linearized version,
namely
ΔM y = f
(2.24)
V y
− y = h ;
this will help us to assess the augmented form of (2.19), by introducing a number
of unknown constants and the same number of additional conditions.
Indeed (2.24) is a regular oblique derivative BVP (remember that by hypoth-
esis ΩM is a starshaped domain) and we expect the Fredholm alternative to hold
for it.
It is immediate to see that the null space of (2.24) is just the family of linear
functions in v ≡ V eσ (e the unit vector pointing to σ on the unit sphere), i.e.,
y =c·v . (2.25)
468 F. Sansò
The conditions corresponding to the introduction of the unknown c ∈ R3 , are

derived naturally from (2.14), i.e.,
Y = μ−1 V + O(V 3 ) .
We make here the realistic hypothesis that μ = GM is in fact known, so that
(2.14) translated for y becomes
y = O(V 2 ) , (2.26)
i.e., y has to satisfy four conditions at the origin, namely
y(0) = 0 , ∇y(0) = 0 . (2.27)
Correspondingly we have to introduce in (2.24) not only the unknown vector
c but also another unknown constant c0 . In this way we arrive at the formulation
⎧ Δ y=f
⎪
⎪
M
⎪
⎨ V y
− y = h + c0 + c · v
(2.28)
⎪
⎪ y(0) = 0
⎪
⎩
∇y(0) = 0 .
Instructed by this discussion on the linearized SGBVP, we stipulate as well
that the non-linear problem will have the augmented form
⎧
⎪ ΔM y = F2 (y) + F3 (y)
⎪
⎪
⎨ V y
− y = h + b2 (y) + b3 (y) + b4 (y) + c0 + c · v
(2.29)
⎪
⎪ y(0) = 0
⎪
⎩
∇y(0) = 0
A choice now has to be done of the space where we want to find y. In [27] the choice
of the complement of {c0 + c · v ; c0 ∈ R, x ∈ R3 } in the Hölder space H2,λ (Ωμ )
has been done and a suitable local theorem of existence and uniqueness of y, c0 , c,
has been derived. This though requires that ΩM ∈ H2,λ (σ) and h ∈ H1,λ (σ) and
indeed that h 1,λ is suitably small. To prove the continuity of the Fréchet
derivative of the non-linear part of (2.29) is a technical but not difficult task, that
we leave to the reader (see [23, 27]). On the other hand a crucial step is to prove
the uniqueness for the homogenous version of (2.28). This is easily achieved by
putting
z = V y
− y , (v ∈ ΩM ) (2.30)
and observing that Δy = 0 ⇒ Δz = 0.
On the other hand y(0) = 0, ∇y(0) = 0 implies, through (2.30), that z(0) = 0,
∇z(0) = 0 too. So we must have

Δz = 0
⇒ z = c0 + c · v (in ΩM ), (2.31)
z|SM = c0 + x · v
but then
z(0) = c0 = 0 ∇z(0) = c = 0 .
Now (2.30) implies that

∂ y
V −2 (V y
− y) = 0 , =0, (2.32)
∂V V
i.e.,
y = a(σ)V (in ΩM ) (2.33)
But since it has to be Δy = 0, the function (2.33) can only have the linear
form
y = a·V
and finally, by the condition
∇y(0) = 0 (2.34)
we get a = 0, i.e., y ≡ 0. Uniqueness is therefore proved.
Remark 2.1 (on natural regularity conditions of data W (σ), g(σ)). By following
strictly the procedure presented in [23] one can extend the above result to the very
interesting case that
ΩM ∈ H1,λ (σ) (i.e., W (σ) ∈ H1,λ (σ)) (2.35)
and
g(σ) ∈ H1,λ (σ) (2.36)
as well. This is derived by the so-called intermediate Schauder estimates [8]. Indeed
in this case one has not λ Hölder continuous second derivatives up to the boundary,
yet y ∈ H1,λ (ΩM ), meaning basically that (recall (2.6), (2.18))
r(σ) = μW −1 (σ)[1 + y(σ)]−1 ,
i.e., the unknown surface of the Earth, is a H1,λ (σ) function too. This might be
an idealization, whereas the result of getting a surface H0,1 (σ) (i.e., a Lipschitz
surface) could be closer to reality. Yet such a strong result is till lacking for the
fully non-linear theory.
So assuming that a regularized Earth surface S ∈ H1,λ (σ) be acceptable we
see that (2.35) and (2.36) become natural regularity conditions.
In fact we know that the actual gravitational potential V (r, σ) is a Newtonian
potential, generated by a bounded mass density. This implies V ∈ H 2,p (Rn ), ∀p ≥
2 (see [18]) and, in view of the well-known Sobolev embedding theorems (see [1],
Chapter V) we have V ∈ H1,λ (Rn ) too. So the trace of V on a surface S ∈ H1,λ (σ)
is expected to be H1,λ (σ), while the trace of |∇V | on S will be in Hλ (σ).
This concludes the discussion and the analysis of the non-linear SGBVP. We
pass now to a linearization of the original problem (2.1) to get the standard linear
version of the so-called Molodensky Problem [15, 20], for which more general and
precise results will be obtained.
470 F. Sansò
3. Linearizing the SGBVP in geometry space

We proceed now to the linearization of the SGBVP in its original form (2.1),
namely formulated as a free boundary BVP in terms of the geometrical coordi-
nates (σ, h) ≡ (λ, ϕ, h). Let us recall that the first rigorous linearization has been
performed by T. Krarup for the VGBVP (see [15]). This results in an oblique
derivative BVP with a small difference in the direction along which the oblique
derivative has to be taken with respect to the SGBVP. For the vector GBVP the
direction is the so-called isozenithal, while for the scalar GBVP it becomes, as
we shall see in this section, the normal vertical direction. The two are very close
to one another and, as a byproduct of the analysis of the next sections, we will
see that there is stability of the solution for perturbations even of the direction
of the oblique derivative. This explains why, although conceptually different, the
two problems have provided in the past almost identical solutions. The focus of
this and next sections is to provide existence, uniqueness and stability of the so-
lution in a suitable Sobolev space when the data are basically in L2 (σ) and the
4
approximate boundary, {h = H(σ)}, is a starshaped Lipschitz function.
In order to linearize the problem (2.1) we need first of all approximate “val-
ues” for our unknowns, namely H(σ) and W (x).
We start form the second and we stipulate that the “approximate” potential
is just the so-called normal potential
1
U (x) = V4 (x) + ω 2 (X 2 + Y 2 ) ,
2
4
where V (x) is harmonic and such that on an ellipsoid E of revolution, with given
radius a and eccentricity e, U (x) = U0 constant. The radius a and the eccentricity
e are adapted to the mean radius and flattening of the Earth. In other words V4 (x)
is such that
ΔV4 (x) = 0 outside E
1
V4 (x)|E = U0 − ω 2 (X 2 + Y 2 )|E .
2
The value U0 is chosen according to the relation (cf. [28], part I, §1.9).
GM E 1 # #
U0 = arctg + ω 2 a2 , (b = 1 − e2 a, E = a2 − b2 ) ,
E b 3
where G is the universal constant of Newton’s law and M is chosen to be equal
to the mass of the Earth. With such a choice, the anomalous potential T (x) =
W (x) − U (x) has no centrifugal component, i.e., it is harmonic outside S, and
when the coordinate system is placed with the origin in the barycentre and the Z
axis along the rotation axis, in addition it satisfies the asymptotic relation

1
T (x) = O . (3.1)
|x|3
Note that it is empirically known that T is 5 orders of magnitude smaller than W
and therefore it is well suited for a linearization purpose.
As for the first unknown, i.e., H(P ), we take an approximate surface S, 4

called telluroid, defined by the so-called Marussi mapping, namely the point P , of
unknown ellipsoidal height hP , but of known ellipsoidal coordinates (λP , ϕP ), is
mapped to P
which is the point along the normal ν to the ellipsoid through P ,
where the condition
4
U (σ, H(σ)) = W (σ) (3.2)
is satisfied.
An important remark is that, since U (x) is a very smooth function, harmonic
and therefore also real analytic even inside E, down to a disk in the equatorial plane
of radius ea, the regularity of H 4 depends strictly on the regularity of W (σ). In
particular it is not difficult to see that if W (σ) is a Lipschitz function, thanks to the
4
implicit function theorem, H(σ) is Lipschitz too, i.e., its horizontal (i.e., parallel to
E) gradient is bounded, so that the surface S4 itself is Lipschitz and satisfies a cone
condition. A property this which is intimately related to the theory of Sobolev
solutions of elliptic equations (see [14, 17, 18]).
Now notice that
W (σ) = W (P ) = U (σ, H(σ)) + T (σ, H(σ)) ≡ U (σ, H(σ)) 4 . (3.3)
Let us put
4
δH = H(σ) − H(σ) = hP − hP = ζ(σ) , (3.4)
a quantity known in geodesy as height anomaly; moreover we denote by γ the
vector
γ (σ, h) = ∇U (σ, h) , (3.5)
called the normal gravity vector, and we set
γ(σ, h) = |γ (σ, h)| , (3.6)
the modulus of normal gravity.
Returning to (3.3) we have, retaining only first-order quantities,
W (σ) ∼ 4 + ζν · γ (σ, H)
= U (σ, H) 4 + T (σ, H)
4 ; (3.7)
namely, defining the vector
γ
4=−
n , (3.8)
γ
pointing in the direction of the normal vertical, and using again (3.3), we get
T
ζ= . (3.9)
4γ
ν · n
4
Since one can prove that, everywhere on S,
4 = cos δ4 , δ4 < 10−5
ν · n
we see that
4 − 1| < 0.510−10 ,
|ν · n
472 F. Sansò
T
a quantity that, multiplied by γ , is absolutely irrelevant. So (3.9) can be written as
T
ζ= (3.10)
γ
called in literature the Bruns’s relation.
To simplify the writing of next formulas we shall often use the symbol Hσ
instead of H(σ) and so forth for functions of σ; this should not be confused with
the horizontal gradient ∇σ H(σ), etc.
We now. go /onto the linearization of the boundary function g(σ). One has,
putting U = ∂x ∂
γ
, the matrix of the second derivatives of U ,
g(σ) = |g(σ, Hσ )| = |∇W | = |γ (σ, Hσ ) + ∇T (σ, Hσ )|
∼ 4 σ ) + Uν ζ + ∇T (σ, H
= |γ (σ, H 4 σ )|
∼
= γ(σ, H 4 σ ) + γ · (Uν ζ + ∇T )
γ
= γ(σ, H 4 σ ) − (4 4 · ∇T .
n · Uν )ζ − n (3.11)
After defining the free air gravity anomaly
4σ )
Δg = g(P ) − γ(P
) = g(σ, Hσ ) − γ(σ, H (3.12)
and recalling (3.10), we can write (3.11) in the form
4 · Uν
n
−4
n · ∇T − T = Δg , (3.13)
γ
where Δg = Δg(σ) is indeed a known datum on S. 4 We reconduct (3.13) to a more
usual form. First we note that
δ4 ∼ n − ν | < 10−5
= |4
and, since1
|∇T | < 102 mGal ,
we have too
n − ν ) · ∇T | < 10−3 mGal = 1μ Gal
|(4
which is a limit value for the error of the most precise absolute gravimeters.
Therefore we put
−4n · ∇T =∼ −ν · ∇T = − ∂T . (3.14)

∂h
Moreover, let us compute
1 1 1
∇γ = ∇γ 2 = 4.
∇(γ · γ ) = Uγ = −U n
2γ 2γ γ
1 Thegravity modulus is measured in Gal units (1Gal= 1 cm s−2 ); in these units g as well as γ
range around 103 Gal on S. Moreover it is 1mGal = 10−3 Gal, 1μ Gal = 10−6 Gal.
Taking the scalar product with ν and recalling that U is a symmetric matrix,
we get
∂γ
ν · ∇γ = 4 = −4
= −ν · U n n · Uν . (3.15)
∂h
Finally, returning to (2.14), we obtain
∂γ
∂T
− + ∂h T = Δg , (3.16)
∂h γ
which is the ordinary form of the boundary condition of the LSMP. Summarizing,
the linearized SGBVP at this stage is defined by: given S4 and Δg(σ) on S,
4 find T
such that ⎧
⎪
⎪ ΔT = 0 outside S4
⎪
⎪
⎪
⎨ ∂γ
∂h 4
− ∂T
∂h + γ T = Δg(σ) on S (3.17)
⎪
⎪
⎪
⎪ S
⎪
⎩ T = O 13
|x| |x| → ∞ ,
and then compute
T (σ) 4
ζ(σ) = ⇒ H(σ) = H(σ) + ζ(σ) . (3.18)
γ
However we have already learnt that to balance the lack of first degree har-
monics in T we have to add to Δg a linear combination of three suitable functions.
The same reasoning can be extended to the zero degree term requiring the addition
of another unknown constant in the second of (3.17). If we can assume that some
asymptotic coefficients {T m } of a harmonic expansion of T , valid outside some
4 are known, we can even extend the above reasoning
sphere S strictly enclosing S,
to degrees > 1.
This is the case, thanks to the space technology which provides us with models
of anomalous potential like
L +1
R
T = T m Ym (σ) . (3.19)
r
=0 m=−
with Ym (σ) the ordinary set of spherical harmonic functions.

The lower the L, the lower is the error (also called commission error ([28],
§3.8) that is introduced by (3.19). However, for low values of L, the over all error
is so tiny that it can be quite safely accepted. For instance for L = 24 the overall
mean commission error in terms of
∂γ
∂T
Δg = − + ∂h T (3.20)
∂h γ
has a standard deviation of σ ∼
= 1μ Gal, which is certainly negligible.
On the other hand, we will see that the introduction of (3.19) as additional
known data, will give us more freedom for the geometric constraints we will be
4
forced to put on S.
474 F. Sansò
So we decide to further modify (3.17) in the following way: we put

T = T + u , Δg = Δg + f ,
moreover we introduce suitable functions ψ jk (x), that will be specified in §4, and
unknown constants ajk , with the ordering 0 ≤ j ≤ L , −j ≤ k ≤ j, like that of
degrees and order of spherical harmonics, and we state the problem
⎧
⎪
⎪ Δu = 0 outside S4
⎪
⎪
⎨ ∂γ
L j

− ∂h
∂u
+ ∂h γ S
u = f (σ) + ajk ψ jk (x)S on S4 (3.21)

⎪
⎪
⎪
⎪ j=0 k=−j
⎩ u=O 1
, |x| → ∞
|x|L+2
Note has to be taken that the harmonic potential u will have a series repre-
sentation, outside a large sphere S, lacking the first L degrees, i.e., starting from
degree L + 1; whence the asymptotic condition in the third relation of (3.21).
The last step we take in this section is to transform the boundary conditions,
namely the second of (3.21), into a very convenient perturbative form.
This will be derived from the estimates contained in the following Proposition,
based on elementary considerations on ellipsoidal geometry and the normal gravity
formulas. A detailed proof is given in ([29] and [28], §15.2).
Proposition 3.1. Let us put (λ, ϕ) = σ for the spherical longitude and latitude of a
point P and call er (σ), ν (σ) the radial unit vector and the ellipsoidal normal unit
vector at P respectively; then one has
ε(σ) = ν (σ) − er (σ) , ε+ = max |ε(σ)| (3.22)
σ
ε+ ∼
= e2 , (e2 ∼
= 6, 7 · 10−3 ) (3.23)
where e is the eccentricity of the Earth ellipsoid E. Moreover, let us put
R(σ) = |x(σ)| , {x(σ) ∈ S}4 (3.24)
and
∂γ
η(σ) = R(σ) (σ) + 2 ; η+ = max |η(σ)| (3.25)
∂h σ
then one has
η+ ∼
= 2e2 . (3.26)
With such estimates in mind we can write
∂u ∂u
= ν · ∇u = er · ∇u + ε · ∇u = + ε · ∇u .
∂h ∂r
For the sake of brevity we shall use in the rest of the work the notation
∂u
= u
, (3.27)
∂r
so that the above relation can be written
∂u
= u
+ ε · ∇u . (3.28)
∂h
Furthermore we have, by using definition (3.27),

γ
∂γ
2 η
∂h
u= u=− u+ u . (3.29)
γ γ r r
So, multiplying the second of (3.21) by r = R(σ), changing sign and redefining
f (σ) = −Rσ f (σ) , ψjk (x) = −rψ jk (x) ,
we get finally

ru
+ 2u + (rε · ∇u − ηu)|S = f (σ) + ajk ψjk . (3.30)

S
We already note that, according to our estimates (3.23), (3.26), the differen-
tial operator
D = rε · ∇ − η (3.31)
satisfies the inequality
|Du| ≤ ε+ r|∇u| + η+ |u| , (3.32)
so that
it∂ can be considered as a perturbation of the main boundary operator
D0 = r ∂r +2 .
Summarizing, we can finally state the Linearized Scalar Molodensky Problem
in the following modified, perturbative form:
⎧
⎪
⎪ Δu = 0 outside S4
⎪
⎪
⎨
L j
ru
+ 2u = f + ajk ψjk − Du on S4 (3.33)
⎪
⎪
⎪
⎪ i=0 x=−j
⎩ u=O 1
, |x| → ∞ .
|x|L+2
In (3.33) the functions ψjk (x) will be chosen in the next section, and will
depend ultimately only on the shape of S,4 while L will be left a free parameter to
be fixed later on, to get convenient conditions for the stability of the solution.
Note that in (3.33) the unknowns are the potential u and the (L + 1)2 con-
stants {ajk }, which in the minimum case reduce to (1 + 1)2 = 4 unknowns.
The analysis of the problem (3.33) will occupy us in the next two sections,
following the ideas of [28], §15.4.
For a more general treatment of the oblique derivative problem, one can
consult the recent book [6].
4. The analysis of the Simple Molodensky Problem

We define the Simple Molodensky Problem as the linearized SGBVP (3.21) when
∂
the ∂h is approximated by ∂r∂
and (γ)−1 · ∂h
∂γ
is computed by using the spherical
expression
∂γ μ −1 μ 2
γ −1 ∼ 2 −2 3 = − . (4.1)
∂h r r r
476 F. Sansò
This leads to the boundary relation (3.30) and then to a problem like (3.33),
where however the perturbative term is disregarded, namely
⎧
⎪ 4
⎨ Δu = 0 in Ω

ru + 2u 4
⎪ =f+ Σajk ψjk on S (4.2)
⎩ u=O 1
|x| → ∞ ;
|x|2+2
this is called in geodetic literature the Simple Molodensky Problem ([15, 20]).
To study the existence, uniqueness and specially stability of the solution of
(4.2) in a suitable Sobolev space, is the task of this section, a first step to pass then
to the same work for the solution of (3.33), obtained by elementary perturbation
techniques. A first goal of the section is to prove that (4.2) has a unique stable
solution in H1 , i.e., a space of harmonic functions with L2 gradient on the boundary
4 when the following hypotheses are fulfilled
S,
S4 ≡ {r = Rσ } is Lipschitz (4.3)
or
|∇σ Rσ | ≤ C σ a.e. (4.4)
and
f ∈ L2 (σ) (4.5)
or
f 2 (σ)dσ < +∞ . (4.6)
σ
Under such hypotheses we shall prove that the solution of (4.2) satisfies
u H1 ≤ C f L2σ (4.7)
and
j = 0, 1, . . . , L , |k| ≤ j , |ajk | ≤ C f L2σ . (4.8)
Unfortunately however we know that the constant C exists finite, but we do not
know its value, so that the result appears to be too weak to establish a perturbation
theory for (3.33). Then we shall derive a more restrictive theorem, where we find
the value of C but we have to suppose that a model of the anomalous potential,
complete up to some degree and order L, is known and the inclination of S4 with
respect to er is bounded above by some suitable angle I+ < π/2. This result will
then be easily carried over to the analysis of (3.33).
Preliminary to the development of the analysis above described, is the proof
of 4 propositions.
Proposition 4.1. Let the hypotheses (4.4) and (4.6) be satisfied. Then there is one
and only one u harmonic in Ω such that
f (σ) = lim u(r, σ) σ a.e. ; (4.9)
r→Rσ
4 i.e.,
in addition, if we call R any Brillouin radius for S,
R > R+ = sup{Rσ } , (4.10)
σ
and we put
M (u, σ) = sup |u(r, σ)| (4.11)
Rσ <r≤R
we have
1
M (u, σ) 2L2σ = M (u, σ)2 dσ ≤ C 2 f 2L2σ , (4.12)
4π
where the constant C in (4.12) is depending only on the shape of S. 4
Proof. This is a generalization of the Fatou theorem ([3]), adapted to the present
configuration, proved by B.E.J. Dahlberg, and reported in [14], §3. The theory
is essentially based on Hardy–Littlewood maximal inequalities. So we shall not
reproduce the proof, but we rather send the interested reader to the mentioned
paper. Here, we note only that (4.12), by applying a simple approximation of f by
a continuous function, implies as well that if Sδ is a family of uniformly Lipschitz
surfaces, Sδ = {r = Rδ,σ = Rδ + hσ ; 0 < hσ ≤ δ}, supσ |∇Rδ,σ | ≤ C, then for any
u harmonic in Ω attaining boundary values f (σ) ∈ L2σ , one has too

lim |u(Rδ,σ , σ) − f (σ)|2 dσ = 0 . (4.13)
δ→0 σ

Remark 4.1. We notice that the first proof of existence of harmonic functions, or
more generally of solutions of Poisson’s equation, when boundary data on S4 are in
4 is due to Cimmino ([5]). However the approach of this author requires the
L2 (S)
boundary S4 to be of class C 2 (σ) so as to guarantee that it has bounded curvature. A
different proof can be given, much simpler than Hardy–Littlewood theory, where
the boundary S4 is required only to be of class H1+λ (σ). This constitutes thus
a generalization of Cimmino’s result, though it is not as general as Dahlberg’s
theorem.
Remark 4.2. As a further remark we observe that a consequence of Proposition
4.1 is that we can define a Hilbert space of harmonic functions in Ω, 4 that we
2
shall call H0 , which is just the isometric image of Lσ , through the solution of the
corresponding Dirichlet problem, namely

4 ↔ f (σ) = u| ∈ L2
u ∈ H0 (S) S σ
2
(4.14)
u 0 ≡ f L2σ .
Note should be taken that with the above definition,

2
u 0 = u2 (Rσ , σ)dσ , (4.15)
σ
2
which is not exactly equal to S u (Rσ , σ)dS, although it is equivalent to the more
common norm when S4 is a starshaped Lipschitz domain, because (see [28], §15.1)

dS = Rσ Jdσ ,
(4.16)
J = (cos I)−1 , cos I = n · er , 1 ≤ J ≤ A < +∞ .
478 F. Sansò
Proposition 4.2. There are functions {ψjk } ∈ H0 such that

1
∀u ∈ H0 , ψjk , u0 = ujk ≡ u(R, σ)Yjk (σ)dσ , (4.17)
4π
that are linearly independent in H0 .
Moreover, if we call
VL ≡ {ψjk ; 0 ≤ i ≤ L , |k| ≤ j} (4.18)
and
H0L = VL⊥ ≡ {u ∈ H0 ; ψjk , u0 = 0 ψjk ∈ VL } (4.19)
then we have
1
u ∈ H0L ⇔ u = O . (4.20)
rL+2
Proof. That {ψjk } exist is just an application of the Riesz theorem when we realize
that, (4.10) being satisfied, we have from (4.17)
1/2
1
|ψjk u0 | ≤ u(R, σ)2 dσ ≤ C u 0 . (4.21)
4π
Moreover {ψjk } are linearly independent because, if for some constants {cjk }

L
j
cjk ψjk ≡ 0
j=0 k=−j
then, by exploiting the definition (4.17) and the orthogonality of surface spherical
harmonics, ∀ ≤ L, |m| ≤ ,
= +1 >
L j
Rσ
0= cjk ψjk , Ym = cm .
r
j=0 k=−j 0
Finally (4.20) holds because the very definition (4.19) yields the double im-
plication

1
u=O ⇔ {um = 0 , ≤ L |m| ≤ } ⇔ u ∈ H0L .
rL+2
Proposition 4.3. Let us consider the modified Dirichlet problem for all L ≥ 0,
⎧
⎪ 4 ,
⎨ Δu = 0 in Ω

L j
(4.22)
⎪ 4 ;
⎩ u|S = f + aik ψjk on Ω
j=0 k=−j
2
when f ∈ L2σ , then (4.22) has one and only one solution u ∈ H0L , {ajk } ∈ R(L+1) ,
such that
u 0 ≤ f L2σ (4.23)
and
|ajk | ≤ C f L2σ . (4.24)
4 such that
Proof. Let us define u

Δ4u=0 4
(in Ω)
(4.25)
4|S = f ,
u
which exists and satisfies
u
4 0 ≡ f L2σ (4.26)
4∈
in force of Remark 4.2. Note that in general u / H0L , not even
for
L = 0, because
functions u ∈ H0,L=0 have the asymptotic behaviour u = O r12 .
Since the second of (4.22) is equivalent to u ∈ H0L , (4.22) itself is just
equivalent to solving
L j
4=u−
u ajk ψjk (4.27)
j=0 k=−j
with
u ∈ H0L = VL⊥ and Σajk ψjk ∈ VL . (4.28)
So the solution of (4.27), introducing the orthogonal projector
PL ≡ orthogonal projector on VL .
is just

L
j
ajk ψjk = −PL u
4, (4.29)
=0 k=−j
u = (I − PL )4
u. (4.30)
In turn, since {ψjk } are linearly independent, (4.29) implies
|ajk | ≤ C PL u
4 0 ≤ C u
4 0 = C f L2σ
which is (4.24); moreover (4.30) implies
u 0 ≤ u
4 0 ≡ f L2σ .
which is (4.23).
Finally we shall prove a proposition that introduces a generalized version of
energy integrals already used in Geodesy for the analysis of the GBVP ([13, 27,
28]).
Proposition 4.4. Let u ∈ H0L and ∇u ∈ H0L+1 , let α be a real number such that
α < 2L + 4 , (4.31)
then the following identity holds

dσRσα+2 |∇u|2 = 2(α − 1) rα−1 u
2 dΩ
Ω
S
(4.32)
−α dΩr α−1 2 4 σα u
un .
|∇u| + 2 dSR

Ω
S
480 F. Sansò

Proof. Note that S dσ in the left-hand side of (4.33) means that |∇u|2 is computed
4 i.e. with r = Rσ , and the resulting function integrated in dσ on the unit
on S,
sphere. Similar is the situation of the right-hand side where we can put
dS4 = Rσ2 Jdσ
and u
un has again to be computed on S. 4 To prove (4.32) one can start from the
differential identity
1 ∂
∇ · (rα u
∇u) = (α − 1)rα−1 u
2 + rα−1 |∇u|2 + rα |∇u|2 . (4.33)
2 ∂r
4 and using Gauss theorem and the integration by
By integrating (4.33) in Ω
parts in dr, one derives (4.32). The condition (4.31) is necessary in order that all
integrals be convergent and integrals on large spheres tend to zero when the radius
tends to infinity.
We only mention here, for some possible future use, that (4.32) can be further

elaborated, trying to bring at least in part the dΩrα−1 |∇u|2 on the surface S; 4
so we arrive at the identity

dσRσα+2 |∇u|2 = 2(α − 1) rα−1 u
2 dΩ + Rσα−1 (2Rσ u
+ αu)un dS4 (4.34)
S
Ω S

α(α − 1) α2 (α − 1)
− Rσα u2 dσ − dΩrα−3 u2 .
2 S 2
Ω
We are ready now to go to the fundamental theorem of this section.

Theorem 4.1. Let S4 and f satisfy (4.4) and (4.5) respectively; then there is one
and only one solution u of the SMP (4.2), such that u ∈ H1 ∩ H0L for all L ≥ 1
and
u 1 ≤ C f L2σ , (4.35)
4
where the constant C depends only on the shape of S, namely on the Lipschitz
4
constant of S.
Proof. Let us call v the unique solution of the generalized Dirichlet problem
⎧
⎪ Δv = 0 4
in Ω
⎪
⎪
⎨ L j
v|S = f + ajk ψjk on S4 (4.36)
⎪
⎪
⎪
⎩ j=0 k=−j
v ∈ H0L ,
which exists and is unique on the basis of Proposition 4.3, in particular when
L ≥ 1.
If you assume that (4.2) has one solution in H1 ∩ H0L , then indeed
v = ru
+ 2u (4.37)
4 is a simple direct computation;
is the solution of (4.6). That v is harmonic in Ω
that v ∈ H0L (L ≥ 1) descends from the fact that u ∈ H1 and that, for r ≥ R,
v can be expressed in spherical harmonics, so that (4.37) implies the well-known

Stokes relation
(j − 1)
v jk = − ujk , (j ≥ L + 1) (4.38)
R
and therefore v jk = 0 for j ≤ L if the same is true for ujk . Note that the condition
L > 1 has to be set up exactly to allow a one-to-one correspondence between v jk
and ujk , for all j ≥ L + 1.
Now we have to reverse the above statement.
Namely given v satisfying (4.36) we start defining u satisfying (4.37), i.e.,

1 +∞
u(r, σ) = − 2 v(s, σ)sds . (4.39)
r r
That u is harmonic in Ω4 is verified by direct calculation of Δu in spherical
1

coordinates; that u = O rL+2 , when L ≥ 1, comes from (4.38) and the fact that
v jk ≡ 0 for j ≤ L. that u ∈ H0 we verify by the following reasoning. Let us put

1 R 1 +∞
(Rσ ≤ r ≤ R) , u(r, σ) = − 2 v(s, σ)sds − 2 v(s, σ)sds
r r r R
1 1
≡ − 2 u1 (r, σ) − 2 u2 (σ) ; (4.40)
r r
since Rσ is bounded above and below, we need only to prove that u1 (Rσ , σ), u2 (σ) ∈
L2σ . On the other hand, recalling (4.11), (4.12), (4.23),
R
2
dσu1 (Rσ , σ) ≤ dσ v 2 (s, σ)s2 ds · (R − Rσ )
Rσ
≤ C v 20 ≤ C f 2L2σ (4.41)
Moreover we have
+∞ +∞ j+1
R
u2 (σ) = v(s, σ)sds = v jk Yjk (σ)sds
R R s
j=(L+1)
2 v jk
=R Yjk (σ) ,
j−1
j=L+1
so that, recalling again (4.11), (4.12), (4.23),

2 1/2 1/2
1
u2 (σ) L2σ ≤ R Σv 2jk =R v(R, σ)2 dσ ≤ C f L2σ . (4.42)
4π
We note that combining (4.41) and (4.42) one has
u 0 ≤ C f L2σ . (4.43)
On the other hand, since u satisfies (4.37), we obviously have
1
u
0 = (v − 2u) 0 ≤ C f L2σ . (4.44)
r
482 F. Sansò
Now we finally prove that u ∈ H1 . In this respect we note that a natural norm in
H1 is
u 1 = |∇u| 0 = |∇u|2 dσ ; (4.45)

S
that this is a true norm is obvious because u 1 = 0 implies |∇u| = 0 on S4 and
then by the maximum principle
sup |∇u|2 ≤ sup |ux |2 + sup |uy |2 + sup |uz |2 = 0 ,

Ω
S
S
S
so that u is constant in Ω4 and being regular at infinity, has to be zero in the

4
whole Ω.
Now we can apply Proposition 4.4 with α = 0 and L ≥ 1, getting

dσRσ2 |∇u|2 = −2 r−1 u
2 dΩ + 2 dSu 4
un ; (4.46)

Ω
recalling (4.16) and putting
J+ = sup J , R− = inf Rσ > 0 , R+ = sup Rσ ,
we find then

2
R− dσ|∇u|2 ≤ 2R+
2
dσJ|u
||un |
1/2 1/2
2
≤ 2R+ J+ dσu
2 dσu2n
1/2 1/2
2
≤ 2R+ J+ dσu
2 dσ|∇u|2 . (4.47)
Combining (4.47) and (4.44) we readily arrive at

u 1 ≤ C f L2σ . (4.48)
Strictly speaking we can apply the above reasoning only if we know a priori
that u ∈ H1 , which is a condition for the validity of Proposition 4.4. Yet we can
easily circumvent the difficulty by the following approximation process.
Let us put
uλ (r, σ) = u(λr, σ) (4.49)
4
with λ ≥ 1; it is obvious that this is a family of functions harmonic in Ω and that
in particular the trace of uλ (r, σ) on S4 is for λ > 1 a smooth function, i.e.,
uλ (Rσ , σ) = u(λRσ , σ)
which is bounded in σ and even with a bounded |∇σ uλ (Rσ , σ)|, since Rσ is Lip-
schitz. So the function
1
λ > 1 , |∇uλ |2 S = u
(λRσ , σ)2 + 2 |uσ (λRσ , σ)|2
Rσ
is certainly bounded and therefore in L2σ , i.e., uλ ∈ H1 , ∀λ > 1.
Let us note for the sake of clarity that indeed ∇σ u(λRσ , σ) = uσ (λRσ , σ)
since the second function is just uσ (λRσ , σ) = ∇σ u(λr, σ)|r=Rσ .
So (4.47) can be legitimately applied to the difference uλ − uλ , i.e.,
uλ − uλ 1 ≤ C u
λ − u
λ 0 ; (4.50)
on the other hand we have already noted that u
∈ H0 , so that u
λ → u
in L2σ .
Therefore from (4.50), if we take a sequence λn → 1 we see that uλn is
a Cauchy sequence in H1 and then uλu → u ∈ H1 . But we already know that
uλn → u in H0 and therefore it has to be u = u, namely u ∈ H1 . So our conclusion
(4.48) holds.
In principle at this point one could be content with the result of Theorem
∂
4.1 considering that indeed a perturbation of the boundary operator r ∂r + 2,
continuous in H1 , can be treated claiming that when the perturbation is “small
enough”, then a solution still exists and is unique in H1 . Yet, since we do not know
the value of C in (4.48), we are not able to specify how small should actually be
the perturbation. So we need to find a value for C. However, now that we know
that u ∈ H1 , we can resort to the generalized energy identities of Proposition 4.4,
but this time with α = 1 and L ≥ 1, as already done in ([27, 29]). In this case
(4.32) reads

3 2
dσRσ |∇u| = − dΩ|∇u| + 2 dσJRσ3 u
un
2
Ω
S
S
= dσJRσ (uun + 2Rσ u

un )
2
(4.51)

S
On the other hand, on S4

Rσ u
= −2u + f ,
so that (4.51) becomes

3 2
dσRσ |∇u| = dσJRσ2 (2f un − 3uun) ; (4.52)

S
S
we are ready now to prove the following Theorem 4.2.
Theorem 4.2. Let us put

R+
R− = inf Rσ , R+ = sup Rσ , k = , δR = R+ − R− ,
R−
and also
1
J+ = sup J(σ) =
cos I+
I+ = sup I = sup arccos n · er
n(σ) = normal to S4 at (Rσ , σ) ;
484 F. Sansò
then, if the condition

δR 4
3J+ k 3 C0L = 3J+
2 3
k + <1 (4.53)
R− 2L + 5
is satisfied, one has
u 1 ≤ C1L f 0 (4.54)
with
1 2J+ k 2
C1L = . (4.55)
R− 1 − 3J+ k 3 C0L
Proof. The proof comes from a combination of (4.52) and the following lemma.
4 and such that u ∈ H1 ∩ H0L , then
Lemma 4.1. Let u be harmonic in Ω
u 0 ≤ R+ C0L u 1 , (4.56)
Proof. Let us call u+ (σ) = u(R+ , σ); then one can write
uσ = uσ − u+ + u+ , u 0 ≤ u − u+ 0 + u+ 0 . (4.57)
Now consider in Ω+ ≡ {r ≥ R+ } the identities
+∞

u n+1
R+
u+ (σ) = u+nm Ynm (σ) , (4.58)
r
u=L+1 m=−n
+∞

n
(n + 1)2
u
(r, σ)2 dΩ = 4πR+ u2+nm , (4.59)
Ω+ 2n + 1
u=L+1 m=−n
Since
(n + 1)2 2L + 5
(n ≥ L + 1) , ≥
2n + 1 4
and
u+ 20 = 4πΣu2+nm ,
from (4.59) we deduce

1 4
u+ 20 ≤ u
(r, σ)2 dΩ . (4.60)
R+ 2L + 5 Ω+
Moreover
2 R+
R+
2
1 1
|u − u+ | = u (s, σ)sdσ ≤ − u
2 s2 ds ,
Rσ Rσ R+ Rσ
1 1
so that, noting that Rσ − R+ ≤ δR
R+ R− ,

δR
u − u+ 2o ≤ u
2 dΩ , (4.61)
R+ R− δΩ
where δΩ ≡ {(r, σ), Rσ ≤ r ≤ R+ }.
From (4.60), (4.61), used in (4.57) by applying the Cauchy–Schwarz inequal-

ity and the Gauss theorem, we get
) *
1 δR 4
u 20 ≤ + u
2 dΩ + u
2 dΩ
R+ R− 2L + 5 δΩ Ω+
) *
1 δR 4
≤ + ∇u2 dΩ
R+ R− 2L + 5
) * Ω
1 δR 4 2
= + R+ J+ u 0 u n 0
R+ R− 2L + 5
≤ R+ C0L u 0 u 1 .
Simplifying by u 0 we get (4.56) and the lemma is proved.
Now we can go back to (4.52) and applying the Schwarz inequality we get
3
R− u 21 ≤ 2R+
2 2
J+ f 0 u 1 +3R+ J+ u 0 u 1 . (4.62)
Simplifying and using (4.56) we obtain
2
R+ 3
3 u 1 ≤ 2J+ f 0 +3k J+ C0L u 1 .
R−
(4.63)
It is then clear that, if (4.53) is satisfied, (4.54) holds too.

We shall discuss in the next section the geometric meaning of (4.53) together
with the parallel condition holding for the complete linearized SGBVP. Here we
just summarize the results of this section by commenting that the SMP has been
proved to have a stable solution in H1 ∩ H0L , unconditionally for any Lipschitzian
S4 and data in L2σ , and that a specific evaluation of the majorization constant has
obliged us to introduce the restrictive condition (4.53).
5. Analysis of the linearized SGBVP

We can turn now to the analysis of the BVP (3.33), that we summarize in the
following Theorem.
Theorem 5.1. Assume that the hypotheses of Theorem 4.1 hold true and that (4.53)
is verified, so that C1L of (4.55) is positive constant. Assume further on that the
condition
C1L R+ (ε+ + η+ C0L ) < 1 (5.1)
is satisfied too. Then the problem (3.33) has a unique solution u ∈ H1 ∩ H0L ,
2
{ajk } ∈ R(L+1) , such that
u 1 ≤ C2L f 0 (5.2)
where
C1L
C2L = . (5.3)
1 − C1L R+ (ε+ + η+ C0L )
486 F. Sansò
Proof. Note that the problem (3.33) is equivalent to the solution of a SMP with
known term f − Du.
Therefore from Theorems (4.1) and (4.2) we know that
u 1 ≤ C1L f − Du 0 (5.4)
|ajk | ≤ C f − Du 0 . (5.5)
Recalling (3.32) and (4.56) we observe that indeed
Du 0 ≤ ε+ R+ ∇u 0 +η+ u 0
≤ ε+ R+ u 1 +η+ R+ C0L u 1 (5.6)
= R+ (ε+ + η+ C0L ) u 1 (5.7)
Therefore, concentrating on (5.4), we find
u 1 ≤ C1L f 0 +C1L R+ (ε+ + η+ C0L ) u 1 . (5.8)
It is then clear that if (5.1) is true, then (5.8) yields (5.2).
This indeed guarantees uniqueness and stability of the solution. As for ex-
istence however, it is enough to observe that if S is the solution operator of the
functional part of the SMP, then we can write
u = Sf − SDu (5.9)
and the condition (5.1) is exactly implying that SD, as an operator from H1 ∩ H0L
into itself, is a contraction. Therefore the existence of u is a consequence of the
well-known theorem on contractions [9, 31].
Remark 5.1. Now that we have defined two conditions for the stability constants
of the SMP and of the linearized SGBVP, namely (4.58) and (5.1), it is interesting
to investigate their functional and geometrical significance. As we see there are
two parameters to play with, namely J+ = (cos I+ )−1 and L. The third constant
R+ δR
k= =1+
R− R−
is in fact fixed by the global geometry of the Earth surface. If we take
R− = b = 6356.91 km
R+ = a + Hc = 6384.52 km
where
Hc = 6.27 km
is the height of the Chimborazo, a high mountain close to the equator, the tip of
which is probably the furthest point of the surface from the center of the Earth,
we get
δR
= 4.34 · 10−3 .
R−
Other considerations of the geometric meaning of functional properties of the

GBVP can be found in ([11]). Now if we want to be able to treat our BVP for
boundaries with inclinations up to 60◦ , one has to put J+ = 2.
With such a constraint one verifies that with L = 23
3J+ k 3 C023 = 0.995 < 1
while
3J+ k 2 C022 = 1.033 < 1
This means that if we want to be sure that the SMP has a stable solution we have
to subtract from the data the contribution of a global model complete up to degree
23, what, as commented in §3, can be realistically done with a negligible error.
Similarly, one can work with the second condition that, recalling Theorem 3.1,
can be written
2J+ k 3
C1L R+ (ε+ + η+ C0L ) = e2 (1 + 2C0L ) < 1 .
1 − 3J+ k 3 C0L
So for L = 23 we obtain
C122 R+ (ε+ + η+ C023 ) = 6.545 > 1
and the condition (5.1) is not satisfied.
However, already for L = 24 one has
C124 R+ (ε+ + η+ C024 ) = 0.874 < 1
meaning that, subtracting a model complete up to degree 24, the linearized SGBVP
becomes tractable for telluroids with inclination up to 60◦ . This is certainly a
realistic and nice result.
6. Conclusions and open questions

Summarizing the state of the art of the analysis of the GBVP we can agree on the
following statements:
• the non-linear problem, when the rotational terms are treated only as per-
turbations, has a satisfactory theory, especially in that data are required to
have a natural degree of regularity, namely W (σ) ∈ H1,λ , g(σ) ∈ H0,λ ,
• for the SMP we have a quite nice result, requiring only that the telluroid S4
be starshaped and Lipschitz, while gravity anomalies on S4 are only required
to be in L2σ ,
• for the linearized GBVP one has to force the theory of the solution of the
SMP, in order to get a quantitative determination of the constants involved
therein. Then the linear GBVP has a theorem of existence uniqueness and
stability with reasonable requirements for the regularity of data, but enforcing
stronger conditions both geometrical (limited inclination of the telluroid S) 4
and on the data (need of a certain realistic number of asymptotic coefficients
of the potential).
488 F. Sansò
The present theory could therefore be improved in some aspects:

• studying the numerical values of the constants involved in the non-linear
theory, maybe generalizing it to other spaces than weighted Hölder spaces,
• to try to develop a theory like that of the SMP, but in ellipsoidal, rather than
in spherical approximation, as in [12],
• to try to develop an H1,p theory for the linearized GBVP, that could allow
4
to significantly relax the constraints on the inclination of S.
In this respect, let us remark that it is the free boundary character of the
GBVP, reflected in the term 2u in the boundary operator, that imposes more
severe constraints. An analogous theory for the fixed boundary GBVP requires
much weaker conditions to obtain a theorem of existence uniqueness and stability
in H1 . In fact it is only required that L = 1, a standard condition for the anomalous
potential, and inclinations can be up to 89◦ , what is even too much for a reasonable
global model of the surface of the Earth.
References
[1] Adams R.A. Sobolev Spaces. Pure and Applied Mathematics, Academic Press, New
York, San Francisco London, 1975.
[2] Anderson O. In: Sansò F., Sideris M. Geoid determination: theory and methods,
Chapter 9, Part II. Springer-Verlag Berlin-Heidelberg (2013).
[3] Axler S., Bourdon P., Ramsey W. Harmonic Function Theory. Springer-Verlag
(2001).
[4] Banach S. Sur les opérations dans les ensembles abstraits et leur application aux
équations intégrales. Fundamenta Mathematicae (1922).
[5] Cimmino G. Spazi hilbertiani di funzioni armoniche e questioni connesse. in:
Equazioni lineari alle derivate parziali. UMI Roma (1955).
[6] Freeden W., Gerhards C. Geomathematically oriented potential theory. Chapman &
Hall/CRC Press, Taylor Francis, Boca Raton (2013).
[7] Friedman A. Variational principles and Free Boundary Value Problems. John Wiley
& Sons, New York (1982).
[8] Gilbarg D., Trudinger N.S. Elliptic Partial Differential Equations of Second Order.
Classics in Mathematics, Springer, Berlin, Heidelberg (2001).
[9] Granas A., Dugundji J. Fixed Point Theory. Springer Monographs in Mathematics,
vol. 41 (2003).
[10] Grothaus M., Raskop T. Oblique Stochastic Boundary Value Problems. Handbook
of GeoMath, pp. 1052–1076 (2010).
[11] Holota P. Coerciveness of the linear gravimetric boundary value problem and a
geometric interpretation. J. of Geod. 71:640–651 (1997).
[12] Holota P., Nesvadba O. Reproducing kernel Hilbert Space for the exterior of an ellip-
soid and the method of successive approximations in solving geodetic boundary value
problems. VIII Hotine Marussi Symp. on Math. Geod. Rome, June 2013. Springer
(IAG Symp. Series, in press).
[13] Hörmander L. The Boundary Value Problems in Physical Geodesy. Arch. Prot. Mech.
Anal., n. 62, pp. 51–52 (1976).
[14] Jerison D.S., Kenig C. Boundary Value Problems on Lipschitz Domains. MAA Stud-
ies in Math. v. 23, pp. 1–68 (1982).
[15] Krarup T. Mathematical Foundation of Geodesy. Springer, Berlin, Heidelberg (2006).
[16] Marussi A. Intrinsic geodesy. Springer, Berlin (1985).
[17] McLean W. Strongly Elliptic Systems and Boundary Integral Equations. Cambridge
University Press (2000).
[18] Miranda C. Partial Differential Equations of Elliptic Type. Springer-Verlag, Berlin-
Heidelberg (1970).
[19] Molodensky M.S., Eremeev V.F., Yurkina M.I. Methods for the study of the gravita-
tional field of the Earth. Transl. Russian Israel Program for Scient. Trans. Jerusalem
(1960).
[20] Moritz H. Advanced Physical Geodesy, Wichmann Verlag, Karlsruhe (1980).
[21] Moser J. A rapidly convergent iteration method and non-linear differential equations.
Acc. Sc. Norm. Sup. Pisa, v. 20. pp. 265–315 (1966).
[22] Pavlis N.K. In: Sansò F., Sideris M. Geoid determination: theory and methods,
Chapter 6, Part II. Springer-Verlag Berlin-Heidelberg (2013).
[23] Otero J., Sansò F. An analysis of the scalar geodetic boundary value problems with
natural regularity results. J. of Geod., v. 73, pp. 437–435 (1999).
[24] Sacerdote F., Sansò F. The Scalar Boundary Value Problem of Physical Geodesy,
Man. Geod. 11, pp. 15–28 (1986).
[25] Sansò F. Discussion on the existence and uniqueness of the solution of Moloden-
sky’s problem in gravity space. Rend. Acc. Naz. Lincei (Sc. Fisiche, Matematiche,
Naturali) V. 61, S. 8 (1976).
[26] Sansò F. The Geodetic Boundary Value Problem in Gravity Space. Memorie Acc.
Lincei V. 14, S. 8, n. 3 (1977).
[27] Sansò F. Geodetic Boundary Value Problem. Handbook of Geomathematics, Sprin-
ger-Verlag, London (2014).
[28] Sansò F., Sideris M. Geoid determination: theory and methods. Springer-Verlag
Berlin-Heidelberg (2013).
[29] Sansò F., Venuti G. On the Explicit Determination of stability Constants for the
Linearized Geodetic Boundary Value Problems. J. Geod. v. 82, pp. 909–916 (2008).
[30] Seitz K., Heck B. Effects of Non-linearity in the Geodetic Boundary Value Problems.
Deutsche Geodätische Kommission, Reihe A, Heft Nr. 109, Verlag der Bayerischen
Ac. der Wissenschaften, Munich (1993).
[31] Yoshida K. Functional Analysis. Classics in Mathematics, Springer-Verlag, Berlin,
Heidelberg (1980).
Fernando Sansò
Politecnico di Milano – DICA
P.zza Leonardo da Vinci, 32
I-20133 Milano, Italy
e-mail: fernando.sanso@polimi.it

Oblique Stochastic Boundary Value Problem

Martin Grothaus and Thomas Raskop
Abstract. Aim of this note is to report the current state of the analysis for
weak solutions to oblique boundary problems for the Poisson equation. In this
paper, as well deterministic as stochastic inhomogeneities are treated and ex-
istence and uniqueness results for corresponding weak solutions are presented.
We consider the problem for inner bounded and outer unbounded domains
in Rn . Main tools for the deterministic inner problem are a Poincaré inequal-
ity and some analysis for Sobolev spaces on submanifolds, in order to use
the Lax–Milgram Lemma. The Kelvin transformation enables us to translate
the outer problem to a corresponding inner problem. Thus we can define a
solution operator by using the solution operator of the inner problem. The
extension to stochastic inhomogeneities is done with help of tensor product
spaces of a probability space with the Sobolev spaces from the deterministic
problems. We can prove a regularization result which shows that the weak
solution fulfills the classical formulation for smooth data. A Ritz–Galerkin
approximation method for numerical computations is available. Finally, we
show that the results are applicable to geomathematical problems.
Keywords. Oblique derivation problem, Ritz–Galerkin approximation, sto-
chastic extensions.
1. Introduction
The main subject of this article are existence results for solutions to oblique bound-
ary problems for the Poisson equation. We start with the deterministic problems.
The Poisson equation in the domain Σ is given by
Δu = f,
and the oblique boundary condition by
a, ∇u + bu = g.
Oblique derivative problems represent classical topics in the context of mathematical geodesy.
Thus, the editors and the publisher have decided to include this chapter, although an earlier
version has been published in “Handbook of Geomathematics, Vol. 3, 2285–2315, 2015”.
492 M. Grothaus and T. Raskop
This condition is called regular if the equation

|a, ν| > C > 0,
holds on ∂Σ for a constant 0 < C < ∞. The problem is called outer problem if
the Poisson equation has to hold on an outer domain Σ ⊂ Rn . This is a domain Σ,
having the representation Σ = Rn \D where 0 ∈ D is a bounded domain. Conse-
quently, ∂Σ divides the Euclidian space Rn into a bounded domain D, called inner
domain, and an unbounded domain Σ, called outer domain. A problem defined on
a bounded domain is called inner problem. A classical solution corresponding to
continuous a, b, g and f of the oblique boundary problem for the Poisson equation
is a function u ∈ C 2 (Σ)∩C 1 (Σ) which fulfills the first two equations. For the outer
problem u must be regular at infinity, i.e., u(x) → 0 for |x| → ∞. Existence and
uniqueness results for a classical solution to regular oblique boundary problems
for the Poisson equation are already available, see, e.g., [11, 13] or [16]. In order
to allow very weak assumptions on boundary, coefficients and inhomogeneities, we
are interested in weak solutions from Sobolev spaces of one times weakly differen-
tiable functions. When facing the deterministic problems, we have to distinguish
the inner and the outer setting. The reason is that a Poincaré inequality, namely

∇u, ∇u dλn + u2 dH n−1 ≥ C u2 dλn + ∇u, ∇u dλn ,
Σ ∂Σ Σ Σ
for all u ∈ H 1,2 (Σ), is only available for bounded Σ. Thus we can only use the Lax–
Milgram Lemma for the inner problem in order to gain a solution operator. For
the outer problem we use the Kelvin transformation to transform the unbounded
domain Σ to a bounded domain ΣK via

x
ΣK := x ∈ Σ ∪ {0}.
|x|2
Additionally we transform coefficients as well as inhomogeneities and end up with
a inner problem, which posses a unique weak solution v. Finally we transform this
function to the outer space by

1 x
u(x) := v ,
|x|n−2 |x|2
for all x ∈ Σ. This u is then the weak solution to the outer problem and it can be
shown that, in the case of existence, u is the classical solution. Additionally the
transformations are continuous and consequently the solution depends continuous
on the data. Before we go on with stochastic inhomogeneities and stochastic weak
solutions we want to mention that we have to assume a regular inner problem, while
we have a transformed regularity condition for the outer problem resulting from
the transformations. Going to a stochastic setting we have to introduce the spaces
of stochastic functions. These are constructed as the tensor product of L2 (Ω, dP ),
with a suitable probability space (Ω, F , P), and the Sobolev spaces used in the
deterministic theory. They are again Hilbert spaces and we have isomorphisms
to Hilbert space-valued random variables. For the stochastic inner problem we
Oblique Stochastic Boundary Value Problem 493
again employ the Lax–Milgram Lemma, while in the outer setting we define the
solution operator pointwisely for almost all ω ∈ Ω. For all solutions, deterministic
as well as stochastic, a Ritz–Galerkin approximation method is available. Finally
we give some examples from Geomathematics, where stochastic inhomogeneities
are implemented. Proofs for the results presented in this report are given in [9] and
[10]. The examples are taken from [3] and [7]. We want to mention that the articles
[15] as well as [17] also deal with solutions to oblique boundary-value problem.
2. Scientifically relevant domains and function spaces

In this article we consider boundary value problems for the Poisson equation. This
means we are searching for a function which satisfies the Poisson equation in a
subset Σ of Rn and an additional condition on the boundary ∂Σ of this set.
Δu = f, in Σ,
a, ∇u + bu = g on ∂Σ,
f and g are called inhomogeneities, a and b are called coefficients and such a
function u is then called solution. Our analysis is motivated by problems from
Geomathematics. Here oblique boundary problems arise frequently, because in
general the normal of the Earth’s surface does not coincide with the direction of
the gravity vector. Therefore, the oblique boundary condition is more suitable then
a Neumann boundary condition. For details see [3] or [12]. We are dealing with
two different types of sets Σ, namely bounded and outer C m,α -domains, which are
introduced by the following definition. In particular the outer problem is of major
interest for applications.
Definition 1. ∂Σ ⊂ Rn is called a C m,α -surface , m ∈ N and 0 ≤ α ≤ 1 and Σ is
called a bounded C m,α -domain , if and only if
• Σ is a bounded subset of Rn which is a domain, i.e., open and connected,
• There exists an open cover (Ui )i=1,...,N of ∂Σ and corresponding C m,α -dif-
n
feomorhisms Ψi : B1R (0) → Ui , i = 1, . . . , N , such that
Ψi : B10 (0) → Ui ∩ ∂Σ,
Ψi : B1+ (0) → Ui ∩ Σ,
Ψi : B1− (0) → Ui ∩ Rn \Σ,
n
where B1R (0) denotes the open unit ball in Rn , i.e., all x ∈ Rn with |x| < 1.
n
B10 (0) denotes the set of all x ∈ B1R (0) with xn = 0, B1+ (0) denotes the set
n n
of all x ∈ B1R (0) with xn > 0 and B1− (0) denotes the set of all x ∈ B1R (0)
with xn < 0.
On the other hand Σ is called an outer C m,α -domain, if and only if Σ ⊂ Rn is open,
connected and representable as Σ := Rn \D, where D is a bounded C m,α -domain
such that 0 ∈ D. Ψi is called C m,α -diffeomorphism if and only if it is bijective,
Figure 1. C m,α -surface

(Ψi )j ∈ C m,α B1R (0) , Ψ−1
n
i j
∈ C m,α Ui , j = 1, . . . , n and we have for the
n
determinant of the Jacobian Matrix of Ψi , Det(DΨi ) = 0 in B1R (0).
In Figure 1, such a C m,α -surface is illustrated. For this definition and further
details see, e.g., [6]. The definition is independent of the mappings chosen. ∂Σ is a
compact closed and doublepointfree (n − 1)-dimensional C m,α -submanifold. The
outer unit normal vector ν is a C m−1 -vector field. Furthermore, we find a C ∞ -
partition of (wi )1≤i≤N on ∂Σ corresponding to the open cover (Ui )1≤i≤N , provided
by Alt [2]. H n−1 denotes the (n − 1)-dimensional Hausdorff measure on ∂Σ and λn
the Lebesgue measure in Rn . Throughout this paper we assume at least a Lipschitz
boundaries, i.e., C 0,1 -boundaries ∂Σ. Then we have ν ∈ L∞ (∂Σ; Rn ). Note that
some geomathematical relevant examples are even C ∞ -surfaces, e.g., a sphere or
an ellipsoid. We will see in the following chapters that the case of bounded and
outer domains have to be treated differently, because the unboundedness causes
problems which do not occur in the bounded setting. Non the less, we are searching
in both cases for solutions under as weak assumptions as possible. More precisely
we are searching for solutions in Sobolev spaces for inhomogeneities from Banach
space duals of Sobolev spaces. These spaces are introduced in the following.
Definition 2. Let Σ be a bounded C 0,1 -domain and r ∈ N. We define
! "
H r,2 (Σ) := F : Σ → R|∂1α1 · · · ∂nαn F ∈ L2 (Σ) for all α1 + · · · + αn ≤ r ,
r N 1/2
2
F H (Σ) :=
r,2 ∂ F L2 (Σ)
α
.
|α|=0 i=1
Let Σ be an outer C 0,1 -domain and 1 , 2 , 3 be continuous, positive functions

defined on Σ. We define

2 2 2
L1 (Σ) := F : Σ → R|F is measurable with F (x)1 (x)dλ (x) < ∞ ,
n
Σ
! "
H1,21 ,2
(Σ) := F ∈ L21 (Σ)|∂i F ∈ L22 (Σ), 1 ≤ i ≤ n ,
! "
H2,2
1 ,2 ,3
(Σ) := F ∈ L21 (Σ)|∂i F ∈ L22 (Σ) and ∂i ∂j F ∈ L23 (Σ), 1 ≤ j, i ≤ n ,
1/2
2 2
F L2 (Σ) := n
F (x)1 (x)dλ (x) ,
1
Σ

n 1/2
F H1,2, (Σ) := F 2L2 (Σ) + ∂i F 2L2 (Σ) ,
1 2 1 2
i=1
n 1/2

n
F H2,2, (Σ) := F 2L2 (Σ) + ∂i F 2L2 (Σ) + ∂i ∂j F 2L2 (Σ) .
1 2 ,3 1 2 3
i=1 j=1
Let ∂Σ be a C 0,1 -surface and (wi )1≤i≤N be the C ∞ -partition of unity of ∂Σ

corresponding to the open cover from Definition 1. For a function F defined on
∂Σ we obtain a function θi F defined on Rn−1 by:
n−1
(wi F )(Ψi (y, 0)) y ∈ B1R (0),
θi F (y) :=
0 otherwise.
Let now ∂Σ be a C m,1 -surface, m ∈ N. Furthermore let s ∈ R, r ∈ N, with

s < m + 1 and 0 ≤ r ≤ m. Then we define
! "
H s,2 (∂Σ) := F : ∂Σ → R|θi F ∈ H s,2 (Rn−1 ), 1 ≤ i ≤ N ,
! "
H r,∞ (∂Σ) := F : Σ → R|θi F ∈ H r,∞ (Rn−1 ), 1 ≤ i ≤ N ,
N 1/2
2
F H s,2 (∂Σ) := θi F H s,2 (Rn−1 ) ,
i=1

sn−1
F H r,∞ (∂Σ) := max0≤|s|≤r,1≤i≤N ess supB Rn−1 (0) |∂1s1 · · · ∂n−1 θi F | ,
1
where H 0,p (∂Σ) is identical with Lp (∂Σ), p ∈ {2, ∞}. The spaces H2,2 1 ,2 ,3
(Σ),
H1,2
1 ,2
(Σ), L 2
1 (Σ), H s,2
(∂Σ) and H r,2
(Σ) are Hilbert spaces, while the spaces
H r,∞ (∂Σ) are Banach spaces with respect to the norms given above, see, e.g.,
[1] or[5].
The spaces H s,2 (Rn−1 ) are defined via the Fourier transformation. Differen-
tiation in the definition above has to be understood in sense of weak differentia-
tion. The definition of the spaces on ∂Σ above is independent from the choice of
(Ui )1≤i≤N , (wi )1≤i≤N and (Ψi )1≤i≤N . It is left to introduce the spaces H s,2 (∂Σ)
on a C m,1 -surface ∂Σ, 0 ≤ s < m + 1. We do this as follows. Identify each function
F ∈ L2 (∂Σ) with a linear continuous functional on H s,2 (∂Σ), defined by

F (G) := F (x) · G(x) dH n−1 (x),
∂Σ

for all G ∈ H s,2 (∂Σ). Then H s,2 (∂Σ) is defined as

s,2

H (∂Σ) := L2 (∂Σ) ,
· (H s,2 (∂Σ))
where
|F (G)|
F (H s,2 (∂Σ)) := sup .
G∈H s,2 (∂Σ) G H s,2 (∂Σ)
In this way we end up with the space H −s,2 (∂Σ) defined in the previous definition.
We get the following chain of rigged Hilbert spaces, called Gelfand triple.
H s,2 (∂Σ) ⊂ L2 (∂Σ) ⊂ H −s,2 (∂Σ),
densely and continuously. Additionally we have for the duality product

H −s,2 (∂Σ) F, G s,2
H (∂Σ) = F (x) · G(x) dH n−1 (x),
∂Σ
2
for all F ∈ L (∂Σ). Analogously, we introduce the Gelfand triples

H 1,2 (Σ) ⊂ L2 (Σ) ⊂ H 1,2 (Σ) ,
for bounded C 0,1 -domains and

1,2 2 1,2
H|x| 2 ,|x|3 (Σ) ⊂ L|x|2 (Σ) ⊂ H|x| 2 ,|x|3 (Σ) ,
for outer C 0,1 -domains.
3. Poincaré inequality as key issue for the inner problem

In this chapter we will show how to derive a weak formulation for the deterministic
inner regular oblique boundary value problem defined on bounded C 1,1 -domains Σ.
The corresponding weak solution will obviously coincide with the classical solution
in the case of existence. First we derive a weak formulation, then a Poincaré in-
equality for the Sobolev space H 1,2 (Σ) allows us to apply the Lax–Milgram Lemma
in order to provide a solution operator. At next we translate a regularization result
for the Neumann boundary value problem to the oblique boundary value problem.
Finally a Ritz–Galerkin method allows us to approximate the weak solutions with
help of numerical calculations. We proceed with the stochastic extensions. This
means we introduce the stochastic function spaces for the inhomogeneities and
corresponding solutions with help of the tensor product. Then the results for the
deterministic problem can be easily extended to the stochastic setting. The chapter
is divided into five sections according to the described approach.
3.1. The weak formulation

In this section we present the theory of weak solutions to the regular oblique
boundary problem for the Poisson equation for inner domains. Although the weak
problem can be formulated for bounded C 0,1 -domains, in order to prove the ex-
istence of a unique weak solution we need at least a bounded C 1,1 -domain. Con-
sequently we assume Σ ⊂ Rn throughout this section to be such a domain, if not
stated otherwise. At first we give the definition of the regular oblique boundary
problem together with the definition of the classical solution.
Definition 3. Let Σ be a bounded C 1,1 -domain, f ∈ C 0 (Σ), g, b ∈ C 0 (∂Σ) and
a ∈ C 0 (Σ; Rn ) be given, such that
|a(x), ν(x)| > C1 > 0, (1)
2 1
for all x ∈ ∂Σ, where 0 < C1 < ∞. Finding a function u ∈ C (Σ) ∩ C (Σ) such
that
Δu = f in Σ,
a, ∇u + bu = g on ∂Σ,
is called inner regular oblique boundary problem for the Poisson equation and u
is called classical solution.
Because of the condition (1) the problem is called regular. It just means that
the vector field a is non-tangential to ∂Σ for all x ∈ ∂Σ. Now we derive the weak
formulation. The fundamental theorem of the calculus of variations gives
Δu = f in Σ
if and only if
ηΔu dλn = ηf dλn for all η ∈ C0∞ (Σ)
Σ Σ
if and only if
ηΔu dλn = ηf dλn for all η ∈ C ∞ (Σ).
Σ Σ
Additionally on Σ the following Green formula is valid

∂ψ
ϕΔψ dλn + ∇ϕ, ∇ψ dλn = ϕ dH n−1 ,
Σ Σ ∂Σ ∂ν
for all ψ ∈ C 2 (Σ) ∩ C 1 (Σ) and ϕ ∈ C ∞ (Σ). This yields for a classical solution

∂u
η dH n−1 − ∇η, ∇u dλn = ηf dλn ,
∂Σ ∂ν Σ Σ
for all η ∈ C ∞ (Σ). Now we transform the boundary condition

a, ∇u + bu = g on ∂Σ,
to the form
∂
a, ν u + a − (a, νν) · ∇∂Σ u + bu = g on ∂Σ.
∂ν
Using equation (1) we divide by a, ν = 0 to get the equivalent boundary

condition
F G
∂ a b g
u+ − ν , ∇∂Σ u + u= on ∂Σ.
∂ν a, ν a, ν a, ν
Plugging this condition into the equation above, we get the following formulation of
the regular oblique boundary problem for the Poisson equation which is equivalent
to the formulation given in Definition 3. We want to find a function u ∈ C 2 (Σ) ∩
C 1 (Σ) such that
F G
g b a
η − u− − ν, ∇∂Σ u dH n−1
a, ν a, ν a, ν
∂Σ

− ∇η, ∇u dλn − ηf dλn = 0 for all η ∈ C ∞ (Σ).
Σ Σ
The transformation of the boundary term is shown in Figure 2.
Figure 2. Transformation of the oblique boundary condition
Finally, we are weakening the assumptions on data, coefficients, test function

and solution. We give the weak formulation of the inner regular oblique boundary
problem to the Poisson equation, summarized in the following definition.
Definition 4. Let Σ be a bounded C 1,1 -domain, a ∈ H 1,∞ (∂Σ; Rn ) fulfilling con-
1
dition (1), b ∈ L∞ (∂Σ), g ∈ H − 2 ,2 (∂Σ) and f ∈ H 1,2 (Σ) . We want to find a

function u ∈ H 1,2 (Σ) such that
F G n F G
g ai
η, − η − νi , (∇∂Σ u)i
1
H 2 ,2 (∂Σ) a, ν H − 12 ,2 (∂Σ) i=1 H 12 ,2 (∂Σ) a, ν 1
H − 2 ,2 (∂Σ)

b
− (∇η · ∇u) dλn − η u dH n−1 − H 1,2 (Σ) η, f (H 1,2 (Σ)) = 0,
Σ ∂Σ a, ν
for all η ∈ H 1,2 (Σ). Then u is called a weak solution of the inner regular oblique
boundary problem for the Poisson equation.
3.2. Existence and uniqueness results for the weak solution

It is possible to prove the following existence and uniqueness result for the weak
solution to the deterministic inner oblique boundary value problem for the Poisson
equation.
Theorem 5. Let Σ be a bounded C 1,1 -domain, a ∈ H 1,∞ (∂Σ; Rn ), fulfilling condi-
tion (1), and b ∈ L∞ (∂Σ) such that:

b 1 a
ess inf − div∂Σ −ν > 0. (2)
∂Σ a, ν 2 a, ν

1
Then for all f ∈ H 1,2 (Σ) and g ∈ H − 2 ,2 (∂Σ) there exists one and only one
weak solution u ∈ H 1,2 (Σ) of the inner regular oblique boundary problem for the
Poisson equation. Additionally we have for a constant 0 < C2 < ∞:

u H 1,2 (Σ) ≤ C2 f (H 1,2 (Σ)) + g − 21 ,2 .
H (∂Σ)
In the proof we apply the Lax–Milgram Lemma, which gives us a unique

u ∈ H 1,2 (Σ) fulfilling the variational equation
F (η) = a(η, u),
1,2
for all η ∈ H (Σ), provided we have that F and a are continuous and addition-
ally a is a coercive bilinear form. F and a can be obtained easily from the weak
formulation as
F G
g
F (η) = 1 η, − 1,2 η, f (H 1,2 (Σ)) ,
H 2 ,2 (∂Σ) a, ν H − 12 ,2 (∂Σ) H (Σ)
n F G
ai
a(η, u) = η − νi , (∇∂Σ u)i
1 ,2
i=1 H 2 (∂Σ)
a, ν 1
H − 2 ,2 (∂Σ)

b
+ (∇η · ∇u) dλn + η u dH n−1 .
Σ ∂Σ a, ν
The continuity can be shown by some results about the Sobolev spaces occurring
in the weak formulation. In order to prove that a is coercive, i.e., |a(u, u)| ≥
C3 u 2H 1,2 (Σ) , the Poincaré inequality

2 2
∇F, ∇F dλ + n
F dH n−1
≥ C4 F dλ + ∇F, ∇F dλ ,
n n
Σ ∂Σ Σ Σ
which is valid for all F ∈ H 1,2 (Σ) and a constant 0 < C4 < ∞, is indispensable.
Finally, the condition

b 1 a
ess inf − div∂Σ −ν > 0,
∂Σ a, ν 2 a, ν
is also essential to ensure the coercivity of a. Condition (2) can be transformed
into the equivalent form

1 2 a
a, νb > a, ν div∂Σ −ν H n−1 -almost everywhere on ∂Σ.
2 a, ν

a
If div∂Σ a,ν − ν = 0 H n−1 -almost everywhere on ∂Σ, we have for H n−1 -almost
all x ∈ ∂Σ the condition from the existence and uniqueness result for the classical
solution. Furthermore for a = ν, i.e., the Robin problem, the condition reduces
to b > 0 H n−1 -almost everywhere on ∂Σ. Finally, we are able to define for each
bounded C 1,1 -domain Σ, a ∈ H 1,∞ (∂Σ; Rn ) and b ∈ L∞ (∂Σ), fulfilling conditions
in
(1) and (2), a continuous invertible linear solution operator Sa,b by
in

1
Sa,b : H 1,2 (Σ) × H − 2 ,2 (∂Σ) → H 1,2 (Σ),
(f, g) → u,
where u is the weak solution provided by Theorem 5. In addition this means that
the inner weak problem is well posed.
3.3. A regularization result
In this section we will show that the weak solution from the previous section is even
an element of H 2,2 (Σ) if we choose the inhomogeneities and coefficients smooth
enough. The result for the oblique boundary problem is based a regularization
result for the weak solution to the Neumann problem for the Poisson equation.
Theorem 6. Let Σ ⊂ Rn be a bounded C 2,1 -domain, a ∈ H 2,∞ (∂Σ; Rn ) fulfilling
1
condition (1) and b ∈ H 1,∞ (∂Σ). Then for all f ∈ L2 (Σ) and g ∈ H 2 ,2 (∂Σ),
the weak solution u ∈ H 1,2 (Σ) to the inner regular oblique boundary problem for
the Poisson equation, provided in Theorem 5, is even in H 2,2 (Σ). Furthermore we
have the a priori estimate

u H 2,2 (Σ) ≤ C5 f L2(Σ) + g 12 ,2 .
H (∂Σ)
for a constant 0 < C5 < ∞.

In order to prove the result it suffices to show that the normal derivative of
1
the weak solution u of the oblique boundary problem is an element of H 2 ,2 (∂Σ).
Therefore we use some results for Sobolev spaces defined on submanifolds. The
weak solution in H 2,2 (Σ) is related to the classical solution in the following way.
Let u ∈ H 2,2 (Σ) be the weak solution to the inner regular oblique boundary
problem for the Poisson equation, provided by Theorem 6. Then we have
Δu = f λn − almost everywhere in Σ,
a, ∇u + bu = g H n−1 − almost everywhere on ∂Σ.
Such a solution we call strong solution to the inner regular oblique boundary
problem for the Poisson equation.
3.4. Ritz–Galerkin approximation
In this section we provide a Ritz–Galerkin method which allows us to approximate
the weak solution with help of a numerical computation. Let a(η, u) and F (η)
be defined as above and the conditions of Theorem 9 be satisfied. Furthermore
let (Un )n∈N be an increasing sequence of finite-dimensional subspaces of H 1,2 (Σ),
;
i.e.,Un ⊂ Un+1 such that n∈N Un = H 1,2 (Σ). Because Un is as a finite-dimensional
subspace of the Hilbert space H 1,2 (Σ) itself a Hilbert space, we find for each n ∈ N
a unique un ∈ Un with
a(η, un ) = F (η) for all η ∈ Un .
Moreover, let d := dim(Un ) and (ϕk )1≤k≤d be a basis of Un . Then un ∈ Un has
the following unique representation

d
un = hi ϕi ,
i=1
where (hi )1≤i≤d is the solution of the linear system of equations given by

d
a(ϕj , ϕi )hi = F (ϕj ) 1 ≤ j ≤ d.
i=1
The following result from Céa proves that the sequence (un )n∈N really approxi-
mates the weak solution u.
Theorem 7. Let u be the weak solution provided by Theorem 5 and (un )n∈N taken
from above. Then:
C6 n→∞
u − un H 1,2 (Σ) ≤dist(u, Un ) −→ 0,
C7
where C6 and C7 are the continuity and the coercivity constants of a.
3.5. Stochastic extensions
First we define the spaces of stochastic functions. We are choosing a probability
space (Ω, F , P), arbitrary but fixed, such that L2 (Ω, dP ) is separable, and define
1,2

H (Σ) Ω := L2 (Ω, P ) ⊗ H 1,2 (Σ) ∼ = L2 (Ω, P ; H 1,2 (Σ) ),

1
− ,2
HΩ 2 (∂Σ) := L2 (Ω, P ) ⊗ H − 2 ,2 (∂Σ) ∼
1 1
= L2 (Ω, P ; H − 2 ,2 (∂Σ)),
1
HΩ2 (∂Σ) := L2 (Ω, P ) ⊗ H 2 ,2 (∂Σ) ∼
1 1
= L2 (Ω, P ; H 2 ,2 (∂Σ)),
,2
L2 (Σ) := L2 (Ω, P ) ⊗ L2 (Σ) ∼

= L2 (Ω, P ; L2 (Σ))
Ω
1,2
HΩ (Σ) := L2 (Ω, P ) ⊗ H 1,2 (Σ) ∼
= L2 (Ω, P ; H 1,2 (Σ)),
HΩ2,2 (Σ) := L2 (Ω, P ) ⊗ H 2,2 (Σ) ∼
= L2 (Ω, dP ; H 2,2 (Σ)),
with help of the tensor product. Now we can investigate the stochastic inner
regular oblique boundary problem for the Poisson equation. We are searching for
a solution u ∈ HΩ1,2 (Σ) of:
Δu(x, ω) = f (x, ω) for all x ∈ Σ, P-a.a. ω ∈ Ω,
(a · ∇u(x, ω)) + bu(x, ω) = g̃(x, ω) for all x ∈ ∂Σ, P-a.a. ω ∈ Ω,
|(a · ν)| ≥ C8 > 0 on ∂Σ.
Using the argumentation from the first section we come immediately to the
weak formulation of the stochastic boundary problem.
Definition 8. Find u ∈ HΩ1,2 (Σ) with:

g b
1 η, − 1 ,2 − η u dH n−1
dP
H 2 ,2 (∂Σ) a, ν H 2 (∂Σ)
Ω ∂Σ a, ν

n
ai
− 1 η , (∇∂Σ u)i − 12 ,2 dP
H 2 ,2 (∂Σ) a, ν H (∂Σ)
Ω i=1

− (∇η · ∇u) dλn −H 1,2 (Σ) η, f (H 1,2 (Σ)) dP = 0
Ω Σ
for all η ∈ HΩ1,2 (Σ).

u is called stochastic weak solution of the stochastic inner
regular oblique boundary problem for the Poisson equation.
Obviously, u ∈ HΩ1,2 (Σ) is a stochastic weak solution of the stochastic regular
oblique boundary problem for the Poisson equation if and only if for P-a.a. ω ∈ Ω,
uω := u(·, ω) is a weak solution of the deterministic problem
Δuω = f (·, ω) on Σ,
a · ∇uω + buω = g(·, ω) on ∂Σ.
The solution operator of the deterministic problem extends to the stochastic
setting in the following way
Theorem 9. Let Σ be a bounded C 1,1 -domain, a ∈ H 1,∞ (∂Σ; Rn ), fulfilling condi-
tion (1), and b ∈ L∞ (∂Σ) such that:

b 1 a
ess inf − div∂Σ −ν > 0.
∂Σ a, ν 2 a, ν

− 1 ,2
Then for all f ∈ H 1,2 (Σ) Ω and g ∈ HΩ 2 (∂Σ) there exists one and only
one stochastic weak solution u ∈ HΩ1,2 (Σ) of the stochastic inner regular oblique
boundary problem for the Poisson equation. Additionally we have for a constant
0 < C9 < ∞

u H 1,2 (Σ) ≤ C9 f (H 1,2 (Σ))Ω + g − 21 ,2 .
Ω HΩ (∂Σ)
In the proof we use the results from the deterministic setting in order to
prove the requirements of the Lax–Milgram Lemma to be fulfilled. Using the iso-
morphisms of the tensor product spaces to spaces of Hilbert space-valued random
variables, also the regularization result translates to the stochastic setting.
Theorem 10. Let Σ ⊂ Rn be a bounded C 2,1 -domain, a ∈ H 2,∞ (∂Σ; Rn ) fulfilling
1
,2
condition (1) and b ∈ H 1,∞ (∂Σ). Then for all f ∈ L2Ω (Σ) and g ∈ HΩ2 (∂Σ),
the weak solution u ∈ HΩ1,2 (Σ) to the inner regular oblique boundary problem for
the Poisson equation, provided in Theorem 5, is even in HΩ2,2 (Σ). Furthermore we

have the a priori estimate

u H 2,2 (Σ) ≤ C10 f L2Ω(Σ) + g 12 ,2 .
Ω HΩ (∂Σ)
for a constant 0 < C10 < ∞. u is called stochastic strong solution and fulfills the
classical problem almost everywhere.
At the end of this section we want to mention that a Ritz–Galerkin approxi-
mation is available also for the stochastic weak solution, repeating the procedure
from the deterministic problem. For details and proofs of the presented results we
refer to [9].
4. Fundamental results for the outer problem

In this chapter we provide a solution operator for the outer oblique boundary prob-
lem for the Poisson equation. The results presented in this chapter are taken from
[10] and as well the proofs as further details can be found in this reference. The
outer problem is defined in an unbounded domain Σ ⊂ Rn which is representable
as Rn \D, where D is a bounded domain. Additionally we assume 0 ∈ D which
is necessary for the Kelvin transformation. For unbounded Σ a Poincaré inequal-
ity is yet missing. Consequently we cannot use the technique used for the inner
problem because we are unable to prove coercivity of the bilinear from a weak
formulation corresponding to the outer problem. Thus we will not derive a weak
formulation for the outer problem and thus we do not have to consider a regular
outer problem. Our approach is to transform the outer problem to a corresponding
inner problem for which a solution operator is available by the results of the previ-
ous chapter. In this way we will construct our weak solution and for this solution
also a Ritz–Galerkin Method is available because of the continuity of the Kelvin
transformation. Finally we again extend our results for stochastic inhomogeneities
as well as stochastic solutions and present some examples from Geomathematics.
The described procedure is presented in the following four sections.
4.1. Transformations to an inner setting
In this section we define the transformations which will be needed in order to
transform the outer oblique boundary problem for the Poisson equation to a cor-
responding regular inner problem. Then we will apply the solution operator in
order to get a weak solution in the inner domain. This solution will be trans-
formed with help of the Kelvin transformation to a function defined in the outer
domain. In the next section we will finally prove that this function solves the outer
problem for sufficiently smooth data almost everywhere, which gives the connec-
tion to the original problem. The whole procedure is illustrated in the following
Table 1.
We proceed in the following way. First we define the Kelvin transformation
KΣ of the outer domain Σ to a corresponding bounded domain ΣK . At next the
out

Sa,b
Outer problem : Σ (f, g) −→ u

↓ KΣ T1 ↓ T2 ↑K

in
ST 3 (a),T4 (b)
Inner problem: ΣK (T1 (f ), T2 (g)) −→ v

Table 1. Transformation procedure
Kelvin transformation K of the solution for the inner problem will be presented.
Finally we define the transformations T1 and T2 for the inhomogeneities as well as
T3 and T4 for the coefficients. We will also show that the operators K, T1 and T2
are continuous. The consequence is that our solution operator

out
Sa,b (f, g) := K STin3 (a),T4 (b) (T1 (f ), T2 (f )) ,
forms a linear and continuous solution operator for the outer problem. Because all
main results assume Σ to be at least an outer C 1,1 -domain, we fix Σ for the rest
of this chapter as such a domain, if not stated otherwise. At first we transform the
outer domain Σ to a bounded domain ΣK . The tool we use is the so called Kelvin
transformation KΣ for domains. We introduce the Kelvin transformation for outer
C 1,1 -domains in the following definition.
Definition 11. Let Σ be an outer C 1,1 -domain and x ∈ Σ be given. Then we define
the Kelvin transformation KΣ (x) of x by
x
KΣ (x) := .
|x|2
Furthermore, we define ΣK as the Kelvin transformation of Σ via

ΣK := KΣ (Σ) ∪ {0} = KΣ (x)x ∈ Σ ∪ {0} .
From this point on, we fix the notation in such a way that ΣK always means
the Kelvin transformation of Σ. Figure 3 illustrates the Kelvin transformation of Σ.
We have KΣ ∈ C ∞ (Rn \{0}; Rn\{0}) with KΣ2 = IdRn \{0} . Furthermore
we obtain by standard calculus, using the Leibnitz formula for the determinant,
|Det(D(KΣ ))(x)| ≤ C11 |x|−2n for all x ∈ Rn \{0}, 1 ≤ i ≤ n. This is one of the
reasons for the weighted measures of the Sobolev spaces introduced later on. More-
over the transformation leaves the regularity of the surface invariant. Let Σ be an
outer C 2,1 -domain. Then ΣK is a bounded C 2,1 -domain. Moreover we have that
∂ΣK = KΣ (∂Σ). Furthermore, if Σ is an outer C 1,1 -domain, we have that ΣK is a
bounded C 1,1 -domain. There are geometric situations in which ∂ΣK can be com-
puted easily. For example if ∂Σ is a sphere around the origin with radius R, then
∂ΣK is a sphere around the origin with radius R−1 . Furthermore, if ∂Σ ⊂ R2 is an
ellipse with semi axes a and b around the origin, then ∂ΣK is also an ellipse around
Figure 3. Kelvin transformation of Σ
the origin with semi axes b−1 and a−1 . At next we present the transformation for
the weak solution of the inner problem back to the outer setting. Therefore we in-
troduce the operator K. This is the so-called Kelvin transformation for functions.
It transforms a given function u, defined in ΣK , to a function K(u), defined in
Σ. In addition, it preserves some properties of the original function. We will state
some of these properties. So, after the following considerations it will be clear why
we choose exactly this transformation. It will also be clear how we have to choose
the transformations T1 , . . . , T4 in the following. We start with a definition.
Definition 12. Let Σ be an outer C 1,1 -domain and u be a function defined on ΣK .
Then we define the Kelvin transformation K(u) of u, which is a function defined
on Σ, via
1 x
K(u)(x) := u ,
|x|n−2 |x|2
for all x ∈ Σ.
Important is, that this transformation acts as a multiplier when applying the
Laplace operator. Note that −(n − 2) is the only exponent for |x| which has this
property. We have for u ∈ C 2 (ΣK ) that K(u) ∈ C 2 (Σ) with

1 x
Δ(K(u))(x) = (Δu) ,
|x|n+2 |x|2
for all x ∈ Σ. As already mentioned above we will apply K to functions from
H 1,2 (ΣK ). So we want to find a normed function space (V, · V ) such that
K : H 1,2 (ΣK ) → V
defines a continuous operator.
It turns out that the weighted Sobolev space H 1,2
1 1
, |x|
(Σ) is a suitable choice.
|x|2
We have the following important result for K acting on H 1,2 (ΣK ).

Theorem 13. Let Σ be an outer C 1,1 -domain. For u ∈ H 1,2 (ΣK ) let K(u) be
defined as above for all x ∈ Σ. Then we have that
K : H 1,2 (ΣK ) → H 1,2
1 1
, |x|
(Σ)
|x|2
is a continuous linear operator. Moreover K is injective.
It is left to provide the remaining transformations T1 , . . . , T4 . In the first part

we treat T1 , which transforms the inhomogeneity f of the outer problem in Σ to
an inhomogeneity of the corresponding inner problem in ΣK . Assume f to be a
function defined on Σ. We want to define the function T1 (f ) on ΣK , such that
Δu(x) = T1 (f )(x), x ∈ ΣK , (3)
implies that
Δ(K(u))(y) = f (y), y ∈ Σ. (4)
We are able to define T1 for functions defined on Σ as follows.
Definition 14. Let Σ be an outer C 1,1 -domain and f be a function defined in Σ.

Then we define a function T1 (f ) on ΣK by

1 x
T1 (f )(x) := f ,
|x|n+2 |x|2
for all x ∈ ΣK \{0} and T1 (f )(0) = 0.
T1 is well defined and fulfills the relation described by equations (3) and (4).
Furthermore, T1 defines a linear continuous isomorphism
T1 : L2|x|2 (Σ) → L2 (ΣK ),
with (T1 )−1 = T1 . We want to generalize our inhomogeneities in a way similar

to the inner problem. This means we have to identify a normed vector space
(W, · W ), such that

T1 : W → H 1,2 (ΣK ) ,
defines a linear continuous operator. Additionally, we want to end up with a
Gelfand triple
U ⊂ L2|x|2 (Σ) ⊂ W.
Consequently L2|x|2 (Σ) should be a dense subspace. It is possible to prove that the

1,2
space H|x| 2 ,|x|3 (Σ) is a suitable choice. Recall the Gelfand triple, given by

1,2 2 1,2
H|x| 2 ,|x|3 (Σ) ⊂ L |x| 2 (Σ) ⊂ H |x|2 ,|x|3 (Σ) .
Theorem 15. We define a continuous linear operator

T1 : L2|x|2 (Σ) → H 1,2 (ΣK ) ,

by
(T1 (f )) (h) := (T1 (f )) (y)h(y) dλn (y), h ∈ H 1,2 (ΣK ),
ΣK
for f ∈ L2|x|2 (Σ), where L2|x|2 (Σ) is equipped with the norm · H 1,2 (Σ)
,
|x|2 ,|x|3
which, by the BLT Theorem, extends uniquely to a linear bounded operator

1,2
T1 : H|x| 2 ,|x|3 (Σ) → H 1,2 (ΣK ) .
Next we define the transformations for the boundary inhomogeneity g and

the coefficients a and b. This means we want to find transformations T2 , T3 and
T4 such that
(T3 (a)) (x), ∇u(x) + (T4 (b)) (x)u(x) = (T2 (g)) (x), (5)
for all x ∈ ∂Σ , yields that
K
a(y), ∇ ((K(u)) (y)) + b(y)u(x) = g(y), (6)

for all y ∈ ∂ΣK . We start with the transformation T2 (g) of g.
Definition 16. Let Σ be an outer C 1,1 -domain and g be a function defined on ∂Σ.
Then we define a function T2 (g) on ∂ΣK by

x
(T2 (g)) (x) := g , x ∈ ∂ΣK .
|x|2
Again we use a Gelfand triple, namely
1 1
H 2 ,2 (∂Σ) ⊂ L2 (∂Σ) ⊂ H − 2 ,2 (∂Σ).
We have that
T2 : L2 (∂Σ) → L2 (∂ΣK ),
1 1
T2 : H 2 ,2 (∂Σ) → H 2 ,2 (∂ΣK ),
define linear, bounded isometries with (T2 )−1 = T2 . Moreover we define a contin-
uous linear operator
1
T2 : L2 (∂Σ) → H − 2 ,2 (∂ΣK ),
by
1
(T2 (g)) (h) := T2 (g)(y)h(y) dH n−1 (y), h ∈ H − 2 ,2 (∂Σ)
∂ΣK
for g ∈ L2 (∂Σ), where L2 (∂Σ) is equipped with the norm · − 21 ,2 . Hence
H (∂Σ)
again the BLT Theorem gives a unique continuous continuation
1 1
T2 : H − 2 ,2 (∂Σ) → H − 2 ,2 (∂ΣK ).
Closing this section, we give the definitions of the transformations T3 and T4 .
Definition 17. Let Σ be an outer C 1,1 -domain and a and b be defined on ∂Σ. We
define the operators T3 and T4 via
F G
x x
(T3 (a)) (x) := |x|n · a − 2 a , e x ex ,
|x|2 |x|2
F G
x x
(T4 (b)) (x) := |x|n−2 · b + (2 − n) a ,x ,
|x|2 |x|2
for all x ∈ ∂ΣK , where ex denotes the unit vector in direction x. Furthermore we
have
T3 : H 1,∞ (∂Σ) → H 1,∞ (∂ΣK ),
T4 : L∞ (∂Σ) → L∞ (∂ΣK ),
if Σ is an outer C 1,1 -domain and a ∈ H 1,∞ (∂Σ) for T4 . All operators are well
defined and give the relation formulated by equations (5) and (6).
These operators have the properties
T3 : H 1,∞ (∂Σ) → H 1,∞ (∂ΣK ),
T4 : L∞ (∂Σ) → L∞ (∂ΣK ),
if Σ is an outer C 1,1 -domain and a ∈ H 1,∞ (∂Σ) for T4 and
T3 : H 2,∞ (∂Σ) → H 2,∞ (∂ΣK ),
T4 : H 1,∞ (∂Σ) → H 1,∞ (∂ΣK ),
if Σ is an outer C 2,1 -domain and a ∈ H 2,∞ (∂Σ) for T4 .
4.2. Solution operator for the outer problem

In this section we want apply the solution operator of the inner regular problem
in order to get a weak solution of the outer problem. Therefore we will use a
combination of all the operators defined in the previous section. In order to avoid
confusion we denote the normal vector of ∂Σ by ν and the normal vector of ∂ΣK by
ν K . We start with the classical formulation of the outer oblique boundary problem
for the Poisson equation in the following definition.
Definition 18. Let Σ be an outer C 1,1 -domain, f ∈ C 0 (Σ), b, g ∈ C 0 (∂Σ) and
a ∈ C 0 (∂Σ; Rn ) be given. A function u ∈ C 2 (Σ) ∩ C 1 (Σ) such that
Δu(x) = f (x), for all x ∈ Σ,
a(x) · ∇u(x) + b · u(x) = g(x), for all x ∈ ∂Σ,
u(x) → 0, for |x| → ∞,
is called classical solution of the outer oblique boundary problem for the Poisson
equation.
Now we state the main result of this section which can be proved by the
results about the transformations above.
Theorem 19. Let Σ be an outer C 1,1 -domain, a ∈ H 1,∞ (∂Σ; Rn ), b ∈ L∞ (∂Σ),

1 1,2
g ∈ H − 2 ,2 (∂Σ) and f ∈ H|x| 2 ,|x|3 (Σ) , such that

(T3 (a)) (y), ν K (y) > C > 0, (7)

T4 (b) 1 T3 (a)
ess inf − div K − νK > 0, (8)
∂ΣK T3 (a), ν K 2 ∂Σ T3 (a), ν K
for all y ∈ ∂ΣK , where 0 < C < ∞. Then we define

out
u := Sa,b (f, g) := K STin3 (a),T4 (b) (T1 (f ), T2 (g)) ,
as the weak solution to the outer oblique boundary problem for the Poisson equation
out
from Definition 18. Sa,b is injective and we have for a constant 0 < C12 < ∞

u H 1,2 (Σ) ≤ C 12 f
H 1,2 (Σ)
+ g 1
− ,2 .
1 , 1
|x|2 ,|x|3
2 H (∂Σ)
|x|2 |x|
We are able to prove that the Kelvin transformation for functions is also a
continuous operator from H 2,2 (ΣK ) to H 2,2
1
, 1 ,1
(Σ). So we can prove the following
|x|2 |x|
regularization result, based on the regularization result for the inner problem,
see Theorem 6. The following theorem shows, that the weak solution, defined by
Theorem 19, is really related to the outer problem, given in Definition 18, although
it is not derived by an own weak formulation.
Theorem 20. Let Σ be an outer C 2,1 -domain, a ∈ H 2,∞ (∂Σ; Rn ), b ∈ H 1,∞ (∂Σ)
1
such that (7) and (8) holds. If f ∈ L2|x|2 (Σ) and g ∈ H 2 ,2 (Σ) then we have that u
provided by Theorem 19 is a strong solution, i.e., u ∈ H 2,2
1 1
, |x| ,1
(Σ), and
|x|2
Δu = f,
a, ∇u + bu = g,
almost everywhere on Σ and ∂Σ, respectively. Furthermore we have an a priori
estimate
u H 2,2 (Σ) ≤ C13 f L 2 (Σ) + g 1 ,2 ,
1 , 1 ,1 2 |x| 2 H (∂Σ)
|x|2 |x|
with a constant 0 < C13 < ∞.

As a consequence we have that if the data in Theorem 20 fulfills the require-
ments of a classical solution, the weak solution u provided by Theorem 19 coincides
with this classical solution. At the end of this section we investigate the conditions
on the oblique vector field. Analogously to the regular inner problem, we have
condition (8), which is a transformed version of (2) and gives a relation between a
and b, depending on the geometry of the surface ∂Σ. Moreover condition (7) is a
transformed version of (1) and gives the non-admissible direction for the oblique
vector field a. For the regular inner problem, (1) states the tangential directions
as non-admissible for the oblique vector field. For the outer problem the direction
depends as well on the direction of the normal vector ν(y) at the point y ∈ ∂Σ as
on the direction of y itself. In this section we will investigate this dependency in
detail. Using the definitions of T3 and T4 , we can rewrite condition (7) into the
equivalent form

cos ∠a(x),ν K ( x 2 ) − 2 · cos ∠a(x),ex · cos ∠ex ,ν K ( x 2 ) > C13 > 0,
|x| |x|
(9)
for all x ∈ ∂Σ and 0 < C13 < ∞ independent of x. We use the formula
y, z
=: cos(∠y,z ),
|y| · |z|
for vectors in Rn , where ∠y,z denotes the angle 0 ≤ ∠y,z ≤ π between y and z.
Going to R2 and setting
C14 (x) := cos(∠ex ,ν K (x) ),

C15 (x) := sin(∠ex ,ν K (x) ).
we can explicitly characterize the non admissible direction as

−1 C14 (x)
∠a(x),ex = tan ,
C15 (x)
if C15 (x) = 0 and ∠a(x),ex = π2 if C15 (x) = 0. Generally, transforming the problem
to an inner setting transforms the conditions for the coefficients a and b. There
are circumstances in which we have the same non-admissible direction as for the
inner problem, i.e., the tangential directions are non-admissible. For example, this
is the case if ∂Σ is a sphere around the origin. In Figure 4 the situation for Σ ⊂ R2
is illustrated, the dashed line indicates the non-admissible direction, which occurs
because of the transformed regularity condition T3 (a), ν K > C14 > 0, see (7).
Figure 4. Non-admissible direction for the outer problem

4.3. Ritz–Galerkin method

In this subsection we provide a Ritz–Galerkin method for the weak solution to the
outer problem. Therefore we use the approximation of the weak solution to the
corresponding inner problem, provided in Chapter 3. Assume Σ to be an outer
1
C 1,1 -domain. Furthermore let a ∈ H 1,∞ (∂Σ; Rn ), b ∈ L∞ (∂Σ), g ∈ H − 2 ,2 (∂Σ)

1,2
and f ∈ H|x| 2 ,|x|3 (Σ) , such that condition (7) and condition (8) is fulfilled. We
want to approximate the weak solution u to the outer oblique boundary problem,
provided by Theorem 19. Let a and F be defined by

n F G
T3 (a)i
a(η, v) := − η − ν K
, (∇ ∂Σ v) i
1 ,2
i=1 H 2 (∂Σ)
T3 (a), ν K i 1
H − 2 ,2 (∂Σ)

T4 (b)
− (∇η, ∇v) dλn − η v dH n−1
∂Σ T3 (a), ν
K
Σ
F G
T2 (g)
F (η) := 1 η, −H 1,2 (Σ) η, T1 (f )(H 1,2 (Σ))
H 2 ,2 (∂Σ) T3 (a), ν K H − 12 ,2 (∂Σ)
for η, v ∈ H 1,2 (ΣK ).

Furthermore, let (Vn )n∈N be an increasing sequence of finite-dimensional sub-
;
spaces of H 1,2 (ΣK ), i.e., Vn ⊂ Vn+1 such that n∈N Vn = H 1,2 (ΣK ). Then there
exists for each n ∈ N a unique vn ∈ Vn with:
a(η, vn ) = F (η) for all η ∈ Vn ,
see Section 3.4. Moreover, vn can be computed explicitly by solving a linear system
of equations. In Section 3.4 we have also seen that
n→∞
v − vn H 1,2 (Σ) ≤ C16 dist(v, Vn ) −→ 0.
So using the continuity of the operator K, see Theorem 13, we consequently get
the following result
Theorem 21. Let u be the weak solution provided by Theorem 19 to the outer
problem and v, (vn )n∈N taken from Theorem 7 and Theorem 5, both corresponding
to a, b, g, f and Σ, given at the beginning of this section. Then:
n→∞
u − K(vn ) H 1,2 (Σ) ≤ C17 dist(v, Vn ) −→ 0.
4.4. Stochastic extensions and examples

In this section we implement stochastic inhomogeneities as well as stochastic weak
solutions for the outer setting. Again we start by defining the spaces of stochastic
functions. So, let Σ be an outer C 1,1 -domain and (Ω, F , P) a probability space,
arbitrary but fixed, such that L2 (Ω, P ) is separable. We define

2,2
H 1 , 1 ,1 (Σ) 2 2,2 ∼ 2 2,2
:= L (Ω, P ) ⊗ H 1 , 1 ,1 (Σ) = L Ω, P ; H 1 , 1 ,1 (Σ) ,
|x|2 |x| |x|2 |x| |x|2 |x|
Ω

H 1,21 1 (Σ) := L 2
(Ω, P ) ⊗ H 1,2
1 1 (Σ)
∼
= L 2
Ω, P ; H 1,2
1 1 (Σ) ,
,
|x|2 |x|
,
|x|2 |x|
,
|x|2 |x|
Ω

L2|x|2 (Σ) := L2 (Ω, P ) ⊗ L2|x|2 (Σ) ∼
= L2 Ω, P ; L2|x|2 (Σ) ,
Ω

1,2 2 1,2 ∼ 2
H|x|2 ,|x|3 (Σ) := L (Ω, P ) ⊗ H|x|2 ,|x|3 (Σ) = L Ω, P ; H|x|2 ,|x|3 (Σ) 1,2
,
Ω
1

1
HΩ2 (∂Σ) := L2 (Ω, P ) ⊗ H 2 ,2 (∂Σ)
,2 ∼ 1
= L2 Ω, P ; H 2 ,2 (∂Σ) ,

L2Ω (∂Σ) := L2 (Ω, P ) ⊗ L2 (∂Σ) ∼
= L2 Ω, P ; L2 (∂Σ) ,

− 1 ,2 1
∼ 1
HΩ 2 (∂Σ) := L2 (Ω, P ) ⊗ H − 2 ,2 (∂Σ) = L2 Ω, P ; H − 2 ,2 (∂Σ) .
Because all spaces above are separable, we can again use the isomorphisms to
Hilbert space-valued random variables. Thus we can prove the following main
result of this section by defining the stochastic solution operator pointwisely.
Theorem 22. Let Σ be an outer C 1,1 -domain, a ∈ H 1,∞ (∂Σ; Rn ), b ∈ L∞ (∂Σ),

− 1 ,2 1,2
g ∈ HΩ 2 (∂Σ) and f ∈ H|x| 2 ,|x|3 (Σ) , such that (7) and (8) holds. Then we
Ω
define
out
u( · , ω) := Sa,b (f ( · , ω), g( · , ω)),
for dP -almost all ω ∈ Ω. u is called stochastic weak solution to the outer oblique
boundary problem for the Poisson equation. Furthermore we have for a constant
0 < C18 < ∞

u 1,2 ≤ C18 f 1,2
(Σ)
+ g
− 1 ,2 .
H 1 , 1 (Σ) H
|x|2 ,|x|3
2
HΩ (∂Σ)
Ω Ω
|x|2 |x|
Moreover, we have the following result for a stochastic strong solution.

Theorem 23. Let Σ be an outer C 2,1-domain, a ∈ H 2,∞ (∂Σ; Rn ), b ∈ H 1,∞ (∂Σ)
1
,2
such that (7) and (8) holds. If f ∈ L2|x|2 (Σ) and g ∈ HΩ2 (Σ) then we have
Ω
u ∈ H 2,2
1
, 1 ,1
(Σ) , for u provided by Theorem 22, and
|x|2 |x| Ω
Δu(x, ω) = f (x, ω),

a(y), ∇u(y, ω) + b(y)u(y, ω) = g(y, ω),
for λ -almost all x ∈ Σ, for H n−1 -almost all y ∈ ∂Σ and for dP -almost all ω ∈ Ω.
n
Furthermore, we have an a priori estimate

u 2,2 ≤ C19 f
2 (Σ)
+ g 1
,2 ,
1 , 1 ,1 (Σ) Ω
H L 2 HΩ (∂Σ)
|x|2 Ω
|x|2 |x|
with a constant 0 < C19 < ∞.

Again a Ritz–Galerkin method is also available also for the stochastic weak
solution. It is left to the reader to write down the details. As mentioned we close
the section with examples for stochastic data. These are used in geomathematical
applications in order to model noise on measured values. In the following we give
the examples for the outer problem. They are also suitable for the inner problem.
4.4.1. Gaussian inhomogeneities. We choose the probability space (Ω, F, P) such

that Xi , 1 ≤ i ≤ n1 , are P ⊗ λn -measurable and Yj , 1 ≤ j ≤ n2 , are P ⊗ H n−1 -
measurable with Xi (·, x), x ∈ Σ, and Yj (·, x), x ∈ ∂Σ, Gaussian random variables
with expectation value 0 and variance fσ2i (x) or variance gσ2j (x), respectively. Here
fσi ∈ L2|x|2 (Σ) and gσj ∈ L2 (∂Σ). We define:

n1
n2
f (ω, x) := fμ (x) + Xi (ω, x), g(ω, x) := gμ (x) + Yj (ω, x),
i=1 j=1
where fμ ∈ L2|x|2 (Σ) and gμ ∈ L2 (∂Σ). To use such kind of inhomogeneities we

must show
f ∈ L2 (Ω × Σ, P ⊗ |x|4 · λn ) and g ∈ L2 (Ω × ∂Σ, P ⊗ H n−1 ).
It is easy to see that the inhomogeneities defined in this way fulfill these require-
ments and the main results are applicable. Such a Gaussian inhomogeneity is
shown in Figure 5.
Figure 5. Data with Gaussian noise
4.4.2. Gauß–Markov model. Here we refer to [7], in which an application of the

example from the previous paragraph can be found. The authors use a random
field
h(ω, x) := H(x) + Z(ω, x)
to model an observation noise, where x ∈ ∂B1 (0) ⊂ R3 and ω ∈ Ω with (Ω, F , P)

a probability space. Here we have that Z(·, x), x ∈ ∂B1 (0), is a Gaussian ran-
dom variable with expectation value 0 and variance σ 2 > 0. Additionally H(x) ∈
L2 (∂B1 (0)) and the covariance is given by:
cov(Z(·, x1 ), Z(·, x2 )) = K(x1 , x2 ),
where K : ∂B1 (0) × ∂B1 (0) → R is a suitable kernel.
Two geophysically relevant kernels are for example
σ2
M
2n + 1
K1 (x1 , x2 ) := 2
Pn ((x1 · x2 )) 0 ≤ M < ∞,
(M + 1) n=1 4π
σ2
K2 (x1 , x2 ) := exp(−c(x1 · x2 )).
exp(−c)
Pn , 1 ≤ n ≤ M , are the Legendre polynomials defined on R. The noise model
corresponding to the second kernel is called first degree Gauß–Markov model. If
one chooses a P ⊗ H n−1 -measurable random field Z, then h fulfills the require-
ments. Existence of a corresponding probability measure P is provided in infinite-
dimensional Gaussian Analysis, see, e.g., [4].
4.4.3. Noise model for satellite data. In this paragraph we give another precise
application, which can be found in [3]. Here the authors are using stochastic inho-
mogeneities to implement a noise model for satellite data. Therefore random fields
of the form
m
h(ω, x) := hi (x)Zi (ω)
i=1
3
are used, where x ∈ ∂Σ ⊂ R and ω ∈ Ω with (Ω, F , P) a suitable probability
space. Here ∂Σ could be for example the Earth’s surface and we are searching for
harmonic functions in the space outside the Earth. Zi are Gaussian random vari-
ables with expectation value 0 and variance σi2 > 0 and hi fulfilling the assumptions
0,σi
of Paragraph 4.4.1. If one chooses (Ω, F , P) as (Rm , B(R), γcov ij
), where:
0,σi 1 1 −1
γcov ij
:= # e− 2 (y,A y) dλm ,
m
(2π) det(A)
aij := cov(Zi , Zj ), 1 ≤ i, j ≤ m,
one has a realization of Zi as the projection on the ith component in the separable
space L2 (Rm , γcov
0,σi
ij
).
5. Future directions
In this chapter we want to point out one direction of further investigations. We have
seen how to provide the existence of a weak solution to the outer oblique boundary
problem for the Poisson equation. Therefore we introduce several transformations.
In Theorem 15, we proved for the transformation of the space inhomogeneity f

1,2
T1 : H|x| 2 ,|x|3 (Σ) → H 1,2 (ΣK ) .

1,2
This transformation is not bijective, i.e., T1 H|x| 2 ,|x|3 (Σ) = H 1,2 (ΣK ) ,

Finding a Hilbert space V , such that T1 : V → H 1,2 (ΣK ) is bijective would

lead to the existence of a weak solution for a even larger class of inhomogeneities.
Moreover we have for the transformation K of the weak solution to the inner
problem

K : H 1,2 (ΣK ) → H 1,2
1
, 1
(Σ), where again K H 1,2 (ΣK ) = H 1,2 1
, 1
(Σ),
|x|2 |x| |x|2 |x|
1,2
see Theorem 13. Finding a Hilbert space W such that K : H (Σ ) → W is K
bijective, would give us uniqueness of the solution and more detailed information
about the behavior of u and its weak derivatives, when x is tending to infinity.
Additionally, we would be able to define a bijective solution operator for the outer
problem. This could be used to find the right Hilbert spaces, such that a Poincaré
inequality is available. Consequently the Lax–Milgram Lemma would be applicable
directly to a weak formulation for the outer setting, which can be derived similar
to the inner problem. Then we might have to consider a regular outer problem,
because the tangential direction is forbidden for the oblique vector field, if we want
to derive a weak formulation. In turn we get rid of the transformed regularity
condition on a. The results presented in this report are then still an alternative in
order to get weak solutions for tangential a. Moreover, the availability of a Poincaré
inequality would lead to existence results for weak solutions to a broader class of
second-order elliptic partial differential operators in outer domains. See, e.g., [2]
for such second-order elliptic partial differential operators for inner domains.
Instead of using the Ritz–Galerkin approximation, it is also possible to ap-
proximate solutions to oblique boundary-value problems for harmonic functions
with the help of geomathematical function systems, e.g., spherical harmonics. For
such an approach, see, e.g., [8].
6. Conclusion
The analysis of inner oblique boundary value problems is rather good understood
and we reached the limit when searching for weak solutions under as weak as-
sumptions as possible. The outer problem causes still problems because of the
unboundedness of the domain. As mentioned in Section 5, finding the right dis-
tribution spaces such that a Poincaré inequality holds, might lead to bijective
solution operators for an even broader class of inhomogeneities. Nevertheless, we
are already able to provide weak solutions to the outer problem as presented in
the previous sections for very general inhomogeneities. Also stochastic weak solu-
tions for stochastic inhomogeneities as used in geomathematical applications can
be provided and approximation methods for the weak solutions are available.
References
[1] Adams, R.A. (1975) Sobolev Spaces. Academic Press, New York-San Francisco-
London
[2] Alt, H.W. (2002) Lineare Funktionalanalysis. Springer, Berlin
[3] Bauer, F. (2004) An Alternative Approach to the Oblique Derivative Problem in
Potential Theory. Shaker Verlag, Aachen
[4] Berezanskij Y.M. (1995) Spectral Methods in Infinite Dimensional Analysis. Kluwer
Academic Publishers, Dordrecht
[5] Dautray, R., Lions, J.-L. (1988) Mathematical Analysis and Numerical Methods for
Science and Technology, Volume 2: Functional and variational methods. Springer,
Berlin
[6] Dobrowolski, M. (2006) Angewandte Funktionalanalysis. Springer, Berlin
[7] Freeden, W., Maier, T. (2002) On Multiscale Denoising of Spherical Functions: Basic
Theory and Numerical Aspects. Electron. Trans. Numer. Anal. 14: 56–78
[8] Freeden, W., Michel, V. (2004) Multiscale Potential Theory (with Applications to
Geoscience). Birkhäuser Verlag, Boston, Basel, Berlin
[9] Grothaus, M., Raskop, T. (2006) On the Oblique Boundary Problem with a Sto-
chastic Inhomogeneity. Stochastics 78(4): 233–257
[10] Grothaus, M., Raskop, T. (2009) The Outer Oblique Boundary Problem of Potential
Theory. Num. Funct. Anal. Optim. 30(7-8): 1–40
[11] Gilbarg, D., Trudinger, N.S. (1998) Elliptic Partial Differential Equations of Second
Order. Springer, Berlin
[12] Gutting, M. (2008) Fast Multipole Methods for Oblique Derivative Problems. Shaker
Verlag, Aachen
[13] Miranda, C. (1970) Partial Differential Equations of Elliptic Type. Springer, Berlin
[14] Reed. M., Simon, B. (1972) Methods of Modern Mathematical Physics, Volume 1:
Functional Analysis. Academic Press, New York-San Francisco-London
[15] Rozanov, Y., Sanso, F. (2001) The Analysis of the Neumann and Oblique Derivative
Problem: The Theory of Regularization and its Stochastic Version. J. Geod. 75(7-8):
391–398
[16] Rozanov, Y., Sanso, F. (2002a) On the Stochastic Versions of Neumann and Oblique
Derivative Problems. Stochastics Rep. 74(1-2): 371–391
[17] Rozanov, Y., Sanso, F. (2002b) The Analysis of the Neumann and Oblique Derivative
Problem: Weak Theory. In Geodesy: Challenge of the 3rd Millenium. Springer, Berlin
Martin Grothaus and Thomas Raskop

Functional Analysis and Stochastic Analysis Group
Fachbereich Mathematik
e-mail: grothaus@mathematik.uni-kl.de

About the Importance of the Runge–Walsh

Concept for Gravitational Field Determination
Matthias Augustin, Willi Freeden, and Helga Nutz
Abstract. On the one hand, the Runge–Walsh theorem plays a particular role
in physical geodesy, because it allows to guarantee a uniform approximation
of the Earth’s gravitational potential within arbitrary accuracy by a harmonic
function showing a larger analyticity domain. On the other hand, there are
some less transparent manifestations of the Runge–Walsh context in the ge-
odetic literature that must be clarified in more detail. Indeed, some authors
make the attempt to apply the Runge–Walsh idea to the gravity potential
of a rotating Earth instead of the gravitational potential in non-rotating sta-
tus. Others doubt about the convergence of series expansions approximating
the Earth’s gravitational potential inside the whole outer space of the actual
Earth.
The goal of this contribution is to provide the conceptual setup of the
Runge–Walsh theorem such that geodetic expectation as well as mathematical
justification become transparent and coincident. Even more, the Runge–Walsh
concept in form of generalized Fourier expansions corresponding to certain
harmonic trial functions (e.g., mono- and/or multi-poles) will be extended
to the topology of Sobolev-like reproducing kernel Hilbert spaces thereby
avoiding any need of (numerical) integration in the occurring spline solution
process.
Keywords. Runge–Walsh theorem in physical geodesy, theoretical background,

constructive solution concepts, (discrete) boundary value problems, general-
ized Fourier series, spline interpolation.
1. Introduction
In the theory of harmonic functions, a result first motivated by C. Runge [62] in
one-dimensional complex analysis and later generalized, e.g., by J.L. Walsh [70],
I.N. Vekua [69], and L. Hörmander [40] to potential theory in three-dimensional
Euclidean space R3 is of basic interest. For geodetically relevant obligations (see,
518 M. Augustin, W. Freeden, and H. Nutz
e.g., [33, 44, 54, 56, 63], and the references therein) it may be formulated in ac-
cordance with [55]:
Geodetic version of the Runge–Walsh theorem. Let G ⊂ R3 be a regular region, i.e.,
a bounded region G ⊂ R3 dividing R3 uniquely into the inner space G and the outer
space G c = R3 \G, G = G ∪ ∂G, such that G contains the origin and the boundary
∂G is an orientable smooth Lipschitzian manifold of dimension 2. Any harmonic
function in G c that is regular at infinity can be approximated by a function that
is harmonic outside an arbitrarily given Runge (i.e., in geodetic nomenclature
sometimes called Bjerhammar) ball A G, i.e., A ⊂ G with dist(A, ∂G) > 0 (see
Figure 1, right illustration) in the sense that, for any given ε > 0, the absolute
error between the two functions is smaller than ε for all points outside and on
any closed surface completely surrounding ∂G in its outer space. The value ε may
be arbitrarily small, and the surrounding surface may be arbitrarily close to the
surface ∂G.
Figure 1. The geometric situation of the Runge–Walsh theorem (with

A an arbitrary regular region such that A G (left) and A an inner
Runge (i.e., Bjerhammar) ball (right)).
Obviously, the Runge–Walsh theorem in the preceding formulation (with G,

e.g., chosen as the interior of the actual Earth) represents a pure existence theorem.
It guarantees only the existence of an approximating function and does not provide
a constructive method to find it. Nothing is said about the approximation proce-
dure and the computational structure and methodology of the approximation. The
theorem merely describes the theoretical background for the approximation of a
potential by another potential defined on a larger harmonicity domain, i.e., the
Runge region outside the sphere ∂A.
The situation, however, is completely different if spherical geometries are ex-
clusively involved in the Runge concept. Assuming that both A, G are concentric
balls around the origin with A G, a constructive approximation of a potential
in the outer space G c is available, e.g., by outer harmonic (orthogonal) expansions
(see, e.g., [30, 41, 47, 48, 60, 71]). More concretely, within the classical context
of a twofold spherical configuration, a constructive version of the Runge–Walsh
theorem can be guaranteed by finite truncations of Fourier expansions in terms of
outer harmonics, where the L2 (∂G)-convergence of the Fourier series implies uni-
About the Importance of the Runge–Walsh Concept 519
form convergence on any point set K G c . The Fourier coefficients are obtained
by integration over the sphere ∂G. The gravitational potential is available (in spec-
tral sense) by tables of the Fourier coefficients. Nowadays, in fact, outer harmonic
expansions constitute the conventional geodetic tools in globally reflected approx-
imation of the Earth’s gravitational potential and its observables.
From a superficial point of view, one could suggest that the standard ap-
proximation by truncated series expansions in terms of outer harmonics is closely
related to spherical geometries ∂A, ∂G. The purpose of our work, however, is to
show that the essential steps to a constructive Fourier approach can be extended
to any regular, i.e., not-necessarily spherical region G and to any regular, i.e., not-
necessarily spherical Runge region A G (see Figure 1, left illustration). As a
matter of fact, the Runge–Walsh approach enables us to avoid any calamities with
the convergence to the gravitational potential by the generalized Fourier series for
arbitrary sets K G c . In analogy to the spherical case, however, it likewise does
not help to specify convergence inside Ac \G c , so that any attempts (see [3]) to
reduce gravitational information via infinite Fourier series downward from ∂G to
the surface ∂A are not justifiable by the Runge–Walsh framework.
In summary, the Runge–Walsh concept as presented in this work reflects con-
structive approximation capabilities of the Earth’s gravitational (and not gravity)
potential even if geoscientifically realistic (i.e., not necessarily spherical) geometries
come into play.
Mathematically, it should be pointed out that the main techniques for assur-
ing the not-necessarily spherical results are the limit and jump relations and their
formulations of potential theory in the Hilbert space nomenclature of (L2 (∂G),
· L2 (∂G) ). The special function systems for use in constructive Runge–Walsh
theorems are manifold. Moreover, all harmonic functions systems that are regu-
lar at infinity can be taken into account, whose restrictions to the boundary ∂G
of a regular region G form an L2 (∂G)-complete system. For numerical efficiency,
however, we restrict ourselves to certain mono-pole and multi-pole configurations.
2. Special function systems

Spherical harmonics are the functions most commonly used in geosciences to repre-
sent scalar fields on the unit sphere Ω ⊂ R3 . They are used extensively in the gravi-
tational and magnetic applications involving Laplace’s equation. The introduction
of (scalar) spherical harmonics and the derivation of some important properties
can be found, e.g., in [16, 26, 59], and the references therein.
2.1. Spherical harmonics

Let Hn : R3 → R be a homogeneous and harmonic polynomial of degree n ∈ N0 ,
i.e., Hn (λx) = λn Hn (x), λ ∈ R, and ΔHn (x) = 0, x ∈ R3 . Then, the restriction
Yn = Hn |Ω of Hn to the unit sphere Ω in R3 is called a (scalar) spherical harmonic
(of degree n). The space of all spherical harmonics of degree n is denoted by
Harm n (Ω). The spherical harmonics of degree n form a space of dimension 2n + 1,

i.e., dim(Harm n (Ω)) = 2n+1, n ∈ N0 . Using the standard method of separation by
spherical coordinates and observing the homogeneity,
d 2we have Hn (x) = rn Yn (ξ),
for r = |x|, ξ = |x| x
∈ Ω. From the identity r12 dr d
r dr rn = n(n + 1)rn−2 it
follows, in connection with the harmonicity of Hn , that 0 = ΔHn (x) = rn−2 n(n +
1)Yn (ξ)+rn−2 Δ∗ξ Yn (ξ). As a consequence, we can state that any spherical harmonic
Yn ∈ Harm n (Ω), n ∈ N0 , is an infinitely often differentiable eigenfunction of the
Beltrami operator corresponding to the eigenvalue −n(n + 1). More explicitly,
Δ∗ Yn (ξ) = −n(n + 1)Yn (ξ), ξ ∈ Ω. Conversely, every infinitely often differentiable
eigenfunction of the Beltrami operator with respect to the eigenvalue −n(n +
1) constitutes a spherical harmonic of degree n. Using Green’s formulas for the
Beltrami operator, this implies that spherical harmonics of different degrees are
orthogonal with respect to the L2 (Ω)-inner product, i.e.,

(Yn , Ym )L2 (Ω) = Yn (ξ)Ym (ξ) dS(ξ) = 0, n = m, (1)
Ω
where dS is the surface element in R3 . The Gram–Schmidt method allows the
orthonormalization of any set of linearly independent spherical harmonics of de-
gree n with respect to the L2 (Ω)-inner product. Throughout this work, a set
{Yn,k }k=1,...,2n+1 ⊂ Harm n (Ω) always denotes an orthonormal basis of Harm n (Ω).
The Legendre polynomials are one-dimensional orthogonal polynomials that
are of great importance when treating spherical harmonics. A polynomial Pn :
[−1, 1] → R of degree n ∈ N0 is called Legendre polynomial (of degree n) if
1
• −1 Pn (t)Pm (t)dt = 0, n = m,
• Pn (1) = 1.
The Legendre polynomials are uniquely determined by these properties. They have
the explicit representation
[n/2]
(2n − 2s)!
Pn (t) = (−1)s tn−2s , t ∈ [−1, +1], (2)
s=0
2n (n − 2s)!(n − s)!s!
where we use the abbreviation
.n/ 1 1

= n − (1 − (−1) ) .
n
(3)
2 2 2
The zonal function η → Pn (ξ · η), η ∈ Ω, is the only normalized (i.e., Pn (ξ · ξ) =
1) spherical harmonic of degree n that is invariant with respect to orthogonal
transformations which leave ξ ∈ Ω fixed. A consequence of this fact is the addition
theorem (see, e.g., [58] and [7] for different approaches) which states the close
relation of (univariate) Legendre polynomials to spherical harmonics: For n ∈ N0
and ξ, η ∈ Ω, we have
2n+1
2n + 1
Yn,k (ξ)Yn,k (η) = Pn (ξ · η). (4)
4π
k=1
As a direct consequence, estimates for Legendre polynomials and spherical har-

monics are derivable by standard arguments. For n ∈ N0 and k = 1, . . . , 2n + 1,
we have |Pn (t)| ≤ 1, t ∈ [−1, 1]. It is well-known that
∞
1
Pn (t)hn = √ , t ∈ [−1, 1], h ∈ (−1, 1). (5)
1 + h 2 − 2ht
n=0
2.2. Mono- and multi-poles

Of special importance for our considerations is the so-called fundamental solution
for the Laplace operator
1 1
G(Δ; |x − y|) = , x, y ∈ R3 , x = y. (6)
4π |x − y|
This function can be regarded as a mono-pole at y, it represents the gravitational
potential between a mass point at y and a point x in the exterior (of the mass).
An easy manipulation involving (5) yields
2 − 12
1 1 |x| |x|
= 1+ −2 ξ·η , (7)
|x − y| |y| |y| |y|
|x|
where x, y ∈ R3 , |x| < |y|, and ξ = x
|x| , η= y
|y| . With t = ξ · η and h = |y| , this
implies the series expansion
∞ n
1 1 |x|
= Pn (ξ · η). (8)
|x − y| |y| n=0 |y|
Moreover, further calculations show that
∞

1 (−1)n n 1
= |x| (ξ · ∇y )n , (9)
|x − y| n=0 n! |y|
where
(−1)n 1 Pn (ξ · η)
(ξ · ∇y )n = , n ∈ N0 . (10)
n! |y| |y|n+1
The identity (10) is known as Maxwell’s representation formula. As y → |y|−1 ,
y = 0, is (apart from a multiplicative constant) the fundamental solution for
the Laplace operator, this representation tells us that the Legendre polynomials
may be obtained by repeated differentiation of the fundamental solution in the
radial direction of ξ. The potential on the right-hand side may be regarded as the
potential of a multi-pole of order n with the axis ξ at the origin.
From any textbook about spherical harmonics (see, e.g., [16, 26]) we know
that the system {Yn,k } n∈N0 ; is closed in C (0) (Ω). That is, for any given ε > 0
k=1,...,2n+1
and each F ∈ C (0) (Ω), there exist coefficients an,k ∈ R such that
- -
- N 2n+1
-
- -
-F − an,k Yn,k - ≤ ε. (11)
- -
n=0 k=1 C (0) (Ω)
This result also enables us to verify the closure in the Hilbert space L2 (Ω) with
respect to the norm · L2 (Ω) . That is, for any given ε > 0 and each F ∈ L2 (Ω),
there exist coefficients bn,k ∈ R such that
- -
- N 2n+1
-
- -
-F − bn,k Yn,k - ≤ ε. (12)
- -
n=0 k=1 L2 (Ω)
Note that, in a Hilbert space, the closure property of a function system is equivalent
to the completeness property (see, e.g., [6]). The completeness means: If F of class
L2 (Ω) has vanishing Fourier coefficients

F ∧ (n, k) = F (η)Yn,k (η)dω(η) = 0, (13)
Ω
for all n ∈ N0 , k = 1, . . . , 2n + 1, then F = 0 in L2 (Ω)-sense. In other words,

F is uniquely determined by its Fourier coefficients. Furthermore, we can state a
constructive version of (12) in terms of the Fourier expansion, i.e., for all F of class
L2 (Ω),
- -
- N 2n+1
-
- ∧ -
lim -F − F (n, k)Yn,k - = 0. (14)
N →∞ - -
n=0 k=1 L2 (Ω)
Next we consider a sphere ΩR around the origin with radius R > 0. By virtue of
the isomorphism ξ → Rξ, ξ ∈ Ω, we can assume a function F : ΩR → R to be
reduced to the unit sphere Ω. Obviously, an L2 (Ω)-orthonormal system of spherical
harmonics forms an L2 (ΩR )-orthogonal system. More explicitly,

y y
(Yn,k , Yp,q )L2 (ΩR ) = Yn,k Yp,q dS(y) = R2 δn,p δk,q . (15)
ΩR |y| |y|

Introducing the system Yn,k R
(x) = R1 Yn,k |x|
x
, x ∈ ΩR , we get an orthonormal
basis {Yn,k
R
}n∈N0 ; k=1,...,2n+1 of the space L2 (ΩR ).
(a) The functions

n
|x|
R
Hn,k (x) = R
Yn,k (x) , x ∈ R3 , (16)
R
for n ∈ N0 , k = 1, . . . , 2n + 1, are called inner harmonics (of degree n and
order k).
(b) The functions
n+1
R
R
H−n−1,k (x) = R
Yn,k (x) , x ∈ R3 \{0}, (17)
|x|
for n ∈ N0 , k = 1, . . . , 2n + 1, are called outer harmonics (of degree n and
order k).
We let BR (0) = {x ∈ R3 , |x| < R} be the ball around the origin with radius R.
Then it is not difficult to see that the inner harmonics are of class Pot (BR (0)),
i.e.,
R
(i) Hn,k is a member of C (2) (BR (0)),
R R
(ii) Hn,k satisfies ΔHn,k (x) = 0, x ∈ BR (0).
Furthermore, the inner harmonics show the following “boundary behavior” on
ΩR = ∂BR (0):
R
R
(iii) Hn,k = Yn,k ,
ΩR
R R
(iv) Hn,k , Hp,q = δn,p δk,q .
L2 (ΩR )
Analogously, the outer harmonics represent those functions that are harmonic in
the exterior of ΩR and regular at infinity, and which coincide with the spheri-
R
cal harmonics Yn,k on the boundary ΩR . In shorthand nomenclature, the outer
harmonics are of class Pot (R3 \BR (0)), i.e.,
R
(i) H−n−1,k is a member of C (2) (R3 \BR (0)),
R
(ii) H−n−1,k R
satisfies ΔH−n−1,k (x) = 0, x ∈ R3 \BR (0),
R R
(iii) H−n−1 is regular at infinity, i.e., H−n−1,k (x) = O(|x|−1 ), |x| → ∞.
Furthermore, the outer harmonics show the following “boundary behavior” on
ΩR = ∂BR (0):

R
(iv) H−n−1,k =YR ,
ΩR n,k
R R
(v) H−n−1,k , H−p−1,q 2 = δn,p δk,q .
L (ΩR )
R
Moreover, it should be noted that an inner harmonic Hn,k is related to its corre-
R
sponding outer harmonic H−n−1,k in the following way:
2n+1
R R R R2
R
H−n−1,k (x) = R
Hn,k (x) = H x , x ∈ R3 \ {0}. (18)
|x| |x| n,k |x|2
This observation leads us to the following interpretation: The mapping

2
R
x → x̌R = x, x ∈ BR (0)\{0}, (19)
|x|
transforms BR (0)\{0} into R3 \BR (0) and ΩR = ∂BR (0) onto itself. Referring to
Figure 2, we observe that the two triangles with edges (x̌R , y, 0) and (x, y, 0) are
similar whenever y ∈ ΩR . Furthermore, the ratios |x| |y|
|y| and |x̌R | are equal, provided
that y ∈ ΩR .
On the one hand, for x = |x|ξ, ξ ∈ Ω, and y = |y|η, R = |y|, η ∈ Ω, we have
|x − y|2 = x2 + y 2 − 2x · y = |x|2 + R2 − 2|x|R ξ · η. (20)

Figure 2. The inversion x → x̌R with respect to the sphere ΩR .
On the other hand, we see that

2 2 2
|x| R
|x|2 R4 2 2 R2
R |x|2 x − y = R2 |x|4 |x| + R − 2 |x|2 x · y
= |x|2 + R2 − 2|x|R ξ · η. (21)
For all y ∈ ΩR and x ∈ BR (0), we have

2
|x|
|x − y| = |x̌R − y|. (22)
R
After these preparations about the inversion of points with respect to a sphere
ΩR , R > 0, we are able to discuss the Kelvin transform (see, e.g., [41]): Assume
that U is of class C (2) (G), G ⊂ R3 \{0} open. Let ǦR be the image of G under the
inversion x → x̌R = R2 |x|−2 x. Denote by Ǔ = KR [U ]: ǦR → R, with

2
R R
Ǔ (x) = K [U ](x) =
R
U x , (23)
|x| |x|
the Kelvin transform of U with respect to ΩR = ∂BR (0). Then

5
2
R R
ΔǓ (x) = ΔU x . (24)
|x| |x|
As a consequence, the outer harmonics are obtainable by the Kelvin transform KR

from their inner counterparts, and vice versa. More precisely, we have
0 R 1
R
H−n−1,k (x) = KR Hn,k (x), (25)
0 1
R
Hn,k (x) = KR H−n−1,k
R
(x), (26)
for x ∈ R3 \ {0}.
3. Runge–Walsh closure theorems

We begin our considerations with a result on the special function system of outer
harmonics (see [8, 15, 23]).
Lemma 3.1 (Linear Independence). Let G ⊂ R3 be a regular region such that
! R "
R < inf x∈∂G |x|, i.e., BR (0) G. Then the sequence H−n−1,j n∈N ; j=1,...,2n+1 is
0
a subsystem of Pot (R3 \BR (0)), and its restriction to ∂G

R
H−n−1,j (27)
∂G n∈N0 ;
j=1,...,2n+1
forms a linearly independent system.
Proof. In order to verify the statement under the assumption BR (0) G we have
to derive that, for any linear combination H of the form
N 2n+1
R
H= an,j H−n−1,j , (28)
n=0 j=1
the condition H|∂G = 0 implies a0,1 = · · · = aN,1 = · · · = aN,2N +1 = 0. Indeed,

from the uniqueness theorem of the exterior Dirichlet problem (see, e.g., [15]) we
know that H|∂G = 0 yields H|G c = 0. Therefore, for every sphere with radius
R > supx∈∂G |x| around the origin 0, it follows that

R
H−n−1,j (x)H(x) dS(x) = 0 (29)
ΩR
for n = 0, . . . , N, j = 1, . . . , 2n + 1. Inserting (28) into (29) yields, in connection

with the well-known completeness property of the spherical harmonics (see, e.g.,
[16]), that an,j = 0 for all n = 0, . . . , N, j = 1, . . . , 2n + 1, as required.
3.1. L2 -closure and truncated Fourier series expansions
Next, our purpose is to prove completeness and closure theorems (see [8]).
Theorem 3.2 (Completeness). Let G ⊂ R3 be a regular region such that BR (0) G.
! R "
Then the restriction of H−n−1,j n∈N0 ,j=1,...,2n+1 to the surface ∂G

R
H−n−1,j (30)
∂G n∈N0 ;
j=1,...,2n+1
is complete in L2 (∂G).
Proof. Suppose that F ∈ L2 (∂G) satisfies

R R
(F, H−n−1,j )L2 (∂G) = F (y)H−n−1,j (y) dS(y)
∂G ∂G
= 0, (31)
2
n ∈ N0 , j = 1, . . . , 2n + 1. We have to show that F = 0 in L (∂G).
We know that the fundamental solution G(Δ; | · −y|) = (4π| · −y|)−1 of the
Laplace operator given in terms of its spherical harmonic series expansion (see,
e.g., [15])
∞
2n+1
1 1 1 |x|n
G(Δ; |x − y|) = = Yn,j (ξ)Yn,j (η), (32)
4π |x − y| n=0 2n + 1 |y|n+1 j=1
x = |x|ξ, y = |y|η, is analytic in the variable x on the ball BR (0) around the origin
0 with radius R, if y is a member of R3 \ BR (0). For all x ∈ BR (0), we thus find
by virtue of (31)

P (x) = F (y)G(Δ; |x − y|) dS(y)
∂G
∞
2n+1
R
R R
= Hn,j (x) F (y)H−n−1,j (y) dS(y) = 0. (33)
n=0
2n + 1 j=1 ∂G
Analytic continuation shows that the single-layer potential P vanishes in G. In

other words, the equations
P (x − τ ν(x)) = 0, (34)
∂P
(x − τ ν(x)) = 0 (35)
∂ν
hold true for all x ∈ ∂G and all sufficiently small τ > 0, where ν(x) is the outer
unit normal at the point x. Therefore, using the L2 -limit and jump relations (see
[8]), we obtain
2

lim P (x + τ ν(x)) dS(x) = 0, (36)
τ →0+ ∂G
2
∂P
lim
τ →0+ ∂ν (x + τ ν(x)) + F (x) dS(x) = 0, (37)
∂G
and ∂P 2
1
lim (x) + F (x) dS(x) = 0. (38)
τ →0+ ∂G ∂ν 2
The limit in the last equation can be omitted. Hence, the identity (38) can
also be understood as

∂ 1
F (y) G(Δ; |x − y|) dS(y) = − F (x), (39)
∂G ∂ν(x) 2
in the sense of L2 (∂G). The left-hand side of (39) constitutes a continuous function
(see, e.g., [15]). Thus, the function F is continuous itself. For continuous functions,
however, the classical limit and jump relations are valid:
lim P (x + τ ν(x)) = 0, x ∈ ∂G, (40)
τ →0+
∂P
lim (x + τ ν(x)) = −F (x), x ∈ ∂G. (41)
τ →0+ ∂ν
Consequently, the uniqueness theorem of the exterior Dirichlet problem (see, e.g.,
[41]) shows us that P (x) = 0 for all x ∈ R3 \G c . But this means that F = 0 on the
surface ∂G, as required.
From approximation theory (see, e.g., [6]) we know that the properties of
completeness and closure are equivalent in a Hilbert space such as L2 (∂G). The
equivalence leads us to the following statement
Corollary 3.3 (Closure). Let G ⊂ R3 be a regular region such that BR (0) G.
! R "
Then the system H−n−1,j ∂G n∈N0 ,j=1,...,2n+1 is closed in L2 (∂G), i.e., for any
given F ∈ L2 (∂G) and arbitrary ε > 0 there exist coefficients an,j , n = 0, . . . , N ,
j = 1, . . . , 2n + 1, constituting the linear combination
N 2n+1

R
FN = an,j H−n−1,j (42)
∂G
n=0 j=1
such that
F − FN L2 (∂G) ≤ ε . (43)
Based on our results on outer harmonics, i.e., multi-pole expansions, a large
variety of countable systems of potentials can be shown to possess the L2 -closure
property on ∂G. Probably best known are mono-poles, (i.e., fundamental solutions
of the Laplace operator). Their L2 (∂G)-closure can be adequately described by
using the concept of fundamental systems, which should be recapitulated briefly
(see, e.g., [8, 15, 23]).
Definition 3.4 (Fundamental System). Let A, G ⊂ R3 be regular regions satisfying
the “Runge condition” A G (cf. Figure 3). A point set Y = {yn }n=0,1,... ⊂ A
(with yn = yl for n = l) is called a fundamental system in G, if for each Q ∈ Pot (G),
i.e., for each Q ∈ C (2) (G) with ΔQ = 0 in G, the condition Q(yn ) = 0 for all
n = 0, 1, . . . implies Q = 0 in G.
Figure 3. Illustration of the positioning of a fundamental system on

∂A in G.
Two examples of fundamental systems in G should be listed: Y = {yn }n=0,1,... ⊂ A

is a fundamental system in G if
(1) Y is a dense subset of points in A G,

(2) Y is a dense subset of points on the boundary ∂A with A G (cf. Figure 3).
Theorem 3.5. Let G be a regular region. Then the following statement is valid: For
every fundamental system Y = {yn }n=0,1,... ⊂ A in G, A G, the system

G(Δ; | · −yn |) (44)
∂G n=0,1,...
2
is closed in L (∂G).
Proof. Since yn = ym for all n = m, it immediately follows that the system

{G(Δ; | · −yn |)}n=0,1,... is linearly independent.
Our purpose is to verify the completeness of the system (44) in L2 (∂G). To
this end, we consider a function F ∈ L2 (∂G) with

F (x)G(Δ; |x − yn |) dS(x) = 0, n ∈ N0 . (45)
∂G
We have to prove that F = 0 in L2 (∂G). We consider the single-layer potential Q

given by
Q(y) = F (x)G(Δ; |x − y|) dS(x). (46)
∂G
Since Q is harmonic in G, the properties of the fundamental system {yn }n=0,1,...
in G imply that Q(y) = 0 for all y ∈ G. Then, the same arguments as given in the
proof of Theorem 3.2 guarantee that F = 0 in the sense of L2 (∂G), as desired.
Besides the outer harmonics, i.e., multi-poles (see Corollary 3.3) and the
mass (single-)poles (see Theorem 3.5), there exist a variety of countable systems
of potentials showing the properties of completeness and closure in L2 (∂G). Many
systems, however, are much more difficult to handle numerically (for instance,
the ellipsoidal systems of Lamé or Mathieu functions). Although they are orig-
inally meant for particular use in series expansions corresponding to ellipsoidal
boundaries (see, e.g., [34]), they can be likewise taken in any regular (i.e., also
in spherical) Runge–Walsh framework. Nonetheless, our particular aim here is to
show that all constructive approximation in physical geodesy can be provided by
mono-pole and multi-pole (i.e., outer harmonics) conglomerates even if the Runge
region inside is an ellipsoid G such that convergence problems do not occur in G if
the approximation method is suitably organized.
Seen from numerical point of view it is preferable to study some further
kernel systems generated by superposition (i.e., infinite sums) of outer harmonics
(as described, e.g., in [15]). Indeed, if they are explicitly available as elementary
functions like kernel representations known from minimum norm interpolation,
i.e., in the jargon of physical geodesy “least squares collocation” (see, e.g., [55]),
these systems turn out to be particularly suitable for numerical purposes because
of their simple implementation.
Theorem 3.6. Let G ⊂ R3 be a regular region such that BR (0) G. Suppose that
the kernel function K(·, ·) : R3 \BR (0) × BR (0) → R is given by
∞ 2k+1

K ∧ (k)H−k−1,l (x)Hk,l (y)
R R
K(x, y) =
k=0 l=1
∞ k
R 2k + 1 ∧ |y| x y
= K (k) Pk · (47)
|x|
k=0
4πR2 |x| |x| |y|
for x ∈ R3 \BR (0), y ∈ BR (0). Let Y = {yn }n=0,1,... be a fundamental system in

BR (0). Suppose that
⎛ ⎞k
∞

R ⎠
(2k + 1) K ∧ (k) ⎝ <∞ (48)
inf |x|
k=0 x∈∂G
with K ∧ (k) = 0 for k ∈ N0 . Then the system

K(·, yn )
∂G n=0,1,...
is closed in L2 (∂G).
Proof. Let F be of class L2 (∂G). The function P given by

P (y) = K(x, y)F (x) dS(x), (49)
∂G
is analytic in BR (0). Indeed, for all y ∈ R3 with |y| < R, it follows from (49) that
∞
2k+1

∧ R R
P (y) = K (k) Hk,j (y) F (x)H−k−1,j (x) dS(x). (50)
k=0 j=1 ∂G
Assume that P (yn ) = 0 for n = 0, 1, . . .. Since Y = {yn }n=0,1,... is a fundamental

system in BR (0), the function P vanishes in BR0 (0) for all R0 ≤ R. This implies
that

R R
(F, H−k−1,j )L2 (∂G) = F (x)H−k−1,j (x) dS(x) = 0, (51)
∂G
for k ∈ N0 , j = 1, . . . , 2k + 1. Hence, by virtue of the completeness of the system

of outer harmonics (Theorem 3.2), we obtain F = 0 in the topology of L2 (∂G), as
required.
Examples of kernel representations (47) are easily obtainable from known

series expansions in terms of Legendre polynomials (see, e.g., elementary repre-
sentations in [5, 12, 16, 23, 55] based on identities as presented in, e.g., [50]).
Applying the Kelvin transform with respect

! to the
" sphere ΩR (0) around the
origin with radius R, we are led to systems Ǩ(·, y̌n ) n=0,1,... with
∞ 2k+1

K ∧ (k)H−k−1,l (x)H−k−1,l (y̌n )
R R
Ǩ(x, y̌n ) =
k=0 l=1
∞ k+1
2k + 1 ∧ R2 x y̌n
= K (k) Pk · , (52)
k=0
4πR2 |x||y̌n | |x| |y̌n |
where Y̌ = {y̌n }n=0,1,... is the point system generated by application of the Kelvin
transform to Y, i.e., by letting
R2
y̌n = yn , n = 0, 1, . . . (53)
|yn |2
(assuming that 0 ∈
/ Y).
Theorem 3.7. Suppose that Y̌ = {y̌n }n=0,1,... is given as described above. Then the
system

Ǩ(·, y̌n ) (54)
∂G n=0,1,...
2
is closed in L (∂G).
Particularly helpful in geosciences is a fundamental system Y = {yn }n=0,1,...
in BR0 (0) that yields Y̌ = {y̌n }n=0,1,... ⊂ ∂G (cf. Figure 4). In other words, the
closure property is related to points lying on the actual (Earth’s) surface ∂G (note
that the Kelvin transform is easily invertible, so that Y and Y̌ can be easily
constructed from each other).
The (Kelvin modified) kernels Ǩ(·, y̌n ), n ∈ N0 , as given by (52) define re-
producing kernel spaces in Sobolev-like Hilbert spaces. They play a central role in
Figure 4. Fundamental system Y and the Kelvin transformed system

Y̌ ⊂ ∂G (see also [15]).
the theory of (Runge-type) harmonic splines (see [9, 12, 13, 15, 64]). Of particu-
lar significance as examples are the spline kernels corresponding to the following
“symbols” {K ∧ (k)}k∈N0 :
(a) Abel–Poisson kernel:
K ∧ (k) = C0 , k ∈ N0 , (55)
(b) singularity kernel:
2C0
K ∧ (k) = , k ∈ N0 , (56)
2k + 1
(c) logarithmic kernel:
C0
K ∧ (k) = , k ∈ N0 , (57)
(k + 1)(2k + 1)
where C0 = 0 is a real constant (see Examples 4.4–4.6).
From potential theory (see, e.g., [15]) we know the following regularity theo-
rem: Suppose that V is of class Pot (0) (G c ), i.e.,
(i) V ∈ C (2) (G c ) ∩ C (0) (G c ),
(ii) ΔV = 0 in G c ,
(iii) V is regular at infinity, i.e., V (x) = O(|x|−1 ), |x| → ∞.
Then, for all k ∈ N0 and all K G c there exist a constant C (dependent on
K, and ∂G) (note that ∇(1) V = ∇V and ∇(2) V = (∇ ⊗ ∇)V, etc.) such that
1/2
sup |(∇(k) V )(x)| ≤ C |V (x)|2 dS(x) (58)
x∈K ∂G
3 2
(note that the norm of the matrix (∇ ⊗ ∇)V is given by i,j=1 (((∇ ⊗ ∇)V )i,j ) ).
Combining the L2 (∂G)-closure (Theorem 3.2) and the regularity theorems

(Theorem 3.7), we obtain the following result for the system of outer harmonics.
Theorem 3.8. Let G ⊂ R3 be a regular region such that BR (0) G. For given
F ∈ C (0) (∂G), let V be the potential of class Pot (0) (G c ) with V |∂G = F . Then, for
any given ε > 0 and any given K G c , there exist an integer N (dependent on ε)
and a set of coefficients a0,1 , . . . , aN,1 , . . . , aN,2N +1 such that
⎛ 2 ⎞ 1/2
N 2n+1

⎜ ⎟
⎝ F (x) − an,j H−n−1,j (x) dS(x)⎠
R
≤ε (59)

∂G n=0 j=1
and

(k) N 2n+1

sup ∇ V (x) − an,j ∇ H−n−1,j (x) ≤ Cε
(k) R
(60)
x∈K n=0 j=1
hold for all k ∈ N0 .
In other words, the L2 -approximation in terms of outer harmonics on ∂G

implies the uniform approximation (in the ordinary sense) on each subset K with
positive distance to ∂G.
Unfortunately, although the Runge–Walsh framework is formulated for spe-
cial trial systems (such as mono- and/or multi-pole systems), the theorems de-
veloped until now are non-constructive since further information about the choice
of the truncation order N and the coefficients of the approximating linear com-
bination is needed. In order to derive a constructive approximation theorem, the
system of potential values and normal derivatives, respectively, can be orthonor-
malized on ∂G. As a result, we obtain a (generalized) Fourier series expansion
(orthogonal Fourier series expansion) that shows locally uniform approximation.
Theorem 3.9. Let G ⊂ R3 be a regular region such that BR (0) G. For given
F ∈ C (0) (∂G), let V be the solution of the Dirichlet problem V ∈ Pot (0) (G c ),
R
V |∂G = F . Corresponding to {H−n−1,j }n∈N0 ,j=1,...,2n+1 there exists a system
{H−n−1,j (∂G; ·)}n∈N0 ,j=1,...,2n+1 ⊂ Pot (0) (R3 \BR (0)) such that
{H−n−1,j (∂G; ·)|∂G } n∈N0 ;
j=1,...,2n+1
is orthonormal in the sense that

H−n−1,j (∂G; y)H−l−1,k (∂G; y) dS(y) = δn,l δj,k . (61)
∂G
Consequently, V is representable in the form

∞ 2n+1
V (x) = F (y)H−n−1,j (∂G; y) dS(y) H−n−1,j (∂G; x) (62)
n=0 j=1 ∂G
for all points x ∈ K G c . Moreover, for each V (N ) given by

N 2n+1

(N )
V (x) = F (y)H−n−1,j (∂G; y) dS(y) H−n−1,j (∂G; x) (63)
n=0 j=1 ∂G
we have the estimate

sup ∇(k) V (x) − ∇(k) V (N ) (x) (64)
x∈K
⎛ ⎞

N 2n+1
2 1/2

≤C⎝
2
|F (y)| dS(y) − F (y)H−n−1,j (∂G; y) dS(y) ⎠ .

∂G n=0 j=1 ∂G
Note that the orthonormalization procedure can be performed (e.g., by the

well-known Gram–Schmidt orthonormalization process) once and for all in the
case that the boundary surface ∂G of a regular region G is specified.
Next, we rewrite our generalized Fourier approach in a more abstract form.
For that purpose we introduce the concept of so-called Dirichlet bases.
Definition 3.10 (Dirichlet Runge Basis). Let A, G ⊂ R3 be arbitrary regular regions

such that A G (cf. Figure 1, left illustration). A linearly independent system
{Dn }n=0,1,... ⊂ Pot (Ac ) is called a (Pot (Ac )-generated) L2 (∂G)-Dirichlet Runge
basis if
·L2 (∂G)
spann∈N0 {Dn |∂G } = L2 (∂G). (65)
Dirichlet Runge bases are constituted by all special function system discussed
earlier (i.e., mono- and muli-pole system, certain kernel function systems, Lamé
and Mathieu systems, etc.)
Corollary 3.11. Let A, G ⊂ R3 be regular regions such that A G holds true.

Let {Dn∗ }n=0,1,... ⊂ Pot (Ac ), be a function system generated by (Gram–Schmidt)
orthonormalization of an L2 (∂G)-Dirichlet Runge basis {Dn }n=0,1,... ⊂ Pot (Ac ),
such that

(Dn∗ , Dm
∗
)L2 (∂G) = Dn∗ (x)Dm∗
(x) dS(x) = δn,m . (66)
∂G
If F ∈ C (0) (∂G), then

2 12
(N )
lim F (x) − F (x) dS(x) = 0, (67)
N →∞ ∂G
where F (N ) denotes the truncated Fourier series expansion

N

F (N ) = (F, Dn∗ )L2 (∂G) Dn∗ ∂G . (68)
n=0

The potential V ∈ Pot (0) (G c ) satisfying V ∂G = F can be represented in the form

lim sup V (x) − V (N ) (x) = 0, (69)
N →∞
x∈K
for every K G c , where

N
V (N ) = (F, Dn∗ )L2 (∂G) Dn∗ . (70)
n=0
The concrete versions of the Runge–Walsh theorem, i.e., the (generalized)

Fourier expansions (68), are indeed constructed to have the permanence property:
The transition from F (N ) to F (N +1) , and therefore from V (N ) to V (N +1) , merely
necessitates the addition of one more term; all the other terms obtained formerly
remain unchanged. This is characteristic of orthogonal expansions.
In connection with the L2 (∂G)-regularity theorems, we additionally find

the following estimate: For given F ∈ C (0) (∂G), let V satisfy V ∈ Pot (0) (G c ),
V |∂G = F . Then

N
(k) ∗ (k) ∗
sup ∇ V (x) − (F, Dn )L2 (∂G) ∇ Dn (x)
x∈K

n=0
1/2 (71)
N
2 ∗ 2
≤ C F L2(∂G) − (F, Dn )L2 (∂G)
n=0
holds for all k ∈ N0 and all subsets K G c . In addition, Corollary 3.11 indicates
that F − F (N ) is L2 (∂G)-orthogonal to all members of the L2 (∂G)-orthonormal
Runge basis up to the index N . This observation is valid for the Dirichlet problem.
Corollary 3.12. Let A, G ⊂ R3 be regular regions such that the “Runge property”
A G holds true. Let {Dn }n=0,1,... ⊂ Pot (Ac ) be an L2 (∂G)-Dirichlet Runge
basis. If F ∈ C (0) (∂G), then
2 12
(N )
lim F (x) − F (x) dS(x) = 0, (72)
N →∞ ∂G
where the coefficients aN N

0 , . . . , aN of the function

N

F (N ) = aN
n Dn ∂G (73)
n=0
satisfy the “normal equations”

N
aN
n (Dk , Dn )L2 (∂G) = (Dk , F )L2 (∂G) , k = 0, . . . , N. (74)
n=0

The potential V ∈ Pot (0) (G c ) satisfying V = F can be represented in the form
∂G

lim sup V (x) − V (N ) (x) = 0, (75)
N →∞
x∈K
where

N
V (N ) = aN
n Dn (76)
n=0
for every K G c .
Remark. Later on, the construction principle based on Corollary 3.12 formulated
in an appropriately defined Sobolev space structure will lead to harmonic spline
interpolation involving (Sobolev-like) reproducing kernel Hilbert space (RKHS)
features.
The approximation of boundary values and the gravitational potential of,
e.g., the Earth by the method of generalized Fourier expansion in terms of outer
harmonics is achieved by superposition of functions with oscillating character.

The oscillations grow in number, but they decrease in size with increasing trunca-
tion order. The oscillating character of the generalized Fourier expansions remains
true (cf. [11]) if other trial bases are used, for example, mono-poles and certain
kernel function representations such as Abel–Poisson (cf. Example 4.4) and sin-
gularity kernels (cf. Example 4.5). Thus, generalized Fourier expansions provide
least squares approximation by successive oscillations, which become larger and
larger in number, but smaller and smaller in amplitude. It is therefore not (as
[65] has pointed out) a technique of osculating character (as, e.g., interpolation in
reproducing Hilbert spaces by harmonic splines (as proposed by W. Freeden [9]
and L. Shure et al. [64])).
Since the time of [30], there is evidence – at least in the spherical context us-
ing multi-poles, i.e., outer harmonics – that a Fourier series expansion provides an
excellent (spherically reflected) trend approximation of a harmonic function such
as the Earth’s gravitational and magnetic potential. The ideal frequency localiza-
tion, more accurately momentum localization of outer harmonics – each of them
referring to a certain degree of oscillation – has proved to be extraordinarily ad-
vantageous due to the physical interpretability and the immediate comparability
of the Fourier coefficients for observables. From a numerical point of view, how-
ever, trial functions would be desirable that show ideal frequency as well as space
localization on the reference sphere. The uncertainty principle (see, e.g., [14, 26]
and the references therein) teaches us that both properties are mutually exclusive
(except in the trivial case). This explains some problems in the Fourier technique
of approximation, at least by means of outer harmonics. Fourier expansions in
terms of outer harmonics are well suited to resolve low-frequency ingredients in
an observable, while their application is critical to obtain high-resolution phe-
nomena. The kernel functions provided by Example 4.4, Example 4.5, and Exam-
ple 4.6 (among others which are similarly constructed) show a reduced frequency
but increased space localization (on a Runge reference surface) as the series con-
glomerates of outer harmonics are constructed to cover various spectral bands
(see, e.g., [26]).
As graphical examples (taken from the Ph.D.-thesis [49]) we illustrate the
members of the Runge–Walsh orthonormalized low-degree outer harmonics on the
International Reference Ellipsoid (IRE) ∂G thereby choosing the standard param-
eters as specified in physical geodesy (for more detailed IRE-parameter specifica-
tions see, e.g., [39]) (see figure 5).
The figures have been obtained via the well-known Gram–Schmidt orthonormal-
izing process in canonical way. Since the ellipsoid is quite close to a sphere, the
phenotype of the Runge–Walsh orthonormal functions does not differ so much
from outer harmonics on a sphere.
It should also be mentioned that the Runge–Walsh concept also leads to
· C (0) (∂G) -approximation. To be more concrete, from our considerations lead-
ing to locally uniform approximation, we know for a given regular region G with
(a) Degree 1 and order 1. (b) Degree 4 and order 2.
(c) Degree 5 and order 7. (d) Degree 6 and order 4.
(e) Degree 6 and order 10. (f) Degree 8 and order 5.
(g) Degree 9 and order 10. (h) Degree 10 and order 15.
Figure 5. Outer harmonics of different degree and order on the Inter-

national Reference Ellipsoid (IRE).
BR (0) G that
·L2 (∂G)
= L2 (∂G).
R
span n=0,1,...; {H−n−1,j |∂G } (77)
j=1,...,2n+1
The same results remain valid when the regular surface ∂G is replaced by any
inner parallel surface ∂G(−τ ) of distance |τ | to ∂G (where |τ | is chosen sufficiently
small). This fact can be exploited to verify the following closure properties (see [8]
for a detailed proof).
Theorem 3.13. Let ∂G be the boundary of a regular region such that BR (0) G.
Then the system {H−n−1,j |∂G }n∈N0 ,j=1,...,2n+1 is closed in C (0) (∂G):
R
·C (0) (∂G)

= C (0) (∂G).
R
spann=0,1,...; {H−n−1,j |∂G } (78)
j=1,...,2n+1
Remark. The same arguments leading to the C (0) (∂G)-closure of outer harmonics
on ∂G apply to all other systems for which the L2 (∂G)-closure is known, e.g.,
ellipsoidal harmonics, Lamé functions, etc.
Combining our results obtained by Theorem 3.13, we easily arrive at the
following statement.
Theorem 3.14. Let G ⊂ R3 be a regular region such that BR (0) G. For a given
(boundary) function F ∈ C (0) (∂G), let V ∈ Pot (0) (G c ) satisfy V |∂G = F . Then,
for every ε > 0, there exist an integer N (depending on ε) and a finite set of real
numbers an,j such that

N 2n+1

sup V (x) − an,j H−n−1,j (x)
R
x∈G c n=0 j=1

N 2n+1

≤ sup F (x) − an,j H−n−1,j (x) ≤ ε.
R
(79)
x∈∂G n=0 j=1
Unfortunately, a constructive procedure of determining best approximate co-

efficients an,j in the C (0) (∂G)-topology seems to be unknown. Therefore, harmonic
splines (see, e.g., [9, 12, 64]) can be introduced in reproducing kernel Hilbert
subspaces of Pot (Ac ) (characterized by variational principles), so that the spline
method can be regarded as an immediate extension of the method of generalized
Fourier series expansions to reproducing kernel subspaces of Pot (Ac ), hence, pro-
viding coefficients that are optimal in a different (Sobolev like) norm. Moreover,
bandlimited as well as non-bandlimited spline wavelets (see [13, 15, 20], and the
references therein) can be introduced to guarantee constructive multiscale variants
of the Runge–Walsh concept (for more details about the role of non-bandlimited
(Runge-type) spline solutions in geodetic boundary value problems the reader is
referred to the next section).
3.2. RKHS-closure and spline interpolation

Next, our purpose is to formulate closure theorems for a Sobolev-like reproduc-
ing kernel Hilbert subspace of Pot (Ac ) (see [10, 12]): Let A, G ⊂ R3 be regular
regions so that A G (see Figure 1, left illustration). Suppose that {Dn∗ }n=0,1,...
is an L2 (∂A)-Dirichlet Runge basis (in the sense of Definition 3.10) obeying the
orthonormality condition

(Dn∗ |∂A , Dm
∗
|∂A )L2 (∂A) = Dn∗ (x)Dm
∗
(x) dS(x) = δn,m . (80)
∂A
Assume that the system {En }n=0,1,... is given by

En = σn Dn∗ , σn ∈ R\{0}, n = 0, 1, . . . , (81)
and that the sequence {σn }n∈N0 is summable in the sense
∞
σn2 < ∞. (82)
n=0
Then, for every k ∈ N0 and F̃ ∈ L2 (∂A), the sum

∞
∇(k) F (x) = (F̃ , Dn∗ )L2 (∂A) ∇(k) En (x) , x ∈ Ac , (83)
n=0
satisfies the estimate
∞ 12 12
∞
2
(k) (k)
sup ∇ F ≤ F̃ , Dn∗ sup ∇ En (x) , (84)
x∈K L2 (∂A) x∈K
n=0 n=0
where K Ac . The L2 -regularity condition implies with a constant C̃ (dependent

on A, K) that
2

sup ∇(k) Dn∗ (x) ≤ C̃ 2 |Dn∗ (x)|2 dS(x) . (85)
x∈K ∂A

=1
Even more, the mean value theorem of multi-variate analysis shows (cf. [15]) that
there exist a positive constant C̃˜ such that

(k) ∗ ˜
∇ Dn (x) − ∇(k) Dn∗ (y) ≤ C̃ 2 |x − y|2 (86)
is valid for x, y ∈ K. From (85) it follows that

∞ 1 ∞ 12
2 2

(k) 2
sup ∇ En (x) ≤ C σn . (87)
x∈K n=0 n=0
In other words, the expansion on the right of (83) exists such that F is harmonic
in Ac and regular at infinity. In particular,
∞ 12 ∞ 12

sup ∇(k) F (x) ≤ C σn2 F̃ , Dn∗ (88)
x∈G c L2 (∂A)
n=0 n=0
(note that G is assumed to be a regular region satisfying A G). All functions F

which can be expressed by such series as given in (83) form a linear space H(Ac )
on which we are able to impose the structure of a separable Hilbert space (cf. [10])
∞

c
H(A ) = F = (F̃ , Dn∗ )L2 (∂A) En , F̃ ∈ L2 (∂A) (89)
n=0
by taking as an inner product, for members F, G ∈ H(Ac ) associated to F̃ , G̃ ∈

L2 (∂A), respectively, the expression
(F, G)H(Ac ) = (F̃ , G̃)L2 (∂A)
∞
= F̃ , Dn∗ G̃, Dn∗
L2 (∂A) L2 (∂A)
n=0
∞
1
= 2
(F, Dn∗ )L2 (∂A) (G, Dn∗ )L2 (∂A) . (90)
σ
n=0 n
Theorem 3.15. (H(Ac ), (·, ·)H(Ac ) ) is a separable Hilbert space possessing the
(uniquely determined) reproducing kernel
∞
∞

KH(Ac ) (x, y) = En (x) En (y) = σn2 Dn∗ (x) Dn∗ (y) (91)
n=0 n=0
for all x, y ∈ Ac .
Suppose that A, G ⊂ R3 are regular regions such that A G. Assume that
∞
:
X = XN , (92)
N =1
where
XN = {x1 , . . . , xN } ⊂ ∂G, xi = xj , i = j (93)
is a countable dense set of points on ∂G. Then we are able to show the following
H(Ac )-closure result:
Lemma 3.16.
·H(Ac )
H(Ac ) = span KH(Ac ) (x, ·) . (94)
x∈X

Proof. Our aim is to prove that F ∈ H(Ac ) and F, KH(Ac ) (x, ·) H(Ac ) = 0 for all
x ∈ X implies
that F = 0. The reproducing kernel Hilbert space structure tells
us that F, KH(Ac ) (x, ·) H(Ac ) = 0 is equivalent to F (x) = 0. According to our
construction, F is continuous on ∂G. Hence, if F (x) = 0 for some x ∈ ∂G, then F
is different from zero for a whole neighborhood of x on ∂G. But this contradicts
the density of X. Hence, F = 0 on ∂G. The analyticity of F finally yields F = 0
in H(Ac ), as desired.
Lemma 3.16 motivates to discuss discrete boundary value problems within

the framework of H(Ac ), in other words, contrary to the classical boundary value
problems, where the solution process is based on the continuous knowledge of the
“boundary function” as a whole (note that, in the case of a discrete boundary
value problem, the boundary information is given only in a set of discrete points).
For simplicity, we start with the exterior Dirichlet boundary value problem
(EDP) in its classical formulation: Let A, G ⊂ R3 be regular regions so that
A G (see Figure 1, left illustration). Given a function F of class C (0) (∂G),
find a function V of class Pot (0) (G c ) satisfying the boundary condition V |∂G (x) =
F (x), x ∈ ∂G. Our goal is to construct an H(Ac )-spline (interpolation) solu-
tion of the discrete exterior Dirichlet problem (DEDP). To this end, we are in-
terested in a (Runge-type) spline potential P ∈ H(Ac ) with observed values
F (xi ) = Exi [P ] = P (xi ) = βi , i = 1, . . . , N, where the points x1 , . . . , xN ∈ ∂G
are assumed to be associated with linearly independent bounded evaluation func-
tionals Ex1 , . . . , ExN (with respect to the H(Ac )-topology) applied to the “(Runge)
restriction” P |G c ∈ H(Ac )|G c of the potential P ∈ H(Ac ). In doing so, we are able
to find a minimum norm solution SN P
∈ H(Ac ) as a linear combination of the rep-
resenters Exi [KH(Ac ) (·, ·)] to the functionals Exi , i.e., SN P
is exactly the projection
of P to the N -dimensional linear subspace spanned by the linearly independent
representers Exi [KH(Ac ) (·, ·)], i = 1, . . . , N (see, e.g., [6]). In other words, the solu-
tion of (DEDP) is sought in the reproducing kernel Hilbert space H(Ac ) under the
assumption that {β1 , . . . , βN } with βi = Exi [P ] = P (xi ), i = 1, . . . , N, is the (ob-
served) given data set for the unknown potential P corresponding to the discrete
set XN = {x1 , . . . , xN } of points on ∂G. All in all, the aim of minimum norm inter-
polation in H(Ac ) as proposed here is to find the “smoothest” SN P
∈ H(Ac ) within
c
the set of all H(A )-interpolants, where the norm is minimized in the metric of
H(Ac ). Equivalently, the problem is to find a function SN P
within the interpolatory
set
! "
IEPx1 ,...,Ex = Q ∈ H(Ac ) : Exi [Q] = Exi [P ] = P (xi ) = βi , xi ∈ ∂G, i = 1, . . . , N ,
N
(95)
such that
- P-
- SN - = inf Q H(Ac ) . (96)
H(Ac ) Q∈I P Ex1 ,...,Ex
N
For any E-unisolvent system XN = {x1 , . . . , xN } ⊂ ∂G, i.e., for any system
XN = {x1 , . . . , xN } such that {Ex1 , . . . , ExN } forms a set of N linearly indepen-
dent bounded linear functionals on H(Ac ) we introduce H(Ac )-splines relative to
{Ex1 , . . . , ExN } in the following way:
Definition 3.17 (Harmonic Splines). Let XN = {x1 , . . . , xN } ⊂ ∂G be an E-
unisolvent system on ∂G. Then, any function S ∈ H(Ac ) given by

N
N
S(x) = ai Exi [KH(Ac ) (·, x)] = ai KH(Ac ) (xi , x) (97)
i=1 i=1
with arbitrarily given (real) coefficients a1 , . . . , aN is called an H(Ac )-spline rela-

tive to {Ex1 , . . . , ExN }.
The space of all H(Ac )-splines relative to {Ex1 , . . . , ExN } is denoted by
Spline H(Ac ) (Ex1 , . . . , ExN ).
Clearly, Spline H(Ac ) (Ex1 , . . . , ExN ) is an N -dimensional subspace of H(Ac ).

Moreover, by virtue of the reproducing property in H(Ac ), we immediately obtain
the so-called H(Ac )-spline formula.
Lemma 3.18. Let S be a function of class Spline H(Ac ) (Ex1 , . . . , ExN ). Then, for
each F ∈ H(Ac ),

N
N
(S, F )H(Ac ) = ai Exi [F ] = ai F (xi ). (98)
i=1 i=1
By virtue of the E-unisolvence of the system XN = {x1 , . . . , xN } ⊂ ∂G it is

not difficult to verify the uniqueness of interpolation.
Lemma 3.19. For a given potential P ∈ H(Ac ), there exist a unique element SN P
characterized by the property SN ∈ Spline H(Ac ) (Ex1 , . . . , ExN ) ∩ IEx ,...,Ex .

P P
1 N
Proof. The application of the N bounded linear functionals Ex1 , . . . , ExN on H(Ac )
to the H(Ac )-spline of the form (97) yields N linear equations in the unknowns
aN N
1 , . . . , aN , i.e.,

N
j Exi Exj [KH(Ac ) (·, ·)] = Exi [P ] = βi ,

aN i = 1, . . . , N, (99)
j=1
where the coefficient matrix is given by

Exi Exj [KH(Ac ) (·, ·)] i,j=1,...,N = KH(Ac ) (xi , xj ) i,j=1,...,N . (100)
From multi-variate interpolation theory (see, e.g., [6]) we know that (100) consti-
tutes a Gram matrix of N linearly independent functions
Ex1 [KH(Ac ) (·, ·)], . . . , ExN [KH(Ac ) (·, ·)],
hence, it is non-singular such that the linear system (99) is uniquely solvable. The
coefficients aN N P
1 , . . . , aN determine the unique interpolating spline SN .
P
The following minimum norm properties for the interpolating spline SN are
easily derivable (see, e.g., [10] for comparable conclusions).
Lemma 3.20 (First Minimum Property). If F ∈ IEPx , then

1 ,...,ExN
||F ||2H(Ac ) = ||SN

P 2
||H(Ac ) + ||SN
P
− F ||2H(Ac ) . (101)
Lemma 3.21 (Second Minimum Property). Suppose that

S ∈ Spline H(Ac ) (Ex1 , . . . , ExN ) and F ∈ IEPx1 ,...,Ex .
N
Then
||S − F ||2H(Ac ) = ||SN
P
− F ||2H(Ac ) + ||S − SN
P 2
||H(Ac ) . (102)
Summarizing our results on H(Ac )-spline interpolation of a finite set of eval-

uation (Dirichlet) functionals we obtain
Theorem 3.22 (Spline Interpolation). The minimum norm interpolation problem
for solving DEDP from N given data P (xi ) = βi , i = 1, . . . , N,
||SN
P
||H(Ac ) = inf ||Q||H(Ac ) (103)
Q∈IEPx
1 ,...,Ex N
is well posed in the sense that its solution exists, is unique, and depends continu-
P
ously on the data β1 , . . . , βN . The uniquely determined solution SN is given in the
explicit form

N
N
i Exi [KH(Ac ) (·, x)] = x ∈ Ac ,

P
SN (x) = aN aN
i KH(Ac ) (xi , x), (104)
i=1 i=1

1 , . . . , aN satisfy the linear equations

N
j Exi Exj [KH(Ac ) (·, ·)] = Exi [P ] = βi ,

aN i = 1, . . . , N. (105)
j=1
Let ϑXN denote the XN -width on ∂G, i.e., the maximal distance for any point
of ∂G to the system XN :

ϑXN = max min |x − y| . (106)
x∈∂G y∈XN
Our interest is the stability of the solution obtained by spline interpolation

by letting ϑXN → 0 as N → ∞. As already known, for every E-unisolvent system
XN = {x1 , . . . , xN } ⊂ ∂G and for every function P ∈ H(Ac ) there exists a unique
element SNP
∈ H(Ac ) satisfying the conditions Exi [P ] = Exi [SN
P
], i = 1, . . . , N.
Lemma 3.23. Let P be a member of class H(Ac ). Suppose that XN ⊂ ∂G is an
E-unisolvent system. Then there exists a constant C > 0 (dependent on ∂G and
A) such that

P
sup Ex [P ] − Ex [SN ] ≤ C ϑXN P H(Ac ) . (107)
x∈∂G
Proof. For x ∈ ∂G, there exists a point y ∈ XN with |x − y| ≤ ϑXN . Observing

the interpolation property Ey [P ] = Ey [SNP
], y ∈ XN , we see that

Ex [P ] − Ex [SN
P
] = (Ex [P ] − Ey [P ]) − Ex [SNP
] − Ey [SN
P
] . (108)
The reproducing kernel structure of H(Ac ) enables us to derive the estimates

1
|Ex [P ] − Ey [P ]| ≤ (κ(x, y)) 2 P H(Ac ) , (109)

Ex [SN
P
] − Ey [SNP 1
] ≤ (κ(x, y)) 2 SNP
H(Ac ) , (110)
where
κ(x, y) = (Ex Ex − 2Ex Ey + Ey Ey ) [KH(Ac ) (·, ·)]. (111)
SNP
is the smoothest H(Ac )-interpolant, i.e., SN
P
H(Ac ) ≤ P H(Ac ) . From (108),
(109), and (110) we therefore obtain
1
sup |Ex [SN
P
] − Ex [P ]| ≤ 2 (κ(x, y)) 2 P H(Ac ) , (112)
x∈∂G
where x ∈ ∂G and y ∈ XN = {x1 , . . . , xN } ⊂ ∂G. More explicitly, we have

∞
2
κ(x, y) = σn2 (Ex Dn∗ − Ey Dn∗ ) . (113)
n=0
By use of (86) we find

|Ex Dn∗ − Ey Dn∗ | ≤ C |x − y|, (114)
where C > 0 is a constant (depending on Ac and ∂G). This proves Lemma 3.23.
Summarizing our results we obtain
Theorem 3.24. Let A, G ⊂ R3 be regular regions so that A G. Suppose that P is

of class H(Ac ). Let XN = {x1 , . . . , xN } be an E-unisolvent system on ∂G. Let SN
P
denote the uniquely determined solution of the spline interpolation problem (103).
Then there exists a constant B (dependent on A and G) such that
sup |SN
P
(x) − P (x)| ≤ B ϑXN P H(Ac ) . (115)
x∈G c
As consequences we are able to deduce the following C (0) (∂G)-closure theorem

·C (0) (∂G) ·C (0) (∂G)
C (0) (∂G) = span E|∂G [KH(Ac ) (·, ·)] = span KH(Ac ) (x, ·) , (116)
x∈X x∈X
as well as the L2 (∂G)-closure theorem

·L2 (∂G) ·L2 (∂G)
L2 (∂G) = span E|∂G [KH(Ac ) (·, ·)] = span KH(Ac ) (x, ·) , (117)
x∈X x∈X
provided that X is the union of a sequence {XN } of E-unisolvent systems XN on

∂G with ϑXN → 0 as N → ∞.
4. Runge–Walsh solution of geodetic boundary value problems
Figure 6. Earth’s surface, geoid, ellipsoid (λ = oblique gravity vector

normal to the geoid, ν = normal vector to the actual Earth’s surface).
Terrestrial observations of the gravity field on the real (known) Earth’s sur-
face do not generally provide normal derivatives (cf. Figure 6). Instead, oblique
derivatives are measured, since the actual Earth’s surface does not coincide with
the equipotential surface of the geoid (at least not for large parts over continents).
In the following we are interested in discussing a locally uniform approximation
implied by generalized L2 -Fourier series expansions with respect to certain trial
functions (such as outer harmonics (multi-poles) or mono-pole configurations).
Classically, a solution procedure for the oblique derivative problem is under-
taken by virtue of integral equations using the potential of a single-layer. These
results were essentially worked out by A.V. Bitzadse [4] and C. Miranda [53] (see
also the references therein). In accordance with this work, K.R. Koch, A.J. Pope
[43] applied the integral equation procedure to the so-called geodetic boundary
value problem using the known surface of the Earth. However, the strong nature
of the singularities demanding Cauchy’s principal integral value understanding
turned out to be a serious obstacle. For numerical computation, alternative tech-
niques have to be taken into account. The integral equation method also represents
the point of departure for some subsequent work by W. Freden and H. Kersten
[17–19]. They provide a new concept of approximation, viz. generalized Runge–
Walsh Fourier expansions, thereby transferring strongly singular integrals into reg-
ular ones. As for the classical Dirichlet and Neumann boundary value problems
(see [8, 11, 12]), the generalized Fourier series approach yields Fourier coefficients
of the boundary values within the L2 (∂G)-framework, and it simultaneously im-
plies locally uniform approximation of the solution for subsets totally contained in
the outer space. Even more, in a series of papers, [17–19, 23], and [15] successfully
provide the basis for closure theorems in oblique derivative problems in different
topologies such as uniform as well as Hölder norms. Additionally, [35] deal with
Sobolev norms. However, to the knowledge of the authors, up to now only if the
approximation of the boundary values is implemented as a generalized Fourier ex-
pansion in the L2 -context or the method of generalized Fourier series expansions is
transferred as spline procedure to a reproducing kernel Hilbert C (0) -substructure,
constructive realizations of the oblique derivative problem have been implemented

successfully and efficiently to (noisy) oblique data sets (see, e.g., [36–38]).
Our work concerning the exterior Dirichlet problem leads to a remarkable
conclusion: The L2 -method of generalized Fourier series expansions involving ker-
nel functions of type (52) in Pot (Ac ) can be seen in parallel to the minimum norm
(spline) interpolation in the Sobolev-like H(Ac )-reproducing kernel Hilbert space.
Methodologically, indeed, the generalized Fourier series expansion as well as min-
imum norm (spline) interpolation constitute the same “Runge manifestations”,
however, corresponding to different topologies.
4.1. Oblique boundary value problem corresponding to the actual Earth’s surface
The oblique boundary value problem (OBVP) can be formulated briefly as follows:
Let A, G ⊂ R3 be regular regions so that A G. Given a function F of class
C (0,α) (∂G), 0 < α < 1 (i.e., Hölder-continuous with Hölder coefficient α), find a
function V of class Pot (1,α) (G c ) = Pot (G c ) ∩ C (1,α) (G c ) satisfying the boundary
condition
∂V
(x) = F (x), x ∈ ∂G, (118)
∂λ
where λ is a c(1,α) (∂G)-(unit) vector field (i.e., a Hölder-continuous vector field
with Hölder coefficient α) satisfying the physically plausible condition
inf (λ(x) · ν(x)) > 0 (119)

x∈∂G
with ν being the (unit) normal field on ∂G directed outward into G c (cf. Figure 6).
Remark. If the field λ coincides with the normal field ν on ∂G, Eq. (118) becomes
the boundary condition of the ordinary exterior Neumann boundary value problem.
In this case, we know from [17, 18] that the smoothness conditions imposed on the
boundary values F may be weakened from Hölder continuity to just continuity.
In order to discuss the well-posedness of the exterior oblique derivative prob-

lem (EODP), we can follow the standard solution theory by use of the potential
of a single-layer. Existence and uniqueness are recapitulated briefly in accordance
with the work of A.V. Bitzadse [4] and C. Miranda [53]. Moreover, as in classical
theory, we are interested in a regularity theorem providing the solution in locally
uniform topology in the outer space from the L2 -approximation of the boundary
values on ∂G (cf. [17–19]).
The point of departure for our considerations concerning L2 (∂G)-approxima-
tion is the following result (for the proof see [18]).
Theorem 4.1. Let A, G ⊂ R3 be regular regions such that A G holds true. If

{Dn }n=0,1,... is an L2 (∂A)-Dirichlet Runge basis (in the sense of Definition 3.10),
then
∂Dn
spann=0,1,... (120)
∂λ ∂G
is dense in (C (0,α) (∂G), · L2 (∂G) ).
For numerical purposes we orthonormalize the members of an L2 (∂A)-Diri-

chlet Runge basis {Dn }n=0,1,... (e.g., certain systems of mono-poles (fundamen-
tal solutions), outer harmonics (multi-poles), and/or appropriate kernel functions
such as Abel–Poisson kernel (55), singularity kernel (56), logarithmic kernel (57),
etc.). We obtain a system {Dn∗ }n=0,1,..., , Dn∗ ∈ Pot (Ac ), Dn∗ ∈ span (D0 , . . . , Dn )
satisfying the orthonormality condition

∂Dn∗ ∂Dm ∗
∂Dn∗ ∂Dm∗
, = (x) (x) dS(x) = δn,m . (121)
∂λ ∂λ L2 (∂G) ∂G ∂λ ∂λ
We are able to derive the following limit relation (see [15, 17, 18]): If F ∈ C (0,α) (∂G),
then
⎛ ⎞ 12
N ∗
∗
2

∂D ∂D
lim ⎝ F (x) − F, n n
(x) dS(x)⎠ = 0. (122)
N →∞ ∂G n=0
∂λ L2 (∂G) ∂λ

∂V
Consequently, the uniquely determined V ∈ Pot (1,α) (G c ), ∂λ ∂G = F , can be
approximated in the form
N
∂Dn∗
V (N ) = F, Dn∗ (123)
n=0
∂λ 2
L (∂G)
with

lim sup V (x) − V (N ) (x) = 0, (124)
N →∞
x∈K
for every K G c . Equivalently, V (N ) can be obtained by use of the L2 (∂A)-

Dirichlet Runge basis {Dn }n=0,1,... , Dn ∈ Pot (Ac ), in the form

N
(N )
V = aN
n Dn , (125)
n=0

0 , . . . , aN satisfy the “normal equations”
N
N ∂Dn ∂Dk ∂Dk
an , = F, , k = 0, . . . , N. (126)
n=0
∂λ ∂λ L2 (∂G) ∂λ L2 (∂G)
Next we come to the H(Ac )-realization of the discrete exterior oblique de-
rivative problem (DEODP) (see also [12, 15] and the references therein). DEODP
demands to study the boundedness of the functional Dx = ∂/∂λ(x) of the oblique
derivative D = ∂/∂λ at a point x ∈ ∂G with respect to the H(Ac )-topology (as

introduced by Lemma 3.16). For x ∈ ∂G and ε > 0 we are able to conclude that
1 1
|F (x) − F (x + ελ(x))| ≤ F H(Ac ) KH(Ac ) (x, ·) − KH(Ac ) (x + ελ(x), ·) H(Ac )
ε ε
≤ C F H(Ac ) (127)
for some constant C > 0 (depending on Ac and ∂G) provided that F is of class
H(Ac ) (cf. [12]). Consequently, by the same minimum norm procedure as for the
evaluation functional we obtain (thereby formally replacing E by D)
Theorem 4.2. Let A, G ⊂ R3 be regular regions so that A G. Let P be a member
∂λ (xi ) = ∂λ (xi ) = Dxi [P ] = βi , i = 1, . . . , N. Then the
of H(Ac ) satisfying ∂V ∂P
minimum norm interpolation problem

SN
P
H(Ac ) = inf F H(Ac ) (128)
P
F ∈ID x 1 ,...,DxN
with
! "
ID
P
x1 ,...,Dx
= Q ∈ H(Ac ) : Dxi [Q] = Dxi [P ] = βi , xi ∈ ∂G, i = 1, . . . , N ,
N
(129)
ously on the data ∂V ∂P
∂λ (xi ) = ∂λ (xi ) = βi , i = 1, . . . , N . The uniquely determined
solution is given in the form

N
i Dxi [KH(Ac ) (x, ·)], x ∈ Ac ,

P
SN (x) = aN (130)
i=1


N
i Dxi Dxj [KH(Ac ) (·, ·)] = βj ,

aN j = 1, . . . , N. (131)
i=1
As in the Dirichlet case the stability should be investigated. Analogously to
Lemma 3.23 we get
1
sup |Dx [SN
P
] − Dx [P ]| ≤ 2 (κ(x, y)) 2 P H(Ac ) (132)
x∈∂G
where
κ(x, y) = (Dx Dx − 2Dx Dy + Dy Dy )[KH(Ac ) (·, ·)], (133)
x ∈ ∂G, and y ∈ XN = {x1 , . . . , xN } ⊂ ∂G. More explicitly, we have
∞ 2
2 ∂Dn∗ ∂Dn∗
κ(x, y) = σn (x) − (y) . (134)
n=0
∂λ ∂λ
By use of (86) we find (cf. [15]) that

∂Dn∗ ∂Dn∗

∂λ (x) − ∂λ (y) ≤ C |x − y|, (135)
where C > 0 is some constant (depending on Ac and ∂G).

Theorem 4.3. Let A, G ⊂ R3 be regular regions so that A G. Suppose that P is

of class H(Ac ). Let XN = {x1 , . . . , xN } be a D-unisolvent system on ∂G. Let SN
P
denote the uniquely determined solution of the spline interpolation problem (128).
Then there exists a constant B (dependent on A and G) such that
sup |SN
P
(x) − P (x)| ≤ B ϑXN P H(Ac ) . (136)
x∈G c
Obviously,
·C (0) (∂G)
C (0) (∂G) = span D|∂G [KH(Ac ) (·, ·)] (137)
x∈X
and
·L2 (∂G)
L2 (∂G) = span D|∂G [KH(Ac ) (·, ·)] , (138)
x∈X
where X is the union of a sequence {XN } of D-unisolvent systems XN on ∂G with

ϑXN → 0 as N → ∞.
For computational reasons, reproducing Hilbert space kernels with closed
expressions in terms of elementary functions are welcome (see, e.g., [10, 12, 68] for
more details). For that purpose, we specialize to an inner Runge-ball BR0 around
the origin, so that BR0 A = BR (0), R < inf x∈∂G |x|. Moreover, in practical
applications, an outer harmonics L2 (∂A)-Dirichlet Runge basis is of frequent use
at least in physical geodesy (see, e.g., [10, 12, 44, 55, 57]). In fact, via the addition
theorem of spherical harmonics, a large number of representations can be derived
from series expansions in terms of Legendre polynomials:
∞
n+1
2n + 1 R2 x y
KH (R3 \BR (0)) (x, y) = σn2 Pn · , x, y ∈ R3 \BR (0).
n=0
4πR2 |x||y| |x| |y|
(139)
In this approach we restrict ourselves to three important cases.
Example 4.4.
n
R20
σn2 = , R0 < R, n = 0, 1, . . . (140)
R2
Abel–Poisson kernel:
|x| |y| |x|2 |y|2 − R40

KH (R3 \BR (0)) (x, y) = (141)
4πR20 (L(x, y)) 2
3
with
L(x, y) = |x|2 |y|2 − 2R20 x · y + R40 (142)
and
∂ ∂
K 3 (x, y)
∂λx ∂λy H (R \BR (0))

1 9|x|2 |y|2 − R40 (λ(x) · x)(λ(y) · y)
=
4πR20 |x| |y|
3
(L(x, y)) 2
2 2 4
3 3|x| |y| − R0 (λ(x) · x)|y| 2 2
+ (R0 (λ(y) · x) − (λ(y) · y)|x| )
4πR0 2 (L(x, y)) 2 |x|
5

3 3|x|2 |y|2 − R0 4 (λ(y) · y)|x| 2 2
+ (R 0 (λ(x) · y) − (λ(x) · x)|y| )
4πR0 2 (L(x, y)) 2
5
|y|
3 |x|2 |y|2 − R0 4
+ 2 5 |x| |y| R0 2 (λ(x) · λ(y)) − 2(λ(y) · y)(λ(x) · x)
4πR0 (L(x, y)) 2
15 |x|2 |y|2 − R0 4
+ |x| |y|(R0 2 (λ(x) · y) − (λ(x) · x)|y|2 )
4πR0 2 (L(x, y))7/2
× (R0 2 (λ(y) · x) − (λ(y) · y)|x|2 ). (143)
Example 4.5.
n
2 R20
σn2 = , R0 < R, n = 0, 1, . . . (144)
2n + 1 R2
Singularity kernel:
1 1
KH (R3 \BR (0)) (x, y) = (145)
2π (L(x, y)) 12
and
∂ ∂
K 3 (x, y) (146)
∂λx ∂λy H (R \BR (0))
1 1
= (R2 (λ(x) · λ(y)) − 2(λ(x) · x)(λ(y) · y))
2π (L(x, y)) 32 0
3 1
+ (R2 (λ(x) · y) − (λ(x) · x)|y|2 )(R20 (λ(y) · x) − (λ(y) · y)|x|2 ).
2π (L(x, y)) 32 0
Example 4.6.
n
1 R20
σn2 = , R0 < R, n = 0, 1 . . . (147)
(2n + 1)(n + 1) R2
Logarithmic kernel:

1 2R20
KH (R3 \BR (0)) (x, y) = ln 1 + (148)
4πR20 M (x, y)
with
1
M (x, y) = (L(x, y)) 2 + |x| |y| − R20 (149)
and
∂ ∂
K 3 (x, y)
∂λx ∂λy H (R \BR (0))
1 1
=
2π (M (x, y)) + 2R20 M (x, y)
2
. 3
/
× (L(x, y))− 2 (R20 (λ(y) · x) − |x|2 (λ(y) · y))(R20 (n(x) · y) − |y|2 (λ(x) · x))
1 1
+
2π (M (x, y))2 + 2R20 M (x, y)
) *
1 (λ(x) · x)(λ(y) · y)
× (L(x, y)− 2 (R20 (λ(x) · λ(y)) − 2(λ(x) · x)(λ(y) · y))) −
|x||y|
1 M (x, y) + R20
+
π ((M (x, y))2 + 2R20 M (x, y))2
) *
− 12 2 2 |x|
× (L(x, y)) (|x| (λ(y) · y) − R0 (λ(y) · x)) (λ(y) · y)
|y|
) *
−2 1
2 2 |y|
× (L(x, y)) (|y| (λ(x) · x) − R0 (λ(x) · y)) + (λ(x) · x) . (150)
|x|
Example 4.7.
n
1 1 R20
σn2 = 3 , R0 < R, n = 0, 1, . . . (151)
R0 (2n + 1)2 (2n + 3)
R2
Newton kernel:
2
1 1
KH (R3 \BR (0)) (x, y) = dV (z) (152)
4π BR0 (0) |x − z||y − z|
and
2
∂ ∂ 1 λ(x) · (x − z) λ(y) · (y − z)
K 3 (x, y) = dV (z).
∂λx ∂λy H (R \BR (0)) 4π BR0 (0) |x − z|3 |y − z|3
(153)
In other words, the iterated Newton kernel leads back to the volume-based
reproducing kernel Hilbert space structure recently developed by W. Freeden, C.
Gerhards [15].
Finally it should be noted that the advantage of a sphere-based reproducing
kernel Hilbert space (H(Ac ), (·, ·)H(Ac ) ) is twofold:
(1) The reproducing kernel contains outer harmonic contributions of any degree
like the Earth’s gravitational potential itself.
(2) The geometry of the regular region G may be arbitrary so that especially the
actual Earth’s surface ∂G can be easily handled in numerical computations
thereby taking advantage from the fact that there is no need for numerical
integration. The coefficient matrix of the occurring linear (spline) systems is
symmetric and positive definite, hence, they are solvable by standard methods
of linear algebra.
Even better, multi-pole (far and near field) methods in combination with suit-
able domain decomposition procedures (see [36–38] and the references therein)
make spline interpolation (and/or smoothing in the case of error affected data) an
efficient as well as economical technique for numerical application.
Nevertheless, it should be mentioned that the particular choice of the repro-
ducing kernel, i.e., the appropriate topology of H(Ac ) is a problem in minimum
norm (spline) interpolation. In principle, seen from a theoretical point of view, all
topologies are equivalent. In practice, however, the reproducing kernel structure
should be in adaptation to the characteristics of the available dataset (if possible).
Altogether, Runge-type spline interpolation is a constructive method for solv-
ing the oblique derivative problem. The difficulties are the suitable choice of an
inner Runge-sphere and the Sobolev structure, the positioning of the point systems
on ∂G, and the efforts to solve the occurring linear systems.
4.2. Molodensky boundary value problem in physical geodesy
The gravimetric determination of the geoid is a current research area in physical
geodesy. It has become even more important, since the GPS techniques deliver
accurate measurements with dense data coverage. In particular, for geodetic pur-
poses, locally reflected approximation methods resulting in high wavelength geoidal
reconstructions are of future significance.
The original problem of Molodensky can briefly be formulated as follows:
Given, at all points on the geoid ∂G, the gravity potential W and the gravity
vector w = ∇W , then the aim is to determine the geoidal surface ∂G. It is clear by
the definition of the geoid, that W is constant on ∂G, such that only a gauge value
W0 has to be given. Furthermore we will not discuss in detail here, how the gravity
vector w is obtained on ∂G from measurements on the real Earth’s surface. For a
detailed discussion of determining w from the Earth’s surface to the (a priori not
known) geoid, the reader is referred to the literature (an important approach is
given, e.g., by L. Hörmander [40], see also the references therein). Our description
of the linearized Molodensky problem essentially follows the conventional concept
of, e.g., [33, 34, 45, 54].
The geoidal height determination is based on the fact that the geoid ∂G is
approximated by the boundary ∂T of a regular region T called the telluroid with
known gravitational potential U in T c (cf. Figure 7). We assume that there exists
a one-to-one correspondence between ∂G and ∂T . W is the actual potential and
U is an approximation of W called the normal potential. We define u = ∇U which
is called the normal gravity and w = ∇W called the actual gravity which is given
on ∂G. Assume that, for given x ∈ ∂T , the point y ∈ ∂G is the one associated
to x by the one-to-one correspondence between ∂G and ∂T (cf. Figure 7). The
two points are connected by the vector d = y − x. A substitute formulation of the
classical Molodensky problem is now to determine the length of d, i.e., the distance
of the geoid and the approximating telluroid along the one-to-one correspondence
Figure 7. Geoid ∂G, telluroid ∂T , and their one-to-one correspondence.
between ∂G and ∂T . To this end we introduce

δW = W |∂G − U |∂T , (154)
δw = w|∂G − u|∂T , (155)
where δW is called the potential anomaly and δw is called gravity anomaly (see
[44, 45, 54]). Furthermore, we define the disturbing potential T by
T =W −U (156)
in G, so that we have
δW = T |∂G + U |∂G − U |∂T , (157)
δw = w|∂G − u|∂T . (158)
Using the Taylor expansion of u and U in terms of d and neglecting terms of higher
order in d (which represents no substantial loss of accuracy if a sufficiently close
telluroid is chosen) we finally arrive at the approximations
δW (x) = T (x) + u(x) · d, (159)
δw(x) = w(y) − u(y) + m(x) d, (160)
x ∈ ∂T , y ∈ ∂G, where we set

∂2U
m = ∇u = . (161)
∂xi ∂xj i,j=1,2,3
Observing the relations

w(y) − u(y) = (∇W )(y) − (∇U )(y) = (∇T )(y) = (∇T )(x) (162)
we arrive at
δW (x) = T (x) + u(x) · d, (163)
δw(x) = (∇T )(x) + m(x) d. (164)
Equation (163) is called Bruns formula. Actually it connects the disturbing poten-
tial T on the telluroid ∂T with the geoid anomalies d, i.e., the anomalies between
the geoid ∂G and the telluroid ∂T . If we assume that m(x) is invertible for all
x ∈ ∂T , we get by virtue of (162)
d = m(x)−1 (δw(x) − (∇T )(x)). (165)
Inserting the identity (165) into equation (163) we end up with
T (x) − u(x) · (m(x))−1 (∇T )(x) = δW (x) − u(x) · m(x)−1 δw(x). (166)
This is the so-called fundamental boundary condition of physical geodesy.
Following [45] (see also the report [54]) the vector u(x)(m(x))−1 can be seen
in first order of d to be oriented in the direction of the exterior unit normal field
ν on the telluroid ∂T . More specifically,
|x|
u(x)(m(x))−1 = − ν(x). (167)
2
Inserting expression (167) into equation (166) therefore results in the identity
2
ν(x) · (∇T )(x) + T (x) = F (x), (168)
|x|
2
F (x) = ν(x) · δw(x) + δW (x) (169)
|x|
(note that the boundary condition (168) can be seen to be equivalent to (166)
transformed in an appropriate coordinate system).
Summarizing all the steps of the linearization procedure we are led to discuss
the following type of a boundary value problem in potential theory. In fact, our
goal is to solve this problem by a constructive Runge approach as presented in
this work (cf. [22]).
Exterior Molodensky Problem (EMP): Find T ∈ Pot (1) (T c ), i.e., T ∈ C (2) (T c ) ∩
C (1) (T c ) with ΔT = 0 in T c and |T (x)| = O(|x|−1 ), |x| → ∞, such that
∂T
(x) + μ(x)T (x) = F (x), x ∈ ∂T , (170)
∂ν
where μ, F ∈ C (0) (∂T ) are known functions on the boundary surface ∂T of the
regular region T (it should be noted that in modern mathematical nomenclature,
the exterior Molodensky problem (EDP) forms a special Robin problem).
Remark. In the case that ∂T is a sphere, the problem becomes the well-known
Stokes problem (see [39] or [54]) and in the case of an ellipsoid it is called ellipsoidal
Stokes problem (see, e.g., [33, 52, 54]). Locally reflected multi-scale solutions of
Stokes’ problem are due to [27, 29] (see also the references in [15]).
Next, we discuss the well-posedness of the Molodensky boundary value prob-
lem corresponding to a regular telluroidal surface ∂T . First, we will reformulate
the problem in our notation.
Exterior Molodensky Problem (EMP): Given F, μ ∈ C (0) (∂T ), find T ∈ Pot (1) (T c )
such that

∂T
+ μT (x) = F (x), x ∈ ∂T . (171)
∂ν
From [22] we borrow the regularity theorem in the L2 (∂T )-context for the
Molodensky problem.
Theorem 4.8. Let T be of class Pot (1) (T c ). Then there exists a constant C(=
C(k; K, ∂T )) such that
- -
- ∂T -
(k) - -
sup ∇ T (x) ≤ C - + μ L2 (∂T ) T L2(∂T ) (172)
x∈K ∂ν -L2 (∂T )
for all K T c and all k ∈ N0 .
For numerical purposes in the sense of the Runge–Walsh approach we again

orthonormalize the members of a Dirichlet Runge basis (Dn )n=0,1,... obtaining
a system {Dn∗ }n=0,1,..., , Dn∗ ∈ Pot (Ac ), Dn∗ ∈ span (D0 , . . . , Dn ), satisfying the
orthonormality condition

∂Dn∗ ∗ ∂Dm
∗
∗
+ μDn , + μDm = δn,m . (173)
∂ν ∂ν L2 (∂T )
In connection with the regularity result we are then able to derive the following
conclusion in the framework of the Molodensky problem: If F ∈ C (0) (∂T ), then
N
∂
∗
lim
N →∞ F (x) − F,
∂ν
+ μ Dn
∂T n=0 L2 (∂T )
2 1/2
∂
× ∗
+ μ(x) Dn (x) dS(x) = 0. (174)
∂ν
∂
Consequently, the potential T ∈ Pot (1) (T c ), ∂ν + μ T = F on ∂T , can be rep-
resented in the form

lim sup T (x) − T (N ) (x) = 0, (175)
N →∞ x∈K
where
N
∂
T (N ) = F, + μ Dn∗ Dn∗ . (176)
n=0
∂ν 2
L (∂T )
Finally it should be noted that the whole solution context developed for the
discrete exterior oblique derivative problem (DEOP) also remains valid for the
discrete exterior Molodensky problem (DEMP) in an obvious way by using the
linear functional M = ∂ν ∂
+ μ instead of D = ∂λ
∂
. We summarize the results.
Theorem 4.9. Let A, G ⊂ R3 be regular regions so that A G. Let P be a mem-

ber of H(Ac ) satisfying Mxi [P ] = βi , i = 1, . . . , N. Then the minimum norm
interpolation problem
SN
P
H(Ac ) = inf F H(Ac ) (177)
P
F ∈IM x 1 ,...,MxN
with
! "
IM
P
x1 ,...,Mx
= Q ∈ H(Ac ) : Mxi [Q] = Mxi [P ] = βi , xi ∈ ∂G, i = 1, . . . , N ,
N
(178)
is well posed in the sense that its solution exists, is unique, and depends con-
tinuously on the data M(V )(xi ) = M(P )(xi ) = βi , i = 1, . . . , N . The uniquely
determined solution is given in the form

N
i Mxi [KH(Ac ) (x, ·)], x ∈ Ac ,

P
SN (x) = aN (179)
i=1


N
i Mxi Mxj [KH(Ac ) (·, ·)] = βj ,

aN j = 1, . . . , N. (180)
i=1
A multi-scale (Runge-type) method for solving (DEMP) obtained by regu-

larization of layer potentials can be found in [21].
Once again, we notice that the exterior Molodensky problem leads to the re-
markable conclusion that the L2 -method of generalized Fourier series expansions
involving kernel functions of type (52) in Pot (Ac ) can be recognized in parallel to
the minimum norm (spline) interpolation in the Sobolev-like H(Ac )-reproducing
kernel Hilbert space. More explicitly, the generalized Fourier series expansion as
well as minimum norm (spline) interpolation constitute the same “Runge man-
ifestations” of solutions for the Molodensky problem, however, corresponding to
different topologies. Nevertheless, the numerical realization is based on different
data assumptions. In case of the L2 -method of generalized Fourier series expan-
sions the data set has to be equidistributed all over the boundary, since integration
and equidistribution are mutually dependent according to the famous Weyl law (cf.
[72]). Spline interpolation may be performed to discrete points, but the stability
of the solution of the occurring linear system and the prevention of oscillations of
the spline approximations imply a reasonable data structure avoiding larger gaps
of the data distribution.
5. Conclusion
This contribution provides constructive realizations of the Runge–Walsh theorem
in order to solve geodetic boundary value problems such as the exterior oblique
derivative problem from discrete data sets. The numerical methods proposed here,
respectively, are Pot (Ac )- and H(Ac )-generalized Fourier series expansions. Par-
ticular kernels serving as trial functions for use in our numerics are mono- and
multi-poles and their Kelvin transforms relative to a “Runge sphere”, i.e., the
boundary of a Runge ball. A mono-pole interrelates the length of its spectral bands
to the distance of the mono-pole from the Runge sphere. The mono-pole, i.e., the
fundamental solution (as well as its Kelvin transformed singularity kernel) is more
and more space localized and simultaneously in accordance with the uncertainty
principle (cf. [14]) less frequency localized, the closer and closer the mono-pole is
positioned to the Runge sphere. As a matter of fact, seen from a methodological
point of view, Pot (Ac )-Fourier approaches using a sequence of kernel functions
corresponding to an inner fundamental system can be realized in a manner equiv-
alent to outer harmonic expansions for completely recovering the gravitational
potential within the framework of boundary value problems. A sequence of kernel
functions (such as the Abel–Poisson, singularity, and logarithmic kernel) is even
conceptually easier to implement than outer harmonic expansions, as long as the
kernels are available in closed form as elementary functions.
Mono-pole, i.e., fundamental solution approximations have a long history.
Early attempts in potential theory to make the so-called method of fundamental
solutions reality date back to the middle of the 19th century (cf. [66]). Related
studies are due to [61, 67]. Further ideas are, e.g., due to [31, 42, 46, 69, 70]. The line
to the Fourier approach as presented here follows [8, 11, 15, 20, 21, 23, 32, 69, 70].
All these approaches take advantage of the Kelvin transform in potential theory
that is not transferable for more general elliptic partial differential equations.
In the meantime, however, generalized Fourier expansions are theoretically
established and practically applied not only to the Laplace equation, but also to
more general elliptic partial differential equations, e.g., the reduced (Helmholtz)
wave equation (see [25, 28]), the Cauchy–Navier equation (see [1, 24]), (reduced)
Maxwell equations (see [22]), the (linear) Stokes equations (see [51] and the refer-
ences therein). [2] used the method of fundamental solutions in poroelasticity to
model stress fields. The drawback of the numerical realization is the need for an
adequate selection of a finite number of points out of the infinite inner fundamental
system. An optimal strategy for positioning a finite system in a computationally
efficient and physically relevant way remains a great challenge for future work.
References
[1] Abeyratne, M.K., Freeden, W., Mayer, C.: Multiscale Deformation Analysis by
Cauchy–Navier Wavelets. J. Appl. Math., 12, 605–645, 2003.
[2] Augustin, M.A.: A Method of Fundamental Solutions in Poroelasticity to Model the
Stress Field in Geothermal Reservoirs. Lecture Notes in Geosystems Mathematics
and Computing, Birkhäuser, Basel, 2015.
[3] Bjerhammar, A.: Gravity Reduction to an Internal Sphere. Division of Geodesy,
Stockholm, Sweden, 1962.
[4] Bitzadse, A.V.: Boundary-Value Problems for Second-Order Elliptic Equations.
North-Holland, Amsterdam, 1968.
[5] Colombo, O.L.: The Convergence of the External Spherical Harmonic Expansion of
the Gravitational Potential. Boll. Geod. Sci. Affin., 42, 221–239, 1983.
[6] Davis, P.J.: Interpolation and Approximation. Blaisdell Publishing Company, Walt-
ham, MA, 1963.
[7] Freeden W.: Über eine Klasse von Integralformeln der Mathematischen Geodäsie.
Veröff. Geod. Inst. RWTH Aachen, Report 27, 1979.
[8] Freeden, W.: On the Approximation of External Gravitational Potential with Closed
Systems of (Trial) Functions. Bull. Géod., 54, 1–20, 1980.
[9] Freeden, W.: On Spherical Spline Interpolation and Approximation. Math. Meth. in
the Appl. Sci., 3, 551–575, 1981.
[10] Freeden, W.: On Approximation by Harmonic Splines. Manuscr. Geod., 6, 193-244,
1981.
[11] Freeden, W.: Least Squares Approximation by Linear Combinations of (Multi)-Poles.
Scientific Report, No. 344, The Ohio State University, Department of Geodetic Sci-
ence, Columbus, 1983.
[12] Freeden, W.: A Spline Interpolation Method for Solving Boundary Value Problems
of Potential Theory from Discretely Given Data. Math. Part. Diff. Equations, 3,
375–398, 1987.
[13] Freeden, W.: Multiscale Modelling of Spaceborne Geodata. B.G. Teubner, Leipzig,
1999.
[14] Freeden, W.: The Uncertainty Principle and Its Role in Physical Geodesy. Progress
in Geodetic Science, 225–237, Shaker, Aachen, 1999.
[15] Freeden, W., Gerhards, C.: Geomathematically Oriented Potential Theory. Chapman
and Hall/CRC Press, Boca Raton, 2013.
[16] Freeden, W., Gutting, M.: Special Functions of Mathematical (Geo-)Physics. Birk-
häuser, Basel, 2013.
Surface of the Earth. Veröff. Geod. Inst. RWTH Aachen, No. 29, 1980.
Derivative Problem in Potential Theory. Math. Meth. Appl. Sci. 3, 104–114, 1981.
[19] Freeden, W., Kersten, H.: An Extended Version of Runge’s Theorem. Manuscr.
Geod., 7, 267–278, 1982.
[20] Freeden, W., Mayer, C.: Wavelets Generated by Layer Potentials. Appl. Comput.
Harm. Anal. (ACHA), 14, 195–237, 2003.
[21] Freeden, W., Mayer, C.: Multiscale Solution for the Molodenski Problem on Regular
Telluroidal Surfaces. Acta Geod. Geophys. Hung., 41, 55–86, 2006.
[22] Freeden, W., Mayer, C.: Modeling Tangential Vector Fields on Regular Surfaces by
Means of Mie Potentials. Int. J. Wavel. Multires. Inf. Process. 5, 417–449, 2007.
[24] Freeden, W., Reuter, R.: A Constructive Method for Solving the Displacement
Boundary-Value Problem of Elastostatics by Use of Global Basis Systems. Math.
Meth. Appl. Sci., 12, 105–128, 2011.
[25] Freeden, W., Schneider, F.: Runge–Walsh Wavelet Approximation for the Helmholtz
Equation. J. Math. Anal. Appl., 235, 533-566, 1999.
[26] Freeden, W., Schreiner, M.: Spherical Functions of Mathematical Geosciences (a
Scalar, Vectorial, and Tensorial Setup). Springer, Heidelberg, 2009.
[27] Freeden, W., Wolf, K.: Klassische Erdschwerefeldbestimmung aus der Sicht moderner
Geomathematik. Math. Semesterb., 56, 53–77, 2008.
[28] Freeden, W., Mayer, C., Schreiner, M.: Tree Algorithms in Wavelet Approximation
by Helmholtz Potential Operators. Num. Funct. Anal. Optim., 24, 747–782, 2003.
[29] Freeden, W., Fehlinger, T., Klug, M., Mathar, D., Wolf, K.: Classical Globally Re-
flected Gravity Field Determination in Modern Locally Oriented Multiscale Frame-
work. J. of Geodesy, 83, 1171–1191, 2009.
[30] Gauss, C.F.: Allgemeine Theorie des Erdmagnetismus. Resultate aus den Beobach-
tungen des magnetischen Vereins, Göttingen, 1838.
[31] Golberg, M.: The Method of Fundamental Solutions for Poisson’s Equation. Eng.
Anal. Bound. Elem. 16, 205–213, 1995.
[32] Golberg, M., Chen, C.: The Method of Fundamental Solutions for Potential,
Helmholtz, and Diffusion Problems. In: Boundary Integral Methods: Numerical and
Mathematical Aspects (Goldberg, M., ed.), 103–176, WIT Press, Computational Me-
chanics Publications, Southampton, 1998.
[33] Grafarend, E.W.: The Reference Figure of the Rotating Earth in Geometry and
Gravity Space and an Attempt to Generalize the Celebrated Runge–Walsh Approx-
imation Theorem for Irregular Surfaces. GEM Int. J. Geomath., 6, 101–140, 2015
[34] Grafarend, E.W., Klapp, M., Martinec. Z.: Spacetime Modelling of the Earth’s Grav-
ity Field by Ellipsoidal Harmonics. In: Handbook of Geomathematics, Vol. 1 (Freeden,
W., Nashed, M.Z., Sonar, T., Eds.), 2nd edition, 381–496, Springer, New York, 2015.
[35] Grothaus, M., Raskop, T.: Oblique Stochastic Boundary-Value Problem. In: Hand-
book of Geomathematics, Vol. 3 (Freeden, W., Nashed, M.Z., Sonar, T., Eds.), 2nd
edition, 2285–2316, Springer, New York, 2015.
[36] Gutting, M., Fast Multipole Methods for Oblique Derivative Problems. Ph.D.-thesis,
Geomathematics Group, University of Kaiserslautern, Shaker, Aachen, 2007.
[37] Gutting, M.: Fast Multipole Accelerated Solution of the Oblique Derivative Bound-
ary-Value Problem. GEM Int. J. Geomath., 3, 223–252, 2012.
[38] Gutting, M.: Fast Spherical/Harmonic Spline Modeling. In: Handbook of Geomathe-
matics, Vol. 3 (Freeden, W., Nashed, M.Z., Sonar, T., Eds.), 2nd edition, 2711–2746,
Springer, New York, 2015.
[39] Heiskanen, W.A., Moritz, H.: Physical Geodesy. W.H. Freeman and Company (1967).
[40] Hörmander, L.: The Boundary Problems of Physical Geodesy. The Royal Institute of
Technology, Division of Geodesy, Stockholm, Report 9, 1975 (also: Arch. Rat. Mech.
Anal., 62, 1–52, 1976).
pany, New York, 1929.
[42] Kita, E., Kamiya, N.: Trefftz Method: An Overview. Adv. Eng. Softw., 24, 3–13,
1995.
Problem Using the Known Surface of the Earth. Bull. Géod., 106, 467–476, 1972.
Danish Geodetic Institute, Report No. 44, Copenhagen, 1969.
[45] Krarup, T.: Letters on Molodensky’s Problem I–IV. Communication to the members
of the IAG Special Study Group 4.31, 1973. In: Mathematical Foundation of Geodesy:
Selected Papers of Torben Krarup (K. Borre, ed.), Springer, Berlin, Heidelberg, 2006
[46] Kupradze, V., Aleksidze M.: The Method of Functional Equations for the Approx-
imate Solution of Certain Boundary Value Problems. USSR Comp. Math. Math.
Phys., 4, 82–126, 1964.
[47] Laplace, P.S. de: Théorie des attractions des sphéroides et de la figure des planètes,
Mèm. de l’Acad., Paris, 1785.
[48] Legendre, A.M.: Sur l’attraction des Sphéroides. Mém. Math. et Phys. présentés á
l’Ac. r. des. sc. par divers savants, 10, 1785.
[49] Litzenberger, R.: Pyramid Schemes for Harmonic Wavelets in Boundary-Value Prob-
lems. Ph.D.-thesis, Geomathematics Group, University of Kaiserslautern, Shaker,
Aachen 2001.
[50] Magnus, W., Oberhettinger, F., Soni, R.P.: Formulas and Theorems for the Spe-
cial Functions of Mathematical Physics. In: Die Grundlehren der Mathematischen
Wissenschaften in Einzeldarstellungen, Band 52, Springer, Berlin, 3. Auflage, 1966.
[51] Mayer, C., Freeden, W.: Stokes Problem, Layer Potentials and Regularizations, and
Multiscale Applications. In: Handbook of Geomathematics, Vol. 2 (Freeden, W.,
Nashed, M.Z., Sonar, T., Eds.), 2nd edition, 1155–1254, Springer, New York, 2015.
[52] Martinec, Z.: Boundary-Value Problems for Gravimetric Determination of a Precise
Geoid. Springer, Berlin, Heidelberg, New York, 1999.
[53] Miranda, C.: Partial Differential Equations of Elliptic Type. Springer, Berlin, 1970.
[54] Moritz, H.: Recent Developments in the Geodetic Boundary Value Problem. Scien-
tif. Report, No. 266, The Ohio State University, Department of Geodetic Science,
Columbus, Ohio, 1977.
[55] Moritz, H.: Advanced Physical Geodesy. Wichmann Verlag, Karlsruhe, 1980.
[56] Moritz, H.: Classical Physical Geodesy. In: Handbook of Geomathematics, Vol. 1
(Freeden, W., Nashed, M.Z., Sonar, T., Eds.), 2nd edition, 253–289, Springer, New
York, 2015.
[57] Moritz, H., Sünkel, H., Eds.: Approximation Methods in Geodesy: Lectures delivered
at 2nd International Summer School in the Mountains on Mathematical Methods in
Physical Geodesy, Wichmann, Karlsruhe, 1978.
[58] Müller, C.: Spherical Harmonics, Lecture Notes in Mathematics. Vol. 17, Springer,
Berlin, 1966.
[59] Müller C.: Foundations of the Mathematical Theory of Electromagnetic Waves.

Springer, Berlin, Heidelberg, New York, 1969.
[60] Neumann, F.: Vorlesungen über die Theorie des Potentials und der Kugelfunktionen.
B.G. Teubner, Leipzig, 1887.
[61] Ritz, W.: Über eine neue Methode zur Lösung gewisser Variationsprobleme der
mathematischen Physik. J. Reine Angew. Math., 135, 1–6, 1909.
[62] Runge, C.: Zur Theorie der eindeutigen analytischen Funktionen. Acta Math., 6,
229–234, 1885.
[63] Sanso, F.: A Note on Density Problems and the Runge Krarup’s Theorem. Boll.
Geod. Sci. Affin., 41, 422–477, 1982.
Phys. Earth Planet. Inter., 28, 215–229, 1982.
[65] Sommerfeld, A.: Vorlesungen der Theoretischen Physik II. 6th edition, Verlag Harri
Deutsch, Frankfurt, 1978.
[66] Thomson, W.: On Electrical Images, Report of the 17th Meeting of the British Asso-
ciation for the Advancement of Science. Oxford, June 1847, John Murray, London,
1848. Notices and Abstracts of Miscellaneous Communications to the Sections, 6,
John Murray, London, 1848.
[67] Trefftz, E.: Ein Gegenstück zum Ritzschen Verfahren. In: Verh. d. 2. Intern. Kongr.
f. Techn. Mech., 131–137, Zürich, 1926.
[68] Tscherning, C.C., Rapp, R.H.: Closed Covariance Expressions for Gravity Anom-
alies, Geoid Undulations, and Deflections of the Vertical Implied by Anomaly Degree
Variance Models, Scientific Report, No. 208, The Ohio State University, Department
of Geodetic Science, Columbus, 1974.
[69] Vekua, I.N.: Über die Vollständigkeit des Systems harmonischer Polynome im Raum.
Dokl. Akad. Nauk, 90, 495–498, 1953.
[70] Walsh, J.L.: The Approximation of Harmonic Functions by Harmonic Polynomials
and by Harmonic Rational Functions. Bull. Amer. Math. Soc., 35, 499–544, 1929.
[71] Wangerin, A.: Theorie des Potentials und der Kugelfunktionen. 1884, published by
G.J. Göschen, Leipzig, 1921.
[72] Weyl, H.: Über die Gleichverteilung von Zahlen mod. Eins. Math. Ann., 77, 313–352,
1916.
Matthias Augustin
Mathematical Image Analysis Group
Saarland University
Fakultät 6
D-66123 Saarbrücken, Germany


Geomathematical Advances in
Satellite Gravity Gradiometry (SGG)
Willi Freeden, Helga Nutz, and Michael Schreiner
Abstract. A promising technique of globally establishing the fine structure

and the characteristics of the external Earth’s gravitational field is satellite
gravity gradiometry (SGG). Satellites such as ESA’s gradiometer satellite
GOCE are able to provide sufficiently large data material of homogeneous
quality and accuracy. In geodesy, traditionally the external Earth’s gravita-
tional potential and its Hesse matrix are described using orthogonal (Fourier)
expansions in terms of (outer) spherical harmonics. Spherical and outer har-
monics are introduced for the global modeling of (scalar / tensor) fields. We
briefly recapitulate the results interconnecting spherically the potential co-
efficients with respect to tensor spherical harmonics at Low Earth Orbiter’s
(LEO) altitude to the corresponding coefficients with respect to scalar spher-
ical harmonics at the Earth’s surface. The relation between the known tenso-
rial measurements g (i.e., gradiometer data) and the gravitational potential
F on the Earth’s surface is expressed by a linear integral equation of the first
kind. This operator equation is discussed in the framework of pseudodiffer-
ential operators as an invertible mapping between Sobolev spaces under the
assumption that the data are not erroneous. In reality, however, the data g
are noisy such that the Sobolev reference space for the (noisy) tensorial data
g must be embedded in a larger Sobolev space. Under these conditions, we
base our inversion process on the fact that the reference Sobolev subspace
is dense in the larger Sobolev space and that, e.g., a smoothing spline pro-
cess or a signal-to-noise procedure in multiscale framework open appropriate
perspectives to approximate F (in suitable accuracy) from noisy data g.
Keywords. Tensorial spherical framework, satellite gravity gradiometry prob-

lem, multiscale spline and wavelet modeling.
1. Introduction
Due to the non-spherical shape, the irregularities of the interior mass distribution,
and the movement of the lithospheric plates as well as volcanic and magmatic
activities, the external gravitational field of the Earth shows significant local vari-
562 W. Freeden, H. Nutz, and M. Schreiner
ations. The recognition of the microstructure of the Earth’s external gravitational

potential is of tremendous importance for a large palette of geoscientific questions,
for example, studies of the processes in the Earth’s interior, models of sea sur-
face topography and circulations of the ocean, and investigations of the (global)
climate change. An important measurement technique to determine locally the
fine structure of the external gravitational field is terrestrial gravimetry, i.e., the
reconstruction of the density variations inside and on the Earth’s surface from
the gravitational potential and its functionals. A promising technique of globally
establishing the fine structure and the characteristics of the external Earth’s gravi-
tational field is satellite gravity gradiometry (SGG) . Its principle can be described
roughly as follows: Several test masses in a low Earth’s orbiting (LEO)-satellite
feel – due to their distinct positions and the local changes of the gravitational
field – different forces, thus yielding different accelerations. The measurements of
the relative accelerations between two test masses provide information about the
second-order partial derivatives of the gravitational potential at flight position.
Assuming an ideal situation, the full Hesse matrix can be observed globally above
the Earth by an array of test masses.
The traditional way to describe the external Earth’s gravitational potential
and its Hesse matrix of second derivatives is to use orthogonal (Fourier) expansions
with (outer) spherical harmonics as reference trial system, a procedure dating back
to C.F. Gauss in the nineteenth century. During the last decades many gravita-
tional models have been enhanced based on the (until now unrealistic) assumption
of continuously improvable terrestrial gravitational observations (with respect to
both in quality and quantity). But the actual non-uniform distribution and the
strong heterogeneity of the data material have set a limit for an intensive im-
provement of such spherically based (outer) harmonic models down to local scale.
A tremendous step forward in the measurement of data for global modeling was
provided by modern satellites such as ESA’s gradiometer satellite GOCE (mission
duration: 17 March 2009 till 11 November 2013). This satellite was able to provide
sufficiently large data material of homogeneous quality and accuracy, at least for
the diagonal components of the Hesse matrix. However, the great drawback of ac-
quiring gravitational data at LEO’s altitude is that the upward continuation of the
gravitation amounts to an exponential spectral smoothing of the potential coeffi-
cients in terms of outer harmonics with increasing height. In other words, satellite
measurements do not contain the same signal information at LEO’s height (i.e.,
200–250 km) as on the Earth’s surface. This is the reason why the gravitational po-
tential is obtainable from satellite data only in an attenuated form when continued
to the Earth’s surface. Even more, it may happen in downward continuation that
the noise in the measurements is amplified. Nevertheless, for satellite gravity gra-
diometry (SGG), as provided by GOCE, an advantage can be taken from the fact
that second derivatives instead of the potential itself are used as observations on
LEO’s orbit. Mathematically, this means that the exponential decay of the outer
harmonic coefficients is reduced polynomially by two degrees. In other words, SGG
takes advantage of the fact that second derivatives produce a rougher data set than
Geomathematical Advances in SGG 563
the potential itself such that the resolution of the gravitational structure is much
finer. In addition, in the frequency context of outer harmonics, the Meissl scheme
(see, e.g., [10, 20, 26, 31]) enables us in spectral nomenclature to relate the or-
thogonal coefficients at LEO’s height to the orthogonal coefficients at the surface
of the Earth, at least in the context of a spherical model and under the restrictive
assumption of bandlimited outer harmonic modeling without observational errors.
In this contribution the relation between the known tensorial measurements
g (i.e., gradiometer data) and the gravitational potential F on the Earth’s surface
is expressed by the linear integral equation of the first kind

ΛR;r
SGG F (x) = ∇x ⊗ ∇x KΛup (x, y) F (y) dωR (y) = g(x), x ∈ Ωr ,
ΩR
where ΩR and Ωr are the spheres with radii R and r, respectively, ∇x ⊗∇x denotes
the Hesse tensor and KΛup is the Abel–Poisson kernel for the upward continuation
(more details are explained in Section 4). This operator equation is discussed in
the framework of pseudodifferential operators in, e.g., [9] as an invertible mapping
between Sobolev spaces under the assumption that the data are not erroneous. In
reality, however, the data g are noisy such that the Sobolev reference space for the
(noisy) tensorial data g must be embedded in a larger Sobolev space. Under these
conditions, we base our inversion process on the fact that the reference Sobolev
subspace is dense in the larger Sobolev space and that, e.g., a smoothing spline
process or a signal-to-noise procedure in multiscale framework (see [7, 8]) open
perspectives to approximate F (in suitable accuracy) from noisy data g.
Our work yields a modified wavelet approach for regularization of the inverse
gradiometer problem based on ideas presented in [9]. Moreover a tree algorithm
for multiscale decorrelation of the Earth’s external gravitational potential is in-
troduced. In adequate consistency with the reality the spaceborne gradiometer
data are assumed to be of tensorial nature. As an essential tool tensorial radial
basis functions (see, e.g., [12, 14, 26]) are used for multiscale regularization of the
exponentially ill-posed downward continuation of satellite gradiometer data.
2. Potential theoretic aspects

Gravity as observed on the Earth’s surface is the combined effect of the gravita-
tional mass attraction and the centrifugal force due to the Earth’s rotation. The
force of gravity provides a directional structure to the space above the Earth’s
surface. It is tangential to the vertical plumb lines and perpendicular to all level
surfaces. Any water surface at rest is part of a level surface. As if the Earth were a
homogeneous, spherical body gravity turns out to be always constant all over the
Earth’s surface. The plumb lines are directed towards the Earth’s center of mass,
and this implies that all level surfaces are nearly spherical, too.
However, the gravity shows local (temporal and spatial) variations due to
mass density inhomogeneities and temporal mass rearrangement in the Earth’s
interior which enable the investigation of geological structures as fault zone, geo-
logical faults, salt domes and volcanic formations as well as mineral deposits. The
strongest variations in the gravity are caused by the periodically changing positions
of moon and sun relative to the Earth and the associated deformations of the body
of the Earth (ocean and Earth tide and the induced loading effects).
The geoid is a virtual surface shaped by the gravity field of the Earth in the
absence of external influences such as winds and tides. The level surfaces are ideal
reference surfaces, for example, for heights. In more detail, the gravity acceleration
(gravity) w is the resultant of gravitation v and centrifugal acceleration c, i.e.,
w = v + c. The centrifugal force c arises as a result of the rotation of the Earth
around its axis. We assume here a rotation of constant angular velocity around the
rotational axis x3 , which is further assumed to be fixed with respect to the Earth.
The direction of the gravity w is known as the direction of the plumb line, the
quantity |w| is called the gravity intensity (often just gravity). The gravity potential
of the Earth can be expressed in the form: W = V + C. The gravity acceleration
w is given by w = ∇W = ∇V + ∇C. The surfaces of constant gravity poten-
tial W (x) = const, x ∈ R3 , are designated as equipotential (level, or geopotential)
surfaces of gravity. The gravity potential W of the Earth is the sum of the gravita-
tional potential V and the centrifugal potential C, i.e., W = V + C. In an Earth’s
fixed coordinate system the centrifugal potential C is explicitly known. Hence, the
determination of equipotential surfaces of the potential W is strongly related to
the knowledge of the potential V . The gravity vector w given by w(x) = ∇x W (x),
where the point x ∈ R3 is located outside or on a sphere around the origin with
Earth’s radius R, is normal to the equipotential surface passing through the same
point. Thus, equipotential surfaces intuitively express the notion of tangential sur-
faces, as they are normal to the plumb lines given by the direction of the gravity
vector (for more details see, e.g., [17]).
A tremendous step forward in the measurement of data for global model-
ing was provided with modern satellites such as CHAMP (2000–2010), GRACE
(launch 2002; designed for a mission lifetime of five years, GRACE is currently op-
erating in an extended mission phase suffering from aging of the components, e.g.,
health of the batteries), and GOCE (2009–2013). These satellites yield sufficiently
large data material of homogeneous quality and accuracy. However, as already
pointed out, the great drawback of acquiring gravitational data at LEO’s altitude
is that the gravitational potential at ground level is obtainable from satellite data
only in a “rough” form if continued downward to the Earth’s surface. Even more,
it is unavoidable for ill-posed problems such as downward continuation that the
noise in the measurements is amplified. For SGG, however, a certain compensation
effect can be taken from the fact that second derivatives instead of the potential
itself are used. As already mentioned, this mathematically means that the expo-
nential decay of the outer harmonic coefficients is reduced polynomially by two
degrees. In other words, SGG takes advantage of the fact that second derivatives
produce a rougher data set than the potential itself such that the resolution of the
gravitational structure is much finer. A mathematical model relating observables
in geodesy to each other in the frequency context of outer harmonics is the Meissl
scheme (see, e.g., [20, 21, 26, 28–31]). By specifying the spectral properties of the
Fourier coefficients, this scheme enables us to relate the orthogonal coefficients at
LEO’s height to the orthogonal coefficients at the Earth’s surface, at least in the
context of a spherical model and under the restrictive assumption of bandlimited
outer harmonic modeling.
Spherical notation
We begin by introducing some basic notation that will be used throughout our
work: Let x, y, . . . represent the elements of the Euclidean space R3 . For all x ∈
R3 , x = (x1 , x2 , x3 )T , different from the origin, we have x = rξ, r = |x| =
#
x21 + x22 + x23 , where ξ = (ξ1 , ξ2 , ξ3 )T is the uniquely determined directional unit
vector of x ∈ R3 . The unit sphere in R3 will be denoted by Ω, whereas Ωα desig-
nates the sphere around the origin with radius α. If the vectors ε1 , ε2 , ε3 form the
canonical orthonormal basis in R3 , we may represent ξ ∈ Ω in polar coordinates by
#
ξ = tε3 + 1 − t2 cos ϕε1 + sin ϕε2 , −1 ≤ t ≤ 1, 0 ≤ ϕ < 2π, t = cos θ. (1)
Inner, vector, and dyadic (tensor) product of two vectors x, y ∈ R3 , respectively,

are denoted by x · y, x ∧ y and x ⊗ y. As usual, a second-order tensor f ∈ R3 ⊗ R3
is understood to be a linear mapping that assigns to each x ∈ R3 a vector y ∈ R3 .
The (cartesian) components Fij of f are defined by Fij = εi · (f εj ) = (εi )T (f εj ), so

that y = f x is equivalent to y · εi = 3j=1 Fij (x · εj ). We write f T for the transpose
of f ; it is the unique tensor satisfying (f y) · x = y · (f T x) for all x, y ∈ R3 . The
dyadic (tensor) product x ⊗ y of two elements x, y ∈ R3 is the tensor that assigns
to each u ∈ R3 the vector (y · u)x. More explicitly, (x ⊗ y)u = (y · u)x for every
u ∈ R3 . The innerproduct f · g of two second-order tensors f , g ∈ R3 ⊗ R3 is
3
defined by f · g = i,j=1 Fij Gij , while |f | = (f · f )1/2 is called the norm of f .
Next we come to some differential operators, which are of particular impor-
tance in the tensorial context. In terms of polar coordinates (1) the gradient ∇
in R3 allows the representation ∇x = ξ∂/∂r + (1/r)∇∗ξ , where ∇∗ is the surface
gradient of the unit sphere Ω ⊂ R3 . The operator Δ∗ = ∇∗ · ∇∗ is called the
Beltrami operator of the unit sphere Ω. Obviously, it can be understood as the
angular part of the Laplace operator.
Note that, throughout this paper, scalar-valued (resp. vector-valued, tensor-
valued) functions are denoted by capital (resp. small, small bold) letters. A function
F : Ω → R (resp. f : Ω → R3 , f : Ω → R3 ⊗R3 ) possessing k continuous derivatives
on the unit sphere Ω is said to be of class C (k) (Ω) (resp. c(k) (Ω), c(k) (Ω)). C (0) (Ω)
(resp. c(0) (Ω), c(0) (Ω)) is the class of real continuous scalar-valued (resp. vector-
valued, tensor-valued) functions on Ω. For F ∈ C (1) (Ω) we introduce the surface
curl gradient L∗ξ by L∗ξ F (ξ) = ξ ∧ ∇∗ξ F (ξ), ξ ∈ Ω, while ∇∗ξ · f (ξ), ξ ∈ Ω, and
L∗ξ · f (ξ), ξ ∈ Ω, respectively, denote the surface divergence and surface curl of the
vector field f at ξ ∈ Ω. For more details the reader is referred to [5].
Scalar spherical harmonics

Scalar spherical harmonics are defined as restrictions of homogeneous harmonic
polynomials to the unit sphere Ω. In all geosciences interested in global modeling,
spherical harmonics are the functions which are usually taken to represent scalar
fields on a spherical surface such as the Earth’s (mean) sphere.
Definition 2.1. Let Hn be a homogeneous harmonic polynomial of degree n in
R3 , n ∈ N0 , i.e., Hn ∈ Harm n (R3 ). The restriction Yn = Hn |Ω is called (scalar)
spherical harmonic of degree n. The space {Yn | Yn = Hn |Ω , Hn ∈ Harm n (R3 )}
of all (scalar) spherical harmonics of degree n is denoted by Harm n (Ω).
Harm n (Ω) is known to be of dimension 2n+1. Spherical harmonics of different
degrees are orthogonal in L2 (Ω)-sense, that is

(Yn , Yñ )L2 (Ω) = Yn (ξ)Yñ (ξ) dω(ξ) = 0, n = ñ,
Ω
where dω is the surface element on Ω. Throughout this text a capital letter Y

followed by one or two indices always denotes a spherical harmonic of the degree
given by the first index and order given by the second index. Two indices mean that
the function, for example Yn,m , is a member of an L2 (Ω)-orthonormal system of
functions {Yn,1 , . . . , Yn,2n+1 }n∈N0 . By use of the scalar spherical harmonics every
function F ∈ L2 (Ω) can be written as a Fourier series
∞ 2n+1

F = F ∧L2 (Ω) (n, m)Yn,m ,
n=0 m=1
2
(in L (Ω)-sense) with Fourier coefficients

∧L2 (Ω)
F (n, m) = (F, Yn,m )L2 (Ω) = F (η)Yn,m (η)dω(η).
Ω
The system {Yn,m }n=0,1,...;m=1,...,2n+1 is closed in C(Ω) with respect to the norm
· C(Ω) , i.e., for any number ε > 0 and any function F ∈ C(Ω), there exists a
N 2n+1
linear combination FN = n=0 m=1 dn,m Yn,m such that F − FN C(Ω) ≤ ε.
The system {Yn,m }n=0,1,...;m=1,...,2n+1 is, furthermore, complete in L2 (Ω) with
respect to (·, ·)L2 (Ω) , i.e., F ∈ L2 (Ω) with F ∧L2 (Ω) (n, m) = 0 for all n = 0, 1, . . .;
m = 1, . . . , 2n + 1 implies F = 0 (see, e.g., [14]).
Theorem 2.2 (Addition Theorem for Scalar Spherical Harmonics). Let the system
{Yn,m }m=1,...,2n+1 be an L2 (Ω)-orthonormal one in Harm n (Ω). Then, for any pair
(ξ, η) ∈ Ω2 , the addition theorem reads
2n+1
2n + 1
Yn,m (ξ)Yn,m (η) = Pn (ξ · η),
m=1
4π
where Pn : [−1, 1] → [−1, 1] is the Legendre polynomial of degree n.

Tensor spherical harmonics

By l2 (Ω) we denote the Hilbert space of square-integrable tensor fields f : Ω →
R3 ⊗ R3 with the inner product

(f , g)l2 (Ω) = f (ξ) · g(ξ) dω(ξ), f , g ∈ l2 (Ω),
Ω
and the associated norm · l2 (Ω) . Note that the space l2 (Ω) is the completion of
c(Ω) with respect to the norm · l2 (Ω) . The operators o(i,k) : C (∞) (Ω) → c(∞) (Ω),
i, k = 1, 2, 3, transform scalar functions into tensor fields (cf. [12]):
o(1,1) F (ξ) = ξ ⊗ ξF (ξ),
o(1,2) F (ξ) = ξ ⊗ ∇∗ξ F (ξ),
o(1,3) F (ξ) = ξ ⊗ L∗ξ F (ξ),
o(2,1) F (ξ) = ∇∗ξ F (ξ) ⊗ ξ,
o(3,1) F (ξ) = L∗ξ F (ξ) ⊗ ξ,
o(2,2) F (ξ) = itan (ξ)F (ξ),

o(2,3) F (ξ) = ∇∗ξ ⊗ ∇∗ξ − L∗ξ ⊗ L∗ξ F (ξ) + 2∇∗ξ F (ξ) ⊗ ξ,

o(3,2) F (ξ) = ∇∗ξ ⊗ L∗ξ + L∗ξ ⊗ ∇∗ξ F (ξ) + 2L∗ξ F (ξ) ⊗ ξ,
o(3,3) F (ξ) = jtan (ξ)F (ξ),
F ∈ C(2) (Ω), ξ ∈ Ω. Note that the tensors itan = i − ξ ⊗ ξ and jtan = ξ ∧ i are the
surface identity tensor and the surface rotation tensor, respectively. The adjoint
operators O(i,k) to o(i,k) satisfying
(i,k)
o F, f l2 (Ω) = F, O(i,k) f L2 (Ω)
for F ∈ C (2) (Ω) and f ∈ c(2) (Ω) are given by

O(1,1) f (ξ) = ξ T f (ξ)ξ,

O(1,2) f (ξ) = − ∇∗ξ · ptan ξ T f (ξ) ,

O(1,3) f (ξ) = − L∗ξ · ptan ξ T f (ξ) ,
O(2,1) f (ξ) = − ∇∗ξ · ptan (f (ξ)ξ) ,
O(3,1) f (ξ) = − L∗ξ · ptan (f (ξ)ξ) ,
O(2,2) f (ξ) = itan (ξ) · f (ξ),

O(2,3) f (ξ) = ∇∗ξ · ptan ∇∗ξ · ptan,∗ f (ξ) − L∗ξ · ptan L∗ξ · ptan,∗ f (ξ)
− 2∇∗ξ · ptan (f (ξ)ξ) ,

O(3,2) f (ξ) = L∗ξ · ptan ∇∗ξ · ptan f (ξ) + ∇∗ξ · ptan L∗ξ · ptan f (ξ)
− 2L∗ξ · ptan (f (ξ)ξ) ,
O(3,3) f (ξ) = jtan (ξ) · f (ξ),
where
ptan f (ξ) = f (ξ) − (ξ · f (ξ))ξ,
ptan,∗ f (ξ) = f (ξ) − ξ ⊗ ((f (ξ))T ξ),
ξ ∈ Ω. With the help of the operators o(i,k) we are able to define a set of tensor
(i,k)
spherical harmonics {yn,m }i,k=1,2,3; n=0ik ,...; m=1,...,2n+1 by setting
−1/2
(i,k)
yn,m = μ(i,k)
n o(i,k) Yn,m , (2)
(i,k)
where the normalization constants μn are given by
⎧
⎪
⎪ 1, (i, k) = (1, 1),
⎨
(i,k) 2, (i, k) ∈ {(2, 2), (3, 3)},
μn =
⎪
⎪ n(n + 1), (i, k) ∈ {(1, 2), (1, 3), (2, 1), (3, 1)},
⎩
2n(n + 1)(n(n + 1) − 2), (i, k) ∈ {(2, 3), (3, 2)}.
For simplicity, we use the abbreviation
⎧
⎨ 0, (i, k) ∈ {(1, 1), (2, 2), (3, 3)},
0ik = 1, (i, k) ∈ {(1, 2), (1, 3), (2, 1), (3, 1)},
⎩
2, (i, k) ∈ {(2, 3), (3, 2)}.
By harm(i,k)n (Ω) we denote the space of all tensor spherical harmonics of de-
(i,k)
gree n and kind (i, k). If {yn,m }m=1,...,2n+1 is an l2 (Ω)-orthonormal basis of
(i,k)
harmn (Ω), then the tensorial addition theorem reads
2n+1

(i,k) (p,q) 2n + 1 (i,k,p,q)
yn,m (ξ) ⊗ yn,m (η) = Pn (ξ, η),
m=1
4π
(i,k,p,q)
i, k, p, q ∈ {1, 2, 3}, where Pn : Ω × Ω → R3 ⊗ R3 ⊗ R3 ⊗ R3 denote the
Legendre tensors of degree n defined by
−1/2 −1/2
(i,k)
P(i,k,p,q)
n = μ(i,k)
n μ(p,q)
n oξ o(p,q)
η Pn (ξ · η), ξ, η ∈ Ω,
(for explicit representations see [12]). Note that, for sufficiently smooth tensor
fields f : Ω → R3 ⊗ R3 of the form
3

f (ξ) = Fi,k (ξ)εi ⊗ εk , x ∈ Ω,
i,k=1
we set 3
(p,q)
(p,q)
oξ f (ξ) = oξ Fi,k (ξ) ⊗ εi ⊗ εk .
i,k=1
By harmn (Ω) we denote the space of all tensor spherical harmonics of degree n.
(i,k)
We have defined the system {yn,m } of tensor spherical harmonics concen-
trating on the fact that the decomposition into normal and tangential tensor fields
is fulfilled (cf. [26]). But one disadvantage of this set of tensor spherical harmonics
is that these functions are no eigenfunctions of the (scalar) Beltrami operator. As

turns out, this property enables us to define so-called outer harmonics in such
a way that they fulfill the Laplace equation in the outer space. To this end we
introduce the operators õ(i,k) : C (∞) (Ω) → c(∞) (Ω), i, k = 1, 2, 3, based on the
operators o(i,k) and O(i,k) by
⎛ (1,1) ⎞ ⎛ ⎞
õ Yn Yn
⎜ õ(1,2) Yn ⎟ ⎜ Yn ⎟
⎜ (2,1) ⎟ ⎜ ⎟
⎜ õ Yn ⎟ = aD ⎜
⎟ ⎟
⎜ ⎜ Yn ⎟ ,
⎝ õ(2,2) Yn ⎠ ⎝ Yn ⎠
õ(3,3) Yn Yn
and ⎛ ⎞ ⎛ ⎞
õ(1,3) Yn Yn
⎜ õ(2,3) Yn ⎟ ⎜ Yn ⎟
⎜ ⎟ ⎜ ⎟,
⎝ õ(3,1) Yn ⎠ = bD ⎝ Yn ⎠
õ(3,2) Yn Yn
where the matrix operators aD and bD are defined by (see [12])
⎛ ⎞
o(1,1) (D + 1)(D + 2) −o(1,2) (D + 2) −o(2,1) (D + 2) − 12 o(2,2) (D + 2)(D + 1) 12 o(2,3)
⎜ o(1,1) D2 o(1,2) D −o(2,1) (D − 1) − 12 o(2,2) D(D − 1) − 12 o(2,3) ⎟
⎜ ⎟
aD =⎜
⎜ o(1,1) (D + 1)2 −o(1,2) (D + 1) o(2,1) (D + 2) 1 (2,2)
2o (D + 2)(D + 1) − 12 o(2,3) ⎟
⎟
⎝ o(1,1) D(D − 1) o(1,2) (D − 1) o(2,1) (D − 1) − 12 o(2,2) D(D − 1) 1 (2,3) ⎠
2o
0 0 o(2,1) − 12 o(2,2) D(D + 1) − 12 o(2,3)
and
⎛ ⎞
o(1,3) (D + 1) o(3,1) − 12 o(3,2) − 12 o(3,3) D(D + 1)
⎜ o(1,3) D −o(3,1) 1 (3,2)
o 1 (3,3)
o D(D + 1) ⎟
bD =⎜
⎝
2 2 ⎟,
0 o(3,1) (D + 2) − 12 o(3,2) 1 (3,3)
2
o (D + 2)(D + 1) ⎠
1 (3,2)
0 o(3,1) (D − 1) 2
o − 12 o(3,3) D(D − 1)
and D is the pseudodifferential operator D = (−Δ + 14 )1/2 − 12 of order 1 satisfying

DYn = D∧ (n)Yn = nYn for all Yn ∈ Harm n (Ω).
The adjoint operators Õ(i,k) : c(∞) (Ω) → C (∞) (Ω), i, k = 1, 2, 3, to the
operators õ(i,k) satisfying the equation (õ(i,k) G, f )l2 (Ω) = (G, Õ(i,k) f )L2 (Ω) , f ∈
c(∞) (Ω), G ∈ C(∞) (Ω), are easily obtainable as follows
⎛ (1,1) ⎞ ⎛ ⎞
Õ Yn Yn
⎜ Õ(1,2) Yn ⎟ ⎜ Yn ⎟
⎜ ⎟ ⎜ ⎟
⎜ Õ(2,1) Yn ⎟ = aD ⎜ Yn ⎟ ,
⎜ ⎟ ⎜ ⎟
⎝ Õ(2,2) Yn ⎠ ⎝ Yn ⎠
Õ(3,3) Yn Yn
and
⎛ ⎞ ⎛ ⎞
Õ(1,3) Yn Yn
⎜ Õ(2,3) Yn ⎟ ⎜ Yn ⎟
⎜ ⎟ ⎜ ⎟.
⎝ Õ(3,1) Yn ⎠ = bD ⎝ Yn ⎠
Õ(3,2) Yn Yn
After these preliminaries we are now in the position to introduce the tensor spher-
ical harmonics
−1/2
(i,k)
ỹn,m = μ̃(i,k)
n õ(i,k) Yn,m , (3)
n = 0̃ik , . . . ; m = 1, . . . , 2n + 1, where we use the abbreviation
⎧
⎨ 0, (i, k) ∈ {(1, 1), (2, 1), (3, 1)},
0̃ik = 1, (i, k) ∈ {(1, 2), (1, 3), (2, 3), (3, 3)},
⎩
2, (i, k) ∈ {(2, 2), (3, 2)},
and the normalization constants
μ̃n(1,1) = (n + 2)(n + 1)(2n + 3)(2n + 1),
μ̃n(1,2) = 3n4 ,
μ̃n(1,3) = n(n + 1)2 (2n + 1),
μ̃n(2,1) = (n + 1)2 (2n + 3)(2n + 1),
μ̃n(2,2) = n(n − 1)(2n + 1)(2n − 1),
μ̃n(2,3) = n2 (n + 1)2 ,
μ̃n(3,1) = n2 (n + 1)(2n + 1),
μ̃n(3,2) = n(n + 1)2 (2n + 1),
μ̃n(3,3) = n2 (n − 1)(2n + 1).
According to this construction, in contrary to the system (2), each member of the
(i,k)
system {ỹn,m } is an eigenfunction of the Beltrami operator. More explicitly, we
have
Theorem 2.3. Let {Yn,m }n=0,1,...; m=1,...,2n+1 be an L2 (Ω)-orthonormal set of sca-
lar spherical harmonics. Then, the set

(i,k)
ỹn,m i,k=1,2,3; ,
n=0̃ik ,...;
m=1,...,2n+1
as defined by (3), forms an l2 (Ω)-orthonormal set of tensor spherical harmonics

which is closed in c(Ω) and l2 (Ω) with respect to · c(Ω) and · l2 (Ω) , respectively,
and complete in l2 (Ω) with respect to (·, ·)l2 (Ω) . Furthermore, we are able to verify
that
Δ∗ξ ỹn,m
(1,1) (1,1)
= −(n + 2)(n + 3)ỹn,m ,
Δ∗ξ ỹn,m
(1,2) (1,2)
= −n(n + 1)ỹn,m ,
Δ∗ξ ỹn,m
(2,1) (2,1)
= −n(n + 1)ỹn,m ,
Δ∗ξ ỹn,m
(2,2) (2,2)
= −(n − 1)(n − 2)ỹn,m ,
Δ∗ξ ỹn,m
(3,3) (3,3)
= −n(n + 1)ỹn,m ,
Δ∗ξ ỹn,m
(1,3) (1,3)
= −(n + 1)(n + 2)ỹn,m ,
Δ∗ξ ỹn,m
(2,3) (2,3)
= −n(n − 1)ỹn,m ,
Δ∗ξ ỹn,m
(3,1) (3,1)
= −(n + 1)(n + 2)ỹn,m ,
Δ∗ξ ỹn,m
(3,2) (3,2)
= −n(n − 1)ỹn,m ,
where the application of the Beltrami operator is understood component-by-compo-
nent.
Because of the completeness of the tensor spherical harmonics every tensor
field f ∈ l2 (Ω) can be written as a Fourier series
3
∞ 2n+1

f= f (i,k)∧l2 (Ω) (n, m)ỹn,m
(i,k)
i,k=1 n=0̃ik m=1
(in · l2 (Ω) -sense), where the Fourier coefficients are given by

(i,k)∧l2 (Ω) (i,k) (i,k)
f (n, m) = (f , ỹn,m )l2 (Ω) = f (ξ) · ỹn,m (ξ)dω(ξ).
Ω
For a more detailed introduction to the theory of scalar and tensor spherical har-
monics including the development of associated addition theorems the reader is
referred to [12].
Outer harmonics
Up to now, we assumed spherical geometry, i.e., we presented spherical harmonics
which are adequate for the conventional approach, where the reference surface
of the Earth is supposed to be a sphere and the data are assumed to be given
on a spherical satellite orbit. Next, we make the first steps to a Runge concept
by specifying two spheres as illustrated in Figure 1, thereby using the specific
properties of outer harmonics.
Figure 1. The geometric Runge concept underlying this paper.

Our idea can be explained as follows: Starting from the data given on the real
satellite orbit Γ (which is not necessarily required to build a closed surface) we pull
down the tensorial information to a sphere Ωτ of radius τ such that dist(Ωτ , Γ) > 0
by use of tensor outer harmonics. By virtue of “downward continuation” from Ωτ
to the sphere ΩR such that dist(ΩR , Σ) > 0 corresponding to the real Earth’s
surface Σ we are able to calculate the desired solution in terms of scalar outer
harmonics, i.e., the gravitational potential on the real Earth’s surface Σ from data
on the real orbit Γ. In consequence, we have to base our considerations on scalar
as well as tensor outer harmonics that are consistently related to each other.
Scalar outer harmonics. We begin our considerations with the introduction of the
scalar outer harmonics
n+1
1 R x
R
Hn,m (x) = Yn,m , x ∈ Ωext
R , n = 0, 1, . . . ; m = 1, . . . , 2n + 1.
R |x| |x|
They obey the following properties:
• Hn,m
R
is of class C (∞) (ΩextR ),
ext
• Hn,m is harmonic in ΩR : Δx Hn,m
R R
(x) = 0 for x ∈ Ωext
R ,
• Hn,m |ΩR = (1/R)Yn,m ,
R

• (Hn,m
R R
, Hl,s )L2 (ΩR ) = ΩR Hn,mR R
(x)Hl,s (x)dω(x) = δn,l δm,s ,
−1

• |Hn,m (x)| = O |x|
R
, |x| → ∞.
Accordingly, the space Harm n (Ωext
R ) is defined by
Harm n (Ωext
R ) = spanm=1,...,2n+1 (Hn,m ),
while Harm 0,...,n (Ωext

R ) denotes the space
5
n
Harm 0,...,n (Ωext
R )= Harm k (Ωext
R ).
k=0
Tensor outer harmonics. Next we introduce an associated class of tensor outer

harmonics by using tensor spherical harmonics as defined in (3)
n+3
R;(1,1) 1 R (1,1) x
hn,m (x) = ỹn,m , (4)
R |x| |x|
n+1
1 R (1,2) x
hR;(1,2) (x) = ỹ , (5)
n,m
R |x| n,m
|x|
n+2
R;(1,3) 1 R (1,3) x
hn,m (x) = ỹn,m , (6)
R |x| |x|
n+1
1 R (2,1) x
hR;(2,1) (x) = ỹ , (7)
n,m
R |x| n,m
|x|
n−1
R;(2,2) 1 R (2,2) x
hn,m (x) = ỹn,m , (8)
R |x| |x|
n
1 R (2,3) x
hR;(2,3) (x) = ỹn,m , (9)
n,m
R |x| |x|
n+2
1 R (3,1) x
hR;(3,1) (x) = ỹn,m , (10)
n,m
R |x| |x|
n
1 R (3,2) x
hR;(3,2) (x) = ỹ , (11)
n,m
R |x| n,m
|x|
n+1
R;(3,3) 1 R (3,3) x
hn,m (x) = ỹn,m , (12)
R |x| |x|
x ∈ Ωext
R , n = 0̃ik , . . . ; m = 1, . . . , 2n + 1. It is not difficult to show that the
following properties are satisfied:
• hn,m is of class c(∞) (Ωext
R;(i,k)
R ),
• Δx hn,m (x) = 0 for x ∈ Ωext
R;(i,k) R;(i,k)
R , i.e., the component functions of hn,m fulfill
the Laplace equation,
R;(i,k) (i,k)
• hn,m |ΩR = (1/R)ỹn,m ,
R;(i,k) R;(p,q) R;(i,k) R;(p,q)
• (hn,m ,hl,s )l2 (ΩR ) = ΩR hn,m (x) · hl,s (x)dωR (x) = δi,p δk,q δn,l δm,s ,
where l2 (ΩR ) is the space of square-integrable tensor fields on ΩR ,

• |hn,m (x)| = O |x|−1 , |x| → ∞.
R;(i,k)
It must be emphasized that the spherically reflected formulation of the tensorial

SGG problem exclusively uses the tensor outer harmonics of kind (1, 1) specified
by (4). However, for reasons of completeness of our tensor spherical approach, we
have listed all kinds of outer harmonics.
Sphere to sphere interconnection between SGG-data and the gravitational

potential
Next we are interested in characterizing the essential players involved in the SGG-
match as members of infinite-dimensional potential spaces. We begin with the
scalar space
−1
Pot (0) (Ωext
R ) = {F ∈ C
(2)
(Ωext ext
R ) : ΔF = 0 in ΩR , F (x) = O |x| , |x| → ∞}.
In addition, Pot (0) (Ωext ext
R ) is the space of continuous functions F : ΩR → R whose
(0) ext
restrictions F |Ωext
R
are members of Pot (ΩR ). In brief,
Pot (0) (Ωext

R ) = Pot
(0)
(Ωext ext
R ) ∩ C(ΩR ).
Clearly, in accordance with the well-posedness of the classical Dirichlet problem

(see, e.g., [6]) on Pot (0) (Ωext
R ) we are able to impose an inner product by

(F, G)H(Ωext ) = (F, G)L2 (ΩR ) = F (x)G(x)dωR (x).
R
ΩR
Furthermore,
Pot (∞) (Ωext ext
R ) = Pot (ΩR ) ∩ C
(∞)
(Ωext
R ).
Now, let A be the linear space consisting of all sequences {An }n∈N of real
numbers An = 0, n ∈ N0 :
A = {{An } : An ∈ R, An = 0, n ∈ N0 } .
Following [5] we consider the set E = E({An }; Ωext

R ) defined by
∞ 2n+1

∧ 2 2
E = F ∈ Pot (∞) (Ωext
R ) : A 2
n F L (ΩR )
(n, m) < ∞ ,
n=0 m=1
where
∧L2 (Ω R
F R) (n, m) = F (y)Hn,m (y) dωR (y).
ΩR
On E we define an inner product (·, ·)H({An };Ωext ) by

R
∞ 2n+1
∧L2 (Ω ∧L2 (Ω
(F, G)H({An };Ωext ) = A2n F R) (n, m)G R) (n, m), F, G ∈ E.
R
n=0 m=1
As usual, the associated norm is given by

1/2
F H({An };Ωext ) = (F, F )H({An };Ωext ) .
R R
Definition 2.4. The (scalar) Sobolev space H(Ωext ext

R ) (= H({An }; ΩR )) is the com-
pletion of E under the norm · H({An };Ωext ) :
R
·
H({An };Ωext )
H({An }; Ωext
R ) =E
R .
H({An }; Ωext
R ) equipped with the inner product (·, ·)H({An };Ωext ) is a Hilbert R
∗{A }
space. The system {Hn,m n (R; ·)} given by
∗{An }
Hn,m (R; x) = A−1 R
n Hn,m (x), x ∈ Ωext
R ,
∗ ∗{A }
is a Hilbert basis. We simply write Hn,m (R; ·) instead of Hn,m n (R; ·) if no con-
fusion is likely to arise. As is well known, any function F ∈ H({An }; ΩextR ) can be
∗{An }
expanded as a Fourier series in terms of the basis Hn,m (R; ·):
∞ 2n+1
∧H({A ext ∗{An }
F = F n };ΩR )
(n, m)Hn,m (R; ·),
n=0 m=1
where
∧H({A ext ∗{An }
F n };ΩR )
(n, m) = (F, Hn,m (R; ·))H({An };Ωext ) .
R
If no confusion is likely to arise we will also use the notation

∧H(Ωext ) ∧H({A ext
n };ΩR )
F R (n, m) = F (n, m).
Our next goal is the introduction of a class of scalar Sobolev spaces based
on the spherical symbol {(Δ∗;R )∧ (n)}n∈N0 of the Beltrami operator Δ∗;R on the
sphere ΩR . We know that
1 1
Δ∗;R Yn,m = 2 Δ∗ Yn,m = − 2 n(n + 1)Yn,m , n ∈ N0 ; m = 1, . . . , 2n + 1.
R R
In particular, we have Δ∗;R Y0,1 = 0, which requires a shift by a constant, for
1
example 4R 2 , to obtain invertibility. We formally have
s/2 s
∗;R 1 n + 1/2
−Δ + Yn,m = Yn,m
4R2 R
and
s/2 ∧ s
∗;R 1 n + 1/2
−Δ + F (n, m) = F ∧ (n, m),
4R2 R
n ∈ N0 , m = 1, . . . , 2n + 1. For s ∈ R we let
s

ext n + 1/2 ext
Hs (ΩR ) = H ; ΩR
R
and the norm in Hs (Ωext
R ) fulfills
1 s/2
F Hs (Ωext ) = (−Δ∗;R + ) F L2 (ΩR ) .
R
x
4R2
Remark. For the space H0 (Ωext
R ) we identify the norm · H0 (Ωext ) with the
R
· L2 (ΩR ) -norm. The space H0 (Ωext

R ) may be understood to be the space of
all solutions of the Dirichlet boundary value problem in Ωext R corresponding to
L2 (ΩR )-boundary values on ΩR . Note that the potential in H0 (Ωext R ) correspond-
ing to the L2 (ΩR )-(Dirichlet) boundary conditions on ΩR is uniquely determined.
Furthermore, if t < s, then F Ht (Ωext ) ≤ F Hs (Ωext ) and Hs (Ωext ext
R ) ⊂ Ht (ΩR ).
R R
In order to formulate some results about the convergence of the expansion

in terms of outer harmonics to a function in ordinary sense (Sobolev Lemma) we
need the concept of summable sequences {An }n∈N0 ∈ A satisfying
∞
2n + 1
< ∞.
n=0
A2n
Theorem 2.5 (Sobolev Lemma). Assume that the sequences {An }n∈N0 , {Bn }n∈N0 ∈
A are given in such a way that {Bn−1 An }n∈N0 is summable. Then each F ∈
(0)
H{Bn−1 An }; Ωext
R corresponds to a potential of class Pot (Ωext
R ).
The Sobolev Lemma (see [5] for its proof) states that in the case of summabil-
ity of the sequence {Bn−1 An }n∈N0 , the Fourier series in terms of the basis functions
∗
Hn,m ∈ H{Bn−1 An }; Ωext
R is continuous on the boundary ΩR . In particular, we have
the following statement.
Theorem 2.6. If F ∈ Hs (Ωext

R ) with s > 1, then F corresponds to a function of
(0) ext
class Pot (ΩR ).
For any F in L2 (ΩR ), there exists one and only one “harmonic continuation”
U ∈ H(Ωext
R ) which is given by
∞ 2n+1
∧L2 (Ω
U (x) = A2n F R)
∗
(n, m)Hn,m (R; x), x ∈ Ωext
R , (13)
n=0 m=1
where

∧L2 (Ω R 1 ∧H(Ωext )
F R) (n, m) = F (y)Hn,m (y)dωR (y) = U R (n, m),
ΩR A2n
n = 1, 2, . . . ; m = 1, . . . , 2n + 1.
The scalar Sobolev space H(Ωext R ) is a separable Hilbert space and the system
∗
{Hn,m (R; ·)} is a Hilbert basis. The space H(Ωext R ) has the reproducing kernel
ext ext
function KH(Ωext ) (·, ·) : ΩR × ΩR → R given by
R
∞ 2n+1

∗ ∗
KH(Ωext ) (x, y) = Hn,m (R; x)Hn,m (R; y), x, y ∈ Ωext
R .
R
n=0 m=1
In analogy to the scalar case we now introduce its tensorial counterpart pot(Ωext
R )
as follows:
pot(Ωext
R ) = {f ∈ c
(1)
(Ωext
R ) : ∇ · f = 0, ∇ ∧ f = 0 in Ωext
R ,
−3
|f (x)| = O |x| , |x| → ∞}.
Similarly, we let
pot(0) (Ωext ext
R ) = pot(ΩR ) ∩ c
(0)
(Ωext
R ),
and
pot(∞) (Ωext ext
R ) = pot(ΩR ) ∩ c
(∞)
(Ωext
R ).
In order to introduce Sobolev spaces for tensor fields we remember the sequences
{An }n∈N0 ∈ A. Then we define
∞ 2n+1

2 ∧l2 (ΩR )
e = f ∈ pot(∞) (Ωext
R ) : An (f (n, m))2
< ∞ ,
n=0 m=1
where
∧l2 (Ω
f R) (n, m) = f (y)hR;(1,1)
n,m (y)dωR (y).
ΩR
Equipped with the inner product
∞ 2n+1

(f , g)h({An };Ωext ) = A2n f ∧l2 (ΩR ) (n, m)g∧l2 (ΩR ) (n, m),
R
n=0 m=1
f , g ∈ e, the space e becomes a pre-Hilbert space. We define the Sobolev space

h(Ωext ext
R ) = h({An }; ΩR ) to be the completion of e under the norm · h(Ωext =
R )
· h({An };Ωext ) , which denotes the norm associated to (·, ·)h(Ωext ) = (·, ·)h({An };Ωext
R )
:
R R
·
h({An };Ωext )
h({An }; Ωext
R )= e
R .
The space h(Ωext ext
R ) (= h({An }; ΩR )) equipped with the inner product
∗{A }
(·, ·)h({An };Ωext ) is a Hilbert space. The system {hn,mn (R; ·)}n∈N0 ;m=1,...,2n+1 ,
R
given by
h∗{A n} −1 R;(1,1)
n,m (R; x) = An hn,m (x), x ∈ Ωext
R ,

R )-orthonormal Hilbert basis in h(ΩR ). We simply write
∗ ∗{An }
hn,m (R; ·) instead of hn,m (R; ·) if no confusion is likely to arise. As a conse-
quence we can expand a function f ∈ h(Ωext R ) as a Fourier series in terms of the
∗{A }
basis hn,mn (R; ·):
∞ 2n+1
∧
f= f h({An };Ωext )
R (n, m)h∗{A n}
n,m (R; ·),
n=0 m=1
where
∧ ∧
f h(Ωext )
R (n, m) = f h({An };Ωext )
R (n, m) = (f , h∗{A n}
n,m (R; ·))h(Ωext ) .
R
Finally, we are led to define

s

n + 1/2
hs (Ωext
R ) =h ; Ωext
R .
R
The Sobolev Lemma (Theorem 2.5) can be extended in the same way to tensor
fields.
Theorem 2.7 (Tensorial Sobolev Lemma). Suppose that the sequences {An }n∈N0 ,
{Bn }n∈N0 ∈ A are given such that {Bn−1 An }n∈N0 ∈ A is summable. Then each

f ∈ h {Bn−1 An }; Ωext
R corresponds to a function of class pot(0) Ωext
R .
For any f ∈ l2 (ΩR ), there exists one and only one tensorial “harmonic upward
continuation” u ∈ h(Ωext
R ) of the form
∞ 2n+1
∧l2 (Ω
u(x) = A2n f R) (n, m)h∗n,m (R; x), x ∈ Ωext
R , (14)
n=0 m=1
where

1 ∧h(Ωext )
f ∧l2 (ΩR ) (n, m) = f (y) · hR;(1,1) (y)dωR (y) = u R (n, m).
ΩR
n,m
A2n
The tensorial Sobolev space h(Ωext
R ) is a separable Hilbert space, and the system
∗;{An }
{hn,m (R; ·)} is a Hilbert basis. The space h(Ωext
R ) has the reproducing kernel
function
3 3 3 3
Kh(Ωext ) (·, ·) : Ωext ext
R ⊗ ΩR → R ⊗ R ⊗ R ⊗ R
R
given by
∞ 2n+1

∗{An } ∗{An }
Kh(Ωext ) (x, y) = hn,m (x) ⊗ hn,m (y),
R
n=0 m=1
x, y ∈ Ωext
R . This means that
(1,1) (1,1)
• for all x ∈ Ωext
R , ÕR Kh(Ωext ) (·, x) ∈ h(Ωext
R ), where the operator ÕR is
R
(1,1)
the extension of the adjoint operator of to tensor fields of rank four,
õR
(1,1) (1,1)
• ÕR f (x) = ÕR Kh(Ωext ) (·, x), f h(Ωext ) for every f ∈ h(Ωext
R ) and all x ∈
R R
Ωext
R .
More detailed information about tensorial Sobolev spaces can be found in the
Ph.D.-thesis [26]. The interrelation between scalar outer harmonics and their Hesse
tensor is known from [9]

1 (1,1)
(∇x ⊗ ∇x )Hn,m
R
(x) = 2 μ̃n hR;(1,1)
n,m (x), n = 0, 1, . . . ; m = 1, . . . , 2n + 1.
R
The last identity enables us to deduce that, for all F ∈ H(Ωext R ), the “Meissl
relation”

n (1,1)
∗
R μ̃n,m ∗

∇ ⊗ ∇F, hn,m (τ ; ·) h({A };Ωext ) = F, Hn,m (R; ·) H(Ωext )
n τ τ τ2 R
holds true for all τ > R and all n, m. This immediately leads us to the scalar outer
harmonic expansion in terms of tensorial Hesse data
τ n
∞ 2n+1
τ2 ∧h(Ωext ) ∗
F = (∇ ⊗ ∇F ) τ (n, m)Hn,m (R; ·). (15)
R (1,1)
n=0 m=1 μ̃n,m
The correlations between the potential F and the full Hesse tensor of F on the
Earth’s surface and the satellite orbit can be presented in a so-called Meissl scheme
(cf. [31]) as shown in Figure 2. Detailed information about Meissl schemes both in
the framework of outer harmonics and multiscale analysis can be found elsewhere
in this handbook.
Clearly, this formula is extremely suitable in the determination of the scalar
gravitational potential on a spherical Earth ΩR from tensorial SGG-data on a
spherical orbit Ωτ . It expresses the gravitational potential F ∈ H(Ωext R ) in terms
of the gravitational tensor ∇ ⊗ ∇F on the spherical satellite orbit Ωτ in terms
of a spherical harmonic expansion, where the convergence of the series (15) is
understood in uniform sense on every subset S ⊂ Ωext R with dist(S, ΩR ) > 0. Even
more, the convergence on ΩR can also be understood in the L2 (ΩR )-topology.
Figure 2. The full Meissl scheme for the Hesse tensor on the Earth’s
surface and on the satellite orbit (see the contribution [10] in this
handbook).
The outer harmonic expansion (15) actually represents the basic setting for
the multiresolution approach by means of (outer) harmonic wavelets as proposed
later on in this work.
Scalar Runge–Walsh approximation property

From [3] we know the following constructive version of the Runge theorem in terms
of outer harmonics
·C(Σext )
Pot (0) (Σext ) = span R
n=0,1,...; (Hn,m )|Σext , (16)
m=1,...,2n+1
where Σ is any regular surface (for example, a sphere, an ellipsoid, a telluroid,

the geoid, or the real Earth’s surface) and ΩR is a sphere inside Σ, such that the
“Runge condition” R < σ = inf x∈Σ |x| is valid.
The Runge–Walsh approximation property(16) justifies the approximation
of the Earth’s gravitational potential on and outside the regular surface by a
linear combination of scalar outer harmonics, i.e., by harmonic “trial functions
of polynomial nature” showing a “harmonicity domain” Ωext R ⊃ Σ
ext . It should
be remarked that the same property holds true, for example, for outer ellipsoidal
harmonics. However, for reasons of numerical economy and efficiency, we restrict
ourselves to outer spherical harmonics.
Tensorial Runge–Walsh approximation property

In an analogous way we are able to deduce from [26] that
·c(Σext )
pot(0) (Σext ) = span
R;(1,1)
n=0,1,...; (hn,m )|Σext .
m=1,...,2n+1
The calamity of evaluating the gravitational potential in terms of outer harmonics

is that these basis functions are globally supported. This is the reason why they
do not show any space localization but ideal frequency (momentum) localization
(for a more detailed description see [12]). This property makes outer harmonics
difficult to use for high resolution modeling at local scale. As a matter of fact,
the uncertainty principle leads us to the conclusion that outer harmonics are well
suited to resolve low-frequency problems at global scale, i.e., to provide trend
approximations.
Roughly spoken, seen from a numerical point of view, suitable ansatz spaces
should consist of harmonic functions possessing ideal space as well as frequency
localization. But this is mutually exclusive for several reasons (see [4]). Neverthe-
less, as well promising compromise, we are able to handle “sum conglomerates” of
outer harmonics, i.e., so-called (outer) harmonic kernel functions, offering a lim-
ited but appropriately balanced range of frequency as well as space localization.
Even better, we can construct families of kernels which control the increase of
space localization at the cost of the decrease of frequency localization by speci-
fying a scale parameter. Before we come to the definition of such families, called
scaling functions, within regularization procedures of the exponentially ill-posed
SGG-problem, however, it is advisable to describe the SGG-problem as a pseudo-
differential equation between the Sobolev spaces H(Ωext ext
R ) and h(Ωτ ).
3. Functional analytic background

For the convenience of the reader, we present here a brief course of basic facts on
regularization in a Hilbert space setting, which is useful to understand the spline
and multiscale solution strategies in the framework of pseudodifferential equations.
The explanations are based on functional analytic tools as described in [1, 2, 18]
where much more additional material can be found even for more general reference
spaces, too.
Let H and K be two Hilbert spaces with inner products (·, ·)H and (·, ·)K ,
respectively. Let
Λ : H −→ K
be a linear bounded operator. Given y ∈ K, we are looking for a solution of
Λx = y. (17)
In accordance to Hadamard (cf. [16]) we call such a problem well-posed , if the
following properties are valid:
• For all admissible data, a solution exists.

• For all admissible data, the solution is unique.
• The solution depends continuously on the data.
In our setting, these requirements can be translated into
• Λ is injective, i.e., R(Λ) = K.
• Λ is surjective, i.e., N (Λ) = {0}.
• Λ−1 is bounded and continuous.
If one of the three conditions is not fulfilled, the problem (17) is called ill-posed .
It will turn out, that the satellite problems we are concerned with are ill posed,
the most critical problem being the unboundedness of the inverse operator Λ−1 .
Let us discuss the consequences of the violations of the above requirements
for the well-posedness of equation (17). The lack of injectivity of Λ is perhaps the
easiest problem. From theoretical point of view, the space H can be replaced by the
orthogonal complement N (Λ)⊥ , and the restriction of the operator Λ to N (Λ)⊥
yields an injective problem. But in practice, one is normally confronted with the
problem that R(Λ) = K, since the right-hand side is given by measurements and
is, therefore, disturbed by errors. Now, we assume that y ∈ R(Λ), but only a
perturbed right-hand side y δ is known. We suppose that
y − y δ K < δ.
Our aim is to solve
Λxδ = y δ . (18)
Since y δ might not be in R(Λ), the solution of this equation might not exist, and
we have to generalize what is meant by a solution. xδ is called least squares solution
of (18), if
Λxδ − y δ K = inf{ Λz − y δ K : z ∈ H}. (19)
The solution of (19) might not be unique, and therefore one looks for the solution
of (19) with minimal norm. xδ is called best approximate solution of Λxδ = y δ , if
xδ is a least squares solution and
xδ H = inf{ z H : z is a least squares solution of Λz = y δ } (20)
holds.
The notion of a best-approximate solution is closely related to the Moore–
Penrose (generalized) inverse of Λ (see [22, 24], and a large amount of subsequent
contributions). We let
Λ̃ : N (Λ)⊥ −→ R(Λ) with Λ̃ = Λ|N (Λ)⊥
and define the Moore–Penrose (generalized) inverse Λ+ to be the unique linear
extension of Λ̃−1 to
D(Λ+ ) := R(Λ) + R(Λ)⊥
with
N (Λ+ ) = R(Λ)⊥ .
A standard result is provided by

Theorem 3.1. If y ∈ D(Λ+ ), then Λx = y has a unique best-approximate solution
which is given by
x+ = Λ+ y.
Note that the best-approximate solution is defined for all perturbed data
y δ ∈ K, whereas the last theorem requires that the right-hand side is an element
of D(Λ+ ).
A serious problem for ill-posed problems occurs when Λ−1 or Λ+ are not
continuous. That means that small errors in the data or even small numerical
noise can cause large errors in the solution. In fact, in most cases the application
of an unbounded Λ−1 or Λ+ does not make any sense. The usual strategy to
overcome this difficulty is to substitute the unbounded inverse operator
Λ−1 : R(Λ) −→ H
by a suitable bounded approximation
R : K −→ H .
The operator R is not chosen to be fix, but dependent on a regularization parameter
α. According to the conventional approach in the theory of ill-posed problems we
are led to introduce the following definition:
Definition 3.2. A regularization strategy is a family of linear bounded operators
Rα : K −→ H, α > 0,
so that
lim Rα Λx = x for all x ∈ H,
α→0
i.e., the operators Rα Λ converge pointwise to the identity.
From the theory of inverse problems (see, e.g., [24, 25]) it is also clear that
if Λ : H → K is compact and H has infinite dimension (as it is the case for the
application we have in mind), then the operators Rα are not uniformly bounded,
i.e., there exists a sequence (αj ) with limj→∞ αj = 0 and
Rαj L(K,H) → ∞ for j → ∞.
Note that the convergence of Rα Λx in Definition 3.2 is based on y = Λx,
i.e., on unperturbed data. In practice, the right-hand side is affected by errors and
then no convergence is achieved. Instead, one is (or has to be) satisfied with an
approximate solution based on a certain choice of the regularization parameter.
Let us discuss the error of the solution. For that purpose, we let y ∈ R(Λ)
be the (unknown) exact right-hand side and y δ ∈ K be the measured data with
y − y δ K < δ.
For a fixed α > 0, we let
xα,δ = Rα y δ ,
and look at xα,δ as an approximation of the solution x of Λx = y. Then the error

can be split as follows:
xα,δ − x H = Rα y δ − x H
≤ Rα y δ − Rα y H + Rα y − x H
≤ Rα L(K,H) y δ − y K + Rα y − x H ,
such that
xα,δ − x H ≤ δ Rα L(K,H) + Rα Λx − x H .
Figure 3. Typical behavior of the total error in a regularization process.
We see that the error between the exact and the approximate solution consists
of two parts: The first term is the product of the bound for the error in the data and
the norm of the regularization parameter Rα . This term will usually tend to infinity
for α → 0 if the inverse Λ−1 is unbounded and Λ is compact (cf. (3)). The second
term denotes the approximation error (Rα − Λ−1 )y H for the exact right-hand
side y = Λx. This error tends to zero as α → 0 by the definition of a regularization
strategy. Thus, both parts of the error show a diametrically oriented behavior. A
typical picture of the errors in dependence on the regularization parameter α is
sketched in Figure 3. Thus, a strategy is needed to choose α dependent on δ in
order to keep the error as small as possible, i.e., we would like to minimize
δ Rα L(K,H) + Rα Λx − x H .
In principle, we distinguish two classes of parameter choice rules: If α = α(δ)
only depends on δ, we call α = α(δ) an a priori parameter choice rule. Otherwise
α depends also on y δ and we call α = α(δ, y δ ) an a posteriori parameter choice
rule. It is usual to say that a parameter choice rule is convergent, if for δ → 0 the
rule fulfills the limit relations
lim sup{ Rα(δ,yδ ) y δ − Λ+ y H : y δ ∈ K, y δ − y K ≤ δ} = 0 (21)
δ→0
and
lim sup{α(δ, y δ ) : y δ ∈ K, y − y δ K ≤ δ} = 0. (22)
δ→0
We stop here the discussion of parameter choice rules. For more material the inter-
ested reader is referred to any textbook on inverse problems, e.g., [2, 18, 19, 27].
The remaining part of this section is devoted to the case that Λ is compact,
since then we gain benefits from the spectral representations of the operators. If Λ :
H → K is compact, a singular system (σn ; vn , un ) is defined as follows: {σn2 }n∈N are
the nonzero eigenvalues of the self-adjoint operator Λ∗ Λ (Λ∗ is the adjoint operator
of Λ), written down in decreasing order with corresponding multiplicity. The family
{vn }n∈N constitutes a corresponding complete orthonormal system of eigenvectors
of Λ∗ Λ. We let σn > 0 and define the family {un }n∈N via un = Λvn / Λvn K . The
sequence {un }n∈N forms a complete orthonormal system of eigenvectors of ΛΛ∗ ,
and the following formulas are valid:
Λvn = σn un , (23)
∗
Λ un = σn vn , (24)
∞

Λx = σn (x, vn )H un , x ∈ H, (25)
n=1
∞
Λ∗ y = σn (y, un )K vn , y ∈ K. (26)
n=1
The convergence of the infinite series is understood with respect to the Hilbert
space norms under consideration. The identities (25) and (26) are called the sin-
gular value expansions of the corresponding operators. If there are infinitely many
singular values, they tend to 0, i.e., limn→∞ σn = 0.
Theorem 3.3. Let (σn ; vn , un ) be a singular system for the compact linear operator
Λ, y ∈ K. Then we have
∞
|(y, un )K |2
y ∈ D(Λ+ ) if and only if < ∞, (27)
n=1
σn2
and for y ∈ D(Λ+ ) it holds

∞
(y, un )K
Λ+ y = vn . (28)
n=1
σn
The condition (27) is the Picard criterion. It says that a best-approximate

solution of Λx = y exists only if the Fourier coefficients of y decrease fast enough
relative to the singular values.
The representation (28) of the best-approximate solution motivates a method
for the construction of regularization operators, namely by damping the factors
1/σn in such a way that the series converges for all y ∈ K. We are looking for
filters
q : (0, ∞) × (0, Λ L(H,K)) −→ R (29)
such that
∞
q(α, σn )
Rα y := (y, un )K vn , y ∈ K,
n=1
σn
is a regularization strategy. The following statement is known from [18].
Theorem 3.4. Let Λ : H → K be compact with singular system (σn ; vn , un ). Assume
that q from (29) has the following properties:
(i) |q(α, σ)| ≤ 1 for all α > 0 and 0 < σ ≤ Λ L(H,K).
(ii) For every α > 0 there exists a c(α) so that |q(α, σ)| ≤ c(α)σ for all
0 < σ ≤ Λ L(H,K).
(iii) lim q(α, σ) = 1 for every 0 ≤ σ ≤ Λ L(H,K).
α→0
Then the operator Rα : K → H, α > 0, defined by
∞
q(α, σn )
Rα y := (y, un )K vn , y ∈ K,
n=1
σn
is a regularization strategy with Rα L(K,H) ≤ c(α).

The function q is called a regularizing filter for Λ. Two important examples
should be mentioned:
σ2
q(α, σ) =
α + σ2
defines the Tikhonov regularization, whereas

1, σ 2 ≥ α,
q(α, σ) =
0, σ 2 < α
leads to the regularization by truncated singular value decomposition.
4. SGG as exponentially ill-posed problem

After the discussion of the spherical settings and the outer harmonic based nomen-
clature leading to orthogonal expansions we are interested in formulating the SGG
problem by use of the concept of pseudodifferential operators. To this end we
shortly introduce the framework of tensorial spherical pseudodifferential operators
(PDO) (for a profound definition the reader is referred to [12]): Let {Λ∧ (n)}n∈N0
be a sequence of real numbers. The operator Λ : H(Ωext ext
R ) → h(Ωτ ), τ ≥ R > 0,
defined by
∞ 2n+1
∧
Λ∧ (n)F H(ΩR ) (n, m)h∗n,m (τ ; ·),
ext
ΛF =
n=0 m=1
is called a tensorial pseudodifferential operator of order t if

|Λ∧ (n)|
lim
n→∞ n + 1 t
= const = 0
2
holds true for some t ∈ R. If the limit

|Λ∧ (n)|
lim =0
n→∞ n + 1 t
2
holds true for all t ∈ R, the operator Λ is called a tensorial pseudodifferential

operator of exponential order. The sequence {Λ∧ (n)}n∈N0 is called the symbol of
the tensorial PDO Λ.
In the following we define scalar and tensorial kernel functions which are of
basic importance for the consideration of the SGG problem in terms of pseudodif-
ferential operators.
Definition 4.1. Suppose that R, τ ∈ R, τ ≥ R ≥ 0 as usual. Then any kernel

K R,R (·, ·) : Ωext ext
R × ΩR → R of the form
∞
2n+1

K R,R (x, y) = K ∧ (n) ∗
Hn,m ∗
(R; x)Hn,m (R; y), x, y ∈ Ωext
R
n=0 m=1
3 3
is called an HR,R -kernel. Any kernel kR,τ (·, ·) : Ωext ext
R × Ωτ → R ⊗ R of the form
∞
2n+1

∧ ∗
kR,τ
(x, y) = k (n) Hn,m (R; x)h∗n,m (τ ; y)
n=0 m=1
(x, y) ∈ Ωext
R × Ωext
τ is called an hR,τ -kernel.
The sequence {K ∧ (n)}n∈N0 is called the symbol of the HR,R -kernel, whereas
∧
k (n) is called the symbol of the hR,τ -kernel.
Definition 4.2. An HR,R -kernel K R,R (·, ·) with the symbol {K ∧(n)}n=0,... is called
∞ ∧ 2
1. n=0 (K (n)) < ∞,
∞ ∧ 2
2 K (n)
2. n=0 (2n + 1) An < ∞.
In analogy, an hR,τ -kernel kR,τ (·, ·) with the symbol {k∧ (n)}n=0,... is called ad-
missible, if these two conditions are satisfied for the symbol k∧ (n).
We define the convolution of an admissible HR,R -kernel against a function

F ∈ H(Ωext
R ) via the Parseval identity as follows
∞ 2n+1
∧H(Ωext )
(K R,R H(Ωext ) F )(x) = K ∧ (n)F R
∗
(n, m)Hn,m (R; x),
R
n=0 m=1
x ∈ Ωext
R . In analogy, we introduce the convolution of an admissible hR,τ -kernel
against F ∈ H(Ωext ext
R ) and f ∈ h(Ωτ ), respectively, as follows
∞ 2n+1
∧
(kR,τ H(Ωext ) F )(x) = k∧ (n)F H(Ωext )
R (n, m)h∗n,m (τ ; x), x ∈ Ωext
τ ,
R
n=0 m=1
∞ 2n+1
∧
(kR,τ ∗h(Ωext ) f )(x) = k∧ (n)f h(Ωext )
R
∗
(n, m)Hn,m (R; x), x ∈ Ωext
R .
τ
n=0 m=1
Within the context of pseudodifferential operators we are able to formulate

the (tensorial) interrelation of SGG presuming tensorial data at the orbital height
and requiring the potential at the Earth’s surface, however, starting first from a
spherical configuration (for more details concerning the tensorial SGG PDO see
also [9, 13]).
The pseudodifferential operator of the Hesse tensor

The operator of the second derivative (∇ ⊗ ∇) : H(Ωext ext
R ) → h(ΩR ) is defined by
#
∗ (n + 2)(n + 1)(2n + 3)(2n + 1) ∗
(∇ ⊗ ∇)Hn,m (R; x) = hn,m (R; x), x ∈ Ωext
R ,
R2
such that the symbol of this operator can be written as
#
∧ (n + 2)(n + 1)(2n + 3)(2n + 1)
(∇ ⊗ ∇) (n, m) = ,
R2
n = 0, 1, . . . ; m = 1, . . . , 2n + 1. For F ∈ H(Ωext
R ), the operator of the Hesse tensor
is given by
∞ 2n+1
∧
(∇ ⊗ ∇)∧ (n, m)F H(ΩR ) (n, m)h∗n,m (R; x).
ext
(∇ ⊗ ∇)F (x) =
n=0 m=1
The associated tensorial kernel can be expressed in the form

∞ 2n+1 #
(n + 2)(n + 1)(2n + 3)(2n + 1) ∗
k∇⊗∇ (x, y) = 2
Hn,m (R; x)h∗n,m (R; y),
n=0 m=1
R
x, y, ∈ Ωext
R , which leads to
(∇ ⊗ ∇)F (x) = (k∇⊗∇ H(Ωext ) F )(x).

R
The tensorial SGG pseudodifferential operator

The SGG operator ΛR,τ ext ext
SGG : H(ΩR ) → h(Ωτ ) can be formulated in terms of outer
harmonics by
# n
∗ (n + 2)(n + 1)(2n + 3)(2n + 1) R
R,τ
ΛSGG Hn,m (R; x) = h∗n,m (τ ; x), x ∈ Ωext
τ .
τ2 τ
The symbol of this operator is given by

# n
R,τ ∧ (n + 2)(n + 1)(2n + 3)(2n + 1) R
ΛSGG (n, m) = , (30)
τ2 τ
n = 0, 1, . . . ; m = 1, . . . , 2n + 1. The SGG operator applied to F ∈ H(Ωext R ) is
expressible as a series as follows
∞ 2n+1 R,τ ∧ ∧
ΛSGG (n, m)F H(ΩR ) (n, m)h∗n,m (τ ; x).
ext
ΛR,τ
SGG F (x) =
n=0 m=1
The associated kernel is given by

kΛR,τ (x, y)
SGG
∞ 2n+1 # n
(n + 2)(n + 1)(2n + 3)(2n + 1) R ∗
= Hn,m (R; x)h∗n,m (τ ; y)
n=0 m=1
τ2 τ
and
ΛR,τ
SGG F (x) = (kΛR,τ H(Ωext ) F )(x). (31)
SGG R
All in all, the inverse problem

ΛR,τ
SGG F = g (32)
of determining F ∈ H(Ωext ext
R ) from a given g ∈ h(Ωτ ) is exponentially ill-posed
(see, e.g., [19, 23, 24] for a more detailed classification in inverse theory and [15, 32]
for more information about SGG as inverse problem).
5. Spline inversion
We are now interested in calculating the gravitational potential F ∈ H(Ωext
R ) from
the observable g = ΛR,τ
SGG F via the tensorial SGG operator Λ R,τ
SGG :
ΛR,τ
SGG F = kΛR,τ H(Ωext ) F = g. (33)
SGG R
As already mentioned, the inverse problem ΛR,τ SGG F = g of determining F ∈

ext ext
H(ΩR ) from g ∈ h(Ωτ ) is exponentially ill posed.
In the following we propose a spline approximation technique for the regu-
larization of the tensorial SGG problem (33).
Definition 5.1. Suppose that YN = {y1 , . . . , yN } ⊂ Σ (with Σ representing the real
Earth’s surface as shown in Fig. 1). Any function UN ∈ H(Ωext R ) of the form

N
UN (x) = KH(Ωext ) (yi , x)ai , x ∈ Ωext
R ,
R
i=1
with arbitrarily given coefficients a1 , . . . , aN ∈ R is called a scalar harmonic spline
in the space H(Ωext
R ) relative to the system YN ⊂ Σ, provided that the functions
KH(Ωext ) (y1 , ·), . . ., KH(Ωext ) (yN , ·) are linearly independent. The class of all scalar
R R
harmonic splines in H(Ωext

R ) relative to YN is denoted by S
YN
.
H(Ωext
R )
Definition 5.2. Suppose that XN = {x1 , . . . , xN } ⊂ Γ (with Γ representing the

real LEO’s orbit). Any function uN ∈ h(Ωext
τ ) of the form

N
uN (x) = Õτ(1,1) Kh(Ωext ) (xi , x)ai , x ∈ Ωext
τ ,
τ
i=1
with arbitrarily given coefficients a1 , . . . , aN ∈ R is called a tensorial harmonic

(1,1)
spline in h(Ωext
τ ) relative to the system XN ⊂ Γ, provided that Õτ Kh(Ωext ) (x1 , ·),
τ
(1,1)
. . ., Õτ Kh(Ωext ) (xN , ·) are linearly independent. The class of all tensorial har-
τ
monic splines in h(Ωext xN

τ ) relative to XN is denoted by sh(Ωext ) .
τ
Following the usual constituents of harmonic spline theory (see [5]) it is not
difficult to verify the following minimum norm interpolation result.
Theorem 5.3. Let there be known from a function g ∈ h(Ωext τ ) the data points
(xi , g(xi )) ⊂ Γ × (R3 ⊗ R3 ), i = 1, . . . , N . Then the spline interpolation problem
ugN h(Ωext ) = infg v h(Ωext )
τ v∈iN τ
with
igN = {v ∈ h(Ωext
τ ) : v(xi ) = g(xi ), i = 1, . . . , N }
ously on the data g(x1 ), . . . , g(xN ). The uniquely determined solution ugN is given
in the explicit form

N
ugN (x) = Õτ(1,1) Kh(Ωext ) (xi , x)ai ,
τ
i=1
where the coefficients a1 , . . . , aN solve the linear equations

N
Õτ(1,1) Kh(Ωext ) (xi , xj )ai = g(xj ). (34)
τ
i=1
The unique solvability of (34) easily follows from the fact that
(1,1)
∞ 2n+1 n+2
(1,1) ∗ ∗ R
Õτ Kh(Ωext ) (·, y) = μ̃n,m Hn,m (R; ·)hn,m (τ ; y)
τ
n=0 m=1
τ
such that

N
K̃(xi , xj )ai = G(xj )
i=1
with
∞ 2n+1

(1,1) ∗ ∗
K̃(x, y) = μ̃n,m Hn,m (R; x)Hn,m (R; y),
n=0 m=1
and
(1,1)
G(x) = ÕR g(x).
Using the same coefficients ai , i = 1, . . . , N , we are led to an interpolating scalar
ext
spline of the gravitational potential F satisfying ΛR,τ SGG F = g in ΩR in the fol-
lowing way:
N
F
UN (x) = KH(Ωext ) (xi , x)ai
R
i=1
such that
ΛR,τ F
SGG UN (xj ) = g(xj )
holds true for all xj ∈ Γ.
We finally mention the tensorial counterpart of the Shannon sampling theorem.
Theorem 5.4. Let f be in sX N

h(Ωext
. Then f can be reconstructed from its samples
τ )
by the following interpolation formula

N
f (x) = Õτ(1,1) f (xk )pN
k (x), x ∈ Ωext
τ , (35)
k=1
k ∈ sh(Ωext ) , k = 1, . . . , N , are given by

YN
where the “Lagrangians” pN
τ

N
pN
k =
N
wl,k Õτ(1,1) Kh(Ωext ) (xl , x), (36)
τ
l=1
N
and the coefficients wl,k have to satisfy the linear equations

N
N
wl,k Õτ(1,1) Õτ(1,1) Kh(Ωext ) (xi , xl ) = δi,k , i, k = 1, . . . , N. (37)
τ
l=1
Proof. As a member of sXN

h(Ωext )
the function f is uniquely determined by
τ

N
f= ak Õτ(1,1) Kh(Ωext ) (xk , x)
τ
k=1
with

N
ak Õτ(1,1) Kh(Ωext ) (xk , xi ) = f (xi ).
τ
k=1
If we define the Lagrangians as in Equation (36) with coefficients given by (37)

we get

N
N
N
Õτ(1,1) N
wk,l f (xk ) = N
wk,l ai Õτ(1,1) Õτ(1,1) Kh(Ωext ) (xi , xk )
τ
k=1 k=1 i=1

N
= ai δil = al .
i=1
N N
Using wl,k = wk,l we get

N
f (x) = al Õτ(1,1) Kh(Ωext ) (xl , x)
τ
l=1

N
N
= N
wk,l Õτ(1,1) f (x)Õτ(1,1) Kh(Ωext ) (xl , x)
τ
l=1 k=1

N
= Õτ(1,1) f (xk )pN
k ,
k=1
which is the desired result.
Remark. In a realistic situation we have to deal with noisy tensorial measurements

g and thus we have to turn over to an operator equation ΛR,τ SGG : H → h̃ involving
a Sobolev space h̃ ⊃ h (e.g., h̃ = h0 ) such that the invertibility of ΛR,τ
SGG cannot
be assumed anymore and regularization has to come into play. In [9] we presented
a multiscale regularization, whereas in case of the spline approximation presented
here we take advantage of the fact that the space h is dense in h̃ such that the
interpolation procedure is replaceable by a smoothing procedure in the framework
of H and h, thereby taking advantage of the denseness of h in h̃.
6. Multiscale inversion
Next we introduce tensorial wavelets for the approximation of tensor fields (a
more detailed presentation can be found in [5] for the scalar case and in [26] for
the tensorial case).
generator of a scaling function if it satisfies the following requirements (see [11]):
(i) For all j ∈ N0
(ϕj (0))2 = 1,
(ii) for all j, j
∈ N0 with j ≤ j
and all n = 1, 2, . . .
2
(ϕj (n))2 ≤ (ϕj (n)) ,
(iii) for all n = 1, 2, . . .

lim (ϕj (n))2 = 1.
j→∞

h(Ωext
τ )-scaling functions.
Definition 6.2. A family {Φj (·, ·)}j∈N0 of hτ,τ -kernels defined by Φ∧

j (n) = ϕj (n),
j ∈ N0 , n ∈ N0 , i.e.,
∞
2n+1

∗
Φj (x, y) = ϕj (n) Hn,m (τ ; x)h∗n,m (τ ; y), x, y ∈ Ωext
τ ,
n=0 m=1
is called an h(Ωext τ )-scaling function, if it satisfies the following properties:

(i) Φj (·, ·) is an admissible h(Ωext τ )-kernel for every j ∈ N0 ,
(ii) {{Φ∧ j (n)} n∈N 0 } j∈N 0 constitutes a generator of a scaling function.
We now come to the definition of the multiresolution analysis.
Definition 6.3. Let {Φj (·, ·)}j∈N0 be a family of admissible hτ,τ -kernels. Then the
family {vj (Ωext ext
τ )}j∈N0 of scale spaces vj (Ωτ ) defined by
vj (Ωext ext
τ ) = {Φj H(Ωext ) Φj ∗h(Ωext ) f : f ∈ h(Ωτ )},
τ τ
is called an h(Ωext
τ )-multiresolution analysis, if the following properties are satis-
fied:
(i) v0 (Ωext ext ext ext
τ ) ⊂ · · · ⊂ vj (Ωτ ) ⊂ vj+1 (Ωτ ) ⊂ · · · ⊂ h(Ωτ ),
; ·h(Ωext )
(ii) vj (Ωext
τ )
τ
= h(Ωext
τ ).
j∈N0
It is not hard to show that the following multiscale approximation theorem

is valid.
Theorem 6.4. Let {Φj (·, ·)}j∈N0 be an h(Ωext
τ )-scaling function. Then
lim f − Φj H(Ωext ) Φj ∗h(Ωext ) f h(Ωext ) = 0
j→∞ τ τ τ
holds for all f ∈ h(Ωext ext

τ ). In conclusion, {vj (Ωτ )}j∈N0 forms an hτ,τ -multireso-
lution analysis.
The next purpose is to introduce wavelets via a tensorial refinement equation.
Definition 6.5. Let {Φj (·, ·)}j∈N0 be an h(Ωext
τ )-scaling function. Then the family
ext
of h(Ωτ )-kernels {Ψj (·, ·)}j∈N0 given by
(Ψj )∧ (n) = ψj (n), j ∈ N 0 ; n ∈ N0 ,
is called h(Ωext
τ )-wavelet, if the h(Ωext Ψj (·, ·), j ∈ N0 , are admissible and
τ )-kernels
the symbols {ψj (n)}, in addition, satisfy the refinement equation
(ψj (n))2 = (ϕj+1 (n))2 − (ϕj (n))2
for all j ∈ N0 , n ∈ N0 .
As usual, we define the detail space wj (Ωext

τ ) at scale j by
! "
wj (Ωext ext
τ ) = Ψj H(Ωext ) Ψj ∗h(Ωext ) f : f ∈ h(Ωτ ) .
τ τ
In case of low-to-medium wavelength approximation bandlimited wavelets

(based on the Shannon kernel and its modifications) are used for the evaluation
of F , whereas the non-bandlimited variant of the wavelet evaluation (using non-
bandlimited kernels such as Tikhonov, rational, exponential, and “locally sup-
ported” kernels) is applied to short wavelength approximation.
Examples for SGG-regularization strategies
We discuss three examples of multiresolution analyses
∧ by specifying the sequence
{ϕj (n)} in more detail. We let σn = σn,m = ΛR,τ
SGG (n, m) (see Eq. (30)).
Tikhonov regularization. The symbol of the scaling function in case of Tikhonov
regularization is given by
σn
ϕj (n) = 2
σn + γj
with limj→∞ γj = 0 and limj→−∞ γj = ∞ (see Figure 4).
15
10
SGG−symbol
Scale 0
Scale15
Scale 30
10 Scale 45
10
Scale 60
5
10
0
10
−5
10
−10
10
−15
10
0 100 200 300 400 500 600 700 800
Figure 4. Symbol of the SGG operator with R = 6378.127, τ =

R + 225.0 (black) and of the scaling function in case of Tikhonov regu-
larization for different dyadic scales γj = 2j .
Tikhonov–Phillips regularization. The symbol of the scaling function in case of

Tikhonov–Phillips regularization is given by
σn
ϕj (n) = 1
σn2 + γj R2 (n(n + 1) + 14 )
with limj→∞ γj = 0 and limj→−∞ γj = ∞ (see Figure 5).
15
10
SGG−symbol
Scale 0
Scale 15
Scale 30
10 Scale 45
10
Scale 60
5
10
0
10
−5
10
−10
10
−15
10
0 100 200 300 400 500 600 700 800
Figure 5. Symbol of the SGG operator with R = 6378.127, τ = R +

225.0 (black) and of the scaling function in case of Tikhonov–Phillips
regularization for different dyadic scales γj = 2j .
Orthogonal truncated singular value decomposition. The symbol of the scaling

function in case of the orthogonal truncated singular value decomposition is
given by
−1
σn , for n = 0, . . . , Nj ,
ϕj (n) =
0, for n ≥ Nj + 1,
with

0, for j < 0,
Nj =
2j − 1 , for j ≥ 0.
15
10
Scale 2
Scale 4
Scale 6
Scale 8
Scale 9
10 SGG Symbol
10
5
10
0
10
−5
10
−10
10
0 100 200 300 400 500 600 700 800
Figure 6. Symbol of the SGG operator with R = 6378.127, τ = R +

225.0 (black) and of the scaling function in case of truncated singular
value decomposition regularization for different scales.
Runge regularization by bandlimited outer harmonic wavelet integration

The idea now is to use a two step strategy for observing real geometries in the
SGG-inverse problem: For that purpose we formally relate a scalar potential F
of class H(Ωext ext
R ) to a tensorial potential g ∈ h(Ωτ ) by multiscale regularization
R,τ
of the spherically based SGG-equation ΛSGG F = g. In doing so, the scalar har-
monic wavelet expansion of F outside ΩR by means of bandlimited wavelets can
be constructed in such a way that its scalar wavelet (potential) coefficients on ΩR
become expressible as integrals over Ωτ convolving g against tensorial bandlimited
harmonic wavelet kernels. Turning over to Step 1 we are led to understand the con-
volution integrals over Ωτ involving the tensorial bandlimited harmonic wavelet
kernels as linear functionals that can be written as linear combinations in terms
of the actual SGG-data on Γ, hence, in combination with the spherical approach,
the scalar wavelet (potential) coefficients on ΩR become available as linear com-
binations of the actual tensorial SGG-data on Γ. Step 2 enables us to establish
the multiscale regularization of F on and outside Σ such that an approximation
of the gravitational potential from SGG-data becomes available obeying the real
geometries.
In order to derive the integration rules for Runge regularization involving the
2
actual geometry we remember F ∈ H(Ωext R ) to possess a restriction F |ΩR ∈ L (ΩR )
ext 2
(see (13)). In the same way g ∈ h(Ωτ ) implies g|Ωτ ∈ l (Ωτ ) (see (14)). Note
that l2 (Ωτ ) is the Hilbert space of square integrable tensor fields f : Ωτ → R3 ⊗ R3 .
We now formulate a two step method for the Runge realization of multiscale
regularization by integration in more detail:
Step 1. We suppose the scale discrete scaling functions and the corresponding
wavelets to be bandlimited, i.e., there exists a sequence
0 ≤ m0 < m1 < m2 < · · · , lim mj = ∞,
j→∞
with

m0 2n+1

ΦR,τ
0 (x, ·) = (Φ0 )∧ (n) R
Hn,m ;(1,1)
(x)hτn,m (·),
n=0 m=1
and

mj 2n+1

ΨR,τ
j (x, ·) = (Ψj )∧ (n) R
Hn,m ;(1,1)
(x)hτn,m (·), j ≥ 0.
n=0 m=1
This implies

ext
g(y) · ΦR,τ
0 (·, y)dωτ (y) ∈ Harm0,...,m0 (ΩR )
Ωτ
as well as
ext
g(y) · ΨR,τ
j (·, y)dωτ (y) ∈ Harm0,...,mj (ΩR ), j ≥ 0.
Ωτ
In consequence we are able to formulate
= {x1 j , . . . , xMj } ⊂ ΩR , Mj = (2mj + 1)2 , denote funda-

R M M
Theorem 6.6. Let XM j
mental systems with respect to Harm0,...,2mj (Ωext R ), j ≥ 0, i.e., the matrix
⎛ M M
⎞
R
H1,1 (x1 j ) · · · H1,1
R
(xMjj )
⎜ . ⎟
⎜ . .. ⎟
⎝ . . ⎠
M M
R
Hm,2m+1 (x1 j ) · · · Hm,2m+1R
(xMjj )
is assumed to be regular. Then, for g ∈ im(ΛR,τ ext

SGG ) the potential F ∈ H(ΩR )
given by

M0
F (z) = a0k g(y) · ΦR,τ M0 R,R M0
0 (xk , y)dωτ (y)Φ0 (xk , z)
k=1 Ωτ
∞
Mj
M M
+ ajk g(y) · ΨR,τ j R,R
j (xk , y)dωτ (y)Ψj (xk j , z),
j=0 k=1 Ωτ
z ∈ Ωext
R , is the solution of the inverse problem
ΛR,τ
SGG F = g, F ∈ H(Ωext
R ), g ∈ h(Ωext
τ ).
In case of arbitrary g ∈ h(Ωext

τ ) we have the regularized solution

M 0
FJ (z) = a0k g(y) · ΦR,τ M0
0 (xk , y)dωτ (y)Φ0
R,R M0
(xk , z)
k=1 Ωτ

J−1 Mj
M M
+ ajk g(y) · ΨR,τ j R,R
j (xk , y)dωτ (y)Ψj (xk j , z),
j=0 k=1 Ωτ
z ∈ Ωext j j
R , where the integration weights a1 , . . . , aM , j = 0, . . . , J − 1, satisfy the
linear systems
Mj
M
ajk Hn,m (xk j ) = Hn,m (x)dωR (x),
k=1 ΩR
n = 0, . . . , 2mj , i = 1, . . . , 2n + 1. (38)
Step 2. The “orbit” Γ is assumed to be totally contained in the exterior of the
sphere Ωτ . It is helpful to introduce the space
∞ 2n+1 2
∧ ext τ
ext ;(1,1)
h̃(Ωτ ) = g h(Ωτ )
(n, m) hτn,m (x) : g ∈ h(Ωext ext .
τ ), x ∈ Ωτ
n=0 m=1
|x|
Indeed, (h̃(Ωext ext

τ ), (·, ·)l2 (Ωτ ) ) is a Hilbert space. Any g ∈ h̃(Ωτ ) can be uniquely
∞ 2n+1
∧ 2
τ ;(1,1)
g(x) = g l2 (Ωτ )
(n, m) hτn,m (x), x ∈ Ωext
τ .
n=0 m=1
|x|
Using the functions
2
;(1,1) τ ;(1,1)
h̃τn,m (x) = hτn,m (x), x ∈ Ωext
τ ,
|x|
we are led to the space
(1,1) ;(1,1)
h̃arm0,...,m (Ωext
τ ) = span n=0,...,m; (h̃τn,m ).
m=1,...,2n+1
Σ
The set XM = {y1M , . . . , yM
M
} ⊂ Σ, M = (m + 1)2 is called a fundamental system
(1,1)
with respect to h̃arm0,...,m (Ωext
τ ), if the matrix
⎛ ⎞
τ ;(1,1) M τ ;(1,1) M
h̃ (y1 ) . . . h̃1,1 (yM )
⎜ . 1,1 ⎟
⎜ . .. ⎟
⎝ . . ⎠
τ ;(1,1) M τ ;(1,1) M
h̃m,2m+1 (y1 ) . . . h̃m,2m+1 (yM )
is regular. These settings allow us to formulate
Σ
Theorem 6.7. Let XM = {y1M , . . . , yM
M
} ⊂ Σ, M = (m + 1)2 be a fundamental
(1,1)
system with respect to h̃arm0,...,m (Ωext
τ ). Furthermore, suppose that
(1,1)
g ∈ h̃arm0,...,m (Ωext
τ ) and Ψ ∈ h(Ωext
τ ).
Then

M
g(y) · Ψ(y)dωτ (y) = ap · g(ypM )
Ωτ p=1
holds true, if a1 , . . . , aM satisfy
M
;(1,1) M ;(1,1)
ap · h̃τn,m (yp ) = Ψ(y) · h̃τn,m (y)dωτ (y),
p=1 Ωτ
n = 0, . . . , m; j = 1, . . . , 2n + 1.
Altogether, we end up with the following theorem about the resulting gravi-
tational potential F on the Earth’s surface Σ computed from SGG data given on
the real orbit Γ:
Theorem 6.8. Let g be a bandlimited function of class h̃(Ωextτ ) that is given at
M = (m + 1)2 points {y1M , . . . , yM
M
} ⊂ Σ which form a fundamental system with
(1,1)
τ ). Furthermore, let {Ψj }, j ≥ 0, denote a scale dis-
R,τ
respect to h̃arm0,...,m (Ωext
crete bandlimited spherical regularization decomposition wavelet of order 0 with
respect to (32) and corresponding decomposition scaling function {ΦR,τ j }, j ≥ 0.
R,R R,R
Let {Ψj } and {Φj } be the corresponding reconstruction wavelet and scaling
function, respectively. Then the regularized solution FJ of (32) is given by

M0
M
FJ (x) = a0k · b̂0,k M R,R
s g(ys )Φ0 (x, xM0
k )
k=1 s=1

J−1 Mj

M
M
+ ajk · bj,k M R,R
s g(ys )Ψj (x, xk j ),
j=0 k=1 s=1
where b̂0,k 0,k

1 , . . . , b̂M , k = 1, . . . , M0 , satisfy

M
τi (1,1) Mj
(yp ) = (ΦR,R )∧ (n)Hn,j
R;(1,1)
ajp · hn,m (xMj
p ),
p=1
n = 0, . . . , m; i = 1, . . . , 2n + 1,
bj,k j,k
1 , . . . , bM , j = 0, . . . , J − 1, k = 1, . . . , Mj , satisfy

M
τ ;(1,1) ∧ R;(1,1) M
bj,k
s h̃n,i (ysM ) = (ΨR,τ
j ) (n)hn,i (xk j ),
s=1
n = 0, . . . , m; i = 1, . . . , 2n + 1,
M M
aj1 , . . . , ajMj , j = 0, . . . , J − 1, satisfy (38), and XM
R
j
= {x1 j , . . . , xMjj } ⊂ ΩR ,
Mj = (2mj + 1)2 denote fundamental systems with respect to Harm 0,...,2mj (Ωext
R ).
At a first glance, the regularization procedure requires the solution of lots

of linear systems. But if we have a careful look we recognize that always the
same matrix has to be inverted, and all weights for the numerical integration are
obtained by a matrix-vector multiplication. In principle, the assumption that g
is bandlimited can be omitted and the calculated weights can be used for the
numerical integration formulas (which, however, are then no longer exact).
Note that our wavelet based regularization technique is presented for the ten-
sorial SGG problem (32) thereby applying new types of appropriate integration
formula involving outer harmonic spaces. The basic concept is to express con-
volution integrals on spheres by linear combinations of data on real geometries
taking strong advantage of the harmonicity of the wavelet kernels. As the SGG
problem demands a global approach we assume the data to be given on discrete
points uniformly distributed over the satellite orbit Γ. Nevertheless, because of
the space localizing properties of scaling and wavelet functions, our theory can be
extended canonically to local areas for which locally given data sets can be used.
An extension to locally oriented multiscale regularization can be found in [9].
7. A tree algorithm based on harmonic spline exact approximation

In what follows, the SGG problem will be solved within a multiscale procedure in-
volving interpolating splines. Suppose that a SGG data set of values corresponding
to NJ points on Γ is known. Let the family {ΦJ (·, ·)}J∈N0 be an h(Ωext τ )-scaling
function which is bandlimited. We formally set
NJ 2n+1

(2)
ΦJ = (ϕJ (n))2 h∗n,m (τ ; ·) ⊗ h∗n,m (τ ; ·)
n=0 m=1
(2)
(with ϕJ (n) = Φ∧J (n)) for the iterated kernel function. Note that ΦJ is a tensorial
kernel function of rank four. In the framework of convolutions we obviously have
(2)
ΦJ ∗h(Ωext ) g = ΦJ H(Ωext ) ΦJ ∗h(Ωext ) g.
R R R
In accordance with the presentation in [5] for the scalar case it follows that,
(2)
in the terminology of Sobolev spaces, ΦJ is the unique reproducing kernel of
h({An /ϕ∧ ext
J (n)}; Ωτ ). For sufficiently large J, corresponding to the tensor data g,
(2)
(2) Φ ∗g
there is in arbitrarily close accuracy to ΦJ ∗ g a spline uNJJ consistent with a
set of NJ original data g(xN J NJ
1 ), . . . , g(xNJ ):
(2)
∗g

NJ
(2)
Õτ(1,1) ΦJ (xN x ∈ Ωext
Φ
uNJJ (x) ≈ J NJ
i , x)ai , τ .
i=1
N
For j = 0, . . . , J, we assume that the coefficients wl,kj have been determined by
solving the linear system (see Theorem 5.4)
Nj
(2) N
Õτ(1,1) Õτ(1,1) Φj (xl j , xi j ) wi,kj = δlk ,
N N
l, k = 1, . . . , Nj . (39)
i=1
Remark. In its tensorial generality as formulated here the condition (39) cer-
tainly is a bottleneck of the presented method seen from numerical point of view.
Nonetheless, our approach is mathematically interesting. Even more, if we restrict
ourselves to constituting ingredients of the Hesse tensor (such as second radial
derivatives), the tree algorithm can be established in the same way requiring much
less numerical effort. Observe that the solution of the linear system (39) has to
be calculated once and can be stored elsewhere, as far as the same nodal system
is in use.
The tree algorithm consists of the following ingredients:
The initial step (or sampling step). The point of departure is the observation that
for sufficiently large J
g(x) = ΦJ H(Ωext ) ΦJ ∗h(Ωext ) g
τ τ
(2)
Φ
≈ uNJJ ∗g

NJ
(2)
= Õτ(1,1) ΦJ (xN J NJ
i , x)ai , x ∈ Ωext
τ ,
i=1
where aN
i
J
are determined by

NJ
(2)

NJ
NJ (1,1) NJ (1,1)
aN
i
J
= wi,k Õτ (ΦJ ∗h(Ωext ) g)(xN
k )≈
J
wi,k Õτ g(xN J
k ).
τ
k=1 k=1
The pyramid step. We start from
(2)
(2)
Φ ∗g
Nj
(2)
Õτ(1,1) Φj (xi j , x)ai j , x ∈ Ωext
N N
(Φj ∗ g)(x) ≈ uNjj (x) = τ ,
i=1
with

Nj
(2)
wi,kj Õτ(1,1) (Φj ∗h(Ωext ) g)(xk j ),
Nj N N
ai = for j = J0 , . . . , J.
τ
k=1
From
(2)

Nj+1
(2)
Õτ(1,1) Φj (·, xl
Nj+1 Nj+1
Φj ∗ g ≈ al )
l=1
we get the following recursion relation

Nj Nj+1
(2)
Õτ(1,1) Õτ(1,1) Φj (xk j , xl
N N Nj+1 N Nj+1
ai j ≈ )wi,kj al ,
k=1 l=1
i = 1, . . . , Nj , which leads us to the following decomposition scheme:

F → aN J → aNJ−1 → · · · aNJ0
↓ ↓ ↓
(2) (2) (2)
ΦJ ∗ g ΦJ−1 ∗ g ΦJ0 ∗ g.
Note that by use of the corresponding iterated scalar scaling function (see,
e.g., [5]) given by
∞ 2n+1

(2)
ΦJ = (ϕJ (n))2 Hn,m
R R
Hn,m
n=0 m=1
we arrive at a representation of the spline approximation of F satisfying ΛR,τ
SGG F =
g on Ωext
R in the following way
Φ
(2)
∗F
Nj
(2)
x ∈ Ωext
N N
UNjj (x) = Φj (xi j , x)ai j , R .
i=1
Note that, in case of bandlimited scaling functions, “≈” may be replaced by “=”.
Remark. Accordingly, the tree algorithm can be realized if only scalar data within
the Hesse tensor (for example, second radial derivatives) are used for approxima-
tion (cf. [5]). The trace of the Hesse tensor (which is equal to zero) offers the
possibility to validate the method. Moreover, observational errors can be handled
by smoothing or filtering techniques within the tree algorithm (see [5, 7, 8]). Since
these procedures are well documented in the literature, they will be not discussed
in this approach.
8. Conclusion
A gradiometer mission ideally produces a coverage of the entire Earth with ten-
sorial measurements at a certain altitude. Our multiscale method yields decorre-
lations of the scalar internal gravitational potential of the Earth by a tree algo-
rithm within the framework of inverse problems under real geometric situations.
Even more, from the SGG-approach presented here, it can be expected that the
knowledge about the representation of the Earth’s gravitational potential will be
improved in considerable way, providing the transition form low- and meso-based
to high(er) reflected resolution at global scale, thereby offering significant local
features.
Although an impressive rate of the Earth’s gravitational potential can be de-
tected globally at the orbit of a satellite (like GOCE), the computational drawback
of satellite techniques in geoscientific research is the fact that measurements must
be performed at a certain altitude. Consequently, a “downward continuation” pro-
cess must be applied to handle the potential at the Earth’s surface, hence, a loss of
information for the signal is unavoidable. Indeed, “downward continuation” causes
severe problems, since the amount of amplification for the potential is not known
suitably (as an a priori amount) and even small errors in the measurements may
produce huge errors in the potential at the Earth’s surface.
However it is of great advantage that satellite data are globally available,

at least in principle. Nevertheless, from mathematical point of view, we are not
confronted with a boundary value of potential theory. Satellite techniques such as
SGG require the solution of an inverse problem to produce gravitational informa-
tion at the Earth’s surface, where it is needed actually. SGG can be formulated
adequately as (Fredholm) pseudodifferential equation of the first kind, which is
exponentially ill posed, and this fact makes the development of suitable math-
ematical methods with strong relation to the nature and structure of the data
indispensable.
In this respect it should be mentioned that each method in approximation
theory has its own aim and character. Even more, it is the essence of any nu-
merical realization that it becomes optimal only with respect to certain specified
features. For example, Fourier expansion methods with polynomial trial functions
(spherical harmonics) offer the canonical “trend-approximation” of low-frequency
phenomena (for global modeling), they offer an excellent control and comparison
of spectral properties of the signal, since any spherical harmonic relates to one
frequency. This is of tremendous advantage for relating data types under spec-
tral aspects. But it is at the price that the polynomials are globally supported
such that local modeling results into serious problems of economy and efficiency.
Bandlimited kernels can be used for the transition from long-wavelength to short-
wavelength phenomena (global to local modeling) in the signal. Because of their
excellent localization properties in the space domain, the non-bandlimited kernels
can be used for the modeling of short-wavelength phenomena. Local modeling is
effective and economic. But the information obtained by kernel approximations is
clustered in frequency bands so that spectral investigations are laborious and time
consuming. In other words, for numerical work to be done, we have to make an a
priori choice. We have to reflect the different stages of space/frequency localiza-
tion so that the modeling process can be adapted to the localization requirements
necessary and sufficient for our geophysical or geodetic interpretation.
In conclusion, an algorithm establishing an approximate solution for the in-
verse SGG-problem has to reflect the intention of the applicant. Different tech-
niques for regularization are at the disposal of the numerical analyst for global
as well as local purposes. Each effort only gives certain progress in the particular
field of pre-defined interest. If a broad field of optimality should be covered, only a
combined approach is the strategic instrument to make an essential step forward.
Thus, for computational aspects of determining the Earth’s gravitational poten-
tial, at least a twofold combination is demanded, viz. combining globally available
satellite data (including the SGG-contribution) with local airborne and/or ter-
restrial data and combining tools and means of constructive approximation such
as polynomials, splines, wavelets, etc. Altogether, in numerical modeling of the
Earth’s gravitational potential, there is no best universal method, there exist only
optimized procedures with respect to certain features and the option and the fea-
sibility for their suitable combination.
Acknowledgment
The first two authors thank the “Federal Ministry for Economic Affairs and En-
ergy, Berlin” and the “Project Management Jülich” for funding the project “SPE”
(funding reference number 0324016, CBM – Gesellschaft für Consulting, Business
und Management mbH, Bexbach, Germany).
References
[1] Eggermont, P.N., LaRiccia, V., Nashed, M.Z. (2014) Noise Models for Ill-Posed Prob-
lems. In: Freeden,W., Nashed, M.Z., Sonar, T. (Eds.) Handbook of Geomathematics,
2nd edition, Springer, New York, 1633–1658
[2] Engl, H., Hanke, M., Neubauer, A. (1996) Regularization of Inverse Problems,
Kluwer, Dordrecht, Boston, London
[3] Freeden, W. (1980a) On the Approximation of External Gravitational Potential With
Closed Systems of (Trial) Functions, Bull. Géod., 54: 1–20
[4] Freeden, W. (1998) The Uncertainty Principle and Its Role in Physical Geodesy. In:
Freeden, W. (ed.) Progress in Geodetic Science, 225–236. Shaker, Aachen
Leipzig
[6] Freeden, W., Gerhards, C. (2013) Geomathematically Oriented Potential Theory,
Chapman & Hall/CRC Press, Boca Raton
[7] Freeden, W., Maier, T. (2002) On Multiscale Denoising of Spherical Functions: Ba-
sic Theory and Numerical Aspects. Electronic Transactions on Numerical Analysis
(ETNA), 14: 40–62
[8] Freeden, W., Maier, T. (2003) Spectral and Multiscale Signal-to-Noise Thresholding
of Spherical Vector Fields. Computational Geosciences, 7: 215–250
[9] Freeden, W., Nutz, H. (2011) Satellite Gravity Gradiometry as Tensorial Inverse
Problem. Int. J. Geomath. 2: 177–218
[10] Freeden, W., Nutz, H. (2018) Geodetic Observables and Their Mathematical Treate-
ment in Multiscale Framework. In: Freeden, W., Nashed , M.Z. (Eds.), Handbook of
Mathematical Geodesy, this volume, Springer International Publishing, 315–458
[11] Freeden, W., Schneider, F. (1998) Regularization Wavelets and Multiresolution, In-
verse Problems, 14: 493–515
[12] Freeden, W., Schreiner, M. (2009) Spherical Functions of Mathematical Geosciences.
Springer, Heidelberg
[13] Freeden, W., Schreiner, M. (2014) Satellite Gravity Gradiometry (SGG): From Scalar
to Tensorial Solution. In: Handbook of Geomathematics, Freeden, W., Nashed, M.Z.,
and Sonar, T. (Eds.), Springer, New York, Heidelberg, 2nd edition, 339–380
[14] Freeden W., Gervens, T., Schreiner, M. (1998) Constructive Approximation on
the Sphere (with Applications to Geomathematics), Oxford Science Publications,
Clarendon, Oxford
[15] Freeden, W., Schneider, F., Schreiner, M. (1997) Gradiometry – An Inverse Problem
in Modern Satellite Geodesy, in: Engl, H.W., Louis, A., Rundell, W. (Eds.), GAMM–
SIAM Symposium on Inverse Problems: Geophysical Applications, 179–239
[16] Hadamard, J. (1923) Lectures on the Cauchy Problem in Linear Partial Differential
Equations, Yale University Press, New Haven
[17] Heiskanen, W.A., Moritz, H. (1967) Physical Geodesy, Freeman, San Francisco, CA
[18] Kirsch, A. (1996) Introduction to the Mathematical Theory of Inverse Problems,
Springer, New York
[19] Louis, A.K. (1989) Inverse und schlecht gestellte Probleme, Teubner, Leipzig
[20] Meissl, P. (1971) A Study of Covariance Functions Related to the Earth’s Disturbing
Columbus, OH
[21] Meissl, P. (1971) On the Linearization of the Geodetic Boundary Value Problem.
Department of Geodetic Science, No. 152, The Ohio State University, Columbus,
OH
[22] Nashed, M.Z. (1976) Generalized Inverses and Applications, Proceedings of an Ad-
vanced Seminar Conducted by the Mathematics Research Center, University of Wis-
consin, Madison
[23] Nashed, M.Z. (1981) Operator-Theoretic and Computational Approaches to Ill-
Posed Problems with Applications to Antenna Theory. IEEE Transactions on An-
tennas and Propagation, 29(2): 220–231
[24] Nashed, M.Z. (1987) A New Approach to Classification and Regularization of
Ill-Posed Operator Equations. In: Inverse and Ill-Posed Problems, Engl, H.W.,
Groetsch, C.W. (Eds.), Notes and Reports in Mathematics in Science and Engi-
neering 4: 53–75, Academic Press, New York
[25] Nashed, M.Z. (1987) Inner, Outer, and Generalized Inverses in Banach and Hilbert
Spaces. Numer. Funct. Anal. and Optimiz., 9: 261–325
[26] Nutz, H. (2002) A Unified Setup of Gravitational Field Observables. Ph.D.-thesis,
University of Kaiserslautern, Geomathematics Group, Shaker, Aachen
[27] Rieder, A (2003) Keine Probleme mit Inversen Problemen. Vieweg, Wiesbaden
[28] Rummel, R. (1997) Spherical Spectral Properties of the Earth’s Gravitational Po-
tential and Its First and Second Derivatives. In Lecture Notes in Earth Science,
Springer, Berlin, 65: 359–404
[29] Rummel, R. (2014) GOCE: Gravitational Gradiometry in a Satellite. In: Handbook
of Geomathematics, Freeden, W., Nashed, M.Z., and Sonar, T. (eds)., Springer, New
York, Heidelberg, 2nd edition, 211–226
[30] Rummel, R., van Gelderen, M. (1992) Spectral Analysis of the Full Gravity Tensor.
Geophys. J. Int., 111: 159–169
[31] Rummel, R., van Gelderen, M. (1995) Meissl Scheme – Spectral Characteristics of
Physical Geodesy. Manuscr. Geod. 20: 379–385
[32] Schreiner, M. (1994) Tensor Spherical Harmonics and Their Application in Satellite
Gradiometry. Ph.D.-thesis, University of Kaiserslautern, Geomathematics Group
Willi Freeden and Helga Nutz Michael Schreiner

Geomathematics Group Institute for Computational Engineering
University of Kaiserslautern University of Buchs
MPI-Gebäude, Paul-Ehrlich-Str. 26 Buchs, Switzerland

Parameter Choices for

Fast Harmonic Spline Approximation
Martin Gutting
Abstract. The approximation by harmonic trial functions allows the construc-

tion of the solution of boundary value problems in geoscience where the bound-
ary is often the known surface of the Earth itself. Using harmonic splines such
a solution can be approximated from discrete data on the surface. Due to their
localizing properties regional modeling or the improvement of a global model
in a part of the Earth’s surface is possible with splines.
Fast multipole methods have been developed for some cases of the oc-
curring kernels to obtain a fast matrix-vector multiplication. The main idea
of the fast multipole algorithm consists of a hierarchical decomposition of the
computational domain into cubes and a kernel approximation for the more
distant points. This reduces the numerical effort of the matrix-vector mul-
tiplication from quadratic to linear in reference to the number of points for
a prescribed accuracy of the kernel approximation. In combination with an
iterative solver this provides a fast computation of the spline coefficients.
The application of the fast multipole method to spline approximation
which also allows the treatment of noisy data requires the choice of a smooth-
ing parameter. We summarize several methods to (ideally automatically)
choose this parameter with and without prior knowledge of the noise level.
Keywords. Spline approximation, fast multipole methods.
1. Introduction
Spherical splines have been developed by Freeden [38] and independently by Wahba
[134] for interpolation and smoothing/approximation problems on the sphere and
have been generalized to harmonic splines by Freeden [37, 39, 40]. These harmonic
splines can be used for interpolation/approximation on regular surfaces, but in
particular for the solution of boundary value problems where the boundary is
a regular surface. Convergence theorems exist for both spherical splines and har-
monic splines (cf. [41, 42, 43, 44]), and the different types of spline spaces and their
606 M. Gutting
reproducing kernels have been investigated (cf. [49] and the references therein). In
geosciences they have found many applications (see, e.g., [45, 46, 48, 107] and the
references therein).
Splines lead to a system of linear equations which in case of harmonic splines
has to be densely populated. This makes the solution expensive in terms of the
numerical effort. Iterative solvers require fast summation methods corresponding
to the reproducing kernels of these splines to be truly efficient. On the sphere there
are several possible ways to achieve a fast summation (the problem that is con-
sidered determines which method should be preferred): spherical panel clustering
(cf. [45, 52] and the references therein), spherical FFT for gridded data points or
spherical NFFT for non-equispaced data (cf. [116, 82]).
[61, 62, 63, 124] have introduced the fast multipole method (FMM) in two
and three dimensions for fast evaluation of potentials corresponding to the Laplace
operator (generalizations to further operators have been introduced later). Since
such potentials are closely related to certain reproducing kernels of our splines,
the FMM allows fast summation of harmonic splines as well as spherical splines.
Such a combination is used in [57] to solve problems of satellite geodesy with
harmonic splines corresponding to the singularity kernel. We have extended this
to the Abel–Poisson kernel and use the accelerated version of the FMM that was
first introduced in [24, 64]. This approach has also been applied to the oblique
boundary value problem of potential theory in [67, 68].
We consider the following (generalized) interpolation problems:
Problem 1.1 (Interpolation on a regular surface). Let Σ be a C (0) -regular surface

(see Definition 2.1 below for details). Let a finite set of points {x1 , . . . , xN } ⊂ Σ
on the surface and data Fi , i = 1, . . . , N corresponding to these points be given.
The aim is to find a function F in a function space of choice such that F (xi ) = Fi ,
i = 1, . . . , N .
If the data Fi are error-affected and strict interpolation is no longer desirable

and the interpolation conditions are reduced to F (xi ) ≈ Fi , i = 1, . . . , N , and F
has to minimize a functional that balances closeness to the data and smoothness
of F , usually with one (or several) parameters.
Closely related to this interpolation/approximation problem is the discrete
version of the Dirichlet boundary value problem which requires only the values of
the boundary function in a finite set of points on the surface. This is also called a
generalized interpolation problem.
Problem 1.2. Let Σ be a C (k) -regular surface with k ≥ 2 (see again Definition
2.1 below for details). Let {x1 , . . . , xN } ⊂ Σ be a discrete set of N points on the
surface. For each point xi let Fi = U (xi ) be given, where
i = 1, . . . , N .
The task is to determine the potential U ∈ C (0) Σext ∩ C (2) (Σext ) which is
harmonic in Σext , i.e., the exterior of the surface Σ, and regular at infinity, i.e., for
Parameter Choices for Fast Harmonic Spline Approximation 607
|x| → ∞,

|U (x)| = O |x|−1 , (1)

|∇U (x)| = O |x|−2 , (2)
or an approximation UN to it which fits the data, i.e., for i = 1, . . . , N ,
UN (xi ) = Fi = U (xi ). (3)
As before the interpolation conditions (3) are relaxed in case of error-affected
data/measurements.
The outline of this paper is as follows: Section 2 summarizes the theory of
harmonic splines and spline approximation. In Section 3 we establish the connec-
tion between harmonic splines and the sums that can be computed by the fast
multipole method, we introduce the adaptive construction of the decomposition of
the computational domain and provide our version of the fast multipole algorithm
for harmonic splines. Section 4 gives an overview of suitable methods to choose the
smoothing parameter of the approximating splines if the data are afflicted with
(stochastic) noise.
2. Preliminaries
Spherical harmonics, which we denote by Yn,m (with degree n ∈ N0 , order m =
−n, . . . , n), are known to form a complete orthonormal basis of the space L2 (S2 )
of square integrable functions on the unit sphere S2 (see, e.g., [30, 47, 130]).
The spherical harmonics {Yn,m }n∈N0 ,m=−n,...,n also form a closed system in C(S2 )
and are closed and complete in L2 (S2 ). This allows the representation of square-
integrable functions on any sphere S2R of radius R > 0 by their Fourier series,
where the Fourier coefficients of F ∈ L2 (S2R ) are denoted by

1 x
F ∧ (n, m) = F (x) Yn,m R dSR (x) . (4)
S2R R
2.1. Regular surfaces and Runge spheres
Due to the Runge–Walsh approximation theorem, we can use functions which
possess a larger domain of harmonicity to approximate the solution of a problem
which requires harmonicity only outside the Earth’s surface (see [45, 48] for an
extensive introduction of this technique). Harmonic splines as introduced in [37,
39, 40, 125] are constructed in such a way that they are subspaces of the space of
harmonic functions on a sphere situated inside the Earth, the so-called Runge (or
Krarup) sphere (see [109]).
The Earth’s surface is considered to be regular as by the following definition.
Definition 2.1. A C (k) -regular surface Σ ⊂ R3 is a surface in R3 which has to fulfill
(i) Σ divides R3 into the interior Σint and the exterior Σext , where Σint is a
bounded region and Σext is an unbounded region.
608 M. Gutting
(ii) The origin is contained in Σint .

(iii) Σ is closed (and therefore compact) and free of double points.
(iv) Σ is a C (k) -surface, i.e., for each x ∈ Σ there exists a neighborhood U ⊂ R3
of x such that Σ ∩ U possesses a C (k) -parametrization.
We can also define C (k,λ) -regular surfaces Σ ⊂ R3 with λ ∈ (0, 1) as a C (k) -
regular surface where every point x ∈ Σ possesses a neighborhood U such that Σ∩U
can locally be parameterized by a k-times λ-Hölder continuously differentiable
parametrization. Such surfaces are required for oblique derivative boundary value
problems as in [68], but not here. Note that any sphere S2R of radius R > 0 is
obviously a C (∞) -regular surface.
For regular surfaces Σ we can define Runge spheres (cf. [45, 46, 48], see also
[109] where it is called Krarup sphere).
Definition 2.2. The Runge sphereS2R is a sphere of radius R around the origin
such that the exterior of the Runge sphere, i.e., S2R,ext , contains the exterior of the
regular surface Σ, i.e., Σext ⊂ S2R,ext .
2.2. Sobolev spaces

Now we briefly introduce the Sobolev spaces of the form H = H {An }; S2R,ext
using the Runge sphere S2R . For more details, the reader is referred to [45, 48, 107]
and the references in these books.
Definition 2.3. Let {An }n∈N0 ⊂ R be a sequence which satisfies the summability
condition
∞
2n + 1
< ∞. (5)
4πA2n
n=0

The Sobolev space H = H {An }; S2R,ext is defined by
·H {An }; S2
( R,ext )
H = H {An }; S2R,ext = E {An }; S2R,ext , (6)

where E {An }; S2R,ext ⊂ C (∞) S2R,ext is the set of all functions that are har-
monic in S2R,ext , infinitely often differentiable on the Runge sphere S2R and regular
at infinity (i.e., (1) and (2) hold) and whose Fourier coefficients F ∧ (n, m) with
respect to L2 (S2R ) (as defined in (4)) fulfill
∞
n
2
F H{A 2
= A2n (F ∧ (n, m)) < ∞ . (7)
n }; SR,ext
n=0 m=−n
H is a Hilbert space with the inner product defined by
∞ n
F, GH{A }; S2 = A2n F ∧ (n, m)G∧ (n, m) (8)
n R,ext
n=0 m=−n

for F, G ∈ H {An }; S2R,ext .
It is well known (cf. [45, 107] and the references therein) that such a space
possesses a so-called reproducing kernel (see [3] for an overview on reproducing
kernels in general).
Definition 2.4. Let U be a non-empty set and (X, ·, ·X ) be a separable Hilbert
space of real-valued functions on U . Let {Bn }n∈N0 be a complete orthonormal
system in (X, ·, ·X ). Any function K : U × U −→ R of the form
∞

K (x, y) = K ∧ (n)Bn (x) Bn (y) (9)
n=0
with x, y ∈ U and K ∧ (n) ∈ R for n ∈ N0 is called an X-product kernel (briefly an

X-kernel).
An X-kernel K (·, ·) : U × U −→ R is called a reproducing kernel (or shortly
repro-kernel ) for (X, ·, ·X ) if:
(i) K (x, ·) ∈ X for all x ∈ U .
(ii) K (x, ·) , F X = F (x) for all x ∈ U and all F ∈ X.
If there exists such a repro-kernel in X, then X is called a reproducing

kernel
Hilbert space
and the repro-kernel is unique (cf. [3]). In the space H =
H {An }; S2R,ext with a summable sequence {An } the repro-kernel (9) can be
represented by its expansion in Legendre polynomials due to the well-known ad-
dition theorem for spherical harmonics:
∞ 2 n
2n + 1 1 R x y
KH (x, y) = Pn · . (10)
n=0
4πA2n |x||y| |x||y| |x| |y|
2.3. Harmonic splines

We use the reproducing kernels of Section 2.2 to define harmonic splines.
Definition 2.5. Let {L1 , . . . , LN } ⊂ H∗ be a set of N linearly independent bounded

linear functionals on the reproducing kernel Hilbert space H. Then any function
S of the form

N
S= ai Li KH (·, ·) (11)
i=1
with a set of so-called spline coefficients {a1 , . . . , aN } ⊂ R is called an H-spline rel-

ative to {L1 , . . . , LN }. The function space of all H-splines relative to {L1 , . . . , LN }
is denoted by SH (L1 , . . . , LN ).
By construction any H-spline is a harmonic function. The interpolating spline

S F for the function F ∈ H has to fulfill the interpolation conditions
Li S F = Li F for i = 1, . . . , N. (12)
610 M. Gutting
The interpolation conditions (12) can be rewritten as a system of linear equations

for the spline coefficients ai :

N
ai Li Lj KH (·, ·) = Lj F, j = 1, . . . , N, (13)
i=1
whose corresponding matrix possesses the entries Li Lj KH (·, ·) and is symmetric

and positive definite (for linear functionals L1 , . . . , LN ∈ H∗ which are linearly
independent).
In this paper, we consider only evaluation functionals Lx , i.e., Lx F = F (x)
where x ∈ Σext . Furthermore, L1 , . . . , LN are given by Li F = F (xi ) where xi ∈ Σ.
For other types of functionals see [43, 44] or [67, 68] (for the case of oblique
derivatives). In the following theorem we summarize the properties of H-splines.
Theorem 2.6. Let F ∈ H and let {L1 , . . . , LN } ⊂ H∗ . Then the H-spline interpo-
lation problem with the interpolation conditions (12) is uniquely solvable and its
solution S F ∈ SH (L1 , . . . , LN ) possesses the following properties:
- the H-orthogonal projection of F onto SH (L1 , . . . , LN ).
S F is
(i) -
(ii) -S F -H ≤ F H .
(iii) If G ∈ H also satisfies the interpolation conditions (12), then the first mini-
mum property holds:
- -2 - -2
G 2H = -S F -H + -G − S F -H , (14)
i.e., S F is the interpolating function of F in H with minimal norm.
(iv) If S ∈ SH (L1 , . . . , LN ) and G ∈ H also satisfies the interpolation conditions
(12), then the second minimum property holds:
- -2 - -2
S − G = -S F − G- + -S − S F - .
2
H H H
(15)
For the proof and for further details on splines, the reader is referred to
[37, 43, 45, 51, 107] and the references therein.
n
Example 2.7. The choice An = h− 2 , h ∈ (0, 1), fulfills (5) and provides us with
the reproducing kernel called Abel–Poisson kernel which is given by
1 |x|2 |y|2 − h2 R4
KH (x, y) = 3 . (16)
4π
(|x|2 |y|2 + h2 R4 − 2hR2 x · y) 2
1 n
Example 2.8. The sequence An = (n + 12 ) 2 h− 2 , h ∈ (0, 1), also satisfies (5) and
leads to the singularity kernel given by
1 1
KH (x, y) = 1 . (17)
2π
(|x|2 |y|2 + h2 R4 − 2hR2 x · y) 2
In [48] the existence of approximations fulfilling interpolation conditions is
shown by the Runge–Walsh approximation theorem and an extension of Helly’s
theorem (cf. [138]). Convergence results for harmonic splines (cf. [43, 44]) can be
derived that show the convergence to the solution of the Dirichlet boundary value
problem for an increasing density of data points, i.e., if the largest data gap goes
to zero (cf. [43]).
We consider two specific types of splines (using Abel–Poisson and singularity
kernels)
and propose the fast multipole method to quickly compute the sums
ai KH (xi , yj ) for many points in Section 3. This can be used to solve the systems
of linear equations (13) that occur in the solution of the interpolation problems
using harmonic splines.
2.4. Spline approximation
For noisy data, i.e., Fi = U (xi ) + δi , i = 1, . . . , N , where the noise δi is mod-
eled by some stochastic process, e.g., white noise (see Section 4.1 for details), in
Problem 1.2, it makes no sense to compute an interpolation problem. We look for
an approximation to U which can be interpreted as a smoothing of the data (see
[38, 51, 50, 135] for the spherical spline approximation, [37, 45] for the case of
harmonic spline approximation). Minimizing the following functional

N
N
μ(S) = (Li S − Fi )Cij (Lj S − Fj ) + β S H (18)
i=1 j=1

in the reproducing kernel Hilbert space H = H {An }; S2R,ext yields the desired
smoothed approximation of the data. C = (Cik ) ∈ RN ×N denotes a positive
definite matrix which allows us to include covariance information on the data if
available. β > 0 is a constant smoothing parameter which balances closeness to
the data and smoothing. The following theorem of [45, 107] (see also the references
therein) summarizes the existence and uniqueness of a spline approximation.
Theorem 2.9. Let Fi , i = 1, . . . , N , correspond to a set of linearly independent
bounded linear functionals L1 , . . . , LN ∈ H∗ .
Then there exists a unique element S ∈ SH (L1 , . . . , LN ) such that
μ(S) ≤ μ(F ) for all F ∈ H (19)
and μ(S) = μ(F ) if and only if S = F . This element is called the smoothing
spline or approximating spline. Its spline coefficients ai , i = 1, . . . , N , are uniquely
determined by the system of linear equations

N

ai Li Lj KH (·, ·) + β(C −1 )ij = Lj F, j = 1, . . . , N. (20)
i=1
The matrix in (20) corresponds to the sum of the matrix in (13) and βC −1 . It
is still positive definite. If C is the identity matrix, there is only the one smoothing
parameter β. Using a diagonal matrix as C it is possible to introduce weights for
the data Fi and include additional information on the noise of the data. The choice
of the smoothing parameter(s) can be interpreted as the application of a parameter
choice method in the regularization theory of ill-posed problems (see Section 4).
612 M. Gutting
3. The fast multipole method for splines

The interpolation conditions (12) as well as the minimization of the smoothing
functional (18) lead to a system of linear equations (13), or (20) respectively, with
a dense matrix whose size is the number of data points. This matrix can be large
and the solution of the corresponding system of linear equations becomes difficult.
Reproducing kernels of a reproducing kernel Hilbert space H defined by the
summable sequence {An } can be expanded in terms of Legendre polynomials as
in (10). The singularity kernel (17) and the Abel–Poisson kernel (16) possess a
representation as an elementary function and both kernels are closely related to the
1
single pole |x−y| by the Kelvin transform. Therefore, we can use the fast multipole
method (FMM), which has been introduced by Greengard [61], Greengard and
Rokhlin [62, 63], and Rokhlin [124], for the fast summation of harmonic splines,
i.e., of the sum ai KH (xi , ·). This corresponds to the matrix-vector products
occurring in an iterative solver for (13) or (20).
The FMM creates a hierarchical subdivision of the computational domain
into nested cubes that are organized in an octtree data structure. Instead of single
point interaction the cubes summarize the part of the kernel sum corresponding
to the points they contain and interact with other cubes via the coefficients of
truncated inner/outer harmonics expansions. This kernel approximation is applied
as often as possible and on the coarsest possible level of the tree data structure.
Direct evaluation is used only for the closest cubes where the approximation is
not accurate enough. The algorithm has been improved several times to increase
its efficiency (cf., e.g., [24, 64, 137]). We summarize our implementation and show
the application of the FMM to harmonic splines (see [67, 69] for a more detailed
analysis).
3.1. Kelvin transform of reproducing kernels
The Kelvin transform yields the connection between the kernels (17) and (16)
and the fundamental solution of the Laplace equation, i.e., the single pole. The
transform can be seen as a reflection on a sphere around the origin and we choose
the Runge sphere S2R of Definition 2.2 for this (see also text books on potential
theory, e.g., [46, 83]).
Definition 3.1. Let Γ ⊆ R3 be a domain, W : Γ −→ R a function. Let the reflection
of Γ on the sphere S2R be given by

R2
ΓKT = xKT ∈ R3 : KT 2 xKT = x ∈ Γ . (21)
|x |
The function
W KT : ΓKT −→ R,

KT KT KT R R2 R
x → W (x ) = KT W xKT = W (x), (22)
|x | |xKT |2 |xKT |
is called the Kelvin transform of W with respect to the sphere of radius R.
The Kelvin transform is applied to the reproducing kernels with respect to

KT
one argument (the other is kept fixed). The Kelvin transform KH (x, y KT ) of the
singularity kernel (17) can be computed, e.g., by its expansion
∞ 2 n
hn 2n + 1 R x y
KH (x, y) = Pn ·
n=0
n + 12 4π|x||y| |x||y| |x| |y|
∞
1 (h|y KT |)n x y KT
= Pn ·
2π|y| n=0 |x|n+1 |x| |y KT |
1 1 |y KT | KT
= = KH (x, y KT ), (23)
2π|y| |x − hy KT | R
R2
where y KT = |y|2 y and
KT 1 1
KH (x, y KT ) = . (24)
2πR |x − hy KT |
KT
The Kelvin transform KH (x, y KT ) of the Abel–Poisson kernel (16) is given by
1 |x|2 |y|2 − h2 R4
KH (x, y) = 3
4π
(|x|2 |y|2 + h2 R4 − 2hR2 x · y) 2
|y KT | 1 |x|2 − h2 |y KT |2 |y KT | KT
= = KH (x, y KT ), (25)
R 4πR |x − hy KT |3 R
which is related to (24) by
KT 1 1
KH (x, y KT ) = −x · ∇x − 12 Id . (26)
2πR |x − hy KT |
We summarize both (24) and (26) by use of the operator Dx such that
KT 1 1
KH (x, y KT ) = Dx , (27)
2πR |x − hy KT |
where Dx = Id (singularity kernel) or Dx = −x · ∇x − 12 Id (Abel–Poisson kernel).
3.2. Adaptive decomposition of the domain
Now we consider the evaluation of the sum
N N
|y KT | KT
ai KH (xi , y) = ai KH (xi , y KT )
i=1 i=1
R

|y KT | ai
N
1
= Dx (28)
R i=1 2πR |x − hy | x=xi
KT
at the points yj , j = 1, . . . , M = O(N ), which can be the points xi as in the

systems of linear equations (13) or (20). To better distinguish the points xi and
2
yj we call the Kelvin transformed points hyjKT = h |yRj |2 yj which are used in (28)
targets.
614 M. Gutting
In the beginning, a bounding cube is determined which is large enough such

that it contains all points and all targets. This single cube forms level 0 of the
octtree structure and is subdivided into eight equally sized cubes of half its edge
length which then form level 1. Each cube is adaptively divided into nested cubes
where a cube of level l has half the edge length of a cube of level l − 1 as proposed,
e.g., by Cheng et al. [24]. Points and targets are sorted into the currently available
cubes. If a cube contains more than the prescribed maximal number of points or
targets m, it is split into eight smaller cubes of the next level. All its points/targets
are redistributed into these eight cubes. We summarize the necessary vocabulary
in some definitions (see also [24, 64]).
Definition 3.2.
(a) A cube C is called child of the cube B if C results from a single subdivision
of B which in return is named the parent of C.
(b) A cube that is not further subdivided is called childless or a leaf.
(c) Cubes are said to be neighbors if they are of the same size (same level) and
share at least one boundary point. Each cube is a neighbor of itself.
(d) If two cubes are at the same level, but are no neighbors, they are called
well-separated, i.e., between these cubes exists at least one cube of their size.
Each cube carries the relevant information about other cubes, in particular
its neighbors, in four lists as suggested by Cheng et al. [24]. Figures 1 and 2 give
a two-dimensional illustration.
Figure 1. Two-dimensional illustration of an adaptive decomposition

for levels 2 (left) and 3 (right). If the marked square is a leaf, the white
cubes form its list 1 and the blue cubes correspond to list 3. If the
marked square is not a leaf, all white and all blue cubes are treated at
the next level. List 2 cubes are red and list 4 cubes are yellow. Cubes
in green have been handled at coarser levels.
Figure 2. Two-dimensional illustration of an adaptive decomposition

for level 4. If the marked square is a leaf, the white cubes form its list
1 and the blue cubes correspond to list 3. If the marked square is not
a leaf, all white and all blue cubes are treated at the next level. List
2 cubes are red and list 4 cubes are yellow. Cubes in green have been
handled at coarser levels.
Definition 3.3.
(a) In list 1 of the childless cube X are all childless cubes directly adjacent to X.
List 1 only contains any cubes if X is a leaf. In this case it always contains
at least X itself.
(b) List 2 of a cube X consists of all children of neighbors of the parent cube of X
which are well separated from X. The cube X does not need to be childless.
(c) Children of neighbors of the leaf X (or smaller cubes descending from neigh-
bors of X) which do not have any point in common with X form list 3. Their
parents have to be adjacent to X. If X is not childless, then list 3 is empty.
(d) List 4 consists of childless cubes which are neighbors of the parent cube of
X, but these childless cubes are not adjacent to X.
Notice the following observations:
(i) List 1 is the list of all neighbors.
(ii) All cubes in list 2 of a cube X are of the same size as X and well separated
from X.
(iii) The elements of list 3 are all smaller than X and the distance between them
and X is at least their side length and at most the side length of X.
(iv) List 4 of a cube X only contains cubes that are larger than X. They are
separated from X by a distance that is at least the side length of X and at
most their own edge length.
(v) A cube X is in list 3 of a cube Y if and only if Y is in list 4 of X.
(vi) All members of list 1 and list 4 are leaves and list 1 as well as list 3 of a cube
X remain empty if X is not childless.
616 M. Gutting
After finishing the adaptive construction of the octtree and sorting all points and
targets into cubes, the algorithm removes childless cubes that contain neither
points nor targets and are no longer required.
3.3. Single pole expansion

In addition to the decomposition of the domain, the other part of the FMM is
the kernel approximation of the single pole by a truncated expansion and the use
of translation theorems to shift the expansion center and to change the type of
expansion. Similar to (23) we use the generating function of the Legendre polyno-
mials Pn and the addition theorem of spherical harmonics (see, e.g., [1, 47, 102])
to expand the single pole.
1 1
=
|x − y| |y − x0 − (x − x0 )|
∞
|x − x0 |n y − x0 x − x0
= Pn ·
n=0
|y − x0 |n+1 |y − x0 | |x − x0 |
∞
n
∗
= In,m (x − x0 )On,m (y − x0 ), (29)
n=0 m=−n
where |y − x0 | > |x − x0 | for the expansion center x0 ∈ R3 . The upper star ∗ in

(29) denotes the complex conjugate. Thereby, we use the (complex-valued) outer
and inner harmonics for n ∈ N0 , m = −n, . . . , n:
#
4π (n + m)!(n − m)! x
On,m (x) = Yn,m |x| , (30)
2n + 1 |x|n+1

4π |x|n
In,m (x) = # Yn,m |x|x
, (31)
2n + 1 (n + m)!(n − m)!
where ϑ ∈ [0, π], ϕ ∈ [0, 2π) are the usual spherical coordinates of |x|
x
and Yn,m :
S2 → C with

2n + 1 (n − m)!
Yn,m (ξ) = (−1) m
Pn,m (cos(ϑ))eimϕ , ξ ∈ S2 , (32)
4π (n + m)!
are complex-valued fully normalized spherical harmonics of degree n and order m.
Pn,m : [−1, 1] → R are the associated Legendre functions with
1 m dn+m 2
Pn,m (t) = (1 − t2 ) 2 (t − 1)n , m = 0, . . . , n. (33)
2n n! dt n+m
The symmetry relation Pn,−m (t) = (−1)m (n−m)!

(n+m)! Pn,m (t) extends them for negative
orders (cf., e.g., [16, 30]).
Well-known translation theorems for these outer and inner harmonics allow
to shift the expansion center (see, e.g., [34] for a detailed derivation).
Theorem 3.4 (Translation Theorem for Outer Harmonics). Let x, y ∈ R3 such

that |x| > |y|. Then the outer harmonic of degree n ∈ N0 and order m ∈ Z,
−n ≤ m ≤ n, at x − y can be expanded in terms of inner and outer harmonics as
follows

∞

n
On,m (x − y) = In∗ ,m (y)On+n ,m+m (x) (34)
n =0 m =−n

∞

n
= In∗ −n,m −m (y)On ,m (x). (35)
n =n m =−n
Note that in (35) we make use of the convention that In,m = 0 if |m| > n.
Obviously, this infinite series as well as the infinite sum in (29) have to be truncated
for the algorithm which are sources of (truncation) errors. Error estimates for the
truncation errors can be found in [61, 63, 64] and the references therein. Another
approach which investigates the combined error of both truncations in (29) and
(35) is considered in [67].
Theorem 3.5 (Translation Theorem for Inner Harmonics). Let x, y ∈ R3 . Then
the inner harmonic of degree n ∈ N0 and order m ∈ Z, −n ≤ m ≤ n, at x − y can
be expanded in a finite sum of inner harmonics

n
n

In,m (x − y) = (−1)n In ,m (y)In−n ,m−m (x). (36)
n =0 m =−n
For orders with |m| > n we have again by convention In,m = 0. Note that
no truncation is necessary for this translation theorem, i.e., no truncation errors
occur.
By applying (35) of Theorem 3.4 we can translate an outer harmonics expan-
sion with expansion center x0 such as
∞
n
F (x) = Fx∧,O
0
(n, m)On,m (x − x0 ) (37)
n=0 m=−n
which converges uniformly for x ∈ S2r0 ,ext (x0 ) with some r0 > 0. S2r0 ,ext (x0 ) denotes
the exterior of the sphere of radius r0 around x0 . The outer harmonics series
resulting from the translation possesses the expansion center x1 and the coefficients

n
n
Fx∧,O
1
(n
, m
) = Fx∧,O
0
(n, m)In∗ −n,m −m (x0 − x1 ). (38)
n=0 m=−n
This expansion converges uniformly for x ∈ S2r1 ,ext (x1 ) where

S2r1 ,ext (x1 ) ⊂ S2r0 ,ext (x0 ).
This translation of coefficients is called multipole to multipole translation (M2M).
Note that one can show that no further errors arise if the series in (37) has already
been truncated before the translation.
618 M. Gutting
By using formulation (34) of Theorem 3.4 we also find that the outer har-
monics expansion with expansion center x1 can be translated into an inner har-
monics series centered around x2 which converges uniformly for x ∈ S2r2 ,int (x2 ) if
the new ball of convergence is situated completely in S2r1 ,ext (x1 ), i.e., S2r1 ,int (x1 ) ∩
S2r2 ,int (x2 ) = ∅. The resulting coefficients of the inner harmonic expansion are
∞
n

Fx∧,I
2
(n
, m
) = Fx∧,O
1
(n, m)(−1)n +m On+n
∗
,m −m (x2 − x1 ) (39)
n=0 m=−n
and this translation is named multipole to local translation (M2L).

Finally, Theorem 3.5 lets us shift the expansion center of such inner harmonics
expansions to the new center x3 which possesses the coefficients
∞

n
Fx∧,I
3
(n
, m
) = Fx∧,I
2
(n, m)In−n ,m−m (x3 − x2 ). (40)
n=n m=−n
and converges uniformly for x ∈ S2r3 ,int (x3 ) ⊂ S2r2 ,int (x2 ). This translation step is
called local to local translation (L2L). For further details we refer to [67] and the
references therein, in particular [34].
3.4. The fast multipole algorithm

To start the algorithm a first set of multipole expansion coefficients for each cube
containing any points has to be computed. We consider only the part of the spline
related to a single cube X, i.e., the kernel functions KH (xi , ·), where xi ∈ X and
y ∈ Σext :
N
N KT
|y | 1 1
F (y) = ai KH (xi , y) = ai Dx . (41)
i=1
R 2πR
i=1
|x − hy KT |
x=xi
xi ∈X xi ∈X
We find the following expansion for |hy KT − x0 | > |xi − x0 |, xi ∈ X, i.e., if x0 is the
center of the cube X, the targets hy KT and the cube X need to fulfill a distance
requirement, i.e., targets must be contained in a well-separated cube.

|y KT | ai
N ∞
n
∗
F (y) = Dx In,m (x − x0 )On,m (hy KT − x0 )
R 2πR
i=1 n=0 m=−n x=xi
xi ∈X
KT ∞ n
|y |
= Fx∧,O
0
(n, m)On,m (hy KT − x0 ) (42)
R n=0 m=−n
where the multipole coefficients Fx∧,O

0
(n, m) of the cube X are given by

∧,O
N
ai ∗

Fx0 (n, m) = Dx In,m (x − x0 ) . (43)
2πR
i=1 x=xi
xi ∈X
This first step is called point to multipole (P2M) step where the infinite sum in (42)
has to be truncated at degree p. The degree p essentially determines the accuracy
of the algorithm. The coefficients Fx∧,O 0
(n, m) can be translated to other cubes via
relations (38), (39) as well as (40) as long as the distance requirements are fulfilled
by the construction of the decomposition of the domain into nested cubes.
At the end of the fast multipole cycle, i.e., after several M2M-, M2L-, L2L-
translations, each cube Y possesses an inner harmonics expansion centered around
the center of the cube. This expansion has to be evaluated at the targets contained
by Y . This evaluation is called the local to targets (L2T) step:

|y KT | ∧,I
p n

Lj F = F (yj ) = Fx0 (n, m)In,m (hy KT − x0 ) , (44)
R m=−n

n=0 y=yj
R2
where the variable y is hidden by y KT = |y|2 y.
Now we briefly summarize the fast multipole algorithm (see, e.g., [20, 24] or
[67, 69] for our specific implementation).
Algorithm 3.6 (Fast Multipole Algorithm).
Input:
• A set of points xi ∈ Σext (often xi ∈ Σ), i = 1, . . . , N ,
• a set of coefficients ai , i = 1, . . . , N ,
• the choice of the type of the reproducing kernel KH (singularity or Abel–
Poisson with the parameter h and the radius of the Runge sphere R),
• a set of evaluation points yj ∈ Σext , j = 1, . . . , M , where M = O(N ),
• the degree of the multipole expansion p,
• the maximal number of points per cube m.
Aim: compute the sum

N
F (yj ) = ai KH (xi , yj ) for each j = 1, . . . , M. (45)
i=1
Initialization:
2
• Compute the targets hyjKT = h |yRj |2 yj , j = 1, . . . , M .
• Create a bounding box that contains all points and all targets, build the
adaptive octtree and sort in all points and targets. Set L as the maximum
level, eliminate all empty cubes.
• Determine list 1 to list 4 of Definition 3.2. Create a list of all cubes of level l
for each level l = 0, . . . , L. Collect all leaves in a list.
• Allocate memory for the different expansion coefficients of each cube X:
multipole expansion (coefficient vector MX ), local expansion (coefficient vec-
tor LX ).
Fast multipole cycle:
1. Generation of the multipole coefficients:
For all leaves X: P2M, i.e., compute the multipole coefficients MX of the multipole
620 M. Gutting
expansion up to degree p around the center of X from the points in X as in (43).

For level l = L − 1, . . . , 2: M2M, i.e., translate the multipole coefficients of the
children of X to X itself for all cubes X of level l via (38).
2. Interaction phase for list 4:
For level l = 2, . . . , L: for all cubes X of level l: compute the expansion coefficients
of an inner harmonics expansion around the center of X from the points in Y for
all cubes Y of list 4 of X and add them to LX – or use direct evaluation of the
kernel sum corresponding to the points in Y to obtain the result at the targets in
X if the number of targets in X ≤ p2 and X is a leaf.
3. Multipole to local translation:
For level l = 2, . . . , L: for all cubes X of level l: use (39) to translate MX to LY
for all cubes Y in list 2 of X.
4. Translation of the inner harmonics expansions:
For level l = 2, . . . , L − 1: L2L, i.e., translate the local coefficients LX to the
children of X (if there are any) via (40) and add the resulting coefficients to LZ
where Z denotes the corresponding child of X for all cubes X of level l.
5. Evaluation of the expansions and direct interaction:
For all leaves X: L2T, i.e., evaluate the inner harmonics expansion of X at all
targets in X as in (44). Store the result in F .
For all cubes Y in list 1 of X: P2T, i.e., add the direct evaluation of the kernel
sum corresponding to the points in Y at the targets in X to F .
For all cubes Y in list 3 of X: evaluate the multipole expansion around the center
of Y (coefficients MY ) at the targets in X and add the results to F – or use direct
evaluation of the kernel sum corresponding to the points in Y to add the result at
the targets in X to F if the number of points in Y ≤ p2 and Y is a leaf.
6. Reverse the effects of the Kelvin transformation:
|y KT |
F4j = jR Fj for j = 1, . . . , M .
Return the result F4 .
For the computation of the spline coefficients of the spline approximation of
Section 2.4 we consider the system of linear equations (20) instead of (13). This
N
means that we have to add β ai (C −1 )ij to the matrix-vector product that is
i=1
computed by the FMM. In order to keep a fast algorithm the matrix C −1 has to
allow a fast summation method or C has to be a sparse matrix. The trivial cases
where C is a diagonal matrix can also be included in the direct evaluation step of
the fast multipole algorithm.
3.5. Acceleration of the translations
Newer iterations of the FMM include several ideas to reduce the numerical effort
of the translations from the original O(p4 ) to O(p3 ) or even O(p2 ) per translation
operation. This includes the ideas of [137] (see also [24, 64]) for the multipole
to multipole (M2M) and the local to local (L2L) steps using Wigner rotation
matrices (cf., e.g., [16, 25, 30, 130]). The main point is to rotate the coordinate
system such that the shift direction becomes the ε3 -axis, shift there and rotate
back. This reduces the numerical costs from O(p4 ) in the M2M- and L2L-steps to
O(p3 ), since each rotation as well as the shift along the ε3 -axis requires an effort of
O(p3 ). For a detailed description we refer to [137] or [67] with all technical details
of our implementation.
For the M2L translation, [64, 24] have replaced this step with exponential
translations which are based on the numerical integration of the integral represen-
tation
∞ 2π
1 1
= e−λ(x3 −y3 ) eiλ((x1 −y1 ) cos α)+(x2 −y2 ) sin α) dα dλ
|x − y| 2π 0 0

s(ε)
wk
Mk
= e−λk (x3 −y3 ) eiλk ((x1 −y1 ) cos αj,k )+(x2 −y2 ) sin αj,k ) + O(ε) (46)
Mk j=1
k=1
for points
# x, y whose Cartesian coordinates
√ satisfy 1 ≤ x3 − y3 ≤ 4 as well as
0 ≤ (x1 − y1 )2 + (x2 − y2 )2 ≤ 4 2. Details as well as integration points λk ,
weights wk and numbers of points Mk for the trapezoidal rule applied to the inner
integral can be found in [24, 64, 139]. The accuracy ε of the numerical integration
is determined by the discretization parameter s(ε) = O(p) of the outer integral
in (46). The total number of numerical integration points, i.e., the number of
exponential functions and coefficients, is supposed to be O(p2 ).
By Hobson’s formula (cf. [76]) a multipole expansion of F is transformed by
(46) into a series of exponentials (multipole to exponential step, briefly M2X),
these exponentials can be translated efficiently by the exponential to exponential
shift (X2X). Afterwards the local coefficients are computed from the exponential
coefficients (X2L). The restrictions on the positions of x and y mean that the
exponential translations are applicable for cubes in list 2 (see Definition 3.3) that
are situated above the current cube with another cube in between.
However, by combining rotations of the multipole expansion using again the
Wigner rotation matrices, the exponential translation can substitute the M2L
translation for all cubes in list 2. Therefore, the list of all well-separated cubes
(list 2) is split into 6 directional lists (up, down, North, South, East and West)
and instead of M2L the following sequence of transformations is used: (rotation),
M2X, X2X, X2L, (inverse rotation).
Each exponential shift requires numerical costs of O(p2 ) and the rotations
can be applied using O(p3 ) operations (as do the M2X and X2L steps). Thus, this
improves the performance compared to the M2L step’s O(p4 ) effort. Moreover,
we can save translations by recombination (see [24, 64, 67, 69] for more on the
technical details). It should also be noted that there are several symmetries in the
coefficients of the exponential expansion since we are dealing with a real-valued
function F . These symmetries can be used to further reduce the constant of the
numerical costs (cf. [24, 64]).
622 M. Gutting
3.6. Parameters and results of our FMM implementation

Here we present just a few of the parameters that we use in our implementation
of the FMM with exponential translations. For more detailed investigations of our
version (in particular of the recombination of exponential translations mentioned
at the end of Section 3.5) we refer to [67, 68, 69].
The truncation degree p is investigated for different accuracies of the expo-
nential translation s(ε). We increase p while s(ε) is kept fixed and determine when
the integration error of the numerical integration in the exponential translation
(46) dominates the truncation error. This leads to the choices of p for different
levels of s(ε) given by Table 1. Note that the kernels of Abel–Poisson type require
a slightly higher degree and therefore a bit more numerical effort.
s(ε) Singularity kernel Abel–Poisson kernel

8 4 5
17 12 13
26 23 25
Table 1. Resulting truncation degrees p for different s(ε) for the two
types of kernels.
The maximal number of points or targets per cube m has a strong influence
on the adaptive octtree construction and the performance of the FMM. If m is too
small, there are many cubes each containing only very few points. Thus, the kernel
expansion coefficients no longer combine the information of enough points to be
efficient. If m is too large, there are only few cubes each with a large number of
points. This means that far too often instead of kernel expansion direct interaction
is used. Therefore, m can be used to balance the effort of the direct interaction
and the kernel approximation. It has been suggested to choose m = O(p3/2 ) (cf.
[24, 64]) which may serve as a guideline. Many empirical tests (cf. [67]) led us to
the conclusion that the choices for m given by Table 2 provide a good performance

8 85 75
17 130 140
26 380 240
Table 2. Chosen maximal numbers of points m per cube for the sin-
gularity kernel and the Abel–Poisson kernel and the different error lev-
els.
Figure 3. Break-even points by comparison of computation times for

direct (solid line) and FMM accelerated (dashed line) computation (top:
Singularity kernel, bottom: Abel–Poisson kernel), the number of points
forms the abscissae.
in our implementation. Note again that there are remarkable differences between
the two different types of kernels under investigation. Using these optimizations of
the parameters of the FMM we compare its performance with direct computation
and find the break-even points of our implementation, i.e., the minimal number
of points that is necessary for our algorithm to be faster than the direct approach
(see Table 3).

8 530 360
17 1,160 960
26 2,670 2,250
Table 3. Break-even points for the singularity kernel and the Abel–
Poisson kernel.
624 M. Gutting
Note that such results are always very dependent on the implementation. Our
implementation turns out to be efficient even for rather small problem sizes. In
general, the Abel–Poisson kernel requires some more computational time since it
leads to a more difficult P2M-step. Finally, we show the linear asymptotic behavior
which we expect from the FMM in Figure 3 compared to the quadratic behavior
of the direct approach.
4. Parameter choice methods for spline approximation

In this section we summarize several parameter choice methods for the determina-
tion of the smoothing parameter in (18) and (20) that are known from the theory
of regularization of inverse problems. For this context the reader is referred to
[9, 14] where also many numerical tests for a wide range of inverse problems are
presented.
We consider the problem of solving Ax = y where A : X → Y is a linear
operator. In our case X = Y = RN and A is the matrix of the linear system (13)
(in [9, 14] this is called case C2). This is in fact not an ill-posed problem, but can be
severely ill-conditioned (depending on the distribution of the data points) and as
seen in Section 2.4 noisy data lead to the use of a form of Tikhonov regularization,
i.e., the linear system (20) with the smoothing parameter β > 0. It should be
noted that we want to use the FMM for a fast matrix-vector-multiplication, i.e.,
A is never given as an actual matrix. This excludes some of the parameter choice
methods in order to keep reasonably low computational costs. Moreover, it should
be pointed out that the use of an iterative solver (e.g., cg-method or GMRES)
requires a stopping criterion and the number of iterations can be seen as another
regularization parameter which needs to be chosen in some way.
The smoothing parameter β > 0 is discretized exponentially, i.e., βn = β0 qβn
with qβ ∈ (0, 1) and n ∈ N (actually only n = 1, . . . , nmax ). The use of a discrete
set of regularization parameters with a fine enough resolution does not alter the
behavior of most parameter choice methods. For the efficient implementation of
these methods, it is useful to have a bound on the value of the optimal parameter
(i.e., a maximal regularization parameter), especially if the method minimizes some
function (see also [9, 14] and Section 4.3). We denote the vector of noisy data by
y δ (see Section 4.1) and xδn ∈ X is the vector of spline coefficients resulting from
the minimization of the smoothing functional (18) using βn and y δ , i.e., xδn solves
the system of linear equations (20):
xδn = (A + βn I)−1 y δ = A−1 δ
n y , (47)
where An = A + βn I. Furthermore, x0n = A−1
n y with noise-free data y ∈ Y.
4.1. Noise models
We investigate additive noise models, i.e., y δ = y + δξ, where ξ is a normalized
noise element and δ > 0 is the noise level. The most common noise model in the
classical inverse problems literature is deterministic noise (cf. [33]), where ξ ∈ Y
with ξ ≤ 1, so y δ − y ≤ δ. This models discretization errors, but only poorly

represents random measurement errors arising in practice.
A stochastic noise model for a discrete data vector y δ ∈ RN (see [135])
uses ξ, where the components ξi are i.i.d. random variables with mean Eξi = 0
and variance Eξi2 = 1. δ is the standard deviation of each error component δξi
and E y δ − y 2 = δ 2 E ξ 2 = N δ 2 . Note that for the number of data N → ∞
this is unbounded. For correlated errors, δξi possesses the covariance matrix C. If
known, this matrix can be used in (20). We restrict ourselves to diagonal C here.
If y δ = y + δξ with ξ ∼ N (0, I), the noise model corresponds to Gaussian white
noise. For colored noise, if the entries Ckk are increasing, it is called blue noise,
and, if they are decreasing, it is called red noise (see [9] and the references therein
for more details).
4.2. Parameter choice methods

A parameter choice method is a rule that assigns a value for the regulariza-
tion/smoothing parameter. In case of a discrete set of parameters, the method
selects a value for the index, which will be denoted by n∗ . Parameter choice meth-
ods can be classified as three basic types by their input (see, e.g., [8, 33]):
• a priori method, i.e., n∗ is a function of δ and information about x which is
not known in practice. Thus, we do not discuss such methods here.
• a posteriori method, i.e., n∗ = n∗ (δ, y δ ) requiring the noise level δ. If δ is not
known, an estimate is used.
• data-driven method, i.e., n∗ = n∗ (y δ ) requiring only the data y δ as input
(sometimes called “heuristic method”).
If y δ contains stochastic noise, then n∗ is a random variable. n∗ is defined as
either the point at which a function F (n) falls below a threshold (Type 1) or the
minimizer of F (Type 2). Some methods need (sensitive) tuning parameters (in
particular methods of type 1).
It should be pointed out that [4] states that, for an ill-posed problem, a
parameter choice rule that does not explicitly use the noise level (e.g., data-driven
methods) cannot yield a regularization method such that the worst case error
converges to 0 as δ → 0. This Bakushinskii veto is important for deterministic
noise, but it is not really appropriate for stochastic noise (cf. [8, 15]). There are
data-driven rules yielding regularization methods that converge and perform very
well in practice (see also [9]).
Since x = A−1 y is unknown (noise-free data y is not available), a practical pa-
rameter choice method must use some other known or easily computed/estimated
quantities such as, e.g., the norm of the residual defined as y δ − Axδn . In our case
this is a Euclidean norm which is easily computed. Splitting the error x − xδn
such that
x − xδn ≤ x − x0n + x0n − xδn , (48)
626 M. Gutting
the first term (regularization error) is usually bounded by a decreasing function

ϕ(n) reflecting smoothness properties and the so-called qualification of the regu-
larization method (see [9, 33] and the references therein). The second term (prop-
agated noise error) on the right-hand side of (48) can often be bounded for regu-
larization methods as
x0n − xδn ≤ δ(n), (49)
where is a known increasing function of n, indicating that, with less smoothing,
there is more influence of the data noise (cf. [33]).
In the case of stochastic noise, the risk, i.e., the expected squared error E x−
xδn 2 , is considered. For noise with zero mean, instead of (48), the risk can be
decomposed exactly into a sum of squared bias x − x0n 2 and variance terms
E x0n − xδn 2 , i.e.
E x − xδn 2 = x − x0n 2 + E x0n − xδn 2 . (50)
The squared bias can be bounded as before and, under suitable assumptions, the
variance can be expressed as δ 2 2 (n) for some increasing function (n). For white
noise, our Tikhonov regularized solution (47) has variance
1
δ 2 2 (n) = δ 2 E A−1 2
n ξ = δ
2
, (51)
(σk + βn )2
k
where σk are the singular values of A (in our case they are eigenvalues of A).
A much more detailed discussion of the above errors (including, e.g., minimax
results) in various situations can be found in [5, 9, 23, 26, 33, 78, 95, 103, 105].
For some methods, there are stronger results involving oracle inequalities (see
[12, 19, 21, 22]), which provide, for any noise level, a bound on the risk E x − xδn 2
relative to the smallest possible value of the risk, and allow the classification of
methods as asymptotically optimal.
4.3. Maximal regularization parameter
Some parameter choice methods, e.g., the balancing principle (Section 4.4.2), re-
quire a maximal index nmax as essential input in the algorithm itself. [9] suggest
to define the maximal index as
nmax = max{n|(n) < 0.5(∞)}, (52)
where E x0n − xδn 2 2 2 2 2
= δ (n) and δ (∞) is the supremum of the variance. This
allows us to expect that the optimal parameter index is smaller than nmax . Also
note that nmax should not be too large for an efficient computation.
We obtain nmax either by an analytic expression for δ 2 2 (n), as in (51) for
white noise, or by a good estimate of it. If there are several independent data sets
available, then a good estimate (for any noise color) is
δ 2 2 (n) ≈ 2−1 mean{ xδn,i − xδn,j 2 , i = j}. (53)
Often two sets of data are sufficient (see [5] for further details). If only a single
data set is available, then it may not be possible to estimate δ 2 2 (n) if the noise is
correlated with unknown covariance. Then one can define a maximal index ñmax by
βñmax ≈ σñmax for our Tikhonov regularization if there are at least good estimates
of the eigenvalues σk of A available. However, methods that perform much worse
without the use of the maximal index nmax , may yield different results for ñmax
instead of nmax (see [9, 14]).
4.4. Description and evaluation of methods

In this section, we describe the origin and idea of the method, list the input
of the method and the algorithm that we use. We also give a brief discussion
of known theoretical and practical issues about the method, including the most
relevant references. Several of the methods use a tuning parameter or some other
parameter that must be chosen (see [9, 14] and the references therein for further
details).
4.4.1. Discrepancy principle. The discrepancy principle of [115, 110, 111] is one
of the oldest and most widely used parameter choice procedures (cf. [33] and
references therein). Its idea is that for a good regularized solution, the norm of
the residual should match the noise level δ of the data. The method needs the
following input:
• Norms of residuals {Axδn − y δ }n≤nmax until a certain bound is satisfied.
• Noise level δ.
• Tuning parameter τ ≥ 1.
In a deterministic setting with y δ − y ≤ δ, the parameter choice n∗ is the first n
such that Axδn − y δ ≤ τ δ. In a stochastic setting, with the error in each element
of y δ ∈ RN having standard deviation δ, the choice n∗ is the first n such that
√
Axδn − y δ ≤ τ δ N . (54)
Originating from a deterministic setting, the discrepancy principle has also been
studied in stochastic settings (see, e.g., [17, 29, 97, 132]) and for many regulariza-
tion methods and many inverse problems. There are many results on convergence
properties of this method for both settings (see, e.g., [33, 65, 79, 104, 111, 112, 114]
for the deterministic case and [29, 97, 98, 132] for the stochastic case).
The discrepancy principle is one of the fastest methods available, since one
only needs to compute the residuals until the bound (54) is satisfied which allows
the use of the FMM of Section 3. However, its drawback is the requirement of an
accurate estimate of the noise level. Estimations that are just slightly off can lead
to very poor solutions (see [73, Chap. 7]).
There are also many variants of the method such as the transformed discrep-
ancy principle (cf. [119, 120, 71]), the modified discrepancy principle (MD rule)
(cf. [32, 53, 117, 118]), or the varying discrepancy principle (cf. [17, 94]). Their
main drawback is that they are no longer easily compatible with the FMM. For
comparative studies in the context of inverse problems with stochastic noise we
refer, e.g., to [9, 14].
628 M. Gutting
4.4.2. Balancing principle. The balancing principle of [90] was originally derived
for statistical estimation from direct observations in a white noise model. Since
then it has been developed further for regularization of linear and nonlinear inverse
problems (see, e.g., [6, 13, 11, 58, 129, 105, 106]) in deterministic and stochastic
settings. The idea is to balance the known propagated noise error bound δ(n) in
(49) with the unknown regularization error (48) by an adaptive procedure that
employs a collection of differences of regularized solutions. As input the balancing
principle needs:
• Maximal index nmax , e.g., as defined in (52).
• All regularized solutions {xδn }n≤nmax up to the index nmax .
• An upper bound δ(n) for the propagated noise error x0n − xδn or a bound
or estimate δ 2 2 (n) of the variance E x0n − xδn 2 .
• Noise level δ (and the covariance in the stochastic setting if known). Then one
can use known expressions for δ(n). Alternatively, if one has two or more
independent sets of data yiδ , then E x0n − xδn 2 can be estimated by (53).
• Tuning constant κ, typically κ ∈ [0.5, 1.5] (cf. [9] and the references therein).
The balancing functional is defined by
! −1 "
b(n) = max 4 xn − xk /(δ(k)) . (55)
n<k≤nmax
The smoothed balancing functional (which is monotonously decreasing) is defined

as B(n) = maxn≤k≤nmax {b(k)}. Then the parameter choice n∗ is the first n such
that B(n) ≤ κ.
The balancing principle is one of the few parameter choice methods for which
oracle inequalities for the error are known (cf. [121, 13]), i.e., there are stronger
results than rates of convergence alone. For variants we refer to [10, 105] and for
comparative studies we mention [9, 14].
One variant should be noted: The hardened balancing principlewhich is a
modified version of the balancing principle in the stochastic setting first proposed
in [5]. The input is the same as before, but without the tuning parameter and the
noise level. Furthermore, an expression or approximation of the scaled variance
2 (n) = δ −2 E x0n − xδn 2 , or any scalar multiple of this (so δ can be unknown),
is required. The balancing functional b(n) is defined as in (55) and the smoothed
balancing functional B(n) = maxn≤k≤nmax {b(k)}. The parameter choice is
#
n∗ = argminn≤nmax B(n) (n) , (56)
where any scalar multiple of (n) gives the same choice. The method has the
advantage that it does not require a tuning parameter. Numerical experiments in
[5, 12, 9, 14] indicate that the method is very stable even for colored noise.
4.4.3. Quasi-optimality criterion. The quasi-optimality criterion by Tikhonov and

Arsenin [127], Tikhonov and Glasko [128] (see also [77]) is one of the oldest and
simplest available parameter choice methods. An overview of the method and its
history can be found, e.g., in [8]. As input for the minimization the following is
required:
• Maximal index nmax , e.g., as defined in (52).
• All regularized solutions {xδn }n≤nmax up to nmax .
The noise level does not need to be known, and there is no tuning parameter. The
parameter choice n∗ is defined simply as
! "
n∗ = argminn≤nmax xδn − xδn+1 . (57)
The well-known continuous version
- for
- Tikhonov regularization defines the param-
- d δ-
eter choice by β∗ = argmin -β dβ xβ -. Using a difference quotient in place of the
derivative for the discrete parameters βn = β0 qβn we obtain (57).
For a discrete set of regularization parameters, the use of a suitable maximal
index nmax is essential, because the method is based on a discrete evaluation
of a differential. Hence is very sensitive to a situation where the regularization
operators A−1n are formally different, but are practically the same (cf. [9] and the
references therein). Convergence results for the Tikhonov regularization with the
quasi-optimality criterion for different settings can be found in [89, 56], for further
convergence properties see [7, 84, 85, 113].
4.4.4. L-curve method. The L-curve method, proposed by Hansen [72, 73] and
Hansen and O’Leary [74], is based on the long-known fact that a log-log parametric
plot of ( Axδn − y δ , xδn ) often has a distinct L-shape (cf. [88]). Points on the
vertical part correspond to large n (under-smoothed solutions) and those on the
horizontal part correspond to small n (over-smoothed solutions), which suggests
that the “corner point” of the L-curve should define a good value of the parameter
n. It is usually applied manually and can provide good results then whereas finding
the L-curve corner is hard to automate. As input to minimize a certain function
the following is used:
• Norms of all residuals {Axδn − y δ }n≤nmax .
• Norms of the regularized solutions {xδn }n≤nmax .
The noise level does not need to be known. The parameter choice can be defined
by the product of the norms of the residual and regularized solution, i.e.,
! "
n∗ = argminn≤nmax Axδn − y δ · xδn . (58)
Here the “corner point” is defined by the slope of its “tangent” being −1 as in
[122] (see also [33]). The generalizations minimize Axδn − y δ xδn τ (see [122, 93]),
where τ is a tuning constant.
Since “corner point” is not a well-defined notion, several algorithms have
evolved with different definitions (see [74, 18, 75, 122]). [93] derived first rigorous
optimality results for the L-curve criterion. In many (but not all) problems, vari-
ants of the L-curve method has been observed to give a reasonably good parameter
choice which can deal with correlated errors. See [9] for an overview of references
where the method works or runs into severe limitations.
630 M. Gutting
4.4.5. Generalized cross-validation. Generalized cross-validation (GCV), due to

[133], is a popular method for problems with discrete data and stochastic noise as
(20). It goes back to ordinary cross-validation, where the idea is to consider all the
“leave-one-out” regularized solutions and choose the parameter that minimizes
the average of the squared prediction errors using each solution to predict the
missing data value. These calculations do not require the computation of all the
regularized solutions. Weighting the prediction errors, [27, 60, 133, 135] derived
the GCV method, which has is invariant under orthogonal transformations of the
data. Some other parameter choice methods proposed in the literature have been
shown to be closely related to GCV, in particular the Akaike information criterion
(AIC) of [2, 35]. As input to minimize a certain function we need:
• Sums of squares of all the residuals {Axδn − y δ }n≤nmax where y δ ∈ RN .
• The trace of the influence matrix AA−1 δ
n mapping y to Axn .
δ
The noise level does not need to be known. The GCV parameter estimate is de-
fined by

Axδn − y δ 2
n∗ = argminn≤nmax . (59)
(N −1 tr(I − AA−1 n ))
2
GCV is closely related to and behaves like the unbiased prediction risk method
(also known as Mallows Cp or CL ; see [31, 92, 135]). It is asymptotically optimal
with respect to the prediction risk as N → ∞ for stochastic white noise and the
Tikhonov regularization (cf. [66, 91, 96, 132, 136]). The GCV method has been
used widely and performs very well for reasonably large data sets with uncorrelated
errors (white noise). However, it is known (see, e.g., [31, 86, 87, 99, 101, 126, 135])
that for smaller data sets or correlated errors of red noise type, the method is
rather unstable, often resulting in under-smoothing.
The term tr(AA−1 n ) in the GCV function is a measure of the degrees of free-
dom in the regularized solution. For its fast computation making use of the FMM
trace estimation methods are needed that use stochastic (Monte-Carlo) algorithms
(cf. [54, 55, 59, 80, 81]).
In order to overcome the instability of GCV, several variants have evolved.
The robust GCV (RGCV) method has been developed and investigated in [99,
100, 123]. It needs the same input as for GCV and additionally:
• The trace of the square of the influence matrix (AA−1 2
n ) .
• A robustness parameter γ ∈ (0, 1). Note that with γ = 1 the RGCV method
is just GCV.
The RGCV parameter estimate is defined by minimizing a certain function:

Axδn − y δ 2 −1 −1 2

n∗ = argmin γ + (1 − γ)N tr((AAn ) ) . (60)
n≤nmax (N −1 tr(I − AA−1n ))
2
The family of robust GCV methods developed in [100] also includes the strong
robust GCV method, denoted R1 GCV. As input one needs the same as for GCV
as well as
• The trace of A−1

n
∗ −1
An .
• A robustness parameter γ ∈ (0, 1). Note that for γ = 1 the R1 GCV method
is just GCV.
The R1 GCV parameter estimate is defined by minimizing a certain function:

Axδn − y δ 2 −1 −1 ∗ −1

n∗ = argmin γ + (1 − γ)N tr((A A ) . (61)
(N −1 tr(I − AA−1 2 n n
n≤nmax n ))
The modified GCV method involves a simple modification of the GCV func-
tion that is designed to stabilize the method (cf. [28, 131]). The inputs are the
same as for GCV plus:
• A stabilization parameter c > 1. For c = 1 the method reduces to GCV.
The noise level does not need to be known. The modified GCV estimate is de-
fined by

Axδn − y δ 2
n∗ = argminn≤nmax . (62)
(N −1 tr(I − cAA−1n ))
2
For comparative studies of these variants and further details we refer to [9, 14]
and the references therein.
5. Conclusion
Using the FMM in an iterative algorithm like, e.g., conjugate gradients or GMRES
is an efficient solution strategy that can treat interpolation problems and Dirichlet
boundary value problems with many data points on regular surfaces (e.g., the ac-
tual topography of the Earth) (see [67, 68, 69]). It should be pointed out that this
spline approach is not restricted to a global treatment, but also applies to regional
domains (cf. [67, 68]). This can lead to a local improvement of the gravitational
field in areas of particular interest. The approach can be extended to spline ap-
proximation (in particular for diagonal covariance matrices) as seen in Section 2.4
and the end of Section 3.4).
The smoothing parameter(s) plays a crucial role in this approach and must
be chosen very carefully or a lot of information is lost to oversmoothing, in partic-
ular the high-frequent details of the signal. We have presented several parameter
choice methods the can be used without losing the advantages of the FMM. Their
performance for the regularization of inverse problems has been investigated in
several studies with different solution techniques (see, e.g., [9, 14, 70] and the ref-
erences therein). Tests of the combination of the FMM with these parameter choice
methods (cf. [9, 14, 70] and the references therein) are an interesting challenge for
the future. In particular the interaction with stopping criteria for iterative solvers
needs further investigation. Note that often the solution (even using the FMM)
requires much more computational effort than the evaluation of the parameter
choice. It can be advisable to apply several parameter choice methods to find the
best choice of the parameter.
632 M. Gutting
For highly irregular distributions of data points, the spline approach reaches
its limits due to large data gaps which result in severe ill-conditioning. Even
smoothing splines cannot completely bridge this gap so far though further in-
vestigation is required. However, functional matching pursuit methods (RFMP or
ROFMP) can result in better approximations (see [36, 70, 108] and the references
therein), but so far these algorithms require high numerical costs. These meth-
ods are also iterative regularizations and the combination of stopping criteria and
regularization parameters has been investigated for a class of ill-posed problems
in [70].
References
[1] M. Abramowitz and I.A. Stegun. Handbook of Mathematical Functions. Dover Pub-
lications, Inc., New York, 1972.
[2] H. Akaike. Information theory and an extension of the maximum likelihood princi-
ple. In Second International Symposium on Information Theory (Tsahkadsor, 1971),
pages 267–281. Akadémiai Kiadó, Budapest, 1973.
[3] N. Aronszajn. Theory of reproducing kernels. Trans. Amer. Math. Soc., 68:337–404,
1950.
[4] A.B. Bakushinskii. Remarks on choosing a regularization parameter using the quasi-
optimality and ratio criterion. U.S.S.R. Comput. Math. Math. Phys., 24:181–182,
1984.
[5] F. Bauer. Some considerations concerning regularization and parameter choice al-
gorithms. Inverse Prob., 23(2):837–858, 2007.
[6] F. Bauer and T. Hohage. A Lepskij-type stopping rule for regularized Newton
methods. Inverse Prob., 21:1975–1991, 2005.
[7] F. Bauer and S. Kindermann. The quasi-optimality criterion for classical inverse
problems. Inverse Prob., 24:035002, 20 pp., 2008.
[8] F. Bauer and S. Kindermann. Recent results on the quasi-optimality principle. J.
Inverse Ill-Posed Prob., 17(1):5–18, 2009.
[9] F. Bauer and M.A. Lukas. Comparing parameter choice methods for regularization
of ill-posed problems. Math. Comput. Simul., 81(9):1795–1841, 2011.
[10] F. Bauer and A. Munk. Optimal regularization for ill-posed problems in metric
spaces. J. Inverse Ill-Posed Probl., 15(2):137–148, 2007.
[11] F. Bauer and S. Pereverzev. Regularization without preliminary knowledge of
smoothness and error behavior. European J. Appl. Math., 16(3):303–317, 2005.
[12] F. Bauer and M. Reiß. Regularization independent of the noise level: an analysis
of quasi-optimality. Inverse Prob., 24:055009, 16 pp., 2008.
[13] F. Bauer, T. Hohage, and A. Munk. Iteratively regularized Gauss–Newton method
for nonlinear inverse problems with random noise. SIAM J. Numer. Anal., 47(3):
1827–1846, 2009.
[14] F. Bauer, M. Gutting, and M.A. Lukas. Evaluation of parameter choice methods
for regularization of ill-posed problems in geomathematics. In W. Freeden, M.Z.
Nashed, and T. Sonar, editors, Handbook of Geomathematics. 2nd ed., pages 1713–
1774. Springer, Heidelberg, 2015.
[15] S.M.A. Becker. Regularization of statistical inverse problems and the Bakushinkii
veto. Inverse Prob., 27:115010, 22 pp., 2011.
[16] L.C. Biedenharn and J.D. Louck. Angular Momentum in Quantum Physics (Theory
and Application). Encyclopedia of Mathematics and its Applications. Addison-
Wesley, Reading, Massachusetts, 1981.
[17] G. Blanchard and P. Mathé. Discrepancy principle for statistical inverse problems
with application to conjugate gradient iteration. Inverse Prob., 28:115011, 23 pp.,
2012.
[18] D. Calvetti, P.C. Hansen, and L. Reichel. L-curve curvature bounds via Lanczos
bidiagonalization. Electron. Trans. Numer. Anal., 14:20–35, 2002.
[19] E.J. Candès. Modern statistical estimation via oracle inequalities. Acta Numerica,
15:257–325, 2006.
[20] J. Carrier, L. Greengard, and V. Rokhlin. A fast adaptive multipole algorithm for
particle simulations. SIAM J. Sci. Stat. Comput., 9(4):669–686, 1988.
[21] L. Cavalier. Nonparametric statistical inverse problems. Inverse Prob., 24(3):
034004, 19 pp., 2008.
[22] L. Cavalier, G.K. Golubev, D. Picard, and A.B. Tsybakov. Oracle inequalities for
inverse problems. Ann. Statist., 30(3):843–874, 2002.
[23] L. Cavalier, Y. Golubev, O. Lepski, and A. Tsybakov. Block thresholding and sharp
adaptive estimation in severely ill-posed inverse problems. Theory Probab. Appl.,
48(3):426–446, 2004.
[24] H. Cheng, L. Greengard, and V. Rokhlin. A fast adaptive multipole algorithm in
three dimensions. J. Comput. Phys., 155:468–498, 1999.
[25] C.H. Choi, J. Ivanic, M.S. Gordon, and K. Ruedenberg. Rapid and staple deter-
mination of rotation matrices between spherical harmonics by direct recursion. J.
Chem. Phys., 111(19):8825–8831, 1999.
[26] D.D. Cox. Approximation of method of regularization estimators. Ann. Statist., 16
(2):694–712, 1988.
[27] P. Craven and G. Wahba. Smoothing noisy data with spline functions. Numer.
Math., 31:377–403, 1979.
[28] D.J. Cummins, T.G. Filloon, and D. Nychka. Confidence intervals for nonpara-
metric curve estimates: Toward more uniform pointwise coverage. J. Amer. Statist.
Assoc., 96(453):233–246, 2001.
[29] A.R. Davies and R.S. Anderssen. Improved estimates of statistical regularization
parameters in Fourier differentiation and smoothing. Numer. Math., 48:671–697,
1986.
[30] A.R. Edmonds. Drehimpulse in der Quantenmechanik. Bibliographisches Institut,
Mannheim, 1964.
[31] B. Efron. Selection criteria for scatterplot smoothers. Ann. Statist., 29(2):470–504,
2001.
634 M. Gutting
[32] H.W. Engl and H. Gfrerer. A posteriori parameter choice for general regularization
methods for solving linear ill-posed problems. Appl. Numer. Math., 4(5):395–417,
1988.
[33] H.W. Engl, H. Hanke, and A. Neubauer. Regularization of Inverse Problems.
Kluwer, Dordrecht, 1996.
[34] M.A. Epton and B. Dembart. Multipole translation theory for the three-dimen-
sional Laplace and Helmholtz equations. SIAM J. Sci. Comput., 16 (4): 865–897,
1995.
[35] R.L. Eubank. Spline Smoothing and Nonparametric Regression. Marcel Dekker,
New York, 1988.
[36] D. Fischer and V. Michel. Sparse regularization of inverse gravimetry – case study:
Spatial and temporal mass variation in South America. Inverse Prob., 28:065012,
34 pp., 2012.
[37] W. Freeden. On approximation by harmonic splines. Manuscripta Geod., 6:193–244,
1981a.
[38] W. Freeden. On spherical spline interpolation and approximation. Math. Method.
Appl. Sci., 3:551–575, 1981b.
[39] W. Freeden. Interpolation and best approximation by harmonic spline functions.
Boll. Geod. Sci. Aff., 1:105–120, 1982a.
[40] W. Freeden. On spline methods in geodetic approximation problems. Math. Method.
Appl. Sci., 4:382–396, 1982b.
[41] W. Freeden. Spherical spline interpolation: Basic theory and computational aspects.
J. Comput. Appl. Math., 11:367–375, 1984a.
[42] W. Freeden. Ein Konvergenzsatz in sphärischer Spline-Interpolation. Z. f. Vermes-
sungswes. (ZfV), 109:569–576, 1984b.
[43] W. Freeden. A spline interpolation method for solving boundary value problems of
potential theory from discretely given data. Numer. Methods Partial Differential
Equations, 3:375–398, 1987a.
[44] W. Freeden. Harmonic splines for solving boundary value problems of potential
theory. In J.C. Mason and M.G. Cox, editors, Algorithms for Approximation, pages
507–529. The Institute of Mathematics and its Applications, Conference Series, Vol.
10, Clarendon Press, Oxford, 1987b.
[45] W. Freeden. Multiscale Modelling of Spaceborne Geodata. B.G. Teubner, Stuttgart,
Leipzig, 1999.
[46] W. Freeden and C. Gerhards. Geomathematically Oriented Potential Theory. Chap-
man & Hall/CRC, Boca Raton, 2013.
[47] W. Freeden and M. Gutting. Special Functions of Mathematical (Geo-)Physics.
Birkhäuser, Basel, 2013.
[48] W. Freeden and V. Michel. Multiscale Potential Theory (with Applications to
Geoscience). Birkhäuser, Boston, Basel, Berlin, 2004.
[49] W. Freeden and M. Schreiner. Special functions in mathematical geosciences: An
attempt at a categorization. In W. Freeden, M.Z. Nashed, and T. Sonar, editors,
Handbook of Geomathematics. 1st ed., pages 925–948. Springer, Heidelberg, 2010.
[50] W. Freeden, M. Schreiner, and R. Franke. A survey on spherical spline approxima-

tion. Surv. Math. Ind., 7:29–85, 1997.
[51] W. Freeden, T. Gervens, and M. Schreiner. Constructive Approximation on
the Sphere (with Applications to Geomathematics). Oxford Science Publications,
Clarendon, 1998a.
[52] W. Freeden, O. Glockner, and M. Schreiner. Spherical panel clustering and its
numerical aspects. J. Geodesy, 72:586–599, 1998b.
[53] H. Gfrerer. An a posteriori parameter choice for ordinary and iterated Tikhonov
regularization of ill-posed problems leading to optimal convergence rates. Math.
Comput., 49(180):507–522, 1987.
[54] D. Girard. A fast “Monte-Carlo cross-validation” procedure for large least squares
problems with noisy data. Numer. Math., 56(1):1–23, 1989.
[55] D. Girard. Asymptotic comparison of (partial) cross-validation, gcv and randomized
gcv in nonparametric regression. Ann. Statist., 26(1):315–334, 1998.
[56] V. Glasko and Y. Kriksin. On the quasioptimality principle for linear ill-posed
problems in Hilbert space. Vychisl. Mat. Mat. Fiz., 24:1603–1613, 1984.
[57] O. Glockner. On Numerical Aspects of Gravitational Field Modelling From SST
and SGG by Harmonic Splines and Wavelets (with Application to CHAMP Data).
Ph.D.-thesis, Geomathematics Group, Department of Mathematics, University of
Kaiserslautern. Shaker, Aachen, 2002.
[58] A. Goldenshluger and S. Pereverzev. Adaptive estimation of linear functionals in
Hilbert scales from indirect white noise observations. Prob. Th. Rel. Fields, 118(2):
169–186, 2000.
[59] G.H. Golub and U. von Matt. Generalized cross-validation for large scale problems.
J. Comput. Graph. Statist., 6(1):1–34, 1997.
[60] G.H. Golub, M. Heath, and G. Wahba. Generalized cross-validation as a method
for choosing a good ridge parameter. Technometrics, 21:215–223, 1979.
[61] L. Greengard. The Rapid Evaluation of Potential Fields in Particle Systems. MIT
Press, Cambridge, MA, 1988.
[62] L. Greengard and V. Rokhlin. A fast algorithm for particle simulations. J. Comput.
Phys., 73(1):325–348, 1987.
[63] L. Greengard and V. Rokhlin. Rapid evaluation of potential fields in three dimen-
sions. In C. Anderson and L. Greengard, editors, Vortex Methods, pages 121–141.
Springer, 1988.
[64] L. Greengard and V. Rokhlin. A new version of the fast multipole method for the
Laplace equation in three dimensions. Acta Numerica, 6:229–269, 1997.
[65] C.W. Groetsch. The Theory of Tikhonov Regularization for Fredholm Equations of
the First Kind. Pitman, Boston, 1984.
[66] C. Gu. Smoothing Spline ANOVA Models. Springer, New York, 2002.
[67] M. Gutting. Fast Multipole Methods for Oblique Derivative Problems. Ph.D.-thesis,
Geomathematics Group, Department of Mathematics, University of Kaiserslautern.
Shaker, Aachen, 2007.
[68] M. Gutting. Fast multipole accelerated solution of the oblique derivative boundary
value problem. GEM – Int. J. Geom., 3(2):223–252, 2012.
636 M. Gutting
[69] M. Gutting. Fast spherical/harmonic spline modeling. In W. Freeden, M.Z. Nashed,

and T. Sonar, editors, Handbook of Geomathematics. 2nd ed., pages 2711–2746.
Springer, Heidelberg, 2015.
[70] M. Gutting, B. Kretz, V. Michel, and R. Telschow. Study on parameter choice
methods for the RFMP with respect to downward continuation. Front. Appl. Math.
Stat., 3, doi:10.3389/fams.2017.00010, 17 pp., 2017.
[71] U. Hämarik and T. Raus. On the choice of the regularization parameter in ill-posed
problems with approximately given noise level of data. J. Inverse Ill-Posed Probl.,
14(3):251–266, 2006.
[72] P.C. Hansen. Analysis of discrete ill-posed problems by means of the L-curve. SIAM
Rev., 34(4):561–580, 1992.
[73] P.C. Hansen. Rank-Deficient and Discrete Ill-Posed Problems. SIAM, Philadelphia,
1998.
[74] P.C. Hansen and D.P. O’Leary. The use of the L-curve in the regularization of
discrete ill-posed problems. SIAM J. Sci. Comput., 14(6):1487–1503, 1993.
[75] P.C. Hansen, T.K. Jensen, and G. Rodriguez. An adaptive pruning algorithm for
the discrete L-curve criterion. J. Comput. Appl. Math., 198(2):483–492, 2007.
[76] E.W. Hobson. The Theory of Spherical and Ellipsoidal Harmonics (Second Reprint).
Chelsea Publishing Company, New York, 1965.
[77] B. Hofmann. Regularization of Applied Inverse and Ill-Posed Problems. Teubner,
Leipzig, 1986.
[78] B. Hofmann and P. Mathé. Analysis of profile functions for general linear regular-
ization methods. SIAM J. of Numer. Anal., 45(3):1122–1141, 2007.
[79] T. Hohage. Regularization of exponentially ill-posed problems. Numer. Funct. Anal.
Optim., 21:439–464, 2000.
[80] M. Hutchinson. A stochastic estimator of the trace of the influence matrix for
Laplacian smoothing splines. Commun. Stat. Simulation Comput., 18(3):1059–1076,
1989.
[81] M.F. Hutchinson and F.R. de Hoog. Smoothing noisy data with spline functions.
Numer. Math., 47:99–106, 1985.
[82] J. Keiner, S. Kunis, and D. Potts. Fast summation of radial functions on the sphere.
Computing, 78:1–15, 2006.
[83] O.D. Kellogg. Foundation of Potential Theory. Springer, Berlin, Heidelberg, New
York, 1967.
[84] S. Kindermann. Convergence analysis of minimization-based noise level-free param-
eter choice rules for linear ill-posed problems. Electron. Trans. Numer. Anal., 38:
233–257, 2011.
[85] S. Kindermann and A. Neubauer. On the convergence of the quasioptimality crite-
rion for (iterated) Tikhonov regularization. Inverse Probl. Imaging, 2(2):291–299,
2008.
[86] R. Kohn, C.F. Ansley, and D. Tharm. The performance of cross-validation and
maximum likelihood estimators of spline smoothing parameters. J. Amer. Statist.
Assoc., 86:1042–1050, 1991.
[87] S.C. Kou and B. Efron. Smoothers and the Cp , generalized maximum likelihood,
and extended exponential criteria: a geometric approach. J. Amer. Statist. Assoc.,
97(459):766–782, 2002.
[88] C.L. Lawson and R.J. Hanson. Solving Least Squares Problems. Prentice-Hall, En-
glewood Cliffs, N.J., 1974.
[89] A. Leonov. Justification of the choice of regularization parameter according to quasi-
optimality and quotient criteria. U.S.S.R. Comput. Math. Math. Phys., 18(6):1–15,
1979.
[90] O. Lepskij. On a problem of adaptive estimation in Gaussian white noise. Theory
Probab. Appl., 35(3):454–466, 1990.
[91] K.-C. Li. Asymptotic optimality of CL and generalized cross-validation in ridge
regression with application to spline smoothing. Ann. Statist., 14:1101–1112, 1986.
[92] K.-C. Li. Asymptotic optimality for Cp , CL , cross-validation and generalized cross-
validation: Discrete index set. Ann. Statist., 15:958–975, 1987.
[93] S. Lu and P. Mathé. Heuristic parameter selection based on functional minimiza-
tion: optimality and model function approach. Math. Comput., 82(283):1609–1630,
2013.
[94] S. Lu and P. Mathé. Discrepancy based model selection in statistical inverse prob-
lems. J. Complexity, 30(3):290–308, 2014.
[95] M.A. Lukas. Convergence rates for regularized solutions. Math. Comput., 51(183):
107–131, 1988.
[96] M.A. Lukas. Asymptotic optimality of generalized cross-validation for choosing the
regularization parameter. Numer. Math., 66(1):41–66, 1993.
[97] M.A. Lukas. On the discrepancy principle and generalized maximum likelihood for
regularization. Bull. Aust. Math. Soc., 52(3):399–424, 1995.
[98] M.A. Lukas. Comparisons of parameter choice methods for regularization with dis-
crete noisy data. Inverse Prob., 14(1):161–184, 1998.
[99] M.A. Lukas. Robust generalized cross-validation for choosing the regularization
parameter. Inverse Prob., 22(5):1883–1902, 2006.
[100] M.A. Lukas. Strong robust generalized cross-validation for choosing the regulariza-
tion parameter. Inverse Prob., 24:034006, 16 pp., 2008.
[101] M.A. Lukas. Robust GCV choice of the regularization parameter for correlated
data. J. Integral Equations Appl., 22(3):519–547, 2010.
[102] W. Magnus, F. Oberhettinger, and R.P. Soni. Formulas and Theorems for the Spe-
cial Functions of Mathematical Physics, volume 52 of Die Grundlehren der mathe-
matischen Wissenschaften. Springer, New York, 3rd edition, 1966.
[103] B.A. Mair and F.H. Ruymgaart. Statistical inverse estimation in Hilbert scales.
SIAM J. Appl. Math., 56(5):1424–1444, 1996.
[104] P. Mathé. What do we learn from the discrepancy principle? Z. Anal. Anwend., 25
(4):411–420, 2006.
[105] P. Mathé and S.V. Pereverzev. Geometry of linear ill-posed problems in variable
Hilbert spaces. Inverse Prob., 19(3):789–803, 2003.
[106] P. Mathé and S.V. Pereverzev. Regularization of some linear ill-posed problems
with discretized random noisy data. Math. Comput., 75(256):1913–1929, 2006.
638 M. Gutting
[107] V. Michel. Lectures on Constructive Approximation – Fourier, Spline, and Wavelet

Methods on the Real Line, the Sphere, and the Ball. Birkhäuser, Boston, 2013.
[108] V. Michel. RFMP – an iterative best basis algorithm for inverse problems in the
geosciences. In W. Freeden, M.Z. Nashed, and T. Sonar, editors, Handbook of Ge-
omathematics, 2nd ed., pages 2121–2147. Springer, Heidelberg, 2015.
[109] H. Moritz. Classical physical geodesy. In W. Freeden, M.Z. Nashed, and T. Sonar,
editors, Handbook of Geomathematics. 1st ed., pages 127–158. Springer, Heidelberg,
2010.
[110] V.A. Morozov. On the solution of functional equations by the method of regular-
ization. Soviet Math. Dokl., 7:414–417, 1966.
[111] V.A. Morozov. Methods for Solving Incorrectly Posed Problems. Springer-Verlag,
New York, 1984.
[112] M.T. Nair, E. Schock, and U. Tautenhahn. Morozov’s discrepancy principle under
general source conditions. Z. Anal. Anwend., 22:199–214, 2003.
[113] A. Neubauer. The convergence of a new heuristic parameter selection criterion for
general regularization methods. Inverse Prob., 24(5):055005, 10 pp., 2008.
[114] S. Pereverzev and E. Schock. Morozov’s discrepancy principle for tikhonov regu-
larization of severely ill-posed problems in finite dimensional subspaces. Numer.
Funct. Anal. Optim., 21:901–916, 2000.
[115] D. Phillips. A technique for the numerical solution of certain integral equations of
the first kind. J. Assoc. Comput. Mach., 9:84–97, 1962.
[116] D. Potts and G. Steidl. Fast summation at nonequispaced knots by NFFTs. SIAM
J. Sci. Comput., 24(6):2013–2037, 2003.
[117] T. Raus. On the discrepancy principle for the solution of ill-posed problems. Uch.
Zap. Tartu. Gos. Univ., 672:16–26, 1984.
[118] T. Raus. The principle of the residual in the solution of ill-posed problems with
nonselfadjoint operator. Tartu Riikl. Ül. Toimetised, 715:12–20, 1985.
[119] T. Raus. An a posteriori choice of the regularization parameter in case of approxi-
mately given error bound of data. In Pedas, A., editor, Collocation and Projection
Methods for Integral Equations and Boundary Value Problems, pages 73–87. Tartu
University, Tartu, 1990.
[120] T. Raus. About regularization parameter choice in case of approximately given error
bounds of data. In G. Vainikko, editor, Methods for solution of integral equations
and ill-posed problems, pages 77–89. Tartu University, Tartu, 1992.
[121] T. Raus and U. Hämarik. On the quasioptimal regularization parameter choices for
solving ill-posed problems. J. Inverse Ill-Posed Probl., 15(4):419–439, 2007.
[122] T. Reginska. A regularization parameter in discrete ill-posed problems. SIAM J.
Sci. Comput., 17(3):740–749, 1996.
[123] T. Robinson and R. Moyeed. Making robust the cross-validatory choice of smooth-
ing parameter in spline smoothing regression. Commun. Stat. Theory Methods, 18
(2):523–539, 1989.
[124] V. Rokhlin. Rapid solution of integral equations of classical potential theory. J.
Comput. Phys., 60:187–207, 1985.
[125] L. Shure, R.L. Parker, and G.E. Backus. Harmonic splines for geomagnetic mod-
elling. Phys. Earth Planet. Inter., 28:215–229, 1982.
[126] A.M. Thompson, J.W. Kay, and D.M. Titterington. A cautionary note about cross-
validatory choice. J. Statist. Comput. Simul., 33:199–216, 1989.
[127] A. Tikhonov and V. Arsenin. Solutions of Ill-Posed Problems. Wiley, New York,
1977.
[128] A. Tikhonov and V. Glasko. Use of the regularization method in non-linear prob-
lems. U.S.S.R. Comput. Math. Math. Phys., 5(3):93–107, 1965.
[129] A. Tsybakov. On the best rate of adaptive estimation in some inverse problems. C.
R. Acad. Sci., Paris, Sér. I, Math., 330(9):835–840, 2000.
[130] D.A. Vars̆alovic̆, A.N. Moskalev, and V.K. Chersonskij. Quantum Theory of Angular
Momentum. World Scientific, Singapore, 1988.
[131] R. Vio, P. Ma, W. Zhong, J. Nagy, L. Tenorio, and W. Wamsteker. Estimation
of regularization parameters in multiple-image deblurring. Astron. Astrophys., 423:
1179–1186, 2004.
[132] C.R. Vogel. Computational Methods for Inverse Problems. SIAM, Philadelphia,
2002.
[133] G. Wahba. Practical approximate solutions to linear operator equations when the
data are noisy. SIAM J. Numer. Anal., 14(4):651–667, 1977.
[134] G. Wahba. Spline interpolation and smoothing on the sphere. SIAM J. Sci. Stat.
Comput., 2:5–16. Also errata: SIAM J. Sci. Stat. Comput., 3, 385–386, 1981.
[135] G. Wahba. Spline Models for Observational Data. SIAM, Philadelphia, 1990.
[136] G. Wahba and Y.H. Wang. When is the optimal regularization parameter insensitive
to the choice of the loss function? Comm. Statist. Theory Methods, 19(5):1685–1700,
1990.
[137] C.A. White and M. Head-Gordon. Rotating around the quartic angular momentum
barrier in fast multipole method calculations. J. Chem. Phys., 105(12):5061–5067,
1996.
[138] H. Yamabe. On an extension of the Helly’s theorem. Osaka Math. J., 2(1):15–17,
1950.
[139] N. Yarvin and V. Rokhlin. Generalized Gaussian quadratures and singular value
decomposition of integral equations. SIAM J. Sci. Comput., 20(2):699–718, 1998.
Martin Gutting
University of Siegen
Walter-Flex-Straße 3
D-57068 Siegen, Germany

Inverse Gravimetry as an Ill-Posed Problem

in Mathematical Geodesy
Willi Freeden and M. Zuhair Nashed
Abstract. The gravitational potential of (a part of) the Earth is assumed to

be available, the inverse gravimetry problem is to determine the density con-
trast function inside (the specified part of) the Earth from known potential
values. This paper deals with the characteristic ill-posed features of transfer-
ring input gravitational information in the form of Newtonian volume integral
values to geological output characteristics of the density contrast function.
Some properties of the Newton volume integral are recapitulated. Different
methodologies of the resolution of the inverse gravimetry problem and their
numerical implementations are examined dependent on the data source. Three
cases of input information may be distinguished, namely internal (borehole),
terrestrial (surface), and/or external (spaceborne) gravitational data sets. Sin-
gular integral theory based inversion of the Newtonian integral equation such
as Haar-type solutions are proposed in a multiscale framework to decorre-
late specific geological signal signatures with respect to inherently available
features. Reproducing kernel Hilbert space regularization techniques are stud-
ied (together with their transition to mollified variants) to provide geological
contrast density distributions by “downward continuation” from terrestrial
and/or spaceborne data. Finally, reproducing kernel Hilbert space solutions
are formulated for use of gravimeter data, independent of a specifically chosen
input area, i.e., in whole Euclidean space R3 .
Keywords. Newton potential, Poisson differential equation, density modeling,
inverse gravimetry, multiscale decorrelation, multiscale spline modeling.
1. Introduction
Gravimetry is a central research area of geodesy, geophysics, and geoexploration. It
is a potential field technique which reflects variations in the Earth’s gravitational
field. These variations are caused by density contrasts inside the Earth. Gravimet-
ric surveys are carried out by use of extremely sensitive instruments capable of
measuring tiny variations in the gravitational field. A gravimeter is a type of an
accelerometer. There are essentially two types of gravimeters, namely relative and
absolute gravimeters. Absolute gravimeters measure the local gravity in absolute
units. They work by directly measuring the acceleration of a mass during free fall
in a vacuum. A new measurement technique is provided by atomic gravimeters.
Relative gravimeters compare the value of gravity at one point with another. Most
common relative gravimeters are spring-based. A spring-based relative gravime-
ter is basically a weight on a spring, and by measuring the amount by which the
weight stretches the spring, gravitation becomes available via Hooke’s Law in lin-
earized form. On global scale gravimetric datasets are used in gravity surveys for
establishing the figure of the geoid. Locally micro-gravimeters are in use, e.g., for
geodetic and geophysical research, geothermal exploration, petroleum and mineral
recovery.
In applied mathematics, inverse gravimetry (in its conventional form) may be
reduced to the following situation: The Newtonian potential of a density contrast
function F is defined as an improper integral over a volume G ⊂ R3 , namely

A[F ](x) = G(Δ; |x − y|) F (y) dy = V (x), x ∈ R3 , (1.1)
G
formed by convolution of F with a kernel function G(Δ; ·) having a singularity,

namely the Newtonian kernel
1 1
G(Δ; |x − y|) = , x ∈ R3 \{y}. (1.2)
4π |x − y|
which apart of a minus sign represents the fundamental solution of the Laplace
equation, i.e.,
Δx G(Δ; |x − y|) = 0, x = y
(or in distributional jargon, −Δx G(Δ; |x − y|) = δ(|x − y|), where δ(·) is the Dirac
distribution). In the region G the volume integral constitutes an operator A that
acts as the inverse to the negative Laplace operator,

−Δx V (x) = −Δx A[F ](x) = − Δx G(Δ; |x − y|) F (y) dy = F (x), x ∈ G,
G
(1.3)
which is to say (at least if the function F is Hölder continuous) that the operation
of taking the Newtonian potential of a function is an inverse operation to the
application of the negative Laplace operator.
Note that the integral (1.1) is named for I. Newton (1642–1720), who first
discovered it and later pioneered the work of P.-S. Laplace (1749–1829) about har-
monic functions. Indeed, the setting (1.1) serves as the fundamental gravitational
potential in Newton’s Law of Gravitation (1687).
Inverse Gravimetry as an Ill-Posed Problem in Mathematical Geodesy 643
Seen from potential theoretic perspective (see, e.g., [29]), Equation (1.1) is
in close relation to the Third Green Theorem

α(x) P (x) = − G(Δ; |x − y|) ΔP (y) dy (1.4)
G

∂ ∂
+ G(Δ; |x − y|) P (y) − P (y) G(Δ; |x − y|) dω(y),
∂G ∂ν(y) ∂ν(y)
that holds true for all twice continuously differentiable functions P on G, where
α(x) is the solid angle subtended by the surface ∂G at the point x ∈ R3 . It should
be mentioned that ⎧
⎪ 1, x ∈ G,
⎪
⎨
1
α(x) = , x ∈ ∂G,
⎪
⎪ 2
⎩
0, x ∈ G c ,
if the boundary surface ∂G is (locally) continuously differentiable. As an immediate
consequence we may expect that the discussion of A[F ](x), x ∈ R3 , actually has
to be split into three cases, dependent on the location of x ∈ R3 as a point of
the inner space G, outer space G c , or the boundary ∂G, i.e., the internal, surface
(terrestrial), and external (spaceborne) input data V (x). Moreover, a solution of
(1.1) in G is not unique, since the addition of any harmonic function to V will not
affect the equation (1.3). In potential theory this observation for the inner space G
can be used to prove existence and uniqueness of solutions to the Dirichlet problem
for the Poisson equation in suitably regular domains and for suitably well-behaved
functions: One first applies a Newtonian potential to obtain a solution, and then
adjusts by adding a harmonic function to get the correct boundary data.
Our intent in this paper is different from solving a boundary-value problem:
We are interested in the inverse gravimetry problem, i.e., the extraction of infor-
mation from the gravitational potential V known for certain locations to determine
the density contrast F inside G. In the language of functional analysis, we have to
solve a Fredholm integral equation of the first kind (1.1) that canonically leads to
the framework of the theory of ill-posed problems (as presented by the authors in
another chapter of this handbook). The main difficulty, however, is that the input
data of the inverse problem are not canonically given in the inner space G, but
usually in G c . As a matter of fact, until now in physical geodesy, only measure-
ments are taken on the surface ∂G (terrestrial measurements) and/or in the outer
space G (spaceborne measurements), i.e., in the set G c . Only in exceptional cases,
e.g., in the neighborhood of “boreholes” of geothermal projects, the gravitational
potential V and the target function F are given inside G, so that the use of the
Poisson differential equation (1.3) becomes applicable in the inversion process.
Typically, for inverse problems, there will also be certain physical constraints
which will be necessary to impose on the potential pattern so that the wanted
geological pattern of the density distribution can be approximated in some ac-
ceptable manner. Such constraints are usually referred to as conditions reflecting
realizability conditions. They will be represented in our mathematical framework

by requiring the density functions to lie in some appropriate subset of the output
space. Under certain conditions these realizability constraints will serve to regular-
ize the originally ill-posed problem, while in others, they will dictate compromises
that will have to be made between requirements for accuracy of the approximating
functions and the demands of meeting such a priori constraints. In this contribu-
tion we are essentially interested in regularization procedures based on mollifier
techniques. Different types of mollifiers will be studied in more detail, namely
Haar-type and singular integral-type mollifiers.
Not only in physical geodesy, but also in inverse problem theory, there is a
huge literature about the character and role of gravimetry in the framework of
ill-posed and inverse problems, from which we only mention a limited selection:
[3, 9–11, 29, 34, 59, 71–75, 90, 93, 99, 100, 107, 114] (for further details the reader
is referred to the references therein). Our paper, however, follows a different ap-
proach, whose point of departure is the introductory chapter [27] of the “Handbook
of Geomathematics”.
2. Newton volume integral

In order to handle the inverse gravimetry problem some potential-theoretic prelim-
inaries are needed: Let G be a regular region in R3 , i.e., a bounded region G dividing
R3 uniquely into the inner space G and the outer space G c = R3 \G, G = G ∪ ∂G,
such that the boundary ∂G is an orientable Lipschitzian manifold of dimension 2
(for example, ball, ellipsoid, cube and other polyhedra, spheroid, telluroid, geoid,
(actual) Earth or appropriate parts of it).
A real-valued function P is called harmonic in G ⊂ R3 if P is of class
C (2) (G) of functions with continuous second-order partial derivatives and satis-
fies the Laplace equation
2 2 2
∂ ∂ ∂
ΔP (x) = + + P (x1 , x2 , x3 ) = 0, x = (x1 , x2 , x3 )T ,
∂x1 ∂x2 ∂x3
(2.1)
for all x ∈ G.
2.1. Basics of potential theory

Some important examples of harmonic functions are given below in the classical
nomenclature of potential theory (see [29]).
(a) Potential of a mass point: According to Newton’s Law of Gravitation two
points x, y with masses Mx , My attract each other with a force given by
γ Mx My
− (x − y), x, y ∈ R3 , x = y. (2.2)
4π |x − y|3
The force is directed along the line connecting the two points x, y . The con-
stant γ denotes Newton’s gravitational constant (note that γ can be assumed
to be equal to one in the theoretical part, but not in numerical applications).
Although the masses Mx , My attract each other in symmetric way, it
is convenient to call one of them the attracting mass and the other one the
attracted mass. Conventionally the attracted mass is set equal to unity and
the attracting mass is denoted by M :
γ M
v(x) = − (x − y), x ∈ R3 \{y}. (2.3)
4π |x − y|3
The formula (2.3) expresses the force exerted by the mass M on a unit mass
located at the distance |x − y| from M . Obviously, the intensity |v(x)| of the
force v(x) is given by
γ M
|v(x)| = , x ∈ R3 \{y}. (2.4)
4π |x − y|2
The scalar function V defined by
1 1
V (x) = γ M G(Δ; |x − y|) = γ M , x ∈ R3 \{y} (2.5)
4π |x − y|
is called the potential of gravitation at y. The force vector v(x) is the gradient
vector of the scalar V (x):
v(x) = ∇V (x), x ∈ R3 \{y}. (2.6)
Calculating the divergence ∇· of the gradient field v, it readily follows that
∇ · v(x) = ∇ · ∇ V (x) = ΔV (x) = 0, x ∈ R3 \{y}. (2.7)
(b) Potential of a finite mass point system: The potential for N points xi with
masses Mi , i = 1, . . . , N , is the sum of the individual contributions

N
V (x) = γ Mi G(Δ; |x − yi |), x ∈ R3 \{y1 , . . . , yn }. (2.8)
i=1
Clearly we have
ΔV (x) = 0, x ∈ R3 \{y1 , . . . , yN }. (2.9)
(c) Potential of a volume: Let G ⊂ R3 be a regular region. The point masses are
distributed continuously over G ⊂ R3 with density F . Then the discrete sum
(2.8) becomes a continuous sum, i.e., an integral over the body G:

V (x) = γ G(Δ; |x − y|)F (y) dy. (2.10)
G
Obviously,
ΔV (x) = 0, x ∈ R3 \G. (2.11)
Note that V is defined on the whole space R3 , however, ΔV (x) may not be
obtained easily by interchanging the Laplace operator and the integral over
G for all points x inside G. At infinity the potential behaves like

1
|V (x)| = O , |x| → ∞, (2.12)
|x|
|x|
uniformly with respect to all directions (note that |y| ≤ 2 implies |x − y| ≥
||x| − |y|| ≥ 12 |x|), i.e., V is regular at infinity.
2.2. Properties of the Newton integral

The Newton (volume) integral (2.10) over a regular region G corresponding to
a mass density distribution F satisfies the Laplace equation in the outer space
G c = R3 \G. Clearly, this property is an immediate consequence of the harmonicity
of the fundamental solution for the Laplace operator (in what follows we restrict
ourselves to a Newton integral (2.10) with γ chosen equal to 1).
Harmonicity in G c . Let F : G → R be an integrable bounded function. Then

V (x) = G(Δ; |x − y|)F (y) dy, x ∈ G c (2.13)
G
satisfies
Δx G(Δ; |x − y|)F (y) dy = 0 (2.14)
G
for all x ∈ G c , i.e., V is harmonic in G c .
Properties in G. By one-dimensional Taylor linearization (cf. [27, 38]) we obtain
1 1 1 1 3 1
√ = √ − (u − u0 ) + (u − u0 )2 (2.15)
u u0 2 u 23 8 (u0 + θ(u − u0 )) 52
0
for some θ ∈ (0, 1). Setting u = r2 and u0 = ρ2 we therefore find

1 1 r2 3 1
= 3− 2 + (r2 − ρ2 )2 . (2.16)
r 2ρ ρ 8 (ρ + θ(r2 − ρ2 )) 52
2
In other words, by letting r = |x − y| we are able to give a simple example for the
“mollification” of the fundamental solution of the Laplace equation
1
G(Δ; r) = , r > 0, (2.17)
4πr
by
⎧
⎪
⎪ 1 1 2
⎨ 3 − 2r , r≤ρ
8πρ ρ
GH
ρ (Δ; r) = (2.18)
⎪
⎪
⎩ 1 , r > ρ.
4πr
such that GHρ (Δ; ·) is continuously differentiable for all r ≥ 0. Obviously, G(Δ; r) =
H
Gρ (Δ; r) for all r > ρ. As a consequence,
1 1
G(Δ; |x − y|) = , |x − y| = 0, (2.19)
4π |x − y|
admits a “mollification” of the form
⎧
⎪ 1 1 2
⎪
⎨ 3 − |x − y| , |x − y| ≤ ρ
8πρ ρ2
GH (Δ; |x − y|) = (2.20)
ρ
⎪
⎪ 1
⎩ , ρ < |x − y|.
4π|x − y|
Let F : G → R be of class C(0) (G). We set

V (x) = G(Δ; |x − y|)F (y) dy (2.21)
G
and
VρH (x) = ρ (Δ; |x − y|)F (y) dy.
GH (2.22)
G
The integrands of V and V ρ differ only in the ball Bρ (x) = {y ∈ R3 : |x − y| < ρ}

around the point x with radius ρ. Because of its continuity the function F : G → R
is uniformly bounded on G. This fact shows that

V (x) − Vρ (x) = O
H
|G(Δ; |x − y|) − Gρ (Δ; |x − y|)| dy
H
Bρ (x)
2
= O(ρ ), ρ → 0. (2.23)
(0)
Therefore, V is of class C (G) as the limit of a uniformly convergent sequence of
continuous functions on G. Furthermore, we let

v(x) = ∇x G(Δ; |x − y|)F (y) dy (2.24)
G
and
vρH (x) = ∇x GH
ρ (Δ; |x − y|)F (y) dy. (2.25)
G
Because of the fact
|∇x G(Δ; |x − y|)| = O(|x − y|−2 ), (2.26)
the integrals v and vρHexist for all x ∈ G. It is not hard to see that

sup v(x) − vρH (x) = sup |∇x V (x) − ∇x VρH (x)| = O(ρ), ρ → 0. (2.27)
x∈G x∈G
Consequently, v is a continuous vector field on G. Moreover, as the relation (2.27)

holds uniformly on G, we obtain

v(x) = ∇V (x) = ∇x G(Δ; |x − y|)F (y) dy. (2.28)
G
Altogether, we are allowed to formulate the following properties:

Let G be a regular region. Let F : G → R be of class C (0) (G). Then V, VρH as
defined by (2.21), (2.22), respectively, are of class C (1) (G), such that
lim sup |V (x) − VρH (x)| = 0. (2.29)
ρ→0
x∈G
Furthermore, ∇V is of class C (0) (G), such that

∇x V (x) = F (y) ∇x G(Δ; |x − y|) dy, x ∈ G. (2.30)
G
and
lim sup |∇x V (x) − ∇x VρH (x)| = 0. (2.31)
ρ→0
x∈G
2.3. Poisson’s differential equation

We come to the Poisson differential equation under the assumption of μ-Hölder
continuity, μ ∈ (0, 1], imposed on the function F on G. For that purpose we note
3
3 −5
that the Taylor linearization of s− 2 is given by s02 − 32 s0 2 (s−s0 ). Hence, by letting
s = r2 and s0 = ρ2 , we are able to replace r−3 by 2ρ13 (5 − ρ32 r2 ). As a consequence,
as “mollification” of
1
Z (Δ; |x − y|) = , |x − y| = 0, (2.32)
4π|x − y|3
we are able to introduce
⎧
⎪ 1 3 2
⎪
⎨ 5 − |x − y| , |x − y| ≤ ρ
8πρ3 ρ2
Zρ (Δ; |x − y|) = (2.33)
⎪
⎪ 1
⎩ , ρ < |x − y|.
4π|x − y|3
The function r → Zρ (Δ; r), r ≥ 0, is continuously differentiable. Moreover, by the
same arguments as above, it can be shown that the vector field

zρ (x) = − Zρ (Δ; |x − y|)(x − y)F (y) dy (2.34)
G
converges uniformly on G with ρ → 0 to the limit field

v(x) = ∇Vx (x) = ∇x G(Δ; |x − y|)F (y) dy. (2.35)
G
For all x ∈ Bρ (x) we obtain by a simple calculation

15 1 |x − y|2
∇x · ( Zρ (Δ; |x − y|)(x − y)) = − . (2.36)
8π ρ3 ρ5
Furthermore, an easy calculation shows that

∇x · (Zρ (Δ; |x − y|) (x − y)) dy = 1. (2.37)
Bρ (x)
Hence, under the additional assumption of μ-Hölder continuity, μ ∈ (0, 1], for the
function F on G, i.e., |F (x) − F (y)| ≤ C |x − y|μ for all x, y ∈ G, we obtain

∇x · zρ (x) = − ∇x · Zρ (Δ; |x − y|)(x − y)F (y) dy (2.38)
G

=− ∇x · (Zρ (Δ; |x − y|)(x − y)) F (y) dy
Bρ (x)

= − α(x) F (x) + (F (x) − F (y)) ∇x · (Zρ (Δ|x − y|)(x − y)) dy.
Bρ (x)
Thus, the μ-Hölder continuity of F guarantees the estimate

sup |∇x · zρ (x) + α(x) F (x)| = O(ρμ ), ρ → 0, (2.39)
x∈G
uniformly with respect to x ∈ G, where α(x) is the solid angle at x subtended

by the surface ∂G. In an analogous way, we are able to show that the first par-
tial derivatives of (2.34) converge uniformly to continuous limit fields. Again, the
uniform convergence shows that ∇V is differentiable in G, and we have

∇x ·v(x) = ΔV (x) = Δx G(Δ; |x−y|)F (y) dy = −α(x) F (x), x ∈ G. (2.40)
G
It should be noted that the assumption of μ-Hölder continuity of F , μ ∈

(0, 1], is needed for the proof of (2.40). Indeed, Petrini [91] showed that the μ-
Hölder continuity of F , μ ∈ (0, 1], is necessary to imply the second continuous
differentiability of the Newton volume potential.
Let G be a regular region. If F is of class C (0,μ) (G), μ ∈ (0, 1], then the
Poisson differential equation

−Δx F (y) G(Δ; |x − y|) dV (y) = α(x) F (x) (2.41)
G
holds true for all x ∈ G, where α(x) is the solid angle subtended by the surface ∂G
at x.
3. Ill-posedness of the gravimetry problem

Contrary to the case of L2 (∂G) (see [23] for its verification), the class L2 (G) of
square-integrable functions on a regular region G is not obtainable only by the L2 -
completion of a countable harmonic function system. In addition, we have to take
into account a so-called “anharmonic function system” (see, e.g., [9, 34, 71, 109]).
This observation should be studied here in a more detailed way, since it explains
the ill-posedness of the gravimetry problem.
Let G ⊂ R3 be a regular region. By E(G) we denote the space of all infinitely
differentiable functions F in R3 possessing a compact support supp(F ) in G. We
equip E(G) with the following topology: a sequence {φn } ⊂ E(G) is called con-
vergent to zero if and only if (1) there exists a bounded B ⊂ R3 such that φn
vanishes outside B, (2) for every differential operator ∇α the sequence {∇α φn } is
convergent to zero with respect to the norm · C (0) (G) . Members of E(G) are called
test functions. Elements of the dual space E ∗ (G), i.e., continuous linear functionals
F : E(G) → R, are called distributions (or generalized functions). Clearly, multi-
plication (by a scalar) and addition are defined canonically for members of the
class E ∗ (G), hence, they are in use for distributions in the same way, too. More
details can be found in any textbook on distributions, e.g., [58].
Let F ∈ E ∗ (G) be a given distribution. Assume that there exists a function
F : G → R that is locally integrable, i.e., F is integrable on every compact subset
of G, such that F (φ) = G F (x)φ(x) dV (x) holds for all test functions φ ∈ E(G).
Then F is called a regular distribution. If F ∈ E ∗ (G) is a regular distribution, then
the associated function F is uniquely determined (except on a set of Lebesgue
measure zero). Note that a well-known distribution that is not regular is the delta
distribution δ given by δ(φ) = φ(0).
A sequence {Fn } ⊂ E ∗ (G) is called convergent to F ∈ E ∗ (G) if and only
if limn→∞ Fn (φ) = F (φ) for all φ ∈ E(G). This definition helps to introduce
derivatives of distributions: If, for a given distribution F ∈ E ∗ (G), there exists
a distribution F̃ ∈ E ∗ (G) such that F̃(φ) = (−1)[α] F (∇α φ), α ∈ N30 , [α] =
α1 + α2 + α3 , for every φ ∈ E(G), then we set F̃ = ∇α F . In our potential theoretic
approach we are particularly interested in Laplace derivatives: A functional F ∈
E ∗ (G) is called distributionally harmonic if and only if ΔF = 0. The set of all
regular harmonic L2 (G)-distributions in E ∗ (G) is denoted by DistHarm(G). The
space DistHarm(G) apparently represents a generalization of the set Harm(G) of
harmonic functions in G. Indeed, the following characterization is valid (see, e.g.,
[16]): The set DistHarm(G) of all regular harmonic L2 (G)-distributions is a closed
subspace of L2 (G).
It is known from the theory of distributions that the set

(2) 2
Harm(G) = {H ∈ C (G) : (H(x)) dx < ∞, ΔH(x) = 0, x ∈ G} (3.1)
G
is a subset of DistHarm(G). Moreover, the so-called Weyl Lemma (see, e.g., [58])
tells us that G F (x)ΔΦ(x) dx = 0 for all Φ ∈ E(G) implies F ∈ Harm(G),
i.e., DistHarm(G) ⊂ Harm(G). As a consequence, we are led to the following
remarkable result:
Let G ⊂ R3 be a regular region. Then,
DistHarm(G) = Harm(G). (3.2)
Harm(G) is a closed linear subspace of L2 (G). Thus, a well-known result of
functional analysis on orthogonal decompositions (see, e.g., [16, 112]) enables us
to formulate the decomposition of L2 (G) in the form:
· 2
· 2
⊥
L2 (G) = Harm(G) L (G) ⊕ Harm(G) L (G) . (3.3)
The following convention is in use (see [109], and also [9] and [71]):
· 2
⊥
AnHarm(G) = Harm(G) L (G) . (3.4)
The members of AnHarm(G) are called anharmonic functions in G.
Summarizing our results, we finally arrive at the following decomposition
theorem:
If G ⊂ R3 is a regular region, then
L2 (G) = Harm(G) ⊕ AnHarm(G). (3.5)
3.1. Hadamard’s classification of the gravimetry problem
In classical nomenclature of physical geodesy, the inversion of “Newton’s Law
of Gravitation” (1.1) from terrestrial and spaceborne gravitational data, i.e., the
determination of the internal density contrast function from potential data on and
outside the boundary ∂G is known as the gravimetry problem. In other words, for a
regular region G ⊂ R3 , we are interested in the problem of determining the density
function F ∈ L2 (G) from (information of) the gravitational potential V on G c in
accordance with the integral equation

V (x) = A[F ](x) = F (y) G(Δ; |x − y|) dy, x ∈ R3 . (3.6)
G
In the sequel we denote the image of X = L2 (G) under the operator A by Y, i.e.,

Y := A[L2 (G)] = {V : V = A[F ] = G(Δ; | · −y|)F (y) dy, F ∈ L2 (G)}. (3.7)
G
3
Furthermore, for any subset H ⊂ R , we introduce the operator
AH : X = L2 (G) → Y |H. (3.8)
(more accurately, AGH ) with Y |H consisting of all AH [F ] given by

H x → AH [F ](x) = G(Δ; |x − y|) F (y) dy, F ∈ L2 (G) (3.9)
G
(note that A may be formally understood as AR3 ). Clearly, Y |H forms a set of

harmonic functions in H, provided that H is a subset of G c .
In shorthand notation, the (terrestrial/spaceborne gravimetry problem)
(TSGP) of classical physical geodesy can be formulated as follows:
(TSGP): Given V ∈ L2 (G c ), find F ∈ L2 (G) with AG c [F ] = V.
In accordance with Hadamard’s classification, TSGP violates all criteria, viz.
uniqueness, existence, and stability:
(i) A solution of the gravimetry problem exists only if V belongs to the space
Y |G c . However, it should be pointed out that this restriction does not cause
any numerical difficulty since, in practice, the information of V is only finite-
dimensional.
(ii) The most serious problem is the non-uniqueness of the solution: The asso-
ciated Fredholm integral operator AG c has a kernel (null space) which is
already known to coincide with the L2 (G)-orthogonal space of the closed lin-
ear subspace of all harmonic functions on G. Unfortunately, the orthogonal
complement, i.e., the class of anharmonic functions, is infinite-dimensional.
More precisely, if F is a member of class L2 (G), then AG c : L2 (G) →
Y |G given by
c

V = AG c [F ] = G(Δ; | · −y|) F (y) dy , F ∈ L2 (G), (3.10)
G G c
defines a linear operator such that AG c [F ] is harmonic in G c and regular at

infinity. The operator AG c as defined by (3.10) is surjective, but it is not
injective. Indeed, the null space (kernel) of AG c
N (AG c ) = AnHarm(G) (3.11)
consists of all functions in L2 (G) that are orthogonal to harmonic functions
in G. N (AG c ) is the space of anharmonic functions in G: Let G ⊂ R3 be a
regular region. Then we have
· 2
· 2
⊥
L2 (G) = Harm(G) L (G) ⊕ Harm(G) L (G) , (3.12)
hence,
L2 (G) = Harm(G) ⊕ AnHarm(G) = Harm(G) ⊕ N (AG c ). (3.13)
(iii) Restricting the operator AG c to Harm(G) leads to an injective mapping which
has a discontinuous inverse.
Concerning the historical background, the problem of non-uniqueness has

been discussed extensively in literature. This problem can be resolved by impos-
ing some reasonable additional condition on the density. As we already saw, a
suitable condition, suggested by the mathematical structure of the Newton poten-
tial operator A is to require that the density be harmonic. In fact, the approximate
calculation of the harmonic density has already been implemented in several pa-
pers (see, e.g., [79] and the references therein), whereas the problem of determining
the anharmonic part seems to be still a great challenge.
3.2. Spectral inversion procedure for balls

The set Harm(Bβ (0)) of harmonic functions in the ball Bβ (0) with radius β around
the origin 0 is a closed subspace of L2 (Bβ (0)) (note that β can be chosen, for exam-
ple, to be the radius of a Runge (Bjerhammar) sphere (see, e.g., [29]) or the (mean)
Earth’s radius). Moreover, the inner harmonics {Hn,j (β; ·)}n=0,1,...;j=1,...,2n+1
given by n
2n + 3 |x| x
Hn,j (β; x) = Yn,j , x ∈ Bβ (0), (3.14)
β3 β |x|
constitute a complete orthonormal system in the Hilbert space

Harm(BBβ (0) ), ·, ·L2 (Bβ (0))
provided that {Yn,j }n=0,1,...;j=1,...,2n+1 is a complete system of spherical harmonics
on the unit sphere (see, e.g., [38]).
The set of square-integrable harmonic functions on the outer space
Bβc (0) = R3 \Bβ (0)
of a sphere ∂Bβ (0) given by Harm(Bβc (0)) is a closed subspace of L2 (Bβc (0)). More-
ext
over, the outer harmonics {H−n−1,j (β; ·)}n=1,2,...;j=1,...,2n+1 given by
n+1
2n − 1 β x
H−n−1,j (β; x) = 3
Y n,j , x ∈ Bβc (0) (3.15)
β |x| |x|
form a complete orthonormal system in the Hilbert space

Harm(Bβc (0)), ·, ·L2 (Bβc (0)) .
It should be remarked that an outer harmonic of degree n = 0 is proportional to

1 x 1
Y0,1 =√ , x ∈ Bβc (0). (3.16)
|x| |x| 4π|x|
This function, however, is not an element of L2 (Bβc (0)).
Harmonic Case. The operator ABc (0) , given by

β

ABc (0) [F ](x) = G(Δ; |x − y|) F (y) dy, x ∈ Bβc (0), (3.17)
β
Bβ (0)
has the null space

N ABc (0) = AnHarm(Bβ (0)). (3.18)
β
For any F ∈ L2 (Bβ (0)), there exists a unique orthogonal decomposition

F = PHarm(Bβ (0)) [F ] + PAnHarm (Bβ (0)) [F ], (3.19)
where PHarm(Bβ (0)) [F ] ∈ Harm(Bβ (0)) and PAnHarm(Bβ (0)) [F ] ∈ AnHarm(Bβ (0)).
We are allowed to represent PHarm(Bβ (0)) [F ] as a Fourier series in terms of inner
harmonics
∞ 2n+1

PHarm(Bβ (0)) [F ] = F, Hn,j (β; ·)L2 (Bβ (0)) Hn,j (β; ·) (3.20)
n=0 j=1
with respect to the topology of L2 (Bβ (0)). Suppose that y ∈ Bβc (0) is arbitrary
but fixed. Then the potential at y corresponding to the mass density distribution
F can be represented in the form

ABc (0) [F ](y)
β
∞ 2n+1
4π 1
= β2 # Hn,j β; ·) , F L2 (Bβ (0)) H−n−1,j (β; y) .
n=0 j=1
2n + 1 (2n − 1)(2n + 3)
(3.21)
A harmonic solution F ∈ Harm(BBβ (0) ) of the problem
ABc (0) [F ] = P, P ∈ Harm(Bβc (0)) (3.22)
β
is unique and is given via its Fourier coefficients

2n + 1 #
F, Hn,j (β; ·)L2 (Bβ (0)) = (2n − 1)(2n + 3) P, H−n−1,j (β; ·)L2 (Bc (0)) ,
4πβ 2 β
(3.23)
n ∈ N, j ∈ {1, . . . , 2n + 1}, and
F, H0,1 (β; ·)L2 (Bβ (0)) = 0, (3.24)
∞ 2n+1
2n + 1 #
F = (2n − 1)(2n + 3) P, H−n−1,j (β; ·)L2 (Bc (0)) Hn,j (β; ·)
n=1 j=1
4πβ 2 β
(3.25)
in the sense of L2 (Bβ (0)).
In accordance with the Picard condition (that appeared in book form in the
chapter [84]) the equation ABc (0) [F ] = P is solvable if and only if P is harmonic
β
and the following series is finite, i.e.,
∞ 2n+1
2
n4 P, H−n−1,j (β; ·)L2 (Bc (0)) < ∞. (3.26)
β
n=1 j=1
Note that
∞ 2n+1
2
(2n + 1)2 (2n − 1)(2n + 3) P, H−n−1,j (β; ·)L2 (Bc (0)) < ∞
β
n=1 j=1
(3.27)
∞ 2n+1
2
⇔ n4 P, H−n−1,j (β; ·)L2 (Bc (0)) < ∞.
β
n=1 j=1
This condition can be also motivated within the framework of harmonics by ob-
serving
·L2 (B
F ∈ Harm(BBβ (0) ) = {Hn,j (β; ·) : n ∈ N0 ; j ∈ {1, . . . , 2n + 1}} β (0))
, (3.28)
that implies
∞ 2n+1
2
F, Hn,j (β; ·)L2 (Bβ (0)) < ∞ . (3.29)
n=0 j=1
Operators of the type ABc (0) are compact, hence, we are confronted with the fact
β
that the restricted operator

A c Harm(BB (0) ) : Harm(BB (0) ) → A c
Bβ (0) β β
Harm(BB (0) )
Bβ (0) β
−1
is invertible, but its inverse operator ABc (0) Harm(B is discontinuous.
β Bβ (0) )
Anharmonic Case. An orthogonal basis for AnHarm(Bβ (0)) (with respect to the
space L2 (Bβ (0))) can be found in [9]. A different non-orthogonal anharmonic basis
has been developed in [71, 72]:
(a) A complete L2 (Bβ (0))-orthogonal system in AnHarm(Bβ (0)) is given by

2 x
x → |x| Pk,n (|x| ) Yn,j
n
, (3.30)
|x| k∈N; n∈N0
j∈{1,...,2n+1}
where {Pk,n }k∈N;n∈N0 is a system of polynomials defined by

2 3 3 t
Pk,n (t) = Gk n + , n + ; 2 . (3.31)
β 2n+3 2 2 β
Here, the functions Gk , k ∈ N0 , are the Jacobi polynomials, which are the
only polynomials on [0, 1] to satisfy the following conditions for all n, m ∈ N0 :
(i) Gn (a, b; ·) is a polynomial of degree n on [0, 1].
(ii) Gn (a, b; 0) = 1.
1
(iii) 0 xa−1 (1 − x)b−a Gn (a, b; x) Gm (a, b; x) dx = 0 for n = m,
provided that a > 0 and b > a − 1.
(b) A closed system in Anharm(Σint ) is given by

(2n + 3)β 2k x
x → |x| n+2k
− |x| Yn,j
n
. (3.32)
2n + 2k + 3 |x| k∈N; n∈N0
j∈{1,...,2n+1}
Moreover, the basis functions are polynomials of degree ≤ N ∈ N\{1} if and

only if the index triple! (k, n, 0 j) is1"within the range n ∈ {0, . . . , N − 2}, j ∈
{1, . . . , 2n + 1}, k ∈ 1, . . . , N −n
2 , where [ · ] is the Gauss bracket, defined
by [x] = max{ν ∈ Z : ν ≤ x}, x ∈ R. The set of anharmonic polynomials
with degrees ≤ N possesses the dimension 16 N 3 − 16 N .
The obvious advantage of the system in (a) is its orthogonality. On the other
hand, the system described in (b) has a radial part (see also [34]), which is explic-
itly given, whereas the radial part of the orthogonal system has to be calculated
iteratively by means of recurrence formulas.
The important role of the anharmonic functions in the theory of the gravime-
try problem is also stressed if we investigate a radially symmetric density distribu-
tion which is given for the mantle and the outer and inner core of the Earth. Such
a structure of spherical layers does not give any information in the gravitational
potential and, therefore, cannot be recovered by means of harmonic functions. The
Ph.D.-thesis of V. Michel [71], indeed, shows that a reconstruction of the (deep)
Earth’s interior with a harmonic function system makes no sense. Therefore, a

reliable method for the (global) approximation of the density distribution of the
Earth requires a treatment of both orthogonal projections: the harmonic part and
the anharmonic part.
Moreover, we recall that the contribution of H−1,1 to an (outer) gravitational
(disturbance) potential can be neglected when applying an appropriate coordinate
transformation (see, e.g., [81] for more details). This operation can, therefore,
physically be interpreted as filtering out the contribution of the radially sym-
metric density structures in the Earth’s interior (note that the total mass of an
anharmonic density function is zero).
3.3. Spectral inversion procedure for regular regions

The above results will now be extended to the investigation of the inverse problem
AG c [F ] = V, where AG c [F ] is the gravitational potential of a regular region G ⊂ R3
and F ∈ L2 (G) is the desired mass density distribution F. As already known from
(3.11), the null space of the operator AG c is given by N AG c = AnHarm(G).
A general complete orthonormal basis system for the harmonic functions
inside or outside an arbitrary regular region is not available. This is the reason
why the following setting is useful: Let the families of functions
{Hn,j (G; ·)} n∈N0 ; and {H−n−1,j (G c ; ·)} n∈N; (3.33)
j=1,...,2n+1 j=1,...,2n+1
be complete orthonormal systems of the Hilbert spaces

Harm(G), ·, ·L2 (G) and Harm(G c ), ·, ·L2 (G c ) , (3.34)
respectively, and {kG∧ (n)}n∈N0 be the symbol of
AG c : L2 (G) → Y |G c = R(AG c ) = AG c (L2 (G), (3.35)
given by
AG c [F ](x)
∞ 2n+1
(3.36)
= kG∧ (n) F, Hn,j (G; ·)L2 (G) H−n−1,j (G c ; x), x ∈ G c , F ∈ L2 (G),
n=0 j=1
where H−1,1 (G; ·) is not an element of L2 (G c )). We assume that kG∧ (n) = 0 for all
n ∈ N0 .
If ∂G is a sphere with radius β around the origin, we let
Hn,j (G; ·) := Hn,j (β; ·); n ∈ N0 , j ∈ {1, . . . , 2n + 1}; (3.37)
H−n−1,j (G c ; ·) := H−n−1,j (β; ·); n ∈ N, j ∈ {1, . . . , 2n + 1} . (3.38)
Moreover, we set
4π β2
kG∧ (n) = kβ∧ (n) = # . (3.39)
2n + 1 (2n − 1)(2n + 3)
The inverse problem AG c [F ] = V with F ∈ Harm(G) unknown, is solvable if

and only if V ∈ Harm(G) with
? @ 2
∞ 2n+1
V, H−n−1,j (G c ; ·) L2 (G c )
<∞ . (3.40)
n=1 j=1
kG∧ (n)
In this case, the harmonic solution F ∈ Harm(G) is uniquely determined and

spectrally given by
F, H0,1 (G; ·)L2 (G) = 0, (3.41)
V, H−n−1,j (G c ; ·)L2 (G c )
F, Hn,j (G; ·)L2 (G) = , (3.42)
kG∧ (n)
for n ∈ N, j ∈ {1, . . . , 2n + 1}.

As already known, the inverse operator (AG c Harm(G))−1 , defined on the
image Y |G c , is unbounded. Due to unavoidable errors in the measurements of the
gravitational field the application of this inverse operator to the observed poten-
tial for a direct reconstruction of the mass density distribution is not senseful.
Therefore, we have to take into account suitable regularizations. Indeed, the re-
sults as presented here enable us to apply projection-, multiscale-, and iteration
regularization techniques in the way indicated, e.g., in our contribution [35].
4. Mollifier methods
Next we deal with space regularization methods for the Newton volume integral
involving singular integral mollification.
4.1. Haar-type mollifier method

We start from the differential equation
ρ (Δ; |y − z|) = − Hρ (|y − z|)
Δy GH (4.1)
with
1 |y−z|2
8πρ (3 − ρ2 ) , |y − z| ≤ ρ,
ρ (Δ; |y
GH − z|) = 1 (4.2)
4π|y−z| , |y − z| > ρ,
where 3
4πρ3 , |y − z| ≤ ρ,
Hρ (|y − z|) = (4.3)
0 , |y − z| > ρ
is the so-called Haar kernel (note that Bρ (0) = 43 πρ3 ).
It is well known (see, e.g., [29]) that the Haar singular integral {Iρ }ρ>0 de-
fined by
H
Iρ [F ] = Fρ = Hρ (| · −z|) F (z) dz, (4.4)
G
with the Haar kernel as mollifier satisfies the limit relation limρ→0+ Iρ [F ] = F, F ∈
L2 (G), in the topology of L2 (G). Moreover, we have
lim Iρ [F ](x) = α(x) F (x), x ∈ G, F ∈ C (0) (G), (4.5)
ρ→0+
where α(x) is the solid angle at x subtended by the surface ∂G.

In constructive approximation, locally supported functions
y → Hρ (|x − y|), ρ > 0, x ∈ R3 ,
are nothing new, with one-dimensional counterparts having been discussed already
by Haar (1910). The primary importance of locally supported Haar kernels in the
classical one-dimensional Euclidean space is that they led to the “birth” to an
entire “basis family” by means of two operations, viz. dilations and translations.
In other words, an entire set of approximants is available from the single locally
supported “Haar mother kernel”, and this basis family provides useful “building
block functions” that enable the multiscale modeling and the decorrelation of data.
Internal/Terrestrial Gravimetry Problem (ITGP). Correspondingly to {Iρ }ρ>0 we
introduce the family {AH ρ }ρ>0 given by

AH ρ (Δ; | · −z|) F (z) dz,
H
ρ [F ] = Vρ = GH (4.6)
G
such that
ΔAH
ρ [F ] =Δ GHρ (Δ; | · −z|) F (z) dz
G

= −Iρ [F ] = −Fρ = − Hρ (| · −z|) F (z) dz.
H
(4.7)
G
Multiscale mollifier approximation. Next we are interested in applying the mul-

tiscale “Haar philosophy” to an approximate determination of the mass density
distribution inside G (cf. [29]): Suppose that {ρj }j∈N0 is a positive, monotonously
decreasing sequence with limj→∞ ρj = 0, for example, the dyadic sequence given
by ρj = 2−j . For j ∈ N0 , we consider the differences
Ψ GH
ρ
(Δ; | · −y|) = GH
ρj+1 (Δ; | · −y|) − Gρj (Δ; | · −y|)
H
(4.8)
j
and
ΨHρj (| · −y|) = Hρj+1 (| · −y|) − Hρj (| · −y|). (4.9)
Ψ GH
ρj
(Δ; ·) and ΨH ρj are called “ρj -fundamental wavelet function” and “ρj -Haar
wavelet function”, respectively. The associated “ρj -potential wavelet functions”
and the “ρj -density wavelet functions” are given by

H
(W V )ρj = Ψ GH
ρ
(Δ; | · −y|) F (y) dy (4.10)
j
G
and
(W F )H
ρj = ΨHρj (| · −y|) F (y) dy, (4.11)
G
respectively. The ρj -potential wavelet functions and the ρj -density wavelet func-
tions, respectively, characterize the successive detail information contained in
VρHj+1 − VρHj and FρHj+1 − FρHj , j ∈ N0 . In other words, we are able to decorre-
late the potential V and the “density signature”F , respectively, in form of ”band
structures”
ρj = Vρj+1 − Vρj ,
(W V )H H H
(4.12)
and
ρj = Fρj+1 − Fρj .
(W F )H H H
(4.13)
The essential problem to be solved in multiscale extraction of geological features

is to identify those detail information, i.e., band structures in (4.12), which con-
tain specifically desired geological (density) characteristics in (4.13). Seen from
a numerical point of view, it is remarkable that both wavelet functions y →
Ψ GHρj
(Δ; | · −y|) and y → ΨHρj (| · −y|) vanish outside a ball around the center
x due to their construction, i.e., these functions are spacelimited showing a ball
as local support. Furthermore, the ball becomes smaller with increasing scale pa-
rameter j, so that more and more high frequency phenomena can be highlighted
without changing the features outside the balls. Forming the telescoping sums

J−1
J−1
(W V )H
ρj = VρHj+1 − VρHj , (4.14)
j=0 j=0
and

J−1
J−1

(W F )H
ρj = Fρj+1 − Fρj , (4.15)
j=0 j=0
we are easily led to the identities

J−1
VρHJ = VρH0 + (W V )H
ρj (4.16)
j=0
and

J−1
FρHJ (x) = FρH0 + (W F )H
ρj . (4.17)
j=0
Thus we finally end up with the following multiscale relations

∞

lim VρHJ = VρH0 + (W V )H
ρj (4.18)
J→∞
j=0
and
∞
∞

lim FρHJ = FρH0 + (W F )H
ρj = lim ΔVρHJ = ΔVρH0 + Δ(W V )H
ρj . (4.19)
J→∞ J→∞
j=0 j=0
Altogether, the potential V as well as the “density signature” F can be

expressed in additive way as a low-pass filtered signals VρH0 and FρH0 and successive
band-pass filtered signals (W V )H H
ρj and (W F )ρj , j = 0, 1, . . . , respectively.
Multiscale mollifier numerics. For a sufficiently large integer J, it follows from

(4.5) that
α(x) F (x) & IρJ [F ](x) = FρHJ (x)

(4.20)
= HρJ (|x − y|)F (y) dy, x ∈ G, F ∈ C (0) (G)
G
(“&” means that the error is negligible). From (4.1) we obtain

Δx AρJ [F ](x) = Δx GHρJ (Δ; |x − y|)F (y) dV (y)
G

= − HρJ (|x − y|)F (y) dy & − α(x) F (x), (4.21)
G
where we are aware of the fact that

V (x) & GHρJ (Δ; |x − y|)F (y) dy, x ∈ G, (4.22)
G
with negligible error. In order to realize a fully discrete approximation of F we

have to apply approximate integration formulas leading to

NJ
V (x) & ρJ (Δ; |x − yi |) wi F (yi ),
GH NJ NJ NJ
x ∈ G, (4.23)
i=1
where wiNJ ∈ R, yiNJ ∈ G, i = 1, . . . , NJ , are the known weights and knots,

respectively.
For numerical realization of mass density modeling by means of Haar kernels
we notice that all coefficients
aN
i
J
= wiNJ F (yiNJ ), i = 1, . . . , NJ , (4.24)
are unknown. Then we have to solve a linear system, namely

NJ
V (xTk J ) = ρJ (Δ; |xk − yi |) ai , xTk J ∈ G, k = 1, . . . , TJ ,
TJ NJ NJ
GH (4.25)
i=1
to determine aN J TJ
i , i = 1, . . . , NJ , from known gravitational values V (xk ) at knots
xk ∈ G, k = 1, . . . , TJ .
TJ
Once all density values F (yiNJ ), i = 1, . . . , NJ , are available (note that the
integration weights wiNJ , i = 1, . . . , NJ , are known), the density distribution F can
be obtained from the formula

NJ
F (x) & FρHJ (x) = HρJ (|x − yiNJ |) wiNJ F (yiNJ ), x ∈ G. (4.26)
i=1

NJ
=ai
Even better, fully discrete Haar filtered versions of F at lower scales can be derived
in accordance with the approximate integration rules

Nj
N Nj N
Hρj (|x − z|) F (z) dV (z) & Hρj (|x − yi j |) wi F (yi j ) (4.27)
G i=1
N N
for j = J0 , . . . , J, where wi j , yi j , i = 1, . . . , Nj , are known weights and knots,
N N
respectively, such that {y1 j , . . . , yNjj } ⊂ {y1NJ , . . . , yN
NJ
J
} ⊂ G, i.e., the sequence
of knots {y1NJ , . . . , yN
NJ
J
} ⊂ G shows a hierarchical positioning. Altogether, our ap-
proach yields Haar filtered versions (4.27) establishing a (space-based) multiscale
decomposition FρHJ , . . . , FρHJ0 of the density distribution F , such that an entire set
of approximations is available from a single locally supported “mother function”,
i.e., the Haar kernel function (4.3), and this set provides useful “building block
functions”, which enable decorrelation of the density signatures and suitable stor-
age and fast decorrelation of density data.
Moreover, fully discrete Haar filtered versions of F at lower scales can be
derived in accordance with the approximate integration rules

Nj
N N N
FρHj (x) = Hρj (|x−y|) F (y) dy & Hρj (|x−yi j |) wi j F (yi j ), x ∈ G, (4.28)
G i=1
N N
for j = J0 , . . . , J, where wi j , yi j , i = 1, . . . , Nj , are known weights and knots,
N N
respectively, such that we can take advantage of the fact that {y1 j , . . . , yNjj } ⊂
{y1NJ , . . . , yN
NJ
J
} ⊂ G.
The serious problem of our multiscale approach, however, is that measure-
ments of gravitation are only available in the interior G in exceptional cases, for
example, locally in geothermal boreholes. However, we are able to take into ac-
count surface measurements on ∂G, but it may be questioned that deep geological
formations can be detected by an exclusive use of terrestrial gravitational data.
Nevertheless, the multiscale method as explained above is an important postpro-
cessing method to improve the interpretability of already available geological mod-
els as well as (wavelet) decorrelation mechanisms to extract certain local features
of practical relevance in density band signatures (see [15, 78]).
4.2. De la Vallée Poussin-type mollifier method

The critical point in the Haar-type approach is the discontinuity of the Laplace
derivative of GHρ (Δ; ·), i.e., the ordinary Haar function Hρ . In what follows we are
therefore interested in a smoothed Haar kernel variant, called de la Vallée Poussin
kernel.
For x, y ∈ R3 we define the de la Vallée Poussin kernel V Pρ = [0, ∞) → R,
ρ > 0, by 2
1 (1 − ρr2 ) , r ≤ ρ
V Pρ (r) = V P (4.29)
Cρ 0 , r > ρ,
where the normalization constant CρV P = 8π 3
15 ρ is chosen in such a way that
ρ
V Pρ (|x − y|) dy = 4π V Pρ (r) r2 dr = 1. (4.30)
R3 0
It is easy to see that r → − 61 r2 + 20ρ 1 4

2 r , r ≥ 0, satisfies
2
d 2 d
− r12 dr r dr − 16 r2 + 20ρ
1
2r
4
= 1 − ρr2 , r ≥ 0, ρ > 0. (4.31)
As a consequence, it follows that

1
− 61 |x − y|2 + 1
20ρ2 |x − y|4 , |x − y| ≤ ρ
Gρ (Δ; |x − y|) =
VP C VP
ρ (4.32)
0 , |x − y| > ρ
satisfies
−Δx GVρ P (Δ; |x − y|) = V Pρ (|x − y|), x, y ∈ R3 . (4.33)
2
An elementary calculation yields − r12 dr
d 2 d
r dr (1 − r
ρ2 ) = 6
ρ2 , so that
−Δx V Pρ (Δ; |x − y|) = DρV P (|x − y|), x, y ∈ R3 , (4.34)
where
1 6 8π
CρV P ρ2
= 3 ρ , |x − y| ≤ ρ
DρV P (|x − y|) = (4.35)
0 , |x − y| > ρ.
Clearly, all methodological concepts developed for the Haar case together
with its multiscale settings remain valid. Their formulations are straightforward.
The following result, however, serves as strategic basis for our approach to den-
sity feature extraction in specific representation within the de la Vallée Poussin
framework.
Theorem 4.1. The “ρ-de la Vallée Poussin potential functions”

VρV P (x) = GVρ P (Δ; |x − y|)F (y) dy (4.36)
G
and the “ρ-de la Vallée Poussin density function”

VP
Fρ (x) = V Pρ (|x − y|)F (y) dy (4.37)
G
satisfy the relations

sup |V (x) − VρV P (x)| = O(ρ2 ), ρ→0 (4.38)
x∈G
and
lim sup |α(x)F (x) − FρV P (x)| = 0, (4.39)
ρ→0
x∈G
where α(x) is the solid angle subtended by the boundary ∂G at x ∈ G.

Unfortunately, de la Vallée Poussin potentials VρV P do not generally show a
faster convergence to V than VρH .
Approximate mollifier solution. In similarity to our previous Haar considerations
we use the operators

AVρ P [F ] = VρV P = GVρ P (Δ; | · −z|) F (z) dz, F ∈ L2 (G), (4.40)
G
and
IρV P [F ] = FρV P = V Pρ (| · −z|) F (z) dz, F ∈ L2 (G). (4.41)
G
We denote the image of X = L2 (G) under the operator AVρ P by YρV P . So, instead

of discussing the integral A[F ](x) = G G(Δ; |x − y|) F (y) dy we choose AVρ P [F ],
F ∈ L2 (G), for some sufficiently small ρ > 0. We take advantage of the fact that

G(Δ; |x − z|) DρV P (|y − z|) dz = V Pρ (|x − y|), x, y ∈ G. (4.42)
G
Note that

Δx G(Δ; |x − z|) DρV P (|y − z|) dz = Δx V Pρ (|x − y|) = DρV P (|x − y|), x, y ∈ G.
G
(4.43)
After these preliminaries we are able to conclude that

VP VP
Iρ [F ](x) = Fρ (x) = V Pρ (|x − w|)F (w) dw
G

= G(Δ; |w − z|) Dρ (|x − z|) dz F (w) dw
VP
G G

= Dρ (|x − z)
VP
G(Δ; |w − z|)F (w) dw dz
G G

= Dρ (|x − z|) A[F ](z) dz =
VP
DρV P (|x − z|) V (z) dz.
G G
holds true for x ∈ G, so that

FρV P (x) = DρV P (|x − z|) V (z) dz, x ∈ G. (4.44)
G
The right-hand side of (4.44) is given analytically when the parameter ρ is chosen
appropriately. So, if we define the operator Sρ : YρV P → X in the form

VP
Fρ = Sρ [V ] = DρV P (| · −z|) V (z) dz, x ∈ G, (4.45)
G
then, by (4.44), this operator maps the gravitational potential to mollified solutions
of (ITGP). This property motivates the term mollified inverse of A used for Sρ .
The discretization of the identity (4.45) given by

N
FρV P (x) & wi DρV P (|x − ziN |) V (ziN ), ziN ∈ G, x ∈ G (4.46)
i=1
may serve as an alternative to improve local density knowledge from given internal
(e.g., borehole) data V (ziN ), i = 1, . . . , N , where wi , i = 1, . . . , N, are the known
integration weights.
Finally, it should be noted that, more generally, any singular integral (cf. [76,
77]) can be chosen in analogy to the de la Vallée Poussin kernel, i.e., smoothed Haar
kernel, as far as its Laplace derivative takes a reasonable role in the mollification
context.
4.3. Singular integral-type mollifier method

First we recapitulate the concept of a singular integral: Let {Kρ }ρ>0 be a family
of functions r → Kρ (r), r ≥ 0, satisfying the following conditions:
(i) Kρ (r) = 0, r > ρ,
(ii) Kρ (r) ≥ 0, r ≥ 0,
(iii) Kρ |[0, ρ] is of class C (∞) ,
(iii) − r12 dr
d 2 d
r dr Kρ (r)|r∈[0,ρ] = 0,
ρ
(iv) 4π 0 Kρ (r) r2 dr = 1.
Then the family {Iρ }ρ>0 of operators Iρ : F → Iρ [F ], F ∈ X, (X = C (0) (R3 ) or
X = L2 (R3 )), given by

Iρ [F ](x) = Fρ (x) = Kρ (|x − y|) F (y) dy = Kρ (|x − y|) F (y) dy (4.47)
R3 Bρ (x)
is called a singular integral in X, if the following approximate identity relation

holds true
lim Iρ [F ] − F X = 0 (4.48)
ρ→0
for all F ∈ X.
Obviously, an example of a singular integral of the aforementioned type is
given by the de la Vallée Poussin kernel.
Let G be a regular region. Suppose that {Kρ }ρ>0 is a kernel constituting a
singular integral in L2 -metric, then it is not difficult to show (see, e.g., [76, 77])
that the limit relation

12
2
lim |Iρ [F ](x) − F (x)| dx =0 (4.49)
ρ→0 G
holds true for all F ∈ L2 (G), while, for all F ∈ C (0) (G), we have
lim sup |Iρ [F ](x) − α(x) F (x)| = 0, (4.50)
ρ→0
x∈G
where α(x) is the solid angle subtended by the boundary ∂G at the point x ∈ G.
Correspondingly to the family {Kρ }ρ>0 we are led to families {Gρ }ρ>0 and
{Dρ }ρ>0 such that
−Δx Gρ (Δ; |x − y|) = Kρ (|x − y|), x, y ∈ R3 (4.51)
and
−Δx Kρ (|x − y|) = Dρ (|x − y|), x, y ∈ R3 . (4.52)
Our interest now is in the terrestrial gravimetry problem (TGP), that may be
regarded as particularly relevant problem in geoscientific practice (our considera-
tions, however, remain valid for (ITGP)). We start from known values V (xi ), xi ∈
∂G, i = 1, . . . , N, given by

A[F ](xi ) = G(Δ; |xi − z|) F (z) dz = V (xi ), xi ∈ ∂G, i = 1, . . . , N, (4.53)
G
which can be thought of as resulting from moment discretization of the gravimetry

integral equation (cf. (1.1))

A[F ](x) = G(Δ; |x − z|) F (z) dz = V (x), x ∈ ∂G, F ∈ L2 (G). (4.54)
G
(TGP) aims at determining an approximation of the function F ∈ L2 (G) from the

N equations (4.53). Introducing the following settings

(N )
A[F ] := G(Δ; |x1 − y|) F (y) dy, . . . , G(Δ; |xN − y|) F (y) dy ,
G G
(4.55)
(N ) T
v := (V (x1 ), . . . , V (xN )) , (4.56)
we are able to rewrite the equations (4.53) in operator form as follows:
(N )
A : L2 (G) → RN , F → (N ) v = (N ) A [F ]. (4.57)
(N )
We look for an approximate inverse S : RN → L2 (G) for (N )
A in the form

N
(N )
S t := V (xi ) D(|xi − ·|) (4.58)

i=1
=ti
in terms of functions D(|xi − ·|) ∈ L2 (G), i = . . . , N, satisfying

N

(N ) (N )
S A [F ] = G(Δ; |xi − z|)F (z) dz D(|xi − ·|)
i=1 G

N
= F (z) G(Δ; |xi − z|) D(|xi − ·|) dz. (4.59)
G k=1
Now the stage is set for explaining the mollifier philosophy, i.e., the sum

N
G(Δ; |x − xi |) D(|xi − y|) (4.60)
i=1
is understood as discrete version of the “continuous expression”

G(Δ; |x − z|) D(|z − y|) dz & δ(|x − y|) (4.61)
G
whose “mollifier version” for some family {Kρ }ρ>0 constituting a singular integral
is given by
G(Δ; |x − z|) Dρ (|z − y|) dz = Kρ (|x − y|), (4.62)
G
with sufficiently small ρ > 0. This observation leads to the sum

N
(N )
Sρ t = V (xi ) Dρ (|xi − ·|) (4.63)
i=1
and
N

(N ) (N )
Sρ A [F ] = G(Δ; |xi − z|)F (z) dz Dρ (|xi − ·|)
i=1 G

N
= F (z) G(Δ; |xi − z|) Dρ (|xi − ·|) dz. (4.64)
G k=1
(N ) (N ) (N )
as approximations to S t and S A [F ], respectively.
Moment method. Next we mention the finite moment problem for (TGP). For that
purpose we assume that the N potential (volume integral) values

G(Δ; |xi − y|) F (y) dy = V (xi ), xi ∈ ∂G, i = 1, . . . , N. (4.65)
G
are known.
The standard solution process (see, e.g., [19, 60]) consists of finding a linear
combination in terms of the functions x → G(Δ; |xi − x|), x ∈ G, xi ∈ ∂G, i =
1, . . . , N. In other words, the moment method looks for a function F ∈ XN satis-
fying the conditions (4.65), where XN is given by
XN := spani=1,...,N G(Δ; |xi − ·|). (4.66)
As a consequence, the moment solution is a harmonic function inside G.
More formally, consider again a semi-discrete observation operator (N ) A :

L2 (G) → RN , F → (N ) v = (N ) A [F ], of type (4.55), (4.56). Remembering F ∈ XN
and choosing F as the linear combination

N
F = βi G(Δ; |xi − ·|) (4.67)
k=1
we are led to a (uniquely solvable) linear system in the unknowns β1 , . . . , βN , viz.

N
βi G(Δ; |xi − y|) G(Δ; |xj − y|) dy = V (xj ), j = 1, . . . , N, (4.68)
k=1 G
that turns out to play a central role in the context of minimum norm (spline)
interpolation in reproducing kernel Hilbert spaces as discussed in Section 5.
4.4. Backus–Gilbert method

The concept originally proposed by Backus and Gilbert (cf. [5–7]) is that one does
not primarily wish to solve the finite moment problem as explained above, but
rather one is interested in how well all possible candidates for solution can be
recovered pointwise. More specifically, the Backus–Gilbert method is based on a
pointwise minimization criterion: Keep y ∈ G fixed and determine the numbers
μi (= μi (y)), i = 1, . . . , N , as the solution of the following minimization problem:
2
N

|z − y|2 μi G(Δ; |xi − z|) dz → min. (4.69)
G
i=1
subject to μ ∈ R , μ = (μ1 , . . . , μN )T with

N
N
μi G(Δ; |xi − z|) dz = 1. (4.70)
G i=1
It should be remarked that the factor z → |z − y|2 , z ∈ G, in the integrand of

N
(4.69) is a measure for the concentration of the sum i=1 μi G(Δ; |xi − y|) around
the point y ∈ G under consideration. In the literature (see, e.g., [66, 92]), more
generally, the term z → |z − y|2ν , z ∈ G, ν ≥ 1, is sometimes chosen. In this case,
N
the larger ν, the more concentrated is the sum i=1 μi G(Δ; |xi − y|) around
y ∈ G.
In matrix-vector nomenclature (thereby omitting the dependence on the

fixed, but arbitrary point y ∈ G) we are able to rewrite the quadratic optimization
problem (4.69), (4.70), in the form
μ· Q μ → min, (4.71)
subject to
κ · μ = 1, (4.72)
where

(Q)i,j := |z − y|2 G(Δ; |xi − z|) G(Δ; |xj − z|) dz, i, j = 1, . . . , N, (4.73)
G
and
κj := G(Δ; |xj − z|) dz, j = 1, . . . , N. (4.74)
G
In fact, (4.71) and (4.72) is a quadratic minimization problem with only one linear
equation constraint. We may assume that κ = (κ1 , . . . , κN )T is different from
0, since otherwise the constraint (4.72) cannot be valid. The introduction of a
Lagrange multiplier well known from optimization theory (see, e.g., [110]) can be
used to characterize the solvability of the resulting linear Qμ − λκ = 0 under the
constraint κ·μ = 1, i.e., existence and uniqueness. In more detail, from the integral

in (4.69), we see that μ· Q μ ≥ 0 and μ· Q μ = 0 implies N i=1 μi G(Δ; |xi −·|) = 0,
so that the linear independence of the system {G(Δ; |xi − ·|)}i=1,...,N shows that
Q is positive definite.
Summarizing our results we therefore obtain the following statement:
The symmetric matrix Q ∈ RN ×N as defined by (4.73)is positive definite for
every y ∈ G. The quadratic minimization problem (4.71) and (4.72) is uniquely
solvable. The vector μ is the unique solution of (4.71) and (4.72) if and only if
there exist a real number λ (the Lagrange multiplier) so that (μ, λ) ∈ RN +1 solves
the linear system Qμ − λκ = 0 under the constraint κ · μ = 1.
The Lagrange multiplier λ = μ · Q μ represents the minimal value of the
quadratic minimization problem.
Consider the unique solution μ ∈ RN , μ = (μ1 , . . . , μN )T , μi = Mi (y), i =
1, . . . , N, of the quadratic minimization problem (4.71) and (4.72). The Backus–
Gilbert solution FN of the discrete version of (TGP)

G(Δ; |xi − y|) FN (y) dy = V (xi ), xi ∈ ∂G, i = 1, . . . , N (4.75)
G
is defined by

N
FN (y) = V (xi ) μi , y ∈ G. (4.76)
i=1
The minimal value λ (more accurately, λ(y)) is called the spread.
As already mentioned, the Backus–Gilbert solution (4.76) generally is not

a solution of the finite moment problem (4.65). This observation is certainly a
disadvantage. Therefore, the question arises if the error may be estimated in
an appropriate way (see [60] for related work in one-dimensional context): Let
F ∈ L2 (G) be any solution of the finite moment problem (4.65). Suppose that FN
given by (4.76) is the Backus–Gilbert solution. Then, in connection with (4.70), it
follows that

N
N
FN (y) − F (y) = V (xi ) μi − F (y) μi G(Δ; |xi − z|) dz
i=1 G i=1
N

= G(Δ; |xi − z|) (F (z) − F (y)) μi dz (4.77)
i=1 G
holds true. Consequently, we obtain

N

|FN (y) − F (y)| ≤ G(Δ; |xi − z|) μi ) |F (z) − F (y)| dz. (4.78)
G
i=1
Under the assumption of Lipschitz-continuity of F in G, i.e., the existence of a

constant CF so that
|F (z) − F (y)| ≤ CF |z − y|, y, z ∈ G, (4.79)
we are able to deduce that
N

|FN (y) − F (y)| ≤ CF G(Δ; |xi − z|) μi |z − y| dz. (4.80)
G i=1

By virtue of the Cauchy–Schwarz inequality we therefore obtain from (4.80)

N

|FN (y) − F (y)| ≤ CF 1· G(Δ; |xi − z|) μi |z − y| dz
G
i=1
N 2 1/2
#
≤ CF G 2
G(Δ; |xi − z|) μi |z − y| dz . (4.81)

G i=1
For N ∈ N, y ∈ G, we set

2 2 2
eN (y) := min |ZN (z)| |z − y| dz : ZN ∈ XN , ZN (z) dz = 1 .
G G
(4.82)
Thus, we finally arrive at
#
|FN (y) − F (y)| ≤ CF G eN (y) (4.83)
as pointwise error estimate of the difference of the solution of the finite moment
problem (4.65) and the Backus–Gilbert solution (4.76).
We conclude our considerations with the question if the Backus–Gilbert

method admits a relation to the mollifier method: Once again, consider the semi-
discrete observation operator
(N )
A : L2 (G) → RN , F → (N ) v = (N ) A [F ], (4.84)
where

(N )
A[F ] := G(Δ; |x1 − y|) F (y) dy, . . . , G(Δ; |xN − y|) F (y) dy ,
G G
(4.85)
(N ) T
v := (V (x1 ), . . . , V (xN )) . (4.86)
By virtue of the operator (N ) S given by
N
(N )
S v (y) = V (xi ) μi (y), y ∈ G, (4.87)
k=1
(N )
we have constructed a left inverse S : RN → L2 (G) such that
N

(N ) (N )
S A [F ](y) = G(Δ; |xi − z|)F (z) dz μi (y)
i=1 G
N

= F (z) G(Δ; |xi − z|) μi (y) dz,
G i=1

δ(|z−y|)
& F (y). (4.88)

Note that we are formally allowed (in distributional context) to formulate

F (y) = F (z) δ(|z − y|) dz
G

& F (z) G(Δ; |x − z|) M (|x − y|) dx dz, (4.89)
G G
where, in analogy to (4.52), we have

N
−Δz δ(z − y|) = M (|z − y|) & −Δz G(Δ; |xi − z|) μi (y). (4.90)
i=1
5. Reproducing Kernel Hilbert Space (RKHS) Methods

Next we consider reproducing kernel Hilbert space solutions. First we discuss the
classical geodetic External/Terrestrial Gravimetry Problem (ETGP). Then we go
over to the Internal/TerrestrialExternal Gravimetry Problem (ITEGP), i.e., the
gravimetry problem in whole Euclidean space R3 .
5.1. External/terrestrial RKHS for regular regions
Let PHarm (G) and PAnHarm (G) be the orthogonal projector of the space L2 (G) to
Harm(G) and N (AG c ) = AnHarm(G), respectively. Then, every function F of the
Hilbert space L2 (G) can be uniquely decomposed in the form
F = PHarm (G) [F ] + PAnHarm (G) [F ] (5.1)
such that
0 1 0 1 0 1
AG c [F ] = AG c PHarm(G) [F ] + AG c PAnHarm (G) [F ] = AG c PHarm(G) [F ] . (5.2)

=0
Furthermore, it is clear that

- -2 - -2
F 2L2 (G) = -PHarm(G) [F ]-L2 (G) + -PAnHarm (G) [F ]-L2 (G) . (5.3)
In conclusion, AG c [PHarm(G) [F ]] is that function of class L2 (G) that has the small-
est L2 (G)-norm among all (density) functions F in L2 (G) generating the same
potential in the space Y |G c = AG c (L2 (G)). Consequently, to every P ∈ Y |G c ,
there corresponds a unique F ∈ Harm(G) such that
AG c [F ] = AG c [PHarm(G) [F ]] = P. (5.4)
The restriction AG c |Harm(G) is a linear bijective operator, i.e., to every P ∈
Y |G c there exists a unique F ∈ Harm(G) such that AG c |Harm(G)[F ] = P .
On the space Y |G c we are able to impose an inner product ·, ·Y |G c by defin-
ing
? @
AG c |Harm(G)[F ], AG c |Harm(G)[G] Y |G c = F, GL2 (G) , (5.5)
where F, G ∈ L2 (G). Y |G c equipped with the inner product ·, ·Y |G c is a Hilbert
space. AG c |Harm(G) is an isometric operator relating L2 (G) to Y |G c . Our goal is
to show that (Y |G c , ·, ·Y |G c ) is a reproducing kernel Hilbert space, i.e., a Hilbert
space equipped with the reproducing kernel KY |G c (·, ·). It is clear that, for every
x ∈ G c , G(Δ; |x − ·|) is an element of Harm(G). From well-known reproducing
Hilbert space theory (see, e.g., [4]), it follows that any given potential P ∈ Y |G c
can be represented in the form
P (x) = AG c |Harm(G)[F ](x) = G(Δ; |x − ·|), F L2 (G) , x ∈ G c , F ∈ Harm(G).
(5.6)
For x ∈ G c , the evaluation functional Ex [P ] = P (x) is a bounded functional on G c .
Indeed, from the Cauchy–Schwarz inequality applied to (5.6) we get
|Ex [P ]| = |P (x)| ≤ ||F ||L2 (G) ||G(Δ; |x − ·|)||L2 (G) . (5.7)
Consequently, we have
|Ex [P ]| = P (x)| ≤ Cx P Y |G c , P ∈ Y |G c , x ∈ G c . (5.8)
Thus, a necessary and sufficient condition for the Hilbert space Y |G c to possess a
reproducing kernel (see, e.g., [4]) is fulfilled. Even more, we are able to find the
explicit expression of the reproducing kernel KY |G c (·, ·) : G c × G c → R for the
Hilbert space Y |G c such that, for every P ∈ Y |G c , the reproducing property
2 3
P (x) = P, KY |G c (x, ·) , x ∈ Gc, (5.9)
Y |G c
is valid. For x ∈ G c and F ∈ Harm(G) such that AG c [F ] = P , we obtain

P (x) = F, G(Δ; |x − ·|)L2 (G) = AG c [F ], AG c [G(Δ; |x − ·|)]Y |G c
= P, AG c [G(Δ; |x − ·|)]Y |G c . (5.10)
Hence, KY |G c (x, ·) = AG c [G(Δ; |x − ·|)], i.e., we have for x, y ∈ G c :
The integral

1 1
KY |G c (x, y) = G(Δ; |x − ·|), G(Δ; |y − ·|)L2 (G) = dz
(4π)2 G |x − z||y − z|
(5.11)
represents the (unique) reproducing kernel of Y |G c .
Clearly, for “geoscientifically relevant geometries” G such as geoid, real Earth,
etc. the integral (5.11) has to be determined by approximate integration rules.
Summarizing our considerations we end up with the following result:

Y |G c , ·, ·Y |G c is a Hilbert space possessing the reproducing kernel (5.11)

KY |G c (x, y) = G(Δ; |x − z|) G(Δ; |y − z|) dz
G

1 1
= 2
dz. (5.12)
(4π) G |x − z||y − z|
Equation (5.12) formally states that, for every fixed x ∈ G c , the function
KY |G c (x, ·) = AG c [G(Δ; |x − ·|)] is the Newtonian potential corresponding to the
“density function” G(Δ; |x − ·|).
5.2. External/terrestrial RKHS for balls

For the special case of a ball Bβ (0) of radius β around the origin the kernel
KY |Bβc (0) (·, ·) given by

1 1
KY |Bβc (0) (x, y) = dz, (5.13)
(4π)2 Bβ (0) |x − z||y − z|
can be expressed as series representation by use of the expansion (see, e.g., [38])
∞
1 |y|n x y
G(Δ; |x − y|) = Pn · , |y| < |x|,
4π n=0 |x|n+1 |x| |y|
where Pn is the Legendre polynomial of degree n.

∞ 2 n+1
β 1 β x y
KY |Bβc (0) (x, y) = Pn · . (5.14)
4π n=0 (2n + 1)(2n + 3) |x||y| |x| |y|
We are interested in an explicit expression of the infinite Legendre sum (5.14).

To this end, we have a closer look at the term
1
(2n + 1)(2n + 3)
that can be decomposed via partial fraction decomposition in the form
1 1 1
= − .
(2n + 1)(2n + 3) 2(2n + 1) 2(2n + 3)
As a consequence, the reproducing kernel can be rewritten in the form
∞ 2 n
β3 1 β x y
KY |Bβ (0) (x, y) =
c Pn ·
8π|x||y| n=0 2n + 1 |x||y| |x| |y|
∞ 2 n
β3 1 β x y
− Pn · . (5.15)
8π|x||y| n=0 2n + 3 |x||y| |x| |y|
We only consider the Legendre expansions
∞
β3 1
Φ1 (h, t) = h2n Pn (t),
8π|x||y| n=0 2n + 1
∞
β3 1
Φ2 (h, t) = h2n Pn (t)
8π|x||y| n=0 2n + 3
for h := √ β < 1 and t := x

|x| · y
|y| ∈ (−1, 1). The remaining cases follow
|x||y|
accordingly. Recalling the generating series in terms of the Legendre polynomials
(see, e.g., [1, 68])
∞
1
h2n Pn (t) = √ , h ∈ [0, 1), t ∈ [−1, 1]. (5.16)
n=0
1 + h4 − 2h2 t
we obtain by integration of both sides of (5.16) with respect to h
∞
1 1
h2n+1 Pn (t) = √ dh = F1 (h, t), (5.17)
n=0
2n + 1 1 + h4 − 2h2 t
where F1 (h, t) is the abbreviation given by
2
2
i −t+h√t2 −1 + 1 − t+√ht2 −1 + 1
F1 (h, t) = − √
1
√
−t+ t2 −1
1 + h4 − 2h2 t
√
−1 1 t − t2 − 1
× E1 i sinh h √ , √ (5.18)
−t + t2 − 1 t + t2 − 1
and E is an elliptic integral of first kind. Then the sum Φ1 is given by
β3 1
Φ1 (h, t) = F1 (h, t).
8π|x||y| h
For the determination of the sum Φ2 we multiply Equation (5.16) by h2 :

∞
∞

2n+2 2 h2
h Pn (t) = h h2n Pn (t) = √ .
n=0 n=0
1 + h4 − 2h2 t
Integrating the last equation with respect to h we obtain
∞
1 2n+3 h2
h Pn (t) = √ dh = F2 (h, t),
n=0
2n + 3 1 + h4 − 2h2 t
where F2 (h, t) is given by
√
2 2
( t2 − 1 + t) −t+h√t2 −1 + 1 − t+√ht2 −1 + 1
F2 (h, t) = √
1
√
2
−t+ t −1
1 + h4 − 2h2 t
√
1 t − t 2−1
× E2 i sinh−1 h √ , √
−t + t2 − 1 t + t2 − 1
√
−1 1 t − t2 − 1
− E3 i sinh h √ , √ ,
−t + t2 − 1 t + t2 − 1
and E2 and E3 are elliptic integrals of first and second kind, respectively. Hence,
it follows that
β3 1
Φ2 (h, t) = F2 (h, t).
8π|x||y| h3
Altogether, for
β x y
# < 1 and · ∈ (−1, 1),
|x||y| |x| |y|
the reproducing kernel KY |Bβc (0) (x, y), x, y ∈ Bβc (0), can be represented in the form
#
β3 |x||y| β x y
KY |Bβc (0) (x, y) = F1 # , ·
8π|x||y| β |x||y| |x| |y|
# 3
|x||y| β x y
− F2 # , · . (5.19)
β3 |x||y| |x| |y|
5.3. External/terrestrial/internal RKHS for regular regions
Looking at the kernel given by

(x, y) → G(Δ; |x − z|) G(Δ; |z − y|) dz (5.20)
G
we see that it is defined for all x, y ∈ R3 (with x, y ∈ G c constituting even a regular

integral expression). Furthermore, to every F ∈ L2 (G) there exists a unique V ∈ Y
of the form

V (x) = A[F ](x) = G(Δ; |x − y|)F (y) dy, x ∈ R3 . (5.21)
G
On the space Y we are able to impose an inner product ·, ·Y by setting
A[F ], A[G]Y = F, GL2 (G) , (5.22)
2
where F, G ∈ L (G). Y equipped with the inner product ·, ·Y is a Hilbert space.
For all x ∈ R3 , the Cauchy–Schwarz inequality yields the estimate

|V (x)| ≤ 2
| G(Δ; |x − y|)| dy |F (y)|2 dy, (5.23)
G G
where we already know that there exists a constant Cx such that

|V (x)| ≤ Cx |F (y)|2 dy, (5.24)
G
holds true for all x ∈ G c . Moreover, for all x ∈ G and some R ≥ d with d =
diam(G) = maxx,y∈G |x − y| we are able to see that

1 1
|G(Δ; |x − y|)|2 dy = dy
G (4π) G |x − y|2
2

1 1
≤ dy
(4π)2 BR (x) |x − y|2
R
1 1
= dS(y) dr
(4π)2 0 |x−y|=r |x − y|2
R
= . (5.25)
4π
All in all, for each fixed x ∈ R3 , the evaluation functional Ex is bounded. Hence, a
necessary and sufficient condition that (Y, ·, ·Y ) be a reproducing kernel Hilbert
space (see, e.g., [4, 16]) is satisfied. In fact, for x ∈ G and F ∈ L2 (G), we obtain
V (x) = G(Δ; |x − ·|), F L2 (G)
= A[G(Δ; |x − ·|)], A[F ]Y
= A[G(Δ; |x − ·|)], V Y , (5.26)
so that

KY (x, y) = G(Δ; |x − z|) G(Δ; |z − y|) dz
G

1 1 1
= dz, x, y ∈ R3 (5.27)
(4π)2 G |x − z| |z − y|
is the unique reproducing kernel of Y . Summarizing our considerations we are
finally allowed to formulate the following statement [98]:
The image space Y = A[L2 (G)] is a reproducing kernel Hilbert space process-
ing the reproducing kernel

KY (x, y) = G(Δ; |x − z|) G(Δ; |z − y|) dz, x, y ∈ R3 . (5.28)
G
Mollifier realization. Denoting by Yρ , ρ > 0, a space of all mollified singular

integral-type Newton integrals Aρ [F ] given by

Aρ [F ] = Gρ (Δ; |x − y|) F (y) dy, F ∈ L2 (G), (5.29)
G
with Gρ (Δ; | · − · |) given by either (2.20) or (4.51), so that Yρ = Aρ [L2 (G)], we are
led to an analogous result in the framework of singular integral-type mollification:
The image space Yρ = Aρ [L2 (G)] is a reproducing kernel Hilbert space pos-
sessing the reproducing kernel

KYρ (x, y) = Gρ (Δ; |x − z|) Gρ (Δ; |z − y|) dz, x, y ∈ R3 . (5.30)
G
Finally, it should be mentioned that

−Δx KYρ (x, y) = Kρ (|x − z|) Gρ (Δ; |z − y|) dz, x, y ∈ R3 . (5.31)
G
Remark. The mathematical structures and results developed for the gravimetry
problem enable us to apply a large variety of ideas and concepts known from the
theory of ill-posed problems (see, e.g., [35] for a geodetically relevant approach).
In our work, we do not consider the details.
However, it should be remarked that reproducing kernel Hilbert space struc-
ture is of particular importance in the inversion of Newton’s Law of Gravitation,
since the reproducing property makes a numerical computation efficient and eco-
nomical (as we shall see from the following gravimetric spline context).
5.4. External/terrestrial/internal spline theory for regular regions
Let G be a regular region. Suppose that {x1 , . . . , xN }, xi = xj , i = j, is a discrete set
of N given points in R3 . Assume that the values γi = V (xi ), xi ∈ R3 , i = 1, . . . , N ,
constitute a given data set from the Newton potential (1.1). We want to find an
V
approximation SN to the potential V such that
V
SN (xi ) = V (xi ) = γi , i = 1, . . . , N. (5.32)
(If the data are noisy, interpolation should be replaced by smoothing (see, e.g.,
[41] and the references therein).) A functional value V (x) at a point x ∈ R3 can
be identified with an evaluation functional
Ex : V → Ex [V ] = V (x), V ∈ Y (G c ). (5.33)
3
For each x ∈ R , the linear functional Ex defined by Ex : V → Ex [V ] = V (x), V ∈ Y,
is bounded on Y , i.e, |Ex [V ]| = |V (x)| ≤ Cx V Y . Moreover, for x ∈ ∂G and for
all V ∈ Y we have Ex [V ] = V (x) = (V, KY (x, ·))Y ) .
Spline method. The Newton potential V , from which the discrete data are known,
is considered as an element of the Hilbert space Y possessing the reproducing kernel
KY (·, ·), while the observed values at the points x1 , . . . , xN ∈ R3 are assumed to be
associated with linearly independent bounded functionals Ex1 , . . . , ExN . In doing

so, we are able to find a minimum norm solution SN V
∈ Y as a linear combination
of the representers Exi [KY (·, ·)] to the functionals Ex1 , . . . , ExN , i.e., SN
V
is meant as
the projection of V to the N -dimensional linear subspace spanned by the linearly
independent representers Exi [KY (·, ·)], i = 1, . . . , N (see, e.g., [16]).
Let {x1 , . . . , xN } ⊂ R3 be a point system, such that the evaluation functionals
Ex1 , . . . , ExN are linearly independent. Then, within the set
IEVx1 ,...,Ex = {U ∈ Y : Exi [U ] = Exi [V ] = γi , i = 1, . . . , N }, (5.34)
N
V
the minimum norm interpolation problem of finding SN that satisfies
SN
V
Y = inf U Y (5.35)
U∈IEVx
1 ,...,ExN
is well posed, i.e., its solution exists, is unique and depends continuously on the
V
data γ1 , . . . , γN . The uniquely determined solution SN is given in the explicit form

N
V
SN (x) = i Exi [KY (x, ·)],
aN x ∈ R3 , (5.36)
i=1

1 , . . . , aN are determined by solving the linear system of
equations
N
i Exi Exj [KY (·, ·)] = γj ,

aN j = 1, . . . , N. (5.37)
i=1
As a consequence of the interpolation procedure, the density inside G is ob-

tained as linear combination in terms of fundamental solutions with singularities
in the points {x1 , . . . , xN } :

N
F
SN (x) = −Δx SN
V
(x) =− i Exi [G(Δ; |x − ·|)]
aN
i=1

N
=− i G(Δ; |x − xi |),
aN x ∈ G\{x1 , . . . , xN }. (5.38)
i=1
F
As a consequence, SN is a harmonic function provided that {x1 , . . . , xN } ⊂ G c .
Spline mollifier method. For sufficiently small ρ, an approximate version of the
kernel

KY (x, y) = G(Δ; |x − z|) G(Δ; |y − z|) dz (5.39)
G
is given by

KYρ (x, y) = Gρ (Δ; |x − z|) Gρ (Δ; |y − z|) dz. (5.40)
G
Note that, from the integral in (5.40), we see that

N N
ak ai Gρ (Δ; |xk − z|) Gρ (Δ; |xi − z|) dz
k=1 i=1 G
2
N
= ak Gρ (Δ; |xk − z|) dz ≥ 0. (5.41)
G k=1
Moreover, the linear independence of the system {Gρ (Δ; |xi − ·|)}i=1,...,N implies
that the Gram matrix

Gρ (Δ; |xi − z|) Gρ (Δ; |xk − z|) dz (5.42)
G k,i=1,...,N
is positive definite, so that KYρ (·, ·) is a positive definite kernel.

In other words, the integral (5.40) defines a Hilbert space Yρ , ·, ·Yρ pos-
sessing (5.40) as the reproducing kernel. In the space Yρ minimum norm (spline)
interpolation as described above can be performed in analogous way:
Let {x1 , . . . , xN } ⊂ R3 be a point system, such that the evaluation functionals
Ex1 , . . . , ExN are linearly independent. Then, within the set

IEVx1 ,...,Ex = {U ∈ YρHJ G c : Exi [U ] = Exi [V ] = γi , i = 1, . . . , N }, (5.43)
N
V
the minimum norm interpolation problem of finding SN that satisfies
SN
V
Yρ = inf U Yρ (5.44)
U∈IEVx
1 ,...,ExN
is well posed, i.e., its solution exists, is unique and depends continuously on the
V
data γ1 , . . . , γN . The uniquely determined solution SN is given in the explicit form

N
V
SN (x) = i Exi [KYρ (x, ·)],
aN x ∈ R3 , (5.45)
i=1

1 , . . . , aN are determined by solving the linear system of
equations
N
i Exi Exj [KYρ (·, ·)] = γj ,

aN j = 1, . . . , N. (5.46)
i=1
In this case, we obtain an approximation of the density distribution as a linear

combination of singular integral-type kernels (5.31) which are not harmonic.
6. Concluding remarks
Beside gravimetry the (Newton) volume potential (2.10) turns out to appear from
different points of view in the context of mathematical geodesy (see, e.g., [26, 29,
31, 32, 34, 38, 52, 55–57, 61, 62, 80, 81, 94]).
Terrestrial measurements of the gravitational force intensity |v(x)| = |∇V (x)|

typically lead to an oblique derivative problem for the Earth’s gravitational poten-
tial V . The obliqueness is a result of the fact that the real Earth’s surface does
not coincide with the geoidal surface (except over certain parts on oceans).
Satellite measurements on orbits lead to vectorial and/or tensorial deriva-
tives of first and second order. Depending on the type of measurements, modern
satellite problems for the determination of the Earth’s external gravitational po-
tential are categorized as Satellite-to-Satellite Tracking (SST) problems (i.e., ∇V
is derivable from orbit deviations) or Satellite Gravity Gradiometry (SGG) prob-
lems (i.e., ∇2 V = (∇ ⊗ ∇) V is available by gradiometer measurements on the
orbit).
References
[1] Abramowitz, M., Stegun, I.A.: Handbook of Mathematical Functions, Dover Publi-
cations, Inc., New York, 1964
[2] Anger, G.: A Characterization of Inverse Gravimetric Source Problem through Ex-
tremal Measures. Rev. Geophys. Space Phys., 19:299–306, 1981.
[3] Anger, G.: Inverse Problems in Differential Equations. Akademie-Verlag, Berlin,
1990.
[4] Aronszajn, N.: Theory of Reproducing Kernels. Trans. Am. Math. Soc., 68:337–404,
1950.
[5] Backus, G.E., Gilbert, F.: Numerical Applications of a Formalism for Geophysical
Inverse Problems. Geophys. J.R. Astron. Soc., 13:247–276, 1967.
[6] Backus, G.E., Gilbert, F.: The Resolving Power of Gross Earth Data. Geophys.
J.R. Astron. Soc., 16:169–205, 1968.
[7] Backus, G.E., Gilbert, F.: Uniqueness of the Inversion of Inaccurate Gross Earth
Data. Philos. Trans. R. Soc. London ., 226:123–197, 1970.
[8] Ballani, L.: Solving the Inverse Gravimetric Problem: On the Benefit of Wavelets.
In: Geodetic Theory Today, Proceedings of the 3rd Hotine–Marussi Symposium on
Mathematical Geodesy 1994 (Sansò, F. Ed.), 151–161, Springer, Berlin, 1995.
[9] Ballani, L., Engels, J., Grafarend, E.W.: Global Base Functions for the Mass Density
in the Interior of a Massive Body (Earth). Manuscr. Geod., 18:99–114, 1993.
[10] Ballani, L., Stromeyer, D.: The Inverse Gravimetric Problem: A Hilbert Space Ap-
proach. In: Proceedings of the International Symposium Figure of the Earth, the
Moon, and Other Planets 1982 (Holota, P. Ed.), 359–373, Prague, 1983.
[11] Ballani, L., Stromeyer, D., Barthelmes, F.: Decomposition Principles for Linear
Source Problems. In: Inverse Problems: Principles and Applications in Geophysics,
Technology, and Medicine, Math. Res. 47 (Anger, G., Gorenflo, R., Jochmann, H.,
Moritz, H., Webers, W., Eds.), Akademie-Verlag, Berlin, 1993.
[12] Barzaghi, R., Sansò, F.: Remarks on the Inverse Gravimetric Problem. Boll. Geod.
Scienze Affini, 45:203–216, 1986.
[13] Berkel, P., Fischer, D., Michel, V. : Spline Multiresolution and Numerical Results
for Joint Gravitation and Normal Mode Inversion with an Outlook on Sparse Reg-
ularization. Int. J. Geomath., 1:167–204, 2010.
[14] Blakely, R.J.: Potential Theory in Gravity and Magnetic Applications. Cambridge
University Press, Cambridge, 1996.
[15] Blick, C.: Multiscale Potential Methods in Geothermal Research: Decorrelation Re-
flected Post-Processing and Locally Based Inversion. Ph.D.-thesis, Geomathematics
[16] Davis, P.J.: Interpolation and Approximation. Blaisdell, New York, 1963.
[17] Dunford N., Schwartz, J.: Linear Operators, Vol. II, Wiley-Interscience, New York,
1963.
[18] Engl, H.: Integralgleichungen. Springer Lehrbuch Mathematik, Wien, 1997.
[19] Engl, H.W., Hanke, M., Neubauer, A.: Regularization of Inverse Problems. Kluwer,
Dordrecht, 1996.
[20] Engl, H., Louis, A.K., Rundell, W. (Eds.): Inverse Problems in Geophysical Appli-
cations. SIAM, Philadelphia, 1997, 231:244–261, 1992.
[21] Folland, G.G.: Introduction to Partial Differential Equations, Princeton University
Press, Princeton, 1995.
[22] Fredholm, I.: Sur une nouvelle méthode pour la résolution du problème de Dirichlet,
Översigt Kongl. Vetenskaps-Akademiens Förhandlingar 57, 39–46, 1900.
[23] Freeden, W.: On the Approximation of External Gravitational Potential With
Closed Systems of (Trial) Functions. Bull. Géod., 54: 1–20, 1980.
1981.
[25] Freeden, W.: Interpolation by Multidimensional Periodic Splines, J. Approx. The-
ory, 55:104–117, 1988.
[26] Freeden, W.: Multiscale Modelling of Spaceborne Geodata, B.G. Teubner, Stuttgart,
Leipzig, 1999.
[27] Freeden, W.: Geomathematics: Its Role, its Aim, and its Potential. In: W. Freeden,
Z. Nashed, and T. Sonar (Eds.), Handbook of Geomathematics. 2nd edition, Vol. 1,
Chapter 1, 3–78, Springer, New York, Heidelberg, 2015.
[28] Freeden, W., Blick, C.: Signal Decorrelation by Means of Multiscale Methods. World
of Mining, 65: 1–15, 2013.
[29] Freeden, W., Gerhards, C.: Geomathematically Oriented Potential Theory. Chap-
man and Hall, CRC Press, Boca Raton, London, New York, 2013.
[30] Freeden, W., Gutting, M.: Special Functions of Mathematical (Geo-)Physics.
Birkhäuser, Basel, 2013.
Surface of the Earth. Veröff. Geod. Inst. RWTH Aachen, 29, 1980.
Derivative Problem in Potential Theory. Math. Meth. Appl. Sci., 4:104–114, 1981.
[33] Freeden, W., Maier, T.: Spectral and Multiscale Signal-to-Noise Thresholding of
Spherical Vector Fields. Computational Geosciences, 7(3):215–250,2003.
[35] Freeden, W., Nashed, M.Z.: Operator-Theoretic and Regularization Approaches to
Ill-Posed Problems. GEM Int. J. Geomath., https://doi.org/10.1007/s13137-017-
0100-0, 2017.
[36] Freeden, W., Nutz, H.: Satellite Gravity Gradiometry as Tensorial Inverse Problem.,
Int. J. Geomath., 2:177–218, 2012.
[37] Freeden, W., Schneider, F.: Regularization Wavelets and Multiresolution, Inverse
Problems, 14:493–515, 1998.
[38] Freeden, W., Schreiner, M.: Spherical Functions of Mathematical Geosciences – A
Scalar, Vectorial, and Tensorial Setup. Springer, Heidelberg, 2009.
Tensorial Solution. In: Handbook of Geomathematics, Vol. 1 (Freeden, W., Nashed,
M.Z., Sonar,T., (Eds.)), 269–302, Springer, Heidelberg, 2010.
Tensorial Solution. In: Handbook of Geomathematics, Freeden, W., Nashed, M.Z.,
and Sonar, T., Eds., Springer, 2nd edition, New York, Heidelberg, 2nd edition,
2015.
[41] Freeden, W. and Witte, B.: A Combined (Spline-)Interpolation and Smoothing
Method for the Determination of the Gravitational Potential from Heterogeneous
Data. Bull. Geod., 56:53–62, 1982.
[42] Freeden, W., Michel, V., Nutz, H.: Satellite-to-Satellite Tracking and Satellite Grav-
ity Gradiometry (Advanced Techniques for High-Resolution Geopotential Field De-
termination). J. Eng. Math., 43: 19–56, 2002.
[43] Freeden, W., Schneider, F., Schreiner, M.: Gradiometry – An Inverse Problem in
Modern Satellite Geodesy, in: Engl, H.W., Louis, A., Rundell, W. (Eds.), GAMM-
SIAM Symposium on Inverse Problems: Geophysical Applications, 179–239, 1997.
[44] Glockner, O.: On Numerical Aspects of Gravitational Field Modelling from SST
and SGG by Harmonic Splines and Wavelets (with Application to CHAMP Data),
Ph.D.-thesis, Geomathematics Group, University of Kaiseslautern, 2001.
[45] Grafarend, E.W.: Six Lectures on Geodesy and Global Geodynamics. In: Pro-
ceedings of the Third International Summer School in the Mountains (Moritz, H.,
Sünkel, H., Eds.), 531–685, 1982.
[46] Groetsch, C.W.: The Theory of Tikhonov Regularization for Fredholm Equations of
the First Kind. Pitman, London-Boston, 1984.
schweig, 1993.
[48] Haar, A.: Zur Theorie der orthogonalen Funktionensysteme. Math. Ann., 69:331–
371, 1910.
[49] Hadamard, J.: Sur les problèmes aux dérivés partielles et leur signification physique.
Princeton Univ. Bull., 13:49–52, 1902.
[50] Hadamard, J.: Lectures on the Cauchy Problem in Linear Partial Differential Equa-
tions, Yale University Press, New Haven, 1923.
[51] Hanson, R.J.: A Numerical Methods for Solving Fredholm Integral Equations of
the First Kind. SIAM J. Numer. Anal. 8:616–662, 1971.
[52] Heiskanen, W.A., Moritz, H.: Physical Geodesy. Freeman, San Francisco, 1967.
[53] Helmert, F.: Die Mathematischen und Physikalischen Theorien der Höheren Geo-
däsie 2. B.G. Teubner, Leipzig, 1884.
[54] Hille, E.: Introduction to the General Theory of Reproducing Kernels, Rocky Moun-
tain, J. Math., 2:321–368, 1972.
[55] Hörmander, L.: Pseudodifferential Operators. Comm. Pure Appl. Math., 18:501–
517, 1965.
[56] Hörmander, L.: The Boundary Problems of Physical Geodesy. Archive for Rational
Mechanics and Analysis, 62:1–52, 1976.
[57] Hofmann-Wellenhof, B., Moritz, H.: Physical Geodesy, Springer, Wien, New York,
2005.
[58] Jantscher, L.: Distributionen. Walter de Gruyter, Berlin, New York, 1971.
pany, New York, 1929.
[60] Kirsch, A.: An Introduction to the Mathematical Theory of Inverse Problems. 2nd
ed., Springer, Heidelberg, 1996.
Problem Using the Known Surface of the Earth. Bulletin Géodésique, 106:467–476,
1972.
Danish Geodetic Institute, Report No. 44, Copenhagen, 1969.
[63] Lavrentiev, M.M.: Some Improperly Posed Problems of Mathematical Physics, Iz-
dat. Sibirsk. Otdel, Akad. Nauk. SSSR, Novosibirsk, 1962, English Transl., Springer
Tracts in Natural Philosophy, Vol. 11, Springer-Verlag, Berlin, 1967.
[64] Lieusternik, L.A., Sobolev, V.J.: Elements of Functional Analysis. New York: Un-
gar, 1961.
[65] Locker J., Prenter, P.M.: Regularization with Differential Operators. J. Math. Anal.
Appl., 74:504–529, 1980.
[66] Louis, A.K.: Inverse und schlecht gestellte Probleme. Teubner, Stuttgart, 1989.
[67] Louis, A.K., Maass, P.: A Mollifier Method for Linear Equations of the First Kind.
Inverse Problems, 6:427–440, 1989.
[68] Magnus, W., Oberhettinger, F., Soni, R.P.: Formulas and Theorems for the Spe-
cial Functions of Mathematical Physics. In: Die Grundlehren der mathematischen
Wissenschaften in Einzeldarstellungen, Band 52, Springer, Berlin, 3. Auflage, 1966.
[69] Meissl, P.A.: A Study of Covariance Functions Related to the Earth’s Disturbing
Columbus, OH, (1971).
[70] Meissl, P.A.: Hilbert Spaces and Their Applications to Geodetic Least Squares
Problems. Boll. Geod. Sci. Aff., N. 1, 181–210, 1976.
[71] Michel, V.: A Multiscale Method for the Gravimetry Problem: Theoretical and Nu-
merical Aspects of Harmonic and Anharmonic Modelling. Ph.D.-thesis, Geomath-
ematics Group, University of Kaiserslautern, Shaker, Aachen, 1999.
[72] Michel, V.: Scale Continuous, Scale Discretized and Scale Discrete Harmonic
Wavelets for the Outer and the Inner Space of a Sphere and Their Application
to an Inverse Problem in Geomathematics. Appl. Comp. Harm. Anal. (ACHA),
12:77–99, 2002.
[73] Michel, V.: A Multiscale Approximation for Operator Equations in Separable Hilbert
Spaces – Case Study: Reconstruction and Description of the Earth’s Interior. Ha-
bilitation Thesis, University of Kaiserslautern, Geomathematics Group, Shaker,
Aachen, 2002a.
[74] Michel, V: Regularized Wavelet-Based Multiresolution Recovery of the Harmonic
Mass Density Distribution from Data of the Earth’s Gravitational Field at Satellite
Height. Inverse Problems, 21:997–1025, 2005.
[75] Michel, V., Fokas, A.S.: A Unified Approach to Various Techniques for the Non-
Uniqueness of the Inverse Gravimetric Problem and Wavelet-Based Methods. In-
verse Problems, 24: doi:10.1088/0266–5611/24/4/045019, 2008.
[76] Michlin, S.G.: Multidimensional Singular Integrals and Integral Equations. Perga-
mon Press, New York, 1965.
[77] Michlin, S.G.: Lehrgang der Mathematischen Physik. 2nd edition, Akademie-Verlag,
Berlin, 1975.
[78] Möhringer, S.: Decorrelation of Gravimetric Data. Ph.D.-thesis, Geomathematics
[79] Moritz, H.: Advanced Physical Geodesy. Herbert Wichmann Verlag, Karlsruhe, Aba-
cus Press, Tunbridge, 1980.
[80] Moritz, H.: The Figure of the Earth. Theoretical Geodesy of the Earth’s Interior.
Wichmann Verlag, Karlsruhe, 1990.
[81] Moritz, H.: Classical Physical Geodesy. In: W. Freeden, Z. Nashed, and T. Sonar,
Eds., Handbook of Geomathematics. Springer, Heidelberg, 253–290, 2015.
[82] Morozov, V.A.: Methods for Solving Incorrectly Posed Problems. Springer, New
York, 1984.
ed., Academic, New York, 311–359, 1971.
Generalized Inverses and Applications. Academic Press, New York, 193–244, 1976.
[85] Nashed, M.Z.: New Applications of Generalized Inverses in System and Control
Theory. In: Proc. 1980 Conf. on Information Sciences and Systems, J.B. Thomas.
ed. Princeton. NJ, Princeton, 353–358, 1980
Problems with Applications to Antenna Theory. IEEE Trans. Antennas Propa-
gation, 29:220–231, 1981.
[87] Nashed, M.Z.: A New Approach to Classification and Regularization of Ill-Posed
Operator Equations. In: H. Engl and C.W. Groetsch, Eds., Inverse and Ill-Posed
Problems, Band 4, Notes and Reports in Mathematics and Science and Engineering,
Academic Press, Boston, 1987.
[88] Nashed, M.Z.: Inverse Problems, Moment Problems and Signal Processing: Un
Menage a Trois, Mathematics in Science and Technology. Siddiqi, A.H., Singh,
R.C. Manchanda, P. (Eds.), World Scientific, New Jersey, 1–19, 2010.
[89] Nashed, M.Z., Votruba, F.G.: A Unified Operator Theory of Generalized Inverses.
In: Generalized Inverses and Applications. M.Z. Nashed ed., Academic Press, New
York, 1–109, 1976.
[90] Parker, R.L.: The Theory of Ideal Bodies for Gravity Interpretation. Geophys. J.
R. Astr. Soc., 42:315–334, 1975.
130:233–235, 1900.
[92] Rieder, A.: Keine Probleme mit Inversen Problemen. Vieweg 2003.
[93] Rummel, R.: Geodesy, In: Encyclopedia of Earth System Science, Volume 2, 253–
262, Academic Press, 1992.
[94] Rummel, R.: GOCE: Gravitational Gradiometry in a Satellite. In: Handbook of
Geomathematics, Vol. 1. (Freeden, W., Nashed, M.Z., Sonar,T., (Eds.)), 98–105,
Springer, Heidelberg, 2010.
[95] Rummel R., Balmino, G., Johannessen, J., Visser, P., Woodworth P.: Dedicated
Gravity Field Missions – Principles and Aims. J. Geodyn., 33:3–20, 2002.
[96] Rummel, R., van Gelderen, M.: Meissl Scheme – Spectral Characteristics of Physical
Geodesy. Manuscr. Geod. 20:379–385, 1995.
[97] Saitoh, S.: Theory of Reproducing Kernels and its Applications. Longman, New
York, 1988.
[98] Sansò, F., Internal Collocation. Atti Della Accademia Nazionale Dei Lincei, Serie
VIII, Volume XVI, 2–52, 1980.
[99] Sansò, F., Rummel, R. (Eds.): Geodetic Boundary Value Problems in View of the
One Centimeter Geoid. Lecture Notes in Earth Sciences, Vol. 65, Springer, Berlin,
Heidelberg, 1997.
[100] Sansò, F., Tscherning, C.C.: The Inverse Gravimetric Problem in Gravity Mod-
elling. In: Festschrift to Torben Krarup (Kejlsø, E., Poder, K., Tscherning, C.C.,
Eds.), 299–334, Geodaetisk Institute, Copenhagen, 1989.
[101] Sansò, F., Barzaghi, R., Tscherning, C.C.: Choice of Norm for the Density Distri-
bution of the Earth. Geophys. J. R. Astr. Soc. 87:123–141, 1986.
Phys. Earth Planet. Inter. 28: 215–229, 1982.
[103] Skorvanek, M.: The Inverse Gravimetric Problem for the Earth. In: Proceedings
of the 4th International Symposium on Geodesy and Physics of the Earth 1980,
464–475, Veröff. Zentralinst. Physik der Erde, Vol. 63, 1981.
[104] Tikhonov, A.N.: On the Stability of Inverse Problems. Dokl. Akad. Nauk SSSR,
39:195–198, 1943.
ularization Method, Dokl. Akad Nauk SSSR, 151:501–504, 1963.
[106] Tscherning, C.C.: Analytical and Discrete Inversion Applied to Gravity Data. In:
Proceedings of the Interdisciplinary Inversion Workshop 1, Methodology and Appli-
cation Perspectives in Geophysics, Astronomy and Geodesy. (Holm Jacobsen, B.,
(Ed.)), 5–8, Aarhus, 1992.
[107] Tscherning, C.C., Strykowski, G.: Quasi-Harmonic Inversion of Gravity Field Data,
Model Optimization in Exploration Geophysics 2. In: Proceedings of the 5th In-
ternational Mathematical Geophysics Seminar (Vogel, A., Ed.), 137–154, Vieweg,
Braunschweig, Wiesbaden, 1987.
[108] Vogel, C.R.: Computational Methods for Inverse Problems. SIAM, Philadelphia,
2002.
[109] Weck, N.: Zwei inverse Probleme in der Potentialtheorie. In: Mitt. Inst. Theor.
Geodäsie, Universität Bonn, 4:27–36, 1972.
[110] Werner, J.: Optimization Theory and Applications. Vieweg-Verlag, Braunschweig,
Wiesbaden 1984.
[111] Xia X.G., Nashed M.Z.: The Backus–Gilbert Method for Signals in Reproducing
Hilbert Spaces and Wavelet Subspaces. Inverse Problems, 10:785–804, 1994.
[112] Yosida, K.: Functional Analysis. 5th ed., Springer, Berlin, 1965.
[113] Zidarov, D.P.: Some Uniqueness Conditions for the Solution of the Inverse Gravi-
metric Problem. Comptes rendus de l’Académie bulgare des Sciences, 33:909–912,
1980.
[114] Zidarov, D.P.: Conditions for Uniqueness of Self-Limiting Solutions of the Inverse
Problems. Comptes rendus de l’Académie bulgare des Sciences, 39:57–60, 1986.
[115] Zidarov, D.P.: Inverse Gravimetric Problem in Geoprospecting and Geodesy. Devel-
opments in Solid Earth Geophysics, 19, Elsevier, Amsterdam, 1990.
Willi Freeden
M. Zuhair Nashed
Mathematics Department
University of Central Florida
Orlando, USA

Gravimetry and Exploration

Abstract. In this work we are especially concerned with the “mathemati-

zation” of gravimetric exploration and prospecting. We investigate the ex-
tractable information of the Earth’s gravitational potential and its observ-
ables obtained by gravimetry for gravitational modeling as well as geological
interpretation. More explicitly, local gravimetric data sets are exploited to
visualize multiscale reconstruction and decorrelation features to be found in
geophysically and geologically relevant signature bands.
Keywords. Absolute and relative gravimetry, Newtonian gravitational ap-

proach, multiscale density field modeling.
Introduction
Newton’s famous law about the mutual attraction of two masses formulated in
“De mundi systemate” 1715 tells us that the attractive force, called gravitation,
is directed along the line connecting the two centers of mass of the objects and is
proportional to two masses as well as to the squared inverse of the distance between
the objects. If the Earth had a perfectly spherical shape and if the mass inside the
Earth were distributed homogeneously or rotationally symmetric, the line along
which an apple fell would indeed be a straight line, directed radially and going
exactly through the Earth’s center of mass. The gravitational field obtained in this
way would be perfectly spherically symmetric. In reality, however, the situation is
more complex. The topographic features, mountains and valleys, are very irregular.
The actual gravitational field is influenced by strong irregularities in density within
the Earth. As a result, the gravitational force deviates from one place to the other
from that of a homogeneous sphere. More explicitly, internal density signatures
are reflected in gravitational field signatures, and orthogonal coefficients in terms
of spherical harmonics of gravitational field signatures smooth out exponentially
with increasing distance from the Earth’s body. As a consequence, positioning
systems are ideally located as far as possible from the Earth, whereas gravity field
sensors are ideally located as close as possible to the Earth. Following these basic
688 C. Blick, W. Freeden, and H. Nutz
principles, various positioning and gravity field determination techniques have been
designed. Sensors may be sensitive to local or global features of the gravity field.
Considering the spatial location of the data points, we may differentiate between
terrestrial (surface), airborne, and spaceborne methods.
Concerning gravity on a global scale, e.g., for global geoid determination
(that will not be investigated here), it should be pointed out (see, e.g., [6–8]) that
the terrestrial distribution of Earth’s gravity data is far from being homogeneous
with large gaps, in particular over oceans but also over land. In addition, the qual-
ity of the data is very distinct. Thus, global terrestrial gravity data coverage now
and in the foreseeable future is far from being satisfactory. This is the reason why
spaceborne measurements come into play for global gravity determination. Until
now, the relatively poor precision of satellite-only spaceborne gravity measure-
ments has hindered a wider use of this type of measurements for local purposes.
Seen from future exploration aspects, however, it must be remarked that only
coordinated research will provide a breakthrough in modeling and understanding
significant structures and processes in the Earth’s interior. In fact, the authors are
convinced that the way forward, even in global modeling, has to be based on two
requirements:
i) combining data from different sensors and sources,
ii) multiscale modeling, i.e., “zooming-in downward continuation” of the differ-
ent data sources starting from globally available spaceborne data as means
for an appropriate trend solution via more accurate (regional) airborne data
down to (local) high-precision gravimetric data sets.
In this contribution, we are especially concerned with the mathematical study
of gravimetry in exploration, in particular postprocessing of all already available
models. To this end we briefly explain the status quo of gravimetric observation
and standard modeling. On the basis of these results we present new multiscale
methods by means of geoscientifically relevant wavelets for the decorrelation of
signatures inherent in geological information.
1. Gravity, gravitation, and gravimetry

The force of gravity provides a directional structure to the space above the Earth’s
surface. It is tangential to the vertical plumb lines and perpendicular to all (level)
equipotential surfaces. Any water surface at rest is part of a level surface. (Level)
equipotential surfaces are ideal reference surfaces, for example, for heights. The
geoid is defined as that level surface of the gravity field which best fits the mean
sea level.
The direction of the gravity vector can be obtained by astronomical posi-
tioning. Measurements are possible on the Earth’s surface. Observations of the
gravity vector are converted into so-called vertical deflections by subtracting a
corresponding reference direction derived from a simple gravity field model as-
sociated to, e.g., an ellipsoidal surface. Vertical deflections are tangential fields
Gravimetry and Exploration 689
of the anomalous potential. Due to the high measurement effort required to ac-
quire these types of data compared to a gravity measurement, the data density of
vertical deflections is much less than that of gravity anomalies. Gravitational field
determination based on the observation of deflections of the vertical and combined
with gravity is feasible in smaller areas with good data coverage.
1.1. Gravitational, centrifugal, and gravity acceleration
The gravity acceleration (gravity) w is the resultant of the gravitation v and the
centrifugal acceleration c such that
w = v + c. (1.1)
The centrifugal force c arises as a result of the rotation of the Earth about
its axis. In this work concerned with local gravity exploration we are allowed to
assume a rotation of constant angular velocity ω. The centrifugal acceleration
acting on a unit mass is directed outward perpendicularly to the spin axis (see
Figure 1.1). Introducing the so-called centrifugal potential C, such that c = ∇C,
the function C turns out to be non-harmonic. The direction of the gravity w is
known as the direction of the plumb line, the quantity |w| is called the gravity
intensity (often also just called gravity and denoted in the geodetic jargon by g).
Altogether, the gravity potential of the Earth can be expressed in the form
W = V + C, (1.2)
and the gravity acceleration w is given by
w = ∇W = ∇V + ∇C. (1.3)
fktgevkqp qh rnwod nkpg
y
d
w
x
egpvgt qh ocuu
Figure 1.1. Gravitation v, centrifugal acceleration c,

gravity acceleration w.
As already pointed out, the surfaces of constant gravity potentials, i.e., W =

const., are designated as equipotential (level, or geopotential) surfaces of gravity
(for more details, the reader is referred to monographs in physical geodesy, e.g.,
[27, 32, 51, 69]).
The actual Earth’s surface (globally available from modern spaceborne tech-
niques such as DOPPLER, GNSS, LASER, VLBI, etc.) does not coincide with an
equipotential surface (i.e., a level surface). The force of gravity is generally not
perpendicular to the actual Earth’s surface (see Figure 1.2). However, we are con-
fronted with the gravity intensity as an oblique derivative on the Earth’s surface.
The gravity vector is an oblique vector at any point on the Earth’s surface and
generally not the normal vector.
w
Figure 1.2. Earth’s surface, geoid, ellipsoid (λ = |w| oblique unit grav-
ity vector, normal vector to the geoid, but usually not normal to the
Earth’s surface).
The determination of equipotential surfaces of the potential W is strongly

related to the knowledge of the potential V . The gravity vector w given by w =
∇W is normal to the equipotential surface passing through the same point. Thus,
equipotential surfaces such as the geoid intuitively express the notion of tangential
surfaces, as they are normal to the plumb lines given by the direction of the gravity
vector.
1.2. Gravimeter and gravimetry
Essentially, the Earth is a “spheroid”, with a slight flattening (0.35%) at the poles,
a mean radius of 6368km, and a mean mass of 5.98 · 1024 kg. At the surface of the
Earth, its mean value of gravity intensity is given by 9.80 sm2 . At the equator, it re-
duces to 9.78 sm2 ; at the poles, it increases to about 9.83 sm2 , reflecting the flattening.
Gravimeters are typically designed to measure very tiny fractional changes of
the Earth’s gravity, caused by nearby geologic structures or the shape of the Earth.
There are two types of gravimeters, viz. relative and absolute gravimeters. Absolute
gravimeters measure the local gravity and are directly based on measuring the
acceleration of free fall (for example, of a test mass in a vacuum tube). Relative
gravimeters compare the value of gravity at one point with another. They must
be calibrated at a location, where the gravity is known accurately and measure
the ratio of the gravity at the two points. Most common relative gravimeters are
spring-based. By determining the amount by which the weight stretches the spring,
gravity becomes available via Hooke’s law (see Figure 1.3). The highest possible
accuracy of relative gravity measurements are conducted at the Earth’s surface.
Measurements on ships and in aircrafts deliver reasonably good data only after the
removal of inertial noise. In addition, when interested in gravimetric exploration,
Figure 1.3. The principle of gravimetry (with kind permission of

Teubner-publishing taken from [38] in modified form).
it should be noted that a high measurement accuracy of at least 0.1mGals, but

more adequately 0.01mGals (1 mGal = 10−5 sm2 , cf. Table 1) has to be achieved (cf.
Figure 1.4).
SI Units Traditional Units

−2 −2
10 ms 1 Gal
10−5 ms−2 1 mGal
10−8 ms−2 1 μGal
Table 1. Traditional units for gravimetric measurements and their SI
unit complement.
By gravimetry, we denote the determination of the Earth’s mass density dis-
tribution from data of the gravitational potential or related quantities. Clearly,
for purposes of exploration, it is obvious that the determination of gravity inten-
sities as well as gravity anomalies of dimension very much larger than the gravity
anomalies caused by, e.g., aquifers, oil and gas structures are of less significance.
The fundamental interest in gravimetric methods in exploration is based on the
measurements of small variations.
Gravity prospecting has been first used in the case of strong density con-
trasts in a geological structure, and the usual approach is to measure differences
in gravity from place to place. Today, the interpretation of gravimetric data is done
by comparing the shape and size of gravity disturbances and anomalies to those
Figure 1.4. Illustration of the components of the gravity acceleration

(ESA medialab, ESA communication production SP–1314).
caused by bodies of various geometrical shapes at different depths and differing

densities.
The observed gravity depends on the following effects to be removed (for
more detailed studies, see, e.g., [53, 59, 64]): attraction of the reference ellipsoid,
elevation above sea level, topography, time dependent variations (tidal), (Eőtvős)
effect of a moving platform, isostatic balance on the lower lithosphere, density
variations inside the upper crust. To isolate the effects of local density variations
from all other contributions, it is necessary to apply a series of reductions: The
attraction of, e.g., the reference ellipsoid or another reference surface has to be
subtracted from the measured values. An elevation correction must be done, i.e.,
the vertical gradient of gravity is multiplied by the elevation of the station and
the result is added. With increasing elevation of the Earth, there is usually an
additional mass between the reference level and the actual level. This additional
mass itself exerts a positive gravitational attraction. Bouguer correction and ter-
rain correction are applied to correct for the attraction of the slab of material
between the observation point and the geoid. A terrain correction accounts for
the effect of nearby masses above or mass deficiencies below the station. Isostatic
correction accounts for the isostatic roots (Moho). Other corrections have to be
applied to the data in order to account for effects not related to the subsurface:
Drift corrections are necessary, since each gravimeter suffers mechanical changes
over time, and so does its output measurement. This change is generally assumed
to be linear. Tidal corrections have to be imposed, i.e., the attraction of the Sun
and Moon has to be calculated and subtracted from the measurements. In case of
acquisition on a moving platform, the motion relative to the surface of the Earth
implies a change in centrifugal acceleration. The Eőtvős correction depends on the
latitude and velocity vector of the moving platform. It should be observed that
free air anomaly does not correct for the first two effects which could mask the
gravity anomalies related to the Bouguer density contrasts in the crust. Complete
Bouguer correction effectively remove the gravity anomalies due to bathymetry,
but still contain the gravity effect of the Moho. Isostatics contain the gravity effect
of the Moho. For more details the reader is referred to geodetic textbooks such as
[32, 36, 68] and to the literature concerned with prospecting and exploration (see,
e.g., [53, 54, 64] and the references therein).
Gravity prospecting can be done over land or sea areas using different tech-
niques and equipment. Terrestrial gravimetry was first applied to prospect for salt
domes (e.g., in the Gulf of Mexico) (an example of the Eastern part of Germany
is shown in Figure 1.5), and later for looking for anticlines in continental areas
(see, e.g., [53, 54], and the references therein). Nowadays, gravimetry is in use all
over the world in diverse applications:
(1) Gravimetric surveys serve regional geological mapping.

(2) Gravimetry is helpful in different phases of the oil exploration and production
processes.
(3) Gravimetric surveys are employed in mineral exploration, for example, to
detect mineral deposits (see Figure 1.5) of economic interest (such as metals,
salt, coal).
(4) Archaeological and geotechnical studies aim at the mapping of subsurface
voids and overburden variations.
(5) Gravimetric campaigns may be applied for groundwater and environmental
studies. They help to map aquifers to provide formations and/or structural
control.
(6) Gravimetric studies give information about tectonically derived changes and
volcanological phenomena.
(7) Gravimetric studies provide useful information on changes in the level of
water in geothermal reservoirs and therefore on the longevity of a geothermal
resource.
It is surprising that the use of gravimetry is in infancy in the German geother-

mal scene, although it has much to offer. Due to (regional) airborne and (global)
spaceborne gravity information such as satellite-to-satellite tracking (SST) and/or
satellite gravity gradiometry (SGG), new promising components in gravimetri-
cally oriented modeling can be expected in the future, for example, based on
multiscale modeling providing reconstruction and decomposition of geological sig-
natures, where seismic modeling is difficult or impossible because of anthropogenic
activities, e.g., in mining areas.
Figure 1.5. Top: Gravity effect in [μm · s−1 ] of the salt dome Werle
(Mecklenburg, Germany); bottom: Geological vertical profile (with kind
permission of Teubner-publishing taken from [38] in modified form).
All in all, nowadays the main applications of gravimetry can be listed as

follows:
(i)definition of geological structural settings,
(ii)faults delineation,
(iii)recovery of salt bodies, metal deposits,
(iii)detection of heap of coal, ore, etc.,
(iv) 2D/3D forward modeling, inversion, and postprocessing to assist seismic mod-
eling,
(v) combination with geomagnetic interpretation,
(vi) 4D monitoring, etc.
Figure 1.6 shows the gravity as well as the vertical/horizontal gradient curves
induced by a simple geological structure in sectional illustration (for similar illus-
trations, the reader is referred to, e.g., [5, 26, 39, 48, 53, 54, 64]). It is remarkable
that the vertical/horizontal gradient curves show significant interactions on density
variations.
gravity
vertical gradient
horizontal gradient
density difference
Figure 1.6. Schematic diagram of the horizontal/vertical gradients vs.

the gravity potential (cf. [54]).
The knowledge of horizontal/vertical derivatives of the gravity potential is

therefore a useful addendum to prospecting and exploration. This is the reason
why we are interested in discussing these derivatives in more detail (based on ideas
and concepts developed in [15]).
2. Surface horizontal/vertical derivatives of the gravity potential

The lines that intersect all equipotential surfaces orthogonally are not exactly
straight but slightly curved (cf. Figure 2.1). They are called lines of gravity force
or plumb lines. The gravity vector at any point is tangential to the plumb line.
Hence, “direction of the gravity vector”, “vertical” and “direction of the plumb
line” are synonymous. As the equipotential surfaces are, so to speak, “horizontal”,
i.e., orthogonal to the plumb lines, they play an important part in our daily life
(e.g., in civil engineering for the purpose of height determination). Equipotential
surfaces of the Earth’s gravity potential W allow, in general, no simple mathe-
matical representation. This is the reason why physical geodesy and geophysics
choose a suitable reference surface for modeling the geoid, i.e., the equipotential
surface at sea level. The reference surface is constructed as an equipotential sur-
face of an artificial normal gravity potential U . Its gradient field, i.e., u = ∇U , is
called normal gravity. For reasons of simplicity, physical geodesy usually uses an
ellipsoid of revolution in such a way that a good adaption to the Earth’s surface
is guaranteed. Closed representations of normal gravity potentials, in considera-
tion of the centrifugal force, can be found extensively in the geodetic literature
(cf. [25, 27, 32, 36, 45, 51, 65]), and the references therein). The deviations of the
gravity field of the Earth from the normal field of such an ellipsoid are small. The
remaining parts of the gravity field are gathered in a so-called disturbing gravity
field ∇T corresponding to the disturbing potential T = W − U .
2.1. Gravity anomalies, gravity disturbances, and vertical deflections

Knowing the gravity potential, all equipotential surfaces (including the geoid) are
given by an equation of the form W (x) = const. By introducing U as the normal
gravity potential corresponding to the ellipsoidal field, the disturbing potential T
is the difference of the gravity potential W and the normal gravity potential U ,
i.e., we are led to a decomposition of the gravity potential in the form W = U + T.
According to the concept developed by Stokes [65], Helmert [33], and Pizzetti
[56, 57] we may assume that
(a) the center of the ellipsoid coincides with the center of gravity of the Earth,
(b) the difference of the mass of the Earth and the mass of the reference body
(ellipsoid) is zero.
>ĞǀĞůƐƵƌĨĂĐĞ >ĞǀĞůƐƵƌĨĂĐĞ
ǆ
ǁ;ǆͿ ǆ
ǁ;ǆͿ
ƐƉŚĞƌĞ
ŐĞŽŝĚ
WůƵŵďůŝŶĞ
WůƵŵďůŝŶĞ
Figure 2.1. Level surfaces and plumb lines for a homogeneous ball
(left) and an Earth-like body (right) (from [15]).
A point x of the geoid can be projected onto its associated point y of the
ellipsoid by means of the ellipsoidal normal. The distance N (x) between x and y
is called the geoidal height or geoidal undulation in x (cf. Figure 2.2). The gravity
anomaly vector a(x) at the point x of the geoid is defined as the difference between
the gravity vector w(x) and the normal gravity vector u(y), i.e.,
a(x) = w(x) − u(y). (2.1)
Another possibility is to form the difference between the vectors w and u at
the same point x such that we get the gravity disturbance vector d(x) defined by
d(x) = w(x) − u(x). (2.2)
In geodesy, several basic mathematical relations between the scalar fields |w| and
|u| as well as between the vector fields a and d are known. In the following, we
only describe the fundamental relations heuristically (see also [27, 32]).
The point of departure for our excursion into geodesy is the observation that
the gravity disturbance vector d(x) at the point x on the geoid can be written as
Figure 2.2. Illustration of the gravity vector w(x), the normal gravity
vector u(x), and the geoidal height N (x). Here, ν and ν
denote the
normal to the geoid and the reference ellipsoid, respectively (follow-
ing [32]).
follows:
d(x) = w(x) − u(x) = ∇ (W (x) − U (x)) = ∇T (x). (2.3)
∂U
According to Taylor’s formula, U (y)+ ∂ν (y)N (x)
is the linearization of U (x),
i.e., by expanding the potential U at the point x and truncating the Taylor series
at the linear term, we get
∂U
U (x) & U (y) + (y)N (x), (2.4)
∂ν
where
u(y)
ν
(y) = − (2.5)
|u(y)|
is the ellipsoidal normal at y and the geoidal undulation N (x) is the aforementioned
distance between x and y (note that the symbol ‘&’ means that the error between
the left- and the right-hand side may be assumed to be insignificantly small). Using
the fact that T (x) = W (x) − U (x) and observing the relations
∂U
|u(y)| = −ν
(y) · u(y) = −ν
(y) · ∇U (y) = − (y), (2.6)
∂ν
we obtain under the assumption of (2.4) that

U (y) − U (x) T (x) − (W (x) − U (y))
N (x) = = . (2.7)
|u(y)| |u(y)|
Finally, considering U (y) = W (x) = const. = W0 , we end up with the so-called
Bruns formula (cf. [4])
T (x)
N (x) = . (2.8)
|u(y)|
This formula relates the physical quantity T (x) to the geometric quantity N (x) for
points x on the geoid.
It is helpful to study the vector field ν(x) in more detail:

w(x)
ν(x) = − . (2.9)
|w(x)|
Due to the definition of the normal vector field (2.9), we obtain the following
identity
w(x) = ∇W (x) = − |w(x)| ν(x). (2.10)
In an analogous way we obtain
u(x) = ∇U (x) = − |u(x)| ν
(x). (2.11)
The vertical deflection Θ(x) at the point x on the geoid is understood to be the
angular (i.e., tangential) difference between the directions ν(x) and ν
(x). More
concretely, the vertical deflection is determined by the angle between the plumb
line and the ellipsoidal normal through the same point:
Θ(x) = ν(x) − ν
(x) − ((ν(x) − ν
(x)) · ν(x)) ν(x). (2.12)
According to its construction, the vertical deflection Θ(x) at x is orthogonal to the
normal vector field ν(x), i.e., Θ(x) · ν(x) = 0. Since the plumb lines are orthogonal
to the equipotential surfaces of the geoid and the ellipsoid, respectively, the ver-
tical deflection gives briefly spoken a measure of the gradient of the equipotential
surfaces (cf. [32]). From (2.10), in connection with (2.12), it follows that
w(x) = −|w(x)| (Θ(x) + ν
(x) + ((ν(x) − ν
(x)) · ν(x)) ν(x)) . (2.13)
Using Eqs. (2.11) and (2.13) we finally obtain for the gravity disturbing vector
d(x) at the point x
d(x) = ∇T (x) = w(x) − u(x) (2.14)

= −|w(x)| (Θ(x) + ν (x) + ((ν(x) − ν (x)) · ν(x)) ν(x)) − (−|u(x)|ν (x))

= −|w(x)| (Θ(x) + ((ν(x) − ν
(x)) · ν(x)) ν(x)) − (|w(x)| − |u(x)|) ν
(x).
The quantity
D(x) = |w(x)| − |u(x)| (2.15)
is called the gravity disturbance, whereas
A(x) = |w(x)| − |u(y)| (2.16)
is called the gravity anomaly.
Splitting the gradient ∇T (x) of the disturbing potential T at x into a normal
part (pointing into the direction of ν(x)) and an angular (tangential) part (using
the representation of the surface gradient ∇∗ ), we have
∂T 1 ∗
∇T (x) = ν(x) (x) + ∇ T (x), (2.17)
∂ν |x|
where ∇∗ is the surface gradient.
Since the gravity disturbances represent at most a factor 10−4 of the Earth’s
gravitational force (for more details see [32]), the error between ν(x) ∂T
∂ν (x) and

∂T
ν (x) ∂ν (x) has no (computational) significance. Consequently, we may assume
∂T 1 ∗
d(x) & ν
(x) (x) + ∇ T (x). (2.18)
∂ν
|x|
Moreover, the scalar product (ν(x) − ν
(x)) · ν(x) can also be neglected. Thus, in
connection with (2.14), we obtain
d(x) & −|w(x)| Θ(x) − D(x)ν
(x). (2.19)
By comparison of (2.18) and (2.19), we therefore get
∂T
D(x) = − (x) = −ν
(x) · d(x), (2.20)
∂ν
1
|w(x)| Θ(x) = − ∇∗ T (x). (2.21)
|x|
In other words, the gravity disturbance D(x), beside being the difference in mag-
nitude of the actual and the normal gravity vector, is also the normal component
of the gravity disturbance vector d(x). In addition, we are led to the angular
differential equation (2.21).
Applying Bruns’ formula (2.8) to Eqs. (2.20) and (2.21) we obtain

∂N
D(x) = |w(x)| − |u(x)| = −|u(y)| (x) (2.22)
∂ν
for the gravity disturbance and

1 1
|w(x)| Θ(x) = − ∇∗ T (x) = − |u(y)| ∇∗ N (x) (2.23)
|x| |x|
for the vertical deflections. Note that Θ(x) may be multiplied (without loss of
(computational) precision) either by |w(x)| or by |u(x)| since it is a small quantity.
Turning over to the gravity anomalies A(x), it follows from the identity (2.20)
by linearization that
∂T ∂|u(y)|
− (x) = D(x) & A(x) − N (x). (2.24)
∂ν
∂ν
Using Bruns’ formula (2.8), we obtain for the gravity anomalies that
∂T 1 ∂|u(y)|
A(x) = − (x) + T (x). (2.25)
∂ν
|u(y)| ∂ν
Summing up our results (2.20) for the gravity disturbance D(x) and (2.25) for
the gravity anomaly A(x), we are led to the so-called fundamental equations of
physical geodesy:
∂T
D(x) = |w(x)| − |u(x)| = − (x), (2.26)
∂ν
∂T 1 ∂|u(y)|
A(x) = |w(x)| − |u(y)| = −
(x) + T (x). (2.27)
∂ν |u(y)| ∂ν
Eqs. (2.26) and (2.27) show the relation between the disturbing potential T and the
gravity disturbance D and the gravity anomaly A, respectively, on the geoid (see,
for example, [27, 32, 45]). They are used as boundary conditions in boundary-value
problems.
Remark 2.1. Following [32], the geoidal heights N , i.e., the deviations of the
equipotential surface on the mean ocean level from the reference ellipsoid, are
extremely small. Their order is of only a factor 10−5 of the Earth’s radius (see
[32] for more details). Even more, the reference ellipsoid only differs from a sphere
ΩR with (mean Earth’s) radius R in the order of the flattening of about 3 · 10−3 .
Therefore, since the time of [65], it is common use that, in theory, an ellipsoidal
reference surface should be taken into account. However, in numerical practice,
the reference ellipsoid is treated as a sphere and the Equations (2.22) and (2.23)
are solved in spherical approximation. In doing so, a relative error of the order
of the flattening of the Earth’s body at the poles, i.e., a relative error of 10−3 ,
is accepted in all equations containing the disturbing potential. Considering ap-
propriately performed reductions in numerical calculations, this error seems to
be quite permissible (cf. [32] and the remarks in [24, 25] for comparison with el-
lipsoidal approaches), and this is certainly the case if local exploration is under
consideration. For local purposes as discussed in this contribution, the problem of
non-ellipticity seems to be obsolete.
Remark 2.2. According to the Pizzetti assumptions (see [56, 57]), it follows that
the first moment integrals of the disturbing potential vanish, i.e.,

R
T (y)H−n−1,k (y) dω(y) = 0, (2.28)
ΩR
R
for n = 0, 1, k = 1, . . . , 2n + 1, where H−n−1,k denotes the system of outer
harmonics and dω is the surface element in R3 . More concretely, if the Earth’s
center of gravity is the origin, there are no first-degree terms in the spherical
harmonic expansion of T . If the mass of the spherical Earth and the mass of
the normal ellipsoid is equal, there is no zero term. In this way, together with the
indicated processes in gravitational modeling, formulas and structures are obtained
that are rigorously valid for the sphere.
In the well-known spherical nomenclature, involving a sphere ΩR as reference

surface (R being the mean Earth’s radius) with a mass M distributed homoge-
neously in its interior, we are simply led to (cf. [32])

γM γM y
U (y) = , u(y) = ∇U (y) = − , (2.29)
|y| |y|2 |y|
where γ is the gravitational constant (γ = 6.6742 · 10−11 m3 kg−1 s−2 ). Hence, we
obtain
γM
|u(y)| = , (2.30)
|y|2
∂|u(y)| u(y) γM
=− · ∇|u(y)| = −2 3 , (2.31)
∂ν
|u(y)| |y|
1 ∂|u(y)| 2
=− , (2.32)
|u(y)| ∂ν
|y|
where x is on the geoid and y is on the reference surface (cf. Fig. 2.2). Furthermore,
in spherical nomenclature, i.e., x ∈ ΩR , we obviously have
∂T x
−
(x) = − · ∇T (x). (2.33)
∂ν |x|
Therefore, we end up with the formulation of the fundamental equations of physical
geodesy for the sphere:
x
D(x) = − · ∇T (x), (2.34)
|x|
x 2
A(x) = − · ∇T (x) − T (x). (2.35)
|x| |x|
In addition, in a vector spherical context (see also [22]), we obtain for the
differential equation (2.21)
γM
−∇∗ T (x) = Θ(x), (2.36)
R
and, by virtue of Bruns’ formula (2.8), we finally find that
−∇∗ N (x) = RΘ(x). (2.37)
Remark 2.3. In physical geodesy (see, e.g., [32, 36]), a componentwise scalar deter-
mination of the vertical deflection is usually used. Our work prefers the vectorial
framework, i.e., the vector equation (2.36). In doing so, we are concerned with an
isotropic vector approach by means of the fundamental solution with respect to
the Beltrami operator (see also [22]) instead of the conventional anisotropic scalar
decomposition into vector components due to [44].
The disturbing potential enables us to make the following geophysical in-
terpretations (for more details the reader is referred, e.g., [25, 41, 60, 63], and
the references therein): Gravity disturbances D and gravity anomalies A (Figure
2.3) represent a relation between the real Earth and an ellipsoidal Earth model.
In accordance with Newton’s Law of Gravitation they therefore show the imbal-
ance of forces in the interior of the Earth due to the irregular density distribution
inside the Earth. Clearly, gravity anomalies and/or gravity disturbances do not
determine uniquely the interior density distribution of the Earth. They may be
interpreted as certain filtered signatures, which give major weight to the density
contrasts close to the surface and simultaneously suppress the influence of deeper
structures inside the Earth.
Geoid undulations provide a measure for the perturbations of the Earth from
a hydrostatic equilibrium. They form the deviations of the equipotential surfaces
at mean sea level from the reference ellipsoid. Geoid undulations show no essential
correlation to the distributions of the continents. They seem to be generated by
density contrasts much deeper inside the Earth.
As already explained, the task of determining the disturbing potential T from
gravity disturbances or gravity anomalies, respectively, leads to boundary-value
problems usually corresponding to a spherical boundary. Numerical realizations
of such boundary-value problems have a long tradition, starting from [65] and
[55]. Nonetheless, our work presents some new aspects in their potential theoretic
treatment by proposing appropriate space-regularization techniques applied to the
resulting integral representations of their solutions. For both boundary-value prob-
lems, viz. the Neumann and the Stokes problem, we are able to present two solution
methods: The disturbing potential may be either solved by a Fourier (orthogonal)
expansion method in terms of spherical harmonics or it can be described by a
singular integral representation over the boundary ΩR .
Remark 2.4. So far, much more data on gravity anomalies A(x) = |w(x)| − |u(y)|
are available than on gravity disturbances D(x) = |w(x)| − |u(x)|. However, by
modern GNSS-technology (see, e.g., [52]), the point x on the geoid is rather deter-
mined than y on the reference ellipsoid. Therefore, in future, it can be expected
that D will become more important than A (as [36] point out in their monograph
on physical geodesy). This is the reason why we continue to work with D. Never-
theless, the results of our (multiscale) approach applied to A are of significance.
Therefore, the key ideas and concepts concerning A can be treated in parallel (see
[9, 15, 73] for explicit details).
In order to formulate some results in 3the language

of potential theory, we
(1)
first introduce the potential space Pot R \BR (0) , where BR (0) is the (open)
ball of radius R around the origin. More concretely, we let Pot(R3 \BR (0)) be the
space of all functions F : R3 \BR (0) → R satisfying
(i) F |R3 \BR (0) is a member of C (2) (R3 \BR (0)),
(ii) F satisfies Laplace’s equation ΔF (x) = 0, x ∈ R3 \BR (0).
(iii) F is regular at infinity, i.e., F (x) = O(|x|−1 ), |x| → ∞.
Pot(1) (R3 \BR (0)) is formally understood to be the space

Pot(1) R3 \BR (0) = C (1) R3 \BR (0) ∩ Pot(R3 \BR (0)). (2.38)
Figure 2.3. EIGEN-GL04C derived gravity anomalies and geoidal un-

dulations (reconstructed by use of (spherical) smoothed Haar scaling
functions from [22] based on computations by Mathar [43]).
In the language of potential theory, the exterior Neumann boundary-value problem

corresponding to known gravity disturbances D (compare (2.34)) reads as follows:
(ENPPG) Let D be a continuous function on ΩR = ∂BR (0), i.e., D ∈ C (0) (ΩR )

with

R
D(y)H−n−1,k (y) dω(y) = 0, (2.39)
ΩR

for n = 0, 1, k = 1, . . ., 2n + 1. Find T ∈ Pot(1) R3 \BR (0) , such that the bound-
ary condition D = ∂T
∂ν ΩR holds true and the potential T fulfills the conditions

R
T (y)H−n−1,k (y) dω(y) = 0 (2.40)
ΩR
for n = 0, 1, k = 1, . . . , 2n + 1.
It is known (see, e.g., [15]) that the solution of the boundary-value problem
(ENPPG) can be represented in the form

1
T (x) = D(y) N (x, y) dω(y), x ∈ R3 \BR (0), (2.41)
4πR ΩR
where the Neumann kernel N (·, ·) in (2.41) possesses the spherical harmonic ex-
pansion
∞ n+1
R2 2n + 1 x y
N (x, y) = Pn · . (2.42)
n=2
|x||y| n+1 |x| |y|
By well-known manipulations, the series in terms of Legendre polynomials can be
expressed as an elementary function leading to the integral representation
⎛ ⎛ ⎞⎞
R2 R2
1 2R |y| + y − |x|2 − |x|
x
T (x) = D(y) ⎝ + ln ⎝

⎠⎠ dω(y). (2.43)
4πR ΩR |x − y| R2 R2
|y| + y − x +
|x|2 |x|
It is not difficult to see that for x ∈ ΩR , the integral (2.43) is equivalent to

1 2R |y| + |x − y| − R
T (x) = D(y) + ln dω(y). (2.44)
4πR ΩR |x − y| |y| + |x − y| + R
Written out in spherical nomenclature x = R |x| x y
, y = R |y| , x = y on ΩR , we find
⎛ ⎞
x y
x y 2 R |x| − |y|
N R ,R = + ln ⎝ ⎠ . (2.45)
|x| |y| y x y
|x|
x
− |y| 2R + R |x| − |y|
If we use
1
x y x·y 2

|x| − |y| = 2 − 2 |x| |y| , (2.46)
then, for x = y, we are led to the identity

x y x y
N R ,R =N ,
|x| |y| |x| |y|
⎛ ⎞
√ √
2 ⎝ 2 ⎠.
= − ln 1 + (2.47)
1 − |x| · |y|
x y
1 − |x|
x
· y
|y|
Consequently, for points x ∈ ΩR , we (formally) get the so-called Neumann formula,

which constitutes an improper integral over ΩR :

x 1 y x y
T R = D R N , dω(y), (2.48)
|x| 4πR ΩR |y| |x| |y|
where the Neumann kernel constitutes a radial basis function due to (2.47).
Once more, in accordance with the conventional approach of physical geodesy,
the Neumann formula (2.48) is valid under the following constraints (see also
[23, 32, 51]):
(i) the mass within the reference ellipsoid is equal to the mass of the Earth,
(ii) the center of the reference ellipsoid coincides with the center of the Earth,
(iii) the formulation is given in the spherical context to guarantee economical and
efficient numerics.
Note that we are able to set N (Rξ, Rη) = N (ξ, η) = N (ξ · η) in terms of the
x y
unit vectors ξ = |x| and η = |y| which simplifies our notation: If we define the
single-layer kernel S : [−1, 1) → R by
√
2
S(t) = √ , t ∈ [−1, 1), (2.49)
1−t
the Neumann kernel is the zonal function of the form
N (ξ · η) = S(ξ · η) − ln (1 + S(ξ · η)) , 1 − ξ · η > 0. (2.50)
An equivalent formulation of the improper integral (2.48) over the unit sphere
Ω = ∂B1 (0) is then given by

R
T (Rξ) = D(Rη)N (ξ · η) dω(η). (2.51)
4π Ω
It should be remarked that the exterior Stokes boundary value problem of
determining the disturbing potential from known gravity anomalies can be handled
in a quite analogous way (see [9, 15, 73]), providing the so-called Stokes integral
associated to the radially symmetric Stokes kernel as an improper integral on ΩR .
Next we deal with the vertical deflections Θ (cf. [15, 21]). Suppose that T
fulfills the conditions (2.40). We consider the differential equation (compare Eq.
(2.36))
γM
∇∗ξ T (Rξ) = − Θ(Rξ), (2.52)
R
where T (R·) represents the disturbing potential and Θ(R·) denotes the vertical
deflection. The differential equation (2.52) can be solved in a unique way by means
of the fundamental solution with respect to the Beltrami operator

γM
T (Rξ) = Θ(Rη) · ∇∗η G (Δ∗ ; ξ · η) dω(η), (2.53)
R Ω
where (ξ, η) → G (Δ∗ ; ξ · η), 1 − ξ · η = 0, is the fundamental solution of the

Beltrami equation on the unit sphere Ω (see, e.g., [11]) given by
1 1
G (Δ∗ ; ξ · η) = ln(1 − ξ · η) + (1 − ln(2)). (2.54)
4π 4π
The identity (2.53) immediately follows from the Third Green Theorem (cf. [15,
22]) for ∇∗ on Ω in connection with (2.40). By virtue of the identity
ξ − (ξ · η)η
∇∗η G (Δ∗ ; ξ · η) = − , ξ = η, (2.55)
4π(1 − ξ · η)
the integral (2.53) can be written in the form

R
T (Rξ) = Θ(Rη) · g (Δ∗ ; ξ, η) dω(η), (2.56)
4π Ω
where the vector kernel g(Δ∗ ; ξ, η), ξ = η, is given by
γM ξ − (ξ · η)η
g (Δ∗ ; ξ, η) = − . (2.57)
R2 1 − ξ · η
Again we are confronted with a representation of the disturbing potential T as an
improper integral over the sphere ΩR .
All our settings leading to the disturbing potential on the sphere ΩR turn
out to be improper integrals. As we have shown they have either the singularity
behavior of the single-layer kernel S (cf. Eq. (2.49)) or the characteristic logarith-
mic singularity of the fundamental solution with respect to the Beltrami operator
G(Δ∗ ; ·, ·) (cf. (2.54)). Indeed, the fundamental solution and the single-layer kernel
are interrelated (see [15]) by the identities
√ ∗ 1
S(ξ · η) = 2 e−2πG(Δ ;ξ·η)+ 2 (2.58)
and
1 1
G(Δ∗ ; ξ · η) = − ln(S(ξ · η)) − (1 − 2 ln(2)). (2.59)
2π 4π
Therefore, we are confronted with the remarkable situation that a (Taylor) reg-
ularization of the single-layer kernel implies a regularization of the fundamental
solution, and vice versa.
2.2. Zooming-in localization of signature bands
Next, we present multiscale representations for the Neumann kernel N (cf. Eq.
(2.50)). Note that all modern multiscale approaches have a conception of wavelets
as constituting multiscale building blocks in common, which provide a fast and
efficient way to decorrelate a given signal data set.
The characterizing properties of the multiscale approach (basis property,
decorrelation, and efficient algorithms) are common features of all wavelets and
these attributes form the key for a variety of applications (see, e.g., [12, 15, 20]),
particularly for signal reconstruction and decomposition, thresholding, data com-
pression, denoising by, e.g., multiscale signal-to-noise ratio, etc. and, in particular,
decorrelation.
Next, we follow the (taylorized) regularization methods presented in [23] for

linear regularization of the single-layer kernel S and [21] for linear regularization
of the fundamental solution G(Δ∗ ; ·, ·) of the Beltrami equation. For higher-order
approximations, the reader is referred to the Ph.D.-theses [9] and [73].
The essential idea is to regularize the single-layer kernel function
√
2
S(t) = √ (2.60)
1−t
by replacing it by a Taylor linearization. To this end, we notice that the first
derivative of the kernel S is given by
1
S
(t) = √ 3 , t ∈ [−1, 1). (2.61)
2(1 − t) 2
Consequently, we obtain as (Taylor) linearized approximation corresponding to
τ2
the expansion point 1 − 2R 2 , τ ∈ (0, 2R],
2

τ
τ2 τ2
S(t) = S 1 − +S 1− t− 1− + ··· . (2.62)
2R2 2R2 2R2
In more detail, the kernel S is replaced by its (Taylor) linearized approximation
τ2
S τ at the point 1 − 2R 2 , τ ∈ (0, 2R], given by
⎧
⎨ R 3 − 2R2 (1 − t) , 0 ≤ 1 − t ≤ 2Rτ2
τ τ τ 2 2,
S (t) = √ (2.63)
⎩√ 2 , τ2
1−t 2R2 < 1 − t ≤ 2.
Note that the expansion point 1 − τ 2 /(2R2 ), τ ∈ (0, 2R], is chosen in consis-
tency with the notation in the initial paper [21] and the subsequent papers [14] and
[23]. A graphical illustration of the original kernel S(t) and a τ -scale dependent
version of its linear space-regularized kernel S τ (t) is shown in Figure 2.4.
Clearly, the function S τ is continuously differentiable on the interval [−1, 1],
and we have 2R3 τ2
τ
τ3 , 0 ≤ 1 − t ≤ 2R 2,
(S ) (t) = 1 τ2 (2.64)
√ 3 ,
2(1−t) 2 2R2 < 1 − t ≤ 2.
Furthermore, the functions S and S τ are monotonously increasing on the interval

[−1, 1), such that S(t) ≥ S τ (t) ≥ S(−1) = S τ (−1) = 1 holds true on the interval
[−1, 1). Considering the difference between the kernel S and its linearly regularized
version S τ , we find
√ 2
2
√ 2 − R 3 − 2R2 (1 − t) , 0 < 1 − t ≤ τ 2 ,
1−t 2R
S(t) − S (t) =
τ τ τ
(2.65)
τ2
0, 2R2 < 1 − t ≤ 2.
By elementary manipulations of one-dimensional analysis we readily obtain
1
|S(t) − S τ (t)| dt = O(τ ). (2.66)
−1
8
S(t)
7
0.5
6 S (t)
5
1
3 S (t)
2
S2(t)
1
0
-1 -0.5 0 0.5 1
Figure 2.4. Single-layer kernel S(t) (continuous black line) and its
Taylor linearized regularization S τ (t), for R = 1 and τ = 12 , 1, 2 (dotted
lines).
As a consequence, we have
Lemma 2.5. For F ∈ C (0) (Ω) and S τ defined by (2.63) the limit relation

lim sup S(ξ · η)F (η) dω(η) − S (ξ · η)(ξ · η)F (η) dω(η) = 0
τ
(2.67)
τ →0+ ξ∈Ω Ω Ω
holds true.
In a similar way, by some elementary calculations, one can find the following
relations that are also of importance for the Stokes boundary value problem (see
also the Ph.D.-theses [9, 73]).
Lemma 2.6. Let S be the single-layer kernel given by (2.60) and let S τ , τ ∈ (0, 2R],
be the corresponding (Taylor) linearized regularized kernel defined by (2.63). Then
1
lim |ln (1 + S(t)) − ln (1 + S τ (t))| dt = 0, (2.68)
τ →0+ −1

1

1 1 1 1
lim ln + − ln + dt = 0, (2.69)
τ →0+ −1 S(t) (S(t))2 S τ (t) (S τ (t))2
1 #
2 2
lim (S(t)) − (S τ (t)) 1 − t2 dt = 0. (2.70)
τ →0+ −1
To study the surface gradient and the surface curl gradient, we let F be of
class C (1) (Ω). Letting tξ ∈ R3×3 be the orthogonal matrix (with det(tξ ) = 1)
leaving ε3 fixed such that tξ ξ = ε3 , we get

∗
∇ξ S(ξ · η)F (η) dω(η) = S(η3 )∇∗ξ F (tTξ η) dω(η) (2.71)
Ω Ω
for ξ ∈ Ω and η = (η1 , η2 , η3 )T . By regularizing the single-layer kernel, we obtain

∇∗ξ S τ (ξ · η)F (η) dω(η) = S τ (η3 )∇∗ξ F (tTξ η) dω(η) (2.72)
Ω Ω
for ξ ∈ Ω. The same argumentation holds true for the operator L∗ . Therefore,
Lemma 2.5 leads us to the following limit relations (see [22]).
Lemma 2.7. Let F be of class C (1) (Ω). Let S τ be given by (2.63). Then

lim sup ∗ τ
∇ξ S (ξ · η)F (η)dω(η) − ∇ξ ∗
S(ξ · η)F (η)dω(η) = 0, (2.73)
τ →0+ ξ∈Ω Ω Ω

lim sup L∗ξ S τ (ξ · η)F (η)dω(η) − L∗ξ S(ξ · η)F (η)dω(η) = 0. (2.74)
τ →0+ ξ∈Ω Ω Ω
Using the kernel Gτ (Δ∗ ; ·), given by (see Eq. (2.59))

1 1
Gτ (Δ∗ ; t) = − ln(S τ (t)) − (1 − 2 ln(2)), −1 ≤ t ≤ 1, (2.75)
2π 4π
as “single-layer kernel regularization” of the fundamental solution G(Δ∗ ; ·), we are
led to the following integral relations.
Lemma 2.8. For F ∈ C (0) (Ω) and Gτ (Δ∗ ; ·) defined by (2.75), we have

lim sup G(Δ∗ ; ξ · η)F (η) dω(η) − Gτ (Δ∗ ; ξ · η)F (η) dω(η) = 0, (2.76)
τ →0+ ξ∈Ω Ω Ω
and

lim sup ∇∗ξ Gτ (Δ∗ ; ξ · η)F (η) dω(η) − ∇∗ξ G(Δ∗ ; ξ · η)F (η) dω(η) = 0, (2.77)
τ →0+ ξ∈Ω Ω Ω

lim sup ∗ τ ∗
Lξ G (Δ ; ξ · η)F (η) dω(η) − Lξ∗
G(Δ ; ξ · η)F (η) dω(η) = 0.
∗
(2.78)
τ →0+ ξ∈Ω Ω Ω
Remark 2.9. Numerical implementations and computational aspects of the Taylor

regularization techniques as presented here have been applied (even for subsets
of ΩR ) to different fields of physical geodesy (see, e.g., [13–15, 21–23] and the
The regularization techniques enable us to formulate multiscale solutions for
the disturbing potential from gravity disturbances or vertical deflections (note
that we need higher-order regularizations whenever gravitational observables con-
taining second or higher-order derivatives come into play; an example is gravity
gradiometry, which will not be discussed here).
As point of departure for our considerations serves the special case study of
the linear regularization of the single-layer kernel in the integral representation of
the solution of the Neumann boundary-value problem (ENPPG).
Disturbing Potential from Gravity Disturbances. As
we already
know, the solution
of the (Earth’s) disturbing potential T ∈ P ot(1) R3 \BR (0) from known vertical

derivatives, i.e., gravity disturbances D = ∂T
∂ν ΩR , satisfying the conditions (2.40)
on the sphere ΩR , can be formulated as an improper integral (see Eq. (2.51))

R
T (Rξ) = D(Rη) N (ξ · η) dω(η), ξ ∈ Ω, (2.79)
4π
Ω
with the Neumann kernel N (cf. (2.50)). Our interest is to formulate regularizations
of the disturbing potential T by use of the (Taylor) linearized approximation of
the singe-layer kernel S τ : [−1, 1] → R, τ ∈ (0, 2R], introduced in (2.63). As a
result, we obtain the regularized Neumann kernels

τ2
S τ (ξ · η) − ln (1 + S τ (ξ · η)) , 0 ≤ 1 − ξ · η ≤ 2R 2,
N (ξ · η) =
τ
2
S(ξ · η) − ln(1 + S(ξ · η)), 2R2 < 1 − ξ · η ≤ 2,
τ
⎧ 2

⎪ 2R 2R2
⎪
⎪
R
3 − 2 (1 − ξ · η) − ln 1 + R
3 − 2 (1 − ξ · η) ,
⎨τ τ τ τ
τ2
= 0 ≤ 1 − ξ · η ≤ 2R2 , (2.80)
⎪
⎪ √ √
⎪
⎩√ 2 2 2
1−ξ·η
− ln 1 + √1−ξ·η , 2R2 < 1 − ξ · η ≤ 2.
τ
In doing so, we are immediately led to the regularized representation of the dis-
turbing potential T corresponding to the known gravity disturbances:

R
τ
T (Rξ) = D(Rη)N τ (ξ · η) dω(η) (2.81)
4π
η∈Ω

R R
= D(Rη)N (ξ · η) dω(η) + D(Rη)N τ (ξ · η) dω(η).
4π 4π
η∈Ω; η∈Ω;
τ2 τ2
1−ξ·η> 2R 2 1−ξ·η≤ 2R 2
The representation (2.81) is remarkable, since the integrands of T and T τ only

differ on the spherical cap

τ2
Γτ 2 /(2R2 ) (ξ) = η ∈ Ω : 1 − ξ · η ≤ . (2.82)
2R2
By aid of Lemma 2.5 and Lemma 2.6, we obtain
Theorem 2.10. Suppose that T is the solution of the Neumann boundary-value
problem (ENPPG) of the form (2.79). Let T τ , τ ∈ (0, 2R], represent its regular-
ization (2.81). Then
lim sup |T (Rξ) − T τ (Rξ)| = 0. (2.83)
τ →0+ ξ∈Ω
For numerical applications, we have to go over to scale-discretized approxima-

tions of the solution to the boundary-value problem (ENPPG). For that purpose,
we choose a monotonously decreasing sequence {τj }j∈N0 , such that
lim τj = 0, τ0 = 2R. (2.84)
j→∞
A particularly important example, that we use in our numerical implementations

below, is the dyadic sequence with
τj = 21−j R, j ∈ N0 . (2.85)
It is easy to see that 2τj+1 = τj , j ∈ N0 , is the relation between two consecutive
elements of the sequence. In correspondence to the sequence {τj }j∈N0 , a sequence
{N τj }j∈N0 of discrete versions of the regularized Neumann kernels (2.80), so-called
Neumann scaling functions, is available. Figure 2.5 (left) shows a graphical illus-
tration of the regularized Neumann kernels for different scales j.
8
7
NW
J WN W
0
6 N WN W
1
WN W
2
0
-1 -0.5 0 0.5 1
Figure 2.5. Illustration of the Neumann kernel N (t) (left, continuous

black line) and its Taylor linearized regularization N τj (t), j = 0, 1, 2,
τj = 21−j R and R = 1 (left, dotted lines). The corresponding Taylor
linearized Neumann wavelets W N τj (t) for scales j = 0, 1, 2, are shown
on the right.
The regularized Neumann wavelets, forming the sequence {W N τj }j∈N0 , are

understood to be the difference of two consecutive regularized Neumann scaling
functions, respectively,
W N τj = N τj+1 − N τj , j ∈ N0 . (2.86)
The Neumann wavelets are illustrated in Figure 2.5 (right). These wavelets possess
the numerically important property of a local support. More concretely,
η → W N τj (ξ·η), η ∈ Ω, vanishes everywhere outside the spherical cap Γτj2 /(2R2 ) (ξ).
Let J ∈ N0 be an arbitrary scale. Suppose that N τJ is the regularized Neu-
mann scaling function at scale J. Furthermore, let W N τj , j = 0, . . . , J, be the
regularized Neumann wavelets as given by (2.86). Then, we obviously have

J−1
N τJ = N τ0 + W N τj . (2.87)
j=0
The local support of the Neumann wavelets within the framework of (2.87) should
be studied in more detail: We start with the globally supported scaling kernel
N τ0 = N 2R . Then we add more and more wavelet kernels W N τj , j = 0, . . . , J − 1,
)c* uecng k ? 2 )d* k ? 3 )e* k ? 4
)f* k ? 4 )g* k ? 5 )h* k ? 6 )i* k ? 7
)j* k ? 7 )k* k ? 8 )l* k ? 9 )m* k ? : )n* k ? ;

Figure 2.6. Illustration of the regularized Neumann wavelets η →
W N τj (ξ · η) for scales j = 0, . . . , 9 to visualize the local supports
Γτj2 /(2R2 ) (ξ) for a fixed ξ (cf. [73]) “zooming in” to the hotspot of the
Galapagos islands.
to achieve the scaling kernel N τJ . It is of particular importance that the kernel

functions η → W N τj (ξ · η), ξ ∈ Ω fixed, are ξ-zonal functions with local support
(spherical caps). Figure 2.6 illustrates the computationally relevant regions for the
different wavelet scales j (more detailed studies are presented in the Ph.D. the-
ses [9, 73]). For a better understanding, the areas outside the caps are chosen to
be uncolored. Clearly, the support of the wavelets W N τj becomes more localized
for increasing scales j. In conclusion, a calculation of an integral representation
for the disturbing potential T starts with a global trend approximation using the
scaling kernel at scale j = 0 (of course, this requires data on the whole sphere,
but the data can be rather sparsely distributed since they only serve as a trend
approximation). Step by step, we are able to refine this approximation by use of
wavelets of increasing scale. The spatial localization of the wavelets successively
allows a better spatial resolution of the disturbing potential T . Additionally, the
local supports of the wavelets provide a computational advantage since the inte-
gration has to be performed on smaller and smaller spherical caps. In consequence,
the presented numerical technique becomes capable of handling heterogeneously

distributed data.
All in all, keeping the space-localizing property of the regularized Neumann
scaling and wavelet functions in mind, we are able to establish an approximation
of the solution of the disturbing potential T from gravity disturbances D in form
of a “zooming-in” multiscale method. A low-pass filtered version of the disturbing
potential T at the scale j in an integral representation over the unit sphere Ω is
given by (compare Eq. (2.81))

R
τj
T (Rξ) = D(Rη) N τj (ξ · η) dω(η), ξ ∈ Ω, (2.88)
4π Ω
while the j-scale band-pass filtered version of T leads to the integral representation
by use of the wavelets

R
W T τj (Rξ) = D(Rη) W N τj (ξ · η) dω(η), ξ ∈ Ω. (2.89)
4π
Γτ 2 /(2R2 ) (ξ)
j
τJ0
Theorem 2.11. Let T be the regularized version of the disturbing potential at
some arbitrary initial scale J0 as given in (2.88), and let W T τJ0 +j , j = 0, 1, . . . ,
be given by (2.89). Then, the following reconstruction formula holds true:
N

lim sup T (Rξ) − T τJ0 (Rξ) + W T τJ0 +j (Rξ) = 0.
N →∞ ξ∈Ω j=0
The multiscale procedure (wavelet reconstruction) as developed here can be

illustrated by the following scheme
W T τJ0 W T τJ0 +1

T τJ0 −→ + −→ T τJ0 +1 −→ + −→ T τJ0 +2 . . . .
As a consequence, a tree algorithm based on the regularization in the space do-
main has been realized for determining the disturbing potential T from locally
available data sets of gravity disturbances D. An example is shown in Figure 2.7
(following [73]).
In order to get a fully discretized solution of the Neumann boundary-value
problem (ENPPG), approximate integration by use of appropriate cubature formu-
las is necessary (see, e.g., [16, 35] for more details about approximate integration
on the (unit) sphere). The fully discretized multiscale approximations have the
following representations
R Nj Nj τj
Nj
N
T τj (Rξ) & wk D Rηk N ξ · ηk j , ξ ∈ Ω, (2.90)
4π
k=1
R Nj Nj
Nj
N
W T τj (Rξ) & wk D Rηk W N τj ξ · ηk j , ξ ∈ Ω, (2.91)
4π
k=1
N N
where ηk j are the Nj integration knots and wk j the integration weights.
→ → →
𝑇 𝜏1 𝑇 𝜏2 𝑇 𝜏3
+ ↗ + ↗ + ↗
𝑊 𝑇 𝜏1 𝑊 𝑇 𝜏2 𝑊 𝑇 𝜏3
→ →
𝑇 𝜏4 𝑇 𝜏5 𝑇 𝜏6
⋅⋅⋅
+ ↗ + ↗ +
𝑊 𝑇 𝜏4 𝑊 𝑇 𝜏5 𝑊 𝑇 𝜏6
Figure 2.7. Illustration of a (global) multiscale approximation of the

2
Earth’s disturbing potential T in [ ms2 ] from gravity disturbances D,
τj
i.e., low-pass filtered versions T and detail information (band-pass
filtered versions) W T τj for scales j = 1, . . . , 6, by use of the linear
Neumann scaling functions and wavelets computed from 4 000 000 data
points distributed over the whole sphere ΩR (from the Ph.D.-thesis [73],
Geomathematics Group, University of Kaiserslautern).
Whereas the sum in (2.90) has to be calculated on the whole sphere Ω, the
summation in (2.91) has to be computed only for the local supports of the wavelets
(note that the symbol & means that the error between the right-hand and the left-
hand side can be neglected).
Figures 2.8 to 2.10 present a decomposition of the Earth’s disturbing potential
T in low-pass and band-pass filtered parts for data sets of increasing data density.
(a) Low pass part T τ4 calculated from 490 000 data points distributed over the whole
sphere ΩR
(b) Details W T τ4 at scale 4 from 281 428 (c) Details W T τ5 at scale 5 from 226 800
data points distributed within the black data points distributed within the gray
bordered region in Figure 2.8(a) bordered region in Figure 2.8(a)
Figure 2.8. Low-pass filtered version T τ4 of the disturbing potential
2
τj
T in [ m
s2 ] and the corresponding band-pass filtered versions W T for
scales j = 4, 5 of the magenta bordered region in subfigure 2.8(a) calcu-
lated from different numbers of data points (from the Ph.D.-thesis [73],
Seen from the geodetic reality, the figures are remarkable in the following sense:
For getting a better accuracy in numerical integration procedures providing the
(a) Low pass part T τ6 of the magenta bordered

region in Figure 2.8(a) computed by the sum of
T τ4 (Figure 2.8(a)), W T τ4 (Figure 2.8(b)), and
W T τ5 (Figure 2.8(c)) in this region
(b) Details W T τ6 at scale 6 from 71 253 (c) Details W T τ7 at scale 7 from 63 190
data points distributed within the black data points distributed within the gray
bordered region in Figure 2.9(a) bordered region in Figure 2.9(a)

2
T in [ m
s2 ] of the magenta bordered region in subfigure 2.8(a) and the
corresponding band-pass filtered versions W T τj for scales j = 6, 7 (from
the Ph.D.-thesis [73], Geomathematics Group, University of Kaisers-
lautern).
(a) Low pass part T τ8 of the magenta bordered

region in Figure 2.9(a) computed by the sum of
T τ6 (Figure 2.9(a)), W T τ6 (Figure 2.9(b)), and
W T τ7 (Figure 2.9(c)) in this region
(b) Details W T τ8 at scale 8 from (c) Details W T τ9 at scale 9 from

71 253 data points distributed with- 63 190 data points distributed with-
in the black bordered region in Fig- in the gray bordered region in Figure
ure 2.10(a) 2.10(a)

2
T in [ m
s2 ] of the magenta bordered region in subfigure 2.9(a) and the
corresponding band-pass filtered versions W T τj for scales j = 8, 9 (from
the Ph.D.-thesis [73], Geomathematics Group, University of Kaisers-
lautern).
(global) solution of the boundary-value problem (ENPPG) as illustrated in Figure

2.8 (a), we need denser, globally over the whole sphere ΩR equidistributed data
sets (most notably, in the sense of Weyl’s Law of Equidistribution). However,
in today’s reality of gravitational field observation, we are confronted with the
problem that terrestrial gravitational data (such as gravity disturbances, gravity
anomalies) of sufficient width and quality are only available for certain parts of
the Earth’s surface (for more details concerning the observational aspects see, e.g.,
[6–8, 61, 62]). As a matter of fact, there are large gaps, particularly at sea, where
no data sets of sufficient quality are available at all. This is the reason why the
observational situation implies the need for specific geodetically oriented modeling
techniques taking the heterogeneous data situation and the local availability of the
data (usually related to latitude-longitude data grids) into consideration. In this
respect, the “zooming-in” realization based on single-layer space-regularization is
a suitable efficient and economic mathematical answer.
Disturbing Potential from Vertical Deflections. As already known from (2.56), the
solution of the surface differential equation (see Eq. (2.36))
γM
∇∗ξ T (Rξ) = − Θ(Rξ), ξ ∈ Ω, (2.92)
R
determining the disturbing potential T from prescribed vertical deflections Θ under
the conditions (2.40) is given by

R
T (Rξ) = Θ(Rη) · g (Δ∗ ; ξ, η) dω(η), (2.93)
4π Ω
where the vector kernel g (Δ∗ ; ξ, η) , 1 − ξ · η > 0, reads as follows (see Eq. (2.57))
1 γM 2
g (Δ∗ ; ξ, η) = − (ξ − (ξ · η)η)
2 R2 1 − ξ · η
1 γM
=− (S(ξ · η))2 (ξ − (ξ · η)η). (2.94)
2 R2
Analogously to the calculation of the disturbing potential T from known
gravity disturbances D (i.e., the Neumann problem (ENPPG)), the numerical
calamities of the improper integral in (2.93) can be circumvented by replacing
the zonal kernel S(ξ · η) by the regularized kernel S τ (ξ · η). This process leads to
space-regularized representations T τ of the disturbing potential T calculated from
vertical deflections Θ within a multiscale “zooming-in” procedure analogous to the
approach for gravity disturbances as input data. To be more concrete, the kernel
function g(Δ∗ ; ·, ·) is replaced by the space-regularized function using Eq. (2.63)
γM
g τ (Δ∗ ; ξ, η) = − 2 (S τ (ξ · η))2 (ξ − (ξ · η)η), (2.95)
2 2R
⎧
γM 9R 12R4 4R6 2 τ2
⎨− 2R 2 τ 2 − τ 4 (1 − ξ · η) + τ 6 (1 − ξ · η) (ξ − (ξ · η)η), 0 ≤ 1−ξ·η ≤ 2R2
,
=
⎩ γM 2 τ2
− 2R2 1−ξ·η (ξ − (ξ · η)η), 2R2
< 1 − ξ · η ≤ 2,
for τ ∈ (0, 2R]. This leads to the following approximative representation of the
disturbing potential T :

R
T τ (Rξ) = Θ(Rη) · g τ (Δ∗ ; ξ, η) dω(η), (2.96)
4π Ω
with g τ (Δ∗ ; ·, ·) given by (2.95). Using Eq. (2.70) from Lemma 2.6 we obtain
Theorem 2.12. Suppose that T is the solution (2.93) of the differential equation
(2.92), with Θ being a member of the class of continuous vector-valued functions
c(0) (ΩR ). Let T τ , τ ∈ (0, 2R], represent its regularized solution of the form (2.96).
Then
lim sup |T (Rξ) − T τ (Rξ)| = 0. (2.97)
τ →0+ ξ∈Ω
By restricting {g τ (Δ∗ ; ·, ·)}τ ∈(0,2R] to the sequence {g τj (Δ∗ ; ·, ·)}j∈N0 , cor-

responding to a set of scaling parameters {τj }j∈N0 satisfying τj ∈ (0, 2R] and
limj→∞ τj = 0, we are canonically led to regularized vector scaling functions such
that a scale-discrete solution method for the differential equation (2.92) can be
formulated. The vector scaling function g τj+1 (Δ∗ ; ·, ·) at scale j + 1 is constituted
by the sum of the vector scaling function g τj (Δ∗ ; ·, ·) and the corresponding dis-
cretized vector wavelet wg τj (Δ∗ ; ·, ·), given by
wg τj (Δ∗ ; ξ, η) = g τj+1 (Δ∗ ; ξ, η) − g τj (Δ∗ ; ξ, η) . (2.98)
Note that (cf. [15])

W T τj (Rξ) = Θ(Rη) · wg τj (Δ∗ ; ξ, η)dω(η).
Ω
Application: Gravitational signatures of mantle plumes

Galapagos: “The Galapagos hotspot (Figures 2.8–2.10) is a volcanic hotspot in the
East Pacific Ocean responsible for the creation of the Galapagos Islands as well as
three major aseismic ridge systems, Carnegie, Cocos and Malpelso which are on
two tectonic plates. The hotspot is located near the Equator on the Nazca Plate
not far from the divergent plate boundary with the Cocos Plate. The tectonic
setting of the hotspot is complicated by the Galapagos Triple Junction of the
Nazca and Cocos plates with the Pacific Plate. The movement of the plates over
the hotspot is determined not solely by the spreading along the ridge but also by
the relative motion between the Pacific Plate and the Cocos and Nazca Plates.
The hotspot is believed to be over 20 million years old and in that time, there
has been interaction between the hotspot, both of these plates, and the divergent
plate boundary, at the Galapagos Spreading Center. Lavas from the hotspot do
not exhibit the homogeneous nature of many hotspots; instead there is evidence
of four major reservoirs feeding the hotspot. These mix to varying degrees at
different locations on the archipelago and also within the Galapagos Spreading
Center.” (from [71]) (for more details the reader is referred, e.g., to [31] and the
Hawaii: [58] believe that a stationary mantle plume located beneath the Hawaiian
Islands created the Hawaii-Emperor seamount chain while the oceanic lithosphere
continuously passed over it. The Hawaii-Emperor chain consists of about 100 vol-
canic islands, atolls, and seamounts that spread nearly 6000km from the active
volcanic island of Hawaii to the 75–80 million year old Emperor seamounts nearby
the Aleutian trench. With moving further south east along the island chain, the
geological age decreases. The interesting area is the relatively young southeastern
part of the chain, situated on the Hawaiian swell, a 1200km broad anomalously
shallow region of the ocean floor, extending from the island of Hawaii to the Mid-
way atoll. Here, a distinct gravity disturbance and geoid anomaly occurs that has
its maximum around the youngest island that coincides with the maximum to-
pography and both decrease in northwestern direction. The progressive decrease
in terms of the geological age is believed to result from the continuous motion of
the underlying plate (cf. [50, 72]).
With seismic tomography, several features of the Hawaiian mantle plume are
gained (cf. [58] and the references therein). They result in a Low Velocity Zone
(LVZ) beneath the lithosphere, starting at a depth of about 130–140km beneath
the central part of the island of Hawaii. So far, plumes have just been identified as
low seismic velocity anomalies in the upper mantle and the transition zone, which
is a fairly new achievement. As plumes are relatively thin with respect to their di-
ameter, they are hard to detect in global tomography models. Hence, despite novel
advances, there is still no general agreement on the fundamental questions con-
cerning mantle plumes, like their depth of origin, their morphology, their longevity,
and even their existence is still discussed controversial. This is due to the fact that
many geophysical as well as geochemical observations can be explained by different
plume models and even by models that do not include plumes at all (e.g., [10]).
With our space-localized multiscale method of deriving gravitational signatures
(more concretely, the disturbing potential) from the vertical deflections, we add a
new component in specifying essential features of plumes. The vertical deflections
of the plume in the region of Hawaii are visualized in Figure 2.11.
From the band-pass filtered detail approximation of the vertical deflections
(Figure 2.12) and the corresponding disturbing potential (Figure 2.13), we are
able to conclude that the Hawaii plume has an oblique layer structure. As can be
seen in the lower scale (for which numerical evidence suggests that they reflect the
higher depths), the strongest signal is located in the ocean in a westward direction
of Hawaii. With increasing scale, i.e., lower depths, it moves more and more to the
Big Island of Hawaii, i.e., in eastward direction.
Iceland: The plume beneath Iceland is a typical example of a ridge-centered mantle

plume. An interaction between the North Atlantic ridge and the mantle plume is
believed to be the reason for the existence of Iceland, resulting in melt produc-
tion and crust generation since the continental break-up in the late Palaeocene
and early Eocene. Nevertheless, there is still no agreement on the location of the
plume before rifting started in the East. Controversial discussions, whether it was
Figure 2.11. Illustration of the vertical deflections Θ in the region of

Hawaii (from the Ph.D.-thesis [9], Geomathematics Group, University
of Kaiserslautern).
located under central or eastern Greenland about 62-64 million years ago are still
in progress (cf. [63] and the references therein).
Iceland itself represents the top of a nearly circular rise topography, with
a maximum of about 2.8km above the surrounding seafloor in the south of the
glacier “Vatnajökull”. Beneath this glacier, several active volcanoes are located,
which are supposed to be fed by a mantle plume. The surrounding oceanic crust
consists of three different types involving a crust thickness that is more than three
times as thick as average oceanic crusts. Seismic tomography provides evidence of
the existence of a mantle plume beneath Iceland, resulting in low velocity zones
in the upper mantle and the transition zone, but also hints for anomalies in the
deeper mantle seem to exist. The low velocity anomalies have been detected in
depths ranging from at least 400km up to about 150km. Above 150km, ambiguous
seismic-velocity structures were obtained involving regions of low velocities covered
by regions of high seismic velocities. For a deeper access into the theory of the
Iceland plume, the interested reader is referred to [58] and the references therein.
From Figures 2.14 to 2.16, it can be seen that the mantle plume in lower
scales, i.e., in higher depths, starts in the North of Iceland and with increasing
scale, i.e., lower depths, it moves to the South. It is remarkable that from scale 13
on, the plume seems to divide into two sectors. Since it is known that the disturbing
+ +
low-pass filtering band-pass filtering

(scale j = 6) (scale j = 6)
+ +
band-pass filtering band-pass filtering

+ +
band-pass filtering band-pass filtering

+ =
band-pass filtering low-pass filtering

Figure 2.12. Approximation of the vector-valued vertical deflections Θ
in [ms−2 ] of the region of Hawaii (compare Fig. 2.11). A rough low-pass
filtering at scale 6 is improved by several band-pass filters of scale j =
6, . . . , 11, the last picture shows the multiscale approximation at scale
j = 12, (from the Ph.D.-thesis [9], Geomathematics Group, University
of Kaiserslautern).
+ +
T τ6 W T τ6
+ +
W T τ7 W T τ8
+ +
W T τ9 W T τ10
+ =
W T τ11 T τ12
Figure 2.13. Multiscale reconstruction of the disturbing potential T in
[m2 s−2 ] from vertical deflections Θ for the Hawaiian (plume) area using
the scaling function g τ (a rough low-pass filtering T τ6 at scale j = 6
is improved by several band-pass filters W T τj at scales j = 6, . . . , 11,
the last illustration shows the approximation T τ12 of the disturbing
potential T at scale j = 12, (from the Ph.D.-thesis [9], Geomathematics
Group, University of Kaiserslautern).
Figure 2.14. Illustration of the vertical deflections Θ in the region of

Iceland (see [15]).
potential of the Earth is influenced by its topography, a look at a topographic map

shows that the sector located more Eastern is (probably) caused by the Vatnajökull
glacier (being the biggest glacier in Europe).
All in all, from our multiscale reconstruction, it can be derived that the deeper
parts of the mantle plume are located in the northern part of Iceland (compare
the lower scales in Figure 2.15) while shallower parts are located further south
(compare the higher scales in Figure 2.15). As the North American plate moves
westward and the Eurasian plate eastward, new crust is generated on both sides
of the Mid-Atlantic Ridge. In the case of Iceland, which lies on the Mid-Atlantic
Ridge, the neovolcanic zones are readily seen in Figure 2.16.
In Iceland, electrical production from geothermal power plants has been de-
veloped rapidly. Reflecting the geological situation, Iceland is a unique country
with regard to utilization of geothermal energy, with more than 50% of its primary
energy consumption coming from geothermal power plants. As shown in Figure
2.17, today’s location of power plants in Iceland fits perfectly with the gravimetric
investigations based on horizontal/vertical derivatives of the Earth’s disturbing
potential. As a matter of fact, only from these results it becomes obvious where
future power plants should be placed for geothermal purposes.
W T τ10 W T τ11
W T τ12 W T τ13
W T τ14 W T τ15
τj
Figure 2.15. Band-pass filtered details W T of the disturbing poten-
tial T in [m2 s−2 ] from vertical deflections Θ in the region of Iceland
with respect to the scales j = 10, . . . , 15, (from [15]).
W T τ14 W T τ15
0 2 −2
Figure 2.16. Band-pass
1 filtered details W T τj of the disturbing poten-
tial T in m s from vertical deflections Θ in the region of Iceland
for j = 14, 15 including the Mid-Atlantic Ridge (gray).
Figure 2.17. Geothermal power plants in Iceland [1, 37].
All in all, by the space-based multiscale techniques initiated by Freeden and

Schreiner [21, 22] in gravitation we are able to come to interpretable results involv-
ing geological obligations in relation to hotspots/mantle plumes based on “surface
interpretations” and just by looking at the anomalous behavior in terms of surface
integrals without using the framework of Newton’s volume integrals.
3. Interior gravitational potential and density distribution

3.1. Newton integral and Poisson equation
Seen from a mathematical point of view, the Earth’s gravitational field v is a
gradient field v = ∇V, where the gravitational potential V is an infinitely often
differentiable harmonic scalar field in the exterior of the Earth. As a consequence,
the Earth’s gravitational field v is an infinitely often differentiable vector field in
the exterior of the Earth satisfying ∇ · v = 0, ∇ ∧ v = 0.
According to the classical Newton Law of Gravitation (1687), knowing the
density distribution of a region G such as the Earth, the gravitational potential
(Newton potential) can be computed everywhere in R3 . More explicitly, the grav-
itational potential V of the Earth’s exterior G c = R3 \G is given by

V (x) = γ F (y)G(Δ; |x − y|) dV (y), x ∈ R3 \G, (3.1)
G
with the so-called fundamental solution G(Δ; ·) of the Laplace equation given by
1 1
G(Δ; |x − y|) = − , (3.2)
4π |x − y|
and the gravitational constant γ, where F is the density function. Since γ is a
constant, it has no effect on any of the following considerations. Hence, from now
on, for the sake of simplicity, we neglect the gravitational constant γ in all equa-
tions, but it will be observed in numerical computations. The properties of the
gravitational potential V in the Earth’s exterior are easily described as follows:
(i) V is harmonic in R3 \G, i.e., Δx V (x) = 0, x ∈ R3 \G.
(ii) V is regular at infinity, i.e., |V (x)| = O |x|−1 , |x| → ∞.
Let G ⊂ R3 be a regular region, i.e., a bounded region G ⊂ R3 dividing R3
uniquely into the inner space G and the outer space G c = R3 \G, G = G ∪ ∂G, such
that the boundary ∂G is an orientable smooth Lipschitzian manifold of dimension
2 (for example, ball, ellipsoid, geoid, Earth or appropriate cuboidal parts of it).
It is already known, that the Newton (volume) integral over a regular region G,
corresponding to a mass density distribution F satisfies the Laplace equation in
the outer space G c = R3 \G. Clearly, this property is an immediate consequence
of the harmonicity of the fundamental solution for the Laplace equation (see, e.g.,
[40]).
Theorem 3.1. Let F : G → R be an integrable, bounded function. Then

V (x) = F (y) G(Δ; |x − y|) dV (y), x ∈ Gc, (3.3)
G
satisfies

Δx V (x) = Δx F (y) G(Δ; |x − y|) dV (y) = 0, x ∈ Gc, (3.4)
G
i.e., V is harmonic in G c .
Next, we are interested in showing that the Newton integral in the inner
space satisfies the Poisson equation at least under some canonical conditions on
the density function (see, e.g., [15]).
Theorem 3.2. Let F : G → R be of class C (0) (G). Then V as defined by (3.3) is of

class C (1) (G). Furthermore, we have

∇x V (x) = F (y) ∇x G(Δ; |x − y|) dV (y), x ∈ G. (3.5)
G
Proof. The fundamental solution

1 1
G(Δ; |x − y|) = − , |x − y| = 0, (3.6)
4π |x − y|
admits a “regularization” (mollification) of the form

⎧
⎪
⎪ 3τ 2 − |x − y|2
⎨ − , |x − y| ≤ τ,
8πτ 3
Gτ0 (Δ; |x − y|) = (3.7)
⎪
⎪ 1
⎩ − , τ < |x − y|.
4π|x − y|
For brevity, we set

V0τ (x) = F (y) Gτ0 (Δ; |x − y|) dV (y), x ∈ G. (3.8)
G
The integrands of V and V0τ only differ in the ball Bτ (x) around the point x with
radius τ . Moreover, the function F : G → R is supposed to be continuous on G.
Hence, it is uniformly bounded on G and we derive

sup |V (x) − V0τ (x)| = O |G(Δ; |x − y|) − Gτ0 (Δ; |x − y|)| dV (y)
x∈G Bτ (x)
2
= O(τ ). (3.9)
Therefore, V is of class C(0) (G) as the limit of a uniformly convergent sequence of

continuous functions on G. We let

v(x) = F (y) ∇x G(Δ; |x − y|) dV (y), x ∈ G, (3.10)
G
and
v0τ (x) = F (y) ∇x Gτ0 (Δ; |x − y|) dV (y), x ∈ G. (3.11)
G
As |∇x G(Δ; |x − y|)| = O(|x − y|−2 ), the integrals v and v0τ exist for all x ∈ G. It
is not hard to see that
sup |v(x) − v0τ (x)| = sup |v(x) − ∇x V0τ (x)| = O(τ ). (3.12)
x∈G x∈G
Consequently, v is a continuous vector field on G. Moreover, as the relation (3.12)

holds uniformly on G, we obtain

v(x) = ∇x V (x) = F (y) ∇x G(Δ; |x − y|) dV (y). (3.13)
G
This is the desired result.
Remark 3.3. The proof is standard (see, e.g., [22]). Its explicit formulation, how-
ever, is helpful to understand the feature extraction method.
Next, we come to the Poisson equation under the assumption of Hölder con-
tinuity for the function F on G.
Theorem 3.4. If F is of class C (0,μ) (G), μ ∈ (0, 1], then the Poisson differential
equation

Δx F (y) G(Δ; |x − y|) dV (y) = F (x) (3.14)
G
holds true for all x ∈ G.
The proof can be found in any textbook on potential theory, e.g., [15]. It is
also part of Chapter 5 of this handbook.
The fundamental solution Gτ0 (Δ; ·) as well as the (ordinary) Haar function
given by

3
4πτ 3 , |x − y| ≤ τ,
H0τ (|x − y|) = Δx Gτ0 (Δ; |x − y|) = (3.15)
0, |x − y| > τ,
are depicted in Figure 3.1 for different values of τ .
Figure 3.1. Sectional profile of the functions Gτ0 (Δ; ·) (left) and H0τ
(right) for the values τ = 2−j , j = 0, 1, 2. The black line in the left figure
indicates the profile of the fundamental solution G(Δ; ·).
The critical point that will be expected in numerics of feature extraction by

means of regularized potentials V0τ is the discontinuity of the Laplace derivative
of Gτ0 (Δ; ·), i.e., the (ordinary) Haar function H0τ . This is the reason why we are
interested in higher-order Taylor expansions of the fundamental solution leading
to a polynomial of degree n + 2 given by r → Gτn (Δ; r), r ∈ [0, ∞, ) with
⎧ 1
⎪
⎪ − 4πr , τ ≤ r,
⎪
⎪
⎪
⎪
n+1 n+2−l l
⎪
⎨ 8πτ1n+3 (−1)l (n + 1) n+2 τ r
l
Gτn (Δ; r) = l=0

n+2−l l (3.16)
⎪
⎪ 1
n+1
⎪
⎪ + 2(−1) l+1 n+2
τ r
⎪
⎪
8πτ n+3 l+1
⎪
⎩
l=0
+(−1) n+2 n+1
8πτ n+3 r
n+2
, 0≤r<τ
instead of
1
− 4πr , τ ≤ r,
Gτ0 (Δ; r) = 2 2
(3.17)
− 3τ8πτ−r3 , 0 ≤ r < τ,
so that

0, τ < r,
Hnτ (r) = Δx Gτn (Δ; r) = (n+1)(n+2)(n+3) (τ −r) n (3.18)
8π τ n+3 , 0 ≤ r ≤ τ.
It is easy to see that r → Gτn (Δ; r), r ∈ [0, ∞), is (n + 1)-times continuously
differentiable and r → Hnτ (r), r ∈ [0, ∞), is (n−1)-times continuously differentiable
(where, by convention in case of H0τ , (−1)-times continuously differentiable means
piecewise continuous). Moreover, we notice that H0τ for n = 0 is the ordinary
(spherically symmetric) τ -Haar function in R3 .
As a consequence of our preparatory considerations we obtain the following
statement that serves as strategic basis for our forthcoming approach to geological
feature extraction.
Theorem 3.5. For n ∈ N0 , the “τ -potential functions” of order n

τ
Vn (x) = Gτn (Δ; |x − y|)F (y) dV (y) (3.19)
G
and the “τ -contrast functions” of order n

Fnτ (x) = Hnτ (|x − y|)F (y) dV (y), (3.20)
G
satisfy the limit relations
lim |V (x) − Vnτ (x)| = 0, x∈G (3.21)

τ →0
and
lim |F (x) − Fnτ (x)| = 0, x ∈ G, (3.22)
τ →0
provided that F is (C(0,μ) -Hölder) continuous in the neighborhood of x ∈ G.
The kernels Gτn (Δ; ·) and Hnτ are called “τ -fundamental scaling function of
order n” and “τ -Haar scaling function of order n”, respectively. It should be re-
marked that Gτn (Δ; ·) is constructed in such a way that the normalization condition

Δx Gτn (Δ; |x|) dV (x) = Hnτ (|x|) dV (x) = 1 (3.23)
R3 R3
holds true for all τ > 0 and all n ∈ N0 .

Unfortunately, τ -potential functions Vnτ do not generally show a faster con-

vergence to V than τ -potential functions V0τ ; more concretely, we have

sup |V (x) − Vn (x)| = O
τ
|G(Δ; |x − y|) − Gn (Δ; |x − y|)| dV (y)
τ
x∈G Bτ (x)
2
= O(τ ) (3.24)
for n ∈ N0 . Finally it should be alluded that
lim sup |α(x)F (x) − Fnτ (x)| = 0, (3.25)
τ →0
x∈G
where α(x) is the solid angle subtended at x ∈ G by the boundary surface ∂G.
Remark 3.6. The solid angle α(x) in Equation (3.25) is necessary due to the fact
that the support of Hnτ (| · −x|) is cut of at the boundary ∂G for all τ > 0 with
x ∈ ∂G.
3.2. Multiscale postprocessing of signature decorrelation
Next we deal with new mathematical mechanisms for a deeper interpretation and a
better understanding of gravimetrically available pre-information inside a regular
region G. In order to make the decorrelation mechanisms transparent, our con-
siderations start from the unrealistic assumption that the potential V is known
everywhere in G. Our purpose is to demonstrate how the multiscale procedure for
the potential canonically transfers to the density by use of “Poisson derivatives”.
All in all, the context of this section is meant as conceptual preparation of the
Haar-type inversion process (see also [3]) discussed later on.
Suppose that {τj }j∈N0 is a positive, monotonously decreasing sequence with
limj→∞ τj = 0. For j ∈ N0 , we consider the differences
ΨGτnj (Δ; |x − y|) = Gτnj+1 (Δ; |x − y|) − Gτnj (Δ; |x − y|) (3.26)
and
ΨHnτj (|x − y|) = Hnτj+1 (|x − y|) − Hnτj (|x − y|). (3.27)
ΨGτnj (Δ; ·) and ΨHnτj are called “τj -fundamental wavelet function of order n” and
“τj -Haar wavelet function of order n”, respectively (see Figure 3.2).
The associated “τj -potential wavelet functions” of order n and the “τj -con-
trast wavelet functions” of order n are given by

(W V )τnj (x) = ΨGτnj (Δ; |x − y|)F (y) dV (y) (3.28)
G
and
(W F )τnj (x) = ΨHnτj (|x − y|)F (y) dV (y). (3.29)
G
The τj -potential wavelet functions of order n and the τj -contrast wavelet functions
of order n, respectively, characterize the successive detail information contained in
τ τ τ τ
Vn j+1 − Vn j and Fnj+1 − Fnj , j ∈ N0 . In other words, we are able to recover the
Figure 3.2. Sectional profile of the wavelet functions ΨGτnj (Δ; ·) (left)
and ΨHnτj (·) (right) for n = 0 and τj = 2−j , j = 0, 1, 2.
potential V and the contrast function, i.e., the “density signature” F , respectively,
in form of “band structures”
(W V )τnj = Vnτj+1 − Vnτj , (3.30)
and
(W F )τnj = Fnτj+1 − Fnτj . (3.31)
As a consequence, the essential problem to be solved in multiscale extraction of
geological features is to identify those detail information, i.e., band structures
in (3.30), which contain specific geological (density) characteristics in (3.31), for
example, aquifers, salt domes, etc.
Seen from a numerical point of view, it is remarkable that both wavelet
functions y → ΨGτnj (Δ; |x − y|) and y → ΨHnτj (|x − y|) vanish outside a ball
around the center x due to their construction, i.e., these functions are spacelimited
showing a ball as local support. Furthermore, the support becomes smaller and
smaller with increasing scale parameter j, so that more and more high frequency
phenomena can be highlighted without changing the features outside the balls.
Explicitly written out in our nomenclature we obtain for x ∈ G

(W V )τnj (x) = ΨGτnj (Δ; |x − y|)F (y) dV (y), (3.32)
Bτj (x)∩G
and
(W F )τnj (x) = ΨHnτj (|x − y|)F (y) dV (y). (3.33)
Bτj (x)∩G
Forming the sums

J−1
J−1
(W V )τnj (x) = (Vnτj+1 (x) − Vnτj (x)) , (3.34)
j=0 j=0
and

J−1
J−1
(W F )τnj (x) = (Fnτj+1 (x) − Fnτj (x)) , (3.35)
j=0 j=0
we are easily led to

J−1
VnτJ (x) = Vnτ0 (x) + (W V )τnj (x) (3.36)
j=0
and

J−1
FnτJ (x) = Fnτ0 (x) + (W F )τnj (x). (3.37)
j=0
Thus, we finally end up with the following multiscale relations

∞

V (x) = lim VnτJ (x) = Vnτ0 (x) + (W V )τnj (x) (3.38)
J→∞
j=0
and
∞

α(x)F (x) = lim FnτJ (x) = Fnτ0 (x) + (W F )τnj (x) = lim Δx VnτJ (x), (3.39)
J→∞ J→∞
j=0
i.e.,
∞

α(x)F (x) = Δx Vnτ0 (x) + Δx (W V )τnj (x). (3.40)
j=0
Altogether, the potential V as well as the contrast function, i.e., the “density
signature” F can be expressed in additive way as a low-pass filtered signal Vnτ0 and
τ τ
Fnτ0 and successive band-pass filtered signals (W V )nj and (W F )nj , j = 0, 1, . . . ,
respectively.
It should be mentioned that our multiscale approach is constructed such that,
within the spectrum of all wavebands (cf. (3.30), (3.31)), certain rock formations
or aquifers, respectively, may be associated to a specific waveband characterizing
typical features within the multiscale reconstruction (see Figure 3.3). Each scale
parameter in the decorrelation is assigned to a data function which corresponds to
the associated waveband and, thus, leads to a low-pass approximation of the data
at a particular resolution.
Finally it should be noted that the key ideas of multiscale approximation as
presented here lead back to evaluation methods proposed by Freeden and Schreiner
[21], Freeden and Blick [13], and particularly Freeden and Gerhards [15]. For the
sake of simplicity, the adaptation of this approach to the requirements of gravita-
tional potential as well as density distribution is explained only in scale discrete
form, a scale continuous formulation as presented in [21] is canonical. A variety
of numerical tests and case studies of our approach are found in the Ph.D.-theses
[2, 49].
ORZSDVV¿OWHU
VFDOHM VFDOHM VFDOHM VFDOHM VFDOHM VFDOHM
9 M 9 M 9 M ) M ) M ) M

RSHUDWRU
EDQGSDVV¿OWHU
VFDOHM VFDOHM VFDOHM VFDOHM
9 M 9 M 9 M ) M ) M ) M

RSHUDWRU
9 M 9 M ) M ) M
9 M 9 M ) M ) M
Figure 3.3. Schematic visualization of the multiscale decorrelation

mechanism (see [3]).
3.3. Decorrelation of the Marmousi test model

Next we deal with the decorrelation of the geological signatures of a test area,
namely the well-known Marmousi potential and density model (we use the canon-
ically constructed 3D-version of the Marmousi model as proposed in the Ph.D.-
thesis [2], see Figures 3.4, 3.5). In accordance with this standard test model (see
also [42, 67]), the contrast function F is available as a fully interpreted 3D Mar-
mousi density model extension (see Figure 3.6).
In order to validate the decorrelation abilities of our multiscale approach pre-
sented in the last section, we first perform a decomposition of the potential based
on Eq. (3.38) (see Figure 3.7). Obviously, the low-pass filtered data, i.e., the τj -
τ
potential functions Vn j (see Eq. (3.19)) provide no essential structural information
(see Figure 3.7, left column). However, for smaller scale values τj , by going over to
finer detail information involving τj -wavelet potential functions, we already notice
essential trends of the geological situation of the original density model (see Figure
3.7, right column).
Keeping the properties of the Newton volume integral in mind, we are not
very surprised that, following the construction principles as proposed in our mul-
tiscale approach, the decomposition of the 3D Marmousi density model F based
on Eq. (3.39) (Figure 3.8, left), in fact, shows a significant correlation to the
decomposition of the τj -wavelet potential functions. Nevertheless, the τj -wavelet
contrast functions (Figure 3.8, right) yield additional information, for example,
the separation of all density transitions can be clearly detected at scale j = 9.
This observation is of great significance in geothermal research, where the fracture
transitions play a particular role for detecting areas of internal water flow.
Next we modify the original potential function V by adding three pertur-
bations in form of mass points, i.e., fundamental solutions, at different locations
obtaining the potential Vmod . Our purpose is to determine the locations of these
three disturbances for the depth detection of geological formations. It should be
noted that a decorrelation of the data with low-frequency wavelets (i.e., low val-
ues j) basically means focusing the multiscale approach on low-frequency signal
components. Wavelets to higher values j allow to focus on the high-frequency
interference.
As a consequence, our multiscale decorrelation mechanism shows that the
low-pass filtered signals of V and Vmod are structurally identical (see Figures 3.7
and 3.9). However, at scale j = 9 (see Figure 3.10), we can identify the exact
location of the centers of the introduced fundamental solutions (i.e., buried mass
points) in the band-pass filtered data.
3.4. Gravimetry and Haar-type inversion
The inversion of Newton’s Law of Gravitation (3.1), i.e., the determination of
the internal “density function” from information of the gravitational potential is
known as the gravimetry problem: To be more concrete, the gravimetry problem
amounts to the problem of determining the “density function” F from (discrete)
information of the gravitational potential V in R3 in accordance with the integral
Figure 3.4. Artificially constructed 3D Marmousi density model.
Figure 3.5. Cross-section of the 3D Marmousi density model (cf. [67]).
Figure 3.6. Marmousi density model and its geological interpretation

(cf. [42]).
k?7
k?8
k?9
k?:
k?;
k ? 32
Figure 3.7. Decomposition of the 3D Marmousi potential in low-pass

τ τ
(Vn j , left) and band-pass filtered parts ((W V )nj , right) for the sequence
−j
τj = 9200m · 2 and n = 0 in [ kg/ m]. The choice of the sequence is
adapted to the length of the density model (from the Ph.D.-thesis [2],
k?7
k?8
k?9
k?:
k?;
k ? 32
Figure 3.8. Decomposition of the 3D Marmousi density model into

τ τ
low-pass (Fnj , left) and band-pass0 filtered1 parts ((W F )nj , right) for
−j 3
the sequence τj = 9200m · 2 in kg/ m and n = 0. The choice of
the sequence is adapted to the length of the density model (from the
Ph.D.-thesis [2], Geomathematics Group, University of Kaiserslautern).
k?9
k?:
k?;
Figure 3.9. Decomposition of the modified 3D Marmousi poten-

τ
tial Vmod in low-pass ((Vmod )nj , left) and band-pass filtered parts
((W Vmod )n , right) for the sequence τj = 9200m · 2−j and n = 0 in
τj
0 kg 1
m . The choice of the sequence is adapted to the length of the density
model.
Figure 3.10. Illustration of the band-pass filtered signal (W Vmod )τn9

at scale j = 9 showing the locations of the three disturbing mass points,
i.e., fundamental solutions and, consequently, the depth of the geological
formations.
equation

V (x) = I[F ](x) = G(Δ; |x − y|)F (y) dV (y), x ∈ R3 (3.41)
G
(note that we omit the gravitational constant γ).
In accordance with the mathematical classification due to Hadamard, the
(classical) gravimetry problem of determining F from potential data on ∂G, i.e.,
terrestrial gravitational data, violates all criteria, viz. existence, uniqueness and
stability:
(i) (Existence) The potential V is harmonic outside G. In accordance with the so-
called Picard condition (see, e.g., [70]), a solution only exists if V belongs to
(an appropriate subset in) the space of harmonic functions. However, it should
be pointed out that this observation does not cause a numerical problem since,
in practice, the information of V is only finite-dimensional. In particular, an
approximation by an appropriate harmonic function is a natural ingredient
of any practical method.
Figure 3.11. Equivalent gravity effect of different “sources” to gener-

ate the same gravitational potential on the Earth’s surface (with kind
permission of Teubner-publishing taken from [38] in modified form).
(ii) (Uniqueness) The most serious problem is the non-uniqueness of the solution
(cf. Figure 3.11): The associated Fredholm integral operator I is of the first
kind and has a kernel (null space) which is known (cf. [15, 20]) to coincide
with the L2 (G)-orthogonal space of the closed linear subspace of all harmonic
functions on G. Unfortunately, this orthogonal complement, i.e., the class of

so-called anharmonic functions, is infinite-dimensional.
(iii) (Stability) Restricting the operator to harmonic densities leads to an injective
mapping which has a discontinuous inverse implying an unstable solution.
Concerning the historical background, the question of the non-uniqueness
for the classical gravimetry problem has been discussed extensively in literature,
starting with a paper by Stokes [66] (for more details see, e.g., [15, 46, 47]). This
calamity can be bypassed by imposing some reasonable additional condition on
the density. A questionable condition, suggested by the mathematical structure of
the Newton potential operator I, is to require that the density is harmonic. The
approximate calculation of the harmonic density has already been implemented
in several papers, whereas the problem of determining the anharmonic part seems
to be still a great challenge. Due to the lack of an appropriate physical interpre-
tation of the harmonic part of the density, various alternative variants have been
discussed in the literature. In general, gravitational data yield significant informa-
tion only about the uppermost part of the Earth’s interior, which is not laterally
homogeneous.
Seen from a mathematical point of view in constructive approximation, lo-
cally supported functions are not new, having been discussed already by Haar
(1910). The importance of spacelimited (or in mathematical jargon locally sup-
ported) Haar kernels in view of a multiscale procedure is the “birth” to an entire
“basis family” (scaling functions) by means of two operations, viz. dilations and
translations.
In what follows, we recapitulate the already discussed “Haar philosophy” to
realize an approximate determination of the mass density distribution inside G
from discrete gravitational information. The essential tool is the regularization
procedure of the Newton potential enabling to replace the integral equation (3.41)
by the Fredholm integral equation of first kind

Vnτ (x) = Gτn (Δ; |x − y|)F (y) dV (y), x ∈ R3 , (3.42)
G
for a sufficiently large scale number J, so that the serious problem of non-unique-
ness caused by the occurrence of anharmonic functions is not existent anymore for
terrestrial potential data, however, at the price of non-harmonicity of the “regu-
larizer” y → GτnJ (Δ; |x − y|) in a neighborhood outside x ∈ ∂G.
In choosing a sufficiently large J we are aware of the fact (cf. Theorem 3.4)
that

V (x) & VnτJ (x) = GτnJ (Δ; |x − y|)F (y) dV (y), x ∈ R3 , (3.43)
G
VnτJ
i.e., provides an approximation of the Newton integral (3.3) with negligible
error. We remember
Δx GτnJ (Δ; |x − z|) = HnτJ (|x − z|) (3.44)
for all x, z ∈ R3 . From (3.44) it therefore follows that

Δx GτnJ (Δ; |x − z|)F (z) dV (z) = FnτJ (x) & F (x), x ∈ G. (3.45)
G
In order to realize a fully discrete approximation of F , we have to apply

approximate integration formulas over BτJ (x) ∩ G leading to

NJ
V (x) & VnτJ (x) & GτnJ (Δ; |x − yiNJ |) wiNJ F (yiNJ ), (3.46)
i=1
where wiNJ , yiNJ ∈ BτJ (x) ∩ G, i = 1, . . . , NJ , are known weights and knots,
respectively.
For the determination of the mass density we are confronted with the situ-
ation that all coefficients aN
i
J
= wiNJ F (yiNJ ), i = 1, . . . , NJ , are unknown. This,
however, means that we have to solve a linear system, namely

NJ
V (xMJ
k ) = GτnJ (Δ; |xM
k
J
− yiNJ |)aN J
i , k = 1, . . . , MJ , (3.47)
i=1
in order to determine the coefficients aN J

i , i = 1, . . . , NJ , from known gravitational
3
values V (xk ) at knots xk ∈ R , k = 1, . . . , MJ .
MJ MJ
Once all density values F (yiNJ ), i = 1, . . . , NJ , are available (note that the in-
tegration weights wiNJ , i = 1, . . . , NJ , are known from the approximate integration
rule), the density distribution F can be obtained from the formula

NJ
F (x) & FnτJ (x) = HnτJ (|x − yiNJ |) wiNJ F (yiNJ ), x ∈ G. (3.48)
i=1
Even more, fully discrete Haar filtered versions of F at lower scales, i.e., feature
extraction, can be derived in accordance with the approximate integration rules

Nj
N N N
τj
Fn (x) = Hn (|x − y|)F (y) dV (y) &
τj
Hnτj (|x − yi j |)wi j F (yi j ) (3.49)
G i=1
N N
for j = J0 , . . . , J, where wi j , yi j ,
i = 1, . . . , Nj , are known weights and knots,
N N
respectively, such that we can take adventage of the fact that {y1 j , . . . , yNjj } ⊂
N N
{y1NJ , . . . , yN
NJ
J
} ⊂ G, i.e., the sequence of knots {y1 j , . . . , yNjj } ⊂ G shows a hier-
archical positioning.
Altogether, our approach yields Haar filtered versions (3.49) establishing a

τ
fully discrete (space-based) multiscale decomposition FnτJ , . . . , FnJ0 of the density
distribution F , such that an entire set of approximations is available from a single
locally supported “mother function”, i.e., the Haar kernel function, and this set
provides useful “building block functions”, which enable suitable storage and fast
decorrelation of density data in consistency with geological formations.
It should be remarked that by discretizing the convolution integral by approx-

imate integration in form of a suitable cubature formula, we arrive at a system
of linear equations, which is sparse, since the utilized wavelet has a local support.
The local support enables us to limit the multiscale technique to a local region,
e.g., only to a relevant borehole area of interest, and guarantees that there is no
change in the signature outside the support of the wavelets. All in all, our approach
is given in such a way that the inversion of the equation system turns out to be
numerically efficient and economical.
Remark 3.7. The linear systems occurring in this section can be handled by, e.g.,
use of domain decomposition techniques (see, e.g., [17, 19, 28–30, 34] and the
Remark 3.8. For simplicity, the input data of this subsection are potential data.
In the same way, a linear system can be established by taking, e.g., free air gravity
anomalies on the Earth’s surface. In this case, however, we need Haar functions
(3.18) of positive degrees n, since free air gravity anomalies are generated by a
first-order derivative applied to the disturbing potential.
3.5. Improvement of in-borehole density signatures

Since both the actual potential V and the actual contrast function F, in general,
cannot be measured directly inside the boundary surface ∂G and outside a neigh-
borhood around the already existing boreholes without additional drilling, the a
priori available potential and density information differ from the actual values and
thus form only an approximation to the reality. If one associates a certain scale
value τj within the multiscale process to the available potential data, we are given

Vnτj (x) = Gτnj (Δ; |x − y|)F (y) dV (y). (3.50)
G
Often, in practice during borehole drilling, additional data are gathered by in-
hole gravimetric measurements, so that we may assign a scale value τj+1 to the
improved potential data. If we now take the difference, we arrive at

Nj
N N N
Vnτj+1 (x) − Vnτj (x) = ΨGτnj Δ; x − yi j wi j F yi j , (3.51)
i=1
and

Nj
N N N
Fnτj+1 (x) − Fnτj (x) = ΨHnτj x − yi j wi j F yi j , (3.52)
i=1
respectively. Once again, it should be emphasized that the linear system (3.51)
τ
can be evaluated efficiently and economically (note that the kernels Hnj as well
as ΨGτnj and ΨHnτj have local support due to their construction and, hence, the
systems of Equations (3.51) and (3.52) are sparse).
The improvement by additional data observation is shown using the example

of the equation system (3.52). The input data for the inversion process are gen-
erated by smoothing of a cutout of the 3D Marmousi density model (see Figure
τ
3.12, left for Fnj+1 ). As a result of the inversion (see Figure 3.12, right), we ob-
tain a sharper density model provided that the wavelet used in the inversion is
sufficiently smooth, i.e., n ≥ 2.
τ
Figure 3.12. Illustration of the “best” data before the inversion Fnj+1
(left) and the inversion result (right) for n = 3 and j = 5. The colors
show the densities in [kg/m3 ] (from the Ph.D.-thesis [2]).
An extensive parameter study in the Ph.D.-thesis [2] demonstrates that the

inversion is numerically stable and efficient for smooth Haar-type kernels. Since
the resulting relative error in the inversion depends continuously on the scale,
there exists a reference interval, such that for each scale value inside this reference
interval, an improvement of the data is achieved. This allows a certain tolerance
in the choice of parameters.
Conclusions
Local knowledge of the gravity potential and its equipotential (level) surfaces giv-
ing information about mass distribution have become an important issue for ex-
ploration and prospecting. Indeed, the gravity field is a key component of future
investigation. Seen from a numerical point of view, however, the way forward has
to focus on two challenges:
(i) In reality, the distribution of geopotential data is far from being homoge-
neous with large gaps even in all European areas. In addition, the quality
of the geopotential data under consideration is very distinct. A terrestrial
data coverage now and in the foreseeable future is far from being satisfac-
tory. For data supplementation and numerical stabilization, airborne and/or
spaceborne data are indispensable. This unfortunate situation causes partic-

ular mathematical attention for homogenization and unification to suppress
undesired oscillation phenomena within the modeling process of the data.
(ii) Nowadays, the knowledge of geopotentials such as the Earth’s gravitational
potential and their anomalies have become an important and cost-effective
issue in exploration technologies. However, it is commonly known that highly
accurate sensors, when operating in an isolated manner, have their short-
comings. Combining globally available satellite data with regional airborne
and/or local terrestrial observations within a physically founded and math-
ematically consistent multiscale process is therefore an essential step for-
ward. In this respect, a “zooming-in” detection of specific geophysical at-
tributes is an outstanding field of interest for validating the multiresolution
method based on heterogeneous datasets and geophysically oriented multi-
scale “downward continuation” modeling of the different data sources starting
from spaceborne data as trend solution via more accurate airborne data down
to high-precision local data sets.
Geophysically relevant signatures are usually decomposed into single fre-
quencies. Geomathematically, these techniques are well suited to resolve low and
medium frequency phenomena, while their application to obtain high resolution
models (such as descriptions of local orebodies, salt deposits, aquifers, etc.) is
critical. Due to the quality of the data, i.e., the intrinsic scale amount of signifi-
cant wave packages within the signal, spaceborne (i.e., satellite) data – continued
downward to the Earth’s surface – are the canonical point of departure for mul-
tiscale approximations of lower scale frequency phenomena, while the quality of
airborne and/or terrestrial data can be associated to medium and/or high(er) scale
frequency bands. So, the whole spectrum of spaceborne/airborne/ground data sys-
tems covers all verifiable wave packages. Actually, the advantage of satellite lower
frequency band data at the ground is their availability everywhere, while (airborne)
medium and (terrestrial) high(er) frequency bands usually are at the disposal for
regional and local occurrence, respectively. In this respect, a helpful tool for de-
termining the depth and size is the introduction of known artificial disturbances
such as monopoles in gravitation which superpose the original wave bands of the
data in an easily predictable and calculable way.
Summarizing our results, we are led to the following conclusion: The multi-
scale approach which is presented in this contribution breaks up a complicated sig-
nal (like the gravitational field, the geomagnetic field) into “wave band signatures”
at different scales, i.e., a certain resolution. To each scale parameter, a scaling func-
tion is defined leading to an approximation of the data at that certain resolution.
The difference between two successive scaling functions, i.e., the wavelets, repre-
sents the corresponding wave bands and, thus, yields the desired geological detail
information. With increasing scale, the approximation is getting finer and finer
starting form a low pass approximation and adding more and more wave bands.
The multiscale approach guarantees that the information contained on a certain
(coarse) level is also contained in the approximations of higher scales. It is advanta-

geous that we are able to analyze the wave bands separately (decorrelation). Thus,
this multiscale concept helps to find adaptive methods to the particular structure
of the data. Additionally, the resolution of the model can be adapted to the spa-
tial structures, i.e., for areas with coarse spatial structures, the resolution of the
model can be chosen to be rather low and for areas with complicated structures
the resolution can be increased accordingly. Consequently, since most data show
correlation both in space as in frequency, the multiscale technique is an appro-
priate method for a simultaneous space and frequency localization. As far as the
numerical realization is concerned, fast wavelet methods (FWT) are applicable.
Considering especially the disturbing potential field approximation in gravi-
tation, we observe – from computational point of view – two main requirements:
First, the field characteristics of geological features are usually of local character
such that the use of local wavelets is evident. Second, in view of physical relevance
of the multiscale approach, we need wavelets which have a certain relation to the
corresponding partial differential equation (here: Laplace equation). Moreover, we
have to be concerned with wavelet types which are manageable from mathematical
point of view and, additionally, show a close relation to the physical model. Indeed,
the developed multiscale method by means of regularizing wavelets using physically
motivated fundamental solutions has its origins in works of the Geomathematics
Group of the University of Kaiserslautern (see, e.g., [12, 15, 18, 20–22] and the list
of references therein). The main results and characteristics of our studies presented
here can be summarized as follows:
• Physically based behavior and appropriate interpretability of the developed
wavelets.
• Numerical efficiency and economy of the wavelets by adaptive choice of the
local support and resulting fast algorithms.
• Scale dependent correlation of wavebands and geological structures in a sys-
tematic “zooming-in / zooming out” decorrelation process.
• Specific transparency of certain geological structures for an appropriate choice
of parameters.
• Depth determination and localization of geological formations by artificial
point source disturbances.
Furthermore seen from the point of mathematical methodology, our multiscale
(postprocessing) approach is not only restricted to potential methods involving the
Laplace operator. Similar approaches can be formulated, e.g., for the Helmholtz
and d’Alembert operators (cf. [2]).
Regarding the signature decorrelation and Haar-Type inversion, we deal with
a construction of physically relevant wavelets based on the regularization of the
fundamental solution for the decomposition of gravimetric data, and analyze dif-
ferent examples occurring in exploration. The decomposition of the 3D Marmousi
density model shows a breakdown of the signals into their constituent components.
Our numerical tests have further shown that the inversion technique described for
the local improvement of records is numerically stable. In particular, the method

can be restricted to the specified local region of interest without changing the re-
maining area due to the local support of the wavelets. The resulting linear equation
systems are sparse, hence, they can thus be solved efficiently and economically.
The particular advantage of the decorrelation method proposed here is the
simultaneous calculation of the potential and the contrast function (density func-
tion) without any requirement of additional mathematical and numerical effort
and this while closely ensuring physical relevance and numerically acceptable ef-
fort. It is therefore expected that the method presented here, in fact, will contribute
substantially to minimizing the exploration risk, for example, in geothermal obli-
gations by providing deeper and more secure geological information.
Acknowledgment
The authors thank the “Federal Ministry for Economic Affairs and Energy, Berlin”
and the “Project Management Jülich” for funding the projects “GEOFÜND”
(funding reference number: 0325512A, PI Prof. Dr. W. Freeden, University of
Kaiserslautern, Germany) and “SPE” (funding reference number: 0324061, PI
Prof. Dr. W. Freeden, CBM – Gesellschaft für Consulting, Business und Man-
agement mbH, Bexbach, Germany, corporate manager Prof. Dr. mult. M. Bauer).
References
[1] Bauer, M., Freeden, W., Jacobi, H., and Neu, T., Eds. (2014). Handbuch Tiefe
Geothermie. Springer, Berlin, Heidelberg.
[2] Blick, C. (2015). Multiscale Potential Methods in Geothermal Research: Decorrelation
Reflected Post-Processing and Locally Based Inversion. Ph.D.-thesis, Geomathemat-
ics Group, University of Kaiserslautern.
[3] Blick, C., Freeden, W., and Nutz, H. (2016). Feature Extraction of Geological Sig-
natures by Multiscale Gravimetry. GEM Int. J. Geomath., 8:57–83, 2017.
[4] Bruns, E.H. (1878). Die Figur der Erde. Königl. Preussisch. Geodätisches Institut,
P. Stankiewicz Buchdruckerei, Berlin.
[5] Ernstson, K. (2018). Geologische und geophysikalische Untersuchungen. In Bauer,
M., Freeden, W., Jacobi, H., and Neu, T., Eds., Handbuch Oberflächennahe Geother-
mie. Springer, Heidelberg.
[6] ESA (1996). The Nine Candidate Earth Explorer Missions. Publications Division
ESTEC, Noordwijk, SP-1196(1).
[7] ESA (1998). European Views on Dedicated Gravity Field Missions: GRACE and
GOCE. An Earth Sciences Division Consultation Document. ESD-MAG-REP-CON-
001.
[8] ESA (1999). Gravity Field and Steady-State Ocean Circulation Mission. Publications
Division ESTEC, Noordwijk, SP-1233(1). Report for mission selection of the four
candidate earth explorer missions.
[9] Fehlinger, T. (2009). Multiscale Formulations for the Disturbing Potential and the
Deflections of the Vertical in Locally Reflected Physical Geodesy. Ph.D.-thesis, Geo-
mathematics Group, University of Kaiserslautern.
[10] Foulger, G., Natland, J., Presnall, D., and Anderson, D., Eds. (2005). Plates, Plumes,
and Paradigms. Geological Society of America.
[11] Freeden, W. (1980). On Integral Formulas of the (Unit) Sphere and Their Applica-
tion to Numerical Computation of Integrals. Computing, 25:131–146.
[12] Freeden, W. (1999). Multiscale Modelling of Spaceborne Geodata. B.G. Teubner,
Leipzig.
[13] Freeden, W. and Blick, C. (2013). Signal Decorrelation by Means of Multiscale Meth-
ods. World of Mining, 65(5):304–317.
[14] Freeden, W., Fehlinger, T., Mayer, C., and Schreiner, M. (2008). On the Local Mul-
tiscale Determination of the Earth’s Disturbing Potential from Discrete Deflections
of the Vertical. Comput. Geosc., 4:473–490.
[15] Freeden, W. and Gerhards, C. (2013). Geomathematically Oriented Potential Theory.
CRC Press, Taylor & Francis, Boca Raton.
[16] Freeden, W., Gervens, T., and Schreiner, M. (1998). Constructive Approximation
on the Sphere (with Applications to Geomathematics). Oxford Science Publications,
Clarendon, Oxford.
[17] Freeden, W., Glockner, O., and Schreiner, M. (1998). Spherical Panel Clustering and
Its Numerical Aspects. J. Geodesy, 72:586–599.
[18] Freeden, W. and Gutting, M. (2013). Special Functions of Mathematical (Geo-)
Physics. Applied and Numerical Harmonic Analysis (ANHA). Birkhäuser, Basel.
[19] Freeden, W. and Hesse, K. (2002). On the Multiscale Solution of Satellite Problems
by use of Locally Supported Kernel Functions Corresponding to Equidistributed
Data on Spherical Orbits. Stud. Sci. Math. Hung., 39:37–74.
[20] Freeden, W. and Michel, V. (2004). Multiscale Potential Theory (with Applications
to Geoscience). Birkhäuser, Boston.
[21] Freeden, W. and Schreiner, M. (2006). Local Multiscale Modelling of Geoid Undu-
lations From Deflections of the Vertical. J. Geodesy, 79:641–651.
[22] Freeden, W. and Schreiner, M. (2009). Spherical Functions of Mathematical Geo-
sciences (A Scalar, Vectorial, and Tensorial Setup). Springer, Heidelberg.
[23] Freeden, W. and Wolf, K. (2008). Klassische Erdschwerefeldbestimmung aus der
Sicht moderner Geomathematik. Math. Semesterb., 56:53–77.
[24] Grafarend, E.W., Aardalan, A.A., and Finn, G. (2006). Ellipsoidal Vertical Deflec-
tions and Ellipsoidal Gravity Disturbances: Case Studies. Stud. Geophys. Geod.,
50:1–57.
[25] Grafarend, E.W., Klapp, M., and Martinec, Z. (2015). Spacetime Modeling of the
Earth’s Gravity Field by Ellipsoidal Harmonics. In Freeden, W., Nashed, M.Z., and
Sonar, T., Eds., Handbook of Geomathematics, volume 1, 381–496. Springer, New
York, 2nd edition.
[26] Grant, E.S. and West, G.F. (1965). Interpolation Theory in Applied Geophysics.
McGraw-Hill Book Company.
[27] Groten, E. (1979). Geodesy and the Earth’s Gravity Field I + II. Dümmler, Bonn.
[28] Gutting, M. (2008). Fast Multipole Methods for Oblique Derivative Problems. Ph.D.-
thesis, Geomathematics Group, University of Kaiserslautern.
[29] Gutting, M. (2012). Fast Multipole Accelerated Solution of the Oblique Derivative
Boundary-Value Problem. GEM Int. J. Geomath., 2:223–252.
[30] Gutting, M. (2015). Fast Spherical/Harmonic Spline Modeling. In Freeden, W.,
Nashed, M.Z., and Sonar, T., Eds., Handbook of Geomathematics, volume 3,2711–
2746. Springer, New York, 2nd. edition.
[31] Harpp, K.S., Fornari, D.J., Geist, D.J., and Kurz, M.D. (2003). A Manifestation
of Plume-Ridge Interaction in the Northern Galapagos Islands. Geochemistry, Geo-
physics, Geosystems, 4(9):DOI: 10.1029/2003GC000531.
[32] Heiskanen, W.A. and Moritz, H. (1967). Physical Geodesy. Freeman, San Francisco.
[33] Helmert, F. (1884). Die Mathematischen und Physikalischen Theorien der Höheren
Geodäsie, I, II. B.G. Teubner, Leipzig.
[34] Hesse, K. (2003). Domain Decomposition Methods in Multiscale Geopotential Deter-
mination from SST and SGG. Ph.D.-thesis, Geomathematics Group, University of
Kaiserslautern.
[35] Hesse, K., Sloan, I.H., and Womersley, R.S. (2015). Numerical Integration on the
Sphere. In Freeden, W., Nashed, M.Z., and Sonar, T., Eds., Handbook of Geomath-
ematics, volume 2, 2671–2710. Springer, New York, 2nd edition.
[36] Hofmann-Wellenhof, B. and Moritz, H. (2005). Physical Geodesy. Springer, Wien,
New York.
[37] IEA (2012). Iceland Country Report 2012. Report, IEA Geothermal.
[38] Jacobs, F. and Meyer, H. (1992). Geophysik-Signale aus der Erde. B.G. Teubner,
Leipzig, and VDF Verlag, Zürich.
[39] Keating, P. (1992). Interpretation of the Gravity Anomaly Field in the Noranda
Val d’Or region, Abitibi Greenstone Belt, Canadian Shield. Can. J. Earth Sciences,
29(5):962–971.
[40] Kellogg, O.D. (1967). Foundations of Potential Theory. Springer Berlin, Heidelberg,
New York.
[41] Kusche, J. (2015). Time-Variable Gravity Field and Global Deformation of the
Earth. In Freeden, W., Nashed, M.Z., and Sonar, T., Eds., Handbook of Geomathe-
matics, volume 1, 321–338. Springer, New York, 2nd edition.
[42] Martin, G.S., Wiley, R., and Marfurt, K.J. (2006). Marmousi2: An Elastic Upgrade
for Marmousi. The Leading Edge, 25(2):156–166.
[43] Mathar, D. (2008). Wavelet Variances and Their Application in Geoscientifically
Relevant Feature Detection. Diploma thesis, University of Kaiserslautern, Geomath-
ematics Group.
[44] Meinesz, F.A.V. (1928). A Formula Expressing the Deflection of the Plumb Line in
the Gravity Anomalies and Some Formulas for the Gravity Field and the Gravity
Potential Outside the Geoid. Proc. Koninklijke Akad. Wet. Amsterdam, 31:315–322.
[45] Meissl, P.A. (1971). A Study of Covariance Functions Related to the Earth’s Dis-
turbing Potential. Scientific Report No. 151, The Ohio State University, Department
of Geodetic Science, Columbus.
[46] Michel, V. (2002). A Multiscale Approximation for Operator Equations in Separable

Hilbert Spaces – Case Study: Reconstruction and Description of the Earth’s Interior.
Habilitation thesis, Geomathematics Group, University of Kaiserslautern. Shaker,
Aachen.
[47] Michel, V. and Fokas, A.S. (2008). A Unified Approach to Various Techniques for the
Non-Uniqueness of the Inverse Gravimetric Problem and Wavelet-Based Methods.
Inverse Problems, 24 045019 (25 pp.), DOI: 10.1088/0266-5611/24/4/045019.
[48] Militzer, H. and Weber, F. (1984). Angewandte Geophysik, Band 1, Gravimetrie und
Magnetik. Springer, Wien, New York.
[49] Möhringer, S. (2014). Decorrelation of Gravimetric Data. Ph.D.-thesis, Geomathe-
matics Group, University of Kaiserslautern.
[50] Morgan, W.J. (1971). Convection Plumes in the Lower Mantle. Nature, 230:42–43.
[51] Moritz, H. (2015). Classical Physical Geodesy. In Freeden, W., Nashed, M.Z., and
Sonar, T., Eds., Handbook of Geomathematics, volume 1, 253–289. Springer, New
York, 2nd. edition.
[52] Nations, U. (2010). Current and Planned Global and Regional Navigation Satellite
Systems and Satellite-Based Augmentations Systems. In Proceedings of ICG.
[53] Nettleton, L.L. (1971). Elementary Gravity and Magnetics for Geologists and Seis-
mologists. Society Of Exploration Geophysicists, Tulsa, OK.
[54] Nettleton, L.L. (1976). Gravity and Magnetics in Oil Prospecting. McGraw-Hill, New
York.
[55] Neumann, F. (1887). Vorlesungen über die Theorie des Potentials und der Kugel-
funktionen. B.G. Teubner, Leipzig.
[56] Pizzetti, P. (1894). Geodesia – sulla espressione della gravità alla superficie del
geoide, supposto ellissoidico. Att. R. Acad. Lincei, 3:331–350.
[57] Pizzetti, P. (1909). Corpi equivalenti rispetto alla attrazione newtoniana esterna.
Rom. Acc. L. Rend., 18:211–215.
[58] Ritter, J. and Christensen, U. (2007). Mantle Plumes, A Multidisciplinary Approach.
Springer, Berlin, Heidelberg.
[59] Rivas, J. (2009). Gravity and Magnetic Methods. In Short Course on Surface Explo-
ration for Geothermal Resources.
[60] Rummel, R. (1997). Spherical Spectral Properties of the Earth’s Gravitational Po-
tential and Its First and Second Derivatives. In Sansò, S. and Rummel, R., Eds.,
Geodetic Boundary Value Problems in View of the One Centimeter Geoid, volume 65
of Lecture Notes in Earth Science, 359–404. Springer, Berlin, Heidelberg.
[61] Rummel, R. (2015). GOCE: Gravitational Gradiometry in a Satellite. In Freeden,
W., Nashed, M.Z., Eds., Handbook of Geomathematics, volume 1, 211–226. Springer,
New York, 2nd edition.
[62] Rummel, R., Balmino, G., Johannessen, J., Visser, P., and Woodworth, P. (2002).
Dedicated Gravity Field Missions – Principles and Aims. J. Geodyn., 33:3–20.
[63] Schuber, D., Turcotte, D., and Olson, P. (2001). Mantle Convection in the Earth and
Planets. Cambridge University Press, Cambridge.
[64] Seigel, H.O. (1995). A Guide to High Precision Land Gravimeter Surveys. Scintrex
Limited, Concord, Ontario.
[65] Stokes, G.G. (1849). On the Variation of Gravity at the Surface of the Earth. Trans.
Cambr. Phil. Soc., 148:672–712.
[66] Stokes, G.G. (1867). On the internal distribution of matter which shall produce a
given potential at the surface of a gravitating mass. Proc.Royal Soc. London, 15:482–
486.
[67] Symes, W.W. (2014). T.R.I.P. the Rice inversion project, Department of Compu-
tational and Applied Mathematics, Rice University, Houston, Texas, USA. URL:
http://www.trip.caam.rice.edu/downloads/downloads.html,
accessed 12. September 2016.
[68] Torge, W. (1989). Gravimetry. de Gruyter, Berlin.
[69] Torge, W. (1991). Geodesy. de Gruyter, Berlin.
[70] Tykhonov, A.N. (1963). Solution of Incorrectly Formulated Problems and the Reg-
ularization Method. Sov. Math., 5:1035–1038. (translation from Dokl. Akad. Nauk,
(15):501–504).
[71] Wikipedia (2016). Galápagos hotspot. URL: https://en.wikipedia.org/wiki/
Galapagos hotspot, accessed 12. May 2016.
[72] Wilson, J. (1963). A Possible Origin of the Hawaiian Island. Can. J. Phys., 41:863–
868.
[73] Wolf, K. (2009). Multiscale Modeling of Classical Boundary Value Problems in Phys-
ical Geodesy by Locally Supported Wavelets. Ph.D.-thesis, Geomathematics Group,
University of Kaiserslautern.


Spherical Harmonics Based

Special Function Systems and
Constructive Approximation Methods
Willi Freeden, Volker Michel, and Frederik J. Simons
Abstract. Special function systems are reviewed that reflect particular prop-
erties of the Legendre polynomials, such as spherical harmonics, zonal kernels,
and Slepian functions. The uncertainty principle is the key to their classifi-
cation with respect to their localization in space and frequency/momentum.
Methods of constructive approximation are outlined such as spherical har-
monic and Slepian expansions, spherical spline and wavelet concepts. Regu-
larized Functional Matching Pursuit is described as an approximation tech-
nique of combining heterogeneous systems of trial functions to a kind of a
‘best basis’.
Keywords. Spherical harmonics procedures, Slepian, spline and wavelet meth-
ods, regularized functional matching pursuit.
1. Introduction
Up until the present time, modeling geoscientific data is often performed on a
global scale by orthogonal expansions in terms of spherical harmonics. However,
in many aspects global spherical harmonic modeling cannot keep pace with the
prospects and the expectations of the ‘Earth system sciences’. In particular, there
is an increasing need for high-precision modeling on local areas. As we shall discuss,
Slepian functions are important tools for this purpose. For their part, zonal kernel
functions – in the jargon of constructive approximation: radial basis functions –
have become more and more important because of their space localizing properties
(even in the vectorial and tensorial context).
The addition theorem for spherical harmonics enables us to express all types
of zonal kernel functions in terms of a one-dimensional function, the Legendre poly-
nomial. Weighted additive clustering of Legendre polynomials generates specific
classes of space localizing zonal kernel functions, i.e., Legendre series expansions,
754 W. Freeden, V. Michel, and F.J. Simons
ready for approximation within the scalar, vectorial, and tensorial framework. The
closer the Legendre series expansion is to the Dirac kernel, the more localized is the
zonal kernel in space, and the more economical is its role in (spatial) local computa-
tion. In addition, the Funk–Hecke formula provides the natural tool for establishing
convolutions of spherical fields against zonal kernels. Consequently, by specifying
Dirac families, i.e., sequences of zonal functions tending to the Dirac kernel, (space-
localized) filtered versions of (square-integrable) spherical fields are obtainable by
convolution, leading to ‘zooming-in’, multiscale approximations. Altogether, the
Legendre polynomial is the keystone of any work about special functions in the
mathematical geosciences. It enables the transition from spherical harmonics via
zonal kernels up to the Dirac kernel. The Funk–Hecke formula and its consequences
for spherical convolutions open new methodological perspectives for global as well
as local approximation in scalar, vectorial and tensorial applications.
In this paper, we discuss selected systems of trial functions on the sphere with
a brief excursion to basis functions on the ball. These spherical function systems are
investigated with respect to their localization in space and frequency/momemtum.
Moreover, we briefly summarize a method of finding a best basis by Regularized
Functional Matching Pursuit.
2. Special function systems on sphere and ball

Because of the nearly spherical shape of the Earth, spherical functions and con-
cepts play an essential part in all of the geosciences. By a spherical variant of the
Weierstraß theorem, spherical polynomials, the spherical harmonics, approximate
continuous functions with respect to different topologies.
2.1. Spherical harmonics

Spherical harmonics are the analogues of trigonometric functions for Fourier ex-
pansion theory on the sphere. They were introduced to study gravitational theory
[61, 62]. Early publications on the theory of spherical harmonics in their original
physical interpretation as ‘multipoles’ are by Clebsch [16], Sylvester [93], Heine
[54], Neumann [77], and Maxwell [66]. Global geomagnetic data and basic spher-
ical harmonic expansions became available in the mid 1800s [50]. Today, the use
of spherical harmonics in all geosciences is well established, particularly for the
representation of scalar potentials. Reference models for the Earth’s gravitational
or magnetic fields are distributed as tables of coefficients for the spherical har-
monic expansion of their potentials. In this approach, each spherical harmonic is
a polynomial ‘ansatz-function’, corresponding to one ‘degree’, or in the jargon of
signal processing, to exactly one ‘frequency’. Thus, orthogonal (Fourier) expansion
in terms of spherical harmonics amounts to the superposition of summands with
an oscillating character determined by the degree of the Legendre polynomial (see
Table 1). The more spherical harmonics are involved in the expansion, the more
the oscillations grow in number, but the smaller are their amplitudes.
Spherical Harmonics Based Special Function Systems 755
Weierstraß approximation theorem

↓ geophysical constraint of harmonicity
spherical harmonics Yn,j as restrictions of homogeneous

harmonic polynomials Hn,j to the unit sphere Ω ⊂ R3
orthonormality and invariance

↓ addition theorem
one-dimensional Legendre polynomial Pn :

2n+1
4π
Pn (ξ · η) = Yn,j (ξ)Yn,j (η), ξ, η ∈ Ω
2n + 1 j=1
convolution by the Legendre kernel

↓ Funk–Hecke formula
Legendre transform of F :

2n + 1
(Pn ∗ F )(ξ) = Pn (ξ · η)F (η) dω(η), ξ∈Ω
4π Ω
superposition over frequencies

↓ orthogonal series expansion
Fourier series of F ∈ L2 (Ω):

∞

2n + 1
F (ξ) = Pn (ξ · η)F (η) dω(η), ξ∈Ω
n=0
4π Ω
Table 1. Fourier expansion of square-integrable scalar functions on the

sphere.
The geosciences deal with the space L2 (Ω) of square-integrable functions on

the unit sphere Ω. The quantity
1/2
F L2 (Ω) = (F (ξ))2 dω(ξ) (1)
Ω
may be understood as the energy of the ‘signal’ F ∈ L2 (Ω). The appropriate repre-
sentation of a finite-energy signal in terms of a countable Hilbert basis is one of the
most centrally important problems in the mathematical geosciences. The spherical
harmonics form a Hilbert basis in L2 (Ω). Suitable systems of spherical harmon-
ics {Yn,k }n=0,1,...; k=1,...,2n+1 are often defined by the restriction of homogeneous
harmonic polynomials to the sphere. The polynomial structure has tremendous

advantages. First, spherical harmonics of different degrees are orthogonal. Sec-
ond, the space Harmn of spherical harmonics of degree n is finite-dimensional:
dim(Harmn ) = 2n + 1. Therefore, the basis property of {Yn,k }n=0,1,...;
(∞ k=1,...,2n+1
is equivalently characterized by the completion of the direct sum n=0 Harmn , i.e.:
∞ ·L2 (Ω)
5
2
L (Ω) = Harmn . (2)
n=0
This is the canonical reason why spherical harmonic (multipole) expansions un-
derlie the classical approaches to geopotentials.
Fourier transform. More explicitly, any ‘signal’ F ∈ L2 (Ω) can be split into ‘or-
thogonal contributions’ involving the Fourier transforms F ∧ (n, k) defined by

F ∧ (n, k) = F (ξ)Yn,k (ξ) dω(ξ), (3)
Ω
2
in terms of L (Ω)-orthonormal spherical harmonics {Yn,k } n=0,1,... . Parseval’s
k=1,...,2n+1
identity identifies the spatial energy of a signal with the spectral energy, decom-
posed orthogonally into single frequency contributions
∞ 2n+1
2
F 2L2 (Ω) = F, F L2 (Ω) = (F ∧ (n, k)) .
n=0 k=1
This explains why the (global) geosciences work more often with the ‘amplitude
spectrum’ {F ∧ (n, k)} n=0,1,... than with the ‘original signal’ F ∈ L2 (Ω).
k=1,...,2n+1
Inverse Fourier transform. The ‘inverse Fourier transform’

∞ 2n+1

F = F ∧ (n, k)Yn,k (4)
n=0 k=1
allows the geoscientist to think of the function (signal) F as a sum of ‘wave func-
tions’ Yn,k corresponding to different frequencies. One can think of measurements
as operating on an ‘input signal’ F to produce an output signal G = ΛF , where Λ
is an operator acting on L2 (Ω). Fortunately, large portions of interest can be well
approximated by linear rotation-invariant pseudodifferential operators (see, e.g.,
[33, 47, 92]). If Λ is such an operator on L2 (Ω), this means that
ΛYn,k = Λ∧ (n)Yn,k , n = 0, 1, . . . ; k = 1, . . . , 2n + 1, (5)
∧
where the ‘symbol’ {Λ (n)}n∈N0 is a sequence of real values (independent of the
order k). Thus, we have the fundamental fact that the spherical harmonics are
the eigenfunctions of the operator Λ. Different pseudodifferential operators Λ are
characterized by their eigenvalues Λ∧ (n). All eigenvalues {Λ∧ (n)}n∈N0 are collected
in the so-called symbol of Λ. The ‘amplitude spectrum’ {G∧ (n, k)} of the response
of Λ is described in terms of the amplitude spectrum of functions (signals) by a
simple multiplication by the ‘transfer’ Λ∧ (n).
Bandlimited/spacelimited functions. Physical devices do not transmit spherical

harmonics of arbitrarily high frequency without severe attenuation. The ‘transfer’
Λ∧ (n) usually tends to zero with increasing n. It follows that the amplitude spec-
tra of the responses (observations) to functions (signals) of finite energy are also
negligibly small beyond some finite frequency. Thus, both because of the frequency
limiting nature of the devices used, and because of the nature of the ‘transmitted
signals’, the geoscientist is soon led to consider bandlimited functions. These are
the functions F ∈ L2 (Ω) whose ‘amplitude spectra’ vanish for all n > N (N ∈ N
fixed). In other words, each bandlimited function F ∈ L2 (Ω) can be written as a fi-
N 2n+1
nite Fourier series. So, any function F of the form F = n=0 k=1 F ∧ (n, k)Yn,k
is said to be bandlimited with the band N , if F ∧ (N, k) = 0 for at least one k.
In analogous manner, F ∈ L2 (Ω) is said to be locally supported (spacelimited)
with spacewidth ρ around an axis η ∈ Ω, if for some ρ ∈ (−1, 1) the function
F vanishes on the set of all ξ ∈ Ω with −1 ≤ ξ · η ≤ ρ (where ρ is the largest
number for which this is the case). Bandlimited functions are infinitely often dif-
ferentiable everywhere. Moreover, it is clear that any bandlimited function F is an
analytic function. From the analyticity, it follows immediately that a non-trivial
bandlimited function cannot vanish on any (non-degenerate) subset of Ω. The only
function that is both bandlimited and spacelimited is the zero function.
In addition to bandlimited but non-spacelimited functions, numerical analysis
would like to deal with spacelimited functions. However, as we have seen, such a
function (signal) of finite (space) support cannot be bandlimited, it must contain
spherical harmonics of arbitrarily large frequencies. Thus, there is a dilemma of
seeking functions that are somehow concentrated in both space and frequency
(more accurately, angular momentum domain). There is a way of mathematically
expressing the impossibility of simultaneous confinement of a function to space
and angular momentum, namely the uncertainty principle.
2.2. Zonal kernel functions

To understand the transition from the theory of spherical harmonics through zonal
kernel functions to the Dirac kernel, we have to realize the relative advantages
of the classical Fourier expansion method by means of spherical harmonics, and
this not only in the frequency domain, but also in the space domain. It is char-
acteristic for Fourier techniques that the spherical harmonics as polynomial trial
functions admit no localization in space domain, while in the frequency domain
(or: angular momentum domain), they always correspond to exactly one degree,
i.e., frequency, and therefore, are said to show ideal frequency localization. Because
of the ideal frequency localization and the simultaneous absence of space localiza-
tion, in fact, local changes of fields (signals) in the space domain affect the whole
table of orthogonal (Fourier) coefficients. This, in turn, causes global changes of
the corresponding (truncated) Fourier series in the space domain. Nevertheless,
ideal frequency localization is often helpful for meaningful physical interpretations
by relating the different observables of a geopotential to each other at a fixed
frequency.
Taking these aspects on spherical harmonic modeling by Fourier series into

account, trial functions which simultaneously show ideal frequency localization as
well as ideal space localization would be a desirable choice. In fact, such an ideal
system of trial functions would admit models of highest spatial resolution which
were expressible in terms of single frequencies. However, from the uncertainty
principle – the connection between space and frequency localization – we will see
that both characteristics are mutually exclusive.
In conclusion, Fourier expansion methods are well suited to resolve low and
medium frequency phenomena, i.e., the ‘trends’ of a signal, while their applica-
tion to obtain high resolution in global or local models is critical. This difficulty
is also well known to theoretical physics, e.g., when describing monochromatic
electromagnetic waves or considering the quantum-mechanical treatment of free
particles. There, plane waves with fixed frequencies (ideal frequency localization,
no space localization) are the solutions of the corresponding differential equations,
but they do certainly not reflect the physical reality. As a remedy, plane waves
of different frequencies are superposed into ‘wave-packages’ that gain a certain
amount of space localization, while losing their ideal spectral localization. In a
similar way, a suitable superposition of polynomial functions leads to so-called
zonal kernel functions, in particular to kernel functions with a reduced frequency,
but increased space localization.
More concretely, any kernel function K : Ω × Ω → R that is characterized by
the property that there exists a function K̃ : [0, 2] → R such that
#
K(ξ, η) = K̃(|ξ − η|) = K̃ 2 − 2ξ · η = K̂(ξ · η), ξ, η ∈ Ω, (6)
is called a (spherical) radial basis function (at least in the theory of constructive
approximation).
Zonal kernels. The application of a rotation (i.e., a 3 × 3 ‘orthogonal’ matrix t

with tT = t−1 ) leads to K(tξ, tη) = K̂((tξ) · (tη)) = K̂(ξ · (tT tη)) = K̂(ξ · η) =
K(ξ, η). In particular, a rotation around the axis ξ ∈ Ω (i.e., tξ = ξ) yields
K(ξ, η) = K(ξ, tη) for all η ∈ Ω. Hence, K(ξ, ·) possesses a rotational symmetry
with respect to the axis ξ. In the theory of special functions of mathematical
physics, a kernel K̂ : Ω × Ω → R satisfying K̂(ξ · η) = K̂(tξ · tη), ξ, η ∈ Ω, for
all orthogonal transformations t is known as a zonal kernel function. To highlight
the reducibility of K̂ to a function defined on the interval [−1, 1], the notation
(ξ, η) → K̂(ξ · η), (ξ, η) ∈ Ω × Ω, is used throughout this chapter (see also (6)).
From the theory of spherical harmonics we get a representation of any L2 (Ω)-
zonal kernel function K in terms of a Legendre expansion
∞
2n + 1 ∧
K(ξ·) = K (n)Pn (ξ·) (7)
n=0
4π
(in the · L2 (Ω) -sense), where the sequence {K ∧ (n)}n∈N0 given by

1
K ∧ (n) = 2π K(t)Pn (t) dt (8)
−1
is called the Legendre symbol of the zonal kernel K(ξ·). A simple but extreme
example (with optimal frequency localization and no space localization) is the
Legendre kernel where K ∧ (n) = 1 for one particular n and K ∧ (m) = 0 for m = n,
i.e., the Legendre kernel is given by
2n + 1
Ω × Ω (ξ, η) → Pn (ξ · η).
4π
In other words, additive clustering of weighted Legendre kernels generates zonal
kernel functions. It is of importance to distinguish bandlimited kernels (i.e.,
K ∧ (n) = 0 for all n ≥ N ) and non-bandlimited ones, for which infinitely many
numbers K ∧ (n) do not vanish. Non-bandlimited kernels show a much stronger
space localization than their bandlimited counterparts. Empirically, if K ∧ (n) ≈
K ∧ (n + 1) ≈ 1 for many successive large integers n, then the support of the se-
ries (7) in the space domain is small, i.e., the kernel is spacelimited (i.e., in the
jargon of approximation theory ‘locally supported’). This leads to the other ex-
tremal kernel (in contrast to the Legendre kernel) which is the Dirac kernel with
optimal space localization but no frequency localization and K ∧ (n) = 1 for all n,
where, however, the Dirac kernel does not exist as a classical function in the math-
ematical sense. Nevertheless, it is well known that, if we have a family of kernels
{KJ }J=0,1,... where limJ→∞ KJ∧ (n) = 1 for each n and an additional (technical)
condition holds, then KJ ∗ F tends to F in the sense of L2 (Ω) for all F ∈ L2 (Ω).
Assuming limn→∞ K ∧ (n) = 0, necessary to get a ‘proper’ function, the slower
the sequence {K ∧ (n)}n=0,1,... converges to zero, the lower the frequency localiza-
tion, and the higher the space localization. A unified scheme is found in Table 2.
Zonal kernel function theory relies on the following principles:
(i) Weighted Legendre kernels are the summands of zonal kernel functions.
(ii) The Legendre kernel is ideally localized in frequency. The Dirac kernel is
ideally localized in space.
(iii) The only frequency- and spacelimited zonal kernel is the zero function.
Legendre Dirac
zonal kernels
kernels kernel
general case
bandlimited spacelimited
Table 2. From Legendre kernels via zonal kernels to the Dirac kernel
2.3. Slepian functions

As we have seen, F 2L2 (Ω) is the energy of a certain function F ∈ L2 (Ω). Suppose
now that there is a particular region C ⊂ Ω, and let us define the ‘local’ energy of
that function as F 2L2 (C) . Functions F that are band limited,
N 2n+1

F = F ∧ (n, k)Yn,k (9)
n=0 k=1
cannot also be spacelimited, but they can be spaceconcentrated.
Bandlimited/spaceconcentrated Slepian functions. By maximizing the spatial en-

ergy ratio
F 2L2 (C)
λC (F ) = , 0 < λC (F ) < 1, (10)
F 2L2 (Ω)
we obtain bandlimited spherical ‘Slepian functions’ [56, 67, 87], named in analogy
with the prolate spheroidal wave functions of Slepian [90]. They are not, in general,
zonal functions.
The Fourier coefficients of the Slepian functions are the (N + 1)2 orthogonal
eigenvectors of the symmetric concentration matrix whose elements are the limited-
domain inner-product terms Ym,j , Yn,k L2 (C) , 0 ≤ m, n ≤ N , i.e.,
N 2n+1

Ym,j , Yn,k L2 (C) F ∧ (n, k) = λC (F )F ∧ (m, j). (11)
n=0 k=1
We will give their associated eigenvalues superscripted labels and rank them in
(1) (α) ((N +1)2 )
decreasing order of concentration, 1 > λC (F ) ≥ λC (F ) ≥ λC (F ) > 0. The
bandlimited Slepian functions can alternatively be obtained by solving a Fredholm
integral equation with a ‘Shannon’ concentration kernel:
N
2n + 1
Pn (ξ · η)F (η) dω(η) = λC (F )F (ξ), ξ ∈ Ω. (12)
n=0
4π
C
Spacelimited/bandconcentrated Slepian functions. We can define spacelimited

Slepian functions which are bandconcentrated. They are obtained by the restriction
of the bandlimited Slepian functions F to the region of interest C, or, equivalently,
their Fourier coefficients are
N 2n+1

Ym,j , Yn,k L2 (C) F ∧ (n, k), (13)
n=0 k=1
extending the F ∧ (n, k) to all degrees m = 0, 1, . . . , ∞ and order indices j =

1, . . . , 2m + 1.
A central concept is the effective dimension of functions that are ‘essentially’

space- and bandlimited. The Shannon number is the trace of the concentration
operators in (11)–(12), given by (using the addition theoremin 14b–14c),
(N +1)2
(α)
NC = λC (F ) (14a)
α=1
N 2n+1

= Yn,k , Yn,k L2 (C) (14b)
n=0 k=1
N
2n + 1
= Pn (η · η) dω(η) (14c)
n=0
4π
C

2 C dω(η)
= (N + 1) . (14d)
4π
(α)
The eigenvalue spectrum λC (F ) has a characteristic step-like shape, with the
(N +1)2 (α) NC (α)
property α=1 λC (F ) ≈ α=1 λC (F ) revealing that NC will be close to
the number of Slepian functions that usefully contribute to the approximation of
arbitrary target functions on domains C ⊂ Ω.
While computation can be carried out via either (11) or (12), when the region
of interest C is a spherical cap (one whose boundary ∂C is a circle and whose half-
opening angle is Θ), the integral equation (12) commutes with a Sturm–Liouville
differential equation whose spectral-domain representation has an extremely sim-
ple analytical form, rendering the computation of Slepian functions of domains
essentially trivial [51]. In that case, the Slepian functions degenerate to being
the solutions of fixed-order (j) versions of equation (11), with a partial Shannon
number given in terms of products of the associated Legendre functions and their
derivatives (primed), namely

(N − j + 1)! 1 0

1
Nj = P (t)PN,j (t) − PN,j (t)PN +1,j (t) dt. (15)
2(N + j)! cos Θ N +1,j
Only on circularly symmetric domains and when the spherical-harmonic order
j = 0 are the Slepian functions zonal, and in that case, the fixed-order partial
Shannon number is well approximated by Wieczorek and Simons [104] as
√
NC Θ
N0 ≈ 2 ≈ (N + 1) . (16)
π π
Figure 1 shows examples of spherical-cap Slepian functions, their power spectra,
and their eigenvalue spectra.
When the concentration domain is a spherical cap, the best-concentrated
(highest-eigenvalue) bandlimited Slepian function is a zonal function that is close
to optimally localized under the uncertainty principle (see Section 3). All the lower-
eigenvalue zonal Slepian functions, and finally, all the non-zonal Slepian functions,
together form a complete orthonormal basis for the space of functions on the unit
Slepian functions power spectra (dB) eigenvalue

0 1
2 j=0
−20
0 0.5
−40
−2
0
−60
1 N1 19
0 1
2 j=1
−20
0 0.5
−40
−2
0
−60
2 N2 18
0 1
2 j=2
−20
0 0.5
−40
−2
0
−60
3 N3 17
0 1
2 j=3
−20
0 0.5
−40
−2
0
−60
0° 40° 0 3 18 N4 16
colatitude degree rank
Figure 1. The four best-concentrated (in decreasing gray shading)

fixed-order (top to bottom, j = 0, 1, 2, 3) Slepian functions and their
power spectra, for a common bandwidth N = 18, with the domain C a
40◦ spherical cap. Also shown are the complete eigenvalue spectra with
the fixed-order (partial) Shannon numbers indicated. Compare to [87]
(their Figs. 5.1, 5.2 and 5.3.).
sphere Ω that are bandlimited to N . The partial Shannon numbers Nj sum to the
full Shannon number NC via
2N
+1
NC = Nj . (17)
j=0
2.4. From the scalar to the vector and tensor context

In the second half of the last century, a physically motivated approach for the de-
composition of spherical vector and tensor fields was presented based on a spherical
variant of the Helmholtz theorem, e.g., [6–8, 75]. Following this concept, the tan-
gential part of a spherical vector field is split up into a curl-free and a divergence-
free field by use of two differential operators, viz. the surface gradient and the
surface curl gradient. Of course, an analogous splitting is valid in tensor theory.
scalar
Legendre → vector
Legendre → tensor
Legendre
↓ ↓ ↓
scalar
zonal → vector
zonal → tensor
zonal
↓ ↓ ↓
scalar
Dirac → vector
Dirac → tensor
Dirac
Table 3. From scalar via vectorial to tensorial kernels
In subsequent publications during the second half of the last century, how-
ever, the vector spherical harmonic theory was usually written in local coordinate
expressions that make mathematical formulations lengthy and hard to read. Ten-
sor spherical harmonic settings were even more difficult to understand. In addition,
when using local coordinates within a global spherical concept, differential geome-
try tells us that there is no representation of vector and tensor spherical harmonics
that is free of singularities. As a consequence, vector and tensor spherical harmon-
ics have suffered from an inadequately complex and inconsistent literature. Absent
coordinate-free explicit formulas, the orthogonal invariance based on specific vec-
tor/tensor extensions of the Legendre polynomials was not worked out suitably in
a unifying scalar/vector/tensor framework, nor was the concept of zonal (kernel)
functions adequately generalized to the spherical vector/tensor case.
All new structures for spherical functions in mathematical (geo)physics were
developed by Freeden and Schreiner [43] and Freeden and Gutting [32]. Two funda-
mental transitions underlie their approach: one from spherical harmonics via zonal
kernel functions to the Dirac kernels, and the other one from scalar to vector and
tensor theory (see Table 3).
Helmholtz decomposition of spherical vector/tensor fields. To explain the tran-

sition from the theory of scalar spherical harmonics to its vectorial and tensorial
extensions, Freeden and Schreiner [43] start from physically motivated dual pairs of
operators (the reference space being always the space of signals with finite energy,
i.e., the space of square-integrable fields). The pair o(i) , O(i) , i ∈ {1, 2, 3}, origi-
nates in the ingredients of the Helmholtz decomposition of a vector field, while
o(i,k) , O(i,k) , i, k ∈ {1, 2, 3}, take the analogous role for the Helmholtz decomposi-
tion of tensor fields (see, e.g., [7, 43, 47]).
For example, in vector theory, o(1) F is the normal field
(1)
ξ → oξ F (ξ) = F (ξ)ξ, ξ ∈ Ω,
whereas o(2) F is the surface gradient field

(2)
ξ → oξ F (ξ) = ∇∗ξ F (ξ), ξ ∈ Ω,
and o(3) F is the surface curl gradient field

(3)
ξ → oξ F (ξ) = L∗ξ F (ξ), L∗ξ = ξ ∧ ∇∗ξ , ξ ∈ Ω,
applied to a scalar function F . In addition, O(1) f is the normal component

(1)
ξ → Oξ f (ξ) = f (ξ) · ξ, ξ ∈ Ω,
while O(2) f is the negative surface divergence

(2)
ξ → Oξ f (ξ) = −∇∗ξ · f (ξ), ξ ∈ Ω,
and O(3) f is the negative surface curl

(3)
ξ → Oξ f (ξ) = −L∗ξ · f (ξ), ξ ∈ Ω,
taken over a vector-valued function f .
Clearly, the operators o(i,k) , O(i,k) are also definable in orientation to the
tensor Helmholtz decomposition theorem (for reasons of simplicity, however, their
explicit description is omitted here). The pairs o(i) , O(i) and o(i,k) , O(i,k) of dual
operators lead us to an associated palette of Legendre kernel functions, all of them
generated by the classical one-dimensional Legendre polynomial Pn of degree n.
To be more specific, three types of Legendre kernels occur in the vectorial as well
as tensorial context (see Table 4).
The Legendre kernels o(i) Pn and o(i) o(i) Pn pertain to the vector approach
for spherical harmonics, whereas o(i,k) Pn and o(i,k) o(i,k) Pn , i, k = 1, 2, 3, form the
analogues in tensorial theory. Corresponding to each Legendre kernel, we are led
to two variants for representing square-integrable fields by orthogonal (Fourier)
expansion, where the reconstruction – as in the scalar case – is undertaken by
superposition over all frequencies.
In a unified notation, the formalism for vector/tensor spherical harmonic
theory is based on the following principles (cf. [43]):
(i) The vector/tensor spherical harmonics involving the o(i) , o(i,k) -operators, re-
spectively, are obtainable as restrictions of three-dimensional homogeneous
harmonic vector/tensor polynomials, respectively.
(ii) The vector/tensor Legendre kernels are obtainable as the outcome of sums
extended over a maximal orthonormal system of vector/tensor spherical har-
monics of degree (frequency) n, respectively.
Scalar Legendre polynomial:

(i,i) (i,k)
O(i) O(i) pn O(i,k) O(i,k) Pn
Pn = (i)
= (i,k)
μn μn
application
of o(i)
↓↑ application
of O(i)
application
of o(i,k)
↓↑ application
of O(i,k)
vector Legendre kernel tensor Legendre kernel (order 2)
(i,i) (i,k)
o(i) Pn O(i) pn o(i,k) Pn O(i,k) Pn
p(i)
n = 1/2 = 1/2 p(i,k)
n = 1/2 = 1/2
(i) (i) (i,k) (i,k)
μn μn μn μn
application
of o(i)
↓↑ application
of O(i)
application
of o(i,k)
↓↑ application
of O(i,k)
tensor Legendre kernel (order 2) tensor Legendre kernel (order 4)
(i) (i,k)
o(i) pn o(i) o(i) Pn o(i,k) pn o(i,k) o(i,k) Pn
p(i,i)
n = 1/2 = (i)
P(i,k,i,k)
n = 1/2 = (i,k)
(i) μn (i,k) μn
μn μn
vectorial context tensorial context
Table 4. Legendre scalar, vectorial, and tensorial kernel functions.
(iii) The vector/tensor Legendre kernels are zonal kernel functions, rotation-invar-
iant (in vector/tensor sense, respectively) with respect to orthogonal trans-
formations (leaving one point of the unit sphere Ω fixed).
(iv) Spherical harmonics of degree (frequency) n form an irreducible subspace of
the reference space of (square-integrable) fields on Ω.
(v) Each Legendre kernel implies an associated Funk–Hecke formula that de-
termines the constituting features of the convolution (filtering) of a square-
integrable field against the Legendre kernel.
(vi) The orthogonal Fourier expansion of a square-integrable field is the sum of
the convolutions of the field against the Legendre kernels being extended over
all frequencies.
To summarize, the theory of spherical harmonics provides us with a frame-
work to unify, review and supplement the different approaches in real scalar, vector,
and tensor theory. The essential tools are the Legendre functions, used in orthog-
onal Fourier expansions and endowed with rotational invariance. The coordinate-
free construction yields a number of formulas and theorems that previously were
derived only in coordinate (e.g., polar) representations. Consequently, any kind
of singularities is avoided at the poles. Finally, our transition from the scalar to
the vectorial as well as the tensorial case opens new promising perspectives of
constructing important zonal classes of spherical trial functions by summing up
Legendre kernel expressions, thereby providing (geo-)physical relevance and in-
creasing local applicability [43]. Similar considerations apply to the construction
of vector/tensor Slepian functions, e.g., [21, 79].
2.5. From the sphere to the ball

The modeling of structures inside the Earth requires basis functions on the ball
B = {x ∈ R3 : |x| ≤ β} with β > 0. Several approaches for the construction
of such basis systems exist. Of course, from the mathematical point of view, one
could easily take a basis {Bk }k=0,1,... on the Cartesian domain [−β, β] to con-
struct a basis on the cube [−β, β]3 by simply taking the tensor product basis
(x1 , x2 , x3 ) → Bk1 (x1 )Bk2 (x2 )Bk3 (x3 ), k1 , k2 , k3 = 0, 1, . . . However, the Earth’s
interior is usually subdivided into structural layers that are approximately bounded
by spheres. In view of this fact, the use of cartesian-coordinate-based trial func-
tions appears to be inappropriate and the spherical harmonics also here play an
essential role.
An intuitive approach is to look for basis functions of the form
Gm,n,k (rξ) = Fm,n (r)Yn,k (ξ), ξ ∈ Ω, r ∈ [0, β],
for m, n = 0, 1, . . . and k = 1, . . . , 2n + 1. Also here, orthogonality appears to be
useful, which leads to the requirement that

Gm1 ,n1 ,k1 (x) Gm2 ,n2 ,k2 (x) dx
B
β
2
= r Fm1 ,n1 (r) Fm2 ,n2 (r) dr Yn1 ,k1 (ξ) Yn2 ,k2 (ξ) dω(ξ)
0 Ω
β
= r2 Fm1 ,n1 (r) Fm2 ,n1 (r) dr = 0, (18)
0
if m1 = m2 or n1 = n2 or k1 = k2 . The weight function r2 in the radial integral

in (18) suggests the use of the Jacobi polynomials as building blocks for Fm,n .
However, there is a notable degree of freedom in the choice of (e.g., polynomial)
functions for Fm,n . This degree of freedom can be used to construct the Gm,n,k in a
manner such that they characterize the non-uniqueness of solutions of tomographic
inverse problems in the geosciences or medical imaging. For further details, see
[9, 63, 67, 68, 70, 97] and the contribution by Leweke, Michel, and Telschow (this
book, pp. 883–919). Note that some of the obtained systems become discontinuous
or even singular at the origin 0 ∈ B but in a way such that they are still elements
of L2 (B).
The fact that such orthonormal basis functions on the ball arise from the
spherical harmonics as orthonormal basis functions on the sphere yields a way
to formulate analogies regarding the methodologies and the associated properties
– though often further difficulties occur due to the additional radial coordinate.
Particular analogies exist with respect to the space and ‘frequency’ localization of
kernels
∞ 2n+1

K(x, y) = K ∧ (m, n) Gm,n,k (x) Gm,n,k (y)
m,n=0 k=1
∞

∧ 2n + 1 x y
= K (m, n)Fm,n (|x|)Fm,n (|y|) Pn · , (19)
m,n=0
4π |x| |y|
x, y ∈ B \ {0}. It should be noted, however, that most of the choices of Fm,n

do not lead to functions Gm,n,k which are algebraic polynomials in x1 , x2 , x3 .
Nevertheless, the properties of the Jacobi polynomials and the spherical harmonics
imply that the Gm,n,k also show an increasing oscillatory behavior for increasing
m or n. Furthermore, the Dirac kernel can also here be associated to the case
where all coefficients satisfy K ∧ (m, n) = 1.
One of the advantages of this approach – in contrast to a cartesian setup –
is that the obtained kernels inherit the rotational invariance from the spherical
kernels in the sense that
∞
2n + 1
K(rtξ, stη) = K ∧ (m, n)Fm,n (r)Fm,n (s) Pn ((tξ) · (tη))
n=0
4π
= K(rξ, sη)
for all 3 × 3-orthogonal matrices (i.e., rotations) t, see Figure 2.
120
16
100 14
12
80
0.5 0.5 10
60 8
0 0
z
40 6
0.5 0.5 4
-0.5 -0.5
20
0 0 2
0.5 0.5
0 -0.5 0 -0.5
-0.5 x -0.5 x
y y
(a) Kernel with K ∧ (m, n) = 0.8m 0.9n (b) Kernel with K ∧ (m, n) = 0.8m 0.7n
Figure 2. The figures show localized trial functions B y →
K(x, y) based on a kernel of the kind in (19) with a fixed point
x = (−0.4, −0.1, 0.5)T. The functions are plotted on the sphere with
radius |x|. Each function is a hat function concentrated around x. Its
restriction to a sphere around 0 is a rotationally symmetric function,
as it is known for the case of spherical kernels. Note that the series
representations were truncated at n = m = 400 in the numerical imple-
mentation.
Moreover, the localization with respect to the radius (or the ‘depth’) can be
separated from the localization with respect to the angular coordinates by taking,
for example, symbols of the form K ∧ (m, n) = Am Bn like K ∧ (m, n) = hm n
r hang
for parameters hr , hang ∈ (0, 1), see Figure 3. This is useful, e.g., for tomographic
problems where it is known that the solution has a finer structure in the angular
domain than in the radial domain (or vice versa).
20
100
0.5 0.5 15
80
0 0
z
z
60 10
-0.5 -0.5
40
-0.5 -0.5 5
20
0 0
0.5 0.5
0.5 0 0.5 0 0
x -0.5 x -0.5
y y
(a) Kernel with hr = 0.8 (b) Kernel with hr = 0.5

and hang = 0.9 and hang = 0.9
16 5
14
0.5 0.5 4
12
0 10 0
3
z
-0.5 8 -0.5
6 2
-0.5 4 -0.5
1
0 2 0
0.5 0.5
0.5 0 0.5 0
x -0.5 x -0.5
y y
(c) Kernel with hr = 0.8 (d) Kernel with hr = 0.5

and hang = 0.7 and hang = 0.7
Figure 3. The figures show localized trial functions B y →
K(x, y) based on a kernel of the kind in (19) with a fixed point
x = (−0.4, −0.1, 0.5)T. The functions are plotted on the planar cross
section with normal vector (1, 1, 1)T . By choosing a symbol K ∧ (m, n) =
hm n
r hang , the localization in radial and in angular domain can be con-
trolled separately. Note that the series representations were truncated
at n = m = 400 in the numerical implementation.
3. Spherical uncertainty principle

As pointed out in Section 2, four classes of zonal kernel functions can be distin-
guished, namely bandlimited and non-bandlimited, spacelimited and non-space-
limited ones. In addition, Slepian functions exist in bandlimited and spacelimited
varieties. What is the right kernel function for the purpose of local approximation?
Of course, the user of a mathematical method is interested in knowing the trial
system which fits ‘adequately’ to the problem. When several choices are possible
or an optimal choice cannot be found it is necessary to choose the trial systems in
close adaptation to the data extent or density, and the required smoothness of the
field to be approximated. This, however, is often a local property, i.e., the data
density can be high in one area and low in another. In addition, the field to be ap-
proximated can have a high-detail structure in some parts of the sphere (e.g., over
mountainous regions) and a low-detail structure elsewhere (e.g., over the oceans).
This makes the selection of appropriate trial functions even more challenging.
3.1. Derivation and basic theory

An uncertainty principle that specifies the degree of space and frequency localiza-
tion is helpful to serve as a decisive criterion. The essential outcome is a better
understanding of the classification of zonal kernel functions, and Slepian functions,
based on the development of suitable bounds for their quantification with respect
to space and frequency localization.
Localization in space. Assume F is of class L2 (Ω) with energy

1/2
F L2 (Ω) = (F (η))2 dω(η) = 1.
Ω
(1)
We associate to F the normal (radial) field η → ηF (η) = oη F (η), η ∈ Ω. This
function maps L2 (Ω) into the associated set of normal fields on Ω. The ‘center of
gravity’ of F is the expectation of the normal operator o(1) on Ω,

(1)
gFo = o(1)
η F (η) F (η) dω(η) = η(F (η))2 dω(η) ∈ R3 (20)
Ω Ω
2
thereby interpreting (F (η)) dω(η) as surface mass distribution over the sphere
(1)
Ω embedded in Cartesian space R3 . It is clear that gFo lies in the closed inner
(1)
space Ωint of Ω: |gFo | ≤ 1. The variance of the operator o(1) is understood in the
canonical sense as the variance in the space domain,
(1) 2 2
o(1)
σFo = o(1)
η − gF F (η) dω(η)
Ω
(1) 2 2
= η − gFo (F (η)) dω(η) ∈ R. (21)
Ω
(1) (1) (1)
Observing the identity (η − gFo )2 = 1 + (gFo )2 − 2η · gFo , η ∈ Ω, it follows
(1) (1) (1)
immediately that (σFo )2 = 1 − (gFo )2 . Naturally, 0 ≤ (σFo )2 ≤ 1.
Since we are particularly interested in zonal functions, some simplifications

can be made. Let K be of class L2 [−1, 1] and K L2[−1,1] = 1, where F L2 [−1,1] =
1
(2π −1 (F (t))2 dt)1/2 for F ∈ L2 [−1, 1]. Then the corresponding center of gravity
can be computed readily as follows (ε3 = (0, 0, 1)T ):
1
o(1)

3 2 2
gK(·ε 3) = η K η · ε dω(η) = 2π t (K(t)) dt ε3 . (22)
Ω −1
(1)
(1) o 1 2
Letting toK = gK(·ε3 ) = 2π −1 t (K(t)) dt ∈ R we find for the variance
(1) 2 2 2
o(1)
o
σK = η − gK(·ε 3) K η · ε3 dω(η)
Ω
(1) 2 (1) 2
= 1 − toK = 1 − gK(·ε
o
3) ∈ R. (23)
(1)
ηFo
C
(1)
gFo
(1)
σFo
1
Figure 4. Localization in a spherical cap.
(1) (1)
Figure 4 gives a geometric interpretation of gFo and σFo . We associate
(1) (1) (1)
to gFo , gFo = 0, and its projection ηFo onto the sphere Ω the spherical cap
(1) (1)
C = {η ∈ Ω | 1 − η · ηFo ≤ 1 − |gFo |}. Then the boundary ∂C is a circle with
(1)
radius σFo . Thinking of a zonal function F as a ‘spherical window function’ on
(1)
Ω, the window is determined by C, and its width is given by σFo .
Localization in frequency (‘momentum space’). The ‘expectation in the frequency
domain’ is introduced as the expectation of the surface curl operator o(3) on Ω.
Then, for F ∈ H(2l) (Ω), l ∈ N, i.e., for all F ∈ L2 (Ω) such that there exists a
function G ∈ L2 (Ω) with G∧ (n, k) = (−n(n + 1))l F ∧ (n, k) for all n = 0, 1, . . .;
k = 1, . . . , 2n + 1, we have

o(3)
gF = o(3) 3
η F (η) F (η) dω(η) = 0 ∈ R . (24)
Ω
operator expectation value

(1)
space o(1) gFo = o(1)
η F (η) F (η) dω(η)
Ω
(3)
frequency o(3) gFo = o(3)
η F (η) F (η) dω(η)
Ω
operator variance

2 2
(1)
o(1)
space o(1) σFo = o(1)
Ω
(3) 2 2
o(3)
frequency o(3) σFo
= o(3)
Ω
Table 5. Localization in terms of the normal and curl operators o(1)

and o(3) .
Correspondingly, the variance in the frequency domain is given by

2 2
(3) (3)
σFo = o(3)
η − gF
o
F (η) dω(η) ∈ R. (25)
Ω
The surface theorem of Stokes shows us that

(3) 2
σFo = o(3) (3)
η F (η) · oη F (η) dω(η)
Ω
∗
= −Δ∗η F (η) F (η) dω(η) = gF−Δ . (26)
Ω
Expressed in terms of spherical harmonics we get via the Parseval identity
(3) 2 ∞ 2n+1
2
σFo = n(n + 1) (F ∧ (n, k)) . (27)
n=0 k=1
2 ∞ 2n+1 2
Note that we require F L2 (Ω) = n=0 k=1 (F ∧ (n, k)) = 1. The meaning of
(3) (3)
σFo as measure of ‘frequency localization’ is as follows: the range of σFo is the
(3)
interval [0, ∞]; a large value of σFo occurs if many Fourier coefficients contribute
(3)
to σFo . In conclusion, relating any spherical harmonic to a ‘single wavelength’, a
(3)
large value σFo informs us that F is spread out widely in ‘frequency domain’. In
(3)
contrast, a small value σFo indicates that only a few number of Fourier coefficients
is significant (see Table 5).
Again we reformulate our quantities in the specific context of zonal functions.

Let K(·ε3 ) be of class H(2) (Ω) satisfying K(·ε3 ) L2 (Ω) = 1, then
(3) 2

o
σK(·ε3) = − Δ∗η K η · ε3 K η · ε3 dω(η)
Ω
1
= −2π K(t)Lt K(t) dt (28)
−1
d d
where Lt denotes the Legendre operator as given by Lt = dt (1 − t2 ) dt .
Uncertainties and the uncertainty principle. The square roots of the variances, i.e.,
(1) (3)
σ o and σ o , are called the uncertainties in o(1) and o(3) , respectively. For these
(1) (3) (1)
quantities we get (see [43, 76]) an estimate given by (σFo )2 (σFo )2 ≥ |gFo |2 . We
summarize our results in Theorem 1. For details on the proof the reader is referred
to [29].
Theorem 1. Let F ∈ H(2) (Ω) satisfy F L2 (Ω) = 1. Then
(1) 2 (3) 2 (1) 2

σFo σFo ≥ gFo . (29)
(1)
If gFo is non-vanishing, then
(1) (3)
ΔoF ΔoF ≥ 1, (30)
where we have used the abbreviations
(1)
(1) σo (3) (3)
ΔoF = F(1) , ΔoF = σFo . (31)
gFo
The uncertainty relation measures the tradeoff between ‘space localization’

and ‘frequency localization’ (‘spread in frequency’). It states that sharp localization
in space and frequency are mutually exclusive.
An immediate consequence of Theorem 1 is its reformulation for zonal func-
tions K(ε3 ·) : η → K(ε3 · η), η ∈ Ω.
(1)
Corollary 2. Let K(ε3 ·) ∈ H(2) (Ω) satisfy K L2 [−1,1] = 1. If toK is non-vanishing,
then
(1) (3)
ΔoK ΔoK ≥ 1, (32)
where
o(1)
o(1) σK (3)
o(3)
ΔK = o(1) , ΔoK = σK . (33)
tK
(3)
The interpretation of (σKo
)2 as variance in ‘total angular momentum’ helped
us to prove Theorem 1. However, this interpretation shows two essential drawbacks:
first, the expectation of the surface curl gradient is a vector which seems to be
inadequate in ‘momentum localization’ in terms of scalar spherical harmonics,
(3)
and secondly the value of gFo vanishes for all candidates F . This means that the
‘center of gravity of the spherical window’ in ‘momentum domain’ is independent

of the function F under consideration. Therefore, we are finally interested in the
variance of the operator −Δ∗
2
∗ 2 ∗
σF−Δ = −Δ∗η − gF−Δ F (η) dω(η) (34)
Ω
which is a measure for the ‘spread in momentum’. Now the corresponding expec-
∗
tation value gF−Δ is scalar-valued and non-vanishing. It can be easily seen that

∗ 2

∗ 2
(−Δ∗ )2
σF−Δ = gF − gF−Δ . (35)
In connection with Theorem 1 this leads to the following result.
Theorem 3. Let F be of class H(4) (Ω) such that F L2 (Ω) = 1. Then

2
(−Δ∗ )2 −Δ∗
(1) 2
∗ 2
g − g
(1) F F
σFo σF−Δ ≥ gFo −Δ∗
(36)
gF
∗
provided that gF−Δ = 0. If the right-hand side of (36) is non-vanishing, then
(1) ∗
ΔoF Δ−Δ
F ≥ 1, (37)
where
⎛ 2 ⎞1/2
∗
∗ ⎜ σF−Δ ⎟
∗ 1/2
=⎜ ⎟ (3)
Δ−Δ
F ⎝ g(−Δ∗ )2 −(g−Δ∗ )2 ⎠ = gF−Δ = ΔoF . (38)
F F
−Δ ∗
gF
3.2. Classification of examples

We continue with some examples of particular interest for geoscientific research.
Localization of the spherical harmonics. We know that

(Yn,k (ξ))2 dω(ξ) = 1 . (39)
Ω
One can prove that

(1) (1)
gYo n,k = 0, σYo n,k = 1. (40)
Moreover, we have
∗ ∗
gY−Δ
n,k
= n(n + 1), σY−Δ
n,k
= 0. (41)
In other words, spherical harmonics show an ideal frequency localization, but no
space localization (see Figure 5 for an illustration of space and frequency localiza-
tion for the Legendre polynomials).
1
o?4 o?4
o?7 1 o?7
o?; o?;
0.5 0.8
0.6
0
0.4
−0.5
0.2
−1 0
−P −P/2 0 P/2 P 0 5 10 15
Figure 5. The Legendre kernel Pn for n = 2, 5, 9, space representation

for ϑ → Pn (cos(ϑ)) (left), and frequency representation m → (Pn )∧ (m)
(right).
Localization
of the ideally bandlimited Legendre kernel. We have, with Pn∗ =
2n+1
4π Pn ,

(Pn∗ (ξ · ζ))2 dω(ζ) = 1 (42)
Ω
for all ξ ∈ Ω, such that
(1) (1)
gPo n∗ (ξ·) = 0, σPo n∗ (ξ·) = 1, (43)
∗ ∗
gP−Δ
∗ (ξ·) = n(n + 1), σP−Δ
∗ (ξ·) = 0. (44)
n n
Localization of the bandlimited Shannon kernel. The Shannon kernel Φρ , ρ > 0,

given by
2n + 1
Φρ (ξ · η) = Pn (ξ · η), ξ, η ∈ Ω, (45)
−1
4π
n≤ρ
may be interpreted as a truncated Dirac kernel. It is not surprising that the Shan-
non kernel as a ‘finite polynomial kernel’ shows strong oscillations in space. This
is the price to be paid for the sharp separation in frequency space.
The investigation of the uncertainty properties of the Shannon kernel starts
from (cf. [43])
ρ−1
2n + 1 1 −1
Φρ 2L2 (Ω) = = (*ρ + + 1) + *ρ−1 +*ρ−1 + 1+ , (46)
n=0
4π 4π
where, as usual, *ρ−1 + is the largest integer which is less or equal ρ−1 . Observing
this result, we introduce the normalized Shannon kernel by
1
Φ̃ρ = Φρ . (47)
Φρ L2 (Ω)
25
? 3>38 ? 3>38
? 3>: 1 ? 3>:
20 ? 3>6 ? 3>6
0.8
15
0.6
10
5 0.4
0 0.2
−5 0
−P −P/2 0 P/2 P 0 5 10 15 20
Figure 6. The Shannon scaling function Φρ for ρ = 1/16, 1/8, 1/4.

Space representation ϑ → Φρ (cos(ϑ)) (left) and frequency representa-
tion n → (Φρ )∧ (n) (right).
Its localization in space satisfies

⎛ −1 ⎞2
(1) 2 ρ −1
1 ⎝ 2n + 2 ⎠
o
σΦ̃ =1−
ρ Φρ 2 n=1
4π
2
2*ρ−1 − 1+ + *ρ−1 +*ρ−1 − 1+
=1− , (48)
*ρ−1 + 1+ + *ρ−1 +*ρ−1 + 1+
so that 6
7
7 1 − 2ρ−1 −1+ρ−1 ρ−1 −1 2
7 −1 −1
ρ +1+ρ ρ +1−1
=8
(1)
o
ΔΦ̃ρ
−1 −1 −1
2ρ −1+ρ ρ −1
. (49)
ρ−1 +1+ρ−1 ρ−1 +1
Moreover, we find
−1
2 4π
ρ
2n + 1
o(3)
σΦ̃ = −1 n(n + 1)
ρ *ρ + + 1 + *ρ−1 +*ρ−1 + 1+ n=0 4π
1 *ρ−1 +(1 + *ρ−1 +)2 (2 + *ρ−1 +)
= (50)
2 *ρ−1 + + 1 + *ρ−1 +*ρ−1 + 1+
such that
(3) 1 *ρ−1 +(1 + *ρ−1 +)2 (2 + *ρ−1 +)
ΔoΦ̃ρ = . (51)
2 *ρ−1 + + 1 + *ρ−1 +*ρ−1 + 1+
The results are graphically illustrated in Figure 7.
Localization of the non-bandlimited/non-spacelimited Abel–Poisson kernel. Let
us consider the function Qh : [−1, 1] → R, h < 1, given by
∞
1 1 − h2 2n + 1 n
Qh (t) = 2 3/2
= h Pn (t). (52)
4π (1 + h − 2ht) n=0
4π
4
10 )3*
p

)5*
p
2
10 )3* )5*
p p

0
10
í2
10
í4
10 í4 í3 í2 í1
10 10 10 10
Figure 7. Uncertainty classification of the normalized Shannon Dirac

(1)
o(3) (1) (3)
family Φ̃ρ . Shown are ΔoΦ̃ , ΔΦ̃ , and the product ΔoΦ̃ ΔoΦ̃ as func-
ρ ρ ρ ρ
tions of ρ in a double logarithmic setting.
An easy calculation gives us

1/2
1/2 1 + h2 1
Qh L2 [−1,1] = (Qh2 (1)) = . (53)
4π 1 − h2
Furthermore, for Q̃h (t) = Qh −1L2 [−1,1] Qh (t), t ∈ [−1, 1], we obtain after an
elementary calculation (see also Figure 8)
√
o(1) 1 − h2 −Δ∗ 6h
ΔQ̃ = , ΔQ̃ = . (54)
h 2h h 1 − h2
Thus, we finally obtain
√
(1) ∗ 6 3
ΔoQ̃h Δ−Δ = = > 1. (55)
Q̃h 2 2
(1) ∗
Here, the value ΔoQ̃ Δ−Δ
Q̃h
is independent of h. All intermediate cases of ‘space-
h
frequency localization’ are realized by the Abel–Poisson kernel, but the Abel–
Poisson kernel does not satisfy a minimum uncertainty state.
Localization of the spacelimited Haar kernel. Let k be a non-negative integer, i.e.,

(k)
k ∈ N0 . The (smoothed) Haar kernel {Bh }h∈(0,1) ⊂ C(k−1) [−1, 1] is defined by
⎧
⎨ 0 , t ∈ [−1, h)
(k)
Bh (t) = (t − h)k (56)
⎩ , t ∈ [h, 1].
(1 − h)k
(k)
By definition, Bh is non-negative and has the support [h, 1]. Obviously, the func-
(0)
tion Bh , h ∈ (−1, 1), represents the (classical) Haar function (cf. [53]). The
15 )3*
pR
i

R i
10
0
0 0.2 0.4 0.6 0.8 1
Figure 8. Abel–Poisson kernel uncertainty classification. The curves

(1) ∗
graphically illustrate the functions h → ΔoQ̃ and h → Δ−Δ
Q̃
.
h h
2
i ? 2;5 i ? 2;5
i ? 2;9 1.2 i ? 2;9
i ? 2;; i ? 2;;
1
1.5
0.8
1 0.6
0.4
0.5 0.2
0
0 í0.2
íP íP/2 0 P/2 P 0 5 10 15 20 25
(0)
Figure 9. The Haar kernelBhfor h = 0.3, 0.7, 0.9. Space representa-
(0)
tion ϑ → Bh (cos(ϑ)), ϑ ∈ [−π, π], (left) and frequency representation
(0)
n → (Bh )∧ (n) (right).
(k)
Legendre coefficients of Bh , h ∈ (−1, 1), k ∈ N0 , can be calculated recursively
(cf. [47]):
∧ 1−h
(k)
Bh (0) = 2π = 0, (57)
k+1
∧
(k) 1−h 1−h
Bh (1) = 2π 1− , (58)
k+1 k+2
∧ 2n + 1 ∧ k + 1 − n (k) ∧
(k) (k)
Bh (n + 1) = h Bh (n) + Bh (n − 1). (59)
n+k+2 n+k+2
An elementary calculation shows
- - 1. /2
- (k) -2 (k)
-Bh - = 2π Bh (t) dt
L2 (Ω) −1
1−h
= 2π . (60)
2k + 1
We define the kernel

(k) 2k + 1 (k)
B̃h = B , (61)
2π(1 − h) h
since the uncertainty properties are normally defined for kernels with norm one.
We find 1 2 1 + h + 2k 3
o(1) (k)
gB̃ (k) 3
= 2π t B̃ h (t) dt ε3 = ε . (62)
h ( ·ε ) −1 2 + 2k
Consequently,
(1) 2 2
1 + h + 2k (1 − h)(h + 4k + 3)
σB̃ (k) = 1 −
o
= . (63)
h 2 + 2k (2k + 2)2
Using (31), we finally arrive at
(1) 1 #
ΔoB̃ (k) = (1 − h)(h + 4k + 3). (64)
h 1 + h + 2k
For the localization in frequency, we assume k ≥ 2. We have
(3) 2 1
(k) (k)
o
σB̃ (k)
( ·ε3 )
= −2π B̃h (t) Lt B̃h (t) dt
h −1
1
2k + 1 −2π
= (t − h)k Lt (t − h)k dt
2π(1 − h) (1 − h)2k h
k(h + 2k)
= , (65)
(1 − h)(2k − 1)
so that
(3) k(h + 2k)
ΔoB̃ (k) = . (66)
h (1 − h)(2k − 1)
The application of Lt requires that the kernel is twice differentiable. However,
using integration by parts, the results immediately carry over to the case k = 1.
Figure 10 gives a graphical impression of these results for the particular cases
k = 1 and k = 3.
Localization of the ideally spacelimited Dirac kernel. Letting h formally tend to
1 in the results provided by the uncertainty principle for the Abel–Poisson kernel
function we are able to interpret the localization properties of the Dirac kernel on
Ω satisfying δ ∧ (n) = 1 for all n ∈ N0 :
∞
2n + 1
δ(ξ · η) = Pn (ξ · η), ξ, η ∈ Ω, (67)
n=0
4π
where the convergence is understood in distributional sense. As a matter of fact,
letting h tend to 1 shows us that the variances in the space domain take the con-
stant value 0. On the other hand, the variances in the frequency domain converge
to ∞. Hence, the Dirac kernel shows ideal space localization, but no frequency
localization.
3 )3*
3 )3*
p )l* p )l*
Ci Ci
2.5 )5*
p )l* 2.5 )5*
p )l*
C i
C i
)3* )5* )3* )5*
2 p )l* p )l* 2 p )l* p )l*
Ci Ci Ci Ci
1.5 1.5
1 1
0.5 0.5
0 0
í1 í0.5 0 0.5 1 í1 í0.5 0 0.5 1
Figure 10. Uncertainty classification of the normalized smoothed Haar

(k) (1) (3)
scaling function B̃h (k = 1, left; k = 3 right). Δo (k) , Δo (k) and the
B̃h B̃h
(1) (3)
product Δo (k) Δo (k) are shown as functions of h.
B̃h B̃h
Localization of the non-bandlimited/non-spacelimited Gaussian function.

The minimum uncertainty state within the uncertainty relation is provided by
the Gaussian probability density function (see [47, 59]). Consider the function Gλ
given by
Gλ (t) = e−(λ/2)(1−t) , t ∈ [−1, 1], λ > 0. (68)
An elementary calculation shows us that
G̃λ (t) = γ(λ)e−(λ/2)(1−t) , (69)
with
√ 1
−1/2
γ(λ) = 1/ 4π 1 − e−2λ , (70)
2λ
(1) ∗
satisfies G̃λ L2 [−1,1] = 1. It is not difficult to deduce that ΔoG̃ Δ−Δ
G̃λ
→ 1 as
λ
λ → ∞: the best value of the uncertainty principle (Theorem 3) is 1.
Localization of Slepian functions. The bandlimited Slepian functions solve the

concentration criterion (10) on general domains C ⊂ Ω. If we restrict our attention
to spherical caps as in Figures 1 and 4, the solutions degenerate and equations (11)
and (12) can be solved for fixed spherical-harmonic orders j, with twice-repeated
eigenvalues for the nonzonal functions at the same nonzero absolute orders.
While the Slepian functions do not formally optimize the uncertainty rela-
tion (38), calculations by Wieczorek and Simons [104] reveal that, again on spher-
ical caps C of various opening half-angles Θ, the values attained by the largest-
(α)
eigenvalue (λC ≈ 1 for α = 1) zonal Slepian functions of varying bandwidths N
are very close to satisfying the bounds (38) for Shannon numbers N0 = (N +1)Θ/π
(see (16)) greater than about 2. Furthermore, for increasing Shannon numbers,
the uncertainty products for the αth best-concentrated Slepian function, when
N0 ≥ α + 1, tend to 2α − 1. This favorable behavior was illustrated by Wieczorek
and Simons [104], see their Figures 5 and 6b. Subsequent work by, among others,
Guilloux et al. [52] and Khalid et al. [57], has substantiated and elaborated on
these early analyses.
Slepian functions vs. the Gaussian. Another way by which the spatiospectral lo-
calization properties of the Slepian functions may be appreciated is by comparing
how close they are to the family of minimum-uncertainty ‘squeezed’ coherent states
(e.g., [15, 58]), a common root for many later developments in spline, Slepian func-
tion, and wavelet analysis [20]. This is of importance because in practical problems
in the geosciences (e.g., [17, 84]), as in cosmology (e.g., [94]), we place as much
value on the precise bandwidth, or bandwidth resolution, of our observations as
on the spatial domain of interest. The Gaussian (68) may satisfy the uncertainty
lower bound exactly, but it is not a bandlimited kernel. In contrast, the Slepian
functions (11–12) can be bandlimited and spaceconcentrated at the same time.
Formally, they are the optimizers of (10), though not of (38).
That they get close is shown in Figure 11. Inspired by Bluhm et al. [14] we
determine the squeeze factor, s that renders the suitably normalized function
Gs (cos θ) = γ(s)es cos θ , 0 ≤ θ ≤ π, (71)
as close as possible, in the mean-squared sense, to the best-concentrated bandlim-

ited zonal Slepian function, concentrated to a spherical cap of a certain radius Θ,
and whereby the tradeoff between spatial (the area of the spherical cap) and spec-
tral concentration (the bandwidth N ) is parameterized via the partial Shannon
number N0 = (N + 1)Θ/π.
3.3. Closing remarks

The uncertainty principle represents a trade-off between two ‘spreads’, in position
and in frequency. Sharp localization in space and in frequency are mutually exclu-
sive. The reason for the validity of the uncertainty relation (Theorem 1) is that
the normal and curl operators o(1) and o(3) do not commute, hence, they cannot
be sharply defined simultaneously. Extremal members of the uncertainty relation
are polynomials (spherical harmonics) and Dirac function(al)s. An asymptotically
optimal kernel is the Gaussian function.
Corollary 2 allows a quantitative classification and a hierarchy of the space
and frequency localization properties of kernel functions of the form
∞
2n + 1 ∧
K(t) = K (n)Pn (t), t = ξ · η, (ξ, η) ∈ Ω × Ω. (72)
n=0
4π
In view of their space/frequency localization, it is also important to distinguish

bandlimited kernels (i.e., K ∧ (n) = 0 for all n ≥ N ∈ N0 ) and non-bandlimited
ones (K ∧ (n) = 0 for an infinite number of integers n). Non-bandlimited kernels
show a much stronger space localization than their bandlimited counterparts. It is
12
150 a b
relative misift norm (\%)

10
squeeze factor s
8
100
6
50 4
0 0
1 2 3 4 5 6 1 2 3 4 5 6
Shannon number N0 Shannon number N0
30 6
N0= 1 1= 20°
25 N0= 2 Slepian, squeezed function 5 1= 25°
N0= 3 1= 30°
20 N0= 4 4
1= 45°
N0= 5
s/N0
15 3 1= 60°
N0= 6
10 2
5 1
c d
0 0
20° 25° 30° 45° 60° 0° 15° 30° 45° 60° 75°
spherical cap radius 1 colatitude Q
Figure 11. Slepian functions compared to Gaussian ‘squeezed coher-

ent states’. The chosen squeeze factors render functions of the type (71)
as close as possible to the best-concentrated zonal Slepian functions in
the relative mean-squared sense, for a variety of spherical cap sizes Θ
and Slepian-function bandwidths N , linked through the Shannon num-
ber N0 .
not difficult to prove that, if K ∈ L2 [−1, 1] with K(ξ· ) L2 (Ω) = 1,

∞ 2
(1) 2 2n + 1
∧ ∧
σK(ξ· ) = 1 −
o
K (n)K (n + 1) . (73)
n=1
4π
If K ∧ (n) ≈ K ∧ (n + 1) ≈ 1 for many successive integers n, the space-domain
support of K(t) in (72) is small.
Space/frequency localization on the sphere can also be illustrated directly
from (72). Choosing K ∧ (n) = δnk we obtain a Legendre kernel of degree k, on
the left in our scheme (Table 6). Setting K ∧ (n) = 1 for n = 0, 1, . . ., we obtain
the Dirac kernel. The slower the sequence {K ∧ (n)}n=0,1,... converges to zero, the
lower the frequency localization, but the higher the space localization.
Altogether, Table 6 gives a qualitative illustration of the consequences of the
uncertainty principle in the theory of zonal kernel functions on the sphere: on the
space localization
-
no space localization ideal space localization
frequency localization
-
ideal frequency localization no frequency localization
kernel type
-
Legendre kernel bandlimited spacelimited Dirac kernel
Table 6. The uncertainty principle and its consequences.
left end of this scheme, we have the Legendre kernels with their ideal frequency
(momentum) localization. However, they show no space localization, as they are of
polynomial nature. Thus, the present standard way in applications of increasing the
accuracy in spherical harmonic (Fourier) expansions is to increase the maximum
degree of the spherical harmonics expansions under consideration. On the right
end of the scheme, there is the Dirac kernel which maps a function to its value at
a certain point. Hence, this (generalized) function has an ideal space localization
but no frequency localization. Consequently, it can be used in a finite pointset
approximation.
4. Constructive approximation on the sphere

In Section 4.1, we discuss an approach using Slepian functions, Section 4.2 is an
approach based on splines, and Section 4.3 treats the case of wavelets. Section 4.4
helps combine benefits of various approaches.
4.1. Approximation by Slepian functions

Given a certain region of interest C on the unit sphere Ω and a certain band-
width N (a limiting spherical-harmonic degree in the sense of (9)), optimization
of a concentration criterion yields linear combinations of spherical harmonics that
we call Slepian functions. In Section 2.3, we gave their formulation in terms of
bandlimited functions that are spaceconcentrated. We shall denote these functions
from now on as GC N (ξ). Of course, we can equally well ask for spacelimited func-
tions that are bandconcentrated – see [84, 87] for details. We shall denote those
functions from now on as HCN (ξ). The Fourier coefficients of the HCN can be calcu-
lated from those of the GCN by extension as in (13). We refer to [60] for an extensive
discussion on the properties of what are, essentially, cases intermediate between
these two endmembers, for functions defined on the real line.
If we introduce the space of all square-integrable scalar spherical functions

that are spacelimited to the region C as SC , and the space of all square-integrable
spherical functions that are bandlimited to the spherical-harmonic degree N as
SN , then it is implied that HCN ∈ SC and GC N ∈ SN .
Reproducing properties. We can show that the spectral-domain kernel that we

first encountered in bandlimited form in (11), and which we now extend to 0 ≤
m, n < ∞,

dC(m,j),(n,k) = Ym,j , Yn,k 2
L (C) = Ym,j (ξ)Yn,k (ξ) dω(ξ), (74)
C
is a reproducing kernel in the space SC . Indeed, for any function F ∈ SC ,

∞ 2n+1 ∞ 2n+1

C ∧ ∧
d(m,j),(n,k) F (n, k) = Ym,j (ξ) Yn,k (ξ)F (n, k) dω(ξ)
n=0 k=1 C n=0 k=1

= Ym,j (ξ)F (ξ) dω(ξ)
C
∧
= F (m, j). (75)
At the same time, the spatial-domain Shannon kernel that we encountered in (12),
and which we rebaptize
N N 2n+1

2n + 1
DN (ξ, η) = Pn (ξ · η) = Yn,k (ξ)Yn,k (η), (76)
n=0
4π n=0 k=1
is a reproducing kernel in the space SN , since, for any function F ∈ SN ,

N
2n + 1
Pn (ξ · η)F (η) dω(η)
n=0
4π
Ω
N 2n+1

= Yn,k (ξ) Yn,k (η)F (η) dω(η)
n=0 k=1 Ω
N 2n+1

= Yn,k (ξ)F ∧ (n, k)
n=0 k=1
= F (ξ). (77)
Equations (75) and (77) hold the key to the approximation properties of the
Slepian functions, since they imply that the spacelimited Slepian functions HCN
provide a complete basis for all spacelimited functions in SC ⊂ L2 (Ω), whereas
the bandlimited Slepian functions GC N are a complete basis for all bandlimited
functions in SN ⊂ L2 (Ω).
Mercer’s theorem. A second set of properties that solidifies these notions is es-
tablished through an identity known as Mercer’s theorem, which in this context
takes the form
(N +1)2
C
DN (ξ, η) = GC
N α (ξ) GN α (η), (78)
α=1
(α)
for all the α-indexed bandlimited Slepian functions GC N , with eigenvalues λC ,
from which we establish, using (76) and as in (14d), that
(N +1)2
2 (N + 1)2 NC
DN (ξ, ξ) = GC
N α (ξ) = = , (79)
α=1
4π C dω(η)
recovering
the spherical Shannon number NC and the area of the domain of inter-
est, C dω(η).
A useful corollary is that the eigenvalue-weighted sum of squares
of the band-
limited Slepian eigenfunctions closely approximates the value NC / C dω(η) when
ξ ∈ C, and vanishes otherwise,
(N +1)2
(α)

C 2
NC
(α) C 2 NC / C dω(η) if ξ ∈ C
λC GN α ≈ λC GN α ≈
0 otherwise,
α=1 α=1
which is a consequence of the step-shaped eigenvalue spectrum that we saw in

Figure 1. Eq. (80) testifies to the fact that the effective dimension of the space
SN of bandlimited functions that are also spaceconcentrated to C, is reduced
from the canonical (N + 1)2 to the Shannon number NC . It is our first clue to the
approximation qualities of the Slepian functions, e.g., for (linear) signal estimation
from regionally available data [85].
Power spectrum. If we furthermore define the power spectrum or degree variance

of the bandlimited Slepian functions as
1
2n+1
∧ 2

Pn(α) = GC
N α (n, k) (80)
2n + 1
k=1
we get, via the spectral theorem, equation (74), and the addition theorem a
spectral-domain equation equivalent to (79)–(80), namely,
(N +1)2 2n+1
(α) 1 dω(η)
(α)
λC Pn = C
d(n,k),(n,k) = C
α=1
2n + 1 4π
k=1

NC
(α)
≈ λC Pn(α) , (81)
α=1
which is suggestive of the spectral -domain approximation properties of the Slepian

functions, as arises, e.g., in the theory of (quadratic) power-spectral estimation
from regionally available data [17].
Equations (80) and (81) together, show that the set of NC < (N + 1)2
Slepian functions provide essentially uniform coverage over the spatial domain C
and spectral bandwidth N . This is of interest when estimating (interpolating,
approximating) functions from observations, as is common to a large number of
research fields, not limited to the geosciences.
Alternative Mercer theorem. We note for completeness, and since the relevant
identities have not been published before, that an alternative version of Mercer’s
theoremwould have transformed (12) and (76) from

DN (ξ, η)F (η) dω(η) = λC (F )F (ξ) (82)
C
into the full-domain

DN (ξ, η)F (η) dω(η) = λC (F )F (ξ), (83)
Ω
which have the same eigenfunctions, but where we have defined
N 2m+1
N 2n+1

DN (ξ, η) = dC
(m,j),(n,k) Ym,j (ξ)Yn,k (η). (84)
m=0 j=1 n=0 k=1
In that case, the equivalent to (78) is the to some more familiar expression
(N +1)2
(α) C
N
D (ξ, η) = λC GN α (ξ) GC
N α (η). (85)
α=1
Approximation 0: Noiseless data (interpolation). Imagine a certain function is

‘known’ as a spherical-harmonic expansion. Clearly, considering such a situation
is merely postponing the problem of how to estimate an unknown function from
observations. However, it is a common occurrence in the geosciences that, for
example, space agencies perform exhaustive satellite data reductions that end up in
the official release of spherical harmonic ‘models’ (typically of gravity or magnetic
fields) that are then available for further research [103]. Another situation is where
spectral forward-modeling codes deliver ‘simulations’ that are subsequently in need
of interpretation and evaluation [102].
Whatever the source, and however large the bandlimit, the key property
of the Slepian function basis is that the function expansion coefficients can be
obtained by a simple transformation. If indeed the known function is F , then
it is immaterial whether it is expressed in the spherical-harmonic basis, or in a
bandlimited Slepian basis designed for whichever region C of interest, as long as
its bandwidth N matches the original :
N 2n+1 (N +1)2

∧
F = F (n, k)Yn,k = F ∧ (α) GC
N α. (86)
n=0 k=1 α=1
The Slepian-function expansion coefficients F ∧ (α), α = 1, . . . , (N + 1)2 are

simply obtained from the spherical-harmonic expansion coefficients F ∧ (n, k),
n = 0, 1, . . . , N and order indices k = 1, . . . , 2n + 1, by the (orthogonal) trans-

formation [85]
N 2n+1
∧
F ∧ (α) = GC ∧
N α (n, k)F (n, k). (87)
n=0 k=1
A linear basis transformation (87) is exact and thus, strictly speaking ‘uninterest-
ing’. However, the properties of the Slepian functions designed for a region C are
such that after a partial Slepian expansion to J < (N + 1)2 terms, denoted FJ (ξ),
equation (86) will hold approximately in the region of interest:
N 2n+1

J

F (ξ) = F ∧ (n, k)Yn,k (ξ) ≈ F ∧ (α) GC
N α (ξ), ξ ∈ C. (88)
n=0 k=1 α=1
Clearly, a truncation of the spherical-harmonic series to its first J terms, however

ordered, would generally result in poor approximations, precisely because of the
non-localized spatial behavior of the basis functions. The eigenvalue-ranked Slepian
transformation (87), on the other hand, has reordered the basis such that its
first J functions increasingly uniformly ‘cover’ the spatial region of interest while
providing an increasingly complete coverage over the entire spectral band, see (80)
and (81). As a measure of approximation quality we take the area-weighted relative
mean-squared error. It can be easily shown to depend on the truncation level in
the manner
(N +1)2
2 (α)
2
(F ∧ (α)) λC
F − FJ L2 (C) α>J
2 = . (89)
F L2 (C) (N +1)2
2 (α)
(F ∧ (α)) λC
α=0
Given the universally favorable decay of the eigenvalue spectrum of the spatio-
spectral concentration problem (11), in this noiseless case, the Shannon number NC
is an obvious practical first choice for the truncation level J, although (89) of course
shows the role played by the spectrum of the signal itself. An illustrative numerical
example is given by Simons et al. [88], their Figure 3.
Approximation 1: Noisy data. We finally turn to the approximation problem that

is most familiar in geophysical inverse theory, namely that of the estimation of
a certain unknown signal from noisily observed data. We will briefly discuss the
traditional spherical-harmonics based approach, and then clarify the beneficial role
that localized basis functions (here: Slepian functions) may play in this context. We
adhere to the continuous viewpoint for notational convenience and to lay bare the
structure of the solutions. In practice, all datasets will be sample values at discrete
geographic locations. As a consequence, the properties derived for constructive
approximation by Slepian functions will themselves hold only approximately – to
the degree by which continuous integrals are (hopefully, well) approximated by
their Riemann sums [17, 84].
However, therein lies the crux of the Slepian-function method: if the data
are regionally (in some region C) and densely (warranting a certain ‘Nyquist’
bandlimit at spherical-harmonic degree N ) available, computing the Slepian basis
for the idealized acquisition geometry ahead of time is what will lead to man-
ageably sized inverse problems (on the order of the Shannon number NC , and
NC , (N + 1)2 when |C| , |Ω|) that solve for the unknown signal from which we
assume the data to have been sampled.
Such a viewpoint, in a sense, embodies a strict geographical prior, and is very
different from the splines and wavelets that will be discussed in the remaining
Sections 4.2 and 4.3. Indeed, in contrast to Slepian functions, splines and wavelets
made from zonal kernel functions do not strictly select for particular regions of
interest, although of course, when particular combinations of any of those con-
structions are sought by optimization, as they are in Section 4.4, effectively, they
do. Simons et al. [89] discuss a hybrid situation termed ‘Slepian trees’, as well as
an alternative spherical wavelet transform obtained via a simple ‘cubed-sphere’
mapping of the ‘usual’ separable Cartesian discrete wavelet transforms [20].
The most detailed and up-to-date discussion of approximation by Slepian
functions (both scalar and vector-valued, and for geomathematics problems in-
volving measurements made by satellites at altitude) is found in the works by
Simons and Plattner [80, 81, 86]. From these references, we retain and present a
few essential points.
Suppose that we have ‘data’, M , consisting of a superposition of ‘signal’, F ,
and ‘noise’, E. What is F ? The measurements are only available over some closed
region C of the unit sphere Ω, i.e.,

F (ξ) + E(ξ) if ξ ∈ C
M (ξ) = (90)
unknown/undesired if ξ ∈ Ω \ C.
We assume that both signal and noise can be represented via an infinite spherical
harmonic expansion as in (4), and we furthermore assume that they are uncor-
related realizations of zero-mean Gaussian random processes. Paying no heed to
the structure of the noise (i.e., without explicit prior information that could be
weighted into the norms in the form of a noise covariance) we elect to seek solutions
to the optimization problem that results in a regularized bandlimited (to N , which
remains to be determined) estimate of the signal, F̂ , in the form of equation (9),
and which solves
- - - -2
-F̂ − M -2 2 + λ-F̂ - 2 = minimum, (91)
L (C) L (Ω\C)
where λ ≥ 0 is a regularization (damping) parameter forcing the solution to vanish

outside of the observation domain. In the following two paragraphs, we distinguish
solutions F̂N and F̂J , both bandlimited.
Approximation 2: Regularized spherical-harmonic expansions. Simons and Dahlen

[85] give the Fourier coefficients that solve equation (91) as

N 2n+1
(Ω\C)
−1
F̂N∧ (m, j) = dC
(m,j),(n,k) + λ d(m,j),(n,k) M, Yn,k L2 (C) . (92)
n=0 k=1
(Ω\C)
We note from equation (74) that dC (m,j),(n,k) + d(m,j),(n,k) is the identity matrix.
Regularization is unavoidable: as we have seen, the eigenvalues of dC (m,j),(n,k) trail
off quickly to nearly zero, see Figure 1. Restricted-region data availability is the
prime reason for our inverse problem to be ill posed – even if no downward con-
tinuation from satellite height is required and if no internal density distributions
(in the case of gravimetry) are being sought.
How well are we doing when accepting (92) as our solution? Rewriting the
(α)
inverse Slepian eigenvalues λC with the damping parameter λ as
∗ . /−1
(α) (α) (α)
λC (λ) = λC + λ 1 − λC , (93)
[85] derive the regional relative mean-squared error, the expected value of the ratio
of approximation-error to signal norms as
- -2
- -
E -F̂N − F - 2
L (C)
(N +1)2 (α) .
λ ∗ /2 ) 2 *
(α) (α) (α)
= C
λC (λ) R −1 λC + λ2 1 − λC .
E F 2L2 (C) α=1
NC
(94)
In the expression above, both signal and noise were assumed to be characterized
by a white (flat) power spectrum (defined in (80)), and we introduced R, the
signal-to-noise ratio. Valid only for this admittedly idealized case, (94) neverthe-
less contains all the elements by which the quality of the approximation can be
appreciated: the bandwidth N and the size and shape of the region C enter through
(α)
the eigenvalues λC and the Shannon number NC , and of course the dependence
on the signal-to-noise ratio R and the damping parameter λ are important control-
ling factors. Minimization of the relative error norm provides an implicit criterion
for the regularization parameter:
+1)2 .
(N ∗ /3 2
(α) (α) (α)
λC (λ) λC 1 − λC
λopt = R −1 α=12 . ∗ /3 2 2 . (95)
(N+1)
(α) (α) (α)
λC (λ) λC 1 − λC
α=1
At high signal-to-noise ratios, (95) is well approximated by λopt ≈ R −1 .

Approximation 3: Truncated Slepian expansions. Where did the Slepian functions
go? We solved (91) using spherical harmonics, but we discussed the statistics of
the solution (92) in terms of the eigenvalues of the Slepian concentration problem.
The link, of course, is that the spherical-harmonic solution is derived via the
Rírms = 34.53, 100.00% Rírms = 33.86, 98.05% Rírms = 34.55, 100.08%

50°
35°
20°
5°
í10°
L = 17í72 , dof = 5329 J = 91 , dírmse = 26.72% J = 182 , dírmse = 0.08%
K = 91 , Nd = 500 , S/N = Inf Rírmse = 13.28, 38.46% Rírmse = 0.22, 0.64%
50°
35°
20°
5°
í10°
í15° 0° 15° 30° 45° í15° 0° 15° 30° 45° í15° 0° 15° 30° 45°
Figure 12. Example of Slepian-basis (Shannon number K = 91) ap-

proximation of a non-white bandlimited (bandpass, spherical-harmonic
degrees L = 17 − 72) geomagnetic field from Nd = 500 noiseless data,
for two truncation levels, J = 91 and J = 182, over a circular domain
R. Top: the field, and the two reconstructions. Bottom: the location of
the data points, and the difference between the truth and the approx-
imation. The relative regional root-mean square signal, reconstruction
and error strengths are indicated.
intermediary of the inverse of the Slepian localization matrix dC

(m,j),(n,k) and, with
(Ω\C)
regularization, its complement, d(m,j),(n,k) . Both of these are large, full (though
banded) matrices whose inverses (especially at large spherical-harmonic degrees
N ) are computed at significant cost. We have previously seen how a partial set
of Slepian functions provides excellent regional approximations in noiseless cases.
To conclude this section, we thus propose an estimator for the situation of the
form (90), where we attempt to reconstruct the unknown signal F from a regionally
observed set of noisy measurements, M .
This time, our estimator does not take the form of a spherical-harmonic
expansion that needs to be regularized (sometimes at great computational cost),
but rather of a Slepian-function expansion which can be truncated (usually without
any difficulty at all). In the context of equation (88): we prefer the approximate
identity over the equality which may well furnish us with a ‘complete’ expansion,
but whose coefficients we can only calculate approximately, after regularization.
Rírms = 34.53, 100.00% Rírms = 32.25, 93.41% Rírms = 34.62, 100.26%

50°
35°
20°
5°
í10°
L = 17í72 , dof = 5329 J = 91 , dírmse = 46.56% J = 113 , dírmse = 31.57%
K = 91 , Nd = 500 , S/N = 10 Rírmse = 14.09, 40.79% Rírmse = 8.04, 23.28%
50°
35°
20°
5°
í10°
í15° 0° 15° 30° 45° í15° 0° 15° 30° 45° í15° 0° 15° 30° 45°
Figure 13. Example of Slepian-basis (Shannon number 91) approxi-

mation of a bandlimited field from data with a signal-to-noise ratio of
10, for two truncation levels, 91 and 113. Layout as in Figure 12.
In the framework of Slepian-function estimation, truncation is our regularization.

The Slepian-basis solutions to the ‘unregularized’ (λ = 0) problem (91) are, quite
simply,
−1 ? @
(α)
F̂ ∧ (α) = λC M, GC N α L2 (C) . (96)
Truncation means that we only compute J of them, which gives us the freedom to
avoid the blowup of the inverse eigenvalues, i.e., the estimate in the Slepian basis
is given by
J

F̂J = F̂ ∧ (α) GC N α. (97)
α=1
By the same metric of (94), we evaluate the quality of this solution as
- -2
- -
E -F̂J − F - 2 (N +1)2
L (C) −1 J 1 (α)
=R + λC . (98)
E F 2L2 (C) NC NC
α>J
As (94), but unlike (89), again (98) is only applicable in the case of white noise and
white signal with a signal-to-noise power ratio R. Of course, the signal contained
in the neglected terms of what should be a complete Slepian expansion exerts
a controlling factor on the mean squared error behavior. [85] show how, in the
Slepian basis, the neglected terms positively affect the variance of the estimate,
but negatively the bias; the mean-squared error being the combination of the
two. Minimization of (98) to determine the optimal truncation level for these
circumstances yields it in terms of the Slepian eigenvalue and the signal-to-noise
level, namely
(J )
λC opt ≈ R −1 . (99)
In other words, we include Slepian functions in the expansion until their ranked
eigenvalues drop below the noise-to-signal ratio.
We have ignored that in order to ‘solve’ data-driven approximation problems,
we need to determine an optimal bandwidth N and an optimal truncation level J
for data situations that are more involved than just being given by white noise
and white signal. Such vital practical matters are discussed by Slobbe et al. [91]
and Plattner and Simons [81]. The solution procedures involved are always cum-
bersome – but the computational complexity, and the overall size of the numerical
problem, of the truncated Slepian-function approach is always smaller than via
regularized spherical-harmonics. Slepian functions lend themselves well to solving
approximation problems involving noisy and partially observed data on the sphere.
Two realistic examples of truncated Slepian-basis approximation problems
are given in Figures 12 and 13.
4.2. Approximation by splines

Only relatively recently have zonal kernel function techniques such as spline in-
terpolation/approximation and wavelet analysis been playing a fundamental role
in numerical analysis on the sphere. Spherical splines (independently introduced
by Freeden [28] andWahba [99] in 1981) are canonical generalizations of ‘spherical
polynomials’ (spherical harmonics) which have desirable characteristics as inter-
polating, smoothing, and best approximating functions (see also [100]). By spline
interpolation we mean the variational problem of minimizing an ‘energy’-norm
of a suitable Sobolev space. Depending on the chosen norm, bandlimited and
non-bandlimited splines are distinguished. Spherical splines have been success-
fully applied to many areas of application in particular in geodesy for gravita-
tional field determination, radio occultation, ocean flow, etc. (for more details see
[29, 36, 47] and the references therein). Spherical splines, especially their counter-
parts on the ball, have been applied to tomographic inverse problems in geophysics
([1–3, 10–12, 22, 74]) and in medical imaging ([27]).
To understand spherical splines, we adopt the idea of one-dimensional cubic
splines to the sphere. Cubic splines in one-dimension are well known for hav-
ing minimal ‘bending energy’ (roughly, minimal ‘curvature energy’ understood
in a linearized sense). More concretely, among all interpolating functions of the
b
Sobolev space H(2) ([a, b]), the integral a |F
(x)| dx becomes minimal, where F

may be physically interpreted as the deflection normal to the rest position which
is supposed to be horizontal. The physical model is suggested by the classical inter-
pretation of the potential energy of a statically deflected thin beam which indeed
is proportional to the integral taken over the square of the linearized curvature of
the elastic
beam. Analogously, the concept can be applied to the sphere by choos-
ing Ω |Δ∗ξ F (ξ)|2 dω(ξ), where F now denotes the deflection of a thin membrane
normal to the rest position supposed to be spherical. In other words, the second
derivative canonically takes on the form of the Beltrami operator Δ∗ . Indeed,
our interest now is to state that the interpolating spline to a given dataset has
minimum ‘bending energy’ for all interpolants within the Sobolev space H(2) (Ω).
Furthermore, the spline functions defined in this section are able to simultaneously
interpolate and smooth the data. Hence, we can decide in our spline application,
which knots of the input data should be strictly interpolated and which ones should
be ‘near’ the interpolating function, i.e., the points subjected to smoothing.
Reproducing kernel Hilbert reference space. As usual (see, e.g., [47]), we introduce
the Sobolev space H(2) (Ω) as the completion of C(2) (Ω) with respect to a specific
scalar product thereby specifying H(2) (Ω) as a certain reproducing kernel space.
In more detail, the inner product ·, ·H(2) (Ω) is defined by

F, GH(2) (Ω) = F (η)Y0,1 (η) dω(η) G(η)Y0,1 (η) dω(η)
Ω Ω

=F,GH0
∞ 2n+1

+ (n(n + 1))2 F (η)Yn,j (η) dω(η) G(η)Yn,j (η) dω(η),
n=1 j=1 Ω Ω

=F,GH⊥
0
(100)
which is equivalent in accordance with Parseval’s identity to

F, GH(2) (Ω) = F (η)Y0,1 (η) dω(η) G(η)Y0,1 (η) dω(η)
Ω Ω
=F,GH0

+ (Δ∗η F (η))(Δ∗η G(η)) dω(η)
Ω

=F,GH⊥
0
=F, GH0 + F, GH⊥

0
(101)
for all F, G ∈ C(2) (Ω). The Sobolev space H(2) (Ω) as defined in Section 3.1 is the
completion of C(2) (Ω) under the norm · H(2) (Ω) , i.e.,
·H(2) (Ω)
H(2) (Ω) = C(2) (Ω) , (102)

where F H(2) (Ω) = F, F H(2) (Ω) .
Consider the kernel K : (ξ, η) → K(ξ, η), (ξ, η) ∈ Ω2 = Ω × Ω given in the

form
∞ 2n+1
1
K(ξ, η) = Y0,1 (ξ)Y0,1 (η) + Yn,j (ξ)Yn,j (η). (103)
n=1 j=1
(n(n + 1))2
Then K(ξ, ·), ξ ∈ Ω fixed, is a member of H(2) (Ω). Inserting K into the inner
product, we see via the orthogonal properties of the spherical harmonics that
F, K(ξ, ·)H(2) (Ω) = F (ξ), ξ ∈ Ω, (104)
for all F ∈ H(2) (Ω). Hence, K(·, ·) is the unique reproducing kernel of the Hilbert
space H(2) (Ω). The reproducing kernel K(·, ·) can be decomposed into the repro-
ducing kernels of the spaces H0 and H⊥ 0 , respectively, via
∞ 2n+1
1
K(ξ, η) = Y0,1 (ξ)Y0,1 (η) + Yn,j (ξ)Yn,j (η) . (105)
n=1 j=1 (n(n + 1))2
1
=K0 (ξ,η)= 4π
=K0⊥ (ξ,η)
Applying the spherical-harmonic addition theorem and comparing with (105),

we get
∞ 2n+1
1
K0⊥ (ξ, η) = Yn,j (ξ)Yn,j (η)
n=1 j=1
(n(n + 1))2
∞
1 2n + 1
= Pn (ξ · η)
4π n=1
(n(n + 1))2
= G((Δ∗ )2 ; ξ, η) (106)
where G((Δ∗ )2 , ·, ·) is the Green function with respect to the iterated Beltrami
operator (Δ∗ )2 = Δ∗ Δ∗ (see [28]). Summarizing our results we, therefore, see that

2
K(ξ, η) = Y0,1 (ξ)Y0,1 (η) + G (Δ∗ ) ; ξ, η , ξ, η ∈ Ω, (107)

1
=K0 (ξ,η)= 4π
=K0⊥ (ξ,η)
is the uniquely determined reprokernel of the space (H(2) (Ω), ·, ·H(2) (Ω) ), i.e.:
(i) For each fixed ξ ∈ Ω, K(ξ, η), a function of η, is an element of H(2) (Ω).
(ii) For every function F ∈ H(2) (Ω) and for every point ξ ∈ Ω, the reproducing
property holds:
F (ξ) = F, K(ξ, ·)H(2) (Ω) . (108)
Explicit representation of the reproducing kernel. Keeping the reprostructure of

H(2) (Ω) in mind, we are able to handle our announced spline interpolation and
smoothing problem. To this end, we follow the concept presented by Freeden [28]
and observe, in addition, the explicit representation of G((Δ∗ )2 ; ·, ·) known from
[41]. In fact, Green’s function corresponding to the iterated Beltrami operator

(Δ∗ )2 is continuous on Ω × Ω and admits the explicit formulation:
⎧ 1
⎪
⎪ 4π , 1−ξ·η =0
⎪
⎪
⎪
⎪
⎨ 1
G((Δ∗ )2 ; ξ, η) = 4π (1 − ln(1 − ξ · η)(ln(1 + ξ · η) − ln(2))
1−t 2
⎪
⎪ −L 2 ( 2 ) − (ln(2)) + ln(2) ln(1 + ξ · η)), 1 ± ξ · η = 0
⎪
⎪
⎪
⎪
⎩ 1
4π − 24 , 1 + ξ · η = 0,
π
(109)
where the function L2 (x) is the dilogarithm given by
x ∞

ln(1 − t) xk
L2 (x) = − dt = . (110)
0 t k2
k=1
Spline concept. We have come to the definition of spherical splines corresponding

to one-dimensional cubic splines (a more general concept involving pseudo-differ-
ential operators is known from [47]). Let M1 , . . . , Mn be a linearly independent
system of bounded linear functionals on H(2) (Ω). Any function S ∈ H(2) (Ω) of the
form

n
S(η) = c0 Y0,1 (η) + ai Mi G((Δ∗ )2 ; η, ·), ai ∈ R, η ∈ Ω (111)
i=1
with

n
ai Mi (Y0,1 ) = 0 (112)
i=1
is called a spherical spline in H(2) (Ω) relative to M1 , . . . , Mn .

By virtue of (111) and (112), we are now prepared to formulate the following
uniqueness result: let M1 , . . . , Mn be a system of bounded linear functionals on
the Sobolev space H(2) (Ω) such that the ((n + 1) × (n + 1))-matrix

α κ
(113)
κT 0
is non-singular, where the matrix α and the vector κ are given as follows:

α = Mi Mj G((Δ∗ )2 ; ·, ·) i=1,...,n , (114)
j=1,...,n
κ = (Mi Y0,1 (·))i=1,...,n . (115)
Then, there exists a unique spline in H(2) (Ω) relative to M1 , . . . , Mn that solves the
interpolation problem Mi S = μi , i = 1, . . . , n. This spline is called the interpolating
spline. The proof easily follows by inserting the representation (111) into Mi S = μi ,
i = 1, . . . , n, resulting in a linear system for the coefficients ai , c0 , whose coefficient
matrix is given by (113). Since the matrix is assumed to be non-singular, the
coefficients are uniquely determined.
The key to spline approximation is the so-called spline integration formula

n
Δ∗η S(η) Δ∗η F (η) dω(η) = ak Mk F, (116)
Ω k=1
valid for the uniquely determined interpolating spline S and all members F ∈
H(2) (Ω), provided that the constraint κT a = 0 is fulfilled. The proof is a di-
rect
n conclusion of the reproducing kernel property. Its idea is to inspect the sum
k=1 ak Mk F and substitute F by the reproducing kernel property

n
n
ak M k F = ak Mk F (·), K(η, ·)H(2) (Ω) . (117)
k=1 k=1
Evaluating the inner product by inserting the reproducing kernel function leads
to the desired result.
Next, we turn to dealing with the ‘minimum energy property’ of strict spline
interpolation.
Theorem 4. Let M1 , . . . , Mn be a linearly independent system of bounded linear
functionals on H(2) (Ω). Let S be the unique spline which solves the interpolation
problem Mi S = μi , i = 1, . . . , n. Then, for all twice continuously differentiable
functions F on Ω, which interpolate the given data, i.e., Mi F = μi , i = 1, . . . , n,
the following inequality

(Δ∗η S(η))2 dω(η) ≤ (Δ∗η F (η))2 dω(η) (118)
Ω Ω
holds true with equality if and only if S = F .
The proof easily follows from arguments given by Freeden [28]. Theorem 4
tells us that the ‘bending energy’ (the integral over the second derivative) of the
spline is minimal among all functions in H(2) (Ω) interpolating the data.
Combined spline interpolation and smoothing. Theorem 4 allows an extension to
include smoothing at predefined points while interpolating the remaining pointset
(in accordance with [46]). This technique was used by Blick and Freeden [13] to
visualize radio occultation data collected by the satellite CHAMP.
Given n = p + q data points, where the data points μi , i = 1, . . . , p, are
subjected to smoothing and the points νi , i = 1, . . . , q, are subjected to strict
interpolation, we are lead to the following result.
Theorem 5. Suppose that δ and β12 , . . . , βp2 are prescribed positive weights and
that μi , i = 1, . . . , p; νj , j = 1, . . . , q are given data points. Let M1 , . . . , Mp and
N1 , . . . , Nq be systems of bounded linear functionals on H(2) (Ω) such that the ((p +
q) + 1) × ((p + q) + 1)-matrix
⎛ ⎞
α β κ
⎝ βT γ ζ ⎠ (119)
κT ζ T 0
is non-singular, where the matrices α, β, γ, κ, ζ are given as follows

2
α = Mi Mj G (Δ∗ ) ; ·, · + δβi2 δij i=1,...,p , (Kronecker δij ) (120)
j=1,...,p
∗ 2
β = Mi Nj G (Δ ) ; ·, · i=1,...,p , (121)
j=1,...,q
∗ 2
γ = Ni Nj G (Δ ) ; ·, · i=1,...,q , (122)
j=1,...,q
κ = (Mi Y0,1 (·))i=1,...,p , (123)

ζ = (Nj Y0,1 (·))j=1,...,q . (124)
Then the smoothing spline function S of the form

p q
2 2
S(ζ) = c0 Y0,1 (ξ) + ai Mi G (Δ∗ ) ; ξ, · + bj Nj G (Δ∗ ) ; ξ, · , ξ ∈ Ω,
i=1 j=1
(125)
T T
with coefficients a ∈ R , a = (a1 , . . . , ap ); b ∈ R , b = (b1 , . . . , bq ) and c0 ∈ R
p q
subjected to the constraint

p
q
ai Mi (Y0,1 ) + bj Nj (Y0,1 ) = 0 (126)
i=1 j=1
is the unique solution of the interpolation and smoothing problem given by

Mi S + δβi2 ai = μi , i = 1, . . . , p,
Nj S = νj , j = 1, . . . , q,
corresponding to the data points μi , i = 1, . . . , p; νj , j = 1, . . . , q and represents
the only element of H(2) (Ω) satisfying
p 2 p 2
M i S − μi M i F − μi
+ δS, SH0⊥ ≤ + δF, F H⊥ (127)
i=1
β i=1
βi 0
for all F ∈ H(2) (Ω) with Nj F = νj , j = 1, . . . , q.

As already mentioned, the proof can be given in parallel to the arguments
stated by Freeden and Witte [46]. Moreover, Theorem 4 leads us to the following
comments:
(i) The values μ1 , . . . , μp , ν1 , . . . , νq are regarded as the observed quantities, e.g.,
geodetic observations and measurements.
(ii) The spline function S ∈ H(2) (Ω) satisfies that Mi S is ‘near’ μi , i = 1, . . . , p
and Nj S is equal to νj , j = 1, . . . , q. The ‘nearness’ of the values Mi S to μi ,
i = 1, . . . , p can be controlled by choosing the constant δ in a suitable way.
A small value of δ emphasizes fidelity to the observed data at the expense of
smoothness, while a large value does the opposite.
(iii) Taking δ = 0 yields Mi S = μi , i = 1, . . . , p, i.e., the combined smoothing and
interpolation procedure leads back to strict interpolation.
(iv) For numerical purposes, it is advantageous to adapt the quantities β12 , . . . , βp2
to the standard deviations of the measured values.
4.3. Approximation by wavelets

As already pointed out, the context of the spectral representation of a square-
integrable function by means of spherical harmonics is essential to solving many
problems in today’s applications. In future research, however, orthogonal (Fourier)
expansions in terms of spherical harmonics {Yn,j } will not be the only way of rep-
resenting a square-integrable function. In order to explain this in more detail, we
think of a square-integrable function as a signal in which the spectrum evolves
over space in significant way. We imagine that, at each point on the sphere Ω, the
function refers to a certain combination of frequencies, and that these frequen-
cies are continuously changing. This space-evolution of the frequencies, however,
is not reflected in the Fourier expansion in terms of non-space localizing spher-
ical harmonics, at least not directly. Therefore, in theory, any member F of the
space L2 (Ω) can be reconstructed from its Fourier transforms, i.e., the ‘amplitude
spectrum’ {F ∧ (n, j)} n=0,1,..., , but the Fourier transform contains information
j=1,...,2n+1
about the frequencies of the function over all positions instead of showing how the
frequencies vary in space.
Dirac families. In what follows, we present a two-parameter, i.e., scale- and space-
dependent method of achieving a reconstruction of a function F ∈ L2 (Ω) involving
(scalar) zonal kernel functions which we refer to as a Dirac family {Φρ }ρ∈(0,∞)
converging to the (zonal) Dirac kernel δ. In other words, a Dirac family is a set of
zonal kernels Φρ : [−1, 1] → R, ρ ∈ (0, ∞), of the form
∞
2n + 1
Φρ (ξ · η) = Φ∧
ρ (n) Pn (ξ · η), ξ, η ∈ Ω, (128)
n=0
4π
converging to the ‘Dirac kernel’ δ as ρ → 0, ρ > 0. Consequently, if {Φρ }ρ∈(0,∞) is
a Dirac family, its ‘symbol’ {Φ∧
ρ (n))}n=0,1,... constitutes a sequence satisfying the
limit relation
lim Φ∧ ρ (n) = 1, n = 0, 1, . . . . (129)
ρ→0, ρ>0
Accordingly, if {Φρ }ρ∈(0,∞) is a scaling kernel function, the convolution integrals

(Φρ ∗ F ) (ξ) = Φρ (ξ · η)F (η) dω(η), ξ ∈ Ω, (130)
Ω
converge (in a certain topology) to the limit

F (ξ) = (δ ∗ F )(ξ) = δ(ξ · η)F (η) dω(η), ξ ∈ Ω, (131)
Ω
for all ξ ∈ Ω as ρ tends to 0 (from the positive side). In more detail, if F is a
function of class L2 (Ω) and {Φρ } is a (suitable) Dirac family (tending to the Dirac
kernel), then the following limit relation holds true:
lim F − Φρ ∗ F L2 (Ω) = 0. (132)
ρ→0, ρ>0
There is a large number of Dirac families that is of interest for geoscien-

tific application (for more details, the reader is referred to, e.g., [39, 44] and the
references therein). Only three prototypes of Dirac families should be mentioned

here: the bandlimited Shannon family, the neither bandlimited nor spacelimited
Abel–Poisson and Gauss–Weierstraß families, and the spacelimited Haar family.
It should be noted that an approximate convolution identity (132) acts as a
space and frequency localization procedure in the following way. As {Φρ }ρ∈(0,∞)
is a Dirac family of zonal scalar kernel functions tending to the Dirac kernel, the
function Φρ (η·), is highly concentrated around the point η ∈ Ω, if the ‘scale pa-
rameter’ is a small positive value. Moreover, as ρ tends to infinity, Φρ (η·) becomes
more and more localized in frequency. Correspondingly, the uncertainty principle
states that the space localization of Φρ (η·) becomes more and more decreasing.
In conclusion, the products η → Φρ (ξ · η)F (η), η ∈ Ω, ξ ∈ Ω, for each fixed
value ρ, display information in F ∈ L2 (Ω) at various levels of spatial resolution
bands. Consequently, as ρ approaches ∞, the convolution integrals
or frequency
Φρ ∗ F = Ω Φρ (·η)F (η) dω(η) display coarser, lower-frequency features. As ρ ap-
proaches 0, the integrals give sharper and sharper spatial resolution. Thus, the
convolution integrals can measure the space-frequency variations of spectral com-
ponents, but they have a different space-frequency resolution.
Scaling and wavelet functions. Next we come to the bilinear theory of scaling and
wavelet functions (note that we only deal with the bilinear theory, for basic aspects
of the linear case the reader is referred to, e.g., [39, 44]).
The point of departure for our multi-scale approach is a particular type of
(2)
a Dirac family: a scaling (kernel) function {Φρ }ρ∈(0,∞) is a set of zonal kernels
(2)
Φρ = Φρ ∗ Φρ : [−1, 1] → R, ρ ∈ (0, ∞), of the form
∞
∧ 2n + 1
Φ(2)
ρ (ξ · η) = Φ(2)
ρ (n) Pn (ξ · η), ξ, η ∈ Ω, (133)
n=0
4π
with
∧
lim Φ(2)
ρ (n) = lim (Φ∧ 2
ρ (n)) = 1, n = 0, 1, . . . . (134)
ρ→0, ρ>0 ρ→0, ρ>0
and
∧
Φ(2)
ρ (0) = 1. (135)
Accordingly, the convolution integrals

Φ(2)
ρ ∗ F (ξ) = Φ(2)
ρ (ξ · η)F (η) dω(η), ξ ∈ Ω, (136)
Ω
converge (in a certain topology) to the limit

F (ξ) = (δ ∗ F )(ξ) = δ(ξ · η)F (η) dω(η), ξ ∈ Ω, (137)
Ω
for all ξ ∈ Ω as ρ tends to 0 (from the positive side). In other words, if F is a

(2)
function of class L2 (Ω) and {Φρ } is a certain Dirac family (tending to the Dirac
kernel), then the approximate identity

- -
- -
lim -F − Φ(2)ρ ∗ F- =0 (138)
ρ→0, ρ>0 L2 (Ω)
holds true.
(2)
Each scale approximation Φρ ∗ F of a function F ∈ L2 (Ω) must be made
directly by computing the relevant convolution integrals. In doing so, however, it
(2)
is inefficient to use no information from the approximation Φρ ∗ F within the
(2)
computation of Φρ ∗ F provided that ρ
< ρ. In fact, the efficient construction of
multiscale approximation based on Dirac families usually begins by a multireso-
lution analysis in terms of wavelets, i.e., a recursive method which is efficient for
computation, but not all economic multiscale approaches constitute multiresolu-
tion procedures (see, e.g., [35, 36, 38, 40–43, 47] and the references therein).
Let Ψρ (ξ, η), (ξ, η) ∈ Ω × Ω, be defined via the series expansion
∞
2n + 1
Ψρ (ξ, η) = Ψρ ∧ (n) Pn (ξ · η), (ξ, η) ∈ Ω × Ω, (139)
n=0
4π
(2) ∧ (2) (2) ∧
such that the symbol {Ψρ (n)}n=0,1,... of Ψρ = Ψρ ∗Ψρ is derived from Φρ (n)
via the differential equation (‘scale equation’)
∧ d (2) ∧
Ψ(2)
ρ (n) = −ρ Φ (n). (140)
dρ ρ
As immediate consequences, we obtain from (135) the properties
Ψρ ∧ (0) = 0 (141)
and
lim Ψ∧
ρ (n) = 0
ρ→0, ρ>0
for n = 1, 2, . . . As in classical one-dimensional theory, the condition (135), there-

fore, justifies the notion wavelet of order 0.
Typically, within wavelet nomenclature, we may write
Ψρ;η : ξ → Ψρ;η (ξ) = Ψρ (ξ · η) = Rη Dρ Ψ(·ξ), ξ ∈ Ω,
to indicate Ψρ;η as generated by two parameters, namely the ‘η-rotation operator ’
Rη and the ‘ρ-dilation operator ’ Dρ , respectively, given by
Rη : Ψ(·ξ) → Rη Ψ(·ξ) = Ψ(η · ξ), (142)
Dρ : Ψ(·ξ) → Dρ Ψ(·ξ) = Ψρ (·ξ). (143)
The function Ψ = Ψ1 (i.e., ρ = 1) is called the mother wavelet.
The wavelet transform WT is defined as the L2 (Ω)-inner product (convolu-
tion) of F ∈ L2 (Ω) with the set of ‘rotations’ and ‘dilations’ of F

(WT)(F )(ρ; η) = (Ψρ;η , F )L2 (Ω) = Ψρ;η (ξ)F (ξ) dω(ξ), (144)
Ω
i.e., the wavelet transform acts as a space and frequency localization operator.
The wavelet transform (WT) is invertible on the space of functions F ∈ L2 (Ω)
satisfying F ∧ (0, 1) = 0, i.e.,
∞
dρ
F = (WT)(F )(ρ; η)Ψρ;η (·) dω(η) (145)
Ω 0 ρ
holds true (in the sense of · L2 (Ω) ) for all F ∈ L2 (Ω) satisfying F ∧ (0, 1) = 0.
The reconstruction formula (145), in fact, is based on the simple idea of
dilation and rotation of the mother wavelet.
Figure 14. Shannon scaling (kernel) functions for decreasing scales ρ.
Figure 15. Shannon wavelet (kernel) functions for decreasing scales ρ.
Spectral interrelation between Fourier and wavelet transform. In terms of filter-

ing, {Φρ }ρ∈(0,∞) and {Ψρ }ρ∈(0,∞) may be interpreted (cf. Figures 14 and 15) as
lowpass filter and bandpass filter, respectively. Correspondingly, the convolution
operators are given by
Φρ ∗ F, F ∈ L2 (Ω), (146)
2
Ψρ ∗ F, F ∈ L (Ω). (147)
The Fourier transforms read as follows:
(Φρ ∗ F )∧ (n, j) = F ∧ (n, j)Φ∧
ρ (n), (148)
∧ ∧
(Ψρ ∗ F ) (n, j) = F (n, j)Ψ∧
ρ (n). (149)
These formulas provide the transition from the wavelet transform to the Fourier
transform. Since all scales ρ are used, the reconstruction is highly redundant.
If F, G ∈ L2 (Ω) have vanishing moments of order 0, i.e., if the property
F (0, 1) = G∧ (0, 1) = 0 is satisfied, then it follows from
∧
∞
2 dρ
(Ψ∧ρ (n)) =1 (150)
0 ρ
and the Parseval identity of the theory of spherical harmonics that

∞
dρ
F, Ψρ;η L2 (Ω) G, Ψρ;η L2 (Ω) dω(η)
Ω 0 ρ
∞ ∞ 2n+1

2 dρ
= F ∧ (n, j)G∧ (n, j)(Ψ∧ ρ (n))
0 n=1 j=1
ρ
∞ 2n+1

= F ∧ (n, j)G∧ (n, j)
n=1 j=1
= F, GL2 (Ω) . (151)
Denote by L2 ((0, ∞) × Ω) the space of all integrable functions H : (0, ∞) ×

Ω → R such that ∞
dρ
|H(ρ; η)|2 dω(η) < ∞. (152)
Ω 0 ρ
On the space L2 ((0, ∞) × Ω), an inner product ·, ·L2 ((0,∞)×Ω) can be imposed
corresponding to the norm
∞ 1/2
2 dρ
H L2 ((0,∞)×Ω) = |H(ρ; η)| dω(η) . (153)
Ω 0 ρ
From (151), it follows that
2 3
F, Ψ·,· L2 (Ω) , G, Ψ·,· L2 (Ω) = F, GL2 (Ω) (154)
L2 ((0,∞)×Ω)
and
F, Ψ·,· L2 (Ω) 2L2 ((0,∞)×Ω) = F 2L2 (Ω) . (155)
In other words, the total energy of a signal can be continuously distributed by the
wavelet transform into scale and spatially dependent ‘signal subenergy’.
Least energy representation. WT is a transformation from the one-parameter

space L2 (Ω) into the two-parameter space L2 ((0, ∞) × Ω). Thus, it is clear that
(WT) is not surjective on L2 ((0, ∞) × Ω). That means that W = (WT)(L2 (Ω)) is
a proper subspace of L2 ((0, ∞) × Ω):
W L2 ((0, ∞) × Ω). (156)
Thus, the problem is to characterize W within the framework of L2 ((0, ∞) × Ω).
For that purpose, we consider the operator P : L2 ((0, ∞) × Ω) → W given by
∞

dρ
(P H)(ρ ; η ) = K(ρ
; η
| ρ; η)H(ρ; η) dω(η) , (157)
0 Ω ρ
where

K(ρ ; η | ρ; η) = Ψρ ;η (ξ)Ψρ;η (ξ) dω(ξ).

Ω
W is characterized as follows: H ∈ W if and only if

∞
dρ
H(ρ
; η
) = K(ρ
; η
| ρ; η)H(ρ; η) dω(η) . (158)
0 Ω ρ
It can easily be seen that K(ρ

; η
| ·; ·) ∈ W and K(·; · | ρ; η) ∈ W. The kernel
K(ρ
; η
| ρ; η) is the reproducing kernel in W. The reproducing property (158)
can also be understood in such a way that H ∈ W is calculable by superpositions
of itself. This shows that there is a kind of linear dependence, which can be in-
terpreted as redundancy. Although it might seem inefficient, such redundancy has
certain advantages. Unlike a non-redundant expansion, errors can be detected and
corrected.
The tendency for correcting errors is expressed in the next result (see [35, 47]):
– Let H be an arbitrary element of L2 ((0, ∞) × Ω). Then the unique function
FH ∈ L2 (Ω) which satisfies the property
- - - -
- - - -
-H − F̃H - 2 = inf2 -H − F̃ -
2
L ((0,∞)×Ω) F ∈L (Ω) L ((0,∞)×Ω)
(with F̃H = (WT)(FH )) is given by

∞
dρ
FH (ξ) = H(ρ; η)Ψρ;η (ξ) dω(η) .
0 Ω ρ
Indeed, F̃H is the orthogonal projection of H onto W, which explains the afore-
mentioned statement.
The linear dependence of F̃ ∈ W leads to the effect that the coefficients in
L2 ((0, ∞) × Ω) for reconstructing a function F ∈ L2 (Ω) are not unique. This can
be easily seen from the following identity:
∞ dρ
F (ξ) = F̃ (ρ; η) + F̃ ⊥ (ρ; η) Ψρ;η (ξ) dω(η)
0 Ω ρ
where F̃ = (WT)(F ) and F̃ ⊥ is an arbitrary member of W ⊥ . Nevertheless,

we are able to deal with the following question: given an arbitrary H(ρ; ξ) =
(WT)(F )(ρ; ξ), ρ ∈ (0, ∞), and ξ ∈ Ω, for some F ∈ L2 (Ω), how can we re-
construct F ? The answer (see [35, 47]) is provided by the so-called least-energy
representation:
– Of all possible functions H ∈ L2 ((0, ∞) × Ω) for F ∈ L2 (Ω), the function
H = (WT)(F ) is unique in that it minimizes the ‘energy’ H 2L2 ((0,∞)×Ω) .
More explicitly,
(WT)(F ) L2 ((0,∞)×Ω) = inf H L2 ((0,∞)×Ω) .

H∈L2 ((0,∞)×Ω)
(WT)−1 (H)=F
Wavelet variants. The construction of spherical wavelets has seen an enormous

increase of activity in the last few years. Three features are essential in the thinking
about georelevant wavelets: basis property, decorrelation, and fast computation.
First, wavelets are building blocks for general datasets derived from functions.
By virtue of the basis property, each element of a general class of functions (e.g.,
a geopotential seen as a member of a set of potentials within a Sobolev space
framework) can be expressed in stable way as a linear combination of dilated and
shifted copies of a ‘mother function’ (see [29, 31, 35, 36] and the references therein).
The role of the wavelet transform as a mapping from the class of functions into an
associated two-parameter family of space and scale dependent functions is properly
characterized by least squares properties.
Second, wavelets have the power to decorrelate. In other words, the rep-
resentation of data in terms of wavelets is somehow ‘more compact’ than the
original representation. We search for an accurate approximation by only using a
small fraction of the original information of a function. Typically, the decorrela-
tion is achieved by building wavelets which have a compact support (localization
in space), which are smooth (decay towards high frequencies), and which have
vanishing moments (decay towards low frequencies). Different types of wavelets
can be found from certain constructions of space/momentum localization. The
uncertainty principle tells us that sharp localization in ‘space and momentum’
are mutually exclusive. Nevertheless, decay towards long and short wavelengths
(i.e., bandpass filtering) can be assured without any difficulty. Moreover, vanishing
moments of wavelets (see, e.g., [45, 47]) enable us to combine (polynomial) outer
harmonic expansions (responsible for the long-wavelength part of a function) with
wavelet multiscale expansions (responsible for the medium-to-short-wavelengths
contributions).
Third, the main question of recovering a function on the sphere, e.g., the
Earth’s gravitational potential, is how to decompose the function into wavelet
coefficients, and how to reconstruct efficiently the potential from the coefficients.
There is a ‘tree algorithm’ or ‘pyramid algorithm’ (cf. [29, 47]) that makes these
steps simple and fast. In this respect, it is desirable to switch between the original
representation of the data and its wavelet representation in a time proportional to
the size of the data. In fact, the fast decorrelation power of wavelets is the key to
applications such as data compression, fast data transmission, noise cancelation,
signal recovery, etc.
In the last years, wavelets on the sphere have been the focus of several research
groups which led to different wavelet approaches. Common to all these proposals
is a multiresolution analysis which enables a balanced amount of both frequency
(more accurately, angular momentum) and space localization (see, e.g., [18, 64, 82,
83, 101]).
A group theoretical approach to a continuous wavelet transform on the sphere
is followed by Antoine and Vandergheynst [5], Antoine et al. [4], and Holschneider
[55]. The parameter choice of their continuous wavelet transform is the product
of SO(3) (for the motion on the sphere) and R+ (for the dilations). A continuous
wavelet transform approach for analyzing functions on the sphere is presented by

Dahlke and Maass [19].
The Kaiserslautern constructions (see, e.g., [32, 39, 43, 47]) are intrinsically
based on the specific properties concerning the theory of spherical harmonics.
Wavelet regularization and multiresolution techniques are applied to ‘downward
continuation’ of spaceborne (satellite) observations (see, e.g., [31, 37, 38, 48] and
the references therein). Multiscale signal-to noise ratio modeling is done by signal
and noise decorrelation Freeden and Maier [33, 34]. Freeden and Schreiner [42] are
interested in a compromise connecting zonal function expressions and structured
grids on the sphere to obtain fast algorithms. Freeden et al. [49] and Freeden and
Gerhards [31] generate locally supported wavelets by regularizing fundamental
solutions to pseudodifferential operators.
Finally, much of the material presented in this paper within a spherical frame-
work can be readily formulated for non-spherical reference surfaces, even for vec-
tor and tensor data. Nevertheless, work remains to be done for more realistic
geometries such as (the actual) Earth’s surface, real satellite orbits, etc. These are
challenges for future research.
4.4. Regularized functional matching pursuit and its variants

The Regularized Functional Matching Pursuit (RFMP) and its variants were de-
veloped by Fischer [23], Fischer and Michel [24], Michel [69], Michel and Telschow
[72, 73], and Telschow [95]. They are based on the Matching Pursuit (MP) and its
enhancements as described by Mallat and Zhang [65] and Vincent and Bengio [98],
where the problem consisted of finding a greedy algorithm for the approximation
of an unknown signal F based on given samples F (xj ), j = 1, . . . , N , usually on
Euclidean domains.
For the RFMP, matching pursuit had to be extended to the inverse problem
FF = y
for a linear and continuous operator F : H(D) → Rl , a Hilbert space H(D) of
(some) functions on D ⊂ Rd (e.g., L2 (D) or, more generally, a Sobolev space),
a given data vector y ∈ Rl and an unknown function F ∈ H(D). Many inverse
problems of this kind, such as the downward continuation (F is the gravitational
potential at the surface D(= Ω) and y is a vector of samples at satellite height)
or the inverse gravimetric problem (F is a volume or a surface mass distribution
and y is a vector of samples of the gravitational potential), are ill posed. For this
reason, a regularization technique also had to be included into the RFMP.
The different algorithms are summarized here starting with a short introduc-
tion of the MP. All algorithms have in common that a set of possibly useful trial
functions, the ‘dictionary’ D ⊂ H(D), is chosen in advance. These trial functions
need not originate from one single basis system: D may be (and is often chosen on
purpose as) overcomplete. If D is heterogeneous, it may contain different kinds of
basis systems (in particular, with different frequency and space localization). For
instance, in several numerical applications of the RFMP, an approximate solution
F was combined from spherical harmonics (for a coarse global approximation)

and radial basis functions with different levels of localization (locally improving
the result). Without loss of generality, one can assume that d H(D) = 1 for all
d ∈ D.
Matching pursuit. Assume that a function (signal) F ∈ H(D) is to be approx-
imated by m elements of D. In this context, the expression of the best-m-term
approximation (see, e.g., [96]) occurs. It means that one looks for m elements
d1 , . . . , dm ∈ D and associated coefficients α1 , . . . , αm ∈ R such that the approxi-
mation error - -
- m -
- -
-F − αk dk -
- -
k=1 H(D)
becomes minimal in comparison to all other choices of dk and αk . In formal lan-
guage, the objective is
- -
-
m -
- -
σm (F, D) = inf -F − αk dk - .
dj ∈D, αj ∈R; j=1,...,m - -
k=1 H(D)
For large m, it is often numerically too expensive to find an exact minimizer.

However, this concept can be a guideline for the construction of a less expensive
algorithm with still ‘good’ results.
The first idea is to construct an iterative algorithm, i.e., to find the pairs
(α1 , d1 ), . . . , (αm , dm ) consecutively. The initial problem is to find α1 ∈ R and
d1 ∈ D such that
J (α1 , d1 ) = F − α1 d1 2H(D) = F 2H(D) − 2α1 F, d1 H(D) + α21 (159)
∂
is minimal. With ∂α1 J(α1 , d1 ) = 0, one obtains
−2 F, d1 H(D) + 2α1 = 0, i.e., α1 = F, d1 H(D) .
Inserting this result in (159), one gets
J(α1 , d1 ) = F 2H(D) − F, d1 2H(D) .
Consequently, this dictionary element d1 ∈ D for which F has the largest projec-
tion, i.e., the dictionary element which is most collinear to F , is the optimal choice
in the first step. The first approximation is, therefore,
F1 = F, d1 H(D) d1 ,
where d1 ∈ D is a maximizer of F, d1 2H(D) , i.e.,
d1 = arg max F, d2H(D) .
d∈D
1
With the residual R = F − F1 , one can analogously proceed. In general, if Rn is
given, then one has to find dn+1 ∈ D such that Rn , dn+1 2H(D) is maximal and
then sets
Fn+1 = Fn + Rn , dn+1 H(D) dn+1 .
Functional matching pursuit. In the case of an inverse problem F F = y, one

minimizes the data misfit
2
J (α1 , d1 ) = y − F (α1 d1 ) Rl
2
= y 2Rl − 2α1 y, F d1 Rl + α21 F d1 Rl ,
∂
which implies that, again by assuming that ∂α1 J(α1 , d1 ) = 0,
y, F d1 Rl
α1 = .
F d1 2Rl
Consequently,
2
y, F d1 Rl
J (α1 , d1 ) = y 2Rl − 2
F d1 Rl
shows that d1 has to be chosen such that
2
y, F d1 Rl
Fd1 2Rl
is maximal. Then,
y, F d1 Rl
F1 = d1
F d1 2Rl
is the first approximation. With the residual R1 = y − FF1 , one proceeds again
analogously. Hence, for a given residual Rn , one chooses dn+1 such that
Rn , F dn+1 2Rl
2
Fdn+1 Rl
is maximal and we set
Rn , F dn+1 Rl
Fn+1 = Fn + 2 dn+1 .
Fdn+1 Rl
Regularized functional matching pursuit. For the handling of ill-posed inverse
problems, the Regularized Functional Matching Pursuit (RFMP) includes a Tikho-
nov-type regularization term
2
λ Fn H(D) ,
where λ ∈ R+ is a regularization parameter. Note that the choice of the (Sobolev)
space H(D) influences the obtained result by requiring a particular kind of ‘smooth-
ness’. For instance, the Sobolev space H(2) (Ω) yields a regularization term which
is not equal but similar to the norm which occurs in the minimum principle of
spherical spline interpolation (see Theorem 4).
In analogy to the above, let Fn ∈ H(D) be the approximation after iteration n
and Rn = y − FFn be the residual, the error on the right-hand side of the inverse
problem F F = y. We find dn+1 ∈ D and αn+1 ∈ R such that
Jλ (αn+1 , dn+1 ) = Rn − F (αn+1 dn+1 ) 2Rl + λ Fn + αn+1 dn+1 2H(D)
is minimal. Treating the functional Jλ like J above, one obtains

2 2
Jλ (αn+1 , dn+1 ) = Rn Rl − 2αn+1 Rn , F dn+1 Rl + α2n+1 Fdn+1 Rl

+ λ Fn 2H(D) + 2αn+1 Fn , dn+1 H(D) + α2n+1 , (160)
∂
where the necessary condition ∂αn+1 Jλ (αn+1 , dn+1 ) = 0 yields
Rn , F dn+1 Rl − λ Fn , dn+1 H(D)
αn+1 = 2 . (161)
F dn+1 Rl + λ
If one inserts (161) into (160), one gets
2
2 2
Jλ (αn+1 , dn+1 ) = Rn Rl + λ Fn H(D) − 2
F dn+1 Rl + λ
such that dn+1 ∈ D has to be chosen as a maximizer of
2
.
F dn+1 2Rl + λ
This yields the following algorithm (where the Functional Matching Pursuit is a
particular case for λ = 0).
Algorithm 6 (RFMP). Let a data vector y ∈ Rl , a linear and continuous operator
F : H(D) → Rl , a dictionary D ⊂ {d ∈ H(D) | d H(D) = 1} and an initial
approximation F0 ∈ H(D) be given.
(i) Initialize the iteration with n = 0 and R0 = y − FF0 and select a stop-
ping criterion (data-misfit-based, i.e., choose ε > 0 to require Rn+1 < ε,
or iteration-based, i.e., choose N ∈ N to require n + 1 ≤ N ) as well as a
regularization parameter λ ∈ R+ 0.
(ii) Determine
2
Rn , F dRl − λ Fn , dH(D)
dn+1 = arg max , (162)
d∈D F d 2Rl + λ
αn+1 = 2 (163)
F dn+1 Rl + λ
and set Fn+1 = Fn + αn+1 dn+1 and Rn+1 = Rn − αn+1 F dn+1 .
(iii) If the stopping criterion is satisfied, then use Fn+1 as an approximate solution
to F F = y. Otherwise, increase n by 1 and go to step (ii).
The algorithm is accelerated if one implements the following procedures.
• Normalize the dictionary: use the assumption above and choose all d ∈ D
such that d H(D) = 1, otherwise the norm of the dictionary elements occurs
in (162) and (163) (see, e.g., [24]).
• Move as much as possible to the preprocessing: calculate F d Rl for all d ∈ D

and the (symmetric) matrices with the components d, d ˜ H(D) and F d,F d
˜ Rl ,
˜
respectively, (with d, d ∈ D) once and store them.
• Use preprocessing for finding dn+1 and αn+1 : note, in particular, that
? @
Rn , F dRl = Rn−1 , F d Rl − αn F dn , F dRl ,
Fn , dH(D) = Fn−1 , dH(D) + αn dn , dH(D) ,
where, in both cases, the first summands on the right-hand side are already
known from the previous iteration step (i.e., step n − 1).
It should be mentioned that, in (162), the maximizer need not be uniquely deter-
mined. In this case, no particular strategy for choosing between several maximizers
has been applied yet.
One essential result is the following convergence theorem.
Theorem 7 (Convergence Theorem). Let the dictionary D satisfy:
(i) ‘semi-frame condition’: Thereexist a constant c > 0 and an integer N such
∞
that, for all expansions H = k=1 βk dk with βk ∈ R and dk ∈ D, where the
dk are not necessarily pairwise distinct but {j ∈ N | dj = dk } is a finite set
with at most N elements for each k ∈ N,
∞
c H 2H(D) ≤ βk2 .
k=1
(ii) d H(D) = 1 for all d ∈ D and, if λ = 0, then inf d∈D F d Rl > 0 is required
additionally.
If the sequence (Fn )n is produced by the RFMP and no dictionary element
∞ is chosen
more than N times, then (Fn )n converges in H(D) to F∞ = F0 + n=1 αn dn ∈
H(D). Moreover, the following holds true:
·H(D)
(a) If span D = H(D) and λ ∈ R+
0 is an arbitrary parameter, then F∞
solves
(F ∗ F + λI)F∞ = F ∗ y ,
where F ∗ is the adjoint operator corresponding to F and I is the identity
operator on H(D). In other words,

2 2 2 2
y − F F∞ Rl + λ F∞ H(D) = min y − F F Rl + λ F H(D) ,
F ∈H(D)
where the minimizer is unique, if λ > 0.

(b) If span {F d | d ∈ D} = Rl and λ = 0, then F∞ solves F F∞ = y.
Note that the semi-frame condition has been changed (including the require-
ment on repeated choices of dictionary elements) in comparison to earlier publica-
tions on the RFMP by Michel [69] and Michel and Telschow [72], since an unlimited
number of equally chosen dictionary elements would allow a counterexample for
which the semi-frame condition could not be achieved, as it was pointed out in
[73]. For a proof of the convergence theorem and additional properties, see [71].
For numerical examples of RFMP applied to geodetic problems, see [23–26, 72].
Regularized orthogonal functional matching pursuit. Numerical experiments show

that the RFMP chooses some dictionary elements several times, which actually
means that some of the previously calculated coefficients α1 , . . . , αn are corrected.
The reason for this phenomenon is that the dictionary elements (or their images in
the data space) are typically non-orthogonal. In the case of the Matching Pursuit
(MP), this effect is compensated for by introducing a particular orthogonal projec-
tion procedure in the Orthogonal Matching Pursuit (OMP, see [78]) and by using
‘prefitting’ (see [98]). However, the OMP requires that the data and the solution
are in the same space for performing the projection and it also does not contain a
regularization.
In [95] and [73], the idea behind OMP and ‘prefitting’ was used to enhance
RFMP to Regularized Orthogonal Functional Matching Pursuit (ROFMP). It is
now possible to update the coefficients αi in every iteration. For this reason, the
approximation after step n is represented by

n
(n)
Fn = αi di .
i=1
If one measures the quality of an approximate solution in the data space, i.e., in the
sense of the data misfit, then the best approximation (without a regularization)
in terms of (fixed) d1 , . . . , dn would be given by requiring that F Fn equals the
orthogonal projection of y onto
Vn = span {F d1 , . . . , Fdn } ,
i.e., F Fn = PVn y. This is equivalent to requiring that the residual Rn = y − F Fn
is orthogonal to Vn . Geometrically speaking, FFn is the projection of y onto the
hyperplane Vn and Rn is the associated plumbline, see Figure 16.
Vn
Rn
P Vn y
Figure 16. Illustration of the orthogonal projection PVn y in Rl .

Consequently, the next summand αn+1 dn+1 should complement the previous ap-
proximation Fn such that F Fn+1 = PVn+1 y. However, in general, PVn (F dn+1 ) = 0.
This projection would, however, deteriorate the previously exact approximation of
PVn y by F Fn . For this reason, this redundant part is subtracted, i.e., one is inter-
ested in
F Fn+1 = F Fn + αn+1 [F dn+1 − PVn (F dn+1 )] .
n (n)
If one sets PVn (F d) = i=1 βi (d) F di , then

n
(n)

n
(n)
F Fn+1 = αi F di − αn+1 βi (dn+1 ) F di + αn+1 F dn+1
i=1 i=1
n
(n) (n)
= αi − αn+1 βi (dn+1 ) F di + αn+1 F dn+1 .
i=1
Hence, the task is now (in step n + 1) to find α ∈ R and d ∈ D such that
- -
- n -
- (n) (n) -
-y − αi − αβi (d) F di − αF d-
- - l
i=1 R
is minimized. As an approximation at step n + 1, one uses then
n
(n) (n)
Fn+1 = αi − αn+1 βi (dn+1 ) di + αn+1 dn+1 ,
i=1
(n+1) (n) (n)
αi = αi − αn+1 βi (dn+1 ) for i = 1, . . . , n
(n+1)
and αn+1 = αn+1 . With the regularization, the functional to minimize is
- -2
- n -
- (n) (n) -
-y − αi − αβi (d) F di − αF d-
- - l
i=1 R
- n -2
- -
- (n) (n) -
+ λ- αi − αβi (d) di + αd- .
- -
i=1 H(D)
This is the principle of the ROFMP. We now introduce some abbreviations.

• The orthogonal complement of Vn in Rl is denoted by Wn , i.e., Vn ⊕Wn = Rl ,
and the projection of F d onto Wn is

n
(n)
PWn (F d) = F d − βi (d) F di .
i=1
• The function associated to PVn (F d) in H(D) is denoted by

n
(n)
Bn (d) = βi (d) di .
i=1
Similar derivations as in the cases above finally yield the following algorithm.
Algorithm 8 (ROFMP). Let a data vector y ∈ Rl , a linear and continuous operator

F : H(D) → Rl and a dictionary D ⊂ H(D) \ {0} be given.
(i) Initialize the iteration with n = 0, F0 = 0 and R0 = y and select a stop-
ping criterion (data-misfit-based, i.e., choose ε > 0 to require Rn+1 < ε,
or iteration-based, i.e., choose N ∈ N to require n + 1 ≤ N ) as well as a
regularization parameter λ ∈ R+ 0.
(ii) Determine
2
Rn , PWn (F d)Rl + λ Fn , Bn (d) − dH(D)
dn+1 = arg max 2 2 ,
d∈D PWn (F d) Rl + λ Bn (d) − d H(D)
Rn , PWn (F dn+1 )Rl + λ Fn , Bn (dn+1 ) − dn+1 H(D)
αn+1 = 2 2 .
PWn (F dn+1 ) Rl + λ Bn (dn+1 ) − dn+1 H(D)
(iii) Update the coefficients as follows:
(n+1) (n) (n)
αi = αi − αn+1 βi (dn+1 ) for i = 1, . . . , n,
(n+1)
αn+1 = αn+1
n+1 (n+1)
and set Fn+1 = i=1 αi di as well as Rn+1 = y − F Fn+1 .
(iv) If the stopping criterion is satisfied, then use Fn+1 as an approximate solution
to F F = y. Otherwise, increase n by 1 and go to step (ii).
Obviously, a normalization of the dictionary elements to d H(D) = 1 does
not yield an improvement for the implementation in the case of the ROFMP.
Note that the orthogonal projection becomes more and more expensive with
an increasing number n. For this reason, it is advisable to restart the algorithm
after a certain number of steps N by using y − F FN as the new data vector
to be approximated and recounting from n = 0. Due to the linearity of F , the
consecutively produced approximations can be summed up in the end to obtain
an approximation of the solution F of F F = y. It turned out to be useful to keep,
after each restart, the previous approximation FN = F̃ in the regularization term
and to regularize with F̃ + Fn 2H(D) , where (Fn ) is the approximating sequence
after the restart.
For details of the implementation, see [95]. For numerical experiments and
theoretical results, see [73, 95]. Note that, in the non-regularized case (λ = 0),
the algorithm is able to produce an exact solution of F F = y in at most l steps,
where y ∈ Rl .
5. Conclusion
For the last decades, the possibilities and challenges which have presented them-
selves to geodesists have changed dramatically. Due to tremendously increased pre-
cisions in measurement technologies and the availability of satellite missions, huge
amounts of highly accurate data related to the Earth have become available. This
has opened previously unexpected options for observing, analyzing and predicting
the processes of the Earth system. Such progresses can be seen in manifold ways,
for example when the ocean dynamics can be understood better, when the mass
transports due to climate change or seasonal climatic phenomena can be better
quantified and localized, when static and dynamic models of the Earth’s interior
can be validated and improved by a more precise model or when unprecedented
ways of determining heights become available to geodesists.
Since mathematics plays a central role in the processes of, e.g., denoising,
analyzing or inverting geoscientific data, the changes in the data situation can
be mapped to changes in the requirements on the methodologies in mathematical
geodesy (see also [30]). In this paper, we focussed on the uncertainty principle
of spherical signal analysis which tells us that precise localization in space and
in frequency/momentum are mutually exclusive. Moreover, we can interpret the
uncertainty principle as a fundamental property of a spectrum ranging from ideal
frequency localization (i.e., no space localization) to ideal space localization (i.e.,
no frequency localization). The former is associated to the use of spherical har-
monics, which have been a common choice as basis system in geodesy. Away from
this extremal case, in order that trial functions possess a space localization, they
need to be sums of several spherical harmonics. The closer we come to the latter
end of the spectrum with ideal space localization, the more spherical harmonics
degrees have to be summed up in a trial function leading, as a limit, finally to the
(only as a theoretical concept existing) Dirac functional which includes all degrees.
The aforementioned new challenges due to today’s data situation can be re-
flected in this spectrum. In former days, when only a few data were available which
allowed a very coarse global modeling only, spherical harmonics were the ultimate
and reasonable choice. Today, the demands on highly accurate models which are,
in particular, provided with a very high resolution in space define the limits of
the use of spherical harmonics. These models can be better constructed with trial
functions which combine certain extents of space and frequency localization.
As we have shown, there are many facets of localized trial functions which can
be positioned in the spectrum of space and momentum localization. They include
basis functions generated from (reproducing) kernels of particular function spaces.
Such tools have successfully been used for spline and wavelet approximations in
the geosciences. They leave sufficient degrees of freedom to control their variance in
space and momentum. Furthermore, also Slepian functions provide another equally
valuable tool for regionally approximating or analyzing a signal. They provide us
with an orthonormal basis which is, in contrast to spherical harmonics, spacecon-
centrated (to a region which can be arbitrarily chosen). Moreover, the Slepian
functions are also orthogonal in the L2 -space of the chosen region, which is essen-
tially useful for the modeling of a signal which is only regionally available. Further-
more, Slepian functions can also be not only spaceconcentrated but even space-
limited with the price (due to the uncertainty principle) that they become non-
bandlimited, i.e., they sum up an infinite number of spherical harmonics degrees.
Certainly, there exist many other systems of trial functions on the sphere but
also the ball, which have their own characteristics regarding space- and frequency
localization. We added some references to other methods in appropriate paragraphs
but do not claim to have provided a complete overview. In general, a wide range of
special functions systems is available for the analysis of geoscientific data. However,
it appears that, still, the main focus of (too) many research projects in geodesy
and other disciplines of Earth sciences lies on the data alone but not on the choice
of the methodology for their handling.
In this paper, we have tried to break new synoptical ground in dealing with
spherical harmonics based special function systems and their role in constructive
approximation methods of mathematical geodesy. We have presented a short in-
sight and guide for the zoo of spherical trial functions to encourage geoscientists
to question the mathematical basis functions which they use for their models and
not to use mathematical tools as ‘blackboxes’. We have also summarized briefly
the possibility that regularized functional matching pursuit and its variants yield
as algorithms for generating a kind of a best basis out of a selection of different
basis systems.
Further research on finding the ‘optimal’ basis system for particular prob-
lems in mathematical geodesy has to be done. However, the present state-of-the-
art shows that there is a high potential in improving (not only) geodetic models
by using sophisticated mathematical methodologies. Obviously, our work as pre-
sented here is selective, but not only with respect to the choice of discussed basis
functions. Also, not all details on the treated topics could be discussed up to an
appropriate extent. For example, most of the proofs have been left out completely,
so that the interested reader is referred to the attached list of literature. Nonethe-
less, we believe that we have provided a deeper insight on how geoscientific and,
particularly, geodetic problems can be attacked in a mathematically systematic
and rigorous way.
References
[1] Amirbekyan, A. (2007) The Application of Reproducing Kernel Based Spline Ap-
proximation to Seismic Surface and Body Wave Tomography: Theoretical Aspects
and Numerical Results. Ph.D.-thesis, Geomathematics Group, Dept. Mathematics,
[2] Amirbekyan, A., Michel, V. (2008) Splines on the Three-dimensional Ball and Their
Application to Seismic Body Wave Tomography, Inverse Probl., 24, 015022 (25pp)
[3] Amirbekyan, A., Michel, V., Simons, F.J. (2008) Parameterizing Surface-wave To-
mographic Models with Harmonic Spherical Splines, Geophys. J. Int., 174, pp.
617–628
[4] Antoine, J.P., Demanet, L., Jaques, L., Vandergheynst, P. (2002) Wavelets on the
Sphere: Implementations and Approximations, Appl. Comput. Harmon. Anal., 13,
pp. 177–200
[5] Antoine, J.P., Vandergheynst, P. (1999) Wavelets on the 2-Sphere: A Group-

theoretic Approach, Appl. Comput. Harmon. Anal., 7, pp. 1–30
[6] Backus, G.E. (1966) Potentials for Tangent Tensor Fields on Spheroids, Arch. Ra-
tion. Mech. Anal., 22, pp. 210–252
[7] Backus, G.E. (1967) Converting Vector and Tensor Equations to Scalar Equations
in Spherical Coordinates, Geophys. J.R. Astron. Soc., 13, pp. 61–101
[8] Backus, G.E. (1986) Poloidal and Toroidal Fields in Geomagnetic Field Modelling,
Reviews of Geophysics, 24, pp. 75–109
[9] Ballani, L., Engels, J., Grafarend, E.W. (1993) Global Base Functions for the Mass
Density in the Interior of a Massive Body (Earth), Manuscr. Geod., 18, pp. 99–114
[10] Berkel, P. (2009) Multiscale Methods for the Combined Inversion of Normal Mode
and Gravity Variations. Ph.D.-thesis, Geomathematics Group, Dept. Mathematics,
University of Kaiserslautern, Shaker, Aachen
[11] Berkel, P., Michel, V. (2010) On Mathematical Aspects of a Combined Inversion of
Gravity and Normal Mode Variations by a Spline Method, Math. Geosci., 42, pp.
795–816
[12] Berkel, P., Fischer, D., Michel, V. (2011) Spline Multiresolution and Numerical
Results for Joint Gravitation and Normal Mode Inversion with an Outlook on
Sparse Regularisation, Int. J. Geomath., 1, pp. 167–204
[13] Blick, C., Freeden, W. (2011) Spherical Spline Application to Radio Occultation
Data, J. Geodetic Science, 1, 379–396
[14] Bluhm, R., Kostelecký, V.A., Tudose, B. (1995) Elliptical Squeezed States and
Rydberg Wave Packets, Phys. Rev. A, 52, pp. 2234–2244
[15] Carruthers, P., Nieto, M.M. (1968) Phase and Angle Variables in Quantum Me-
chanics, Rev. Mod. Phys., 40, pp. 411–440
[16] Clebsch, R.F.A. (1862) Ueber eine Eigenschaft der Kugelfunctionen. J. Reine
Angew. Math., 1862(60), pp. 343–350
[17] Dahlen, F.A., Simons, F.J. (2008) Spectral Estimation on a Sphere in Geophysics
and Cosmology, Geoph. J. Int., 174, pp. 774–807
[18] Dahlke, S., Dahmen, W., Schmitt, W., Weinreich, I. (1995) Multiresolution Analysis
and Wavelets on S 2 and S 3 , Numer. Funct. Anal. Optimiz., 16, pp. 19–41
[19] Dahlke, S., Maass, P. (1996) Continuous Wavelet Transforms with Applications to
Analyzing Functions on Spheres, J. Fourier Anal. Appl., 2, pp. 379–396
[20] Daubechies, I. (1992) Ten Lectures on Wavelets, CBMS-NSF Regional Conference
Series in Applied Mathematics, 61, SIAM, Philadelphia
[21] Eshagh, M. (2009) Spatially Restricted Integrals in Gradiometric Boundary Value
Problems, Artif. Sat., 44, pp. 131–148
[22] Fengler, M., Michel, D., Michel, V. (2006) Harmonic Spline-wavelets on the 3-
dimensional Ball and Their Application to the Reconstruction of the Earth’s Den-
sity Distribution From Gravitational Data at Arbitrarily Shaped Satellite Orbits,
ZAMM-Z. Angew. Math. Me., 86, pp. 856–873
[23] Fischer, D. (2011) Sparse Regularization of a Joint Inversion of Gravitational Data
and Normal Mode Anomalies. Ph.D.-thesis, Geomathematics Group, Dept. Math-
ematics, University of Siegen, Verlag Dr. Hut, Munich
[24] Fischer, D., Michel, V. (2012) Sparse Regularization of Inverse Gravimetry – Case
Study: Spatial and Temporal Mass Variations in South America. Inverse Probl., 28,
065012 (34pp.)
[25] Fischer, D., Michel, V. (2013a) Automatic Best-basis Selection for Geophysical
Tomographic Inverse Problems. Geophys. J. Int., 193, pp. 1291–1299
[26] Fischer, D., Michel, V. (2013b) Inverting GRACE Gravity Data for Local Climate
Effects, J. Geodetic Science, 3, pp. 151–162
[27] Fokas, A.S., Hauk, O., Michel, V. (2012) Electro-Magneto-Encephalography for the
Three-shell Model: Numerical Implementation via Splines for Distributed Current
in Spherical Geometry, Inverse Probl., 28, 035009 (28pp.)
[28] Freeden, W. (1981) On Spherical Spline Interpolation and Approximation, Math.
Meth. in the Appl. Sci., 3, pp. 551–575
Leipzig
[30] Freeden, W. (2015) Geomathematics: Its Role, Its Aim, and Its Potential, in: Hand-
book of Geomathematics (W. Freeden, M.Z. Nashed, and T. Sonar, Eds.), 2nd ed.,
Springer, Berlin, Heidelberg, pp. 3–79
[31] Freeden, W., Gerhards, C. (2013) Geomathematically Oriented Potential Theory.
Chapman and Hall, CRC Press, Boca Raton, London, New York
[32] Freeden, W., Gutting, M. (2013) Special Functions of Mathematical (Geo)Physics.
Birkhäuser, Basel
[33] Freeden, W., Maier, T. (2002) On Multiscale Denoising of Spherical Functions:
Basic Theory and Numerical Aspects, Electr. Transact. on Numer. Anal. (ETNA),
14, pp. 40–62
[34] Freeden, W., Maier, T. (2003) Spectral and Multiscale Signal-to-Noise Thresholding
of Spherical Vector Fields, Comput. Geosciences, 7, pp. 215–250
[35] Freeden, W., Michel, V. (2000) Least-Squares Geopotential Approximation by Win-
dowed Fourier Transform and Wavelet Transform, in: Wavelets in the Geosciences
(R. Klees and R. Haagmans, Eds.), Lecture Notes in Earth Sciences, Springer, 90,
pp. 189–241
[36] Freeden, W., Michel, V. (2004) Multiscale Potential Theory (with Applications to
Geoscience), Birkhäuser Verlag, Boston, Basel, Berlin
[37] Freeden, W., Nutz, H. (2011) Satellite Gravity Gradiometry as Tensorial Inverse
Problem, Int. J. Geomath., 2, pp. 177–218
[38] Freeden, W., Schneider, F. (1998) Regularization Wavelets and Multiresolution,
Inverse Probl., 14, pp. 225–243
[39] Freeden, W., Schreiner, M. (1995) Non-Orthogonal Expansions on the Sphere,
Math. Meth. Appl. Sci., 18, pp. 83–120
[40] Freeden, W., Schreiner, M. (1998) Orthogonal and Non-Orthogonal Multiresolution
Analysis, Scale Discrete and Exact Fully Discrete Wavelet Transform on the Sphere.
Constr. Approx., 14, pp. 493–515
[41] Freeden, W., Schreiner, M. (2006), Local Multiscale Modelling of Geoid Undulations
from Deflections of the Vertical, J. Geodesy, 79, 641–651
[42] Freeden, W., Schreiner, M. (2007) Biorthogonal Locally Supported Wavelets on the
Sphere Based on Zonal Kernel Functions, J. Fourier Anal. Appl., 13, pp. 693–709
[43] Freeden, W., Schreiner, M. (2009) Spherical Functions of Mathematical Geosciences
– A Scalar, Vectorial, and Tensorial Setup, Springer, Berlin
[44] Freeden, W., Windheuser, U. (1996) Spherical Wavelet Transform and its Dis-
cretization, Adv. Comput. Math., 5, pp. 51–94
[45] Freeden, W., Windheuser, U. (1997) Combined Spherical Harmonic and Wavelet
Expansion. Appl. Comput. Harmon. Anal., 4, pp. 1–37
[46] Freeden, W., Witte, B. (1982) A Combined (Spline-) Interpolation and Smooth-
ing Method for the Determination of the External Gravitational Potential from
Heterogeneous Data, Bull. Geod., 56, pp. 53–62
[47] Freeden, W., Gervens, T., Schreiner, M. (1998) Constructive Approximation on
the Sphere (with Applications to Geomathematics), Oxford Science Publications,
Clarendon, Oxford
[48] Freeden, W., Michel, V., Nutz, H. (2002) Satellite-to-Satellite Tracking and Satel-
lite Gravity Gradiometry (Advanced Techniques for High-Resolution Geopotential
Field Determination), J. Eng. Math., 43, pp. 19–56
[49] Freeden, W., Fehlinger, T., Klug, M., Mathar, D., and Wolf, K. (2009) Classical
Globally Reflected Gravity Field Determination in Modern Locally Oriented Mul-
tiscale Framework, J. Geodesy, 83, pp. 1171–1191
[50] Gauß, C.F. (1838) Allgemeine Theorie des Erdmagnetismus, Resultate aus den
Beobachtungen des magnetischen Vereins, Göttingen
[51] Grünbaum, F.A., Longhi, L., Perlstadt, M. (1982) Differential Operators Com-
muting with Finite Convolution Integral Operators: Some Non-Abelian Examples,
SIAM J. Appl. Math., 42, pp. 941–955
[52] Guilloux, F., Faÿ, G., Cardoso, J.F. (2009) Practical Wavelet Design on the Sphere,
Appl. Comput. Harmon. Anal., 26, pp. 143–160
[53] Haar, A. (1910) Zur Theorie der orthogonalen Funktionensysteme, Math. Ann., 69,
pp. 331–371
[54] Heine, E. (1878) Handbuch der Kugelfunktionen, Reimer, Berlin
[55] Holschneider M (1996) Continuous Wavelet Transforms on the Sphere, J. Math.
Phys., 37, pp. 4156–4165
[56] Kennedy, R.A. and Sadeghi, P. (2013) Hilbert Space Methods in Signal Processing,
Cambridge University Press, Cambridge, UK
[57] Khalid, Z., Durrani, S., Sadeghi, P., Kennedy, R.A. (2012) Spatio-spectral Analysis
on the Sphere Using Spatially Localized Spherical Harmonics Transform, IEEE
Trans. Signal Process., 60, pp. 1487–1492
[58] Kowalski, K., Rembieliński, J. (2000) Quantum Mechanics on a Sphere and Coher-
ent States, J. Phys. A: Math. Gen., 33, pp. 6035–6048
[59] Laı́n Fernández, N. (2003) Polynomial Bases on the Sphere, Ph.D.-thesis, University
of Lübeck
[60] Landau, H.J. and Pollak, H.O. (1961) Prolate spheroidal wave functions, Fourier
analysis and uncertainty – II, Bell Syst. Tech. J., 40, pp. 65–84
[61] Laplace, P.S. de (1785) Théorie des Attractions des Sphéroides et de la Figure des
Planètes, Mém. de l’Acad., Paris
[62] Legendre, A.M. (1785) Recherches sur l’Attraction des Sphéroides Homogènes,
Mém. math. phys. prés. à l’Acad. Aci. par divers savants, 10, pp. 411–434
[63] Leistedt, B., McEwen, J.D. (2012) Exact Wavelets on the Ball, IEEE Trans. Signal
Process., 60, pp. 6257–6269
[64] Lyche, T., Schumaker, L. (2000) A Multiresolution Tensor Spline Method for Fitting
Functions on the Sphere, SIAM J. Sci. Comput., 22, pp. 724–746
[65] Mallat, S.G., Zhang, Z. (1993) Matching Pursuits with Time-Frequency Dictionar-
ies, IEEE Trans. Signal Process., 41, pp. 3397–3415
[66] Maxwell, J.C. (1891) A Treatise on Electricity and Magnetism (1873, 1881, 1891)
Bd. 1 und 2 Ungekürzter Nachdruck der letzten Auflage (1954), Dover, New York
[67] Michel, V. (2013) Lectures on Constructive Approximation – Fourier, Spline, and
Wavelet Methods on the Real Line, the Sphere, and the Ball, Birkhäuser, Boston
[68] Michel, V. (2015a) Tomography – Problems and Multiscale Solutions, in: Hand-
book of Geomathematics (W. Freeden, M.Z. Nashed, and T. Sonar, Eds.), 2nd ed.,
Springer, Berlin, Heidelberg, pp. 2087–2119
[69] Michel, V. (2015b) RFMP – An Iterative Best Basis Algorithm for Inverse Problems
in the Geosciences, in: Handbook of Geomathematics (W. Freeden, M.Z. Nashed,
and T. Sonar, Eds.), 2nd ed., Springer, Berlin, Heidelberg, pp. 2121–2147
[70] Michel, V., Orzlowski, S. (2016) On the Null Space of a Class of Fredholm Integral
Equations of the First Kind, J. Inverse and Ill-Posed Problems, 24, pp. 687–710
[71] Michel, V., Orzlowski, S. (2017) On the Convergence Theorem for the Regularized
Functional Matching Pursuit (RFMP) Algorithm, Int. J. Geomath., 8, pp. 183–190
[72] Michel, V., Telschow, R. (2014) A Non-linear Approximation Method on the Sphere,
Int. J. Geomath., 5, pp. 195–224
[73] Michel, V., Telschow, R. (2016) The Regularized Orthogonal Functional Matching
Pursuit for Ill-posed Inverse Problems, SIAM J. Numer. Anal., 54, pp. 262–287
[74] Michel, V., Wolf, K. (2008) Numerical Aspects of a Spline-based Multiresolution
Recovery of the Harmonic Mass Density out of Gravity Functionals, Geophys. J.
Int., 173, pp. 1–16
[75] Morse, P.M., Feshbach, H. (1953) Methods of Theoretical Physics, McGraw-Hill,
New York
[76] Narcowich, F.J., Ward, J.D. (1996) Nonstationary Wavelets on the m-Sphere for
Scattered Data, Appl. Comput. Harmon. Anal., 3, pp. 324–336
[77] Neumann, F. (1887) Vorlesungen über die Theorie des Potentials und der Kugel-
funktionen, Teubner, Leipzig, pp. 135–154
[78] Pati, Y.C., Rezaiifar, R., Krishnaprasad, P.S. (1993) Orthogonal Matching Pursuit:
Recursive Function Approximation with Applications to Wavelet Decomposition,
in: Asilomar Conference on Signals, Systems and Computers, IEEE Conf. Pub., 1,
pp. 40–44
[79] Plattner, A., Simons, F.J. (2014) Spatiospectral Concentration of Vector Fields on
a Sphere, Appl. Comput. Harmon. Anal., 36, pp. 1–22
[80] Plattner, A., Simons, F.J. (2015a) Potential-Field Estimation Using Scalar and Vec-
tor Slepian Functions at Satellite Altitude, in: Handbook of Geomathematics (W.
Freeden, M.Z. Nashed, and T. Sonar, Eds.), 2nd ed., Springer, Berlin, Heidelberg,
pp. 2003–2055
[81] Plattner, A., Simons, F.J. (2015b) High-Resolution Local Magnetic Field Models
for the Martian South Pole From Mars Global Surveyor Data, J. Geophys. Res.,
pp. 1543–1566
[82] Potts, D., Tasche, M. (1995) Interpolatory Wavelets on the Sphere, in: Approxima-
tion Theory VIII (C.K. Chui, L.L. Schumaker, Eds.), World Scientific, Singapore,
pp. 335–342
[83] Schröder, P., Sweldens, W. (1995) Spherical Wavelets: Efficiently Representing
Functions on the Sphere, in: Computer Graphics Proceedings (SIGGRAPH95), pp.
161–175
[84] Simons, F.J. (2010) Slepian Functions and Their Use in Signal Estimation and
Spectral Analysis, in: Handbook of Geomathematics (W. Freeden, M.Z. Nashed,
and T. Sonar, Eds.), 1st ed., Springer, Heidelberg, pp. 891–923
[85] Simons, F.J., Dahlen, F.A. (2006) Spherical Slepian Functions and the Polar Gap
in Geodesy, Geoph. J. Int., 166, pp. 1039–1061
[86] Simons, F.J., Plattner, A. (2015) Scalar and Vector Slepian Functions, Spherical
Signal Estimation and Spectral Analysis, in: Handbook of Geomathematics (W.
Freeden, M.Z. Nashed, and T. Sonar, Eds.), 2nd ed., Springer, Berlin, Heidelberg,
pp. 2563–2608
[87] Simons, F.J., Dahlen, F.A., Wieczorek, M. (2006) Spatiospectral Concentration on
a Sphere, SIAM Rev., 48, pp. 504–536
[88] Simons, F.J., Hawthorne, J.C., Beggan, C.D. (2009) Efficient Analysis and Repre-
sentation of Geophysical Processes Using Localized Spherical Basis Functions, in:
Wavelets XIII, Proc. SPIE, (V. Goyal, M. Papadakis and D. Van de Ville, Eds.),
7446(74460G)
[89] Simons, F.J., Loris, I., Brevdo, E., Daubechies, I.C. (2011) Wavelets and Wavelet-
like Transforms on the Sphere and Their Application to Geophysical Data Inversion,
in: Wavelets and Sparsity XIV, Proc. SPIE, (M. Papadakis, V. Goyal, and D. Van
de Ville, Eds.), 8138(81380X)
[90] Slepian, D. (1983) Some Comments on Fourier Analysis, Uncertainty and Modeling,
SIAM Rev., 25, pp. 379–393
[91] Slobbe, D.C., Simons, F.J., Klees, R. (2012) The Spherical Slepian Basis as a Means
to Obtain Spectral Consistency Between Mean Sea Level and the Geoid, J. Geod.,
86, pp. 609–628
[92] Svensson, S.L. (1983) Pseudodifferential Operators – A New Approach to the
Boundary Value Problems of Physical Geodesy, Manuscr. Geod., 8, pp. 1–40
[93] Sylvester, T. (1876) Note on Spherical Harmonics, Phil. Mag., II, 291 and 400
[94] Tegmark, M. (1996) A Method for Extracting Maximum Resolution Power Spectra
from Microwave Sky Maps, Mon. Not. Roy. Astron. Soc., 280, pp. 299–308
[95] Telschow, R. (2015) An Orthogonal Matching Pursuit for the Regularization of
Spherical Inverse Problems, Ph.D.-thesis, Geomathematics Group, Dept. Mathe-
matics, University of Siegen, Verlag Dr. Hut, Munich
[96] Temlyakov, V.N. (2003) Nonlinear Methods of Approximation, Found. Comput.

Math. 3, pp. 33–107
[97] Tscherning, C.C. (1996) Isotropic Reproducing Kernels for the Inner of a Sphere
or Spherical Shell and Their Use as Density Covariance Functions. Math. Geol. 28,
pp. 161–168
[98] Vincent, P., Bengio, Y. (2002) Kernel Matching Pursuit, Mach. Learn., 48, pp.
169–191
[99] Wahba, G. (1981) Spline Interpolation and Smoothing on the Sphere, SIAM J.
Sci. Stat. Comput., 2, pp. 5–16, (also errata: SIAM J. Sci. Stat. Comput., 3, pp.
385–386)
[100] Wahba, G. (1990) Spline Models for Observational Data, CBMS-NSF Regional
Conference Series in Applied Mathematics, 59, SIAM, Philadelphia
[101] Weinreich, I. (2001) A Construction of C(1) -Wavelets on the Two-dimensional
Sphere, Appl. Comput. Harmon. Anal., 10, pp. 1–26
[102] Wicht, J., Stellmach, S., Harder, H. (2010) Numerical Dynamo Simulations: From
Basic Concepts to Realistic Models, in: Handbook of Geomathematics (W. Freeden,
M.Z. Nashed, and T. Sonar, Eds.), 1st ed., pp. 459–502, Springer, Heidelberg
[103] Wieczorek, M.A. (2015) The Gravity and Topography of the Terrestrial Planets,
in: Treatise on Geophysics (T. Spohn, ed.), 2nd ed., vol. 10, pp. 153–2193, Elsevier,
Amsterdam
[104] Wieczorek, M.A., Simons, F.J. (2005) Localized Spectral Analysis on the Sphere,
Geophys. J. Int., 162, pp. 655–675
Willi Freeden
Volker Michel
University of Siegen
Walter-Flex-Str. 3
D-57068 Siegen, Germany
Frederik J. Simons
Princeton University
Department of Geosciences
Princeton, NJ, USA

Spherical Potential Theory:

Tools and Applications
Christian Gerhards
Abstract. In the current chapter, we transfer classical potential theoretic con-

cepts from the Euclidean space R3 to a setting intrinsic on the sphere. We
present uniqueness results for the Poisson equation on the sphere, explicitly
construct Green functions for spherical caps and complete function systems
for harmonic approximation, and elaborate on decompositions of vector fields
on the sphere. Among the intended applications are problems from oceanog-
raphy, geodesy, and geomagnetism. Some examples are presented at the end.
Keywords. Green’s functions, Helmholtz decomposition, Hardy–Hodge de-

composition, fundamental solutions, complete function systems, boundary
value problems, surface potentials, curve potentials,limit- and jump-relations,
Green’s formulas, vertical deflections, disturbing potential, multiscale approx-
imation, geostrophic flow, mean dynamic topography, vortex motion.
Classical potential theoretic concepts in the Euclidean space R3 have been de-
scribed in an earlier chapter of this handbook. They appear frequently in geodesy
when treating the harmonic gravitational potential in the exterior of the spherical
Earth. The sphere ΩR = {x ∈ R3 : |x| = R} occurs as the boundary surface of a
subdomain in R3 . Opposed to this, in the present chapter, the sphere is not rep-
resenting a boundary surface, it is rather regarded as the underlying domain on
which a problem is formulated. Examples for this are the spherical Navier–Stokes
equations and shallow water equations in meteorology and ocean modeling (see,
e.g., [5, 12, 19, 29, 42, 44]). But also simpler spherical differential equations occur
in geodesy and geomagnetism (see, e.g., [2, 10, 11, 14, 16, 23, 25]) and vortex
dynamics (see, e.g., [32, 33, 42–44]), more precisely, those based on the Beltrami
operator Δ∗ (the spherical counterpart to the Laplace operator Δ). Latter is going
to be the focus of this chapter. In particular, we are interested in the Beltrami
equation on subdomains ΓR ⊂ ΩR of the sphere, which eventually leads to po-
tential theoretic concepts analogous to those of the Euclidean case. Subdomains
appear naturally, e.g., due to only regionally available data or coastal/continental
boundaries. The problems we take a closer look at are the following (note that
822 C. Gerhards
Ω and Γ are simply abbreviations for the unit sphere Ω1 and a corresponding
subdomain Γ1 ):
Poisson Problem (PP): Let H be of class C(1) (Γ). We are looking for a function
U of class C(2) (Γ) such that
Δ∗ U (ξ) = H(ξ), ξ ∈ Γ. (1)
Dirichlet Problem (DP): Let F be of class C(0) (∂Γ). We are looking for a function
U of class C(2) (Γ) ∩ C(0) (Γ) such that
Δ∗ U (ξ) = 0, ξ ∈ Γ, (2)
U − (ξ) = F (ξ), ξ ∈ ∂Γ. (3)
Neumann Problem (NP): Let F be of class C(0) (∂Γ). We are looking for a function
U of class C(2) (Γ) ∩ C(0) (Γ), with a well-defined normal derivative ∂ν
∂
U − on
∂Γ, such that
Δ∗ U (ξ) = 0, ξ ∈ Γ, (4)
∂ −
U (ξ) = F (ξ), ξ ∈ ∂Γ. (5)
∂ν
In the setting above, ∂Γ denotes the boundary curve of Γ, ν(ξ) the outward
directed unit normal vector at ξ ∈ ∂Γ, and ∂ν∂
the corresponding normal derivative.
−
The minus of U simply indicates that we are approaching the boundary ∂Γ from
within Γ.
Certainly, the problems above and its potential theoretic consequences can
be and have been treated on more general manifolds than the sphere (e.g., in [8,
37, 38]). However, we focus on the geophysically relevant case of the sphere where
explicit representations of the fundamental solution and some Green’s functions
are known. In large parts, we follow the course of [15] and emphasize similarities
and differences to the Euclidean case.
The first section supplies the reader with necessary notations and several
mathematical tools related to spherical potential theory. In Section 2, we treat the
problems (PP), (DP), and (NP). In particular, we are interested in integral rep-
resentations of their solutions. In Section 3, we turn towards spherical differential
operators of order one, namely the surface gradient ∇∗ (the spherical counterpart
to the gradient ∇) and the surface curl gradient L∗ . We investigate the corre-
sponding differential equations on Γ as well as the so-called spherical Helmholtz
decomposition and the spherical Hardy–Hodge decomposition. Section 4 comments
briefly on complete function systems and approximation methods on the sphere.
Finally, in Section 5, applications of the previous concepts to some geophysical
problems are discussed, namely, vertical deflections, (geostrophic) ocean flow, and
a toy problem for point vortex motion.
Spherical Potential Theory: Tools and Applications 823
1. Fundamental tools
Of fundamental importance to us is the Beltrami operator Δ∗ which denotes the
tangential contribution to the Euclidean Laplace operator Δ. More precisely,
1 ∂ ∂ 1
Δx = 2 r2 + Δ∗ , (6)
r ∂r ∂r r2 ξ
where Δx acts on x ∈ R3 while Δ∗ξ acts on ξ = |x| x
∈ Ω. The length |x| is usually
∗
denoted by r. Furthermore, ∇ stands for the (spherical) surface gradient, which
denotes the tangential contribution to the gradient ∇:
∂ 1
∇x = ξ + ∇∗ξ . (7)
∂r r
The occasionally occurring (spherical) surface curl gradient L∗ acts via ξ ∧ ∇∗ξ at
a point ξ ∈ Ω (“∧” denotes the vector product). It should be noted that Δ∗ =
∇∗ · ∇∗ = L∗ · L∗ (“ · ” denotes the Euclidean inner product). If it is clear on
which variables the operators act, we usually omit the subindices ξ and x. For
convenience, we typically use Greek letters ξ, η to indicate unit vectors in Ω while
we use latin letters x, y for general vectors in R3 . Upper case letters F, G denote
scalar-valued functions mapping Γ ⊂ Ω into R while lower case letters f, g denote
vector-valued functions mapping Γ ⊂ Ω into R3 . Correspondingly, the set of k-
times continuously differentiable scalar-valued functions on Γ is designated by
C(k) (Γ) and the set of k-times continuously differentiable vector-valued functions
on Γ by c(k) (Γ). The closure of Γ is denoted by Γ and the open complement by
Γc = Ω \ Γ.
Whenever we talk about subdomains Γ ⊂ Ω in this chapter, we mean, with-
out further mention, regular regions, i.e., subdomains with a sufficiently smooth
boundary curve ∂Γ (for details, the reader is referred to [15]; an exemplary illus-
tration is supplied in Figure 1). For such regular regions, the positively oriented
unit tangential vector τ (ξ) at a point ξ ∈ ∂Γ is well defined. The unit normal
vector ν(ξ) at ξ ∈ ∂Γ points into the exterior of Γ and is perpendicular to τ (ξ)
Gr (x)
G
x t(x)
¶G
n(x) Gc
Figure 1. Examples for a general regular regionΓ (left) and a spherical

cap Γρ (ξ) with center ξ and radius ρ (right).
824 C. Gerhards
and ξ (i.e., ν(ξ) is perpendicular to the boundary curve ∂Γ but tangential to the
unit sphere Ω). The normal derivative of a scalar-valued function F at ξ ∈ ∂Γ is
defined as
∂
F (ξ) = ν(ξ) · ∇∗ξ F (ξ). (8)
∂ν
1.1. Green’s formulas
We frequently need integral expressions that describe the shifting of differential
operators from one integrand to another, so-called Green formulas. Some spherical
versions are stated in the next theorem.
Theorem 1.1 (Spherical Green Formulas I).
(a) If f is of class c(1) (Γ) and tangential, i.e., ξ · f (ξ) = 0 for ξ ∈ Γ, then

∇∗ · f (η)dω(η) = ν(η) · f (η)dσ(η), (9)
Γ
∂Γ
L∗ · f (η)dω(η) = τ (η) · f (η)dσ(η). (10)

Γ ∂Γ
(b) If F is of class C (Γ) and f of class c(1) (Γ), then

(1)

f (η) · ∇∗ F (η)dω(η) + F (η)∇∗ · f (η)dω(η)
Γ Γ

= ν(η) · (F (η)f (η)) dσ(η) + 2 η · (F (η)f (η)) dω(η), (11)
∂Γ Γ

f (η) · L F (η)dω(η) + F (η)L∗ · f (η)dω(η)
∗
Γ Γ

= τ (η) · (F (η)f (η)) dσ(η). (12)
∂Γ
(c) If F, H are functions of class C(2) (Γ), then

F (η)Δ∗ H(η)dω(η) − H(η)Δ∗ F (η)dω(η)
Γ Γ

∂ ∂
= F (η) H(η)dσ(η) − H(η) F (η)dσ(η). (13)
∂Γ ∂ν ∂Γ ∂ν
Generally, ‘dω’ denotes the surface element in Γ ⊂ Ω and ‘dσ’ the line element
on ∂Γ.
Remark 1.2. The formulas (11)–(13) are direct consequences of (9) and (10). Drop-
ping the boundary terms ∂Γ . . . dσ, all of these formulas also hold true for the
choice Γ = Ω.
A crucial step for later considerations is the combination of Green’s formulas
with the fundamental solution for the Beltrami operator G(Δ∗ ; ·) : [−1, 1) → R,
which is uniquely determined by the following properties:
1. For any fixed ξ ∈ Ω, the function η → G(Δ∗ ; ξ · η) is twice continuously

differentiable on Ω \ {ξ} and
1
Δ∗η G(Δ∗ ; ξ · η) = − , η ∈ Ω \ {ξ}. (14)
4π
1
2. For any fixed ξ ∈ Ω, the function η → G(Δ∗ ; ξ · η) − 4π ln(1 − ξ · η) is
continuously differentiable on Ω.
1 ∗
3. For any fixed ξ ∈ Ω, it holds 4π Ω G(Δ ; ξ · η)dω(η) = 0.
Some basic calculations show that the function given by
1 1
G(Δ∗ ; t) = ln(1 − t) + (1 − ln(2)), t ∈ [−1, 1), (15)
4π 4π
satisfies the properties (i)–(iii). The property (i) denotes the major difference be-
tween the fundamental solution for the Laplace operator G(Δ; ·) and its spherical
counterpart. While G(Δ; ·) generates a ’true’ Dirac distribution in the sense that
Δy G(Δ; |x − y|) = 0, y ∈ R3 \ {x}, the fundamental solution G(Δ∗ ; ·) only gener-
ates a Dirac distribution up to an additive constant (reflecting the null space of
the Beltrami operator Δ∗ ). Eventually, applying Green’s formulas from Theorem
1.1, the properties of G(Δ∗ ; ·) lead to the following integral representations.
Theorem 1.3 (Spherical Green Formulas II).
(a) If F is of class C(2) (Γ), then we have for ξ ∈ Ω,

α(ξ) 1
F (ξ) = F (η)dω(η) + G(Δ∗ ; ξ · η)Δ∗η F (η)dω(η)
2π 4π Γ Γ

∂ ∗ ∂
+ F (η) G(Δ ; ξ · η)dσ(η) − G(Δ∗ ; ξ · η) F (η)dσ(η).
∂Γ ∂ν(η) ∂Γ ∂ν(η)
(16)
(b) If F is of class C(1) (Γ), then we have for ξ ∈ Ω,

α(ξ) 1
F (ξ) = F (η)dω(η) − ∇∗η G(Δ∗ ; ξ · η) · ∇∗η F (η)dω(η)
2π 4π Γ Γ

∂
+ F (η) G(Δ∗ ; ξ · η)dσ(η),
∂ν(η)

∂Γ

1
= F (η)dω(η) − L∗η G(Δ∗ ; ξ · η) · L∗η F (η)dω(η)
4π Γ Γ

∂
+ F (η) G(Δ∗ ; ξ · η)dσ(η). (17)
∂Γ ∂ν(η)
The solid angle α of a regular region Γ is defined such that α(ξ) = 2π for ξ ∈ Γ,
α(ξ) = π for ξ ∈ ∂Γ, and α(ξ) = 0 for ξ ∈ Γc .

Remark 1.4. Again, dropping the boundary terms ∂Γ . . . dσ in the expressions of
Theorem 1.3 leads to results that hold true for the global choice Γ = Ω.
826 C. Gerhards
Remark 1.5. Choosing F ≡ 1 in any of the formulas in Theorem 1.3 implies

⎧
⎪ Γ
⎨ 1 − 4π , ξ ∈ Γ,
∂
G(Δ∗ ; ξ · η)dσ(η) = 1 Γ
2 − 4π , ξ ∈ ∂Γ,
(18)
∂Γ ∂ν(η) ⎪
⎩ − Γ ,
4π ξ ∈ Γ c
,
where Γ denotes the surface area of Γ. The behaviour of (18) across the bound-
ary ∂Γ states a first hint at the limit and jump relations of the layer potentials
in Section 1.4. Apart from the additive constant Γ
4π , they are identical to the
Euclidean setting.
1.2. Harmonic functions
In this subsection, we turn towards functions that are harmonic (with respect to
the Beltrami operator) in Γ, i.e., functions U of class C(2) (Γ) that satisfy
Δ∗ U = 0 in Γ. (19)
If no confusion with the Euclidean case is likely to arise, we just say that U is
harmonic. Plugging such functions into (16), together with the choice of Γ being
a spherical cap Γρ (ξ) = {η ∈ Ω : 1 − ξ · η < ρ} with center ξ ∈ Ω and radius
ρ ∈ (0, 2), we end up with the mean value property for harmonic functions.
Theorem 1.6 (Mean Value Property I). A function U of class C(0) (Γ) is harmonic
if and only if
√
1 2−ρ
U (ξ) = U (η)dω(η) + √ U (η)dσ(η), ξ ∈ Γ, (20)
4π Γρ (ξ) 4π ρ ∂Γρ (ξ)
for any spherical cap Γρ (ξ) ⊂ Γ.

The mean value property above contains the typical additive term for spheri-
cal problems. However, we can get rid of this additive constant when using Green’s
functions for spherical caps as described later on in Section 2.3. We are led to the
following representation which resembles a Mean Value Property that is more
closely related to the Euclidean case of functions that are harmonic with respect
to the Laplace operator.
Theorem 1.7 (Mean Value Property II). A function U of class C(0) (Γ) is harmonic
if and only if

1
U (ξ) = # U (η)dσ(η), ξ ∈ Γ, (21)
2π ρ(2 − ρ) ∂Γρ (ξ)
for any spherical cap Γρ (ξ) ⊂ Γ.

Once a Mean Value Property is established, it can be used to derive a Max-
imum Principle. For details, we refer to [15] or, in the Euclidean case, any book
on classical potential theory such as [17, 26, 30, 45].
Theorem 1.8 (Maximum Principle). If U of class C(2) (Γ) ∩ C(0) (Γ) is harmonic,
then
sup |U (ξ)| ≤ sup |U (ξ)|. (22)
ξ∈Γ ξ∈∂Γ
1.3. Surface potentials

Analogous to the Euclidean setting, we can define a Newton potential and layer
potentials for the spherical setting which take over the corresponding roles. The
obvious difference is that now the Newton potential is a surface potential and
the layer potentials represent curve potentials. Throughout this section, we take a
closer look at the surface potential

U (ξ) = G(Δ∗ ; ξ · η)H(η)dω(η), ξ ∈ Ω, (23)
Γ
From the properties of the fundamental solution for the Beltrami operator it be-
comes directly clear that U is of class C(2) (Γc ) and that

1
Δ∗ξ G(Δ∗ ; ξ · η)H(η)dω(η) = − H(η)dω(η), ξ ∈ Γc . (24)
Γ 4π Γ
Yet, the interesting question is what happens if ξ ∈ Γ, i.e., when the integration
region contains the singularity of G(Δ∗ ; ·).
Theorem 1.9. If H is of class C(0) (Γ) and U is given by (23), then U is of class
C(1) (Ω) and

∇∗ξ G(Δ∗ ; ξ · η)H(η)dω(η) = ∇∗ξ G(Δ∗ ; ξ · η)H(η)dω(η), ξ ∈ Ω. (25)
Γ Γ
The proof of the theorem above can be based on a regularization of the
fundamental solution G(Δ∗ ; ·). This approach also works for the application of the
Beltrami operator to U . However, in connection with Theorem 1.3, we find that
Δ∗ U is not continuous across ∂Γ anymore (compare equation (24) and Theorem
1.10). For brevity, we do not supply the proofs at this point but refer the reader,
e.g., to [15]. A related regularized Green function plays an important role in the
applications in Section 5 and is explained in more detail later on.
Theorem 1.10. If H is of class C(1) (Γ) and U is given by (23), then U is of class
C(2) (Γ) and satisfies

∗ ∗ 1
Δξ G(Δ ; ξ · η)H(η)dω(η) = H(ξ) − H(η)dω(η), ξ ∈ Γ. (26)
Γ 4π Γ
1.4. Curve potentials
While surface potentials are useful to deal with the Poisson problem(PP), curve
potentials are particularly useful when dealing with functions that are harmonic
(with respect to the Beltrami operator). More precisely, we take a closer look at
the two layer potentials

U1 [Q̃](ξ) = G(Δ∗ ; ξ · η)Q̃(η) dσ(η), ξ ∈ Γ, (27)
∂Γ
828 C. Gerhards
and

∂
U2 [Q](ξ) = G(Δ∗ ; ξ · η) Q(η) dσ(η), ξ ∈ Γ. (28)
∂Γ ∂ν(η)
From the properties of the fundamental solution G(Δ∗ ; ·) it can be seen that the
so-called double-layer potential U2 [Q] is harmonic in Γ for any Q of class C(0) (∂Γ).
The single-layer potential U1 [Q̃] is harmonic
in Γ if Q̃ is of class C(0) (∂Γ) and if
the integral over ∂Γ vanishes, i.e., if ∂Γ Q̃(η)dσ(η) = 0 (we say that Q̃ is of class
(0)
C0 (∂Γ)). Therefore, these two potentials represent good candidates for solutions
to the boundary value problems (DP) and (NP). The aim of the present section is
to investigate the behaviour of the single- and double-layer potentials U1 [Q] and
U2 [Q̃], respectively, when they approach the boundary ∂Γ. The essential behaviour
of the double-layer potential U2 [Q] is already reflected by the relation (18). Based
on this relation and a set of several more technical estimates, one can prove the
following set of limit- and jump-relations at the boundary ∂Γ.
Theorem 1.11 (Limit- and Jump-Relations). Let Q, Q̃ be of class C(0) (∂Γ) and
U1 , U2 be given as in (27) and (28), respectively. Furthermore, let ξ ∈ ∂Γ.
(a) For the single-layer potential, we have the limit-relations

ξ ± τ ν(ξ)
lim U1 [Q̃] √ − U1 [Q̃](ξ) = 0, (29)
τ →0+ 1 + τ2

∂ ξ ± τ ν(ξ) ∂ 1
lim U1 [Q̃] √ − U1 [Q̃] (ξ) = ± Q̃(ξ). (30)
τ →0+ ∂ν 1+τ 2 ∂ν 2
For the double-layer potential, we have

ξ ± τ ν(ξ) 1
lim U2 [Q] √ − U2 [Q](ξ) = ∓ Q(ξ). (31)
τ →0+ 1+τ 2 2
(b) For the single-layer potential, we have the jump-relations

ξ + τ ν(ξ) ξ − τ ν(ξ)
lim U1 [Q̃] √ − U1 [Q̃] √ = 0, (32)
τ →0+ 1 + τ2 1 + τ2

∂ ξ + τ ν(ξ) ∂ ξ − τ ν(ξ)
lim U1 [Q̃] √ − U1 [Q̃] √ = Q̃(ξ). (33)
τ →0+ ∂ν 1 + τ2 ∂ν 1 + τ2
For the double-layer potential, we have

ξ + τ ν(ξ) ξ − τ ν(ξ)
lim U2 [Q] √ − U2 [Q] √ = −Q(ξ), (34)
τ →0+ 1 + τ2 1 + τ2

∂ ξ + τ ν(ξ) ∂ ξ − τ ν(ξ)
lim U2 [Q] √ − U2 [Q] √ = 0. (35)
τ →0+ ∂ν 1+τ 2 ∂ν 1 + τ2
All of the relations above hold uniformly with respect to ξ ∈ ∂Γ.
Remark 1.12. Theorem 1.11 essentially tells us that the single-layer potential U1 [Q̃]
∂
and the normal derivative of the double-layer potential ∂ν U2 [Q] are continuous
across the boundary ∂Γ while the double-layer potential U2 [Q] and the normal
∂
derivative of the single-layer potential ∂ν U1 [Q̃] are not. However, one has to be
∂
careful about ∂ν U2 [Q]: it is only well-defined on ∂Γ under higher smoothness as-
sumptions on Q than just C(0) (∂Γ). Therefore, we only supplied the jump relation
for this particular case but not the limit relation, which is sufficient for most
theoretical considerations.
Remark 1.13. The relations in Theorem 1.11 were formulated with respect to the
uniform topology for Q, Q̃ ∈ C(0) (Ω). However, they can also be formulated with
respect to the L2 (Ω)-topology for Q, Q̃ ∈ L2 (Ω). For details, the reader is again
referred to [15, 17] and earlier references therein.
2. Boundary value problems for the Beltrami operator

In this section, we investigate the problems (PP), (DP), and (NP) and try to
obtain integral representations of their solutions.
2.1. Poisson problem
We remember the Poisson problem (PP) from the beginning of this chapter: Let
H be of class C(1) (Γ), then we are looking for a function U of class C(2) (Γ) such
that
Δ∗ U (ξ) = H(ξ), ξ ∈ Γ. (36)
1

If we choose H̄ = H − Γ Γ
H(η)d(η), we find that Γ H̄(η)dω(η) = 0 and, by
Theorem 1.10, that

Ū (ξ) = G(Δ∗ ; ξ · η)H̄(η)dω(η), ξ ∈ Γ, (37)
Γ
satisfies Δ∗ Ū (ξ) = H̄(ξ), for ξ ∈ Γ.
1
Setting U (ξ) = Ū (ξ) − Γ ln(1 − ξ · ξ̄) Γ H(η)dω(η), for some fixed ξ¯ ∈ Γc ,
we eventually obtain the desired solution satisfying

1
Δ∗ U (ξ) = H̄(ξ) + H(η)dω(η) = H(ξ), ξ ∈ Γ. (38)
Γ Γ
The solution of (36), however, is not unique. Subscribing further boundary values
on U , e.g., Dirichlet boundary values U − (ξ) = F (ξ), for ξ ∈ ∂Γ, it is possible to
obtain uniqueness. Letting Ũ denote the function U from (38) that we constructed
before, we can formulate the boundary value problem of finding a function Ũ ˜ that
solves
˜ (ξ) = 0, ξ ∈ Γ,
Δ∗ Ũ (39)
˜ − (ξ) = F (ξ) − Ũ − (ξ),
Ũ ξ ∈ ∂Γ. (40)
The newly obtained function U = Ũ + Ũ ˜ would then satisfy the desired differential
equation (36) and the desired Dirichlet boundary values. Boundary value problems
such as (39), (40) are studied in more detail in the upcoming section.
830 C. Gerhards
2.2. Dirichlet and Neumann problem

We take a closer look at the following boundary value problems that have already
been mentioned in the introduction:
Dirichlet Problem (DP): Let F be of class C(0) (∂Γ). We are looking for a func-
tion U of class C(2) (Γ) ∩ C(0) (Γ) such that
Δ∗ U (ξ) = 0, ξ ∈ Γ, (41)
−
U (ξ) = F (ξ), ξ ∈ ∂Γ. (42)
Neumann Problem (NP): Let F be of class C(0) (∂Γ). We are looking for a func-
tion U of class C(2) (Γ) ∩ C(0) (Γ), with a well-defined normal derivative ∂ν
∂
U−
on ∂Γ, such that
Δ∗ U (ξ) = 0, ξ ∈ Γ, (43)
∂ −
U (ξ) = F (ξ), ξ ∈ ∂Γ. (44)
∂ν
First, we formalize the term U − (ξ). For ξ ∈ ∂Γ, it is meant in the sense

ξ − τ ν(ξ)
U − (ξ) = lim U √ , (45)
τ →0+ 1 + τ2
i.e., we approach the boundary ∂Γ in normal direction from within Γ. The term
U + (ξ) is meant in the sense

+ ξ + τ ν(ξ)
U (ξ) = lim U √ , (46)
τ →0+ 1 + τ2
i.e., we approach the boundary ∂Γ in normal direction from the outside of Γ (or,
in other words, from within Γc ). The expressions ∂ν ∂
U ± (ξ) are meant analogously.
We can already see the connection to the limit- and jump-relations from Theorem
1.11. More precisely, making the ansatz U = U2 [Q] for the Dirichlet problem (DP)
and U = U1 [Q̃] for the Neumann problem (NP), Theorem 1.11 yields the following
closely related problems:
Integral Dirichlet Problem (IDP): Let F be of class C(0) (∂Γ). We are looking
for some Q of class C(0) (∂Γ) that satisfies
1
F (ξ) = U2 [Q](ξ) + Q(ξ), ξ ∈ ∂Γ. (47)
2
Integral Neumann Problem (INP): Let F be of class C(0) (∂Γ). We are looking
(0)
for some Q̃ of class C0 (∂Γ) that satisfies
1
F (ξ) = U1 [Q̃](ξ) − Q̃(ξ), ξ ∈ ∂Γ. (48)
2
In other words, the Dirichlet problem (DP) and the Neumann problem (NP) have
been reduced to the Fredholm equations (IDP) and (INP). These boundary inte-
gral formulations have been used, e.g., in [20, 34] to numerically solve the original
boundary value problems for the Beltrami operator. In this section, however, we
are mainly interested in (IDP) and (INP) as tools to guarantee the existence of
solutions to (DP) and (NP) via the Fredholm alternative. Uniqueness of the solu-
tions can be obtained via the application of the maximum principle from Theorem
1.8 and the Green formulas from Theorem 1.1.
Remark 2.1. There are two noteworthy differences in comparison to the Euclidean
case. First, considerations on the sphere do not require a clear distinction between
interior and exterior problems since the open complement Γc of a bounded regular
region Γ ⊂ Ω is again a bounded regular region. Second, the single-layer potential
(0) (0)
U1 [Q̃] is only harmonic if Q̃ ∈ C0 (∂Γ). A solution of (48) in C0 (∂Γ) exists if
(0)
and only if F is of class C0 (∂Γ), which suits the general necessary condition for
the existence of a solution to (NP) that can be obtained from Green’s formulas.
However, it should be mentioned that the integral equation (48) additionally has
a unique solution Q̃ ∈ C(0) (∂Γ) if F is of class C(0) (∂Γ). This is not true for the
Euclidean counterpart.
Summarizing, and including the considerations from Section 2.1, we obtain
the following results. For details, the reader is again referred to [15] and, for the
Euclidean counterparts, to [17, 26, 30, 45].
Theorem 2.2 (Uniqueness).
(a) A solution of (DP) is uniquely determined.
(b) A solution of (NP) is uniquely determined up to an additive constant.
Theorem 2.3 (Existence for Generalized (DP)). Let F be of class C(0) (∂Γ) and H
of class C(1) (Γ). Then there exists a unique solution U of class C(2) (Γ) ∩ C(0) (Γ)
of the Dirichlet problem
Δ∗ U (ξ) = H(ξ), ξ ∈ Γ, (49)
−
U (ξ) = F (ξ), ξ ∈ ∂Γ. (50)
Theorem 2.4 (Existence for Generalized (NP)). Let F be of class C(0) (∂Γ) and H
of class C(1) (Γ). Then there exists an up to an additive constant uniquely deter-
mined solution U of class C(2) (Γ) ∩ C(0) (Γ), with a well-defined normal derivative
∂ −
∂ν U on ∂Γ, to the Neumann problem
Δ∗ U (ξ) = H(ξ), ξ ∈ Γ, (51)
∂ −
U (ξ) = F (ξ), ξ ∈ ∂Γ, (52)
∂ν
if and only if

F (η)dσ(η) − H(η)dω(η) = 0. (53)
∂Γ Γ
Proof. The condition (53) is a simple consequence from

∗ ∂
H(η)dω(η) = Δ U (η)dω(η) = U (η)dσ(η) = F (η)dσ(η), (54)
Γ Γ ∂Γ ∂ν ∂Γ
832 C. Gerhards
where Green’s formulas have been used for the second equation. The general exis-
tence follows from the application of the Fredholm alternative to (INP).
2.3. Green’s functions
Next, we are interested in the representation of a solution to (DP) and (NP). A
possibility is indicated in Theorem 1.3(a). However, this representation requires the
∂
simultaneous knowledge of U and ∂ν U on the boundary ∂Γ, which is not necessary
and can be problematic since the two quantities are not independent from each
other. As a remedy, Green’s functions for Dirichlet and Neumann boundary values
can be used.
More precisely, a function GD (Δ∗ ; ·, ·) is called a Dirichlet Green function
(with respect to the Beltrami operator) if it can be decomposed in the form
GD (Δ∗ ; ξ, η) = G(Δ∗ ; ξ · η) − ΦD (ξ, η), η ∈ Γ, ξ ∈ Γ, ξ = η, (55)
where ΦD (ξ, ·) is of class C(2) (Γ) ∩ C(1) (Γ) and satisfies
1
Δ∗η ΦD (ξ, η) = − , η ∈ Γ, (56)
4π
Φ−
D (ξ, η) = G(Δ ∗
; ξ · η), η ∈ ∂Γ, (57)
∗
for every ξ ∈ Γ. Analogously, a function GN (Δ ; ·, ·) is called a Neumann Green
function (with respect to the Beltrami operator) if it can be decomposed in the
form
GN (Δ∗ ; ξ, η) = G(Δ∗ ; ξ · η) − ΦN (ξ, η), η ∈ Γ, ξ ∈ Γ, ξ = η, (58)
where ΦN (ξ, ·) is of class C(2) (Γ) ∩ C(1) (Γ) and satisfies the conditions
1 1
Δ∗η ΦN (ξ, η) = − , η ∈ Γ, (59)
Γ 4π
∂ ∂
Φ− (ξ, η) = G(Δ∗ ; ξ · η), η ∈ ∂Γ, (60)
∂ν(η) N ∂ν(η)
for every ξ ∈ Γ. Using Theorem 1.3(a) and Theorem 1.1(c) for ΦD and ΦN , we
eventually achieve the representations

∂
U (ξ) = GD (Δ∗ ; ξ, η)Δ∗η U (η)dω(η) + U (η) GD (Δ∗ ; ξ, η)dσ(η) (61)
Γ ∂Γ ∂ν(η)
and

1
U (ξ) = U (η)dω(η) + GN (Δ∗ ; ξ, η)Δ∗η U (η)dω(η)
Γ Γ Γ

∂
− GN (Δ∗ ; ξ, η) U (η)dσ(η), (62)
∂Γ ∂ν(η)
which yield integral representations for solutions to (DP) and (NP), respectively,
under the condition that U is of class C(2) (Γ). It remains to construct the auxiliary
functions ΦD and ΦN . Some general construction principles on the sphere can be
found, e.g., in [24, 32]. In this chapter, we focus on spherical caps Γρ (ζ). The
Kelvin transf.
Ú
2 z pstereo(x) pstereo(h) pstereo(x)
R
x
Gr(z)
h
Ú
x
-z
Figure 2. Schematic description of the construction of the reflection
point ξ̌.
procedure is similar to the construction of a Dirichlet Green function for a disc in

R2 . For ξ ∈ Γρ (ζ), we need to find a reflection point ξˇ ∈ (Γρ (ζ))c and a scaling
factor ř ∈ R such that

1 − ξ · η = ř 1 − ξˇ · η , η ∈ ∂Γρ (ζ), ξ ∈ Γρ (ζ). (63)
Indeed, under this assumption, it is clear that
1 1
ΦD (ξ, η) = ln(ř(1 − ξˇ · η)) + (1 − ln(2)) (64)
4π 4π
satisfies the desired conditions (56) and (57). The reflection point ξˇ can be obtained
by a stereographic projection of ξ onto R2 , then applying a Kelvin transform to
the projection point, and eventually projecting it back to the sphere (cf. Figure 2
for an illustration). ξˇ represents the spherical Kelvin transformation of ξ. The
scaling factor ř is obtained by solving (63). Alternatively, the entire Dirichlet
Green function GD (Δ∗ ; ·, ·) can be obtained from a stereographic projection of the
Dirichlet Green function for the Laplace operator on a disc in R2 . But this route
would not supply us with a spherical counterpart to the Kelvin transform. We can
conclude our considerations with the following theorem.
Theorem 2.5. Let Γ = Γρ (ζ) be a spherical cap with center ζ ∈ Ω and radius
ρ ∈ (0, 2). Furthermore, for ξ ∈ Γρ (ζ) we set
1 ř − 1
ξ̌ = ξ − ζ, (65)
ř ř(ρ − 1)
1 + 2ξ · ζ(ρ − 1) + (ρ − 1)2
ř = − . (66)
ρ(ρ − 2)
Then
1 1
GD (Δ∗ ; ξ, η) = ln(1 − ξ · η) − ln(ř(1 − ξ̌ · η)), (67)
4π 4π
834 C. Gerhards
and a solution U ∈ C(2) (Γ) of the Dirichlet problem (DP) can be represented by

1 ξ·ζ +ρ−1 1
U (ξ) = # F (η) dσ(η), ξ ∈ Γρ (ζ). (68)
2π ρ(2 − ρ) ∂Γρ (ζ) 1 − ξ·η
Remark 2.6. Applying Theorem 2.5 for ζ = ξ leads to the Mean Value Property
II from Theorem 1.7.
A Neumann Green function for the Beltrami operator cannot be obtained by
a simple stereographic projection of the Neumann Green function for the Laplace
operator on a disc in R2 . But some computations based on the previously obtained
auxiliary function ΦD yield the following theorem.
Theorem 2.7. Let Γ = Γρ (ζ) be a spherical cap with center ζ ∈ Ω and radius
ρ ∈ (0, 2). Furthermore, let ξˇ and ř be given as in Theorem 2.5. Then, a Neumann
Green function is given by
1 1 1−ρ
GN (Δ∗ ; ξ, η) = ln(1 − ξ · η) + ln(ř(1 − ξ̌ · η)) + ln(1 + ζ · η). (69)
4π 4π 2πρ
A solution U ∈ C(2) (Γ) of the Neumann problem (NP) can be represented by

1
U (ξ) = U (η)dω(η) (70)
2πρ Γρ (ζ)

1 1−ρ
− ln(1 − ξ · η) + ln(2 − ρ) F (η) dσ(η), ξ ∈ Γρ (ζ).
∂Γρ (ζ) 2π 2πρ
3. Spherical decompositions and first-order differential equations

In this section we treat differential equations for the surface gradient ∇∗ and the
surface curl gradient L∗ . They come up, e.g., when dealing with vertical deflec-
tions and geostrophic ocean flow. Additionally, we take a look at some spherical
decompositions of vector fields that are particularly useful in geosciences.
3.1. Surface gradient and surface curl gradient
Different from the Poisson equation, solutions of the differential equations with
respect to the surface gradient and the surface curl gradient are uniquely deter-
mined up to an additive constant on regular regions Γ ⊂ Ω, without the necessity
of boundary values. Also the existence of a solution can be easily guaranteed. This
is summarized in the following two lemmas.
Lemma 3.1 (Uniqueness). Let U be of class C(1) (Γ). Then
∇∗ U (ξ) = 0, ξ ∈ Γ, (71)
∗
if and only if U is constant on Γ. The same holds true for L U (ξ) = 0, ξ ∈ Γ.
Theorem 3.2 (Existence).
(a) Let f ∈ c(1) (Γ) be a tangential vector field satisfying
L∗ · f (ξ) = 0, ξ ∈ Γ. (72)
Then there exists a function U of class C(2) (Γ), which is uniquely determined
up to an additive constant, such that
f (ξ) = ∇∗ U (ξ), ξ ∈ Γ. (73)
(1)
(b) Let f ∈ c (Γ) be a tangential vector field satisfying
∇∗ · f (ξ) = 0, ξ ∈ Γ. (74)
Then there exists a function U of class C(2) (Γ), which is uniquely determined
up to an additive constant, such that
f (ξ) = L∗ U (ξ), ξ ∈ Γ. (75)
From Theorem 1.3(b), we know a possible expression of the solutions to the
differential equations for the surface gradient and the surface curl gradient. How-
ever, this representation requires the knowledge of U on the boundary ∂Γ, which
is actually not necessary according to Theorem 3.2. Using a Neumann Green func-
tion together with the identities in Theorem 1.3(b) directly implies the following
results.
Theorem 3.3.
(a) Let f of class c(1) (Γ) be a tangential vector field satisfying L∗ · f (ξ) = 0,
ξ ∈ Γ. Then a solution of
f (ξ) = ∇∗ U (ξ), ξ ∈ Γ, (76)
is given by

1 ∗
U (ξ) = U (η)dω(η) − ∇η GN (Δ∗ ; ξ, η) · f (η)dω(η). (77)
Γ Γ Γ
(b) Let f of class c(1) (Γ) be a tangential vector field satisfying ∇∗ · f (ξ) = 0,
ξ ∈ Γ. Then a solution of
f (ξ) = L∗ U (ξ), ξ ∈ Γ, (78)
is given by

1 ∗
U (ξ) = U (η)dω(η) − Lη GN (Δ∗ ; ξ, η) · f (η)dω(η). (79)
Γ Γ Γ
Remark 3.4. If we deal with the entire sphere Γ = Ω, the same results as in the
preceding theorem hold true. For the integral representations, one simply has to
substitute the Neumann Green function by the fundamental solution G(Δ∗ ; ·).
3.2. Helmholtz and Hardy–Hodge decomposition
We begin with the spherical Helmholtz decomposition of a vector field f . It essen-
tially describes the split-up of the vector field into a radial and two tangential com-
ponents, of which one is surface curl-free and the other one surface divergence-free.
In geomagnetism, this has applications, e.g., in the separation of polar ionospheric
current systems into field-aligned currents (which are nearly radial in polar re-
gions) and Pedersen and Hall currents (see, e.g., [1, 2, 21, 39]). In other areas, the
836 C. Gerhards
spherical Helmholtz decomposition has a natural connection as well: geostrophic

ocean flow, e.g., is purely tangential and surface divergence-free while the vertical
deflection of the geoidal normal vector is approximately purely tangential and sur-
face curl-free. For convenience, we use the following notations for the Helmholtz
operators acting on a scalar function F at a point ξ ∈ Ω:
o(1) F (ξ) = ξF (ξ), o(2) F (ξ) = ∇∗ F (ξ), o(3) F (ξ) = L∗ F (ξ). (80)
Writing f = o(1) F1 +o(2) F2 +o(3) F3 on a subdomain Γ and using the orthogonality
of the three operators, we obtain Δ∗ F3 (ξ) = L∗ · f (ξ), ξ ∈ Γ. Latter can be solved
by the methods of the previous section. We need to prescribe boundary data on F3
in order to obtain uniqueness of the scalar function F3 . All in all, we can formulate
Decomposition Theorem 3.5. More details can be found, e.g., in [15, 21].
Theorem 3.5 (Spherical Helmholtz Decomposition). Let f be of class c(2) (Γ). Then
there exist scalar fields F1 of class C(2) (Γ) and F2 , F3 of class C(2) (Γ) such that
f (ξ) = o(1) F1 (ξ) + o(2) F2 (ξ) + o(3) F3 (ξ), ξ ∈ Γ. (81)
Uniqueness of F1 , F2 , F3 is guaranteed by the properties

F2 (η)dω(η) = 0 (82)
Γ
and
F3− (ξ) = F (ξ), ξ ∈ ∂Γ, (83)
for a fixed function F of class C(0) (∂Γ). The Helmholtz scalars F1 , F2 , and F3 can
be then represented by

∗
F2 (ξ) = − ∇η GN (Δ∗ ; ξ, η) · f (η)dω(η)
Γ
+ F (η) τη · ∇∗η GN (Δ∗ ; ξ, η)dσ(η), ξ ∈ Γ (84)
∂Γ
∗
F3 (ξ) = − Lη GD (Δ∗ ; ξ, η) · f (η)dω(η)
Γ
+ GD (Δ∗ ; ξ, η)τη · f (η)dσ(η)
∂Γ
∂
+ F (η) GD (Δ∗ ; ξ, η)dσ(η), ξ ∈ Γ (85)
∂Γ ∂ν η

for ξ ∈ Γ. Additionally, if Γ F1 (η)dω(η) = 0, then

F1 (ξ) = ξ · f (ξ) = Δ∗ξ G(Δ∗ ; ξ · η) η · f (η)dω(η), ξ ∈ Γ. (86)
Γ
Remark 3.6. Clearly, the type of boundary conditions that have to be prescribed to
obtain uniqueness of the Helmholtz decomposition can be varied. They can be im-
posed on F2 instead of F3 , or the Dirichlet boundary conditions can be substituted
by Neumann boundary conditions. Neumann boundary conditions are occasion-

ally more advantageous as they allow the imposition of boundary information on
the normal and tangential direction of the vectorial quantities o(2) F2 and o(3) F3 ,
respectively, which are in some cases better accessible from the given data than
the scalars F2 or F3 . Representations analogous to Theorem 3.5 can be derived by
Green’s formulas and the results from Section 2.
Remark 3.7. For the particular caseΓ = Ω, the results from Theorem 3.5 hold
true as well if the boundary integrals ∂Γ . . . dσ are dropped and the Neumann and
Dirichlet Green functions are substituted by the fundamental solution G(Δ∗ ; ·).
For the uniqueness, condition (83) has to be substituted by Ω F3 (η)dω(η) = 0.
We then obtain

∗
F2 (ξ) = − ∇η G(Δ∗ ; ξ · η) · f (η)dω(η), ξ ∈ Ω, (87)
Ω
∗
F3 (ξ) = − Lη G(Δ∗ ; ξ · η) · f (η)dω(η), ξ ∈ Ω. (88)
Ω
Thus, in the global case Γ = Ω, the Helmholtz scalars F2 and F3 are determined
uniquely up to an additive constant without further constraints. The vectorial
quantities o(2) F2 and o(3) F3 are actually uniquely determined. This is not true for
general subdomains Γ ⊂ Ω.
Next, we turn to a different spherical decomposition, the so-called spherical
Hardy–Hodge decomposition (the name is adopted from the Euclidean decomposi-
tion presented in [3], although its spherical version is known and used significantly
longer, e.g., in [2, 18, 21, 36, 40] and references therein). It is based on the set of
operators

1 1
õ(1) = o(1) D + − o(2) , õ(2) = o(1) D − + o(2) , õ(3) = o(3) , (89)
2 2
1
where the operator D is given by D = −Δ∗ + 14 2 . A decomposition in terms
of these operators can be interpreted as a decomposition of a spherical vectorial
signal with respect to sources lying inside a given sphere (reflected by the õ(1) -
contributions), sources lying in the exterior of the sphere (õ(2) -contributions), and
sources on the sphere (õ(3) -contributions). For the gravitational field measured at
satellite altitude, e.g., only the õ(1) -contribution is of relevance. Concerning the
Earth’s crustal magnetization, only the õ(2) -contribution of the magnetization gen-
erates a magnetic effect at satellite altitude. The generated magnetic field itself,
however, only consists of õ(1) -contributions since its source (i.e., the magnetization)
is located inside the satellite’s orbit. The decomposition and the integral repre-
sentation of its scalar functions can be closely related to the spherical Helmholtz
decomposition. For details, we refer the reader to [15, 18, 21, 22]. Yet, the non-local
structure of the operator D makes it very difficult to obtain results on subdomains
Γ ⊂ Ω. Therefore, the following theorem only treats the decomposition for the case
Γ = Ω.
838 C. Gerhards
Theorem 3.8 (Spherical Hardy–Hodge Decomposition). Let f be of class c(1) (Ω).

Then there exist scalar fields F̃1 , F̃2 , F̃3 of class C(2) (Ω) such that
f (ξ) = õ(1) F̃1 (ξ) + õ(2) F̃2 (ξ) + õ(3) F̃3 (ξ), ξ ∈ Ω. (90)
Uniqueness of F̃1 , F̃2 , F̃3 is guaranteed by the properties

F̃3 (η)dω(η) = 0, (91)
Ω

F̃1 (η) − F̃2 (η)dω(η) = 0. (92)
Ω
The Hardy–Hodge scalars F̃1 , F̃2 , and F̃3 can then be represented by
1 1 1
F̃1 = D−1 F1 + D−1 F2 − F2 , (93)
2 4 2
1 1 1
F̃2 = D−1 F1 + D−1 F2 + F2 , (94)
2 4 2
F̃3 = F3 , (95)
where F1 , F2 , F3 are the Helmholtz scalars from Theorem 3.5 and Remark 3.7.
Remark 3.9. The operator D−1 can be represented as the convolution operator

1 1
D−1 F (ξ) = # F (η)dω(η), ξ ∈ Ω, (96)
2π Ω 2(1 − ξ · η)
acting on a function F of class C(0) (Ω). Thus, equations (93)–(95) together with
Theorem 3.5 and Remark 3.7 form integral representations of the Hardy–Hodge
scalars.
4. Complete function systems

In the Euclidean setting, spherical harmonics form a complete function system in
L2 (ΩR ), and their harmonic extensions into the ball BR = {x ∈ R3 : |x| < R}
and its exterior BcR = {x ∈ R3 : |x| > R} (so-called inner and outer harmonics,
respectively) form suitable function systems to approximate functions that are
harmonic with respect to the Laplace operator. The limit- and jump-relations of
layer potentials enable the extension of the completeness results to more general
manifolds than the sphere. With the considerations of the previous sections at
hand, we are now able to formulate analogous completeness results for function
systems on general curves ∂Γ. We obtain completeness for certain function systems
in L2 (∂Γ) whose harmonic extensions into Γ ⊂ Ω are particularly well suited for
the approximation of functions that are harmonic with respect to the Beltrami
operator.
First, we need the notion of a fundamental system: Suppose that {ξk }k∈N ⊂ Γ
is a set of points satisfying
dist({ξk }k∈N , ∂Γ) > 0. (97)
xk
¶G
Figure 3. Example for a fundamental system {ξk }k∈N (with respect to Γ).
If, for any harmonic function F in Γ, the condition F (ξk ) = 0, k ∈ N, implies that
F (ξ) = 0 for all ξ ∈ Γ, then we call {ξk }k∈N a fundamental system (with respect
to Γ). Assuming that Σ ⊂ Γ is a regular region with dist(Σ, ∂Γ) > 0, an example
for such a fundamental system is given by a dense point set {ξk }k∈N ⊂ ∂Σ. A
particularly simple choice for Σ is a spherical cap within Γ (cf. Figure 3).
We begin with the completeness of function systems based on the fundamen-
tal solution for the Beltrami operator.
Theorem 4.1. Let {ξk }k∈N be a fundamental system with respect to Γ. Then the
following statements hold true:
(a) The function system {Gk }k∈N0 given by
1 1
Gk (ξ) = ln(1 − ξk · ξ), k ∈ N, G0 (ξ) = ,
4π 4π
is complete, and hence closed in L2 (∂Γ).
(b) The function system {G̃k }k∈N0 , given by
1 ∂ 1
G̃k (ξ) = ln(1 − ξk · ξ), k ∈ N, G̃0 (ξ) = ,
4π ∂ν(ξ) 4π
is complete, and hence closed in L2 (∂Γ).
Remark 4.2. Let us assume that {ξk }k∈N is a fundamental system with respect to
Γc . Then the functions G̃k from Theorem 4.1 are harmonic in Γ and, thus, partic-
ularly suitably for the approximation of harmonic functions in Γ. The functions
1
Gk from Theorem 4.1 need to be modified since they only satisfy Δ∗ Gk (ξ) = − 4π ,
for ξ ∈ Γ and k ∈ N. Any auxiliary function G of class C(2) (Γ) that satisfies
1
Δ∗ G(ξ) = 4π , ξ ∈ Γ, can be added to Gk without changing the completeness
property. In other words, e.g.,
(mod) 1 (mod)
Gk (ξ) = Gk (ξ) − ln(1 − ξ · ξ̄), k ∈ N, G0 (ξ) = G0 (ξ),
4π
with a fixed ξ¯ ∈ Γc , forms a complete function system in L2 (∂Γ) that additionally
(mod)
satisfies Δ∗ Gk (ξ) = 0, ξ ∈ Γ.
840 C. Gerhards
Next, we want to transfer the results from Theorem 4.1 to inner harmonics
for spherical caps. In order to achieve this, we first need to clarify what we mean by
inner harmonics for spherical caps. The sine and cosine functions obviously take
the role of spherical harmonics on a circle in R2 . Their harmonic continuations
into the disc DR = {x ∈ R2 : |x| < R} with radius R > 0 and into its exterior
DcR = {x ∈ R2 : |x| > R} (the so-called inner and outer harmonics, respectively)
are given by
(int) 1 r n
Hn,1 (R; x) = √ cos(nϕ), n ∈ N0 , x ∈ DR , (98)
R π R
1
r n
(int)
Hn,2 (R; x) = √ sin(nϕ), n ∈ N, x ∈ DR , (99)
R π R
n
(ext) 1 R
Hn,1 (R; x) = √ cos(nϕ), n ∈ N0 , x ∈ DcR , (100)
R π r
n
(ext) 1 R
Hn,2 (R; x) = √ sin(nϕ), n ∈ N, x ∈ DcR , (101)
R π r
where x = (r cos(ϕ), r sin(ϕ))T , r ≥ 0, ϕ ∈ [0, 2π). Inner harmonics on a spherical

cap Γρ (ζ) with radius ρ ∈ (0, 2) and center ζ ∈ Ω can then be obtained by a simple
stereographic projection. More precisely,
1
ρ,ζ (int) 1
Hn,k (ξ) = Hn,k ρ 4 (2 − ρ) 4 ; pstereo (ζ; ξ) , ξ ∈ Γρ (ζ), (102)
denotes an inner harmonic (of degree n and order k) on Γρ (ζ). The applied stere-
ographic projection pstereo (ζ; ·) : Ω \ {−ζ} → R2 is defined via

2ξ · (tε1 ) 2ξ · (tε2 )
pstereo (ζ; ξ) = , , (103)
1+ξ·ζ 1+ξ·ζ
where ε1 = (1, 0, 0)T , ε2 = (0, 1, 0)T , ε3 = (0, 0, 1)T denotes the canonical ba-
sis in R3 and t ∈ R3×3 a rotation matrix with tε3 = ζ. From the harmonicity
(int) ρ,ζ
of Hn,k (R; ·) in DR it follows that Hn,k is harmonic in Γρ (ζ). Note that, as al-
ways, harmonicity in the Euclidean space R2 is meant with respect to the Laplace
operator while it is meant with respect to the Beltrami operator when we are
intrinsic on the sphere Ω. Opposed to the Euclidean case, outer harmonics for
spherical caps do not play a distinct role. Actually, for a spherical cap Γρ (ζ), the
corresponding outer harmonics coincide with the inner harmonics for the spherical
c
cap (Γρ (ζ)) = Γ2−ρ (−ζ), which is why we do not consider them separately. The
relation
ln(1 − ξ · η) = − ln(2) + ln(1 + ξ · ζ) + ln(1 − η · ζ)

# ∞ 2
2 ρ,ζ 2−ρ,−ζ
− ρ(2 − ρ)π Hn,k (ξ)Hn,k (η), (104)
n=1
n
k=1
for ξ ∈ Ω\{−ζ}, η ∈ Ω\{ζ}, and |pstereo (ζ; ξ)| < |pstereo (ζ; η)|, eventually allows to
transfer the completeness results from Theorem 4.1 to inner harmonics on spherical
caps (for details, the reader is referred to [15]).
Theorem 4.3. Let Γρ (ζ) be a spherical cap with Γ ⊂ Γρ (ζ). Then the following
statements hold true:
! ρ,ζ " ! ρ,ζ "
(a) The inner harmonics H0,1 ∪ Hn,k n∈N,k=1,2 form a complete, and hence
closed function system in L2 (∂Γ).
(b) The normal derivatives of the inner harmonics, i.e.,
! ρ,ζ " ! ∂ ρ,ζ "
H0,1 ∪ ∂ν Hn,k n∈N,k=1,2
form a complete and hence closed function system in L2 (∂Γ).
We conclude this section by stating the use of the function systems from above
for the approximation of solutions to the spherical boundary value problems (DP)
and (NP) from Section 2.
Theorem 4.4. Let {Φk }k∈N0 denote one of the function systems introduced in The-
orem 4.1(b), Remark 4.2, or Theorem 4.3, and U ∈ C(2) (Γ) ∩ C(0) (Γ) be a solution
of one of the boundary value problems (DP) or (NP). Then, for every ε > 0, there
exist M ∈ N0 and coefficients ak ∈ R, k = 0, 1, . . . , M , such that
- -
- M -
- -
-U − a k Φk - < ε. (105)
- -
k=0 L2 (Γ)
The choice of M and the coefficients ak , k = 1, . . . , M , can be based solely on an
∂
approximation of U or ∂ν U on the boundary ∂Γ.
Remark 4.5. All the density and approximation results that were obtained in this
section in an L2 -context also hold true in a C(0) -context with respect to the uniform
topology and can be shown by the tools supplied throughout this chapter (see, e.g.,
[15, 17]).
5. Applications in geoscience
In this section, we present some applications of the previous tools to the approx-
imation of different quantities of interest in physical geodesy. More precisely, we
use techniques from Section 3.1 to reconstruct the disturbing potential from given
vertical deflections over South America and the mean dynamic ocean topography
(MDT) from given geostrophic ocean flow patterns over the Pacific Ocean, respec-
tively. We will be rather brief about the geophysical derivations of the underlying
spherical differential equations and refer the reader to classical literature such as
[25, 28, 42, 44]. The particular formulations of our setting can also be found, e.g.,
in [10, 11, 13, 18]. Opposed to the latter, our reconstructions in Sections 5.1 and
5.2 are based on the approach in Section 3.1 via Neumann Green functions and
does not require boundary information for the spherical caps under consideration.
842 C. Gerhards
Scale J=4 Scale J=6 Scale J=10
120o W 90o W 60o W 30o W 120o W 90o W 60o W 30o W 120o W 90o W 60o W 30o W
0.55
30o N 30o N 1 30o N
3.5
0.5
0.9
o
15 N 15o N 15o N
0.45 3
0.8
o
0 0.4 0o 0o
0.7 2.5
0.35
o
15 S 15o S 0.6 15o S
0.3 2
o 0.5
30o S 0.25 30 S 30o S
1.5
0.4
0.2
45o S 45o S 0.3
o
45 S
0.15 1
0.1 0.2
0.5
o o
60 S 0.05 60 S 0.1 60o S
Figure 4. The kernel ∇∗ GJN (Δ∗ ; ξ, ·) for scales J = 4, 6, 10 and a fixed

evaluation point ξ located at 7◦ N, 74◦ W (colors indicate the absolute
value and arrows the orientation).
In Section 5.3, based on the results from Section 4, we address a model problem
motivated by point vortex motion on the sphere.
5.1. Vertical deflections

The Earth’s gravity potential W = U + T is typically split into a normal gravity
potential U corresponding to a reference ellipsoid E (i.e., U (x) = const. for x ∈
E) and a smaller remaining disturbing potential T . The vertical deflection Θ(x)
measures the angular distance between the normal vector νG (x) at a point x on the
geoid G (i.e., W (x) = const. for x ∈ G) and the corresponding ellipsoidal normal
vector νE (x) with respect to E. Assuming that νG −νE and νE are nearly orthogonal
and that the deviation of the reference ellipsoid from a sphere is negligible, one can
derive the following relation between the disturbing potential and the deflections
of the vertical:
GM
∇∗ T (Rξ) = − Θ(Rξ), ξ ∈ Ω, (106)
R
where R is the Earth’s mean radius, G the gravitational constant, and M the
Earth’s mass. For more details, the reader is referred to, e.g., [18, 25, 28]. We are
particularly interested in solving (106) for the disturbing potential T in a subregion
Γ ⊂ Ω (or, in other words, in a subregion ΓR of the spherical Earth’s surface ΩR )
from knowledge of the vertical deflections Θ only in that subregion. Theorem 3.3
yields the representation

1 GM ∗
T (Rξ) = T (Rη)dω(η) + ∇η GN (Δ∗ ; ξ, η) · Θ(Rη)dω(η), ξ ∈ Γ,
Γ Γ R Γ
(107)
of which the first summand on the right-hand side simply represents the constant
Γ
mean disturbing potential Tmean in ΓR . We focus on the special case that Γ = Γρ (ζ)
is a spherical cap with center ζ ∈ Ω and radius ρ ∈ (0, 2), so that Theorem
2.7 supplies us with an explicit representation of the Neumann Green function
GN (Δ∗ ; ·, ·).
Concerning the numerical evaluation of (107), we first need to discretize the
integral since Θ is typically only available in a discrete set of measurement points.
Disturbing Potential Vertical Deflections

−5
o o o o o o o o x 10
120 W 90 W 60 W 30 W 120 W 90 W 60 W 30 W
o o
30 N 300 30 N 14
o o
15 N 15 N 12
200
o o
0 0 10
100
o
15 S 15o S
8
m2/s2
rad
0
30o S 30o S
6
−100
o o
45 S 45 S 4
−200
2
o
60o S 60 S
−300
Figure 5. The ‘true’ disturbing potential T (left) and the correspond-

ing vertical deflections Θ (right; colors indicate the absolute value and
arrows the orientation).
For the tests in this section, we assume Θ to be given on a Gauss–Legendre grid

in the spherical cap Γρ (ζ), so that we can use the quadrature rule from [27].
Second, the numerical integration can become instable due to the singularity of
the Neumann Green function GN (Δ∗ ; ξ, η) at ξ = η (originating in its contribution
1
4π ln(1−ξ·η)). This can be circumvented by a regularization around this singularity
via a truncated Taylor expansion. More precisely, for scaling parameters J =
0, 1, 2, . . ., we define the regularized Neumann Green function
⎧ 1 1 ˇ
⎪
⎪ 4π ln(1 − ξ · η) + 4π ln(ř(1 − ξ · η))
⎪ +
⎨ 1−ρ
1 − ξ · η ≥ 2−J ,
2πρ ln(1 + ζ · η),
GJN (Δ∗ ; ξ, η) =
⎪
⎪ 2J
(1 − ξ · η) − 4π
J 1
ln(2) − 4π
⎪
⎩ 4π 1 1−ρ
+ 4π ln(ř(1 − ξ · η)) + 2πρ ln(1 + ζ · η), 1 − ξ · η < 2−J .
ˇ
(108)
The regularization GJN (Δ∗ ; ·, ·) of the Neumann Green function GN (Δ∗ ; ·, ·) closely
relates to the regularization of the fundamental solution G(Δ∗ ; ·) briefly mentioned
after Theorem 1.9. A stable approximation of T at scale J is then given by

Γ GM ∗ J
TJ (Rξ) = Tmean + ∇η GN (Δ∗ ; ξ, η) · Θ(Rη)dω(η), ξ ∈ Γρ (ζ),
R Γρ (ζ)
(109)
and satisfies limJ→∞ supξ∈Γ̃ |TJ (Rξ) − T (Rξ)| = 0 for every subset Γ̃ ⊂ Γρ (ζ) with
dist(Γ̃, ∂Γρ (ζ)) > 0. Thus, higher scales J yield a more precise approximation of
T and the difference TJ+1 − TJ between two consecutive scales reveals features of
more and more local origin. The kernel ∇∗ GJN (Δ∗ ; ξ, ·) is illustrated in Figure 4.
In order to illustrate the reconstruction of the disturbing potential by the
approximations TJ , we first compute a ‘true’ disturbing potential T from EGM2008
844 C. Gerhards
Disturbing Potential, Scale J=6 Disturbing Potential, Difference Scale J=8 and Scale J=6
120o W 90o W 60o W 30o W o
120 W
o
90 W
o
60 W
o
30 W
200 50
o o
30 N 30 N
150 40
o
15o N 15 N
100
30
0o 0o
50
20
15o S 15o S
0
m2/s2
m2/s2
10
o
30 S −50 30o S
0
−100
45 S
o
45o S −10
−150
−20
o −200 o
60 S 60 S
−30
o o o o o o o o
120 W 90 W 60 W 30 W 120 W 90 W 60 W 30 W
250
o
30 N 30o N
200 40
o o
15 N 150 15 N
30
0o 100 0o
20
50
o o
15 S 15 S
m2/s2
m2/s2
0 10
30o S 30o S
−50
0
o −100
45 S 45o S −10
−150
−20
−200
o o
60 S 60 S
−250 −30
120o W 90o W 60o W 30o W 120o W 90o W 60o W 30o W
30o N 30o N 60
250
15 N
o 200 o
15 N
40
150
0o 0o
100 20
15o S 50 15o S
m2/s2
m2/s2
0
0
o o
30 S 30 S
−50 −20
45o S −100 45o S
−150 −40
−200
o o
60 S 60 S −60
−250
Disturbing Potential, Scale J=15

120o W 90o W 60o W 30o W
o 300
30 N
o
15 N
200
0o
100
15o S
m2/s2
0
o
30 S Figure 6. Reconstructions of the dis-
45o S
−100 turbing potential TJ at scales J = 6, 8,
10, 15 (left) and the differences T8 − T6 ,
−200
60 S
o T10 − T8 , T15 − T10 between the recon-
−300 structions at these scales (right).
Disturbing Potential Disturbing Potential, Error Scale J=15

o o o o
120 W 90 W 60 W 30 W 120o W 90o W 60o W 30o W
10
30o N 300 30o N
8
15o N o
15 N 6
200
0o 0o 4
100
o o 2
15 S 15 S
m2/s2
m2/s2
0 0
o
30 S 30o S
−2
−100
o −4
45o S 45 S
−6
−200
o −8
60o S 60 S
−300
−10
Figure 7. The ‘true’ disturbing potential T (left) and the reconstruc-

tion error T − T15 (right).
(cf. [41]∗ ) as a reference, using spherical harmonic degrees n = 3, . . . , 250. From

this T , we obtain our input vertical deflections Θ via (106) on a Gauss–Legendre
grid of 63,252 points in a spherical cap over South America (cf. Figure 5). The
approximations TJ for different scales J are shown in Figure 6. One can clearly
see the refinement of the local features in the differences of the reconstructions TJ .
Furthermore, the error T − T15 in Figure 7 indicates a good approximation of T
and does not reveal any artefacts due to the local reconstruction without use of
any boundary information.
5.2. Geostrophic ocean flow
In subregions ΓR ⊂ ΩR of the ocean with a sufficiently large horizontal extent,
away from the top and bottom Ekman layers and coastal regions, the geostrophic
balance holds true: the horizontal pressure gradients in the ocean balance the
Coriolis force resulting from horizontal currents. The Coriolis force term in a point
x ∈ ΓR is given as the tangential contribution of −2Rρ w ∧ v(x), where v(x) is the
horizontal ocean flow velocity and w = |w|ε3 the Earth’s rotation vector. ρ denotes
the density and is assumed to be constant. The pressure P (x) in x ∈ ΓR can be
regarded as being proportional to the mean dynamic topography (MDT) H(x),
which denotes the height of the sea surface relative to the Geoid G and can be
determined from altimetry measurements. More precisely, P (x) = ρGH(x), where
G denotes the gravitational constant. Using the geostrophic balance, we therefore
obtain
−2Rρ(w · ξ) ξ ∧ v(Rξ) = ρG∇∗ H(Rξ), ξ ∈ Γ, (110)
or, equivalently,
2R
|w|(ξ · ε3 )v(Rξ) = L∗ H(Rξ), ξ ∈ Γ. (111)
G
∗ dataaccessed via
http://earth-info.nga.mil/GandG/wgs84/gravitymod/egm2008/egm08 wgs84.html
846 C. Gerhards
For more details on the geophysical background, the reader is referred, e.g., to
[42, 44]. In order to compute the MDT H from knowledge of the ocean flow velocity
v in Γ, we need to solve Equation (111). Theorem 3.3 yields the representation

1 2R
H(Rξ) = H(Rη)dω(η) − |w| (η · ε3 ) L∗η GN (Δ∗ ; ξ, η) · v(Rη)dω(η),
Γ Γ G Γ
ξ ∈ Γ,
of which the first summand on the right-hand side simply represents the constant
Γ
mean MDT Hmean in ΓR . Again, we focus on the special case that Γ = Γρ (ζ) is
a spherical cap with center ζ ∈ Ω and radius ρ ∈ (0, 2), so that we can apply
the considerations from the previous section, i.e., we obtain an approximation at
scale J by

Γ 2R
HJ (Rξ) = Hmean − |w| (η · ε3 ) L∗η GJN (Δ∗ ; ξ, η) · v(Rη)dω(η),
G Γρ (ζ)
ξ ∈ Γρ (ζ), (112)
where GJN (Δ∗ ; ·, ·) is given as in (108).

In order to illustrate the reconstruction of the MDT by the approximations
HJ , we first compute a ‘true’ MDT H from [35]† as a reference. From this H, we
can obtain our input ocean flow velocity v via (111) on a Gauss–Legendre grid
of 63,252 points in a spherical cap over the Western Pacific Ocean (cf. Figure 8).
The approximations HJ for different scales J are shown in Figure 9. The error
H − H15 in Figure 10 indicates a good approximation of H with larger errors only
around the Hawaiian islands (where the geostrophic balance does not hold in the
first place).
MDT Geostrophic Flow

o o o o o o
180 W 150 W 120 W 180 W 150 W 120 W
70 3.5
o o
45 N 45 N
60 3
50 2.5
o o
30 N 30 N
2
cm/s
40
cm
o o
15 N 15 N 1.5
30
1
0o 20 0o
0.5
10
o o
15 S 15 S
Figure 8. The ‘true’ MDT H (left) and the corresponding scaled

geostrophic ocean flow velocity ξ → (ξ · ε3 )v(ξ) (right; colors indicate
the absolute values and arrows the orientation).
† data accessed via http://apdrc.soest.hawaii.edu/projects/DOT

MDT, Scale J=6 MDT, Difference Scale J=8 and Scale J=6
o o o o o o
180 W 150 W 120 W 180 W 150 W 120 W
65
o
45 N o 4
60 45 N
55
2
50
o o
30 N 30 N
45
0
cm
cm
40
o o
15 N 35 15 N
−2
30
25
0o 0
o
−4
20
15
15o S 15o S −6
180o W 150o W 120o W 180 W
o o
150 W
o
120 W
70
3
o o
45 N 45 N
60
2
50 1
o o
30 N 30 N
0
cm
40
cm
o o
15 N 15 N −1
30
−2
o o
0 0
20
−3
o 10 o −4
15 S 15 S
o o o o o o
180 W 150 W 120 W 180 W 150 W 120 W
70
o o
45 N 45 N
3
60
2
50
o o
30 N 30 N
1
40
cm
cm
15o N o
15 N
0
30
o −1
0 20 0o
10 −2
15o S o
15 S
MDT, Scale J=15
o o o
180 W 150 W 120 W
70
o
45 N
60
o
50
30 N
40
cm
15o N
30
Figure 9. MDT reconstructions HJ at
scales J = 6, 8, 10, 15 (left) and the dif-
0
o 20
ferences H8 −H6 , H10 −H8 , H15 −H10 be-
10
tween the reconstructions at these scales
15o S (right).
848 C. Gerhards
MDT MDT, Error Scale J=15

o o o o o o
180 W 150 W 120 W 180 W 150 W 120 W
0.5
70
o o 0.4
45 N 45 N
60 0.3
0.2
o 50 o
30 N 30 N
0.1
40
cm
cm
0
15o N 15o N −0.1

30
−0.2
0o 20 0o −0.3
−0.4
10
15o S 15o S −0.5
Figure 10. The true MDT H (left) and the reconstruction error H −
H15 (right).
5.3. Point vortex motion

Vorticity describes the rotational motion of a fluid. In the ocean, for horizontal
flows v which extend over regions ΓR ⊂ ΩR at spatial scales of several tens or
hundreds of kilometers, the following relation for the vorticity ω holds true:
ω(Rξ) = L∗ · v(Rξ), ξ ∈ Γ. (113)
The incompressible horizontal flow v itself can be represented by a stream function
Ψ via v = L∗ Ψ, so that we obtain
ω(Rξ) = Δ∗ Ψ(Rξ), ξ ∈ Γ. (114)
The geostrophic flow from Section 5.2 is an example for such a current. For more
geophysical background on vorticity, the reader is again referred to [42, 44].
A single point vortex at location Rη on the sphere is associated with a vor-
1
ticity ω(Rξ) = ω̄ δ(1 − ξ · η) − 4πR of strength ω̄ ∈ R (by δ we denote the
Dirac distribution) and a corresponding stream function Ψ(Rξ) = R ω̄
G(Δ∗ ; ξ · η),
ξ ∈ Ω \ {η}. If we consider a point vortex at location Rη in a subdomain ΓR ⊂ ΩR
that produces no flow across the boundary ∂ΓR (e.g., a coastline), the vortic-
ity would be ω(Rξ) = ω̄ (δ(1 − ξ · η)) and the corresponding stream function
Ψ(Rξ) = R ω̄
GD (Δ∗ ; η, ξ), ξ ∈ Γ \ {η}. In [20, 34], this motivated solving the model
problem
Δ∗ Ψ̃(Rξ) = 0, ξ ∈ Γ, (115)

N
ω̄i ω̄i
Ψ̃− (Rξ) = G(Δ∗ ; ξ · ηi ) − ¯
ln(1 − ξ · ξ), ξ ∈ ∂Γ, (116)
i=1
R 4πR
for a fixed ξ¯ ∈ Γc and point vortices of strengths ω̄i located at Rηi ∈ ΓR , i =

N
1, . . . , N . The actual stream function is then given by Ψ(Rξ) = i=1 ω̄Ri G(Δ∗ ; ξ ·
N ω̄i ∗
ηi ) − 4πR ω̄i
ln(1 − ξ · ξ̄) − Ψ̃(Rξ) = i=1 R GD (Δ ; ηi , ξ), ξ ∈ Γ \ {η1 , . . . , ηN }.
More details on point vortex motion on the entire sphere (and more general closed
Figure 11. The ‘true’ potential Ψ (left) and the corresponding hor-
izontal flow velocity v (right; colors indicate the absolute values and
arrows the orientation).
manifolds) can be found, e.g., in [6, 7, 31, 33], and details on point vortex motion
on subdomains of the sphere with impenetrable boundaries, e.g., in [24, 32].
In this section, we focus on the model problem (115), (116). Opposed to
[20, 34], where boundary integral methods have been used, we want to solve it
by the method of fundamental solutions based on the results of Section 4. More
precisely, we choose Γ to be a spherical cap in the Northern hemisphere: Γ =
Γρ (ζ) with center ζ = (0, 0, 1)T and radius ρ = 0.9. For simplicity, we set R =
1. The centers ηi ∈ Γρ (ζ), i = 1, . . . , N , and the corresponding strengths ω̄i of
the point vortices are chosen randomly. The point ξ¯ ∈ (Γρ (ζ))c from (116) is
set to ξ¯ = (0, 0, −1)T . Furthermore, we assume the boundary data (116) to be
given in equidistantly distributed points ξi ∈ ∂Γρ (ζ), i = 1, . . . , M . Eventually, we
(mod)
interpolate the data by the functions Gk , k = 0, . . . , M − 1, from Theorem 4.1
and Remark 4.2, i.e.,
(mod) 1 1
Gk (ξ) = ln(1 − ξ · ξ̄k ) − ln(1 − ξ · ξ̄), k = 1, . . . , M − 1, (117)
4π 4π
where the center points ξ¯k , k = 1, . . . , M − 1, are chosen to be equidistantly
distributed on ∂Γρ̄ (ζ), for a radius ρ̄ > ρ. The resulting approximation ΨM,N,ρ̄ of
Ψ in Γρ (ζ) is given by

N
1
ΨM,N,ρ̄(ξ) = ω̄i G(Δ∗ ; ξ · ηi ) − ln(1 − ξ · ξ̄) − Ψ̃M,N,ρ̄ (ξ), ξ ∈ Γρ (ζ),
i=1
4π
(118)

M−1
(mod)
Ψ̃M,N,ρ̄(ξ) = ak Gk (ξ), ξ ∈ Γρ (ζ), (119)
k=0
where the coefficients ak , k = 0, . . . , M − 1, are obtained from the approximate
solution of (115), (116) via interpolation of the boundary data. The resulting
ΨM,N,ρ̄ and the corresponding reconstruction errors are plotted in Figure 12 for
850 C. Gerhards
different settings of M, ρ̄ (we fix the number of point vortices to N = 40). The

actual potential Ψ and the underlying horizontal flow v are shown in Figure 11.
We restrict our test example to a spherical cap Γ = Γρ (ζ) because we then know
an explicit representation of Ψ via the Dirichlet Green function GD (Δ∗ ; ·, ·) from
Section 4 and can compute the reconstruction errors. However, the approach can
be easily adapted to more complex geometries of Γ.
The results in Figure 12 show a good performance for the test example of
this easy to implement technique. The influence of the parameter ρ̄ turns out to
be fairly harmless for M = 1000 source points. A significant deterioration of the
reconstruction error does not occur before ρ̄ = 0.968 (cf. Figure 12). However,
in general, the method of fundamental solutions can be rather sensitive to the
choice of the involved parameters, in particular of the source points ξ¯k and the
collocation points ξk . Furthermore, it can be advantageous to use a regularized
least squares method instead of a simple interpolation. An overview on the method
of fundamental solutions in general and its recent developments can be found, e.g.,
in [4, 9]. Latter, however, treat only the Euclidean setting. The current section is
meant as a basic illustration of the method of fundamental solution for boundary
value problems intrinsic on the sphere based on the techniques described in this
chapter.
References
[1] O. Amm. Elementary currents for ionospheric fields. J. Geomag. Geoelectr., 49:947–
955, 1997.
[2] G. Backus, R. Parker, and C. Constable. Foundations of Geomagnetism. Cambridge
University Press, 1996.
[3] L. Baratchart, D.P. Hardin, E.A. Lima, E.B. Saff, and B.P. Weiss. Characterizing
kernels of operators related to thin plate magnetizations via generalizations of Hodge
decompositions. Inverse Problems, 29:015004, 2013.
[4] C.S. Chen, A. Karageorghis, and Y.S. Smyrlis. The Method of Fundamental Solutions
– A Meshless Method. Dynamic Publishers, Inc., 2008.
[5] R. Comblen, S. Legrand, E. Deleersnijdera, and V. Legata. A finite element method
for solving the shallow water equations on the sphere. Ocean Modelling, 28:12–23,
2009.
[6] D.G. Dritschel. Contour dynamics/surgery on the sphere. J. Comp. Phys., 78:477–
483, 1988.
[7] D.G. Dritschel and S. Boatto. The motion of point vortices on closed surfaces. Proc.
R. Soc. A, 471:20140890, 2015.
[8] R.L. Duduchava, D. Mitrea, and M. Mitrea. Differential operators and boundary
value problems on hypersurfaces. Math. Nachr., 279:996–1023, 2006.
[9] G. Fairweather and A. Karageorghis. The method of fundamental solutions for el-
liptic boundary value problems. Adv. Comp. Math., 9:69–95, 1998.
Reconstruted Potential, M=1000, rho=0.900005 Reconstruction Error for Potential, M=1000, rho=0.900005
0.02
180oo W
180 E 5 180oo W
180 E
150o W 150o E 150o W 150o E
o 4 o
30 N 30 N 0.01
120 W
o o
120 E 3 120 W
o o
120 E
o
60 N
o
60 N 0
2
m2/s
m2/s
o
W 90o N 90o E 90o W 90o N 90o E −0.01
−0.02
−1
o o o o
W 60 E 60 W 60 E
−2
−0.03
30o W 30o E −3 30o W 30o E

o o
0 0
−4 −0.04
0.02
180oo W
180 E 5 180oo W
180 E
o o
150 W 150 E 150o W 150o E
o 4 o
30 N 30 N 0.01
120o W 120o E 3 120o W 120o E

o
60 N
o
60 N 0
2
1
m2/s
m2/s
90o W 90o N 90o E 90o W 90o N 90o E −0.01
−0.02
−1
o o o o
60 W 60 E 60 W 60 E
−2
−0.03
30o W 30o E −3 30o W 30o E

o o
0 0
−4 −0.04
180oo W
180 E 5 180oo W
180 E
o o 0.08
150 W 150 E 150o W 150o E
o 4 o
30 N 30 N
0.07
120o W o 3
120 E 120o W 120o E 0.06
o o
60 N 60 N
2 0.05
0.04
1
m2/s
m2/s
90o W 90o N 90o E 90o W 90o N 90o E
0.03
0
0.02
−1
o o o o 0.01
60 W 60 E 60 W 60 E
−2
0
30o W 30o E −3 30o W 30o E −0.01

o o
0 0
−4 −0.02
Reconstruted Potential, M=30000, rho=0.900005 Reconstruction Error for Potential, M=30000, rho=0.900005 x 10
−3
1
180oo W
180 E 5 180 W
180 E
oo
o o
150 W 150 E 150o W 150o E
o
30 N 4 30o N 0.5
120o W 120o E 3
120o W 120o E
o 0
60 N 60o N
2
1
m2/s
m2/s
90o W 90o N 90o E o

90 W
o
90 N
o
90 E −0.5
−1 −1
o o o o
60 W 60 E 60 W 60 E
−2
−1.5
30o W 30o E −3 o
30 W
o
30 E
0o 0
o
−4 −2
Figure 12. The reconstructed potential ΨM,N,ρ̄ (left) and the corre-
sponding reconstruction errors Ψ − ΨM,N,ρ̄ (right) for M = 1000, 30000
and ρ̄ = 0.900005, 0.905, 0.968.
852 C. Gerhards
[10] T. Fehlinger, W. Freeden, C. Mayer, and M. Schreiner. On the local multiscale deter-
mination of the earths disturbing potential from discrete deflections of the vertical.
Comp. Geosc., 12:473–490, 2009.
[11] T. Fehlinger, W.S. Freeden, W. Freeden, C. Mayer, D. Michel, and M. Schreiner.
Local modelling of sea surface topography from (geostrophic) ocean flow. ZAMM,
87:775–791, 2007.
[12] M.J. Fengler and W. Freeden. A nonlinear galerkin scheme involving vector and
tensor spherical harmonics for solving the incompressible Navier–Stokes equation on
the sphere. SIAM J. Sci. Comp., 27:967–994, 2005.
[13] W. Freeden. Geomathematics: Its role, its aim, and its potential. In W. Freeden,
M.Z. Nashed, and T. Sonar, editors, Handbook of Geomathematics. Springer, 2nd
edition, 2015.
[14] W. Freeden, T. Fehlinger, M. Klug, D. Mathar, and K. Wolf. Classical globally
reflected gravity field determination in modern locally oriented multiscale framework.
J. Geod., 83:1171–1191, 2009.
[15] W. Freeden and C. Gerhards. Geomathematically Oriented Potential Theory. Pure
and Applied Mathematics. Chapman & Hall/CRC, 2012.
[16] W. Freeden, D. Michel, and V. Michel. Local multiscale approximations of geo-
strophic flow: Theoretical background and aspects of scientific computing. Marine
Geodesy, 28:313–329, 2005.
[17] W. Freeden and V. Michel. Multiscale Potential Theory (with Applications to
Geoscience). Birkhäuser, 2004.
[18] W. Freeden and M. Schreiner. Spherical Functions of Mathematical Geosciences.
Springer, 2009.
[19] M. Ganesh, Q.T. LeGia, and I.H. Sloan. A pseudospectral quadrature method for
Navier–Stokes equations on rotating spheres. Math. Comp., 80:1397–1430, 2011.
[20] S. Gemmrich, N. Nigam, and O. Steinbach. Boundary integral equations for the
Laplace–Beltrami operator. In H. Munthe-Kaas and B. Owren, editors, Mathematics
and Computation, a Contemporary View. Proceedings of the Abel Symposium 2006.
Springer, 2008.
[21] C. Gerhards. Spherical decompositions in a global and local framework: Theory and
an application to geomagnetic modeling. Int. J. Geomath., 1:205–256, 2011.
[22] C. Gerhards. Locally supported wavelets for the separation of spherical vector fields
with respect to their sources. Int. J. Wavel. Multires. Inf. Process., 10:1250034, 2012.
[23] C. Gerhards. Multiscale modeling of the geomagnetic field and ionospheric currents.
In W. Freeden, M.Z. Nashed, and T. Sonar, editors, Handbook of Geomathematics.
Springer, 2nd edition, 2015.
[24] E. Gutkin and K.P. Newton. The method of images and Green’s function for spherical
domains. J. Phys. A: Math. Gen., 37:11989–12003, 2004.
[25] W.A. Heiskanen and H. Moritz. Physical Geodesy. W.H. Freeman and Company,
1967.
[26] L.L. Helms. Introduction to Potential Theory. Wiley-Interscience, 1969.
[27] K. Hesse and R.S. Womersley. Numerical integration with polynomial exactness over
a spherical cap. Adv. Comp. Math., 36:451–483, 2012.
[28] B. Hofmann-Wellenhof and H. Moritz. Physical Geodesy. Springer, 2nd edition, 2005.
[29] A.A. Il’in. The Navier–Stokes and Euler equations on two-dimensional closed mani-
folds. Math. USSR Sb., 69:559–579, 1991.
[30] O.D. Kellogg. Foundations of Potential Theory. Springer, reprint edition, 1967.
[31] R. Kidambi and K.P. Newton. Motion of three point vortices on a sphere. Physica
D, 116:143–175, 1998.
[32] R. Kidambi and K.P. Newton. Point vortex motion on a sphere with solid boundaries.
Phys. Fluids, 12:581–588, 2000.
[33] Y. Kimura and H. Okamoto. Vortex motion on a sphere. J. Phys. Soc. Jpn., 56:4203–
4206, 1987.
[34] M.C.A. Kropinski and N. Nigam. Fast integral equation methods for the Laplace–
Beltrami equation on the sphere. Adv. Comp. Math., 40:577–596, 2014.
[35] N. Maximenko, P. Niiler, M.-H. Rio, L.C. Melnichenko, D. Chambers, V. Zlotnicki,
and B. Galperin. Mean dynamic topography of the ocean derived from satellite and
drifting buoy data using three different techniques. J. Atm. Ocean. Tech., 26:1910–
1919, 2009.
[36] C. Mayer and T. Maier. Separating inner and outer Earth’s magnetic field from
CHAMP satellite measurements by means of vector scaling functions and wavelets.
Geophys. J. Int., 167:1188–1203, 2006.
[37] M. Mitrea and M. Taylor. Boundary layer methods for Lipschitz domains in Rie-
mannian manifolds. J. Func. Anal., 163:181–251, 1999.
[38] M. Mitrea and M. Taylor. Potential theory on Lipschitz domains in Riemannian
manifolds: Sobolev–Besov space results and the Poisson problem. J. Func. Anal.,
176:1–79, 2000.
[39] N. Olsen. Ionospheric F-region currents at middle and low latitudes estimated from
MAGSAT data. J. Geophys. Res., 102:4564–4576, 1997.
[40] N. Olsen, K.-H. Glassmeier, and X. Jia. Separation of the magnetic field into external
and internal parts. Space Sci. Rev., 152:135–157, 2010.
[41] N.K. Pavlis, S.A. Holmes, S.C. Kenyon, and J.K. Factor. The development and
evaluation of the Earth Gravitational Model 2008 (EGM2008). J. Geophys. Res.,
117:B04406, 2012.
[42] J. Pedlosky. Geophysical Fluid Dynamics. Springer, 1979.
[43] L.M. Polvani and D.G. Dritschel. Wave and vortex dynamics on the surface of a
sphere. J. Fluid Mech., 255:35–64, 1993.
[44] R.H. Stewart. Introduction to physical oceanography. Online.
[45] J. Wermer. Potential Theory. Springer, 1974.
Christian Gerhards
Computational Science Center
University of Vienna
A-1090 Vienna, Autria
e-mail: christian.gerhards@univie.ac.at

Joint Inversion of Multiple Observations

Christian Gerhards, Sergiy Pereverzyev Jr., and Pavlo Tkachenko
Abstract. Joint inversion becomes increasingly important with the availability

of various types of measurements related to the same quantity. Questions
arising in this context are how to combine the different data sets in the first
place and, secondly, how to choose the multiple parameters that naturally
occur in such a combination. This chapter discusses some recently proposed
techniques addressing these issues. Additionally, we distinguish the two cases
when all underlying problems are ill posed (e.g., satellite data only) and when
some of them are not ill posed (e.g., satellite data is complemented by data at
the Earth surface). Theoretical discussions of the topics above are presented
as well as numerical experiments with different settings of simulated data.
Keywords. Aggregation methods, combination of satellite and ground models.
1. Introduction
In various applications, especially in geoscience, one is provided with several data
sets of observations of the same quantity of interest. These data sets may contain
observations based on different physical principles (e.g., satellite-to-satellite track-
ing (SST) data as in the case of the GRACE satellite mission [29] and satellite
gravity gradiometry (SGG) data as in the case of the GOCE satellite mission [27])
or observations of the same quantity at different locations (e.g., global magnetic
field satellite measurements as by the CHAMP and Swarm missions [11, 23] and
local magnetic field measurements at or near the Earth’s surface). Depending on
the kind of data, either the physical measurement principles have to be taken
into account for the combination or properties originating from the location of the
observations or possibly both.
The ill-posed problem of downward continuation of potential field data is
an ever-present topic with gravitational and magnetic satellite data. It has been
studied intensively, e.g., in [4, 18, 20, 21, 44, 46, 53, 56, 67, 68]. However, those
studies treat downward continuation only for a single set of measurements. With
multiple satellite observation models, such as SST and SGG, each providing ap-
proximations of the same quantity of interest, one is left with the choice of which to
trust. A more advanced question is what to do with a less trustable approximation
856 C. Gerhards, S. Pereverzyev Jr., and P. Tkachenko
or, in other words, whether an approximation that involves all available observa-
tions may actually serve as an effective way to reduce uncertainties in independent
models. This question was and is discussed quite intensively in the geophysical lit-
erature, where the term ‘joint inversion’ was introduced by the authors of [69] for
methods which provide a solution of various types of observation equations that
are inverted simultaneously. A short overview about the application of joint inver-
sion in geophysics may be found in [30]. To distinguish inversion methods based
on different data combinations, researchers have introduced different names, e.g.,
aggregation [51], which was also used in the context of statistical regression analy-
sis [37, 38]. Regardless of their names, what the above-mentioned approaches have
in common is that they induce stability by simultaneously utilizing different types
of indirect observations of the same phenomenon, which essentially limits the size
of the class of possible solutions [2].
When additional data at or near the Earth’s surface comes into play, not
only ill-posed downward continuation influences the quality of the models but
also the restricted local availability of data. A vast amount of research has been
addressed to localization on the sphere. This involves, e.g., the development of
spherical splines (e.g., [17, 61]), spherical cap harmonics (e.g., [31, 65]), Slepian
functions (e.g., [58, 59, 62, 63]), as well as spherical multiscale methods (e.g.,
[7, 10, 14, 19, 22, 24, 34, 35, 39, 50, 52, 60]). However, methods that simultaneously
address ill-posedness and localization are rather rare (e.g., [25, 59]). So, the task
remains to find adequate ways of combining these different types of data and to
appropriately choose the involved parameters.
We split this chapter into three sections that we have labeled according to
what we believe their area of application might be rather than what mathematical
methods have been used. Each of these brief sections is written in a way that
should make it understandable without reading the other two.
Section 2 focuses on the combination of different models obtained from dif-
ferent ill-posed problems (e.g., the computation of the gravity potential at the
Earth’s surface from SST or SGG satellite measurements). It does not aim at
obtaining the model directly from the supplied data but rather at aggregating dif-
ferent models into a more stable, trustworthy one. Joint inversion naturally leads
to multi-parameter regularization (see, e.g., [40] for the regularization of geopoten-
tial determination from different types of satellite observations). At this point it is
important to note that one should distinguish between multi-parameter schemes,
where the regularization parameters penalize the norms of the approximant in dif-
ferent spaces and schemes where the parameters weigh the data misfits in different
observation spaces. In the former schemes an observation space is fixed, and by
changing the regularization parameters we try to find a suitable norm for the so-
lution space, while in the latter schemes the situation is opposite: by changing the
parameters we try to construct a common observation space as a weighted direct
sum of given spaces. The choice of the regularization parameters for the former
schemes has been extensively discussed in the literature. A few selected references
are [8, 9, 13, 36, 45]. As to the latter schemes (schemes with a fixed solution space),
Joint Inversion of Multiple Observations 857
to the best of our knowledge, we can indicate only the paper [40], where a heuristic
parameter choice rule is discussed, and the papers [15, 41], where the parameter
choice is considered as a learning problem under the assumption that for similar
inverse problems a suitable parameter choice has been known. It is clear that such
approaches can be used only for particular classes of problems. In this section, we
give a fairly general method that can be regarded as a parameter choice based on
linear aggregation.
Section 3 provides a method on how to improve a global potential field model
based on satellite data with additional discrete local data at the Earth’s surface.
The choice of the involved parameters is based on a method similar to that of
Section 2.
Section 4 has a similar setup as Section 3. But while in Section 3 the goal is
to obtain a global model that is refined by local data, the goal in Section 4 is to
obtain a local model at the Earth’s surface that is refined by global satellite data
(and which eventually reveals a better reconstruction of the coarse features than
a model purely based on local data). The presented method for the combination
of the two data sets is based on the construction of convolution kernels that pay
tribute to localization at the Earth’s surface as well as to the ill-posedness of down-
ward continuation (cf. [25]). Concerning the choice of the involved regularization
parameters, the linear aggregation from Section 2 could generally be applied but
the obtained error estimates do not hold true anymore when using the well-posed
local ground data as a reference. The parameter choice method that we provide
here is tailored for problems where direct measurements of the modeled quantity
are available (cf. [26]).
2. Global combination of satellite models

In this section, we assume to have Hilbert spaces X , Yi and noisy observations
yiεi ∈ Yi , i = 1, . . . , m, that are connected to the quantity of interest x ∈ X via
yiεi = Ai x + ei , i = 1, . . . , m. (1)
The functions ei represent the additive noise contained in the data and εi > 0 the
noise level:
Ai x − yiεi Yi εi , i = 1, . . . , m, (2)
The operators Ai : X → Yi are assumed to be compact, linear, and injective. The
exact noise-free solution x of (1) will be denoted by x† . Classical choices for the
situation above are X = L2 (Ωr ) and Yi = L2 (Ωρi ), where Ωr = {ξ ∈ R3 : |x| = r}
denotes a spherical Earth surface of radius r and Ωρi spherical satellite orbits with
radii ρi > r. Upward continuation operators can, e.g., have the form

Ai x = Ki (·, η)x(η)dS(η), i = 1, 2, (3)
Ωr
with kernels

1 ∂ ρ21 − r2
K1 (ξ, η) = − , ξ ∈ Ωρ1 , η ∈ Ωr ,
4πr ∂ρ1 (ρ21 + r2 − 2ξ · η)3/2
for satellite-to-satellite tracking (SST) or

1 ∂2 ρ22 − r2
K2 (ξ, η) = , ξ ∈ Ωρ2 , η ∈ Ωr ,
4πr ∂ρ22 (ρ22 + r2 − 2ξ · η)3/2
for satellite gravity gradiometry (SGG). This setting will be used in the numerical
examples in Section 2.2 later on.
The joint inversion of the multiple observation models (1), (2) can be formu-
lated as the minimization problem

m
min λi Ai x − yiεi 2Yi + x 2X . (4)
x∈X
i=1
The regularization parameters λi > 0 are introduced to adjust the contribu-

tions of the data misfit from the different observations. It is convenient to rewrite
the objective
(m functional of (4) in a compact form by introducing the direct sum
Yλ = i=1 Yi,λ i mof the observation spaces. The inner product on Yλ is defined
as y, ȳYλ = i=1 λi yi , ȳi Yi , for y = (yi )i=1,...,m , ȳ = (ȳi )i=1,...,m ∈ Yλ , and
λ = (λi )i=1,...,m . The norm · Yλ is defined according to this inner product. With
these notations at hand, the minimization (4) can be rewritten as
min Aλ x − y ε 2Yλ + x 2X , (5)

x∈X
where ε = (εi )i=1,...,m and y ε = (yiεi )i=1,...,m ∈ Yλ .

The operator Aλ = (Ai )i=1,...,m : X → Yλ is again compact, linear, and
injective. Representation (5) allows the classical Tikhonov–Phillips form of the
minimizer xελ of (4) and (5):
xελ = (I + A∗λ Aλ )−1 A∗λ y ε , (6)
where I : X → X is the identity operator and A∗λ : Yλ → X is the adjoint of Aλ .

A similar representation as in (5) was obtained in [40] by Bayesian reasoning.
It is also suggested in [40] to relate the values of the regularization parameters λi
with the observation noise levels (variances) εi as follows:
ε21
λi = λ1 , i = 1, . . . , m. (7)
ε2i
Note that this relation reduces the multi-parameter regularization (4) to a single-
parameter regularization since only λ1 needs to be chosen. The heuristic rule (7)
can be motivated from a bound for the noise propagation error. Specifically, we
have that
x0λ − xελ X = (I + A∗λ Aλ )−1 A∗λ (y − y ε ) X
(I + A∗λ Aλ )−1 A∗λ Yλ →X y − y ε Yλ
1
Aλ x† − y ε Yλ ,
2
where y 0 := (yi0 )i=1,...,m = (Ai x† )i=1,...,m ∈ Yλ denotes noise-free input data. It
follows from Assumption (2) that
m 12
0 ε 1 2
xλ − xλ X λi εi . (8)
2 i=1
The heuristics behind the rule (7) is now clear: The choice (7) equates all the
terms on the right-hand side of (8) and balances the data misfits against each
other. The final balance may be achieved by making a choice of the remaining
parameter λ = λ1 . The latter can be chosen by known single-parameter choice
rules such as the quasi-optimality criterion [66] (let us label this strategy by M1
for later reference). Within this strategy, we choose the value λ from
ΛqN = {λ̃k = λ̄q k : k = 0, 1, . . . , N },
for some fixed q > 1 and λ̄ > 0 such that

ε
xλ − xελq−1 = min xελ̃k − xελ̃k−1 : λ̃k , λ̃k−1 ∈ ΛN .
q
(9)
A multi-parameter version of the quasi-optimality (QO) criterion can be used
as an alternative to the single-parameter reduction (7). We denote this strategy
by M2 for later reference. For the sake of clarity we describe it here only for the
case of two parameters: for each fixed regularization parameter, say λ1 ∈ ΛqN , we
chose the value λ2 = λ2 (λ1 ) ∈ ΛqN according to the one-parameter QO criterion
(9) with respect to λ = λ2 . Then we repeat the strategy (9) for the set of pairs
(λ1 , λ2 (λ1 )) with respect to λ = λ1 .
2.1. The linear aggregation method
If a priori information like noise level of the measurements is given, then many
studies are available on the choice of the regularization parameters. If this is not
the case, one has to fall back on heuristic rules, e.g., quasi-optimality as mentioned
above. In this section, we propose to ‘aggregate’ various approximations of x† based
on such heuristic rules in order to obtain an improved approximation. More details
on the study of this scheme can be found in [12].
Let us assume for now that, from somewhere, we have obtained M different
approximations xεj , j = 1, . . . , M . The goal is to find an optimal linear combination
(‘aggregation’)

M
xεag,β opt = βjopt xεj , (10)
j=1
in the sense that β opt = (βjopt )j=1,...,M minimizes

- -
- -
- † M
-
-
min -x − ε-
βj xj - . (11)
β - -
j=1
X
The solution β opt of (11) is determined by solving a system of linear equations.
To state this system of linear equations, we need the Gram matrix

G = xεi , xεj X i,j=1,...,M and κ = xεj , x† X j=1,...,M .
Then
Gβ opt = κ. (12)
Remark 2.1. If the approximations xεj , j = 1, . . . , M , are linearly independent,
then the Gram matrix G is positive definite and invertible such that
G−1 RM →RM c1 ,
for some constant c1 > 0.
At the technical level, we are actually more concerned with the degree of the
linear correlation of xεj , j = 1, . . . , M , and the condition number of the Gram ma-
trix G. In principle, one may control this by excluding those members of the family
XM = {xεj : j = 1, . . . , M } that are close to be linearly dependent of others. It is
clear that their exclusion does not significantly change the value of (11). Gram–
Schmidt orthogonalization combined with a thresholding technique can achieve
this. The details can be found in [12].
The problem with the approach (10)–(12) is that the right-hand side κ of
(12) is not known since x† is not accessible. Summarizing our results so far, we
propose the following method to obtain an improved approximation of x† .
Method 2.2. From somewhere, we are given M different approximations xεj , j =
1, . . . , M of x† (based on the m observation equations (1), (2)). Beyond that, we as-
sume to have a particularly trustworthy approximation x̃ ∈ X that we use to define
κ̃ = xεj , x̃X j=1,...,M . The Gram matrix G is given by G = xεi , xεj X i,j=1,...,M .
The parameter β ∗ = (βj∗ )j=1,...,M is chosen as the solution to
Gβ ∗ = κ̃ (13)
†
and the corresponding aggregated approximation of x is given by

M
xεag,β ∗ = βj∗ xεj ,
j=1
It remains to study the influence of the approximation κ̃, i.e., how well does
xεag,β ∗ perform in comparison to the optimal aggregation xεag,β opt (based on the
exact κ)? To answer this, we regard the linear functional strategy as introduced
in [1, 3, 42]. The essence of this strategy is that one is not interested in completely
knowing x† but only in knowing some quantity derived from it, such as the value
of a bounded linear functional x̄, ·X applied to the solution x† . This quantity
can be estimated more accurately than x† itself. In order to allow a theoretically
rigorous study, we will choose a particular κ̃:
First, we pick the most trustworthy observation equation among those m ones
available from (1), (2). The choice might, e.g., be the problem with the operator
ε
Aj that is least ill posed or the one with the data yj j ∈ Yj that has the lowest
noise level εj . We abbreviate the chosen observation equation by
y ε = Ax + e, (14)
Ax − y Y ε.
ε
(15)
Furthermore, xεα denotes the Tikhonov–Phillips regularized solution to this prob-
lem, i.e.,
xεα = xεα (y ε ) = (αI + A∗ A)−1 A∗ y ε , (16)
for some parameter α > 0. From [32, 48, 49], we know that if ϕ : [0, A 2X →Y ]
→
[0, ∞) is an index function (i.e., ϕ is continuous, strictly increasing, and satisfies
t
ϕ(0) = 0), if ϕ(t) is non-decreasing, and if Aϕ,R denotes the source condition set
Aϕ,R = {x̄ ∈ X : x̄ = ϕ(A∗ A)v, v X R} , R > 0 fixed, (17)
then the convergence rate
sup sup x† − xεα (y ε ) X = O(ϕ(θ−1 (ε))), (18)
ε
x† ∈Aϕ,R y ∈Y;
Ax† −y ε Y ε
√
holds true for the choice α = θ−1 (ε), with θ(t) = ϕ(t) t. Note that O(ϕ(θ−1 (ε)))
is the best guaranteed order of accuracy for the reconstruction of x† ∈ Aϕ,R from
the observation (14) and (15). Now, for each of the M available approximations
xεj , j = 1, . . . , M , we can approximate the component κj of the vector κ by
κ̃j = xεj , xεαj X , (19)
where αj = α = θ−1 (ε). The approximation κ̃ of κ is then defined via κ̃ =
(κ̃j )j=1,...,M . In [12] it has been shown that under the conditions above we have

ε †
xj , x X − xεj , xεαj X = o(ϕ(θ−1 (ε))) (20)
However, in practice, the function ϕ describing the smoothness of the unknown

solution x† is not known. As a result, one cannot implement the a priori parame-
ter choice αj = α = θ−1 (ε). In principle, this difficulty may be resolved by use of
the so-called Lepskii-type balancing principle, introduced in [6, 28] in the context
of the linear functional strategy. But, the a posteriori parameter choice strategy
presented in those papers requires the knowledge of an index functions ψj describ-
ing the smoothness of xεj in terms of the source condition xεj ∈ Range(ψj (A∗ A)).
This requirement may be too restrictive in some applications. To overcome this
difficulty, we consider below a modification of the balancing principle to achieve
the error bounds (20) without requiring the knowledge of ϕ or ψj .
The balancing principle is well known in the literature (see, for example, [28,
43] Section 1.1.5, and [47]). Following the general theory, we formulate a version
of the balancing principle suitable for our context: We define a parameter set
ΛN = {ε2 = α̃1 < α̃2 < · · · < α̃N = 1}
and choose αj , j = 1, . . . , M , according to the balancing principle

αj = max α̃k ∈ ΛN : xεj , xα̃k X − xεj , xα̃l X
- -
4ε -(α̃l I + AA∗ )−1 Axεj -Y , l = 1, . . . , k − 1 . (21)
Obviously, the choice of αj as above does not require the knowledge of the
index function ϕ. These considerations culminate in the following theorem. A
detailed derivation can found in [12].
Theorem 2.3. Suppose that xεag,β opt is the optimal aggregator in the sense of (10)–
(12) and xεag,β ∗ is its approximation according to Method 2.2. The approximation κ̃
required in Method 2.2 shall be constructed from (16), (19), (21), using only y ε , A
from (14), (15), and xεj , j = 1, . . . , M . If x† ∈ Aϕ,R , where ϕ is an index function
t
such that the function ϕ(t) is increasing and takes zero value at the origin, then
x† − xεag,β ∗ X − x† − xεag,β opt X = o(ϕ(θ−1 (ε))).

Theorem 2.3 tells us that the coefficients βj∗ of the aggregator xεag,β ∗ can be
effectively obtained from the input data in such a way that the error x† −xεag,β ∗ X
differs from the optimal error x† − xεag,β opt X by a quantity of higher order than
the best guaranteed accuracy of the reconstruction of x† from the most trustable
observation (14), (15).
Remark 2.4. On the one hand, it is clear that involving more linearly independent
solutions in the aggregation helps in reducing the approximation error. On the
other hand, from the proof of Theorem 2.3 in [12], it can be seen that the √ coefficient
implicitly involved in o(ϕ(θ−1 (ε))) increases with M at least as fast as M . This
means that in order to be effective, the aggregator xεag,β ∗ should be built on the
basis of a modest number of approximations xεj . In our numerical illustrations, we
use M = 2.
Also note that in our analysis, the balancing principle (21) has been used
mainly for theoretical reasons. In the numerical experiments below, the regular-
ization parameters αj are chosen by the quasi-optimality criterion (compare (9))
such that

ε ε
xj , xαj X − xεj , xεαj q−1 X

= min xεj , xεα̃k X − xεj , xεα̃k−1 X : α̃k , α̃k−1 ∈ ΛqN . (22)
A practical advantage of the quasi-optimality criterion is that it does not require
the knowledge of the noise level ε. At the same time, as it was shown in [5], under
some assumptions on the noise spectral properties and on x† , the quasi-optimality

criterion allows optimal order error bounds.
The strategy presented in this subsection, i.e., Method 2.2 with κ̃ chosen
according to (16), (19), and (22), can be seen as the third method for the parameter
choice. We label this strategy by M3 for later use.
2.2. Numerical examples
We present in this section numerical experiments to demonstrate the efficiency of
the proposed aggregation methodand to compare it with other known methods in
the literature. All data are simulated in a way that they mimic the inputs of the
SST-problem and the SGG-problem described by the equations (1), (3).
It is well known (see, e.g., [18]) that the integral operators Ai defined by (3)
with the kernels Ki , i = 1, 2, act between the Hilbert spaces X = L2 (Ωr ) and
Yi = L2 (Ωρi ) of square-summable functions on the spheres Ωr , Ωρi , i = 1, 2, and
admit the singular value expansions
∞
2k+1
1 F · G
(i) ξ 1
Ai x(ξ) = ak Yk,l Yk,l ,x , ξ ∈ Ωρi , (23)
ρi ρi r r L2 (Ωr )
k=0 l=0
where the Yk,l denotes the orthonormal spherical harmonic of degree k and order
l on the unit sphere Ω1 , and
k k
(1) r k + 1 (2) r (k + 1)(k + 2)
ak = , ak = , k ∈ N0 . (24)
ρ1 ρ1 ρ2 ρ22
The solution x† to (1), (3) (i.e., the restriction of x to Ωr ) models the grav-
itational potential measured at the sphere Ωr , that is expected to belong to the
spherical Sobolev space Hs (Ωr ) with s = 32 (see, e.g., [64]), which means that
H(k, l) = 1r Yk,l r· , xL2 (Ωr ) asymptotically decay at least
its Fourier coefficients x
− 32
as (k + 1) . Therefore, to produce the data for our numerical experiments we
simulate the Fourier coefficients x H(k, l) of the solution x† via
3
H(k, l) = (k + 1)− 2 gk,l ,
x k ∈ N0 , l = 1, . . . , 2k + 1,
where gk,l are uniformly distributed random values in [−1, 1]. In view of (23),
· εi
the Fourier coefficients yI 1
i (k, l) = ρi Yk,l ρi , yi L2 (Ωρi ) of the noisy data yi are
εi εi
simulated as
yI
(i) (i)
εi
i (k, l) = a x
k H(k, l) + e , k ∈ N0 , l = 1, . . . , 2k + 1, i = 1, 2,
k,l
(i)
where ek,l is Gaussian white noise which roughly correspond to (2) with noise
level
ratio εε12 = 3. All random simulations are performed 500 times such that we
have
data for 1000 problems of the form (1), (3). Moreover, we take r = 6371km for
the radius of the Earth, and ρ1 = 6621km, ρ2 = 6771km. All spherical Fourier
coefficients are simulated up to the degree N = 300, which is in agreement with
the dimension of the existing models, such as Earth Gravity Model 96 (EGM96).
Thus, the set of simulated problems consists of 500 pairs of the SGG- and SST-
type problems (1), (3). In our experiments, each pair is inverted jointly by means
of Tikhonov–Phillips regularization (4), (5) performed in a direct weighted sum
of the observation spaces Yi = L2 (Ωρi ), i = 1, 2, and we use three methods for
choosing the regularization parameters (weights) λ1 , λ2 :
In the first method (i.e., M1), we relate them according to (7). Recall that
the data are simulated such that εε12 = 3. Therefore, we have λ2 = 9λ1 . Then
the parameter λ1 is chosen according to the standard quasi-optimality criterion
40+j
from the geometric sequence Λ30 = {10 8 : j = 0, . . . , 30}. As a result, for each
of 500 pairs of the simulated problems we apply M1 and obtain a regularized
approximation to the solution x† that will play the role of the approximant xε1 .
In the second method (i.e., M2), the parameters λ1 , λ2 are selected from Λ30
according to the multi-parameter version of the quasi-optimality criterion. In this
way, for each of 500 pairs of the simulated problems we apply M2 and obtain the
second approximant xε2 .
The third method (i.e., M3) consists in aggregating the approximants xε1 , xε2
according to the methodology described at the end of the previous subsection. In
our experiments the role of the most trustable observation equation (14) is played
by the equations of the SGG-type (23), i = 2, and we label the aggregation based
on them as M3(2). We choose these equations because the data for them are sim-
ulated with smaller noise intensity. Then the required regularization parameters
α1 , α2 are selected according to the quasi-optimality criterion (22) from the geo-
metric sequence Λ30 in such a way that α11 , α12 ∈ Λ30 . Note that in general, no
specific relation is required between the sets of possible values of the regularization
parameters λj and αj . In this test, we use the same set Λ30 for the sake of simplicity.
We have to admit that the decision, which model to select as the most
trustable one, may contribute to the performance of the aggregation method M3.
In our discussion, the ‘most trustable model’ might be either the least ill-posed
observation equation or the equation with the smallest noise level. If one has a
model with both of these features, then one can choose it. However, it may hap-
pen that the above features are not attributed to the same observation equation.
For example, in our numerical illustrations for (3), (23), (24), the SST-type equa-
tion (3), (23), i = 1, is contaminated by more intensive noise, but it is less ill-posed
than the SGG-type equation (3), (23), i = 2, which has been chosen by us as the
most trustable model. This can be seen from (24) if one compares the rates of
(1) (2)
the decrease of the singular values ak and ak as k → ∞: for the considered
(i)
values r = 6371 km, ρ1 = 6621 km, ρ2 = 6771 km both ak , i = 1, 2, decrease
(1) (2)
exponentially fast, but ak decreases slower than ak .
To illustrate what happens when an alternative model is chosen as the most
trustable one, we implement the aggregation method M3 on the base of the SST-
type equation (23), i = 1, and label it as M3(1). All other implementation details
are exactly as described for M3(2).
The performance of all four methods is compared in terms of the relative
errors x† − xεj X / x† X , j = 1, 2, and x† − xεag,β ∗ X / x† X . The results are
M3(2)
M3(1)
M2
M1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 2.1. Examples of a joint regularization of two observation mod-

els. Relative errors of the regularization by a reduction to a single reg-
ularization parameter (M1), the regularization with a multi-parameter
quasi-optimality criterion (M2), and the regularization by aggregation
(M3(1), M3(2)).
displayed in Figure 2.1, where the projection of each circle onto the horizontal
axis exhibits a value of the corresponding relative error of one of the methods M1,
M2, M3(1), and M3(2), in the joint inversion of one of 500 pairs of the simulated
problems. From this figure we can conclude that the aggregation by the linear
functional strategy can essentially improve the accuracy of the joint inversion
compared to M1 and M2. This conclusion is in agreement with our Theorem 2.3.
At the same time, Figure 2.1 also presents an evidence of the reliability of the
proposed approach. Indeed, in the considered case, even with the use of an alterna-
tive (i.e., suboptimal) trusted reference model, the aggregation, this time M3(1),
performs at least at the level of the best among the approximants M1 and M2.
3. Global combination of satellite and ground models

We now consider the situation when the global satellite data is complemented by
regional/local measurements at or near the Earth’s surface. The satellite data are
well suited for the reconstruction of large-scale structures, but it fails for spatially
localized features (due to the involved downward continuation). The opposite is
true for locally/regionally available ground data. They are well suited to capture
local phenomena but fails for global trends. Therefore, in order to obtain high-
resolution gravitational models, such as EGM2008 [55], or geomagnetic models,
such as NGDC-720 [54], it becomes necessary to combine both types of data. It
is the goal of this section to introduce a two-parameter approximation reflecting
such a combination.
Figure 3.1. Schematic Description of the given data situation. The

red areas indicate regions where data is available.
In the exterior Ωext

r = {x ∈ R3 : |x| > r} of the Earth, the gravity field
and the crustal magnetic field can be described by a harmonic potential x. From
satellite measurements we obtain data y1 on a spherical orbit Ωρ = {x ∈ R3 :
|x| = ρ} and from ground or near-ground measurements we obtain data y2 in a
subregion Γr of the spherical Earth surface Ωr of radius r < ρ (cf. Figure 3.1).
Of interest to us is the harmonic potential x, in particular, its restriction to
the Earth surface Ωr . It satisfies
Δx = 0, in Ωext
r , (25)
x = y1 , on Ωρ , (26)
x = y2 , on Γr . (27)
The knowledge of y1 on Ωρ already supplies information that can be used for ap-
proximating x. However, only a coarse structure of x can be reconstructed because
of the presence of noise and ill-posedness of the involved downward continuation.
Additional discrete measurements y2 of y2 in Γr may improve the situation. By
y1 , y2 , we typically denote functions while y1 ∈ RL1 , y2 ∈ RL2 denote discrete
evaluations of these functions at locations ξi ∈ Ωρ , i = 1, . . . , L1 , and ηi ∈ Γr ,
i = 1, . . . , L2 , respectively.
3.1. The inversion procedure
The goal is to approximate x globally on Ωr (or, since x is harmonic, in all of
Ωext
r ) from the situation (25)–(27). In general, we assume the following setting to
be given:
Setting 3.1.
(a) We have discrete measurements y2ε2 ∈ RL2 of y2ε2 on Γr and we know y1ε1 ∈
L2 (Ωρ ). Both quantities contain additive noise, i.e., y1ε1 = y1 + e1 and y2ε2 =
y2 + e2 with e1 L2 (Ωρ ) ε1 and e2 L2 (Ωr ) ε2 .
(b) The discretization operator D : L2 (Γr ) → RL 2

w maps a function x̄ ∈ L (Γr )
2
to the corresponding measurements x̄ = Dx̄ = x̄(η1 ), . . . , x̄(ηL2 ) ∈ RL 2

w . By
Rw we denote the space R equipped with the inner product
L2 L2

L2
x̄, ȳw = wi x̄i ȳi , x̄, ȳ ∈ RL2 , (28)
i=1
and the corresponding norm · w .

(c) Here, D and wi are such that
x̄, ȳL2 (Γr ) = Dx̄, Dȳw , for all x̄, ȳ ∈ PN , (29)
where PN is the space of spherical polynomials up to degree N .
(d) HK ⊂ L2 (Ωr ) is a reproducing kernel Hilbert space (RKHS) with reproducing
kernel of the form
N 2k+1
1 η
ξ
K(ξ, η) = dN (k) 2
Yk,l Yk,l , ξ, η ∈ Ωr , (30)
r r r
k=0 l=1
where dN = (dN (1), dN (2), . . . , dN (N )) is a monotone sequence. An inner

product on HK is defined by

N
1 2
2k+1 · 3 2 · 3
x̄, ȳHK = Yk,l , x̄ Yk,l , ȳ ,
r2 dN (k) r L2 (Ωr ) r L2 (Ωr )
k=0 l=1
x̄, ȳ ∈ HK .
To combine global data on Ωρ and local data on Γr for the approximation of
x in Ωext
r , we consider the following procedure:
Method 3.2. Let x satisfy (25)–(27) and let the conditions of Setting 3.1 be satis-
fied. Then, the approximation xεN,λ
1 ,ε2
1 ,λ2
of x is defined as the minimizer of
2 2 2
min λ1 Ax̄ − y1ε1 L2 (Ωρ ) + λ2 Dx̄ − Dy2ε2 w + x̄ HK , (31)
x̄∈HK
where λ1 , λ2 > 0 are the regularization parameters. A : HK → L2 (Ωext

r ) denotes
the upward continuation operator, i.e., Ax̄ is harmonic in Ωext
r .
Note that, when only global data are used, i.e., λ2 = 0, then (31) is reduced
to the Regularized Collocation (RC) method [53]. On the other hand, if λ1 = 0
(this means that we use only local data in Γr ) then (31) defines the solution after
denoising and is, in some sense, extrapolation. The solution of the minimization
problem (31) is given by the following theorem:
Theorem 3.3. Let Setting 3.1 hold true. Then the minimizer xεN,λ
1 ,ε2
1 ,λ2
of (31) on
Ωr has the form
N 2k+1
1 ·
xεN,λ
1 ,ε2
1 ,λ2
= Hελ11,ε
x 2
,λ2 (k, l) Yk,l ,
r r
k=0 l=1
where
F · G
1
Hελ11,ε
x 2
,λ2 (k, l) = xεN,λ
1 ,ε2
1 ,λ2
, Yk,l
r r L2 (Ωr )
λ1 ak yI
ε1 I ε2
1 (k, l) + λ2 dN (k)y2 (k, l)
= 2n+1
,
λ1 a2k + λ2 dN (k) n=0 m=1 ck,l,n,m + 1
N
F G
Iε1 ε1 1 ·
y1 (k, l) = y1 , Yk,j ,
ρ ρ L2 (Ωρ )

L2
1 η
yI
ε2 i ε2
2 (k, l) = wi Yk,l y2,i ,
i=1
r r
F · 1 · G
1
ck,l,n,m = Yn,m , Yk,j . (32)
r r r r L2 (Γr )
As mentioned earlier, y2,i ε2

denotes the evaluation of y2ε2 at the location ηi ∈ Γr ,
k
ε2
i.e., y2,i = y2ε2 (ηi ) = Dy2ε2 i , and ak = ρr represent the symbols of the upward
continuation operator A.
3.2. Numerical example

In a test example, we use the following reference potential
r k
30 2k+1
1

ξ
x(ξ) = H(k, l) Yk,l
x , ξ ∈ Ωext
r , (33)
|ξ| |ξ| |ξ|
k=0 l=1
as our true solution to (25)–(27). The Fourier coefficients of x are given by x H(k, l) =
3
(k + 12 )− 2 gk,l , where the gk,l are uniformly distributed random values from [−1, 1].
This means that x mimics the Sobolev smoothness s = 32 of the Earth’s gravita-
tional potential. The radius ρ is chosen such that ρr = 1.48 (if r = 6371km is the
mean Earth radius, then ρ = 9429km).
The function y1 on Ωρ is given as the restriction of x to Ωρ . Its noisy counter-
part y1ε1 is generated by adding Gaussian white noise of intensity ε1 = 0.05 at the
knots {ξi }i=1,...,L1 of a Gauss–Legendre cubature grid on the sphere Ωρ . For the
subregion Γr we choose the spherical cap Γr = Γr (ξ, ¯ r̄) = {η ∈ Ωr : 1 − ξ̄ · η < r̄}
r r
with spherical radius r̄ ∈ (0, 2) and the North Pole ξ¯ = (0, 0, r) as center. The
function y2 denotes the restriction of x to Γr . The noisy discrete values y2ε2 ∈ RL2
are simulated by adding a Gaussian white noise of intensity ε2 = 0.1 to the values
of y2 at the knots {ηi }i=1,...,L2 of a Gauss–Legendre cubature grid on the cap Γr .
For Gauss–Legendre grids, positive cubature weights are known that satisfy the
polynomial exactness condition required in (29).
We apply Method 3.2 to obtain approximations xεN,λ 1 ,ε1
1 ,λ2
of x in the setting
above. N = 30 is fixed and λ1 , λ2 are chosen from the set Λ200 = {106 × 0.95j :
j = 0, . . . , 200}. To assess the performance of the considered schemes, we measure
the relative error

2 1/2
-
-
-
- 30 2k+1 ε1 ,ε2
ε1 ,ε1
-x − xN,λ1 ,λ2 - 2 k=0 l=1 H
x (k, l) − H
xλ1 ,λ2 (k, l)
L (Ωr )
errrel =
x L2 (Ωr )
= 30 2k+1 1/2 ,
2
k=0 l=1 (H
x (k, l))
where xHελ11,ε2
,λ2 (k, l) are given by (32). We test method for different radii r̄ of the
spherical cap Γr = Γr (ξ, ¯ r̄). Spherical radius r̄ = 2 means that we regard the entire
sphere Γr = Ωr .
0.8
best choice of O1 and O2
0.7
O2 =0 and best choice of O1
0.6 O1 =0 and best choice of O2
0.5
0.4
0.3
0.2
0.1
0 r̄
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 3.2. Optimal relative errors errrel for the method (32) with
RC (λ2 = 0), denoising + extrapolation (λ1 = 0), and no additional
conditions on λ1 , λ2 .
The results are displayed in Figure 3.2, where the vertical axis represents the
global relative errors errrel . The relative errors are plotted for each of the three
methods: regularized collocation (RC) method corresponding to the case when
λ2 = 0, denoising + extrapolation (the case when λ1 = 0), and the two-parameter
scheme (32). For all methods, the ‘optimal’ regularization parameters are chosen,
i.e., the choice of the parameter is performed such that errrel is minimal (with
respect to the possible constraints λ1 = 0 or λ2 = 0).
Of course, the relative errors displayed in Figure 3.2 require the knowledge
of the ‘true’ solution x, which cannot be obtained in practice. However, Figure
3.2 shows that even for rather small subregions Γr (i.e., small spherical radii r̄)
of the sphere, where the additional noisy data y2ε2 are available, the use of these
additional data allows to improve the reconstruction compared to the standard
Tikhonov method (which only uses information y1ε1 at satellite altitude).
Next, we consider the more realistic case when the regularization parameters
are chosen by an a posteriori procedure, which does not require the knowledge of
the ‘true’ x. We can choose the regularization parameters by the analogue of the
well-known quasi-optimality criterion for the two-parameter case (see strategy M2
in Section 2.1).
However, our tests show that within this strategy we may only reach the
accuracy corresponding to local data denoising + extrapolation (λ1 = 0). In this
situation the use of the quasi-optimality criterion makes sense only for a large
amount of regional data (almost full coverage of the sphere Ωr ).
Potentially, one may perform global data smoothing with direct inversion.
This strategy has been presented in [57] for pointwise computations. However, it
does not use local data and its accuracy cannot be improved regardless the amount
of local information.
In this context, the idea is to use M procedures and aggregate them by a linear
2
! ε as presented in" Section 2 as Method 2.2, where X = L (Ωr )
functional strategy
and x̃ = xj̃ ∈ xj , j = 1, 2, . . . , M is a trustworthy approximant that we use to
ε
define κ̃ in (13).
We now apply this version of Method 2.2 to our previous test example. In
particular, we will aggregate M = 2 solutions xε1 , xε2 : the first one is given by the
two-parameter quasi-optimality (QO) criterion (strategy M2 in Section 2) applied
to the already obtained approximations xεN,λ 1 ,ε2
1 ,λ2
of x, the second one is the solu-
tion after global data smoothing with direct inversion. This choice leads to a linear
system of two equations with two unknowns β1 and β2 . For the approximation κ̃
of the right-hand side of (13) we use the second involved approximation (smooth-
ing+inversion). In Figure 3.3, we compare the relative error of the aggregated
0.8
best choice of O1 and O2
0.7 O2 =0 and best choice of O 1

O1 and O 2 according to QO
0.6 Smoothing on : U + direct inversion
Aggregation
0.5
0.4
0.3
0.2
0.1
0 r̄
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 3.3. Relative errors errrel for RC (λ2 = 0) with the best choice
of regularization parameters (based on the knowledge of x), for global
data smoothing + direct inversion, for QO, and for the aggregation of
the latter two.
∗ ∗
approximation xεag,β ∗ with the approximation xεN,λ
1 ,ε2
∗ ∗ (where λ1 , λ2 are chosen
1 ,λ2
ε1 ,ε2
according to QO) and with the actual best choice of xN,λopt ,λopt (based on the
1 2
knowledge of the ‘true’ x). One can conclude that in most situations the aggrega-
tion of the two considered regularization methods provides a better approximation
than each of the involved strategies separately. QO is only better than the aggre-
gated result if Γr is close to the entire sphere Ωr (i.e., if the spherical radius r̄
is close to 2). Moreover, it is instructive to observe that the aggregated solution

“ignores” the two-parameter QO for small spherical caps and uses it only for larger
spherical caps.
4. Local combination of satellite and ground models

We assume to be in the same situation as in Section 3 and have data y1 and y2
given on Ωρ and Γr , respectively (cf. Figure 3.1). Yet, opposed to Section 3, we are
not interested in a global model on Ωr but in a local model of on Γr . One might
say that Section 3 uses local data in Γr to refine a global model while now we use
global data on Ωρ to improve the coarse features of a local model.
A possibility to incorporate the local data set on Γr and the global data set
on Ωρ is to use spherical basis functions that are built upon satisfying certain lo-
calization constraints as well as ameliorating the ill-posed downward continuation
(see, e.g., [59]). However, in this section, we are aiming at designing convolution
kernels KN , K̃N that reflect the trade-off between localization and downward con-
tinuation (see, e.g., [25]). An approximation of x is then given by the convolution of
the data against these kernels. Opposed to the spherical basis function approach,
the convolution approach allows to treat the input data y1 and y2 consecutively
and indicates the influence of each one to the overall approximation of x. The
approximation does not hold globally but only in Γr . Therefore, we denote the
restriction of x to Γr by x† .
Just like in Section 3, we illustrate our approach for the Dirichlet problem
(25)–(27). The approximation of x† (which now denotes the restriction of x to
Γr ) from noisy data is denoted by xεN,λ1 ,ε2
1 ,λ2
, with unknown parameters N, λ1 , λ2 ,
and will be explained in Section 4.1. The approach is not limited to the problem
(25)–(27) but can also be applied to other problems where the singular values for
the underlying problems are known. Vectorial problems that are of relevance in
geomagnetic applications can be handled as well. For details on the theoretical
aspects of this section, the reader is referred to [25].
4.1. The inversion procedure
We first regard only the problem (25), (26). Then x can be approximated on Ωr by

xN = TN [y1 ] = KN (·, η)y1 (η)dS(η), (34)
Ωρ
with a kernel KN of the form

N 2k+1

1 ξ 1 η
KN (ξ, η) = dN (k) Yk,l Yk,l , ξ ∈ Ωr , η ∈ Ωρ , (35)
r r ρ ρ
k=0 l=1
and symbols dN (k) satisfying

ρ k
lim dN (k) = , (36)
N →∞ r
i.e., they converge towards the singular values of downward continuation. On the
other hand, regarding only the problem (25), (27), with y2 given on all of Ωr , x can
be approximated on Ωr by

xN = T̃N [y2 ] = K̃N (·, η)y2 (η)dS(η), (37)
Ωr
with a kernel K̃N of the form

N 2k+1
η
1 ξ 1
K̃N (ξ, η) = dÑ (k) Yk,l Yk,l , ξ ∈ Ω r , η ∈ Ωr , (38)
r r r r
k=0 l=1
and symbols dÑ (k) satisfying

lim dÑ (k) = 1. (39)
N →∞
However, according to (25)–(27), ground data y2 is only given in the subregion

Γr ⊂ Ωr . Thus, the symbols dÑ (k) do not only have to be chosen to satisfy (39)
but also such that K̃N (ξ, ·) is fairly well localized in Γr for every ξ ∈ Γr . To
compensate this localization constraint, we have to include an approximation of
the form (34), (35) that uses satellite data on Ωρ . Therefore, we use the following
approximation xN,λ1 ,λ2 of x† :
Method 4.1. Let x† denote the restriction of x to Γr , where x satisfies (25)–(27).

The approximation xN,λ1 ,λ2 of x† is defined by
xN,λ1 ,λ2 = TN [y1 ] + T̃N [y2 ]. (40)
TN and the corresponding kernel KN are given as in (34), (35). The operator T̃N
and the kernel K̃N are slightly modified in the sense

T̃N [y2 ] = K̃N (·, η)y2 (η)dS(η),
Γr
k
κN 2k+1
r 1 ξ 1 η
K̃N (ξ, η) = dÑ (k) − dN (k) Yk,l Yk,l ,
ρ r r r r
k=0 l=1
ξ ∈ Ω r , η ∈ Ωr ,
for some fixed κ > 1. The symbols dN (k), dÑ (k) of KN and K̃N , respectively, are
chosen to minimize the functional
κN

N k 2
2 r
˜
F (dN , dN ) = λ1 ˜
1 − dN (k) + λ1 1 − dN (n)
ρ
k=0 k=0

N
- -
+ λ2 dN (k)2 + -K̃N -2 2 . (41)
L ([−1,1−)
k=0
The parameters λ1 , λ2 > 0 are not known a priori. The radius > 0 is fixed in
advance and chosen to reflect the ’spherical radius’ of the set Γr , and
1−
- -
-K̃N -2 2 = |K̃N (t)|2 dt,
L ([−1,1−])
−1
where it has to be noted that the kernel K̃N (ξ, η) actually only depends on the
scalar product t = ξr · ηr , i.e.,
κN
k
2k + 1 r
K̃N (ξ, η) = K̃N (t) = dÑ (k) − dN (k) Pk (t),
4πr2 ρ
k=0
where Pk is the Legendre polynomial of degree k.
Remark 4.2. The functional F in (41) reflects the properties that we would like
to imply on the kernels KN and K̃N . The first term on the right-hand side of
(41) represents the overall approximation error (under the assumption that undis-
turbed global data is available on Ωρ as well as on Ωr ), the second term reflects
the approximation error under the assumption that only undisturbed data on Ωρ
is available. The third and fourth term act as penalty terms. While the third
term is meant to regularize the ill-posed downward continuation, the fourth term
penalizes the kernel K̃N if it is not localized in Γr (more precisely, it penalizes
the contributions of K̃N (·, η) in the exterior spherical cap Ωr \ Γr (η, ), where
Γr (η, ) = {ξ ∈ Ωr : 1 − ξr · ηr < } is a spherical cap with center η ∈ Γr and fixed
radius > 0 such that Γr (η, ) ⊂ Γr ). The parameters λ1 , λ2 allow to weigh the
approximation property against the regularization and the localization penalty.
Eventually, the obtained symbols dN (k) of KN reflect the regularization of
downward continuation while the dÑ (k) offer some control over the localization of
K̃N in Γr . An illustration of the optimized symbols is shown in Figure 4.1. One
can see that large parameters λ2 (relative to λ1 ) typically cause a strong damping
of dN (k) (i.e., satellite data y1 is damped and ground data y2 has more influence)
while small parameters λ2 (again, relative to λ1 ) are in favour of dN (k) (i.e.,
satellite data y1 has more influence). The question whether a stronger influence of
satellite data or a stronger influence of ground data is adequate depends on the
noise levels ε1 , ε2 . However, the noise levels are typically not known, so that other
means of determining λ1 , λ2 become necessary.
The minimizer of (41) can easily be computed by solving the system of linear
equations
Md = λ,
N
where d = dN (0), dN (1) ρr , . . . , dN (N ) ρr , dÑ (0), . . . , dÑ (*κN +)) ∈ RN +κN +2 ,
λ = (λ1 , . . . , λ1 ) ∈ RN +κN +2 , and

D1 + P1 −P2
M= .
−P3 D2 + P4
Figure 4.1. The symbols dN (k) (left) and dÑ (k) (right) for different
choices of λ1 , λ2 .
The diagonal matrices D1 , D2 are given by

2k
D1 = diag λ2 + λ1 , D2 = diag λ1 k=0,...,κN ,
r k=0,...,N

whereas P1 ,. . . , P4 are submatrices of the Gram matrix Pk,l k,l=0,...,κN . More
precisely,

P1 = Pk,l k,l=0,...,N , P2 = Pk,l k=0,...,N ; ,
l=0,...,κN

P3 = Pk,l k=0,...,κN ; , P4 = Pk,l k,l=0,...,κN .
l=0,...,N
with
1−
(2k + 1)(2l + 1)
Pk,l = Pk (t)Pl (t)dt.
2 −1
Under appropriate conditions on λ1 , λ2 , it can be shown that xN,λ1 ,λ2 converges
to x† as N → ∞. This convergence also holds true for xεN,λ 1 ,ε2
1 ,λ2
(the approximation
based on noisy input data y1 , y2 ) as ε1 , ε2 → 0 if an appropriate connection of
ε1 ε2
ε1 , ε2 , and N is assumed. Details can be found in [25].

If only discrete data y1ε1 , y2ε2 is available, then the evaluation of the integrals
ε1 ,ε2
in xN,λ1 ,λ2 has to be done numerically. In the examples later on we use the cubature
rules [16, 33].
4.2. A parameter choice method for direct observations
Let us now assume that we have obtained a set of various approximations
{xεN,λ
1 ,ε2
}
1 ,λ2 (N,λ1 ,λ2 )∈Λ
with Λ ⊂ N × R+ × R+ denoting the finite set of param-
eters. The standard question in such a situation is which xεN,λ1 ,ε2
1 ,λ2
to choose as a
‘good’ approximation of x† . Typically, we do not know the noise levels ε1 , ε2 and
obviously not the true x† . But we have the direct discrete observations y2ε2 of x† in
Γr as a reference. Since those are the non-ill-posed contributions to our problem,

a natural choice would be to choose N ∗ , λ∗1 , λ∗2 as the minimizers of
min xεN,λ
1 ,ε2
1 ,λ2
− y2ε2 w , (42)
(N,λ1 ,λ2 )∈Λ
1 ,ε2
where xεN,λ
1 ,ε2
1 ,λ2
= xεN,λ 1 ,λ2
(ηi ) i=1,...,L2 ∈ RL2 represents the vector of the evalu-
ations of xεN,λ
1 ,ε2
1 ,λ2
at the locations ηi ∈ Γr , i = 1, . . . , L2 . However, although it is
easy to find a minimizer of (42), it is not clear how to derive an adequate estimate
of the error x† − xεN1∗,ε,λ2 ∗ ,λ∗ L2 (Γr ) . Therefore, instead of (42), we use a modified
1 2
procedure to choose N ∗ , λ∗1 , λ∗2 that allows such an error estimate (cf. [26]).
We begin by formulating a more general setting than in the previous para-
graph:
Setting 4.3.
(a) Let Γr ⊂ Ωr be a subdomain where discrete measurements y ∈ RL of the
underlying quantity x are available. Let D : L2 (Γr ) → RL
w be the discretization
operator of Setting 3.1 and assume that it satisfies condition (29).
(b) The measurements y in Γr may be blurred by additive noise (i.e., y ε = y + e
for some e ∈ L2 (Γr )) and we assume that
- † -
-Dx − Dy ε - ε.
w
(c) We assume that from somewhere, a set {xεj }j=1,2,...,M of approximations of
x† is available and that all these approximations belong to the space PκN of
spherical polynomials up to degree *κN +.
Remark 4.4. If Γr is, for instance, a spherical cap, one can find a system of nodes
{ηi }i=1,...,L ⊂ Γr and positive weights w = (w1 , . . . , wL ) such that
L
x̄(ζ)dS(ζ) = wi x̄(ηi ), for all x̄ ∈ P2N .
Γr i=1
Then it is clear that the discretization operator

Dx̄ = x̄ = x̄(ηi ) i=1,...,L ∈ RL , x̄ ∈ L2 (Γr ),
meets condition. If x satisfies the problem (25)–(27) and the approximations xεj of
x† are chosen to be of the form xεN,λ
1 ,ε2
1 ,λ2
as in Method 4.1, then we see that the
parameter choice problem for the inversion procedure described in Section 4.1 fits
into Setting 4.3.
The optimal parameter jopt is the one that minimizes minj=1,2,...,M x† −
Of course, xεjopt cannot be found without knowledge of x† . But we
xεj L2 (Γr ) .
observe
- - 2 3
- ε -
-xj − xεjopt - = sup xεj − xεjopt , a
L2 (Γr ) a∈L2 (Γr ),aL2 (Γr ) =1 L2 (Γr )
2 3
= max xεj − xεjopt , a , (43)
a∈A L2 (Γr )
where the finite set A is defined as follows

xεj − xεi
A= - -
-xε − xε - 2 : i, j = 1, 2, . . . , M ⊂ PκN . (44)
j i L (Γr )
Then, for any j = 1, 2, . . . , M and a ∈ A, the right-hand side of (43) can be

rewritten
2 3 ? @ 2 3
xεj − xεjopt , a 2 = xεj , a L2 (Γ ) − xεjopt , a 2
L (Γr ) r L (Γr )
? ε @ 2 3
= xj , a L2 (Γ ) − Dxεjopt , Da
r
? @ w
≈ xεj , a L2 (Γ ) − Dy ε , Daw .
r
The computations above motivate the following choice of j ∗ as a substitute to the

choice j = jopt :
Method 4.5. Let Setting 4.3 hold true and the set A be given as in (44). Then we
choose the parameter j ∗ to be the minimizer of
? @

min max xεj , a L2 (Γ ) − Dy ε , Daw
j=1,...,M a∈A r
Choosing j ∗ according to Method 4.5 eventually allows the estimate:

Theorem 4.6. Let Setting 4.3 hold true. Then the minimizer j ∗ according to Method
4.5 satisfies
- † - - - - -
-x − xε∗ - 2 - † ε - - † -
j L (Γr )
-x − xjopt - + 2 -Dx − Dx jopt - + 2ε.
ε
2 L (Γr ) w
In other words, we can reasonably bound the error of xεj∗

against the error of
the optimal xεjopt among all available approximations xεj . Knowledge of the noise
level ε is not required for the choice of the parameter j ∗ . More details can be found
in [26].
4.3. Numerical example

In this section, we illustrate the numerical performance of the previously described
joint inversion and parameter choice method for the problem (25)–(27). The ap-
proximations xεN,λ1 ,ε2
1 ,λ2
are obtained as described in Method 4.1, the choice of λ∗1 ,
∗
λ2 is performed according to Method 4.5 and Remark 4.4. The precise procedure
for the numerical test is as follows:
(a) From the EGM2008 gravity potential model (cf. [55]), we generate a reference
potential x up to spherical harmonic degree 130 on a sphere Ωρ , ρ = 7, 071km,
¯ r̄), r = 6, 371km, with center ξ¯ = (0, 0, r)
and on a spherical cap Γr = Γr (ξ,
and spherical radius r̄ = 0.3 (corresponding to a spherical radius of ap-
proximately 5, 000km around the North Pole at the Earth’s surface). The
truncation degree in Method 4.1 will be fixed to *κN + = 150.
(b) From EGM2008, we generate different sets of corresponding noisy measure-

ments y1ε1 , y2ε2 , where the noise levels ε2 of the ground data and ε1 of the
satellite data are varied among 0.001, 0.1. The data on Ωρ are computed on
an equiangular grid according to [16] while the data on the spherical cap
Γr are computed on a Gauss–Legendre grid according to [33]. In a first set
of examples we choose the cubature rules [16, 33] to be polynomially exact
up to degree 300, which guarantees condition (29) and implies y1ε1 ∈ RL1 ,
y2ε2 ∈ RL2 with L1 ≈ L2 ≈ 90, 000. In two further examples, we choose the
polynomial exactness of the cubature rules to be only of degree 130 and de-
gree 80, respectively (i.e., y1ε1 ∈ RL1 , y2ε2 ∈ RL2 with L1 ≈ L2 ≈ 18, 000 and
L1 ≈ L2 ≈ 6, 500, respectively). In these cases, the condition (29) is violated.
(c) For the different sets of input data from part (b), we compute approximations
xεN,λ
1 ,ε2
1 ,λ2
of x† on Γr via the procedure of Method 4.1. The parameter λ1 is
varied in the interval [101 , 108 ] and λ2 is varied in the interval [10−2 , 103 ].
This way, we obtain M = 100 different approximations xεN,λ 1 ,ε2
1 ,λ2
for each fixed
ε1 ε2
noise level ε1 , ε2 and for each set of input data y1 , y2 .
(d) Among the M approximations xεN,λ 1 ,ε2
1 ,λ2
, we use the proposed procedure from
Method 4.5 to choose a ‘good’ approximation xεN,λ 1 ,ε2
∗ ,λ∗ . Afterwards, we com-
1 2
pare the relative error of our parameter choice, i.e.,
† †
err∗ = xεN,λ
1 ,ε2
∗ ,λ∗ − x L2 (Γr ) / x L2 (Γr ) ,
1 2
with the actual best relative error, i.e.,

erropt = xεN,λ
1 ,ε2
opt
,λopt
− x† L2 (Γr ) / x† L2 (Γr ) ,
1 2
which can be computed because we know x† from part (a).

The results of the tests are shown in Figures 4.2 and 4.3. Each figure shows
the relative errors err∗ and erropt for every test run. Additionally, we plotted
errmax = maxλ1 ,λ2 xεN,λ
1 ,ε2
− x† L2 (Γr ) / x† L2 (Γr ) and the average errors errav =
1
ε1 ,ε2 †
1 ,λ2
†
M λ1 ,λ2 xN,λ1 ,λ2 − x L2 (Γr ) / x L2 (Γr ) in order to illustrate the performance.
R=7071.2, r=6371.2, epsilon2=0.0010, epsilon1=0.0010 R=7071.2, r=6371.2, epsilon2=0.0010, epsilon1=0.1000
0 0
10 10
errmax errmax
errav errav
erropt erropt
errk* errk*
−1
10
−2
10
−1
10
−3
10
Figure 4.2. Relative errors for ε1 = ε2 = 0.001 (left) and ε1 = 0.1,

ε2 = 0.001 (right; the dotted black line marks the noise level ε1 = 0.1)
and cubature rules with polynomial exactness of degree 300.
In Figure 4.2, it can be seen that the algorithm works particularly well for the
setting ε1 = ε2 and that the oracle error err∗ is nearly identical with the minimum
error erropt . The situation is different when ε2 . ε1 . The minimum error erropt is
smaller than the noise level ε2 . Thus, since our parameter choice strategy is based
on comparing xεN,λ1 ,ε2
1 ,λ2
to y2ε2 , we cannot expect that err∗ is as good as erropt . Yet,
astonishingly enough, it seems that err∗ is still slightly smaller than ε2 for our test
setting.
R=7071.2, r=6371.2, epsilon2=0.0010, epsilon1=0.0010, quadrature degree=130 R=7071.2, r=6371.2, epsilon2=0.0010, epsilon1=0.0010, quadrature degree=80
0
10 0.45
errmax errmax
errav errav
erropt erropt
0.4
errk* errk*
−1
10
0.35
0.3
−2
10
0.25
0.2
−3
10
Figure 4.3. Relative errors for ε1 = ε2 = 0.001 and cubature rules

with polynomial exactness of degree 130 (left) and of degree 80 (right).
Figure 4.3 show that the parameter choice algorithm is fairly stable with
respect to violation of condition (29). For cubature rules with polynomial exact-
ness only of degree 130, we see that err∗ is still nearly identical to erropt. For
exactness of degree 80, however, we see that the parameter choice rules fails. The
good stability of the rule in our example with respect to violation of (29) is due
to the fact that the strongest contributions of the EGM2008 model are at low
spherical harmonic degrees. For tests with synthetic data that has approximately
equal strength at all spherical harmonic degrees, we refer the reader to [26]. In
such situations the parameter choice rule is less stable but ‘small’ violations of
condition (29) still yield good results.
Acknowledgment
Pavlo Tkachenko gratefully acknowledges the support of the Austrian Science Fund
(FWF): projects P25424, I1669, and of the consortium AMMODIT funded within
EU H2020-MSCA-RICE. Sergiy Pereverzyev Jr. gratefully acknowledges the sup-
port of the Austrian Science Fund (FWF): project P29514-N32.
References
[1] R.S. Anderssen. The linear functional strategy for improperly posed problems. In
J.R. Cannon and U. Hornung, editors, Proceedings of the Conference held at the
Mathematical Research Institute at Oberwolfach, Black Forest, May 18–24, 1986,
pages 11–30, 1986.
[2] R.S. Anderssen. Inverse problems: A pragmatist’s approach to the recovery of infor-
mation from indirect measurements. ANZIAM, 45:C588–C622, 2005.
[3] R.S. Anderssen and H.W. Engl. The role of linear functionals in improving conver-
gence rates for parameter identification via Tikhonov regularization. In M. Yamaguti,
K. Hayakawa, Y. Iso, M. Mori, T. Nishida, K. Tomoeda, and M. Yamamoto, editors,
Inverse Problems in Engineering Sciences, ICM-90 Satellite Conf. Proc., pages 1–10,
1991.
[4] F. Bauer, M. Gutting, and M.A. Lukas. Evaluation of parameter choice methods for
regularization of ill-posed problems in geomathematics. In W. Freeden, M.Z. Nashed,
and T. Sonar, editors, Handbook of Geomathematics. Springer, 2nd edition, 2015.
[5] F. Bauer and S. Kindermann. The quasi-optimality criterion for classical inverse
problems. Inverse Problems, 24:035002, 2009.
[6] F. Bauer, P. Mathe, and S. Pereverzev. Local solutions to inverse problems in
geodesy. J. Geod., 81:39–51, 2007.
[7] M. Bayer, W. Freeden, and T. Maier. A vector wavelet approach in iono- and mag-
netospheric geomagnetic satellite data. J. Atm. Solar-Terr. Phys., 63:581–597, 2001.
[8] M. Brenner, Y. Jiang, and Y. Xu. Multiparameter regularization for Volterra kernel
identification via multiscale collocation methods. Adv. Comp. Math., 31:421–455,
2008.
[9] C. Brezinski, M. Redivo-Zaglia, G. Rodriguez, and S. Seatzu. Multi-parameter reg-
ularization techniques for ill-conditioned linear systems. Numerische Mathematik,
94:203–228, 2003.
[10] A. Chambodut, I. Panet, M. Mandea, M. Diamet, M. Holschneider, and O. Jamet.
Wavelet frames: an alternative to spherical harmonic representation of potential
fields. Geophys. J. Int., 163:875–899, 2005.
[11] CHAMP. http://www.gfz-potsdam.de/champ/.
[12] J. Chen, S. Pereverzyev Jr., and Y. Xu. Aggregation of regularized solutions from
multiple observation models. Inverse Problems, 31:075005, 2015.
[13] Z. Chen, Y. Lu, Y. Xu, and H. Yang. Multi-parameter Tikhonov regularization for
linear ill-posed operator equations. J. Comp. Math, 26:37–55, 2008.
[14] S. Dahlke, W. Dahmen, E. Schmitt, and I. Weinreich. Multiresolution analysis and
wavelets on s2 and s3 . Num. Func. Anal. Appl., 16:19–41, 1995.
[15] J.C. De Los Reyes and C.-B. Schönlieb. Image denoising: Learning the noise model
via nonsmooth PDE-constrained optimization. Inverse Probl. Imaging, 7:1183–1214,
2013.
[16] J.R. Driscoll and M.H. Healy Jr. Computing fourier transforms and convolutions on
the 2-sphere. Adv. Appl. Math., 15:202–250, 1994.
[17] W. Freeden. On approximation by harmonic splines. Manuscr. Geod., 6:193–244,
1981.
[18] W. Freeden. Multiscale Modelling of Spaceborne Geodata. Teubner, 1999.
[19] W. Freeden and C. Gerhards. Poloidal and toroidal field modeling in terms of locally
supported vector wavelets. Math. Geosc., 42:818–838, 2010.
[20] W. Freeden and S. Pereverzev. Spherical Tikhonov regularization wavelets in satellite
gravity gradiometry with random noise. J. Geod., 74:730–736, 2001.
[21] W. Freeden and F. Schneider. Regularization wavelets and multiresolution. Inverse

Problems, 14:225–243, 1998.
[22] W. Freeden and U. Windheuser. Combined spherical harmonics and wavelet expan-
sion – a future concept in earth’s gravitational potential determination. Appl. Comp.
Harm. Anal., 4:1–37, 1997.
[23] E. Friis-Christensen, H. Lühr, and G. Hulot. Swarm: A constellation to study the
Earth’s magnetic field. Earth Planets Space, 58:351–358, 2006.
[24] C. Gerhards. Locally supported wavelets for the separation of spherical vector fields
with respect to their sources. Int. J. Wavel. Multires. Inf. Process., 10:1250034, 2012.
[25] C. Gerhards. A combination of downward continuation and local approximation for
harmonic potentials. Inverse Problems, 30:085004, 2014.
[26] C. Gerhards, S. Pereverzyev Jr., and P. Tkachenko. A parameter choice strategy for
the inversion of multiple observations. Adv. Comput. Math., 43:101–112, 2017.
[27] GOCE. http://www.esa.int/our activities/observing the earth/goce.
[28] A. Goldenshluger and S. Pereverzev. Adaptive estimation of linear functionals in
Hilbert scales from indirect white noise observations. Probab. Theory Rel. Fields,
118:169–186, 2000.
[29] GRACE. http://www.jpl.nasa.gov/missions/gravity-recovery-and-climate-
experiment-grace/.
[30] A. Gyulai, M.K. Baracza, and E.E. Tolnai. The application of joint inversion in
geophysical exploration. Int. J. Geosc., 4:283–289, 2013.
[31] G.V. Haines. Spherical cap harmonic analysis. J. Geophys. Res., 90:2583–2591, 1985.
[32] M. Hegland. An optimal order regularization method which does not use additional
smoothness assumptions. SIAM J. Num. Anal., 29:1446–1461, 1992.
[33] K. Hesse and R.S. Womersley. Numerical integration with polynomial exactness over
a spherical cap. Adv. Comp. Math., 36:451–483, 2012.
[34] M. Holschneider. Continuous wavelet transforms on the sphere. J. Math. Phys.,
37:4156–4165, 1996.
[35] M. Holschneider, A. Chambodut, and M. Mandea. From global to regional analysis
of the magnetic field on the sphere using wavelet frames. Phys. Earth Planet. Inter.,
135:107124, 2003.
[36] K. Ito, B. Jin, and Takeuchi. Multi-parameter Tikhonov regularization. Technical
report, arXiv 1102.1173, 2011.
[37] L.K. Jones. A simple lemma on greedy approximation in Hilbert space and con-
vergence rates for projection pursuit regression and neural network training. Ann.
Statist., 20:608–613, 1992.
[38] A. Juditsky and A. Nemirovski. Functional aggregation for nonparametric regression.
Ann. Statist., 28:681–712, 2000.
[39] R. Klees and T. Wittwer. Local gravity field modeling with multipole wavelets. In
P. Tregoning and C. Rizos, editors, Dynamic Planet, International Association of
Geodesy Symposia, volume 130. Springer, 2007.
[40] K.R. Koch and J. Kusche. Regularization of geopotential determination from satellite
data by variance components. J. Geod., 76:259.268, 2002.
[41] K. Kunisch and T. Pock. A bilevel optimization approach for parameter learning in
variational models. SIAM J. Imag. Sci., 6:938–983, 2013.
[42] A.K. Louis. Approximate inverse for linear and nonlinear problems. Inverse Prob-
lems, 12:175–190, 1996.
[43] S. Lu and S. Pereverzev. Regularization Theory for ill-posed problems – Selected
Topics. de Gruyter, 2013.
[44] S. Lu and S. Pereverzyev. Multiparameter regularization in downward continuation
of satellite data. In W. Freeden, M.Z. Nashed, and T. Sonar, editors, Handbook of
Geomathematics. Springer, 2nd edition, 2015.
[45] S. Lu, S. Pereverzyev Jr., and S. Sampath. Multiparameter regularization for con-
struction of extrapolating estimators in statistical learning theory. In X. Shen and
A.I. Zayed, editors, Multiscale Signal Analysis and Modeling. Springer, 2013.
[46] T. Maier and C. Mayer. Multiscale downward continuation of champ fgm-data for
crustal field modelling. In C. Rizos, H. Lühr, and P. Schwintzer, editors, First
CHAMP Mission Results for Gravity, Magnetic and Atmospheric Studies. Springer,
2003.
[47] P. Mathe. The Lepskii principle revisited. Inverse Problems, 22:L11–L15, 2006.
[48] P. Mathe and B. Hofmann. How general are general source conditions? Inverse Prob-
lems, 24:015009, 2008.
[49] P. Mathe and S. Pereverzev. Geometry of linear ill-posed problems in variable Hilbert
scales. Inverse Problems, 19:789, 2003.
[50] C. Mayer and T. Maier. Separating inner and outer Earth’s magnetic field from
CHAMP satellite measurements by means of vector scaling functions and wavelets.
Geophys. J. Int., 167:11881203, 2006.
[51] J.R.W. Merrick. Aggregation of forecasts from multiple simulation models. In Pro-
ceedings of the 2013 Winter Simulation Conference, pages 533–542. IEEE, 2013.
[52] V. Michel. Scale continuous, scale discretized and scale discrete harmonic wavelets
for the outer and the inner space of a sphere and their application to an inverse
problem in geomathematics. Appl. Comp. Harm. Anal., 12:77–99, 2002.
[53] V. Naumova, S. Pereverzev, and P. Tkachenko. Regularized collocation for spherical
harmonics gravitational field modeling. Int. J. Geomath., 5:81–98, 2014.
[54] NGDC-720. http://geomag.org/models/ngdc720.html.
[55] N.K. Pavlis, S.A. Holmes, S.C. Kenyon, and Factor J.K. The development and
evaluation of the Earth Gravitational Model 2008 (EGM2008). J. Geophys. Res.,
117:B04406, 2012.
[56] S. Pereverzyev and E. Schock. Error estimates for band-limited spherical regulariza-
tion wavelets in an inverse problem of satellite geodesy. Inverse Problems, 15:881–890,
1999.
[57] S. Pereverzyev and P. Tkachenko. Pointwise computation in an ill-posed spherical
pseudo-differential equation. Comp. Meth. Appl. Math., 15:213–219, 2015.
[58] A. Plattner and F.J. Simons. Spatiospectral concentration of vector fields on a
sphere. Appl. Comp. Harm. Anal., 36:1–22, 2014.
[59] A. Plattner and F.J. Simons. Potential-field estimation from satellite data using
scalar and vector Slepian functions. In W. Freeden, M.Z. Nashed, and T. Sonar,
editors, Handbook of Geomathematics. Springer, 2nd edition, 2015.
[60] P. Schröder and W. Sweldens. Spherical wavelets: Efficiently representing functions
on the sphere. In S.G. Mair and R. Cook, editors, Proceedings of the 22nd Annual
Conference on Computer Graphics and Interactive Techniques. ACM Press, 1995.
[61] L. Shure, R.L. Parker, and G.E. Backus. Harmonic splines for geomagnetic modeling.
Phys. Earth Planet. Inter., 28:215–229, 1982.
[62] F.J. Simons, F.A. Dahlen, and M.A. Wieczorek. Spatiospectral localization on a
sphere. SIAM Review, 48:505–536, 2006.
[63] F.J. Simons and A. Plattner. Scalar and vector Slepian functions, spherical signal
estimation and spectral analysis. In W. Freeden, M.Z. Nashed, and T. Sonar, editors,
Handbook of Geomathematics. Springer, 2nd edition, 2015.
[64] S.L. Svensson. Pseudodifferential operators – a new approach to the boundary prob-
lems of physical geodesy. Manuscr. Geod., 8:1–40, 1983.
[65] E. Thébault, J.J. Schott, and M. Mandea. Revised spherical cap harmonic analysis
(R-SCHA): Validation and properties. J. Geophys. Res., 111:B01102, 2006.
[66] A.N. Tikhonov and V.B. Glasko. Use of the regularization method in non-linear
problems. USSR Comput. Math. Math. Phys., 5:93–107, 1965.
[67] H. Trompat, F. Boschetti, and P. Hornby. Improved downward continuation of po-
tential field data. Explor. Geophys., 34:249–256, 2003.
[68] I.N. Tziavos, V.D. Andritsanos, R. Forsberg, and A.V. Olesen. Numerical investi-
gation of downward continuation methods for airborne gravity data. In C. Jekeli,
L. Bastos, and J. Fernandes, editors, Gravity, Geoid and Space Missions – GGSM
2004 IAG International Symposium Porto. Springer, 2005.
[69] K. Vozoff and D.L.B. Jupp. Joint inversion of geophysical data. Geophys. J. Int.,
42:977–991, 1975.
Christian Gerhards
Computational Science Center
University of Vienna
A-1090 Vienna, Austria
e-mail: christian.gerhards@univie.ac.at
Sergiy Pereverzyev Jr.
Department of Mathematics
University of Innsbruck
A-6020 Innsbruck, Austria
e-mail: sergiy.pereverzyev@uibk.ac.at
Pavlo Tkachenko
Institute for Design and Control
of Mechatronical Systems
Johannes Kepler University
A-4040 Linz, Austria
e-mail: pavlo.tkachenko@jku.at

On the Non-uniqueness of Gravitational and

Magnetic Field Data Inversion (Survey Article)
Sarah Leweke, Volker Michel, and Roger Telschow
Abstract. The gravitational and the magnetic field of the Earth represent
some of the most important observables of the geosystem. The inversion of
these fields reveals hidden structures and dynamics at the surface or in the in-
terior of the Earth (or other celestial bodies). However, the inversions of both
fields suffer from a severe non-uniqueness of the solutions. In this paper, we
present a generalized approach which includes the inversion of gravitational
and magnetic field data. Amongst others, uniqueness constraints are proposed
and compared. This includes the surface density ansatz (also known as the
thin layer assumption). We characterize the null space of the considered class
of inverse problems via an appropriate orthonormal basis system. Further,
we expand the reconstructable part of the solution by means of orthonormal
bases and reproducing kernels. One result is that information on the radial
dependence of the solution is lost in the observables. As an illustration of the
non-uniqueness, we show examples of anomalies which cannot be disclosed
from the inversion of gravitational data. This paper is intended to be a theo-
retical reference work on the inversion of gravitational but also magnetic field
data of the Earth.
1. Introduction
Numerous tasks in mathematical geodesy involve the regularization of ill-posed
inverse problems. The reason is obvious: neither the interior of the Earth nor the
Earth’s surface in its entirety are accessible for exploration. However, the demand
for more accurate and more localized models has dramatically increased for the
last decades. As a consequence, numerous large data sets of various observables
have been generated, also by means of satellite missions. These data sets often
provide us with the possibility to derive models for non-observable, but urgently
needed geodetic fields. Examples are the quantification of mass transports due
to climate change or other phenomena (GRACE (Gravity Recovery And Climate
Experiment) data are well appropriate for this purpose, see, e.g., [11, 26, 29, 52, 53])
884 S. Leweke, V. Michel, and R. Telschow
and the modeling of those layers of the Earth which contribute to the magnetic
field (this can be done with SWARM data, see, e.g., [37, 43, 50]).
This survey article presents a generalized approach which comprises, in par-
ticular, the inversion of gravitational or magnetic field data. In the former case,
the unknown is the mass density distribution of the Earth’s body or its surface. In
the latter case, the unknown is considered to be the electric current distribution
inside. In this sense, this paper is an extension of the survey article [33] on inverse
gravimetry. One benefit of the generalized approach is that it makes it easier to
transfer theoretical knowledge and numerical methods from one problem to the
other within the considered class of problems. For example, it was shown in [23]
and [33] that such a transfer yields novel achievements. Furthermore, our gen-
eralized approach also enables us to set the surface mass density approach (also
known as the thin layer assumption) into the same concept with the inversion for
volumetric density distribution – two approaches which have often been used par-
allelly and independently (see, e.g., [33, 52]). Since this paper addresses primarily
a geodetic audience, we focus on the relevant facts and their interpretation. For
the detailed mathematical theory including the proofs, we recommend to use the
paper [34] as a supplement.
Note that the considered inversion of magnetic field data is motivated by the
inversion of MEG (magnetoencephalography) data, as it occurs in medical imaging
(see also [23] and the references therein). Thus, it does not represent a typical
inverse problem in geomagnetics, where, for instance, material parameters like the
magnetization or the susceptibility are the unknowns and not the current (see, e.g.,
[46]). However, the inversion of the magnetic field for currents in the interior might
be interesting for investigating the outer core. Nevertheless, there is still an obvious
limitation of our generalized approach with respect to the practical applicability
in geomagnetics. On the other hand, reversing the point of view, the generalized
approach shows a perspective how methods from medical imaging (which exist in
a vast variety) could be transferred to geodetic and geophysical inverse problems.
The content and the outline of the paper are as follows: in Section 2, we
summarize some basic fundamentals, like the definition of the function spaces and
the orthogonal polynomials which we need.
In Section 3, we formulate the generalized class of inverse problems which
represents the central theme of this paper. Then, we discuss two particular cases:
the inversion of the gravitational field (this is known as the inverse gravimetric
problem) and the inversion of the Bio–Savart operator of a magnetic field for get-
ting the current distribution inside (we call this the inverse magnetic problem).
With this in mind, every theoretical result that we present here for the generalized
problem is valid for these two particular applications, and the derived formulae can
be directly used for the precise problem by inserting the associated parameters. In
Subsection 3.2, we derive a spectral relation between the given field and the un-
known field. This relation directly shows the problem of the non-uniqueness which
is linked to the insufficiently identifiable radial parametrization of the solution.
Non-uniqueness of Gravitational and Magnetic Field Data 885
In Section 4, we introduce a class of orthonormal basis systems on a 3-

dimensional ball. One particular instance of this class yields the well-known system
of harmonic and anharmonic functions which have been used for the inverse gravi-
metric problem. We include some plots of the basis functions and show that the
basis is appropriate for separating the solution into its projections on the null space
of the solution (i.e., the indeterminable part of the solution) and on the orthogonal
complement (the components of the solution which are uniquely constrained by
the given data). We also show graphical illustrations of phantoms which occur,
that is, examples of anomalies inside the Earth which cannot be distinguished if
only gravitational data are available.
In Section 5, we discuss several modeling assumptions which can be used to
obtain a unique solution: a minimum norm constraint, a harmonicity constraint, a
layer density constraint and the surface density (i.e., thin layer) constraint, which
is common for the identification of water mass transports.
2. Preliminaries
In this work, the set of positive integers is denoted by N, where N0 := N ∪ {0}.
Moreover, R represents the set of real numbers. The Euclidean standard R3 -scalar
product (dot product) is denoted by · and the cross product by√×. The norm
3
associated to the Euclidean dot product is represented by |x|!:= x · x, x ∈ R ".
Furthermore, the sphere with radius R is denoted by ΩR := x ∈ R |x| = R 3
! "
and the corresponding (closed) ball is denoted by B := x ∈ R3 |x| ≤ R . For
R = 1, we often use the abbreviation Ω := Ω1 . By S := Ωβ , with β > R, we
denote a particular sphere in the exterior of B. This could, for example, represent
a satellite altitude or the location of airborne data.
A function F : G → R possessing k continuous derivatives on the open set
G ⊂ Rn is of class C(k) (G), for 0 ≤ k ≤ ∞. Furthermore, for a measurable
set G ⊂ Rn , L2 (G) stands for the space of all square-integrable functions (more
precisely, some equivalence classes of such functions). L2 (G) is a Hilbert space with
the inner product

F, GL2 (G) := F (x)G(x) dμ(x), F, G ∈ L2 (G),
G
and the norm

1/2
F L2 (G) = F (x)2 dμ(x) , F ∈ L2 (G),
G
where μ is an appropriate measure, like a surface measure ω if G is a surface. For

a mathematically accurate definition of the space, see, for example, [42].
(α,β)
With Pm , we denote the Jacobi polynomials, where α, β > −1. They are
uniquely determined by the conditions that
(α,β)
1. each Pm is a polynomial of degree m,
2. for all m, n ∈ N0 with m = n,
2 3 1
(α,β) (α,β) (α,β)
Pm , Pn := (1 − x)α (1 + x)β Pm (x)Pn(α,β) (x) dx = 0, (1)
α,β −1
(α,β) m+α
3. and for each m ∈ N0 , we set Pm (1) = m .
For α = β = 0, the Jacobi polynomials coincide with the Legendre polynomials.
For further properties and the L2 [0, R]-norm of Legendre, or (more generally)
Jacobi polynomials, see [24, 36, 49].
3. Generalization of gravitational and magnetic field inversion

3.1. A class of inverse problems and examples
Within this paper, we consider a class of inverse problems which are given by a
Fredholm integral operator of the first kind T : L2 (B) → L2 (S)

T : D → D(x)k(x, ·) dx = V (2)
B
with an integral kernel k : B × S → R of the form
∞ l
|x| i x y
k(x, y) := ci i+1 Pi · , (3)
i=0 |y| |x| |y|
which is defined for all (x, y) ∈ dom(k), where the domain of the kernel k is given
by
dom(k) := {(x, y) ∈ B × S | x = 0 if there exists i ∈ N0 with li < 0} .
In this setting, the right-hand side V in Equation (2) is given and the function D
is unknown. It is the aim to reconstruct D in B from knowledge of V on S. In order
to have a well-defined integral kernel, which means that the series representation
in (3) converges, k has to fulfil certain assumptions:
Assumption 3.1. teset
1. The sequence (ci )i∈N0 is a real and bounded sequence (i.e., there exists c ∈ R+
such that supi∈N0 |ci | ≤ c).
2. The sequence of real exponents (li )i∈N0 satisfies inf i∈N0 li ≥ −1.
3. The sequence (li )i∈N0 fulfils the condition supi∈N0 Rli −i < ∞.
Note, that the third condition implies
1 1
Ri−li = ≥ > 0.
Rli −i supi∈N Rli −i
This kind of integral equation arises in many areas, for example, in geosciences
and medical imaging. Two examples for this inverse problem are given below.
For both, Example 3.2 (inverse gravimetric problem) and Example 3.3 (inverse
magnetic problem), the conditions of Assumption 3.1 are fulfilled. In the first
particular case, that is, li = i and ci = γ for all i ∈ N0 , the integral kernel is well
known. In this case, we directly obtain k(x, y) = |x−y| γ
for |x| < |y|, due to the
identity
∞ i
|x| x y 1
i+1 iP · = for |x| < |y|. (4)
i=0 |y|
|x| |y| |x − y|
Example 3.2 (The Inverse Gravimetric Problem). For the inverse gravimetric
problem, the kernel and the integral operator are given by

T G : D → D(x)k G (x, ·) dx,
B
∞ i
G γ |x| x y
k (x, y) := =γ P
i+1 i
· ,
|x − y| i=0 |y|
|x| |y|
where x ∈ B, y ∈ S, Pi denotes the Legendre polynomial of degree i and γ is

the gravitational constant. T G D is known as the gravitational potential or the
Newton potential. The associated inverse problem T G D = V represents the recon-
struction of a (volumetric) mass density function from the gravitational potential,
which is important, for example, for the detection of particular anomalies or mass
transports. For the latter, time series of potential models have been provided, for
instance, by the GRACE mission, see [10]. Note that the determination of a surface
density can be regarded as a particular modeling in this context.
This problem first occurs in the works of Stokes [47] and has been widely
discussed since then (see also the survey article [33]).
Example 3.3 (The Inverse Magnetic Problem). To compute the magnetic field
B caused by electric sources inside a body, the quasi-static approximation of
Maxwell’s equation is often used, see [39].
E = −∇U on B, ∇·B = 0 on B,
T
∇×B =0 on S, ∇ × B = μ0 J on B,
where E is the electric field, U is the electric potential, J T = J P + σE is the total

current with the primary current J P and the Ohmic current σE, σ is the conduc-
tivity, and μ0 is the permeability. It is common to use the Biot–Savart operator
instead of Maxwell’s equations to describe the relation between the current and
the magnetic field

μ0 x−y
B(x) = J T (y) × dy. (5)
4π B |x − y|
In this case, we want to recover a particular component of the electric current
inside B (which could be the Earth (in particular the outer core)). Note that
this geophysical problem is closely related to a problem in medical imaging, where

neuronal currents are determined from magnetoencephalography (MEG) data, see,
for example, [19]. In some applications, only the reconstruction of the primary
current instead of the total current or the induced current is of interest. After
splitting the current in this sense and assuming a ball-shaped conductor consisting
of spherical shells Ωj with constant conductivities σj , one obtains the Geselowitz’
formula (see [25])

μ0 x−y
B(x) = J P (y) × dy
4π B |x − y|3

μ0 x−y
− (σj−1 − σj ) V (y)n(y) × dω(y),
4π j Ωj−1 |x − y|3
where n is the normal vector on the surface Ωj . With the identity in (4) and
after further calculations, see [23], one gets a relation for the magnetic potential
(B = ∇V )
∞ i
1 |x| x y
V (y) = ∇x · (J P (x)×x) Pi · dx.
4π B i=0 |y|
i+1
(i + 1) |x| |y|
More precisely, the vectorial current J P inside B can be decomposed via two scalar-
valued (up to an additional constant unique) functions F and G and a scalar-valued
unique function J r (see, e.g., [23]) as follows:
1 ∗ 1
J P (rξ) = ∇ G(rξ) − L∗ξ F (rξ) + J r (rξ)ξ.
r ξ r
Here, B \{0} x = rξ with ξ ∈ Ω and r = |x|, ∇∗ξ is the surface gradient, and
L∗ξ := ξ × ∇∗ξ is the surface curl operator on the unit sphere. Due to [45] and the
above decomposition, the relation between the current and the magnetic potential
V in a spherical model can be described by
∞ i−1
1 |x| x y
V (y) = Δ∗|x|
x F (x) Pi · dx,
4π B i=0 |y|
i+1
(i + 1) |x| |y|
where Δ∗x denotes the Beltrami operator.
|x|
Hence, only the function F and, therefore, only one tangential component of
the current can be reconstructed. We use now the abbreviation D(x) := Δ∗x F (x)
|x|
such that for the inverse magnetic problem (as we call the problem here), the
kernel and the integral operator are given by

T M : D → D(x)k M (x, ·) dx, (6)
B
∞
1 |x|i−1 x y
k M (x, y) := Pi · , (7)
4π i=0 |y|i+1 (i + 1) |x| |y|
where x ∈ B \{0}, y ∈ S.
This particular modeling of the inversion of magnetic data has been used for
data from MEG, as we indicated above. For two reasons, we consider a discussion to
be useful: The magnetic field of a ball-shaped domain with a current in the interior
is also relevant in geodesy, and there is a close link to the inverse gravimetric
problem as our generalized approach suggests.
We can find further properties of the integral kernel in (3). An estimate shows
that the kernel function k(·, y), for each fixed y ∈ S, is a function in L2 (B). Indeed
(with x = rξ, r ∈ [0, R], ξ ∈ Ω) we get, using Assumption 3.1 and the fact that
|Pi (t)| ≤ 1 for all i ∈ N0 and all t ∈ [−1, 1], the estimate

∞ 2
2 |x|li x y
(k(x, y)) dx = ci i+1 Pi · dx
B B i=0 |y| |x| |y|
∞ l
2 R ∞
2
2 |x| i 2 2 r li
≤c i+1
dx = 4πc r i+1
dr
B i=0 |y| 0 i=0 |y|
R ∞
2 2 ∞
2
2 rli +1 2 ln −n Ri+1
= 4πc i+1 dr ≤ 4πRc sup R i+1 < ∞.
0 i=0 |y| n∈N0 i=0 |y|
The last series is convergent and, hence, finite, since it is a geometric series. With
similar calculations one can prove that the interchanging between the series and
the integration over B was allowed.
Besides the well-definition of the integral kernel, we need the existence of
the integral in (2) to obtain a well-defined problem. We will later see that this
is achieved if some technical conditions are fulfilled. On the other hand, for the
well-posedness of the problem (in the sense of Hadamard), three questions are
important.
• Does, for every right-hand side V in (2), a solution D exist?
• Is there not more than one solution D for a given V ?
• Is the problem stable, that is, does D depend continuously on the data V ?
The question about the non-uniqueness of the solution for the above men-
tioned problems has been discussed comprehensively in literature. One of the first
works is the paper due to Stokes [47] for the inverse gravimetric problem. Further
publications are, for example, [4, 6, 8, 48]. For a survey article on this topic, see
[33]. For the inverse magnetic problem (with a focus on medical imaging), see
[13–15, 19–22, 45].
In the following sections, we want to derive a possibility to characterize the
null space, or in other words we want to describe the part of the solution which is
non-reconstructable. We also want to formulate additional conditions to guarantee
the uniqueness of the solution. For this, we need more knowledge of the forward
problem.
3.2. Derivation of a spectral relation

In this subsection, it is our aim to derive an equation which connects the spherical
harmonics coefficients of the given function V and the unknown function D. With
this spectral relation, we are able to give answers to the questions concerning the
ill-posedness of the problem. For this purpose, we analyze the forward problem.
The following considerations are motivated by a similar result for the particular
case of the inverse gravimetric problem, see [33]. We assume that we can choose
basis functions for D which are separable into a radial and an angular part such
that D is expandable in an L2 (B)-convergent spherical harmonics series
∞ 2n+1

x
D(x) = Dn,j (|x|)Yn,j . (8)
n=0 j=1
|x|
Here, Yn,j denotes the spherical harmonics of degree n and order j, which are
an orthonormal basis for L2 (Ω). Furthermore, Dn,j (r), r ∈ [0, R], represents the
spherical harmonics coefficients for the case that D is restricted to the sphere
around the origin with radius r.
By virtue of the weak convergence in Hilbert spaces, we know that
∞ 2n+1

x
D(x)F (x) dx = Dn,j (|x|)Yn,j F (x) dx
B B n=0 j=1 |x|

∞ 2n+1

x
= Dn,j (|x|)Yn,j F (x) dx
n=0 j=1 B |x|
for all functions F ∈ L2 (B). In particular, this holds true for the integral kernel
k(·, y) ∈ L2 (B) for all y ∈ S. Inserting the expansion (8) in (2) and using the
abbreviation y = |y| η, x = rξ with η, ξ ∈ Ω, we get
∞ 2n+1
∞ R
c i r li
V (y) = r2 Dn,j (r) i+1 dr Pi (ξ · η) Yn,j (ξ) dω(ξ)
n=0 j=1 i=0 0 |y| Ω
∞ 2n+1
∞ R
ci 4π
= i+1 rli +2 Dn,j (r) dr δi,n Yn,j (η)
n=0 j=1 i=0 |y| 0 2n + 1

∞ 2n+1 R
4πcn
ln +2
= r Dn,j (r) dr Yn,j (η) . (9)
n=0 j=1 0 (2n + 1) |y|n+1
In the first step the reproducing property of the reproducing kernel for the spherical
harmonics of degree n, given by
2n + 1
Ω2 (ξ, η) → Pn (ξ · η),
4π
is used. More precisely,

2n + 1
Pi (ξ · η)Yn,j (ξ) dω(ξ) = Yn,j (η)δi,n (10)
4π Ω
for all η ∈ Ω. We also remark that the existence of the integral in (2) only depends
on the existence of the integral of the radial part and the convergence of the series
in (9). Regarding the latter, we obtain a pointwise convergence of (9) for y ∈ S,
since the following estimate of the summands in (9) (note that maxξ∈Ω |Yn,j (ξ)| ≤
#
(2n + 1)/(4π) for all n ∈ N0 ) holds true:

R
ln +2 4πcn −n−1 y
r Dn,j (r) dr |y| Yn,j
0 2n + 1 |y|
1/2
R2ln +3 R 2 2 4πc −n−1 2n + 1
≤ r (Dn,j (r)) dr |y|
2ln + 3 0 2n + 1 4π
R 1/2 n+1
R2ln +3 2 2 4π R
≤c r (D n,j (r)) dr .
R2n+2 (2ln + 3) 0 2n + 1 |y|
The right-hand side is bounded for all n ∈ N0 , due to the conditions on (ln )n∈N
(see Assumption 3.1, items 2 and 3) and the convergence of the Parseval identity
of D ∈ L2 (B). Hence, the series (9) is dominated by a geometric series for all y ∈ S
(i.e., |y| > R).
We are also able to extend the function V onto ΩR . In addition, for V |ΩR , we
obtain the L2 (ΩR )-convergence of the series representation in Equation (9). This
convergence is a direct consequence of the Cauchy–Schwarz inequality and the
1
Parseval identity (note that { R Yn,j ( R· )}n∈N0 ,j=1,...,2n+1 is an orthonormal basis
2
of L (ΩR )), since
2 2
2
R
∞ 2n+1
4πcn
ln +2
V |ΩR L2 (ΩR ) = r Dn,j (r) dr
n=0 j=1 0 (2n + 1)Rn
2
R r2ln +2
∞ 2n+1 R
4πc
2 2
≤ dr r (D n,j (r)) dr
n=0 j=1 0 R2n 0 2n + 1

R2ln +3−2n R
∞ 2n+1
≤ 16π 2 c2 r2 (Dn,j (r))2 dr
n=0 j=1
2l n + 3 0
R2ln −2n 2
≤ 16π 2 c2 R3 sup D L2 (B) < ∞.
n∈N0 2ln + 3
Hence, Equation (9) is valid pointwise on S and in the sense of L2 (ΩR ) on ΩR .
In order to find a direct relation between the Fourier coefficients of the given
function V and the unknown function D, we consider the Fourier coefficients of V
restricted to the sphere ΩR . This relation can be seen directly from (9).
Theorem 3.4. Consider the orthonormal basis system on ΩR given by the set of
functions {?R1 Yn,j ( R· )}n∈N
0· ,j=1,...,2n+1
@ . Then, the Fourier coefficients of V defined
1
by Vn,j := V |ΩR , R Yn,j R L2 (ΩR ) satisfy the identity

R
ln +2 4πcn
Vn,j = r Dn,j (r) dr .
0 (2n + 1)Rn
for all n ∈ N0 , j = 1, . . . , 2n + 1. This yields the equation
R
(2n + 1)Rn
Vn,j = rln +2 Dn,j (r) dr, if cn = 0, (11)
4πcn 0
otherwise Vn,j = 0 with j = 1, . . . , 2n + 1, respectively.
The relation from Theorem 3.4 allows an infinite number of choices for Dn,j
and, hence, the solution D cannot be uniquely determined by the function V |ΩR .
For the inverse gravimetric problem, the last relation is well known, see, for exam-
ple, [35, 38, 41], and for the inverse magnetic problem for R = 1, see for instance
[21]. Analogously, we obtain with (remember that S = Ωβ , β > R) for all n ∈ N0 ,
j = 1, . . . , 2n + 1
F G
S 1 ·
Vn,j := V |S , Yn,j
β β L2 (S)
n
R
ln +2 4πcn R
= r Dn,j (r) dr n
= Vn,j (12)
0 (2n + 1)β β
the spherical harmonics coefficients of V with respect to an orthonormal basis
system on S. Hence, we have a direct relation between the singular values of the
Fredholm integral operator T and the spherical harmonic coefficients Vn,j . The
β ) symbolizes the upward continuation from S to ΩR . The
additional factor ( R n
upward continuation does not effect the null space of the operator T at all. Due
to this property and the aim to keep the formulae simple, we analyze Equation
(11) further and keep in mind that we can consequently deduce properties of T
via Equation (12).
Note that (11) shows, in particular, the degree of freedom with respect to
the radial part of D, since Vn,j is some weighted radial mean of Dn,j (r). On the
other hand, one can expect a one-to-one relation for the angular dependence of
V and D.
4. Investigation of the homogeneous problem

In order to obtain a unique solution, an appropriate modeling is required, that is,
the solution space has to be restricted by certain constraints. Before this can be
done (in Section 5), we have to study the null space ker T , that is, the space of all
D with T D = 0. Note that, due to the linearity, all solutions of T D = V are given
by D̃ + D0 , with arbitrary D0 ∈ ker T , for a particular solution D̃ of T D = V .
4.1. Some orthonormal basis functions on the ball

It is our aim to characterize the null space, that is, the so-called kernel of the Fred-
holm integral operator of the first kind, in order to describe the non-reconstructable
parts of the solution. For the separation of L2 (B) into the null space and the or-
thogonal complement we need an appropriate basis for L2 (B).
For the ball, there are several known basis systems available. For the con-
struction of these systems see, for example, [1, 7, 17, 30, 32, 51]. We analogously use
the idea to combine an orthonormal basis system on the unit sphere with one on
the interval [0, R], to construct a basis system on the ball. The L2 (B)-orthonormal
system used here is a generalization of the system which was introduced in [17]
and [7].
For x ∈ B \{0}, it is given by

2 l
(0,ln +1/2) |x| |x| n x
Gm,n,j (x) := γm,n Pm 2 2 −1 Yn,j , (13)
R R ln |x|
(α,β)
with m, n ∈ N0 , j = 1, . . . , 2n+ 1, where {Pm }m∈N0 are the Jacobi polynomials
and γm,n are normalization constants with

4m + 2ln + 3
γm,n := . (14)
R3
(0,l +1/2)
Since α = 0 in Equation (13) and Pm n (1) = 1 for all m, n ∈ N0 , we get
Gm,n,j |ΩR = γm,n Yn,j ( R· ).
The functions in (13) were called GIm,n,j in [31] and [32] in the case of ln = n
(remember that this setting corresponds to the inverse gravimetric problem).
A continuous expansion of our functions Gm,n,j on the domain B is possible,
if all exponents ln , n ∈ N, are positive. Otherwise we obtain a singularity at the
origin of the functions Gm,n,j for negative values of ln and a discontinuity at the
very same place in the case ln = 0 for n > 0. For the theory stated in this paper,
this is not a problem, since Gm,n,j remains square-integrable for −1 ≤ ln (as we
required).
As we claimed above, the functions Gm,n,j for m, n ∈ N0 , j = 1, . . . , 2n + 1
given in (13) build an orthonormal basis for L2 (B). This property can easily be
verified by calculating the inner products and using a formula for a weighted L2 -
2
norm of Jacobi polynomials (see, e.g., [36]). With # the L (Ω)-orthogonality of the
spherical harmonics and the substitution r = R (1 + z)/2, we obtain
Gm,n,j , Gμ,ν,ι L2 (B)
2 2
r2ln +2 (0,ln +1/2)
R
r (0,ln +1/2) r
= γm,n γμ,ν δν,n δι,j P 2 − 1 P 2 − 1 dr
0 R2ln m R2 μ
R2
1
R3
= γm,n γμ,n δν,n δι,j l +5/2
(1 + z)ln +1/2 Pm(0,ln +1/2)
(z)Pμ(0,ln +1/2) (z) dz
2n −1
R3 2ln +3/2
= γm,n γμ,n δν,n δι,j δμ,m
2ln +5/2 2m + ln + 3/2
2 R3
= γm,n δν,n δι,j δμ,m = δμ,m δν,n δι,j .
4m + 2ln + 3
Thus, the set {Gm,n,j }m,n∈N0 ,j=1,...,2n+1 is L2 (B)-orthonormal. Moreover,
the spherical harmonics are complete in L2 (Ω) and the Jacobi polynomials are
complete with respect to the inner product in (1) such that the system
{Gm,n,j }m,n∈N0 ,j=1,...,2n+1
2
is complete in L (B) and constitutes an orthonormal basis.
Some of the functions GIm,n,j (i.e., in the case of ln = n) are shown in Fig-
ures 1 and 2. For m = 0, the functions GI0,n,j are inner harmonics, hence they are
harmonic, and attain their maximum and minimum on the boundary. A selection
of the functions corresponding to the inverse magnetic problem, where ln = n − 1,
is shown in Figures 3 and 4. The singularity (for n = 0, i.e., l0 = −1) at the origin
is visible in Figures 3 (A) and (C) and Figure 4 (B).
4.2. Splitting the basis into the null space and its complement
With the orthonormal basis introduced in Subsection 4.1, we are now able to
expand the functions Dn,j in (8) for all n ∈ N0 and j = 1, . . . , 2n + 1 and we
obtain
∞ 2
r ln (0,ln +1/2) r
Dn,j (r) = ln dm,n,j γm,n Pm 2 2 −1 , (15)
R m=0 R
where dm,n,j := D, Gm,n,j L2 (B) and γm,n is given in (14).
For further investigations of the forward problem, we use the representation
of (the known function) V in (9), where we have already calculated the integral
over the angular part. For the remaining integral over the radial part, we use
the precise representation of Dn,j in (15) and the orthogonality of the Jacobi
# 1/2
2
polynomials.With the substitution r = R (1 + z)/2, dr = R4 1+z dz, we get
∞ 2
R
r2ln +2
R
r
rln +2 Dn,j (r) dr = d γ
m,n,j m,n m P (0,ln +1/2)
2 − 1 dr
0 0 Rln m=0 R2
∞ 1
R3+ln R3+ln
= dm,n,j γm,n (1 + z)ln +1/2 Pm(0,ln +1/2)
(z) dz = d0,n,j γ0,n .
2ln +5/2 m=0 −1 2ln + 3
Inserting the latter result in (9), we eventually obtain (remember the defini-
tion of γm,n in (14))

R
∞ 2n+1
4πcn y
ln +2
V (y) = r Dn,j (r) dr n+1 Yn,j
n=0 j=1 0 (2n + 1) |y| |y|
1 1
z
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(a) GIm,n,j for m = 0, n = 0, j = 1 (b) GIm,n,j for m = 0, n = 2, j = 5
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(c) GIm,n,j for m = 1, n = 0, j = 1 (d) GIm,n,j for m = 1, n = 1, j = 3
Figure 1. The functions Gm,n,j in the case ln = n (also called GIm,n,j )
for different parameters m, n, j are plotted at the plane through the
origin with normal vector (1, 1, −1)T . For the particular parameters, see
the respective caption. The maximum is always yellow and the minimum
is blue (see also [32, 34]).
∞ 2n+1

R3+ln 4πcn y
= d0,n,j γ0,n Yn,j
n=0 j=1
(2ln + 3)(2n + 1)|y|n+1 |y|
∞ 2n+1

4πcn Rln −1 y
= n+1 d0,n,j γ0,n Yn,j . (16)
n=0 j=1 (2n + 1) |y| |y|
1 1
z
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(a) GIm,n,j for m = 1, n = 2, j = 5 (b) GIm,n,j for m = 2, n = 0, j = 1
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(c) GIm,n,j for m = 2, n = 1, j = 3 (d) GIm,n,j for m = 2, n = 2, j = 5
Figure 2. The functions Gm,n,j in the case ln = n (also called GIm,n,j )
for different parameters m, n, j are plotted at the plane through the
origin with normal vector (1, 1, −1)T . For the particular parameters, see
the respective caption. The maximum is always yellow and the minimum
is blue (see also [32, 34]).
Hence, Gm,n,j is in the null space of the operator T with the kernel from (3),
if and only if m > 0 or cn = 0. Examples of functions in the null space are given
in Figures 2 and 4 (for different inverse problems). The function plotted in Figure
5 is not in the null space.
Since L2 (B) is the direct sum of the null space ker T and its orthogonal
complement, the obtained result allows a precise characterization of the null space
1 1
z
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(a) Gm,n,j for m = 0, n = 0, j = 1 (b) Gm,n,j for m = 0, n = 2, j = 5
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(c) Gm,n,j for m = 1, n = 0, j = 1 (d) Gm,n,j for m = 1, n = 1, j = 3
Figure 3. The functions Gm,n,j in the case ln = n − 1 for different
parameters m, n, j are plotted at the plane through the origin with nor-
mal vector (1, 1, −1)T . For the particular parameters, see the respective
caption. The maximum is always yellow and the minimum is blue (see
also [34]).
of the corresponding Fredholm integral operator as

! "·L2 (B)
ker T = span Gm,n,j m ≥ 1, n ∈ N0 , j = 1, . . . , 2n + 1 or cn = 0 . (17)
For the inverse gravimetric problem (ln = n), we can deduce the well-known
fact that the null space can be described as the set of all anharmonic functions,
which are the elements of the orthogonal complement of the set of all harmonic
1 1
z
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(a) Gm,n,j for m = 1, n = 2, j = 5 (b) Gm,n,j for m = 2, n = 0, j = 1
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(c) Gm,n,j for m = 2, n = 1, j = 3 (d) Gm,n,j for m = 2, n = 2, j = 5
Figure 4. The functions Gm,n,j in the case ln = n − 1 for different
parameters m, n, j are plotted at the plane through the origin with nor-
mal vector (1, 1, −1)T . For the particular parameters, see the respective
caption. The maximum is always yellow and the minimum is blue (see
also [34]).
functions. That is,

·L2 (B) ⊥L2 (B)

ker T G = span GIm,n,j m ≥ 1 = F ∈ C(2) (B) ΔF = 0 ,
∂2 ∂2 ∂2
where Δ = ∂x21
+ ∂x22
+ ∂x23
represents the Laplace operator. In this case, the
functions GI0,n,j , n ∈ N0 , j = 1, . . . , 2n + 1 are the inner harmonics and, therefore,
z
0
−1
1
1
0
0
y −1 −1 x
Figure 5. The function G0,4,8 , which is not in the null space of the
Fredholm integral operator T for ln = n − 1.
form a basis for the set of all harmonic functions on the ball:
n
I 2n + 3 |x| x
G0,n,j (x) = Yn,j , x ∈ B.
R Rn+1 |x|
For some particular cases of the considered Fredholm integral operators, we are also
able to find a characterization of the null space via an elliptic partial differential
equation.
For this purpose, we consider the particular integral kernel
∞
i+κ
|x| x y
k(x, y) := ci Pi · , (x, y) ∈ dom(k),
i=0 |y|i+1 |x| |y|
for a fixed κ ∈ [−1, ∞) and ci = 0 for all i ∈ N0 . Note that in the case of the
inverse gravimetric problem κ = 0 and in the case of the inverse magnetic problem
κ = −1. We have already proven that the orthogonal complement of the null space
of the corresponding operator T is given by the set
(ker T )⊥L2 (B)

·L2 (B)

= span Gm,n,j m = 0, n ∈ N0 , j = 1, . . . , 2n + 1 and cn = 0 .
Now, we define an elliptic partial differential operator Δ̃ by

2
−κ ∂ 2 ∂ 1 ∗ −κ
Δ̃F (rξ) := Δ r F (rξ) = + + Δ r F (rξ) .
∂r2 r ∂r r2 ξ
Using the product rule for the derivative, we get

∂ ∂2
Δ̃F (rξ) = −κ(−κ − 1)r−κ−2 − 2κr−κ−1 + r−κ 2 − 2κr−κ−2
∂r ∂r

∂
+2r−κ−1 + r−κ−2 Δ∗ξ F (rξ)
∂r

∂2 ∂
= r−κ 2 + 2(1 − κ)r−κ−1 + κ(κ − 1)r−κ−2 + r−κ−2 Δ∗ξ F (rξ).
∂r ∂r
In the particular case of the inverse gravimetric problem (i.e., κ = 0) this reduces
to
2
∂ 2 ∂ 1 ∗
Δ̃F (rξ) = + + Δ F (rξ) = ΔF (rξ),
∂r2 r ∂r r2 ξ
and the differential operator corresponding to the inverse magnetic problem (i.e.,
κ = −1) is given by
2
∂ ∂ 2 1 ∗
Δ̃F (rξ) = r 2 + 4 + + Δξ F (rξ).
∂r ∂r r r
In order to get a new characterization of the null space, we apply the differential
operator to the basis functions G0,n,j for n ∈ N0 , j = 1, . . . , 2n + 1 and obtain
r n+κ r n+κ
Δ̃G0,n,j (rξ) = Δ̃ γ0,n Yn,j (ξ) = γ0,n Δ r−κ Yn,j (ξ)
R R
γ0,n n
= n+κ Δ (r Yn,j (ξ)) = 0,
R
since the mapping rξ → rn Yn,j (ξ) is a harmonic function for all n ∈ N0 , j =
1, . . . , 2n + 1. In analogy, Δ̃Gm,n,j ≡ 0 for m ≥ 1, n ∈ N0 , j = 1, . . . , 2n + 1
follows by similar considerations. This means that Δ̃F is equal to zero if and only
if rξ → r−κ F (rξ) is a harmonic function, that is, is contained in
! "
span GI0,n,j n∈N ,j=1...,2n+1 .
0
Since this is equivalent to expanding F (rξ) in terms of rκ GI0,n,j (rξ) and ln = n + κ

here, our definition in (13) leads us to the following result.
Theorem 4.1. If we assume that there exists a fixed parameter κ ≥ −1 such that
ln = n + κ for all n ∈ N0 and that cn = 0 for all n ∈ N0 , then
· 2
ker T = span {Gm,n,j | m > 0, n ∈ N0 , j = 1, . . . , 2n + 1} L (B)
⊥L2 (B)

= F : B → R (rξ → r−κ F (rξ)) ∈ C(2) (B) and Δ̃F = 0 . (18)
After having given two mathematical characterizations of the null space in

Equation (18) for a particular case (i.e., ln = n + κ) and one characterization in
Equation (17) for the general case, we want to demonstrate what kind of functions
D generate the same forward solution V .
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x −1 −1
(a) G1,5,10 (b) Sum of the function in Fig. 5 and (A)
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x −1 −1
(c) G10,2,4 (d) Sum of the function in (B) and (C)
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x −1 −1
(e) G7,10,4 (f) Sum of the function in (D) and (E)

Figure 6. Several functions from the null space of T , that is, they gen-
erate the solution V = 0 (left column), and the sum of these functions
with G0,4,8 ∈ ker T (right column) which generate the same right-hand
side V = T G0,4,8 , that is, the same data for the inverse problem.
1 1
z
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(a) The PREM model (b) Sum of PREM and K(z1 , ·)
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(c) Sum of PREM and K(z2 , ·) (d) Sum of PREM and K(z1 , ·)+K(z2 , ·)+
K(z3 , ·)
Figure 7. The density of the PREM model added to several functions
from the null space of T G . They all generate the same gravitational
potential. Here, zi ∈ B, i = 1, 2, 3 are fixed.
For this purpose, we consider the function G0,4,8 plotted in Figure 5, which
is not in the null space of the operator, that means this function generates the
result T G0,4,8 = V = 0. Then, we add several functions from the null space (see
Figures 6 (A), (C), and (E)) to G0,4,8 . The results are shown in Figures 6 (B), (D),
and (F). Keep in mind that all functions in the left column of Figure 6 generate
the zero potential and all functions in the right column of Figure 6 generate the
same forward solution V = T G0,4,8 . Similarly, we proceed in Figure 7, where linear
combinations of functions K(zi , ·), zi ∈ B \{0}, with

100 2n+1

K(zi , x) := (0.95)1+n G1,n,j (x)G1,n,j (zi ), x ∈ B \{0},
n=0 j=1
are added to the density D of the PREM model, see [18]. Again, K(zi , ·) can
be extended onto B, if ln ≥ 0 for all n ∈ N0 . Note that K(zi , ·) ∈ ker T for all
zi ∈ B\{0} such that, again, there is no difference between the potentials generated
by PREM (see Figure 7 (A)) and the potentials generated by the perturbed mass
densities in Figures 7 (B), (C), and (D).
Hence, the solution of the inverse problem from Equation (2) is not unique,
since we can always add functions from the null space to it without changing the
function V . In particular, Figure 7 shows that certain kinds of mass anomalies (in
the interior of the Earth) remain completely concealed if gravitational data are
used solely.
Now we can sum up our results and give an answer to the three questions
about the well-posedness of the problem posed in Section 3.
Theorem 4.2. Let the operator T : L2 (B) → L2 (S) be given by

T : D → D(x)k(x, ·) dx. (19)
B
with an integral kernel k : B × S → R of the form
∞
|x|li x y
k(x, y) := ci i+1 Pi · , x ∈ B \{0}, y ∈ S,
i=0 |y| |x| |y|
satisfying Assumption 3.1. Moreover, let the following three conditions be fulfilled
(by the function V ):
• The restriction V |ΩR of V is an L2 (ΩR )-function.
• The spherical harmonics coefficients Vn,j of V fulfil a summability condition
∞
2n+1

n2 (2ln + 3)R2n−2ln c−2
n
2
Vn,j < ∞.
n=0 j=1
cn =0
• The function V is harmonic in the exterior of B, that is, ΔV (y) = 0 for all
y ∈ R3 \B, and regular at infinity, that is, |V (y)| = O(|y|−1 ) and |∇V (y)| =
O(|y|−2 ) for |y| → ∞.
Then both inverse problems, which are, the recovery of D ∈ L2 (B) from either
given values of V |ΩR or the upward continued potential V |S are ill posed, since
their solutions are not unique. However, in both cases, the solution exists under
these conditions but is not stable.
The second condition in Theorem 4.2 is also known as the Picard condition.In
several cases, for example, the inverse gravimetric problem (i.e., ln = n, cn = γ
for all n ∈ N0 ), the Picard condition implies V |ΩR ∈ L2 (ΩR ). For the inverse
gravimetric problem the Picard condition is satisfied, since the (empirical) Kaula
rule of thumb holds:
2n+1

V |ΩR , Yn,j 2L2 (ΩR ) = O(ϑn+1 n−3 ), n → ∞,
j=1
for a constant ϑ ∈]0, 1[, see, for example, [28] or [44]. Note that the Picard condition
is necessary for the existence of the solution. Since this condition is not necessarily
satisfied by every V |ΩR ∈ L2 (ΩR ), also this criterion by Hadamard may be violated.
We want to discuss the instability of the solution in detail using the following
example.
Example 4.3. Let a family of functions be defined by

1 βn y
Vn (y) := Yn,1 , y ∈ S, for all n ∈ N0 .
n |y|n+1 |y|
Since { β1 Yn,1 ( β· )}n∈N0 is an L2 (S)-orthonormal system, we get
1
Vn L2 (S) =
→ 0 as n → ∞.
n
Hence, the norms build a null sequence. Using Equation (16), we see that
√ n
2n + 3 (2n + 1) β
Dn (x) := G0,n,1 (x)
4πR3/2 n R
yields T Dn = Vn in the case of ln = n, cn = 1. In addition, we obtain that the
sequence of norms diverges, since β > R and
√ n
2n + 3(2n + 1) β
Dn L2 (B) = G0,n,1 L2 (B)
4πR3/2 n R
√ n
2n + 3(2n + 1) β
= → ∞ as n → ∞.
4πR3/2 n R
Thus, small changes in the potential V yield large changes in the solution D and,
hence, the problem is not stable. Note that this instability is already given for the
case of terrestrial data, which means that it is not (only) caused by the instability
of the downward continuation.
4.3. Expansion of the solution in reproducing kernel based functions
In certain cases, it can be of interest to expand the unknown function D in terms
of appropriate reproducing kernels instead of orthonormal basis functions. Repro-
ducing kernels are localized in contrast to the global orthonormal basis functions
from the previous subsection (see also the paper by Freeden, Michel and Simons
in this handbook). In addition, the problems due to the discontinuity at the origin
can be avoided by using this approach. For a more general introduction into re-
producing kernels and reproducing kernel Hilbert spaces, see, for a general setting,
[2, 3, 5, 16], for reproducing kernel Hilbert spaces on the ball.
Let H := H((Am,n ), B) ⊂ L2 (B), with the real sequence (Am,n )m,n∈N0 , be

defined as
⎧ ⎫·H
⎨ ∞ 2n+1
⎬

H((Am,n ), B) := F ∈ L2 (B) A2m,n F, Gm,n,j 2L2 (B) <∞ ,
⎩ ⎭
m,n=0 j=1
with
∞ 2n+1

F 2H := A2m,n F, Gm,n,j 2L2 (B) , F ∈ H.
m,n=0 j=1
The inner product in H is then given by

∞ 2n+1

F, GH = A2m,n F, Gm,n,j L2 (B) G, Gm,n,j L2 (B) (20)
m,n=0 j=1
for all F , G ∈ H.
If the sequence (Am,n )m,n∈N0 fulfils a certain summability condition, see, for
more details, [34], then H is a reproducing kernel Hilbert space. Due to the property
of the sequence (Am,n )m,n∈N0 , the evaluation functional in H is continuous. The
reproducing kernel of H is given by K : (B \{0}) × (B \{0}) → R with
∞
2n+1

K(z, x) := A−2
m,n Gm,n,j (x)Gm,n,j (z), z, x ∈ B \{0}. (21)
m,n=0; j=1
Am,n =0
Again, in certain cases of Gm,n,j , the definition of K on B × B is valid.

The kernel K has the reproducing property, that is,
F, K(z, ·)H = F (z) for all F ∈ H and all z ∈ B \{0}.
In our setting, the first input argument z denotes the (fixed) centre of the kernel,
that is, the position in the ball where the kernel is located. Some examples of
reproducing kernels with the same centre and different sequences (Am,n )m,n∈N0
are plotted in Figure 8. As one can see, the discontinuity at the origin is, at least
visibly, smoothed away.
Let the set {y1 , . . . , y } ⊂ S, ∈ N, contain our measuring positions. We
define linear functionals by F ν F := (T F )(yν ) for ν = 1, . . . , . In other words,
the functionals F ν are the evaluations of our operator T applied to an (unknown)
function F at the measuring positions yν , ν = 1, . . . , . The data collected at the
sensor positions are given by vν = V (yν ). The functionals F ν are linear, since they
are the composition of the linear operator T and the linear evaluation functional.
If the function F is an element of the Sobolev space H((Am,n ), B) with a
sequence (Am,n )m,n∈N0 fulfilling the summability condition
∞
1
A−2
0,n < ∞,
n=0
(2n + 1)(2ln + 3)
1 1
z
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(a) h = 0.75 (b) h = 0.8
1 1
z
0 0
−1 −1
1 1
1 1
0 0
0 0
y −1 −1 x y −1 −1 x
(c) h = 0.9 (d) h = 0.92
Figure 8. Reproducing kernel K(z1 , ·) for several (Am,n )m,n∈N0 with
A−2
m,n = (Cn + 1)h
2(m+n)
δm,0 at a fixed centre z1 ∈ B\{0}, a sufficiently
large constant C, and the functions Gm,n,j in the case ln = n − 1.
then the functionals F ν are also continuous (with yν = rν ξν ) for ν = 1, . . . , , since
∞ ∞ 2n+1 2

ν 2 2
|F F | = |(T F )(yν )| = T F, Gm,n,j L2 (B) Gm,n,j (yν )
∞ 2n+1 m=0 n=0 j=1
2

2 c n R ln −1 A0,n
= (4π) F, G0,n,j L2 (B) γ Y (ξ )
(2n + 1)|rν |n+1 0,n A0,n
n,j ν
n=0 j=1
∞ 2n+1

∞ 2n+1 2
2 2 2 c n R ln −1
≤ (4π) F, G0,n,j L2 (B) A0,n γ Yn,j (ξν )
n=0 j=1 n=0 j=1
A0,n (2n + 1)|rν |n+1 0,n
2
∞
R 2n + 1
≤ (4πc)2 F 2H sup Rln −n
n∈N0 A2 (2n
n=0 0,n
2
+ 1) (2ln + 3) 4π
2 ∞
1
≤ 4πc2 R F 2H sup Rln −n 2 (2n + 1)(2l + 3) < ∞,
n∈N0 A
n=0 0,n n
due to (16), the Cauchy–Schwarz inequality, the definition of the inner product in
H in (20), and Assumption 3.1.
We can apply these functionals to the kernel with respect to z and obtain
the following result by using Equation (16) and the addition theorem for spherical
harmonics. The interchanging of limits (in the series) and the integral, which is
needed in this calculation, is allowed due to the previous estimates. Hence,
) ∞ *
|z|li z y
Fz K(z, x) =
ν
K(z, x) ci i+1 Pi · dz
B i=0
|y| |z| |y| y=yν
∞
2n+1

= A−2 ν
m,n Gm,n,j (x)Fz Gm,n,j (z)
m,n=0; j=1
Am,n =0
∞ 2n+1

−2 |x|ln x −1 4πcn Rln yν
= A0,n γ0,n ln Yn,j γ0,n Yn,j
n=0 j=1
R |x| (2n + 1)|yν |n+1 |yν |
∞
−2 |x|ln x yν
= A0,n cn Pn · .
n=0
|yν |n+1 |x| |yν |
It is known that we can construct an expansion for the solution, see [3, 23] by

D(x) = aν Fzν K(z, x). (22)
ν=1
Our aim is to determine the corresponding coefficients aν , ν = 1, . . . , . Applying

the functional on both sides, we obtain for ι = 1, . . . ,

Fxι D(x) = vι = aν Fxι Fzν K(z, x)
ν=1

∞

4πc2n R2ln yι yν
= aν A−2 −2
0,n γ0,n Pn · .
ν=1 n=0
2n + 1 |yι |n+1 |yν |n+1 |yι | |yν |
This linear system is uniquely solvable, which means that the expansion in (22)
is unique, if the linear and continuous functionals F ν , ν = 1, . . . , are linearly
independent, see [2]. Among all solutions D ∈ H with F ν D = vν for ν = 1, . . . , ,
the solution in (22) uniquely minimizes the norm · H induced by the inner
product in (20). These are basically the ideas of a spline interpolation method (for
further details, see [2] and [9]).
5. Constraints for the uniqueness of the solution

In the previous section, we have shown that we cannot expect a unique solution of
the Fredholm integral equation of the first kind stated in (2). Hence, in practice,
additional conditions are necessary to impose uniqueness. Some possible unique-
ness constraints are now discussed. The most approaches are generalizations of
the results in [33]. More precisely, we present the minimum norm condition, a
generalization of the harmonicity constraint, and the layer density constraint. In
addition, we discuss the surface density approach.
5.1. Minimum norm constraint

As we have seen, we are not able to obtain a uniquely determined solution without
additional assumptions or information. A widespread approach to force uniqueness
is the minimum norm condition (see, e.g., [40]). The following result is a gener-
alization of the theorem concerning the minimum norm solution of the inverse
gravimetric problem, see [33] and the references therein. Throughout this subsec-
tion, we assume that the conditions in Theorem 4.2 are fulfilled and, hence, a
solution of the inverse problem exists.
Recall Equation (11), which is repeated below for convenience:
R
(2n + 1)Rn
Vn,j = rln +2 Dn,j (r) dr, if cn = 0,
4πcn 0
Vn,j = 0 for all j = 1, . . . , 2n + 1 otherwise. Dn,j is originated by the (in L2 (B)

convergent) series
∞ 2n+1

x
D(x) = Dn,j (|x|)Yn,j .
n=0 j=1
|x|
The minimum norm conditionis fulfilled, if among all D ∈ L2 (B) with V =

B
D(x)k(x, ·) dx, we choose the one with the minimum (squared) norm

∞ 2n+1
R
2 2
D L2 (B) = (D(x)) dx = r2 (Dn,j (r))2 dr.
B n=0 j=1 0
If we minimize this expression, we obtain the following minimization problem for

each n ∈ N0 and j = 1, . . . , 2n + 1:
R
minimize r2 (Dn,j (r))2 dr,
0
R
2n + 1 n
subject to rln +2 Dn,j (r) dr = R Vn,j , if cn = 0.
0 4πcn
Note that the side condition drops out in the case cn = 0 such that the uncon-
strained minimizer Dn,j ≡ 0 occurs. With the substitution Fn,j (r) := rDn,j (r),
the problem above is equivalent to

R
minimize (Fn,j (r))2 dr,
0
R
2n + 1 n
subject to rln +1 Fn,j (r) dr = R Vn,j , if cn = 0.
0 4πcn
We now apply an orthogonal decomposition in L2 [0, R] to Fn,j in the sense that
R
Fn,j (r) = αn,j rln +1 + Hn,j (r), where 0 rln +1 Hn,j (r) dr = 0. With this ansatz,
our minimization problem reads
R
2
minimize αn,j r2ln +2 dr + Hn,j 2L2 [0,R] ,
0
R
2n + 1 n
subject to αn,j r2ln +2 dr = R Vn,j , if cn = 0.
0 4πcn
Since the side condition is independent of Hn,j , we see that Hn,j ≡ 0 yields the
unique minimum, for which we have
2n + 1 Rn
αn,j = (2ln + 3) Vn,j , if cn = 0
4πcn R2ln +3
and αn,j = 0 for all j = 1, . . . , 2n + 1, if cn = 0. We summarize our results in the
following theorem.
Theorem 5.1. Let the conditions

on V from Theorem 4.2 be fulfilled. Then, among
all D ∈ L2 (B) with V = B D(x)k(x, ·) dx, the L2 (B)-convergent series,
∞ 2n+1
l
2n + 1 n−ln −3 |x| n x
D(x) = (2ln + 3) R Vn,j ln Yn,j
n=0 j=1
4πcn R |x|
cn =0
∞ 2n+1
2ln + 3 2n + 1 n−ln
= R Vn,j G0,n,j (x), (23)
n=0 j=1
R3 4πcn
cn =0
is the unique minimizer of the functional

F (D) := (D(x))2 dx.
B
In the particular case of ln = n and cn = γ for all n ∈ N0 , it can be proven

that the harmonic solution is equivalent to the minimum norm solution, see [33].
This particular solution of the inverse gravimetric problem is then given by
∞ 2n+1
1 2n + 3 2n + 1
D(x) = Vn,j GI0,n,j (x), x ∈ B.
γ n=0 j=1 R3 4π
The convergence of the series in (23) can be proven using the orthonormality
of the Gm,n,j functions, since the Parseval identity yields
2ln + 3 2n + 1 2
∞ 2n+1

D 2L2 (B) = R2n−2ln Vn,j
2
.
n=0 j=1
R3 4πcn
cn =0
Comparing this with Theorem 4.2, we achieve that the series in (23) converges if
and only if V fulfils the Picard condition, that is,
∞ 2n+1
n2 (2ln + 3) 2
V < ∞. (24)
c2 R2(ln −n) j=1 n,j
n=0 n
cn =0
5.2. A generalization of the harmonicity constraint

In [33], the quasi-harmonic solution, which had already been discussed in the
literature, was seized on. In this case, functions of the kind x → |x|n+p Yn,j ( |x|
x
),
+
x ∈ B, for a fixed p ∈ R0 are used as basis functions. We consider here the
generalized case of a basis {Bn,j }n∈N0 ,j=1,...,2n+1 given by

|x|kn x
Bn,j (x) := kn +1 Yn,j , n ∈ N0 , j = 1, . . . , 2n + 1
R |x|
with a preliminarily chosen sequence (kn )n∈N0 ⊂ R and the additional condition
that 2kn + 3 > 0 for all n ∈ N0 . This condition guarantees that these functions
have a finite L2 (B)-norm. The orthogonality is a direct consequence of the L2 (Ω)-
orthogonality of the spherical harmonics Yn,j , since
k k
|x| n x |x| ν x
Bn,j , Bν,ι L2 (B) = Y
kn +1 n,j
Y
kν +1 ν,ι
dx
B R |x| R |x|
R 2kn +2
r
= 2kn +2
dr δn,ν δj,ι
0 R
R2kn +3
= δn,ν δj,ι
(2kn + 3)R2kn +2
R
= δn,ν δj,ι .
2kn + 3
In the case kn = n, the subspace spanned by this basis is the set of all harmonic
functions and in the case kn = n + p we get the quasi-harmonic setting.
In contrast to the previous subsection, we have to assume slightly different
properties of V . However, note that Assumption 3.1 is still valid.
Assumption 5.2. We suppose that

• the restriction V |ΩR of V is an L2 (ΩR )-function,
• the summability condition

∞
2n+1
n2 (ln + kn + 3)2 2
R2n−2ln V <∞
n=0
c2n (2kn + 3) j=1 n,j
cn =0
is fulfilled,
• V is harmonic in the outer space, that is, ΔV (y) = 0 for all y ∈ R3 \B,
• V is regular at infinity.
With the orthogonal basis {Bn,j }n∈N0 ,j=1,...,2n+1 , the density D can be rep-
resented by the expansion
∞ 2n+1
2kn + 3
D(x) = dn,j Bn,j (x), x ∈ B \{0}, (25)
n=0 j=1
R
in the sense of L2 (B). In accordance with the notations above, we have

2kn + 3 rkn
Dn,j (r) = dn,j , r ∈ [0, R].
R Rkn +1
Thus, the relation between the Fourier coefficients of V and Dn,j in (11) becomes
for all j = 1, . . . , 2n + 1
R
(2n + 1)Rn 2kn + 3 rln +kn +2
Vn,j = dn,j dr
4πcn 0 R Rkn +1

2kn + 3 Rln +kn +3
= dn,j
R (ln + kn + 3)Rkn +1

2kn + 3 Rln +2
= dn,j , if cn = 0, (26)
R (ln + kn + 3)
and Vn,j = 0, if cn = 0. Solving (26) for dn,j and inserting the result in (25), we
obtain
∞ 2n+1
2kn + 3 |x|kn x
D(x) = dn,j +1
Yn,j
n=0 j=1
R R k n |x|
∞ kn 2n+1

2n + 1 n−ln −2 |x| x
= (ln + kn + 3)R +1
Vn,j Yn,j + D̃
n=0
4πcn R kn
j=1
|x|
cn =0
∞ 2n+1

2n + 1
= (ln + kn + 3)Rn−ln −2 Vn,j Bn,j (x) + D̃,
n=0
4πcn j=1
cn =0
· 2
where D̃ ∈ span{Bn,j | n ∈ N0 with cn = 0, j = 1, . . . , 2n + 1} L (B) is arbitrary.
The convergence of the series is guaranteed by the summability conditions on V .
Summarizing these results, we get the next theorem.
Theorem 5.3. Let cn = 0 for all n ∈ N0 , and let Assumptions 3.1 and 5.2 be
fulfilled. Then the unique solution D ∈ U , where the L2 (B)-subspace U has the
basis {Bn,j }n∈N0 ,j=1,...,2n+1 , of the inverse problem

D(x)k(x, y) dx = V (y) in R3 \B,
B
with (x, y) ∈ dom(k) is given by
∞ 2n+1
2n + 1 |x|kn x
D(x) = (ln + kn + 3)Rn−ln kn +3 Vn,j Yn,j ,
n=0
4πcn R j=1
|x|
in the sense of L2 (B).

In [33], the biharmonic solution was also considered. In this case, the needed
radial basis is given by the sum of two radial parts. An approach for a general-
ization of this ansatz is given by the sum of K ∈ N different radial parts, that

is, {( Ki=1 | · |
ki,n
)Yn,j ( |·|· )}n∈N0 ,j=1,...,2n+1 . However, without any additional in-
formation, a unique solution cannot be obtained in this case (see also the result
for the biharmonic solution in [33]).
5.3. Layer density constraint
As we have seen above, the non-uniqueness is primarily a matter of the radial
parametrization of the solution D. For this reason and in view of the fact that,
for example, lithospheric heterogeneities are particularly interesting with respect
to their lateral structure, we consider here the (thin) spherical shell
! "
Ω[τ,τ +ε] := x ∈ R3 : 0 < τ ≤ |x| ≤ τ + ε ≤ R ,
for τ > 0 and ε > 0. We are interested in finding a solution D which consists
of purely laterally inhomogeneous anomalies in Ω[τ,τ +ε] . This kind of uniqueness
constraint was, for example, used in [23] for the inverse magnetic problem.
For the layer density constraint, we assume that the density D ∈ L2 (B) has
(again) the form
∞ 2n+1

x
D(x) = Dn,j (|x|)Yn,j , x ∈ B, (27)
n=0 j=1
|x|
where now
Dn,j (r) := κdn,j χ[τ,τ +ε] (r), r ∈ [0, R], (28)
for all n ∈ N0 , j = 1, . . . , 2n+1, and χ is the characteristic function (i.e., χA (x) = 0
if x ∈ A and χA (x) = 1 if x ∈ A). The normalization constant κ is chosen as

3
κ := .
(τ + ε)3 − τ 3
Assumption 5.4. For the function V , we now assume that
• the restriction V |ΩR of V is an L2 (ΩR )-function,
• the summability condition

∞
2n+1

n2 ln2 R2n
+3 2 2
V2 <∞
n=0
((τ + ε) n − τ n ) cn j=1 n,j
l l
cn =0
is fulfilled,
• V is harmonic in the outer space, that is, ΔV (y) = 0 for all y ∈ R3 \B,
• V is regular at infinity.
Using (11) and the desired representation of D, we have

R
(2n + 1)Rn
Vn,j = rln +2 Dn,j (r) dr
4πcn 0
R
=κ rln +2 dn,j χ[τ,τ +ε] (r) dr
0
(τ + ε)ln +3 − τ ln +3
= κdn,j .
ln + 3
This yields, for all j = 1, . . . , 2n + 1,
(2n + 1)Rn ln + 3
κdn,j = Vn,j , if cn = 0,
4πcn (τ + ε)ln +3 − τ ln +3
and Vn,j = 0, if cn = 0. We insert this in Equations (27) and (28) and directly
obtain, for all x ∈ B,
∞ 2n+1

x
D(x) = dn,j κχ[τ,τ +ε](|x|)Yn,j
n=0 j=1
|x|
∞ 2n+1
(2n + 1)(ln + 3)
Rn x
= Vn,j χ[τ,τ +ε] (|x|)Yn,j + D̃,
n=0 j=1
4πcn (τ + ε)ln +3 − τ ln +3 |x|
cn =0
where

·L2 (B)
·
D̃ ∈ span Dn,j (| · |)Yn,j n ∈ N0 with cn = 0, j = 1, . . . , 2n + 1
|·|
can be chosen arbitrarily.
Theorem 5.5. Let cn = 0 for all n ∈ N0 and let Assumptions 3.1 and 5.4 be
fulfilled. Then the unique solution under the layer density constraint is given by
∞ 2n+1
Rn (2n + 1)(ln + 3)
x
D(x) = V χ
ln +3 − τ ln +3 n,j [τ,τ +ε]
(|x|)Yn,j (29)
n=0 j=1
4πc n (τ + ε) |x|
in the sense of L2 (B).

Moreover, under the conditions in Assumption 5.4, the corresponding poten-

tial V possesses the following outer harmonics expansion
∞
4πcn ln +3 ln +3
−n−1 2n+1
y
V (y) = κ (τ + ε) −τ |y| dn,j Yn,j .
n=0
(2n + 1)(l n + 3) j=1
|y|
This series fulfils the condition of Assumption 5.4, that is, V |ΩR ∈ L2 (ΩR ):
∞
2 2
4πcn ln +3 ln +3
2 −2n 2n+1

V |ΩR L2 (ΩR ) = κ (τ + ε) −τ R d2n,j
n=0
(2n + 1)(l n + 3) j=1
∞ l +3 2 2n+1
R n
+R l n +3
≤ 16π 2 c2 κ2 2 3 2n
d2n,j
n=0
(2n + 1) (l n + 3) R j=1

∞ 2n+1

2 2 2 R2ln +6−2n
≤ 64π c κ sup d2n,j < ∞.
n∈N0 (2n + 1)2 (ln + 3)2 n=0 j=1
For this estimate, we used the boundedness of the sequence (cn )n∈N0 (given by
Assumption 3.1, item 1), the boundedness of the supremum in the latter estimate
(given by Assumption 3.1, items 2 and 3), and the square-integrability of D.
5.4. Surface density
In inverse gravimetry, in particular, it is reasonable to consider a surface density
instead of a density on the entire ball B. In a time-variable gravity field (with
relatively short time scales) most of the changes occur on the (Earth’s) surface or
at least on layers very close to it. So, if one is interested in anomalies as devia-
tions from a reference model, which could be an annual mean, for instance, these
anomalies can be typically found on the surface of the underlying body.
So far, in our general setup, we have

V (y) = (T D)(y) = D(x)k(x, y) dx. (30)
B
Since the operator T is linear and continuous, we can also read the equation above
in distributional sense. For the mathematical theory of distributions and, in this
context, the definition of test functions, the reader is referred to [27]. In other
words, we can look at Equation (30) as an application of a regular distribution D
applied to the kernel k, that is
V (y) = (T D)(y) = D, k(·, y). (31)
Actually, we have a regular distribution D with
Dϕ := D, ϕ
for all test functions1 ϕ, which is uniquely determined by the function D and
vice versa (at least almost everywhere). Thus, the distribution can be, in fact,
1 Actually, the function k(·, y) is not a test function, but the domain of D can be extended such
that Dk(·, y) makes sense and equals (T D)(y).

represented by the function D itself and the distinction is commonly omitted.

Now, one can think of replacing the regular distribution and also allow singular
distributions. For our purposes, a very useful singular distribution is F δΩR , which
is a variation of the well known delta distribution and is given by

F δΩR , ϕ := F (x)ϕ(x) dω(x),
ΩR
for an arbitrary, over ΩR square-integrable, function F and for every test function
ϕ. In that case, we have (cf. Equation (31))

Ṽ (y) := DδΩR , k(·, y) = D(x)k(x, y) dω(x).
ΩR
Conclusively, with our previous considerations, we get
∞
R ln x y
Ṽ (y) = cn n+1 D(x)Pn · dω(x)
n=0
|y| ΩR |x| |y|
∞
R ln y
= cn n+1 D(Rξ)Pn ξ · R2 dω(ξ).
n=0
|y| Ω |y|
With the addition theorem for spherical harmonics and the ansatz (8), it follows
that
∞ 2n+1
4πcn Rln +2
y
Ṽ (y) = Y
n+1 n,j
D(Rξ)Yn,j (ξ) dω(ξ)
n=0 j=1
2n + 1 |y| |y| Ω
∞ 2n+1
4πcn n
ln −n+2 R 1 y
= R Dn,j (R) Yn,j .
n=0 j=1
2n + 1 |y| |y| |y|
Consequently, we find the Fourier coefficients
4πcn ln −n+2
Ṽn,j = R Dn,j (R)
2n + 1
which in other words means that, for cn = 0,
(2n + 1)Rn
Ṽn,j = Rln +2 Dn,j (R). (32)
4πcn
As we see, this problem is again uniquely solvable (if cn = 0 for all n ∈ N0 ) and
in the particular case of the inverse gravimetric problem, the coefficients read
2n + 1
Ṽn,j = R2 Dn,j (R). (33)
4πγ
Theorem 5.6. Let Dn,j be given according to (32) and cn = 0 for all n ∈ N0 .
Further, let Ṽ be a harmonic function in the exterior of ΩR which is regular at
infinity with Ṽ |ΩR ∈ L2 (ΩR ) and
∞ 2n+1
n2 R2n−2ln
2
Ṽn,j < ∞.
n=0 j=1
c2n
Then a distributional solution of the Fredholm integral equation of the first kind in
(2) is given by
∞ 2n+1
(2n + 1)Rn−ln ·
DδΩR = Ṽn,j Yn,j δΩR .
n=0 j=1
4πR2 cn R
In the inverse gravimetric problem, as the typical application of the surface

density approach, we have the following setting. Let ρ̄ : B → R be a density given
by an arbitrary reference model of the Earth, for example, the Preliminary Refer-
ence Earth Model (PREM), see [18]. The corresponding gravitational potential is
given by
ρ̄(x)
V̄ = γ dx
B |x − ·|
and describes a part of the potential that does not change in the associated time
span. The entire measured potential is given by V = V̄ + Ṽ , where Ṽ are the
relevant occurring changes in the gravitational potential. That is, we are here
looking for a surface density σ : ΩR → R with

σ(x)
Ṽ = V − V̄ = γR2 dω(x),
ΩR |x − ·|
which causes these changes of the potential. By virtue of Equation (33), we know
that the Fourier coefficients of the surface density are given by
(2n + 1)
σn,j = Ṽn,j (34)
4πγR2
for all n ∈ N0 and all j = 1, . . . , 2n + 1. Chao [12] also proved that this problem is
uniquely solvable. The obtained formula (34) coincides with the formulae which are
commonly used in geodesy for a surface density ansatz or thin layer assumption,
respectively, as originally proposed in [52].
6. Conclusions
We observed similarities between the inverse gravimetric and the inverse magnetic
problem by considering both as particular cases of a kind of a master inverse prob-
lem. With this approach, a larger class of data inversion problems can be analyzed
and solved all at once. A particular focus of the paper was the complete analysis of
the non-uniqueness of the solution of all inverse problems of the investigated type.
This analysis was based on something like a fundamental equation for the Fourier
coefficients of the given data and the solution. The construction of a particular
and appropriate orthonormal system on the ball enabled us to further understand
the relation of the solution and the data. With this basis system and an adequate
expansion in the data space, we characterized the null space of the Fredholm in-
tegral operator of the first kind in detail and calculated the singular system. Such
a knowledge is an essential prerequisite for a series of regularization methods for
inverse problems.
Furthermore, using the derived singular value decomposition, we also proved

that this kind of inverse problem is unstable, that is, the inverse operator is un-
bounded. It also turned out that all considered problems have in common that most
of the radial information gets lost. The ill-posedness of the considered problems
is severely aggravated by the fact that the null space of the operator is infinite-
dimensional, and, hence, the solution of the inverse problem is not unique. For
this reason, we discussed four different additional conditions in order to obtain a
unique solution: the minimum norm condition, a generalization of the harmonicity
constraint, the layer density condition, and the surface density approach. In the
particular case of the inverse gravimetric problem, our results coincide with the
corresponding well-known results and in the case of the inverse magnetic problem,
we found new results.
References
[1] Abramo, L.R., Reimberg, P.H., Xavier, H.S. (2010) CMB in a box: causal structure
and the Fourier–Bessel expansion. Phys. Rev. D 82:043510
[2] Amirbekyan, A. (2007) The Application of Reproducing Kernel Based Spline Ap-
proximation to Seismic Surface and Body Wave Tomography: Theoretical Aspects
and Numerical Results. Ph.D.-thesis, University of Kaiserslautern, Geomathematics
Group, https://kluedo.ub.uni-kl.de/frontdoor/index/index/docId/1872
[3] Amirbekyan, A., Michel, V. (2008) Splines on the three-dimensional ball and their
application to seismic body wave tomography. Inverse Probl. 24:015022
[4] Anger, G. (1990) Inverse Problems in Differential Equations. Akademie-Verlag,
Berlin
[5] Aronszajn, N. (1950) Theory of reproducing kernels. Trans. Am. Math. Soc. 68:337–
404
[6] Ballani, L., Stromeyer, D. (1992) On the structure of uniqueness in linear inverse
source problems, Vieweg, Braunschweig, pp. 85–98. Theory and Practice of Geophys-
ical Data Inversion
[7] Ballani, L., Engels, J., Grafarend, E.W. (1993) Global base functions for the mass
density in the interior of a massive body (Earth). Manuscr. Geodaet. 18:99–114
[8] Barzaghi, R., Sansò, F. (1986) Remarks on the inverse gravimetric problem. Boll.
geod. e sci. affini XLV:203–216
[9] Berkel, P. (2009) Multiscale Methods for the Combined Inversion of Normal Mode
and Gravity Variations. Ph.D.-thesis, University of Kaiserslautern, Geomathematics
Group
[10] Center for Space Research (2002) GRACE.
http://www.csr.utexas.edu/grace/overview.html, [Online; accessed 30-August-
2015]
[11] Chao, B.F., Dehant, V., Gross, R.S., Ray, R.D., Salstein, D., Watkins, M., Wilson,
C. (2000) Space geodesy monitors mass transports in global geophysical fluids. EOS
81:247–250
[12] Chao, B.F. (2005) On inversion for mass distribution from global (time-variable)
gravity field, J. Geodyn., 39:223–230
[13] Dassios, G., Fokas, A.S. (2009) Electro-magneto-encephalography for a three-shell

model: dipoles and beyond for the spherical geometry. Inverse Probl. 25:035001
[14] Dassios, G., Fokas, A.S. (2013) The definite non-uniqueness results for deterministic
EEG and MEG data. Inverse Probl. 29:065012
[15] Dassios, G., Fokas, A.S., Kariotou, F. (2005) On the non-uniqueness of the inverse
MEG problem. Inverse Probl. 21:L1–L5
[16] Davis, P.J. (1975) Interpolation and Approximation. Dover Publications, New York
[17] Dufour, H.M. (1977) Fonctions orthogonales dans la sphère. Résolution théorique du
problème du potential terrestre. B. Geod. 51:227–237
[18] Dziewonski, A.M., Anderson, D.L. (1981) Preliminary reference Earth model.
Physics of the Earth and Planetary Interiors 25:297–356
[19] Fokas, A.S. (2009) Electro-magneto-encephalography for a three-shell model: dis-
tributed current in arbitrary, spherical and ellipsoidal geometries. J. R. Soc. Interface
6:479–488
[20] Fokas, A.S., Kurylev, Y. (2012) Electro-magneto-encephalography for the three-shell
model: minimal L2 -norm in spherical geometry. Inverse Probl. 28:035010
[21] Fokas, A.S., Gel-fand, I.M., Kurylev, Y. (1996) Inversion method for magnetoen-
cephalography. Inverse Probl. 2:L9–L11
[22] Fokas, A.S., Kurylev, Y., Marinakis, V. (2004) The unique determination of neuronal
currents in the brain via magnetoencephalography. Inverse Probl. 20:1067–1082
[23] Fokas, A.S., Hauk, O., Michel, V. (2012) Electro-magneto-encephalography for the
three-shell model: numerical implementation via splines for distributed current in
spherical geometry. Inverse Probl. 28:035009
[24] Freeden, W., Gutting, M. (2013) Special Functions of Mathematical (Geo-)Physics.
Birkhäuser, Basel
[25] Geselowitz, D.B. (1970) On the magnetic field generated outside an inhomogeneous
volume conductor by internal current sources. IEEE Trans. Magn. 6:346–347
[26] Han, S., Shum, C., Bevis, M., Ji, C., Kuo, C. (2006) Crustal dilatation observed by
GRACE after the 2004 Sumatra-Andaman earthquake. Science 313:658–662
[27] Hörmander, L. (1983) The Analysis of Linear Partial Differential Operators I.
Springer, Berlin
[28] Kaula, W.M. (1966) Theory of Satellite Geodesy. Blaisdell, Waltham
[29] Kusche, J., Schrama, E. (2005) Surface mass redistribution inversion from global
GPS deformation and Gravity Recovery and Climate Experiment (GRACE) gravity
data. J. Geophys. Res. 110:B09409
[30] Leistedt, B., McEwen, J.D. (2012) Exact wavelets on the ball. IEEE Trans. Signal
Process 60:6257–6269
[31] Michel, V. (2005) Wavelets on the 3-dimensional ball. Proc. Appl. Math. Mech.
5:775–776
[32] Michel, V. (2013) Lectures on Constructive Approximation. Fourier, Spline, and
Wavelet Methods on the Real Line, the Sphere, and the Ball. Birkhäuser, Boston
[33] Michel, V., Fokas, A.S. (2008) A unified approach to various techniques for the non-
uniqueness of the inverse gravimetric problem and wavelet-based methods. Inverse
Probl. 24:045019
[34] Michel, V., Orzlowski, S. (2016) On the null space of a class of Fredholm integral
equations of the first kind. J Inverse Ill-Posed Probl 24:687–710
[35] Moritz, H. (1990) The Figure of the Earth. Theoretical Geodesy of the Earth’s
Interior. Wichmann Verlag, Karlsruhe
[36] Nikiforov, A.F., Uvarov, V.B. (1988) Special Functions of Mathematical Physics. A
Unified Introduction with Applications. Birkhäuser, Basel
[37] Olsen, N., et al (2013) The Swarm Satellite Constellation Application and Research
Facility (SCARF) and Swarm data products. Earth Planets Space 65:1189–1200
[38] Pizzetti, P. (1910) Intorno alle possibili distribuzioni della massa nell’interno della
terra. Annali di Mat., Milano XVII:225–258
[39] Plonsey, R. (1969) Biomagnetic Phenomena. McGraw-Hill, New York
[40] Rieder, A. (2003) Keine Probleme mit Inversen Problemen. Vieweg, Wiesbaden
[41] Rubincam, D.P. (1979) Gravitational potential energy of the Earth: a spherical har-
monics approach. J. Geophys. Res.-Sol. Ea 84:6219–6225
[42] Rudin, W. (1991) Functional Analysis, 2nd edn. McGraw-Hill, Inc.
[43] Sabaka, T., Olsen, N. (2006) Enhancing comprehensive inversions using the Swarm
constellation. Earth Planets Space 58:371–395
[44] Sansò, F., Rummel, R. (1997) Geodetic boundary value problems in view of the one
centimeter geoid, Lect. Notes Earth Sci., vol 65. Springer, Berlin, Heidelberg
[45] Sarvas, J. (1987) Basic mathematical and electromagnetic concepts of the biomag-
netic inverse problem. Phy. Med. Biol. 32:11–22
[46] Schnetzler, C. (1985) An estimation of continental crust magnetization and sus-
ceptibility from Magsat data for the conterminous United States. J. Geophys. Res.
90:2617–2620
[47] Stokes, G.G. (1867) On the internal distribution of matter which shall produce a
given potential at the surface of a gravitating mass. Proc. Royal Soc. 15:482–486
[48] Stromeyer, D., Ballani L. (1984) Uniqueness of the inverse gravimetric problem for
point mass models. Manuscr. Geodaet. 9:125–136
[49] Szegö, G. (1975) Orthogonal Polynomials. American Mathematical Society, Provi-
dence, Rhode Island
[50] Thébault, E., Purucker, M., Whaler, K.A., Langlais, B., Sabako, T.J. (2010) The
magnetic field of the Earth’s lithosphere. Space Sci. Rev. 155:95–127
[51] Tscherning, C.C. (1996) Isotropic reproducing kernels for the inner of a sphere or
spherical shell and their use as density covariance functions. Math. Geol. 28:161–168
[52] Wahr, J., Molenaar, M., Bryan, F. (1998) Time variability of the Earth’s gravity
field: hydrological and oceanic effects and their possible detection using GRACE. J.
Geophys. Res. 103B(12):30205–30229
[53] Wouters, B., Chambers, D., Schrama, E.J.O., (2008) GRACE observes small-scale
mass loss in Greenland. Geophys. Res. Lett. 35:L20501
Sarah Leweke and Volker Michel Roger Telschow

University of Siegen Computational Science Center
Geomathematics Group University of Vienna
Walter-Flex-Str. 3 Oskar Morgenstern-Platz 1, Room 07.131
D-57068 Siegen, Germany A-1090 Vienna, Austria
emails: leweke@mathematik.uni-siegen.de e-mail: roger.telschow@univie.ac.at
michel@mathematik.uni-siegen.de

Index
Abel–Poisson kernel, 352, 531, 548, 610, Hilbert ∼, 173
775 property, 323
Abel–Poisson scaling function, 383 Schauder ∼ in Banach space, 169
absolute value, 240 system, 653, 893
addition theorem, 42–44, 566, 568, 761, Bayes estimate, 256, 294
783, 784, 793 bell curve, 80, 94
admissible, 359, 364, 368, 586 Beltrami operator, 565, 821, 823–827, 829,
aggregation, 856, 864, 865, 870 830, 832, 834, 838–840
linear ∼, 857, 859 Bessel’s inequality, 172
method, 863, 864 best approximate solution, 581
optimal ∼, 860 best worst case error, 243
Akaike information criterion (AIC), 630 bias, 626
algorithm Biot–Savart operator, 887
efficient ∼, 324 Bjerhammar ball, 518
fast multipole ∼, see fast Boscovic–Laplace method, 79
tree ∼, see tree Bouguer anomaly, 693
altimetry, 316, 319 Bouguer correction, 693
anharmonic, 650 boundary value problem, 69
basis, 655 discrete version, 606
function, 650, 885, 897 for the Beltrami operator, 829
anomalous potential, 67, 413 geodetic ∼, see geodetic
anomaly Molodensky ∼, see Molodensky
Bouguer ∼, 693 Neumann ∼, see Neumann
free air ∼, 693 oblique ∼, see oblique
gravity ∼, see gravitiy Stokes ∼, 69
potential ∼, 67 bounded sesquilinear form, see form,
antenna problem, 19 continuous sesquilinear ∼
multivariate ∼, 19 bridge between
spline interpolation ∼, 19 least squares solutions and
aperture distribution, 19 pseudoinverses, 94
approximate identity, 130, 799 mathematics and geodesy, xii, 5
approximate right inverse, 237 Bruns’s formula, 697
approximation error, 583 Bruns’s relation, 472
Aronszajn’s theorem, 195 Bubnov–Galerkin method, 268, 271, 272
Arzelà–Ascoli theorem, 114, 216 convergence of the ∼, 273
asymptotic regularization, 221, 247
Cauchy sequence, 168
Backus–Gilbert method, 667 Cauchy–Kovalevskaya theorem, 194
Bakushinskii veto, 625 Cauchy–Schwarz inequality, 172
balancing principle, 628 centrifugal potential, 62, 564, 689
ball, 168 Chebyshev principle, 83, 88
bandlimited, 323, 757 Christoffel–Darboux formula, 24
kernel functions, 349 circle problem, 12, 13
scaling functions, 377 classification
signal, 17 Hadamard, 205
wavelets, 412 Nashed, 205
bandpass, 371, 388 closed, 525, 527, 528, 545
filter, 324, 432 closure, 527
basis coherent state, 780
922 Index
combination decomposition, 322

of satellite and ground models, 865, 871 decorrelation, 323
of satellite models, 857 deflection of the vertical, 70, 71
compact operator, 114, 215, 655 dense subset, 168
compactness, 228 derivative
complete, 525, 527 directional ∼, 177
function systems, 838 divergence, 177
orthogonal anharmonic systems, 655 gradient, 177
system, 169 Laplace operator, 177
completeness, 168, 525 partial ∼, 177
property, 522 strong ∼, 176
concretization, x weak ∼, 184
condition detail space, 373, 376, 390, 401
equation, 83 determination of π, 11
Marussi ∼, 64 dilation operator, 372
minimum norm ∼, see minimum dilogarithm, 794
oblique boundary ∼, see oblique Dirac
Picard ∼, see Picard delta distribution, 182
semi-frame ∼, 808 family, 797, 798
summability ∼, 608 kernel, see kernel
cone property, 179 direct problem, 104, 204
conformal mapping, 143 direction of the vertical, 461
contamination error, 226 Dirichlet
continuous Green function, 832, 833, 837, 850
Hölder ∼, 178 problem, 532, 822, 830
Lipschitz ∼, 178 integral ∼, 830
convergence theorem, 808 Runge basis, 533
convolution, 183, 276 discovery of Ceres, 2, 77
core, 655 discrepancy principle, 249, 627
correction Morozov’s ∼, 250
Bouguer ∼, 693 distribution, see also space of
drift ∼, 692 distributions, 649
Eőtvős ∼, 693 Dirac ∼, 650
isostatic ∼, 692 equi∼, 51
terrain ∼, 692 regular ∼, 182, 650
tidal ∼, 693 tempered ∼, 191
covariance, 83, 87 disturbing
criterion gravity field, 695
Akaike information ∼, 630 potential, 67, 316, 340, 343, 439, 695,
Hadamard’s ∼, see Hadamard’s criteria 714
Picard ∼, 584 divergence, see also derivative
quasi-optimality ∼, 628 divergence theorem, 54
cubed sphere, 787 domain, 170
cubic polynomial scaling function, 379 bounded ∼, 494
curve potentials, 827 outer ∼, 494
downward continuation, 562, 564, 572
de la Vallée Poussin drift correction, 692
-type singular integral, 662 duality, 175
generator, 131 product, 496
kernel, 662
mollifier method, 662 Earth gravity, 726
wavelet, 131 model 2008, 865, 876–878
Index 923
model 96, 863 positive definite ∼, 172

efficient algorithm, 324 sesquilinear ∼, 171
eigenspectrum, 24, 35 symmetric ∼, 171
ellipsoidal formula
orthogonal coordinates, 146 Bruns’s ∼, 697
shape of the Earth, 4 Christoffel–Darboux ∼, 24
Eőtvős correction, 693 Euler summation ∼, see Euler
equation Gaussian n-point ∼, see Gaussian
condition ∼, 83 Geselowitz’ ∼, 888
ill-conditioned matrix ∼, see integral ∼ for the Legendre (Green)
ill-conditioned function, 27
Maxwell’s ∼, 887 Rodriguez ∼, 22
normal ∼, see normal forward problem, 204
partial differential ∼, see partial Fourier
Poisson ∼, see Poisson coefficients, 172, 566
pseudodifferential ∼, see on a sphere, 607
pseudodifferential expansion
refinement ∼, see refinement Legendre ∼, 758
equidistribution, 51 spherical harmonics, 755
equipotential surface, 689 series, 566, 571, 574, 577
error transform
approximation ∼, 583 on L2 , 191
best worst case ∼, 243 on rapidly decreasing functions, 190
contamination ∼, 226 on tempered distributions, 191
mean square ∼, 78 frame, 323
propagated noise ∼, 626 Fredholm integral
reconstruction ∼, 246 equation of the 1st kind, 904, 908, 916
regularization ∼, see regularization operator of the 1st kind, 916
total ∼, 583 free air anomaly, 693
worst case ∼, 243 functional, 169
Euler summation formula, 36, 37 functional matching pursuit, 806
exponential fundamental
kernel, 351 equations of physical geodesy, 700
scaling function, 382 for the sphere, 701
to exponential, 621 solution, 525
to local, 621 linearly regularized ∼, 709
exterior system, 527, 596
oblique derivative problem, 689
Stokes problem, 69 Galerkin method, 267, 268
Gauss’s error law, 81
fast multipole algorithm, 619 Gauss’s theorem, 54
break-even points of the ∼, 623 Gauss–Legendre sampling, 23, 24
truncation degree of the ∼, 622 Gauss–Seidel method, 283
fast multipole method, 612 Gauss–Weierstraß scaling function, 383
filter qualification, 246 Gaussian
filter regularization, 244 adjustment, 81
filtering, 324, 435 circle problem, 8
form distribution function, 80
bilinear ∼, 171 geometry, 147
coercive sesquilinear ∼, 172 integration, 29
continuous sesquilinear ∼, 172 polynomial exactness of the ∼, 29
hermitian ∼, 171 remainder estimation of the ∼, 32
924 Index
n-point formula, 29, 32 gravity, 315, 563, 689

probability distribution, 79 acceleration, 689
process, 256, 293 anomaly, 67, 318, 340, 343, 693, 698
quadrature, 34 field, 322
convergence of the ∼, 33 disturbance, 698
summability, 15 disturbance vector, 67
GCV, 630 field, 316, 318, 319, 461, 564
generalized cross-validation, 630 intensity, 62, 689
modified GCV method, 631 potential, 61, 62, 460, 564, 689
Robust GCV method (RGCV), 630 normal ∼, 695
Strong robust GCV (R1 GCV), 630 vector, 62
Gelfand triple, 272, 496, 506, 507 normal ∼, 471
generalized Green’s formulas, 824, 825, 831, 832, 837
cross-validation, 630 Green’s functions, 821, 822, 826, 827, 832
Fourier expansion Beltrami operator of ∼, 793
Dirichlet problem of the ∼, 532 Legendre operator of ∼, 24
Molodensky problem of the ∼, 554 ground data, 857, 865, 872, 873, 877
oblique derivative problem of the ∼, ground model, 865, 871
546
inverse, 89, 92, 95, 97, 106, 209, 212, Haar
219, 257 -type mollification, 646
Robinson’s interpretation of the ∼, 88 -type singular integral, 657
generator, 371, 385, 388, 395, 399, 406 function, 660
geodesic, 142 kernel, 776, 777
geodetic smoothed ∼, 354
boundary value problem, 62, 69 wavelet, 731
bvp Hadamard’s classification, 104, 205
linearized ∼, 470 Hadamard’s criteria for well-posed and
scalar ∼, 463 ill-posed problems, 889, 903, 904
vectorial ∼, 461 hard thresholding, 136
circuit, x hardened balancing principle, 628
observables, 318 Hardy’s conjecture, 12
geoid, 64, 564, 690, 697 Hardy–Hodge decomposition, 822, 835,
geoidal height, 696 837, 838
geoidal undulation, 64, 71, 344, 696 Hardy–Landau identity, 13
geomagnetism, 4 harmonic, 325
geostrophic ocean flow, 822, 834, 836, 841, continuation, 576
845, 846 expansion, 473
Geselowitz’ formula, 888 function, 650
GOCE, 562, 564 solution, 654, 657
gradient, see also derivative splines, 609
gradient method, 289 Dirichlet problem of ∼, 541
gravimetry, 316, 319, 321, 322, 562, 691 Molodensky problem of ∼, 555
mollifier solution, 657 oblique derivative problem of ∼, 547
reproducing kernel structure, 670 harmonicity constraint, 885, 908, 910, 917
spectral solution, 652 Harnack’s convergence theorem, 326
spline solution, 677, 678 Hausdorff measure, see also measurable
gravitation of Earth’s body, 726 Helmert’s definition, x
gravitational Helmholtz
field, 316, 318, 319, 321, 411, 562 decomposition, 822, 835–837
potential, 61, 562, 564, 887, 902, 916 operator, 37
tensor, 578 scalars, 836–838
Index 925
Hesse matrix, 562 Kaula’s rule of thumb, 904

Hesse tensor, 578, 587 Kelvin transformation, 504, 505, 530, 612
Hölder’s inequality, 185 kernel
Abel–Poisson ∼, see Abel
ill-conditioned matrix equation, 96, 206 bandlimited ∼, 757
ill-posed problem, 104, 106, 581, 585, 788, de la Vallée Poussin ∼, 662
804, 806, 883, 890, 917 Dirac ∼, 778
ill-posedness, 104, 106, 357 exponential ∼, 351
inequality Gauss’s ∼, 779
Bessel’s ∼, 172 Haar ∼, see Haar
Cauchy–Schwarz ∼, 172 Legendre ∼, see Legendre
Hölder’s ∼, 185 locally supported ∼, 757
Poincaré ∼, 499 logarithmic ∼, 549
initial sampling step, 138 Neumann ∼, see Neumann
initial step, 600 Newton ∼, 550
inner of an operator, see also null space
harmonics, 616, 652 on the ball, 767
translation theorem, 617 product ∼, 609
inverse, 232 rational ∼, 351
product reproducing ∼, see reproducing
Euclidean ∼, 167 Shannon ∼, see Shannon
general ∼, 171 single-layer ∼, 707
integrable, 182 singularity ∼, see singularity
integral formula for the Legendre (Green) spacelimited ∼, 757
function, 27 summable ∼, 134
integrated concept, 317 symbol, 759, 780
International Reference Ellipsoid (IRE), Tikhonov ∼, 127
535, 536 zonal ∼, 758
interpolating spline, 609, 794, 796 Krarup sphere, 607, 608
interpolation, 589 Kronecker delta, 168
operator, 264
problem on a regular surface, 606 L-curve method, 629
interrelation, 5 Lamé systems, 533
intrinsic coordinates, 460, 466 Landweber iteration, 283, 284
inverse Laplace operator, see also derivative, 565,
generalized ∼, see generalized 821, 823, 825, 826, 833, 834, 838, 840
Moore–Penrose ∼, see Moore, 106 latitude integration, 48
problem, 883, 884, 886, 887, 896, 901, lattice, 13
903, 908, 912, 916, 917 fundamental cell ∼, 14
gravimetric ∼, 884, 885, 887, 889, 890, Green function, 36
892, 893, 897, 899, 900, 903, 904, periodical Z-∼, 35
908, 909, 915–917 point discrepancy, 11
magnetic ∼, 884, 887–889, 892, 894, points
899, 900, 912, 916, 917 in circles, 10
pseudo∼, see pseudoinverse in spheres, 10
inversion, 563 Lavrentiev regularization, 220
isometric linear operator, 170 Lax–Milgram lemma, 499
isostatic correction, 692 Lax–Milgram theorem, 174
iterative regularization, 282 layer density constraint, 885, 908, 912, 913
layers, 655
Jacobi polynomials, 655, 885, 886, 893, 894 least squares
jump-relations, 828 method, 77, 269
926 Index
dual ∼, 270 space, 466

problem, 212 mass poles, 527
weighted ∼, 214 configurations, 528
solution, 581 mass transports, 883, 885, 887
Legendre matching pursuit, 805
(Green) function functional ∼, 806
1D Legendre operator, 24 orthogonal regularized functional ∼, 811
1D bilinear expansion, 25 regularized functional ∼, 807
integral formula for the ∼, 27 mathematical
function circuit
associated ∼, 41 abstraction, 5
scalar ∼, 764 concretization, 5
tensor ∼, 764 model, 103
vector ∼, 764 Mathieu systems, 533
harmonic, 40 matrix equation, see ill-conditioned
associated ∼, 41 maximal index, 626
kernel, 764, 774 maximum principle, 827
polynomial, 21, 759, 886, 887 maximum/minimum principle, 326
3D theory, 21 Maxwell’s equation, 887
associated ∼, 40 quasi-static approximation of ∼, 887
orthonormal system of ∼s, 23 Mazur’s lemma, 176
zeros of a ∼, 22 mean square error, 78
symbol, 759 mean value property, 826
tensors, 568 measurable, 181
lemma Hausdorff ∼, 188
Lax–Milgram ∼, 499 Meissl relation, 578
Mazur’s ∼, 176 Meissl scheme, 317, 324, 439, 563, 578
Sobolev ∼, see Sobolev Mercer’s theorem, 784, 785
level surface, 689 method
limit and jump relations, 526 aggregation ∼, see aggregation
limit-relations, 828 Backus–Gilbert ∼, 667
linear independence, 525 Boscovic–Laplace ∼, 79
Lipschitz continuous, see also continuous Bubnov–Galerkin ∼, see Bubnov
Lipschitz property, 179 de la Vallée Poussin mollifier ∼, 662
local to targets (L2T), 619 fast multipole ∼, 612
localization Galerkin ∼, see Galerkin
frequency, 770 Gauss–Seidel ∼, 283
space, 769 gradient ∼, 289
locally supported, 757 Haar mollifier ∼, 658
logarithmic kernel, 353, 531, 549 iterative ∼, 282
longitude integration, 46 L-curve ∼, 629
longitude-latitude data systems, 45 modified GCV ∼, 631
lowpass, 371, 388 mollifier ∼, see mollifier
filter, 324, 432 multiscale ∼, see multiscale
of least squares, see least squares
mantle, 655 parameter choice ∼, 625
mapping Rayleigh–Ritz ∼, 272
conformal ∼, 143 reconstruction ∼, 243
Marussi ∼, 471 regularization ∼, 239
Marussi Ritz–Galerkin ∼, 500
condition, 64 semi-iterative ∼, 286
mapping, 471 singular integral mollifier ∼, 665
Index 927
unbiased prediction risk ∼, 630 deterministic ∼, 624

metric space, 168 models, 624
minimum norm condition, 908, 917 stochastic ∼, 625
model observation ∼, 134
earth gravity ∼, see earth strong ∼, 225
ground ∼, see ground weak ∼, 227
mathematical ∼, 103 non-bandlimited kernel functions, 351
noise ∼, see noise non-bandlimited scaling functions, 380
satellite ∼, see satellite non-Euclidean geometry, 149
mollifier method, 295, 666 norm, see also normed space
Molodensky Euclidean ∼, 167
boundary value problem, 69, 551 Hölder norm, see Hölder space
problem, 61, 63, 469, 551 Lebesgue norm, see Lebesgue space
exterior ∼, 553, 554 operator ∼, 170
linearization of the ∼, 65 Sobolev norm, see Sobolev space
of harmonic splines, 555 normal
of the generalized Fourier expansion, equation, 97, 107, 248
554 L2 (∂G)-theory, 534
simple ∼, 68, 465, 475, 476 SplineH(Ac ) -theory, 541
Moore–Penrose inverse, 89, 93, 95, 97, 106 gravity potential, 695
generalized ∼, 581 potential, 63, 470
Morozov’s discrepancy principle, 250 null space, 170, 653, 656, 883, 885, 889,
multipole to exponential, 621 892–894, 896, 897, 899–903, 917
multiresolution, 799, 803 of a Fredholm int. op. 1st kind, 916
analysis, 592 number π
scheme, 281 determination of the ∼, 11
multiscale different approaches, 11
approach, 370 Gaussian approach, 11
approximation theorem, 592 numerical dilemmas, 298
estimator, 136 numerical methodology, 298
inversion, 591
method, 274, 713 objective
mollifier method, 660 ill-posed problem, 204
inverse problem, 204
Nashed’s bibliography, 89, 209 oblique boundary condition, 498, 510
Nashed’s classification, 106, 205 oblique boundary value problem
natural regularity conditions, 469 boundary condition, 498
near/far field methods, 551 formulation
Neumann classical ∼, 497, 508
problem, 830 stochastic ∼, 501
integral ∼, 830 weak ∼, 499
Neumann boundary value problem, 70, 72 outer problem, 508
of physical geodesy, 703 regular inner problem, 497
Neumann Green function, 832, 834, 835, solution
841–843 classical ∼, 497
Neumann kernel, 704 operator, 500, 504, 509
linearly regularized ∼, 710 stochastic strong ∼, 503, 513
Newton stochastic weak ∼, 502, 512
integral, 642, 645 strong ∼, 500, 509
kernel, 550 weak ∼, 499, 509
potential, 887 transformed condition, 510
noise oblique derivative problem, 69, 70, 545
928 Index
observation noise, 134 Picard criterion, 584

octtree Pizzetti concept, 72, 124, 340
child (cube), 614 plumb line, 62, 689, 695
construction, 614 Poincaré inequality, 499
directional lists, 621 point to multipole (P2M), 619
lists of each cube, 615 point vortex motion, 822, 842, 848, 849
neighbor (cube), 614 Poisson equation, 648
parent (cube), 614 Euclidean differential ∼, 728
well-separated (cube), 614 Poisson problem, 822, 827, 829
operator, 169 polynomial
bounded ∼, 170 harmonic ∼, 40
continuous ∼, 170 homogeneous ∼, 40
continuous linear ∼, 171 Jacobi ∼s, see Jacobi
dual space adjoint ∼, 171 Legendre ∼, see Legendre
Hilbert space adjoint ∼, 174 positive definite form, 172
inverse ∼, 170 potential
isometric linear ∼, 170 anomalous ∼, see anomalous
linear ∼, 170 anomaly, 67
projection ∼, 263 centrifugal ∼, see centrifugal
trace ∼, 188 disturbing ∼, see disturbing
oracle inequalities, 626 gravitational ∼, see gravitational
orthogonal, 172 gravity ∼, see gravity
projection, 263 Newton ∼, 887
regularized functional matching pursuit, normal ∼, see normal
811 normal gravity ∼, 695
orthonormal, 172 prime number assumption, 1
basis system, 883, 885, 892, 893 principle
on the ball, 885 balancing ∼, 628
system of Legendre polynomials, 23 hardened ∼, 628
outer harmonics, 525, 563, 571, 616, 653 Chebyshev ∼, see Chebyshev
expansions, 518 discrepancy ∼, see discrepancy
translation theorem for ∼, 617 maximum ∼, 827
outer inverse, 232 maximum/minimum ∼, 326
uncertainty ∼, 580
Paley–Wiener space, 18 probability distribution function, 94
Paley–Wiener splines, 18 problem
parallelogram identity, 172 antenna ∼, see antenna
parameter choice, 584 boundary value ∼, see boundary value
method, 625 circle ∼, see circle
a posteriori ∼, 625 direct ∼, see direct
a priori ∼, 625 Dirichlet ∼, see Dirichlet
data-driven ∼, 625 forward ∼, 204
heuristic ∼, 625 Gaussian circle ∼, 8
Parseval identity, 173 ill-posed ∼, see ill-posed
partial differential equation, 192 inverse ∼, see inverse
elliptic ∼, 192 Molodensky ∼, see Molodensky
hyperbolic ∼, 193 Neumann ∼, see Neumann
parabolic ∼, 193 Neumann boundary value ∼, see
periodic integration, 35, 39 Neumann
error estimation, 39 oblique boundary value ∼, see oblique
permanence property, 533 oblique derivative ∼, see oblique
Picard condition, 274, 903, 904, 910 of least squares, see least squares
Index 929
Poisson ∼, see Poisson surface, 688

properly-posed ∼, 205 terrestrial ∼ system, 461
Ritz–Galerkin approximation ∼, see Ritz refinement equation, 371, 592
SGG ∼, see SGG tensorial ∼, 401
spline interpolation ∼, 589 vectorial ∼, 390
SST ∼, see SST reflexive space, 171
Stokes ∼, 69 region, 177
well-posed ∼, see well-posed regular ∼, see regular
product kernel, 609 Runge ∼, 518
projection regular
methods, 263 region, 15
operator, 263 regular region, 215, 644, 727, 823, 825,
orthogonal ∼, 263 831, 834, 839
propagated noise error, 626 regularity
property at infinity, 727
basis ∼, 323 property, 180
completeness ∼, 522 theorem, 533, 537
cone ∼, 179 regularization, 323, 385, 395, 406, 588, 787,
Lipschitz ∼, 179 788, 804, 806, 807, 810, 811
mean value ∼, 826 asymptotic ∼, 247
permanence ∼, 533 error, 226, 626
regularity ∼, 180 iterative ∼, 282
Runge–Walsh approximation ∼, see method, 239
Runge parameter, 582
segment ∼, 179 chosen ∼, 625
pseudodifferential maximal ∼, 626
equation, 124, 340 Runge ∼, 595
order, 125 strategy, 127, 582
symbol, 125 Tikhonov ∼, see Tikhonov
operator, 340, 563, 585, 587, 756 wavelet, 381
pseudoinverse, 89, 92, 93, 95, 97, 99, 106, regularized functional matching pursuit,
109, 209, 219 807
pyramid scheme, 425 orthogonal ∼, 811
pyramid step, 600 regularizer, 225
reproducing kernel, 195, 347, 430, 609, 783,
qualification, 246
793
quasi-optimality criterion, 628
function, 348, 377, 576
radial basis functions, 754 property, 795
radially symmetric density distribution, space, 254, 792
655 Riesz representation theorem, 173
range, 170 risk, 626
rational kernel, 351 Ritz–Galerkin approximation
rational scaling function, 382 inner problem, 501
reality space, x outer problem, 511
reconstruction error, 246 Ritz–Galerkin method, 500
reconstruction method, 243 Rodriguez formula, 22
recursion step, 138 Rodriguez rule, 22
reference Runge, 571
ellipsoid, 68 region, 518
International ∼ Ellipsoid (IRE), see regularization, 595
International, 536 sphere, 325, 607, 608
sphere, 68 Runge–Walsh
930 Index
approximation property, 333, 337, 579 generator, 131

theorem, 75, 518 kernel, 349, 760, 774, 783
C-topology, 537 smoothed ∼, 350
L2 -topology, 532 number, 761, 762, 779, 780, 784
trial systems sampling, 15
closure, 527 sampling theorem, 590
completeness, 525 scaling function, 377
linear independence, 525 smoothed ∼, 379
wavelet, 131, 800
sampling shifting operator, 372
data system sign expansion, 47
Gaussian grid, 51 signal-to-noise ratio, 134, 135, 137, 804
longitude-latitude ∼, 48 multi-scale, 134
longitude-latitude grid, 51 sampling, 134
latitude-longitude ∼, 49 single-layer kernel, 707
Shannon ∼ theorem, 590 singular
theorem, 256 integral, 657, 664
satellite Haar-type ∼, 657
-to-satellite tracking, 316, 319, 322, 855, value decomposition, 115, 217
858 value expansions, 584
data, 855, 857, 865, 872, 873, 877 singularity kernel, 353, 531, 549, 610
gravity gradiometry, 316, 319, 322, 562, Slepian function, 760, 779
855, 858 bandconcentrated ∼, 760
model, 857, 865, 871 spaceconcentrated ∼, 760
scalar Slepian trees, 787
geodetic observables, 341 smoothed
outer harmonics, 572 Haar kernel, 354
spherical harmonics, 326, 566 Shannon kernel, 350
wavelet theory, 371 Shannon scaling function, 379
scale smoothing parameter, 611
and space error covariance, 135 Sobolev
and space error variance, 135 –Slobodeckij space, 186
space, 371, 386, 388, 399 embedding theorem, 187
thresholding, 139 lemma, 331, 575, 577
scaling function, 324, 371, 591, 798 tensorial ∼, 339
decomposition regularization ∼, 277 vectorial ∼, 335
fundamental ∼, 730 space, 186, 324, 328, 494, 563, 574, 577,
Haar ∼, 730 608
Neumann ∼, 711 harmonic functions, 608
reconstruction regularization ∼, 278 submanifold, 495
Stokes ∼, 707 vectorial ∼, 333
vector ∼, 719 weighted ∼, 494
Schauder estimates, 464, 469 soft thresholding, 136
segment property, 179 solid angle, 14, 15
semi-frame condition, 808 space
semi-iterative method, 286 Banach ∼, 169
separable space, 169 detail ∼, see detail
SGG, 316, 319, 322, 341, 573, 587 function ∼, 176
operator, 439 Hilbert ∼, 172
problem, 344, 357, 374, 388, 399, 585 Hölder ∼, 178
tensorial ∼, 573 inner product ∼, 172
Shannon K-vector ∼, 167
Index 931
Lebesgue ∼, 184 SST, 316, 319, 322, 341

metric ∼, 168 operator, 439
normed ∼, 169 problem, 344, 346, 357, 374, 388
null ∼, see null stability theorem, 542
of continuously differentiable functions, stochastic functions, 501, 512
177 stochastic inhomogeneities, 513, 514
of distributions, 181 Stokes
pre-Hilbert ∼, 172 boundary value problem, 69
reality ∼, x operator, 344
reflexive ∼, 171 problem, 69
reproducing kernel Hilbert ∼, 195 scaling function, 707
scale ∼, see scale strong convergence, 168
Schwartz ∼, 190 strong noise, 225
separable ∼, 169 summability condition, 608
Sobolev ∼, see Sobolev summable kernel, 134
Sobolev–Slobodeckij ∼, see Sobolev summable sequence, 330
tensor product ∼, 501 support, 177
virtuality ∼, x compact, 177
spacelimited, 757 distribution, 182
sphere surface, 494
cubed ∼, 787 curl, 565
Fourier coefficients on a ∼, 607 curl gradient, 565
fundamental equations of physical divergence, 565
geodesy for the ∼, 701 gradient, 71, 565
Krarup ∼, see Krarup potentials, 827
reference ∼, 68 regular ∼, 607
Runge ∼, see Runge symbol, 126, 586, 656
spherical decomposition, 834, 837 kernel ∼, see kernel
spherical harmonics, 197, 319, 323, 562, Legendre ∼, 759
566, 568, 570
3D complex-valued, 44 telluroid, 63, 471
3D real-valued, 44 tensor
addition theorem, 42 outer harmonics, 335, 572
longitude-latitude, 40 product space, 501
of degree n and order j, 44 spherical harmonics, 568, 570, 764
scalar ∼, see scalar tensorial
tensor ∼, 764 geodetic observables, 343
vector ∼, 764 pseudodifferential operator, 346
spherical spline, 791, 794 refinement equation, 401
smoothing, 796 scaling function, 399
spline, 609, 791, 794, 796 SGG problem, 573
approximation, 588, 611 Sobolev lemma, 339
coefficients, 609 Sobolev space, 337
cubic ∼, 791 wavelet theory, 399
harmonic, 609 terrain correction, 692
interpolating ∼, see interpolating terrestrial reference system, 461
interpolation, 793 test function, 649
interpolation problem, 589 theorem
minimum properties, 610 addition ∼, see addition
smoothing, 611, 620, 793 Aronszajn’s ∼, 195
spherical ∼, see spherical Arzelà–Ascoli ∼, see Arzelà
squeeze factor, 780 Cauchy–Kovalevskaya ∼, 194
932 Index
convergence ∼, 808 UTM-coordinate system, 146

divergence ∼, 54
Gauss’s ∼, 54 variance, 626
Lax–Milgram ∼, 174 vector outer harmonics, 331
Mercer’s ∼, see Mercer vector spherical harmonics, 764
multiscale approximation ∼, 592 vectorial
regularity ∼, see regularity geodetic observables, 342
Riesz representation ∼, 173 pseudodifferential operators, 345
Runge–Walsh ∼, see Runge–Walsh refinement equation, 390
scaling function, 388
sampling ∼, see sampling
Sobolev lemma, 335
Sobolev embedding ∼, 187
Sobolev space, 333
stability ∼, 542
wavelet theory, 388
translation ∼, 617
vertical deflection, 71, 318, 698, 705, 718,
thresholding, 136
822, 834, 836, 841–843, 845
tidal correction, 693
virtuality space, x
Tikhonov
–Phillips regularization, 294, 594 wavelet, 323, 372, 592, 706
kernel, 127 approach, 563
regularization, 127, 221, 229, 248, 585, approximation, 370
593 dilation, 799
scaling function, 381 function, 280, 799
total error, 583 fundamental ∼, 731
total mass (density distribution), 656 Haar ∼, 731
trace operator, 188 least energy representation, 802
transformations, 504 Neumann wavelets, 711
coefficients, 508 reconstruction, 713
domain, 504 redundancy, 802
inhomogeneities, 506, 507 rotation, 799
solution, 505 spherical, 799
translation transformation, 324, 801
exponential, 621 variants, 803
L2L with rotations, 620 vector ∼, 719
local to local (L2L), 618 weak
M2M with rotations, 620 compactness, 175
multipole to local (M2L), 618 convergence, 175
multipole to multipole (M2M), 617 noise, 227
theorem, 617 weakstar
trapezoidal rule, 37 weak compactness, 175
trapezoidal sum, 35 weak convergence, 175
tree algorithm, 563, 599 weighted least squares, 90
bandlimited wavelets, 138 well-posed problem, 104, 106, 205, 580
tree sampling, 137 well-posedness, 104, 106
triangulation, 88, 145 due to Hadamard, 740
truncated singular value decomposition, width, 542
585, 594 worst case bound, 251
worst case error, 243
unbiased prediction risk method, 630
zeros of a Legendre polynomial, 22
uncertainties, 772
zonal functions, 754
uncertainty principle, 580, 772–773
unisolvent, 541
upward continuation, 563, 577
operator, 344

Geodesia

Uploaded by

Copyright:

Available Formats

Geodesia

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Geodesia

Uploaded by

Copyright:

Available Formats

Geosystems Mathematics

More information about this series at http://www.springer.com/series/13389

ISSN 2510-1544 ISSN 2510-1552 (electronic)

Library of Congress Control Number: 2018940865

© Springer International Publishing AG, part of Springer Nature 2018

Cover illustration: designed by Prof. Dr. Schreiner, Buchs

Printed on acid-free paper

W. Freeden, T. Sonar, and B. Witte

M. Augustin, S. Eberle, and M. Grothaus

W. Freeden and M.Z. Nashed

W. Freeden and H. Nutz

M. Grothaus and T. Raskop

M. Augustin, W. Freeden, and H. Nutz

W. Freeden, H. Nutz, and M. Schreiner

W. Freeden and M. Zuhair Nashed

C. Blick, W. Freeden, and H. Nutz

W. Freeden, V. Michel, and F.J. Simons

C. Gerhards, S. Pereverzyev Jr., and P. Tkachenko

S. Leweke, V. Michel, and R. Telschow

Geodesy, as most other disciplines, spans activities ranging from theoretical to

February 2017 Willi Freeden, Kaiserslautern

Observation ( = measurement) Modeling ( = solution)

transfer (bridging process)

Figure 1. The circuit.

day’s geodesy. What is frequently overlooked by potential geodetic users of

As a consequence, in the sense of Moritz’ explications, today’s circuits should

This “Handbook of Mathematical Geodesy” (HbMG) aims at providing in-

The handbook is meant as a mathematical addendum to the foundations in

Gauss as Scientiﬁc Mediator

Abstract. The objective of the paper is to document the pioneer dimension of

Keywords. Gauss, geometric number theory, numerical integration, integral

of mathematics and its current development as a common language and indepen-

1.4. Speciﬁc strategies imposed on our work

2. From Gaussian circle problem to geosampling

Figure 2.1. Title-page of “Disquisitiones arithmeticae”, original latin

B2N = {x ∈ R2 : |x| ≤ N } (2.2)

Figure 2.2. Lattice points inside a circle.

reaches back to L. Euler [33]. In today’s nomenclature it can be equivalently ex-

Figure 2.3. The polyhedral set P2N .

2.2. Circle problem and Hardy’s conjecture

A ﬁrst remarkable result is due to W. Sierpinski [201], who proved by use of a

(note that 139 131

0.250000 Gauss (1801)

Table 1. Incremental improvements for the value ε2 in the estimate (2.17).

2.3. Variants of the circle problem

Figure 2.4. Two-dimensional lattice Λ generated by g1 , g2 ∈ R2 .

• The remainder term can be represented as alternating series, called Hardy–

where Λ is an arbitrary lattice in R2 and

where dx is the volume element, a, y ∈ R2 , F is twice continuously diﬀeren-

Note that, for F = 1, this formula leads back to

For a = y = 0 we obtain the classical Hardy–Landau identity, i.e., the identity

holds true. Observe that J1 satisﬁes the asymptotic relation J1 (r) = r

• Generalizations to lattices Λ ⊂ Rq and regular regions G ⊂ Rq , q ≥ 2, and

where a regular region G in Rq is understood to be an open and connected

with α(a + g) denoting the solid angle subtended by ∂G at a + g (note that,

2.4. Multivariate Shannon sampling

as well as the ﬁnite sum

A simple, but signiﬁcant case of (2.37) in Euclidean space Rq showing no phenom-

In fact, the identity (2.38) is a multi-variate variant of the Shannon sampling