Mathematics Key

Mathematics – Key Technology for the Future
Willi Jäger · Hans-Joachim Krebs

Editors
Mathematics –
Key Technology
for the Future
Joint Projects Between Universities
and Industry 2004–2007
123
Willi Jäger Hans-Joachim Krebs
IWR, Universität Heidelberg Projektträger Jülich
Im Neuenheimer Feld 368 Forschungszentrum Jülich GmbH
69120 Heidelberg 52425 Jülich
Germany Germany
jaeger@iwr.uni-heidelberg.de h.-j.krebs@fz-juelich.de
ISBN 978-3-540-77202-6 e-ISBN 978-3-540-77203-3
DOI 10.1007/978-3-540-77203-3
Library of Congress Control Number: 2008923103
Mathematics Subject Classification (2000): 34-XX, 35-XX, 45-XX, 46-XX, 49-XX, 60-XX, 62-XX,
65-XX, 74-XX, 76-XX, 78-XX, 80-XX, 90-XX, 91-XX, 92-XX, 93-XX, 94-XX

c 2008 Springer-Verlag Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, 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.
Cover design: WMXDesign, Heidelberg

Typesetting and production: le-tex publishing services oHG, Leipzig, Germany
Printed on acid-free paper
987654321
springer.com
Solving real problems requires teams – mathematicians,
engineers, scientists and users from various fields, and – last but
not least – it requires money!
Preface
In 1993, the Federal Ministry of Research and Education of the Federal Re-
public of Germany (BMBF) started the first of now five periods of funding
mathematics for industry and services. To date, its efforts have supported
approximately 280 projects investigating complex problems in industry and
services. Whereas standard problems can be solved using standard mathemat-
ical methods and software off the shelf, complex problems arising from e.g.
industrial research and developments require advanced and innovative math-
ematical approaches and methods. Therefore, the BMBF funding programme
focuses on the transfer of the latest developments in mathematical research
to industrial applications. This initiative has proved to be highly successful
in promoting mathematical modelling, simulation and optimization in science
and technology.
Substantial contributions to the solution of complex problems have been
made in several areas of industry and services.
Results from the first funding period were published in “Mathematik –
Schlüsseltechnologie für die Zukunft, Verbundprojekte zwischen Universität
und Industrie” (K.-H. Hoffmann, W. Jäger, T. Lohmann, H. Schunck (Edi-
tors), Springer 1996). The second publication “Mathematics – Key Technology
for the Future, Joint Projects between Universities and Industry” (W. Jäger,
H.-J. Krebs (Editors), Springer 2003) covered the period 1997 to 2000. Both
books were out of print shortly after publication.
This volume presents the results from the BMBF’s fourth funding period
(2004 to 2007) and contains a similar spectrum of industrial and mathematical
problems as described in the previous publications, but with one additional
new topic in the funding programme: risk management in finance and insur-
ance.
Other topics covered are mathematical modelling and numerical simulation
in microelectronics, thin films, biochemical reactions and transport, computer-
aided medicine, transport, traffic and energy.
As in the preceding funding periods, novel mathematical theories and
methods as well as a close cooperation with industrial partners are essen-
VIII Preface
tial components of the funded projects. In order to strengthen and document

this cooperation with industry, the projects starting in the 5th period are
based on consortium agreements between universities, industry and services,
though there is no direct funding for the industrial partners.
The BMBF has declared the year 2008 as “The Year of Mathematics” in
Germany. This book is an excellent contribution to this special year, demon-
strating the usefulness of and prospects for mathematics in technology and
economics. It is widely acknowledged that the support provided by the Min-
istry has been substantial in advancing the transfer of mathematics to indus-
try and services and in opening academic mathematical research for challenges
arising in industrial and economic applications.
In all areas of society, global aspects are becoming more and more im-
portant. Therefore it is extremely important to promote mathematics and its
transfer as a key technology to sciences and industry both on the national
and on the international level. Here, we take the opportunity to acknowledge
the assistance of the BMBF, providing not only necessary national funding,
but also emphasizing the importance of mathematics and the necessity of
promoting its applications in national and international cooperation. By ini-
tiating an OECD Global Science Forum on “Mathematics for Industry” and
by dedicating the scientific year 2008 to mathematics, the BMBF has made
a clear statement. Special thanks for their support go to Dr. H.-F. Wagner,
Dr. R. Koepke and Professor Dr. J. Richter from the BMBF. Last but not
least, we would also like to acknowledge the valued support of Dr. Stefan
Michalowski, representing the OECD in the Global Science Forum.
We also greatly appreciate the excellent cooperation with Springer Ver-
lag, Heidelberg in publishing this volume – the third one to collect and
present results from the BMBF mathematics programme – in its interna-
tionally renowned series of scientific publications.
Heidelberg, Jülich Willi Jäger

March 2008 Hans-Joachim Krebs
Contents
Part I Microelectronics
Numerical Simulation of Multiscale Models

for Radio Frequency Circuits in the Time Domain
U. Feldmann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Numerical Simulation of High-Frequency Circuits

in Telecommunication
M. Bodestedt, C. Tischendorf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Wavelet-Collocation of Multirate PDAEs for the Simulation
of Radio Frequency Circuits
S. Knorr, R. Pulch, M. Günther . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Numerical Simulation of Thermal Effects
in Coupled Optoelectronic Device-Circuit Systems
M. Brunk, A. Jüngel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Efficient Transient Noise Analysis in Circuit Simulation
G. Denk, W. Römisch, T. Sickenberger, R. Winkler . . . . . . . . . . . . . . . . . . 39
Part II Thin Films
Numerical Methods for the Simulation of Epitaxial Growth

and Their Application in the Study of a Meander Instability
F. Haußer, F. Otto, P. Penzler, A. Voigt . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Micro Structures in Thin Coating Layers:
Micro Structure Evolution and Macroscopic Contact Angle
J. Dohmen, N. Grunewald, F. Otto, M. Rumpf . . . . . . . . . . . . . . . . . . . . . . 75
X Contents
Part III Biochemical Reactions and Transport
Modeling and Simulation of Hairy Root Growth

P. Bastian, J. Bauer, A. Chavarría-Krauser, C. Engwer, W. Jäger,
S. Marnach, M. Ptashnyk, B. Wetterauer . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Simulation and Optimization of Bio-Chemical Microreactors
R. Rannacher, M. Schmich . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Part IV Computeraided Medicine
Modeling and Optimization of Correction Measures

for Human Extremities
R. Brandenberg, T. Gerken, P. Gritzmann, L. Roth . . . . . . . . . . . . . . . . . . 131
Image Segmentation for the Investigation
of Scattered-Light Images when Laser-Optically Diagnosing
Rheumatoid Arthritis
H. Gajewski, J. A. Griepentrog, A. Mielke, J. Beuthan, U. Zabarylo,
O. Minet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Part V Transport, Traffic, Energy
Dynamic Routing of Automated Guided Vehicles in

Real-Time
E. Gawrilow, E. Köhler, R. H. Möhring, B. Stenzel . . . . . . . . . . . . . . . . . . 165
Optimization of Signalized Traffic Networks
E. Köhler, R. H. Möhring, K. Nökel, G. Wünsch . . . . . . . . . . . . . . . . . . . . 179
Optimal Sorting of Rolling Stock at Hump Yards
R. S. Hansmann, U. T. Zimmermann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Stochastic Models and Algorithms for the Optimal Operation
of a Dispersed Generation System Under Uncertainty
E. Handschin, F. Neise, H. Neumann, R. Schultz . . . . . . . . . . . . . . . . . . . . 205
Parallel Adaptive Simulation of PEM Fuel Cells
R. Klöfkorn, D. Kröner, M. Ohlberger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Contents XI
Part VI Risk Management in Finance and Insurance
Advanced Credit Portfolio Modeling and CDO Pricing

E. Eberlein, R. Frey, E. A. von Hammerstein . . . . . . . . . . . . . . . . . . . . . . . . 253
Contributions to Multivariate Structural Approaches
in Credit Risk Modeling
S. Becker, S. Kammer, L. Overbeck . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Economic Capital Modelling and Basel II Compliance
in the Banking Industry
K. Böcker, C. Klüppelberg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Numerical Simulation for Asset-Liability Management
in Life Insurance
T. Gerstner, M. Griebel, M. Holtz, R. Goschnick, M. Haep . . . . . . . . . . . . 319
On the Dynamics of the Forward Interest Rate Curve
and the Evaluation of Interest Rate Derivatives
and Their Sensitivities
C. Croitoru, C. Fries, W. Jäger, J. Kampen, D.-J. Nonnenmacher . . . . . 343
List of Contributors
Peter Bastian Martin Bodestedt

Universität Stuttgart, IPVS Universität zu Köln
Universitätsstraße 38 Weyertal 86–90
D-70569 Stuttgart D-50931 Köln
peter.bastian@ipvs.uni-stuttgart.de mbodeste@math.uni-koeln.de
Jenny Bauer René Brandenberg
Universität Heidelberg, IPMB Technische Universität München
INF 364 Zentrum für Mathematik
D-69120 Heidelberg Boltzmannstr. 3
jenny_bauerbiol@yahoo.de D-85747 Garching bei München
Swantje Becker brandenb@ma.tum.de
Universität Gießen
Markus Brunk
Arndtstraße 2
Universität Mainz
D-35392 Gießen
Staudingerweg 9
swantje.becker@math.uni-giessen.de
D-55099 Mainz
Jürgen Beuthan brunk@mathematik.uni-mainz.de
Charité Universitätsmedizin Berlin
Campus Benjamin Franklin Andrés Chavarría-Krauser
Institut für Medizinische Physik und Universität Heidelberg, IAM
Lasermedizin INF 294
Fabeckstraße 60-62 D-69120 Heidelberg
D-14195 Berlin andres.chavarria.Krauser@iwr.uni-
juergen.beuthan@charite.de heidelberg.de
Klaus Böcker Cristian Croitoru

Risk Integration, Reporting & Universität Heidelberg, IWR
Policies INF 368
Risk Analytics and Methods D-69120 Heidelberg
UniCredit Group, Munich Branch cristian.croitoru
klaus.boecker@hvb.de @iwr.uni-heidelberg.de
XIV List of Contributors
Georg Denk Herbert Gajewski

Qimonda AG Forschungsverbund Berlin e.V.
Am Campeon 1–12 WIAS
D-85579 Neubiberg Mohrenstraße 39
georg.denk@qimoda.com D-10117 Berlin
gajewski@wias-berlin.de
Julia Dohmen
Universität Bonn Ewgenij Gawrilow
Institut für Numerische Simulation Technische Universität Berlin
Nussallee 15 Institut für Mathematik, MA 6-1
D-53115 Bonn Straße des 17. Juni 136
julia.dohmen@ins.uni-bonn.de D-10623 Berlin
gawrilow@math.tu-berlin.de
Ernst Eberlein
Universität Freiburg Jens A. Griepentrog
Abteilung für Mathematische Forschungsverbund Berlin e.V.
Stochastik WIAS
Eckerstraße 1 Mohrenstraße 39
D-79104 Freiburg D-10117 Berlin
eberlein@stochastik.uni-freiburg.de griepent@wias-berlin.de
Christian Engwer Ralf Goschnik

Universität Stuttgart, IPVS Zürich Gruppe Deutschland
Universitätsstraße 38 Poppelsdorfer Allee 25–33
D-70569 Stuttgart D-53115 Bonn
christian.engwer@ipvs.uni- r.goschnik@zurich.com
stuttgart.de
Michael Günther
Rüdiger Frey Bergische Universität Wuppertal
Universität Leipzig Fachbereich C, Gaußstr. 20
Abteilung für Mathematik D-42119 Wuppertal
D-04081 Leipzig guenther@math.uni-wuppertal.de
ruediger.frey@math.uni-leipzig.de
Michael Griebel
Uwe Feldmann Universität Bonn
Qimoda AG Institut für Numerische Simulation
Am Campeon 1–12 Nussallee 15
D-85579 Neubiberg D-53115 Bonn
uwe.feldmann@online.de griebel@ins.uni-bonn.de
Christian Fries Natalie Grunewald

DZBank AG Universität Bonn, IAM
Platz der Republik Wegelerstr. 10
D-60265 Frankfurt am Main D-53115 Bonn
mail@christian-fries.de grunewald@ins.uni-bonn.de
List of Contributors XV
Tobias Gerken Frank Haußer

Technische Universität München Technische Fachhochschule Berlin,
Zentrum für Mathematik FB II
Boltzmannstr. 3 Luxemburger Str. 10
D-85747 Garching bei München D-13353 Berlin
brandenb@ma.tum.de hausser@tfh-berlin.de
Marcus Haep
Peter Gritzmann Zürich Gruppe Deutschland
Technische Universität München Poppelsdorfer Allee 25–33
Zentrum für Mathematik D-53115 Bonn
Boltzmannstr. 3 marcus.haep@zurich.com
D-85747 Garching bei München
Ernst August von Hammerstein
brandenb@ma.tum.de
Universität Freiburg
Abteilung für Mathematische
Thomas Gerstner Stochastik
Universität Bonn Eckerstraße 1
Institut für Numerische Simulation D-79104 Freiburg
Nussallee 15 hammer@stochastik.uni-freiburg.de
D-53115 Bonn
gerstner@ins.uni-bonn.de Willi Jäger
IWR, Universität Heidelberg
Im Neuenheimer Feld 368
Ronny Hansmann D-69120 Heidelberg
Technische Universität Braunschweig jaeger@iwr.uni-heidelberg.de
Institut für Mathematische
Optimierung Ansgar Jüngel
Pockelstraße 14 Technische Universität Wien
D-38106 Braunschweig Institut für Analysis und Scientific
r.hansmann@tu-bs.de Computing
Wiedner Hauptstr. 8–10
1040 Wien, Austria
Edmund Handschin
juengel@anum.tuwien.ac.at
Universität Dortmund
Institut für Energiesysteme und Jörg Kampen
Energiewirtschaft Forschungsverbund Berlin e.V.
Emil-Figge-Straße 70 WIAS
D-44227 Dortmund Mohrenstraße 39
edmund.handschin@udo.edu D-10117 Berlin
kampen@wias-berlin.de
Markus Holtz Stefanie Kammer
Universität Bonn Universität Gießen
Institut für Numerische Simulation Arndtstraße 2
Nussallee 15 D-35392 Gießen
D-53115 Bonn stefanie.kammer
holtz@ins.uni-bonn.de @math.uni-giessen.de
XVI List of Contributors
Robert Klöfkorn Alexander Mielke

Mathematisches Institut Forschungsverbund Berlin e.V.
Abteilung für Angewandte WIAS
Mathematik Mohrenstraße 39
Universität Freiburg D-10117 Berlin
Hermann-Herder-Str. 10 mielke@wias-berlin.de
D-79104 Freiburg i. Br.
robertk@mathematik.uni-freiburg.de Olaf Minet
Charité Universitätsmedizin Berlin
Claudia Klüppelberg Campus Benjamin Franklin
Technische Universität München Institut für Medizinische Physik und
Zentrum für Mathematische Lasermedizin
Wissenschaften Fabeckstraße 60-62
D-85747 Garching bei München D-14195 Berlin
cklu@ma-tum.de olaf.minet@charite.de
Stephanie Knorr Rolf Möhring

Bergische Universität Wuppertal Technische Universität Berlin
Fachbereich C, Gaußstr. 20 Institut für Mathematik, MA 6-1
D-42119 Wuppertal
Straße des 17. Juni 136
knorr@math.uni-wuppertal.de
D-10623 Berlin
moehring@math.tu-berlin.de
Ekkehard Köhler
Brandenburgische Technische
Universität Cottbus Klaus Nökel
Institut für Mathematik PTV AG
Postfach 10 13 44 Stumpfstraße 1
D-03013 Cottbus D-76131 Karlsruhe
ekoehler@math.tu-cottbus.de klaus.noekel@ptv.de
Dietmar Kröner Frederike Neise

Mathematisches Institut Universität Duisburg-Essen
Abteilung für Angewandte Abteilung für Mathematik
Mathematik Forsthausweg 2
Universität Freiburg D-47048 Duisburg
Hermann-Herder-Str. 10 neise@math.uni-duisburg.de
D-79104 Freiburg i. Br.
dietmar@mathematik.uni-freiburg.de Hendrik Neumann
Universität Dortmund
Sven Marnach Institut für Energiesysteme und
Universität Stuttgart, IPVS Energiewirtschaft
Universitätsstraße 38 Emil-Figge-Straße 70
D-70569 Stuttgart D-44227 Dortmund
sven.marnach@ipvs.uni-stuttgart.de hendrik.neumann@udo.edu
List of Contributors XVII
Dirk-Jens Nonnenmacher Werner Römisch

HSH Nordbank AG Humboldt-Universität zu Berlin
Gerhart-Hauptmann-Platz 50 Institute of Mathematics
D-20095 Hamburg D-10099 Berlin
dirk.jens.nonnenmacher@hsh- romisch@math.hu-berlin.de
nordbank.com
Martin Rumpf
Mario Ohlberger Universität Bonn
Institut für Numerische Institut für Numerische Simulation
und Angewandte Mathematik Nussallee 15
Universität Münster D-53115 Bonn
Einsteinstraße 62 martin.rumpf@ins.uni-bonn.de
D-48149 Münster
mario.ohlberger Lucia Roth
@math.uni-muenster.de Technische Universität München
Zentrum für Mathematik
Felix Otto Boltzmannstr. 3
Universität Bonn, IAM D-85747 Garching bei München
Wegelerstr. 10 brandenb@ma.tum.de
D-53115 Bonn
otto@iam.uni-bonn.de Rolf Rannacher
Universität Heidelberg, IAM
Ludger Overbeck
Im Neuenheimer Feld 293/294
Universität Gießen
D-69120 Heidelberg
Arndtstraße 2
rolf.rannacher@iwr.uni-heidelberg.de
D-35392 Gießen
ludger.overbeck@math.uni-giessen.de
Michael Schmich
Mariya Ptashnyk Universität Heidelberg, IAM
Universität Heidelberg, IAM Im Neuenheimer Feld 293/294
INF 294 D-69120 Heidelberg
D-69120 Heidelberg michael.schmich
bernhard.wetterauer @iwr.uni-heidelberg.de
@urz.uni-heidelberg.de
Rüdiger Schulz
Patrick Penzler Universität Duisburg-Essen
Universität Bonn, IAM Abteilung für Mathematik
Wegelerstr. 10 Forsthausweg 2
D-53115 Bonn D-47048 Duisburg
penzler@iam.uni-bonn.de schultz@math.uni-duisburg.de
Roland Pulch Thorsten Sickenberger

Bergische Universität Wuppertal Humboldt-Universität zu Berlin
Fachbereich C, Gaußstr. 20 Institute of Mathematics
D-42119 Wuppertal D-10099 Berlin
pulch@math.uni-wuppertal.de sickenberger@math.hu-berlin.de
XVIII List of Contributors
Björn Stenzel Gregor Wünsch

Technische Universität Berlin Technische Universität Berlin
Institut für Mathematik, MA 6-1 Institut für Mathematik, MA 6-1
Straße des 17. Juni 136 Straße des 17. Juni 136
D-10623 Berlin D-10623 Berlin
stenzel@math.tu-berlin.de wuensch@math.tu-berlin.de
Caren Tischendorf Urszula Zabarylo

Universität zu Köln Charité Universitätsmedizin Berlin
Weyertal 86-90 Campus Benjamin Franklin
D-50931 Köln Institut für Medizinische Physik und
ctischen@math.uni-koeln.de Lasermedizin
Fabeckstraße 60–62
Axel Voigt D-14195 Berlin
Technische Universität Dresden urszula.zabarylo@charite.de
IWR
Zellescher Weg 12–14 Uwe T. Zimmermann
D-01062 Dresden Technische Universität Braunschweig
axel.voigt@tu-dresden.de Institut für Mathematische
Optimierung
Renate Winkler Pockelstraße 14
Humboldt-Universität zu Berlin D-38106 Braunschweig
Institute of Mathematics u.zimmermann@tu-bs.de
D-10099 Berlin
winkler@math.hu-berlin.de
Bernhard Wetterauer
Universität Heidelberg, IPMB
INF 364
D-69120 Heidelberg
bernhard.wetterauer
@urz.uni-heidelberg.de
Part I
Microelectronics
Numerical Simulation of Multiscale Models
for Radio Frequency Circuits
in the Time Domain
Uwe Feldmann
Qimonda AG, Am Campeon 1–12, 85579 Neubiberg, Germany

uwe.feldmann@online.de
1 Introduction
Broadband data communication via high frequent (RF) carrier signals has
become a prerequisite for successful introduction of new applications and ser-
vices in the hightech domain, like cellular phones, broadband internet services,
GPS, and radar sensors for automotive collision control. It is driven by the
progress in microelectronics, i. e. by scaling down from micrometer dimen-
sions into the nanometer range. Due to decreasing feature size and increasing
operating frequency very powerful electronic systems can be realized on inte-
grated circuits, which can be produced for mass applications at very moderate
cost. However, technological progress also opens clearly a design gap, and in
particular a simulation gap:
• Systems get much more complex, with stronger interaction between digital
and analog parts.
• Parasitic effects become predominant, and neither mutual interactions nor
the spatial extension of circuit elements can be further neglected.
• The signal-to-noise ratio decreases and statistical fluctuations in the fab-
rication lines increase, thus enhancing the risk of circuit failures and yield
reduction.
Currently available industrial simulation tools can not cope with all of these
challenges, since they are mostly decoupled, and adequate models are not yet
completey established, or too expensive to evaluate. The purpose of this paper
is to demonstrate that joint mathematical research can significantly contribute
to improve simulation capabilities, by proper mathematical modelling and
development of numerical methods which exploit the particular structure of
the problems. The depth of mathematical research for achieving such progress
is beyond industrial capabilities; so academic research groups are strongly
involved. However, industry takes care that the problems being solved are of
industrial relevance, and that the project results are driven into industrial use.
4 U. Feldmann
The challenging simulation problems mentioned above accumulate in RF

transmitter-receiver pairs (transceivers), which constitute the core function-
ality of most RF communication systems. Although not being very large in
terms of transistor count, transceivers are often extremely hard to simulate
in practice. Mathematically, they exhibit widely separated timescales, they
require coupling of semiconductor device equations with thermal equations
and standard circuit equations, and they are sensitive with respect to de-
vice noise, which is stochastic by nature. For the purpose of this project, a
simplified, but typical representative of a CMOS transceiver was chosen as
a common benchmark. So this transceiver constitutes a common framework
for the research activities within this project. The global objective was to
extend standard simulation methods towards more accurate, comprehensive
and efficient simulation of this transceiver in a large digital circuit environ-
ment.
Since in this setting frequency domain methods are not very helpful, re-
search is focused on time domain models and methods.
For a further discussion of the mathematical problems involved, the
transceiver is shortly described in the next Subsect.
1.1 RF Transceiver Blocks
Systems for RF data transmission usually have comprehensive parts for digital
system processing which work on a sufficiently large number of parallel bits at
conventional clock rates. For data transmission the signals are condensed by
multiplexing and modulation onto a very high frequent analog carrier signal.
This is illustrated in the upper part of Fig. 11 .
Modulation is done by the multiplexer. The latter gets the high frequent
carrier signal from an RF clock generator, which is usually built as a phase
locked loop (PLL). The RF signal being modulated with the data is fed into
the RF transmitter, which – for optical data transmission – may be a laser
diode or – for wireless data transmission – an amplifier with a resonator and
antenna.
A rough scheme for the clock generating PLL is given in Fig. 2. Its core is
a voltage controlled oscillator (VCO), which generates the high frequent har-
monic clock pulses. For stabilization, this RF clock is frequency divided down
onto system frequency and fed into the phase detector. The latter compares
it with the system clock and generates a controlling signal for the VCO: If
the VCO is ahead then its frequency is reduced, and if the VCO lags behind
then its frequency is increased. The number of cycles for the PLL to ‘lock in’
is usually rather large (up to 104 . . . 105 ), which gives rise to very challenging
simulation tasks.
1
In the Figs. bold lines denote RF signals (with frequencies in the range of
10 . . . 50 GHz), while the thinner lines denote signals at standard frequencies
(currently about 1 . . . 2 GHz).
Numerical Simulation of Multiscale Models for RF Circuits 5
Fig. 1. RF data transmission with a pair of transmitter (top) and receiver (bottom)
Fig. 2. PLL for generating RF clock
The receiver part of a transceiver system first has to amplify the signal
and to synchronize itself with the incoming signal, before it can re-extract
the digital information for further processing at standard clock rates, see the
bottom part of Fig. 1.
The RF receiver will be a photo diode in case of optical transmission, and
an antenna with resonator in case of electromagnetic transmission. The high
frequent and noisy low power input signal is amplified in a transimpedance
amplifier and then fed into the demultiplexer, which extracts the digital data
from it and puts them on a low frequent parallel bus for the digital part. A
second PLL takes care that the incoming signal and the receiver’s system clock
are well synchronized. Typically, in this PLL both VCO and phase detector
operate at high frequency, see Fig. 3.
Finally, the VCO’s output signal is down converted onto the base frequency
for delivering a synchronized system clock into the digital part.
Usually, transmitter and receiver are realized on one single chip, in order
to enable a handshaking mode between the server and the host system. Fur-
6 U. Feldmann
Fig. 3. PLL for clock synchronization
Fig. 4. A GSM transceiver circuit
thermore, both make use of the same building blocks, such that e. g. only one
VCO is needed. Figure 4 shows a photo of a GSM transceiver circuit, inclusive
digital signal processing. The inductor windings of the VCO can be clearly
identified in the bottom right corner.
1.2 New Mathematics for Simulating Transceivers
Traditionally, transceivers have been simulated with a standard circuit sim-

ulator, which employs Kirchhoffs equations to set up a system of ordinary
nonlinear differential algebraic equations of DAE index 2, and solves them
either in the time or in the nonlinear frequency domain. However, technolog-
ical progress with decreasing feature sizes and increasing frequency bandwith
drives this approach beyond its limits.
Efficient solvers for multiscale problems in the time domain.
Due to widely separated frequencies of the VCO and the digital system
clock we have a multiscale system, which has to be analyzed over many
Numerical Simulation of Multiscale Models for RF Circuits 7
clock cycles for complete verification e. g. of the lock in of the PLL. With
single rate integration, this may require weeks of simulation time. Recently
developed schemes for separating time scales are more promising, but they
cannot directly be used due to very nonharmonic shape of the digital clock
signals.
Coupling of device and circuit simulation.
Spatial extension of some of the devices (transistors, diodes, . . . ) oper-
ating in the critical path at their technological limits can no longer be
neglected. Hence it is advisable to substitute their formerly used compact
models by sets of semiconductor equations, thus requiring coupled cir-
cuit device simulation. The latter is already available in some commercial
packages, but here the focus is different: The new objective is to simulate
some few transistors and diodes on the device level efficiently together
with thousands of circuit level transistors. This requires extremely robust
coupling schemes.
Interaction with thermal effects in device modelling.
In particular for the emitting laser diode and the receiving photo diode in
opto-electronic data transmission there is interaction with thermal effects,
whose impact on the transceiver chain has not yet been taken into account.
So thermal feedback has to be incorporated into device simulation models
and schemes.
Efficient solvers for stochastic differential algebraic equations.
Widely opened eye-diagrams of the noisy signals on the receiver side are
essential for reliable signal detection and low bit error rates. Hence noise
effects are to be considered very carefully in particular on the receiver
part. Up to now, only frequency domain methods have been developed
for this purpose. However, time domain based methods should be more
generally applicable and hence more reliable in the setting to be considered
here. Therefore, efficient numerical solvers for large systems of stochastic
differential algebraic equations of index 2 are necessary. The focus is here
on efficient calculation of some hundred or even thousand solution pathes,
as are necessary for getting sufficient numerical confidence about opening
of eye-diagrams, etc.
Although the items of mathematical activity look rather widespread here, all
of them serve just to improve simulation capabilities for the RF transceiver
circuitry in advanced CMOS technologies. Therefore all of their results will
be combined in one single simulation environment which is part of or linked
to industrial circuit simulation.
Numerical Simulation of High-Frequency
Circuits in Telecommunication
Martin Bodestedt and Caren Tischendorf
Universität zu Köln, Mathematisches Institut, Weyertal 86–90, 50931 Köln,

Germany, {mbodeste,ctischen}@math.uni-koeln.de
Summary. Driving circuits with very high frequencies may influence the standard
switching behavior of the incorporated semiconductor devices. This motivates us to
use refined semiconductor models within circuit simulation. Lumped circuit models
are combined with distributed semiconductor models which leads to coupled sys-
tems of differential-algebraic equations and partial differential equations (PDAEs).
Firstly, we present results of a detailed perturbation analysis for such PDAE sys-
tems. Secondly, we explain a robust simulation strategy for such coupled systems.
Finally, a multiphysical electric circuit simulation package (MECS) is introduced
including results for typical circuit structures used in chip design.
1 Introduction
In the development of integrated circuit so-called compact models are used
to describe the transistors in the circuit simulations that are performed by
the chip designer. Compact models are small circuits whose voltage-current
characteristics are fitted to the ones of the real device by parameter tuning.
Unfortunately, not even the computationally most expensive compact models
are able to fully capture the switching behavior of field effect transistors when
very high frequencies are used in RF-transceivers.
A remedy to this problem is to employ a more physical modeling ap-
proach and model the critical transistors with distributed device equations.
Here we consider the drift-diffusion equations which is a system of mixed ellip-
tic/parabolic partial differential equations (PDEs) describing the development
of the electrostatic potential and the charge carrier densities in the transistor
region.
For the non-critical devices a lumped modeling approach is followed with
the modified nodal analysis (MNA) [13]. The electric network equations are
in this case a differential algebraic equation (DAE) with node potentials and
some of the branch currents as unknowns. In this article we discuss the ana-
lytical and numerical treatment of the partial differential algebraic equation
10 M. Bodestedt, C. Tischendorf
(PDAE) that is obtained when the MNA network equations are coupled with
the drift-diffusion device equations.
This article is organize as follows. We start by discussing the the refined
circuit PDAE in particular and the perturbation sensitivity of PDAEs in gen-
eral. We state results categorizing the perturbations sensitivity of the PDAE
in terms of the circuit topology. Then, we present the software package MECS
(Multiphysical Electric Circuit Simulator) which makes it possible to perform
transient simulations of circuits with a mixed lumped/distributed modeling
approach for the semiconductor devices. The numerical approximation is ob-
tained by the method of lines (MOL) combined with time-step controlled
time-integration scheme especially suited for circuit simulations.
2 The Refined Circuit Model

For brevity the refined circuit model consisting of the MNA network equations
coupled with the drift-diffusion device equations is summarized in Box 1 [2].
Electric Network Equations:

“ ” “ ” “ ”
AC qC ATC e +AR g ATR e +AL jL + AV jV + AIc ic ATIc e + ĀS JS = −AIs is , (1a)
φ(jL ) − ATL e = 0, (1b)

ATV e = vs . (1c)
Semiconductor Device Equations:
εm Δψ = q(n − p − N ), (1d)
q∂t n − divjn = −qR, where jn = qUT μn grad n − qμn n grad ψ, (1e)
q∂t p + divjp = −qR, where jp = −qUT μp grad p − qμp p grad ψ. (1f)
Coupling interface:
Z
JS = (jn + jp − εm ∂t grad ψ) · grad hi dx (1g)
Ω
“ ”
ψ|ΓD = hĀTS e + ψbi |ΓD , n|ΓD = nD |ΓD , p|ΓD = pD |ΓD , (1h)
grad ψ · ν|ΓN = grad n · ν|ΓN = grad p · ν|ΓN = 0. (1i)
Initial conditions:
(e(t0 ), jL (t0 ), jV (t0 )) = (e0 , jL
0
, jV0 ). (1j)
n(x, t0 ) = n0 (x), p(x, t0 ) = p0 (x), a.e. in Ω, (1k)
Box 1. The refined circuit model
The topology of the circuit is described by a network graph consisting of

network nodes and element branches. It contains the element types: resistors
R, inductors L, capacitors C, independent voltage sources V , independent
Numerical Simulation of High-Frequency Circuits in Telecommunication 11
current sources Is , controlled current sources Ic and semiconductor devices S.

The positions of the different elements are described by the incidence matrices
A accompanied by the corresponding subscript.
The unknowns in the network equations (1a–1c) are node potentials e(t),
inductor currents jL (t) and independent voltage source currents jV (t). The
functions qC (ATC e), g(ATR e) and ic (ATIc e) model capacitor charges, resistor cur-
rents and currents through controlled current sources. They are dependent of
their respective applied voltage. The function φ(jL ) describe the electromag-
netic fluxes in the inductors. Data are the functions describing the independent
voltage and current sources vs (t) and is (t).
For the drift-diffusion equations (1d–1e), after the insertion of the expres-
sions for the charge carrier densities jn (x, t) and jp (x, t) into the two balance
equations we have the unknowns electrostatic potential ψ(x, t), electron den-
sity n(x, t) and hole density p(x, t). The function R(n, p) is a source term
modeling the recombination of the charge carriers with the semiconductor
substrate. The carrier mobilities μn (x) and μp (x) and the doping profile of
the semiconductor substrate N (x) are considered to be space dependent func-
tions. The dielectric permittivity εm , the unit charge q, the intrinsic carrier
density ni and the thermal potential UT are constants.
The semiconductor region Ω ∈ Rd , d ∈ {1, 2, 3} has a boundary Γ which
nD nN
is the union of Dirichlet ∪˙ i=1 ΓDi and Neumann segments ∪˙ i=1 ΓNi . In the
simulations, we additionally to Neumann and Dirichlet conditions consider
mixed boundary conditions at the so-called gate contacts.
The network and device equations are coupled in two ways at the Dirichlet
boundaries. The currents flowing over the semiconductor contacts JS (t) are
evaluated in (1g) and accounted for in Kirchoff’s current law (1a). There, the
functions hi fulfill the Laplace equation and homogeneous Neumann respec-
tively Dirichlet boundary conditions, except at ΓDi where they equal 1. We
have a coupling in the other direction as well as the values of electrostatic
potential ψ at the Dirichlet boundaries ΓDi depend on the node potentials e
in the network (1h).
Beside the dynamic drift-diffusion model (1d–1f) we also consider the sta-
tionary version, which is obtained by putting ∂t n = ∂t p = 0 in (1e–1f):
εm Δψ = q(n − p − N ), (1.1d’)
qdiv(UT μn grad n − μn n grad ψ) = qR, (1.1e’)
qdiv(UT μp grad p + μp p grad ψ) = qR. (1.1f’)
In this case the dynamical term in the current evaluation is neglected. Instead
of (1g) we have

JS = (jn + jp ) · grad hi dx (1.1g’)
Ω
The complete circuit model consisting of the MNA equations coupled with
the stationary drift-diffusion model is (1.1a–c, 1.1d’–g’, 1.1h–k).
3 Perturbation and Index Analysis
Differential algebraic equations (DAEs), but also PDAEs, which can be seen
as abstract DAEs, may have solution operators that contain differential opera-
tors. Now, numerical differentiation is an ill-posed problem, wherefore numer-
ical differentiation of small errors that arise in the approximation process can
lead to large deviations of the approximative solution form the exact one. In
order to successfully integrate PDAEs in time it is important to have knowl-
edge of perturbation sensitivity of the solutions. A measure for this sensitivity
is the perturbation index.
Definition 1. Let X, Z and Y be real Hilbert spaces, I = [t0 , T ], F : Z ×
X × I → Y , δ(t) ∈ Y for all t ∈ I and let wδ and w solve the perturbed,
respectively unperturbed problem:
F (∂t wδ (t), wδ (t), t) = δ(t), F (∂t w(t), w(t), t) = 0.
F = 0 is said to have the perturbation index ν if it is the lowest integer such

that an estimate of the form

ν−1
max ||wδ (t)−w(t)||X ≤ c ||wδ (t0 )−w(t0 )||X+ max ||(∂t )i (wδ (t)−w(t))||Y
t∈I t∈I
i=0
holds for some constant c.
Example 1. Consider the PDAE
∂t u2 (t) − ∂xx
2
u1 (x, t) = 0, u2 (t) = 0, (x, t) ∈ Ω × I, (2)
where u1 (t) ∈ Rn and u2 (·, t) ∈ H01 (Ω). If the two equations are perturbed
with δ1 ∈ C(I, H −1 (Ω)) and δ2 ∈ C 1 (I, Rn ) we can derive the following
bounds for the deviation from the exact solution
max ||u1δ (t) − u1 (t)||H01 ≤ c max (||δ1 (t)||H −1 + |∂t δ2 (t)|) ,

t∈I t∈I
max |u2δ (t) − u2 (t)| = max |δ2 (t)|.
t∈I t∈I
According to Definition 1 the perturbation index of (2) is two since a first

order time-derivative of δ2 appears in the first bound.
Before we turn to the results, we briefly summarize the assumptions needed

for the mathematical analysis of the coupled system.
Assumption 1. The electric network is consistent, its elements are passive
and its data smooth. The controlled sources are shunted parallel with capaci-
tors. The recombination is of Shockley-Read-Hall type and the doping fulfills
N ∈ H01 (Ω ∩ ΓN ) ∩ W 1,4 (Ω) ∩ H 2 (Ω).
The domain Ω ∈ Rd for d ∈ {1, 2, 3} has a Lipschitz boundary. The Neu-

nN
mann boundary ΓN = ∪˙ i=1 ΓNi is the union of C 2 -parameterizable segments.
Further, the (d − 1)-dimensional Lebesgue measure of the Dirichlet bound-
nD
ary ΓD = ∪˙ i=1 ΓDi is positive. The angles between Neumann and Dirichlet
segments do not succeed π/2.
For a more exhaustive discussion of Assumption 1 we refer to [2] and the
references therein. Next we will see that the perturbation sensitivity of the
refined network model strongly depends on the topology of the electric network
graph.
Theorem 1. Let Assumption 1 hold and assume that the contacts of the de-
vice is connected by a capacitive paths.
Then, the perturbation index of the PDAE (1.1a–c, 1.1d’–g’, 1.1h–k) is 1 if
and only if the network graph contains neither loops of capacitors and at least
one voltage source nor cutsets of inductors and independent current sources.
Otherwise, the perturbation index is 2.
The proof can be found in [2]. This result generalizes known index criteria for
the MNA equations [13, 14, 11, 3] of that PDAE with stationary drift-diffusion
equations.
Since we are interested in high-frequency applications it is important to
account for the dynamical behavior of the devices, especially when a stationary
description is used. This is done by the assumption that the contacts are
connected by a capacitive path, which models the reactive or charge storing
behavior of the device. In our next theorem we give a first perturbation result
for a refined circuit model with dynamical device equations.
Theorem 2. Let Assumption 1 hold, the network equations be linear and with-
out loops of capacitors, semiconductors and at least one voltage source and also
without cutsets of inductors and independent current sources. Assume that the
domain Ω is one-dimensional. If the network equations are perturbed by con-
tinuous sufficiently small perturbations the perturbed solutions exist and the
deviations fulfill the bounds
“ ” Z T “ ”
max |ȳ(t)|2 + ||n̄(t)||2L2 + ||p̄(t)||2L2 + ||∂x n̄(τ )||2L2 + ||∂x p̄(τ )||2L2 dτ
t∈I t0
„ Z T «
≤ Cynp |δy0 |2 + ||n0δ ||2L2 + ||p0δ ||2L2 + |δ̃P |2 dτ
t0
„ Z T «
2
max |z̄(t)| ≤ Cz |δy0 |2 + ||n0δ ||2L2 + ||p0δ ||2L2 + 2
|δ̃p | dτ + max |δQ | 2
t∈I t0 t∈I
„ Z T «
max max |∂x ψ̄(x, t)| ≤ Cψ |δy0 |2 + ||n0δ ||2L2 + ||p0δ ||2L2 + |δ̃p |2 dτ .
t∈I x∈Ω t0
In order to split the variables into dynamical and algebraic parts we have
put y := (PCS e, jL ) and z := (QCS e, jV ) where QCS is a projector onto
ker(AC AS )T and PCS = I−QCS . Here, the bar¯denotes the deviation between
the exact and perturbed solution.
The proof can be found in [2]. This result is in good correspondence with
the index criteria in the previous theorem as well as with the ones in [13, 14,
11, 3].
If perturbations in the drift-diffusion equations are allowed one cannot
(at least in the standard way [5, 1]) prove the non-negativity of the charge
carriers and the charge preservation in the diode anymore. These properties
are essential for the a priori estimates of the perturbed solutions, which in
turn are necessary for the estimation of the nonlinear drift currents.
4 Numerical Simulation
Concerning the existence of well-established circuit simulation and device
simulation packages alone, the first natural idea would be to couple these
simulation packages in order to solve the circuit PDAE system described by
(1). However, this approach turned out to involve persistent difficulties. The
main problem consists of the adaption of the different time step controls within
both simulations. This can be handled for low frequency circuits since time
constants for circuits on the one hand and devices on the other hand differ
by several magnitudes in such cases. Our main goal is to investigate high fre-
quency circuits. Here, the pulsing of the circuit is driven near to the switching
time of the device elements. Our coupling of circuit and device simulations
often failed for such cases. Whereas the time step control of the circuit simu-
lation works well, the device simulation does not find suitable stepsizes to
provide sufficiently accurate solutions when higher frequencies are applied.
Therefore, we pursue a different strategy to solve the circuit PDAE system
described by (1) numerically. In order to control the stepsize for the whole
system at once we choose a method of lines approach. First, we discretize the
system with respect to space. Then we use an integration method for DAEs
for the resulting differential-algebraic equation system. Consequently, we take
the same time discretization for the circuit as for the device part.
Space discretization is needed for the device equations (1d)–(1f) as well
as for the coupling interface equations (1g)–(1i). The former ones represent
a coupled system of one elliptic and two parabolic equations for each semi-
conductor. We use here finite element methods leading to
εm Th ψh + qSh (nh − ph − Nh ) = 0, (3a)
∂n
Mn,h + gn,h (jn.h , nh , ph ) = 0, (3b)
∂t
∂n
Mp,h + gp,h (jp.h , nh , ph ) = 0, (3c)
∂t
The coupling interface equations are handled as follows. The Neumann bound-
ary conditions (1i) are already considered in (3a)–(3c) by choosing proper test
functions leading to the discrete defined functions ψh , nh and ph . The Dirich-

let boundary conditions (1h) are evaluated at the grid points of the Dirichlet
boundary. For the approximation of the current equation (1g) we use Gauss
quadrature.
Under the assumptions of Theorem 1, we obtain as DAE index for the
resulting differential-algebraic equation (after space discretization of the cou-
pled problem) exactly the same as the perturbation index for the system
(1.1a–c, 1.1d’–g’, 1.1h–k). Furthermore, under the assumptions of Theorem
2, we obtain DAE index 1 for the system (1.1a–k) as expected concering the
perturbation result given in Theorem 2. For a proof we refer to [11].
4.1 The Software MECS
The multiphysical electric circuit simulator MECS [7] allows the time inte-
gration of electrical circuits using different models for semiconductor devices.
Beside the standard use of lumped models the user can choose distributed
models. So far, drift diffusion models are implemented. On the one hand, the
standard model equations are used as described in (1d)–(1f). On the other
hand one can also select the drift diffusion model where the Poisson equation
(1d) is replaced by the current conservation equation

∂
div jn + jp − εm ∇ψ = 0. (4)
∂t
For the space discretization, the standard finite element method as well as
a mixed finite element [8] is implemented. For the time integration of the
whole system, the user can choose between BDF methods [4], Runge Kutta
methods [12] and a general linear method [15].
4.2 Flip Flop Circuitry
Flip flop circuits represent a basic circuit structure of digital circuits serving
one bit of memory. Depending on the two input signals (set and reset), a flip
flop switches between two stable states. The circuit in Fig. 1 shows a realiza-
tion containing four MOSFETs (Metal Oxid Field Effect Transistors).
Fig. 1. Schematic diagram of a flip flop circuit

Fig. 2. Flip flop simulation results. On the left, the input signals Set and Reset. On
the right, the output realising two more or less stable states when applying different
frequencies
In Fig. 2, we see the output voltage for different applied frequencies de-
pending on the two given input signals. It shows that the stable states 0V
and 5V are not so stable when increasing higher frequencies, namely when
applying 1GHz. The simulation results may provide a more stable behavior
when decreasing the gate length or changing the doping profile. The advantage
of the simulation here is the possibility to study the influence of the dimen-
sions/doping/geometry of the semiconductors onto the switching behavior.
4.3 Voltage Controlled Oscillator (VCO)
As real benchmark we have tested a voltage controlled oscillator from the

transceiver described in the preceding introduction by Uwe Feldmann. It gen-
erates a 1.8 GHz signal with the amplitude of 2.5V (see simulation result in
Fig. 3 on the right). Continuing the simulation over longer time periods shows
the expected behavior to hold the frequency and amplitude as the last periods
on the figure indicate. The tested circuitry contained six MOSFETs that have
been simulated using the drift diffusion model.
On the left of Fig. 3, the electrostatic potential of one of the transistors
is shown at a randomly chosen time point. The simulation package MECS
provides the data at each time discretization point for each semiconductor for
the electrostatic potential and the charge carrier densities.
Fig. 3. VCO simulation results. On the left, the electrostatic potential of the third
MOSFET at a certain time point. On the right, the generated oscillator voltage
References
1. Alı̀ G., Bartel A. and Günther M.: Parabolic differential-algebraic models in
electrical network design, Multiscale Model. Simul. 4, No. 3, 813–838 (2005).
2. Bodestedt M.: Perturbation Analysis of Refined Models in Circuit Simula-
tion. Ph.D. thesis, Technical University Berlin, Berlin (2007).
3. Bodestedt M. and Tischendorf C.: PDAE models of integrated circuits and
index analysis. Math. Comput. Model. Dyn. Syst. 13, No. 1, 1–17 (2007).
4. Hanke M.: A New Implementation of a BDF method Within the Method of
Lines. http://www.nada.kth.se/∼hanke/ps/fldae.1.ps.
5. Gajewski H.: On existence, uniqueness and asymptotic behavior of solutions
of the basic equations for carrier transport in semiconductors. Z. Angew.
Math. Mech. 65, 101–108 (1985).
6. Griepentrog E. and März R.: Differential-algebraic equations and their nu-
merical treatment. Teubner, Leipzig (1986).
7. Guhlke C., Selva M.: Multiphysical electric circuit simulation (MECS) man-
ual. http://www.mi.uni-koeln.de/∼mselva/MECS.html.
8. Guhlke C.: Ein gemischter Finite Elemente-Ansatz zur gekoppelten Schal-
tungs- und Bauelementsimulation. Diploma thesis, Humboldt University of
Berlin (2006).
9. März. R.: Solvability of linear differential algebraic equations with properly
stated leading terms. Results Math. 45, pp. 88–105 (2004).
10. R. März.: Nonlinear differential-algebraic equations with properly formulated
leading term. Technical Report 01-3, Institute of Mathematics, Humboldt-
University of Berlin (2001).
11. Selva M., Tischendorf C.: Numerical Analysis of DAEs from Coupled Circuit
and Semiconductor Simulation. Appl. Numer. Math. 53, No. 2–4, 471–488
(2005).
12. Teigtmeier S.: Numerische Lösung von Algebro-Differentialgleichungen mit
proper formuliertem Hauptterm durch Runge-Kutta-Verfahren. Diploma
thesis. Humboldt University of Berlin (2002).
13. Tischendorf C.: Topological index calculation of DAEs in circuit simulation,

Surv. Math. Ind. 8, No. 3–4, 187–199 (1999).
14. Tischendorf C.: Coupled Systems of Differential Algebraic and Partial Differ-
ential Equations in Circuit and Device Simulation. Modeling and Numerical
Analysis, Habilitation thesis, Humboldt-University of Berlin (2004).
15. Voigtmann S.: GLIMDA solver. General LInear Methods for Differential Al-
gebraic equations. http://www.math.hu-berlin.de/∼steffen/software.html.
Wavelet-Collocation of Multirate PDAEs
for the Simulation of Radio Frequency Circuits
Stephanie Knorr, Roland Pulch, and Michael Günther
Bergische Universität Wuppertal, Fachbereich C, Gaußstr. 20, 42119 Wuppertal,

Germany, {knorr,pulch,guenther}@math.uni-wuppertal.de
1 Motivation
As already explained in the preceding introduction, radio frequency (RF) cir-
cuits introduce several difficulties for their numerical simulation, e.g. widely
separated time scales and a nonharmonic shape of digital signals. Multiscale
signals require huge computational effort in numerical integration schemes,
since the fast time scale restricts the step sizes, whereas the slow time scale
determines a relatively long integration interval. The occurrence of steep gra-
dients in digital signal structures demands an additional refinement of grids
in time domain methods. Moreover, the low smoothness of pulsed signals pos-
sibly causes further difficulties in the numerical simulation.
We present a wavelet-collocation scheme based on a multivariate modeling
of different time scales. Simulation results for a switched-capacitor circuit show
the efficient adaptive grid generation.
2 Multivariate Signal Model

In this section, a multivariate model is introduced for amplitude modulated
(AM) signals. It is applied to the differential-algebraic network equations and
the special structure of the resulting system is investigated.
2.1 Multivariate Model for AM Signals
We consider circuits with quasiperiodic steady state responses including

widely separated time scales. The core idea to efficiently describe multitone
signals is to decouple the time scales by associating a corresponding variable
to each of them. An m-tone quasiperiodic signal x : R → Cd reads

+∞
+∞

x(t) = ··· Xj1 ,...,jm exp i(j1 ω1 + · · · + jm ωm ) t (1)
j1 =−∞ jm =−∞
20 S. Knorr, R. Pulch, M. Günther
Fig. 1. AM signal x (left) and its multivariate function x̂ (right)
√
with Fourier coefficients Xj1 ,...,jm ∈ Cd , the imaginary unit i = −1 and
the fundamental frequencies ωl = 2π/Tl with according time scales Tl ,
l = 1, . . . , m. The multitone structure of the quasiperiodic signal (1) nat-
urally leads to the corresponding m-periodic multivariate function (MVF)
x̂ : Rm → Cd with
∞
∞

x̂(t1 , . . . , tm ) = ··· Xj1 ,...,jm exp i(j1 ω1 t1 +· · ·+jm ωm tm ) . (2)
j1 =−∞ jm =−∞
The MVF (2) is periodic in each time variable and the original signal can be
reconstructed following the ‘diagonal’ direction (t1 = t2 = · · · = tm ):
x(t) = x̂(t, . . . , t). (3)
To illustrate the efficiency of this multivariate modeling, Fig. 1 (left) shows

a high frequency pulse function x : R → R, whose amplitude is modulated by
a low frequency oscillation. Its MVF x̂ (right) exhibits a simple structure and
is determined in the rectangle [0, T1 ] × [0, T2 ] with T1 T2 .
2.2 MPDAE Model
The above multidimensional signal model is now applied to represent solutions

of the differential-algebraic network equations [GF99]. Below, we write the
system in the compact form
d

dt q x(t) = f b(t), x(t) , (4)
where q : Rd → Rd denotes charges and fluxes, b : R → Rs comprises

independent input signals (voltage and current sources) and unknown node
potentials and branch currents are collected in x : R → Rd . We are interested
in systems (4) with quasiperiodic solutions (1).
Wavelet-Collocation of MPDAEs for the Simulation of RF Circuits 21
Assuming m different time scales, we introduce MVFs x̂ : Rm → Rd of the

state variables and b̂ : Rm → Rs of the input signals. Regarding the DAE (4),
Brachtendorf et al. [BWLB96] introduce the corresponding multirate partial
differential-algebraic equation (MPDAE)
∂q(x̂) ∂q(x̂)
+ ···+ = f b̂(t1 , . . . , tm ), x̂(t1 , . . . , tm ) . (5)
∂t1 ∂tm
Given a solution x̂ of the MPDAE (5), a solution of the original DAE (4) can
be obtained via the reconstruction formula (3). For a detailed description of
the relation between DAE and MPDAE solutions, we refer to [Roy01].
To solve the MPDAE (5), we have to impose boundary conditions (BCs),
which determine the structure of the corresponding solution. Looking for an
m-tone quasiperiodic solution (1) of the DAE with periods T1 , . . . , Tm known
from input signals, we solve the MPDAE for an m-periodic MVF with BCs
x̂(t1 , . . . , tm ) = x̂(t1 + k1 T1 , . . . , tm + km Tm ) for all t1 , . . . , tm ∈ R
(6)
and all k1 , . . . , km ∈ Z .
2.3 Characteristic System
Starting from network equations that only comprise ordinary differential

equations (ODEs), the corresponding multirate partial differential equations
(MPDEs) are of hyperbolic type, where each component of the system exhibits
a derivative in the direction of the diagonal, see [PG02]. Thus, the information
transport takes place along characteristic projections, which are straight lines
in diagonal direction. The algebraic constraints in case of DAEs do not af-
fect this information transport and we are able to formulate the characteristic
system of the MPDAE:
d
dτ tl (τ ) = 1, l = 1, . . . , m,
d
(7)
dτ q x̄(τ ) = f b̂(t1 (τ ), . . . , tm (τ )), x̄(τ ) .
Thereby, the restrictions x̄(τ ) = x̂(t1 (τ ), . . . , tm (τ )) as well as the time vari-

ables depend on a parameter τ ∈ R. The part corresponding to the time
variables can be solved explicitly, leading to the characteristic projections
(t1 (τ ), . . . , tm (τ )) = (τ + c1 , . . . , τ + cm ) for arbitrary c1 , . . . , cm ∈ R. These
characteristic projections represent a continuum of parallel straight lines in
the domain of dependence. Inserting this result in the last equation of the
characteristic system (7) yields
d

dτ q x̄(τ ) = f b̂(τ + c1 , . . . , τ + cm ), x̄(τ ) . (8)
This family of DAE systems completely describes the transport of information

in the MPDAE (5).
The special hyperbolic structure of the MPDAE can be exploited to set up
a method of characteristics, see [PG02, Pul03]. In the following section, this
scheme will be demonstrated in detail. The well-posedness of the MPDAE

system is investigated in [KG06].
Furthermore, the idea of the multivariate modeling can also be carried
forward to frequency modulated signals. For a detailed survey on the MPDAE-
modeling, see [PGK07].
3 Wavelet-Collocation
In this section, we present the method of characteristics to obtain m-periodic

solutions x̂ of (5,6). Corresponding boundary value problems along the
characteristic projections are solved by a collocation scheme, where the
time-frequency localization of a wavelet basis is used to generate adaptive
grids.
3.1 Method of Characteristics
To shorten notations, we restrict ourselves to m = 2 different time scales and

consider the biperiodic boundary value problem (BVP) of the MPDAE
∂q(x̂) ∂q(x̂)
+ = f b̂(t1 , t2 ), x̂(t1 , t2 )
∂t1 ∂t2
with boundary conditions
x̂(t1 , t2 ) = x̂(t1 + T1 , t2 ) = x̂(t1 , t2 + T2 ) for all t1 , t2 ∈ R.
We exploit that the biperiodic solution is uniquely defined by its initial values
on the manifold {(t1 , t2 ) ∈ R2 : t2 = 0} and use the initial points
(t1 , t2 ) = ((j1 − 1)h1 , 0) for h1 := T1

n1 and j1 = 1, . . . , n1 .
As we have seen in the previous section, the information transport takes place
along parallel straight lines in direction of the diagonal. Figure 2 shows the
respective domain [0, T1 ] × [0, T2 ] with T1 T2 , where these characteristic
projections are indicated by dotted lines. Thus, we consider the unknown
functions
x̄j1 (τ ) := x̂((j1 − 1)h1 + τ, τ ) for j1 = 1, . . . , n1 , (9)
which extend along the diagonal direction. The corresponding systems (8)
now read
dq(x̄j1 )
(τ ) = f b̂((j1 − 1)h1 + τ, τ ), x̄j1 (τ ) for j1 = 1, . . . , n1 (10)
dτ
and have to be solved for τ ∈ [0, T2 ].
Fig. 2. Characteristic projections of the MPDAE in the domain of dependence
The periodicity in the second time scale leads to linear boundary conditions

x̄1 (T2 ) , . . . , x̄n1 (T2 ) = B x̄1 (0) , . . . , x̄n1 (0) , (11)
where the constant matrix B ∈ Rn1 d×n1 d describes an interpolation scheme

to determine the values x̄j1 (T2 ) using the values x̂((j1 − 1)h1 , T2 ) = x̄j1 (0).
As the separate characteristic systems (10) are only coupled by the bound-
ary conditions (11), this approach is much more efficient than a discretization
on a uniform grid, which performs an unnecessary strong coupling in both
coordinate directions, see [Pul03].
As we assume a strongly nonlinear behavior in the fast time scale t2 (recall
the MVF in Fig. 1 (right) with a pulsed oscillation in t2 -direction), an adaptive
discretization along the characteristic projections is essential to obtain an
efficient simulation technique.
3.2 Wavelet-Based Grid Generation
We aim at the detection of steep gradients in the solution for the generation
of an adaptive grid. The use of wavelets allows us both time and frequency
localization (whereas Fourier transforms are designed for frequency extraction,
only): the discrete wavelet transform of a signal x : R → R reads

wj,k (x) := x(t) ψj,k (t) dt, with ψj,k (t) := 2j/2 ψ(2j t − k). (12)
R
Thereby, the dilation and translation of the ‘mother-wavelet’ ψ is controlled

by the parameters j, k ∈ Z, which specify the frequency range and the time
localization, respectively.
Since we consider pulsed waveforms, wavelets of low order are an adequate
choice. We use
⎧
⎨ 1.5 − 4 |t|, |t| ≤ 0.5
ψ(t) = 0.5 |t| − 0.75, 0.5 < |t| ≤ 1.5 ,
⎩
0, elsewhere
which we call hat-wavelet in the following. It is obtained from a so-called

scaling function, which is the linear B-spline in our case.
Note that in contrast to frequently used orthonormal bases, our wavelet
basis gives rise to a biorthogonal system. The main difference is that for the
synthesis of a wavelet-transformed signal, the dual basis applies1 . So in our
case, the dual wavelet basis in fact undertakes the time-frequency localization
of the coefficients in the wavelet-representation of the respective signal x:

x(t) = j,k ψj,k ,
w
j,k∈Z
where w j,k is the discrete wavelet transform (12) of x with respect to the dual
wavelet ψ. For more details on biorthogonal wavelets, see e.g. [Coh92].
As already mentioned in the previous section, we have to solve BVPs
(10,11) and thus, we restrict our basis functions to a compact interval [0, L]
by ‘folding’ the hat-functions at the interval boundaries, see [CDV93]. In this
way, we obtain a so-called multiresolution analysis of L2 ([0, L]). Without loss
of generality we choose L ∈ R such that supp ψ ⊆ [0, L]. Then, L defines the
number of dilated basis functions on the bounded domain, which of course
can be transformed to the desired interval [0, T2 ].
[0,L]
For our numerical approximation, we regard a subspace VJ ⊂ L2 ([0, L])
[0,L]
of finite dimension, which is composed by a direct sum of a central space V0
2
(spanned by integer translates of the scaling function ) and hat-wavelet spaces
[0,L]
Wj , j = 0, . . . , J − 1 with according frequency localizations:
[0,L] [0,L]

J−1
[0,L]
VJ = V0 ⊕ Wj .
j=0
Due to the bounded domain, the number of translated wavelets ψj,k (12) for
[0,L]
k ∈ Ij ⊂ N, spanning the spaces Wj is finite:
[0,L]
Wj = span ψj,k (t) | k ∈ Ij for j = 0, . . . , J − 1.
To illustrate the capabilities of the hat-wavelets for our digital-like signals,

[0,20]
we approximate a pulse function p : [0, 1] → R in the subspace V2 , i.e.,
using the refinement of L + 1 = 21 basis functions in the central space. The
pulse function p is given in Fig. 3 as dashed line. There we show in three
[0,20] [0,20] [0,20]
plots the coefficients for the corresponding spaces V0 , W0 and W1 ,
respectively. The different markers indicate the size of these coefficients versus
the corresponding time localization centers, which are the locations of the basis
functions’ maximal values.
1
We choose the lowest order wavelet ψe out of the set of possible dual wavelets,
which supplement the hat-wavelets to a biorthogonal basis.
2
The B-spline scaling functions give us the coarsest approximation.
[0,20] [0,20] [0,20]

Fig. 3. Coefficients in V0 (left), W0 (middle) and W1 (right)
[0,20]
Notice that the shape of the pulse is well represented by V0 (scaling
functions are normalized to 1), only at the location of the steep gradients the
[0,20] [0,20]
refinement levels W0 , W1 have contributions, but attenuate fast. Thus,
we can employ this localization of the steep gradients to define an adaptive
grid: given a guess of the solution, we ‘simply’ have to inspect the size of the
respective wavelet transforms.
Grid points associated with ‘small’ coefficients (smaller than a given
threshold) are left out and grid points are added, where the coefficients ex-
ceed a given upper threshold. Of course, the relative level of these coefficients
(compared within the same subspace) is most instructive. A more detailed
description of the grid generation algorithm can be found in [BKP07].
3.3 Wavelet-Based Collocation
Now, we combine the method of characteristics introduced in Sect. 3.1 and

the wavelet-based grid generation described in Sect. 3.2 by solving the DAE
BVP (10, 11) with a wavelet-collocation scheme.
The hat-wavelets serve as ansatz functions to approximate the solutions (9)
along each characteristic projection for τ ∈ [0, T2 ]:
. N
j1 (τ ) :=
x̄j1 (τ ) = x cj1 ,i Ψ i (τ ). (13)
i=1
This approximation holds for each component of x̄j1 : [0, T2 ] → Rd , cj1 ,i ∈ Rd

for all j1 = 1, . . . , n1 and i = 1, . . . , N . Thereby, all N = 2J L + 1 basis
[0,L]
functions of VJ (see Sect. 3.2) are denoted by Ψi : [0, T2 ] → R, i = 1, . . . , N .
We demand that for each characteristic projection, the DAE (10) has to
be fulfilled in n2 = N − 1 collocation points τm ∈ [0, T2 ], m = 1, . . . , n2 :

dτ q xj1 (τ )
d
= f b̂((j1 − 1)h1 + τm , τm ), xj1 (τm ) (14)
τ =τm
for j1 = 1, . . . , n1 . Together with the n1 BCs (11) for the approximations (13),

x1 (0) , . . . , x
n1 (0) = B x1 (T2 ) , . . . , x
n1 (T2 ) , (15)
we end up with n1 · N nonlinear equations for the unknown coefficients

cj1 ,i .
Of course, we want to use the time-frequency localization property of our
wavelets to determine an adaptive set of collocation points for the numerical
simulation. We establish an equidistant start grid in [0, T2 ]:
τm = (m − 1)h2 , h2 = T2
N −1 , m = 1, . . . , N = 2J L + 1,
which corresponds to the time localization centers of the basis functions, cf.
Sect. 3.2. Then, we solve the nonlinear system (14, 15) using a Newton-type
method. Starting values have to be determined for all grid points in the domain
[0, T1 ]× [0, T2 ]. After a few iterations (in our test examples two iterations were
sufficient), the wavelet coefficients already contain enough information about
the structure of the solution to determine an adaptive grid as outlined in
Sect. 3.2. Then, the Newton iteration is continued on the new mesh to solve
for the respective wavelet coefficients.
4 Simulation Results
We investigate a switched-capacitor circuit, the Miller integrator, which is de-
picted in Fig. 4 (left). It contains two MOS-transistors3 M1 and M2 driven by
two complementary pulse functions pa and pb . The output at node 3 approx-
imates the negative integral4 of the input vin . Applying a sinusoidal input
signal, the respective output voltage uref3 (reference solution) can be seen in
Fig. 4 (right). The discrete sampling via the pulse functions causes the signal
to be rough, which is revealed by the zoom in this figure.
The index-1 differential-algebraic model describing the network behavior
can be found in [KF06].
The sinusoidal input signal vin determines the slow time scale T1 = 10−5 s,
whereas the pulses exhibit a period of T2 = 2.5 · 10−8 s.
Fig. 4. Miller integrator circuit (left) and output voltage uref

3 (right)
3
MOS stands for metal oxide semiconductor.
4
The scaling of the output voltage depends on the design parameters.
We apply the multidimensional signal model from Sect. 2 and solve the
resulting biperiodic BVP (10, 11) of the MPDAE by the wavelet-collocation
introduced in Sect. 3. We discretize on n1 = 30 characteristic projections
and use an equidistant start grid of N = 121 points (with J = 2 and L =
30). A different adaptive grid is determined for each characteristic projection,
which results in an average of n2 = 60 mesh points (with a finest resolution
as for an equidistant grid with 241 points).
In Fig. 5 (left), we display the first component û1 of the MPDAE solution,
which shows a strong influence of the pulse functions. The steep gradients
are sharply detected by the wavelet basis, which results in the adaptive grid
depicted in Fig. 5 (right). The MVF û1 is plotted on a common grid for
all characteristic projections (using respective evaluations of the basis func-
tions).
In comparison to the reference solution uref
3 in Fig. 4 (right), the recon-
structed DAE solution u3 is depicted in Fig. 6 (right). The reference solution
to the reconstructed DAE solution u1 (Fig. 6, left) is shown in Fig. 1 (left) in
Sect. 2.1. Both components of the approximative solution show a good agree-
ment with the reference solution, which is confirmed by a discrete L2 -error of
only about 2 %. The pulsed structure of the signals is sharply resolved and no
undesired peaks occur.
Fig. 5. Miller Integrator: MPDAE solution û1 (left) and adaptive grid (right)
Fig. 6. Miller Integrator: Reconstructed DAE solutions u1 (left) and u3 (right)

5 Conclusions
A multivariate modeling is applied to decouple widely separated time scales.

We presented a novel algorithm to solve the resulting MPDAE system. An
adaptive wavelet-based collocation scheme yields an efficient simulation of RF
circuits including analog and digital signal structures. The capability of the
wavelet basis to detect steep gradients and the efficiency of the numerical
scheme are verified by simulation results for a switched-capacitor circuit.
References
[BKP07] Bartel, A., Knorr, S., Pulch, R.: Wavelet-based adaptive grids for the
simulation of multirate partial differential-algebraic equations. To ap-
pear in: Appl. Numer. Math.
[BWLB96] Brachtendorf, H.G., Welsch, G., Laur, R., Bunse-Gerstner, A.: Numer-
ical steady state analysis of electronic circuits driven by multi-tone sig-
nals. Electrical Engineering, 79, 103–112 (1996)
[Coh92] Cohen, A.: Biorthogonal wavelets. In: Chui, C. (ed.) Wavelet Analysis
and its Applications II: Wavelets – A Tutorial in Theory and Applica-
tions, Academic Press, New York, 123–152 (1992)
[CDV93] Cohen, A., Daubechies, I., Vial, P.: Wavelets on the interval and fast
wavelet transforms. Appl. Comput. Harm. Anal., 1, 54–81 (1993)
[GF99] Günther, M., Feldmann, U.: CAD based electric circuit modeling I:
mathematical structure and index of network equations. Surv. Math.
Ind. 8:1, 97–129 (1999)
[KF06] Knorr, S., Feldmann, U.: Simulation of pulsed signals in MPDAE-
modelled SC-circuits. In: Di Bucchianico, A., Mattheij, R.M.M.,
Peletier, M.A. (eds.) Progress in Industrial Mathematics at ECMI 2004,
Mathematics in Industry 8, Springer, Berlin, 159–163 (2006)
[KG06] Knorr, S., Günther, M.: Index analysis of multirate partial differential-
algebraic systems in RF-circuits. In: Anile, A.M., Alı̀, G., Mascali, G.
(eds.) Scientific Computing in Electrical Engineering, Mathematics in
Industry 9, Springer, Berlin, 93–100 (2006)
[Pul03] Pulch, R.: Finite difference methods for multi time scale differential
algebraic equations. Z. Angew. Math. Mech., 83:9, 571–583 (2003)
[PG02] Pulch, R., Günther, M.: A method of characteristics for solving multi-
rate partial differential equations in radio frequency application. Appl.
Numer. Math., 42, 397–409 (2002)
[PGK07] Pulch, R., Günther, M., Knorr, S.: Multirate partial differential alge-
braic equations for simulating radio frequency signals. To appear in:
Euro. Jour. Appl. Math.
[Roy01] Roychowdhury, J.: Analyzing circuits with widely-separated time scales
using numerical PDE methods. IEEE Trans. CAS I, 48, 578–594 (2001)
Numerical Simulation of Thermal Effects
in Coupled Optoelectronic
Device-Circuit Systems
Markus Brunk1 and Ansgar Jüngel2

1
Institut für Mathematik, Universität Mainz, Staudingerweg 9, 55099 Mainz,
Germany, brunk@mathematik.uni-mainz.de
2
Institut für Analysis und Scientific Computing, Technische Universität Wien,
Wiedner Hauptstr. 8–10, 1040 Wien, Austria
juengel@anum.tuwien.ac.at
1 Introduction
The control of thermal effects becomes more and more important in modern
semiconductor circuits like in the simplified CMOS transceiver representa-
tion described by U. Feldmann in the above article Numerical simulation of
multiscale models for radio frequency circuits in the time domain. The stan-
dard approach for modeling integrated circuits is to replace the semiconduc-
tor devices by equivalent circuits consisting of basic elements and resulting in
so-called compact models. Parasitic thermal effects, however, require a very
large number of basic elements and a careful adjustment of the resulting large
number of parameters in order to achieve the needed accuracy.
Therefore, it is preferable to model those semiconductor devices which are
critical for the parasitic effects by semiconductor transport equations. The
transport of electrons in the devices is modeled here by the one-dimensional
energy-transport model allowing for the simulation of the electron temper-
ature. The electric circuits are described by modified nodal analysis. Thus,
the devices are modeled by (nonlinear) partial differential equations, whereas
the circuit is described by differential-algebraic equations. The coupled model,
which becomes a system of (nonlinear) partial differential-algebraic equations,
is numerically discretized in time by the 2-stage backward difference formula
(BDF2), since this scheme allows to maintain the M-matrix property, and the
semi-discrete equations are approximated by a mixed finite-element method.
The objective is the simulation of a benchmark high-frequency transceiver
circuit, using a laser diode as transmitter and a photo diode as receiver. The
optical field in the laser diode is modeled by recombination terms and a rate
equation for the number of photons in the device. The optical effects in the photo
diode are described by generation terms. The numerical results show that the
thermal effects can modify significantly the behavior of the transmitter circuit.
30 M. Brunk, A. Jüngel
2 Modeling
Circuit Modeling
A well-established mathematical description of electric circuits, consisting of

resistors, capacitors, and inductors (RCL circuit) is the modified nodal analy-
sis (MNA) which can be easily extended to circuits containing semiconductor
devices. In the following, the circuit model is described.
The circuit is replaced by a directed graph. The RLC branches are char-
acterized by the incidence matrix A, and the semiconductor branches are
characterized by the semiconductor incidence matrix AS . The basic tools for
the MNA are the Kirchhoff laws and the current-voltage characteristics for
the basic elements,
dq dΦ
Ai + AS jS = 0, v = A e, iR = g(vR ), iC = (vC ), vL = (iL ),
dt dt
where i, v, and e are the vectors of branch currents, branch voltages, and
node potentials, respectively, and jS denotes the semiconductor current (see
below). The variable g denotes the conductivity of the resistor, q is the charge
of the capacitor, and Φ the flux of the inductor. The incidence matrix A is
assumed to consist of the block matrices AR , AC , AL , Ai , and Av , where
the indices i and v indicate the current source and voltage source branches,
respectively.
Denoting by is = is (t) and vs = vs (t) the given input functions for the
sources, we obtain the system for the charge-oriented MNA [13],
dq
AC (A e) + AR g(A
R e) + AL iL + Av iv + AS jS = −Ai is , (1)
dt C
dΦ
(iL ) − A
L e = 0, A
v e = vs , (2)
dt
for the unknowns e(t), iL (t), and iv (t). Equation (1) expresses the Kirchhoff
current law, the first equation in (2) is the voltage-current characteristic for
inductors, and the last equation allows to compute the node potentials.
Semiconductor Device Modeling
The flow of minority charge carriers (holes) in the device is modeled by the
drift-diffusion model for the hole density p. The electron flow is described by
the energy-transport equations [8]. The first model consists of the conservation
law for the hole mass, together with a constitutive relation for the hole current
density. The latter model also includes the conservation law for the electron
energy and a constitutive relation for the energy flux. Both models can be
derived from the semiconductor Boltzmann equation (see [8] and references
therein). They are coupled through recombination-generation terms and the
Coupled Optoelectonic Device-Circuit Systems 31
Poisson equation for the electric potential. More precisely, the electron density
n, the hole density p, and the electron temperature T are obtained from the
parabolic equations
μ−1 −1
n ∂t g1 − divJn = −R(μn g1 , p), ∂t p + divJp = −R(μ−1
n g1 , p) (3)
3
μ−1 −1 −1
n ∂t g2 − divJw = −Jn · ∇V + W (μn g1 , T ) − T R(μn g1 , p), (4)
2
where g1 = μn n and g2 = μn w are auxiliary variables allowing for a drift-
diffusion-type formulation of the fluxes [8], w = 32 nT is the thermal energy,
and μn and μp are the electron and hole mobilities, respectively. The electron
current density Jn , the energy flux Jw , and the hole current density Jp are
given by
g1 g2
Jn = ∇g1 − ∇V, Jw = ∇g2 − ∇V, Jp = −μp (∇p + p∇V ). (5)
T T
The equations are coupled self-consistently to the Poisson equation for the
electric potential V ,
λ2 ΔV = μ−1n g1 − p − C(x), (6)
where λ is the scaled Debye length and the given function C(x) models the
doping profile. The functions
3 n(T − TL ) np − n2i
W (n, T ) = − and R(n, p) = (7)
2 τ0 τp (n + ni ) + τn (p + ni )
with the (scaled) energy relaxation time τ0 and lattice temperature TL =

1 describe the relaxation to the equilibrium energy and Shockley-Read-Hall
recombination-generation processes with intrinsic density ni and electron and
hole lifetimes τn and τp , respectively.
Equations (3)–(6) are solved in the bounded semiconductor domain Ω,
where some initial values nI , pI , and TI are imposed. The boundary of Ω is
assumed to split into two parts. On the insulating parts of the boundary ΓN ,
it is assumed that the normal components of the current densities and of the
electric field vanish. For the temperature, homogenous Neumann boundary
conditions are assumed as in [1]. We have shown in [5] that boundary layers
for the particle densities can be avoided if Robin-type boundary conditions
similar as in [14] are employed on the remaining boundary parts,
n − θn Jn · ν = na and p + θp Jp · ν = pa on ∂Ω\ΓN , (8)
where θn and θp are some parameters and na and pa are ambient particle
densities. Notice that in the one-dimensional simulations presented below,
ΓN = ∅.
Coupling to the Circuit
The boundary conditions for the electric potential at the contacts are deter-
mined by the circuit and are given as

C
V = ei + Vbi on Γk , t > 0, where Vbi = arsinh , (9)
2ni
if the terminal k of the semiconductor is connected to the circuit node i.
The semiconductor current entering the circuit consists of the electron cur-
rent Jn , the hole current Jp , and the displacement current Jd = −λ2 ∂t ∇V ,
guaranteeing charge conservation. The current leaving the semiconductor de-
vice at terminal k, corresponding to the boundary part Γk , is defined by

jk = (Jn + Jp + Jd ) · ν ds,
Γk
where ν is the exterior unit normal vector to ∂Ω. We denote by jS the vector
of all terminal currents except the reference terminal. In the one-dimensional
case, there remains only one terminal, and the current through the terminal
at x = 0 is given by
jS (t) − (Jn (0, t) + Jp (0, t) − ∂t jd,S (0, t)) = 0, jd,S − λ2 Vx = 0, (10)
where the circuit equations (1)–(2) have to be appropriately scaled [6].
The complete coupled system consists of equations (1)–(10) forming an
initial boundary-value problem of partial differential-algebraic equations. The
system resulting from the coupled circuit drift-diffusion equations has at most
index 2 and it has index 1 under some topological assumptions [3, 13]. No
analytical results are available for the coupled circuit energy-transport system.
Optoelectronic Device Modeling
The interaction between optical and electrical effects is modeled by recom-

bination-generation terms appearing in (7). In the following, we present the
model used in the numerical simulations and we refer to [6] for a discussion
about the model simplifications.
For a vertical photo diode, the supplied photons generate free charge car-
riers generating the photo current. We model this effect by adding to the
Shockley-Read-Hall term (7) the generation rate Gopt (x) of free carriers at
depth x, caused by the (scaled) optical irradiation power Pin with angular
frequency ω [10],
Pin
Gopt (x) = η(1 − r) αab e−αab x , (11)
ωA
where the physical parameters are the quantum efficiency η, the reflectivity r
of the irradiated surface with area A, the reduced Planck constant = h/2π,
and the optical absorption αab .
The laser diode is modeled by a pin heterostructure diode in which the

intrinsic (active) region consists of a low-band gap material causing carrier
confinement. The active region works as a Fabry-Perot laser cavity and can be
modeled as a single mode laser. The band discontinuities are simply described
by adding a constant band potential to the electric potential V in the active
region [9]. Additionally to (7), spontaneous and stimulated recombination is
introduced,
c
Rspon = Bnp and Rstim = g(n)|Ξ|2 S, (12)
μopt
respectively, where B is the spontaneous recombination parameter, c the speed
of light μopt the refractive index of the material, g(n) the optical gain depend-
ing on the electron density, |Ξ|2 the intensity distribution of the optical field,
which is a solution of the waveguide equation [6], and S = S(x, t) is the
number of photons in the device.
The optical gain is approximated by g(n) = g0 (n−nth) [7], with differential
gain g0 and threshold density nth . In the lasing mode we can employ the quasi-
neutral assumption n ≈ p in the active region such that the gain becomes
approximately g(p) = g0 (p − nth ) in the recombination term occurring in the
hole equation (3). This allows for a discretization that guarantees positivity
of the discretized hole density [6]. The number of photons S is balanced by
the rate equation

∂t S = vg (β − α)S + Rspon , where β = (g(n) − αbg )|Ξ|2 ds,
Ωa
α is the total loss by external output and scattering, αbg denotes the back-
ground loss, and Ωa is the transverse cross section of the active region. We
prescribe the initial condition S(·, 0) = SI in Ω. Finally, the output power is
computed from of the number of photons by
c
Pout = ω αf |Ξ|2 S (13)
μopt
(see [2]), where αf denotes the facet loss of the laser cavity.
3 Numerical Simulations
The system of coupled partial differential-algebraic equations is first dis-
cretized in time by the BDF2 method since this scheme allows to maintain
the M-matrix property of the final discrete system. The Poisson equation
is discretized in space by the linear finite-element method. Then the discrete
electric potential is piecewise linear and the approximation of the electric field
−Vx is piecewise constant.
The semi-discrete continuity equations at one time step are of the form
gj
−Jj,x + σj gj = fj , Jj = gj,x − Vx , j = 1, 2, (14)
T
with the current densities J1 = Jn and J2 = Jw and some expressions σj and

fj . These equations are discretized in space by a hybridized exponentially
fitted mixed finite-element method [8]. We employ the finite elements of [12]
since they guarantee the positivity of the discrete variables if positive initial
and Dirichlet boundary data are prescribed and if σj ≥ 0, fj ≥ 0 for j = 1, 2.
This property also holds for the Robin conditions (8) [6]. Finally, the nonlinear
discrete system is solved by Newton’s method.
Rectifying Circuit
As a test example we consider a rectifying circuit containing four silicon pn

diodes as in [5] (Fig. 1). Each of the diodes has the length L = 0.1 μm (and
Ly = 0.1 μm, Lz = 2 μm) or L = 1 μm (and Ly = 1 μm, Lz = 20 μm) and
a maximum doping of 1022 m−3 . We have chosen the resistance R = 100 Ω
and the voltage source v(t) = U0 sin(2πωt) with U0 = 5 V and ω = 1 GHz
or ω = 10 GHz. The remaining physical parameters are listed in Table 1.
As initial conditions we take thermal equilibrium densities in the device and
vanishing node potentials and branch currents in the circuit. The initial value
for the displacement current is determined by (10). A computation according
to [11] shows that these values are consistent for the coupled DAE system.
Fig. 1. Left: Graetz circuit. Right: Thermal energy in a pn diode during one oscil-
lation of Vin
Table 1. Physical parameters for a silicon pn-junction diode

Parameter Physical meaning Numerical value
q elementary charge 1.6 · 10−19 As
s permittivity constant 1.05 · 10−10 As/Vm
UT thermal voltage at 300 K 0.026 V
μn /μp low-field carrier mobilities 1500/450 cm2 /Vs
τn /τp carrier lifetimes 10−6 /10−5 s
ni intrinsic density 1016 m−3
τ0 energy relaxation time 4 · 10−13 s
Fig. 2. Output signal of the Graetz circuit for two frequencies of the voltage source.
Left: ω = 1 GHz and L = 0.1 μm. Right: ω = 10 GHz and L = 1 μm
In Fig. 1 the energy density in one of the diodes during one oscillation
is presented. As expected, we observe a high thermal energy in forward bias
(t ∈ [0, 50 ps]), whereas it is negligible in backward bias (t ∈ [50 ps, 100 ps])
although the electron temperature (not shown) may be very large around the
junction [4].
The impact of the thermal effects on the electrical behavior of the cir-
cuit is shown in Fig. 2. The figure clearly shows the rectifying behaviour of
the cuircuit. The largest current is obtained from the drift-diffusion model
since we have assumed a constant electron mobility such that the drift is
unbounded with respect to the modulus of the electric field. The stationary
energy-transport model is not able to catch the capacitive effect at the junc-
tion which is particularly remarkable at higher frequencies.
Optoelectronic Circuit
Next we consider a AlGaAs/GaAs laser diode with a digital input signal.

The transmitted signal is received by a silicon photo diode coupled to a high-
pass filter (see Fig. 3). We have taken a capacitance of 10 pF, the resistances
Fig. 3. Left: Laser and photo diode with a high-pass filter. Right: energy density in
the laser diode for signal v(t) = 2 sin(2πt109 ) V
Table 2. Physical parameters for a laser diode of Al0.7 Ga0.3 As (superscript A) and
GaAs (superscript G). Parameters without superscript are taken for both materials
Parameter Physical meaning Numerical value
Ly /Lz extension of device in y/z-direction 10−6 /10−5 m
Un /Up band potentials in active region 0.1/ − 0.1 V
B spontaneous recombination parameter 10−16 m3 /s
nth threshold density 1024 m−3
αf /αbg mirror/optical background loss 5000/4000 m−1
A
s /s
G
material permittivity 1.08 · 10−10 /1.14 · 10−10 As/Vm
μAn /μ G
n electron mobilities 2300/8300 cm2 /Vs
μp /μG
A
p hole mobilities 145/400 cm2 /Vs
μA G
opt /μopt refractive index 3.3/3.15
A
ni /ni G
intrinsic density 2.1 · 109 /2.1 · 1012 m−3
g0G differential gain in GaAs 10−20 m2
R1 = 1 MΩ, R2 = 100 Ω, and R3 = 1 kΩ, and a backward bias of 0.2 V. The

laser diode has the length of 1 μm with an intrinsic region of 0.1 μm length in
the center of the device. The doping concentration is −1024 m−3 in the p-doped
region, 1024 m−3 in the n-doped region, and 1018 m−3 in the intrinsic region.
The photo diode has a size of L = 6 μm, Ly = 10−5 m and Lz = 10−4 m. For
the quantum efficiency we assume η = 0.5, for the surface reflectivity r = 0.3
and α = 5000 m−1 for the absorption. The remaining parameters are taken
from Tables 1 and 2.
In Fig. 3 the energy density in the laser diode during one half oscillation
is shown. After having passed the threshold, the energy density increases
tremendously in the active region. This is due to carrier confinement in the
heterostructure, as in the lasing mode the carrier density is very high in the
active region.
Finally, we operate the transmitter with a 1 GHz digital signal of 2 V.
In Fig. 4 the light output signal and the received signal by the high-pass
Fig. 4. Output of the laser diode and the high-pass filter for the stationary and
transient energy-transport model with a digital input signal of 1 GHz
filter is presented. Again we observe that the transient energy-transport model

responds better to the capacitive effects in high-frequency circuits than the
stationary model.
4 Conclusion
We have presented a coupled model consisting of the circuit equations from
modified nodal analysis and the energy-transport model for semiconductor
devices, resulting in a system of nonlinear partial differential-algebraic equa-
tions. This system allows for a direct simulation of thermal effects and can
help to improve compact models of integrated circuits. The coupled model is
tested on a Graetz circuit and a high-frequency transmitter with laser and
photo diodes. The results show the impact of the thermal energy on the cir-
cuit. Compared to the constant-temperature drift-diffusion model, the output
signal is smaller due to thermal effects.
With decreasing size of the basic components in integrated circuits and
special power devices, the thermal interaction between circuit elements will
increase in importance in the near future. Therefore, we need to model not only
the carrier temperature but also the device temperature and the interaction
between the circuit elements. Thus, a heat equation for the temperature of
the semiconductor lattice needs to be included in the presented model. This
extension is currently under investigation. We expect that the resulting model
will improve significantly the prediction of hot-electron effects and hot spots
in integrated circuits.
References
1. A. M. Anile, V. Romano, and G. Russo. Extendend hydrodynamic model of
carrier transport in semiconductors. SIAM J. Appl. Math. 61 (2000), 74–101.
2. U. Bandelow, H. Gajewski, and R. Hünlich. Fabry-Perot lasers: thermodynamic-
based modeling. In: J. Piprek (ed.), Optoelectronic Devices. Advanced Simula-
tion and Analysis. Springer, Berlin (2005), 63–85.
3. M. Bodestedt. Index Analysis of Coupled Systems in Circuit Simulatiom. Licen-
tiate Thesis, Lund University, Sweden, 2004.
4. F. Brezzi, L. Marini, S. Micheletti, P. Pietra, R. Sacco, and S. Wang. Discretiza-
tion of semiconductor device problems. In: W. Schilders and E. ter Maten (eds.),
Handbook of Numerical Analysis. Numerical Methods in Electromagnetics. El-
sevier, Amsterdam, Vol. 13 (2005), 317–441.
5. M. Brunk and A. Jüngel. Numerical coupling of electric circuit equations and
energy-transport models for semiconductors. To appear in SIAM J. Sci. Com-
put., 2007.
6. M. Brunk and A. Jüngel. Simulation of thermal effects in optoelectronic devices
using energy-transport equations. In preparation, 2007.
7. S. L. Chuang. Physics of Optoelectronic Devices. Wiley, New York, 1995.
8. P. Degond, A. Jüngel, and P. Pietra. Numerical discretization of energy-

transport models for semiconductors with non-parabolic band structure. SIAM
J. Sci. Comp. 22 (2000), 986–1007.
9. M. Friedrich. Ein analytisches Modell zur Simulation von Silizium-Germanium
Heterojunction-Bipolartransistoren in integrierten Schaltungen. PhD Thesis,
Ruhr-Universität Bochum, Germany, 2002.
10. G. R. Jones, R. J. Jones, and W. French. Infrared HgCdTe Optical Detectors.
In: J. Piprek (ed.), Optoelectronic Devices. Advanced Simulation and Analysis.
Springer, Berlin (2005), 381–403.
11. R. Lamour. Index determination and calculation of consistent initial values for
DAEs. Computers Math. Appl. 50 (2005), 1125–1140.
12. L. D. Marini and P. Pietra. New mixed finite element schemes for current con-
tinuity equations. COMPEL 9 (1990), 257–268.
13. C. Tischendorf. Modeling circuit systems coupled with distributed semiconduc-
tor equations. In: K. Antreich, R. Bulirsch, A. Gilg, and P. Rentrop (eds.),
Modeling, Simulation, and Optimization of Integrated Circuits, Internat. Series
Numer. Math. 146 (2003), 229–247.
14. A. Yamnahakki. Second-order boundary conditions for the drift-diffusion equa-
tions for semiconductors. Math. Models Meth. Appl. Sci. 5 (1995), 429–455.
Efficient Transient Noise Analysis
in Circuit Simulation
Georg Denk1 , Werner Römisch2 , Thorsten Sickenberger2, and Renate

Winkler2
1
Qimonda AG, Products, München, Germany, georg.denk@qimonda.com
2
Institute of Mathematics, Humboldt-Universität zu Berlin, Germany
{romisch,sickenberger,winkler}@math.hu-berlin.de
Transient noise analysis means time domain simulation of noisy electronic

circuits. We consider mathematical models where the noise is taken into ac-
count by means of sources of Gaussian white noise that are added to the
deterministic network equations, leading to systems of stochastic differential
algebraic equations (SDAEs). A crucial property of the arising SDAEs is the
large number of small noise sources that are included. As efficient means of
their integration we discuss adaptive linear multi-step methods, in particu-
lar stochastic analogues of the trapezoidal rule and the two-step backward
differentiation formula, together with a new step-size control strategy. Test
results including real-life problems illustrate the performance of the presented
methods.
1 Transient Noise Analysis in Circuit Simulation

The increasing scale of integration, high clock frequencies and low supply volt-
ages cause smaller signal-to-noise ratios. Reduced signal-to-noise ratio means
that the difference between the wanted signal and noise is getting smaller.
A consequence of this is that the circuit simulation has to take noise into ac-
count. In several applications the noise influences the system behaviour in an
essentially nonlinear way such that linear noise analysis is no longer satisfac-
tory and transient noise analysis, i.e., the simulation of noisy systems in the
time domain, becomes necessary (see [4, 16]). For an implementation of an
efficient transient noise analysis in an analog simulator, both an appropriate
modelling and integration scheme is necessary (see [3]).
Here we deal with the thermal noise of resistors as well as the shot noise
of semiconductors that are modelled by additional sources of additive or mul-
tiplicative Gaussian white noise currents that are shunt in parallel to the
noise-free elements. Thermal noise ith of resistors is caused by the thermal
motion of electrons and is described by Nyquist’s theorem. Shot noise ishot of
40 G. Denk et al.
pn-junctions, caused by the discrete nature of currents due to the elementary

charge, is modelled by Schottky’s formula and inherits noise intensities that
depend on the deterministic currents:

2kT
ith = ξ(t) , ishot = qe |idet (u)|ξ(t) . (1)
R
Here ξ(t) is a standard Gaussian white noise process, R denotes the resistance,
T is the temperature, k = 1.38 · 10−23 is Boltzmann’s constant, idet (u) is
the characteristic of the noise-free current through the pn-junction and qe =
1.60 · 10−19 is the elementary charge.
Combining Kirchhoff’s current law with the element characteristics and
using the charge-oriented formulation yields a stochastic differential-algebraic
equation (SDAE) of the form (see e.g. [15], or for the deterministic case [6])
d m
A q(x(t)) + f (x(t), t) + gr (x(t), t)ξr (t) = 0 , (2)
dt r=1
where A is a constant singular incidence matrix determined by the topology

of the dynamic circuit parts, the vector q(x) consists of the charges and the
fluxes, and x is the vector of unknowns consisting of the nodal potentials
and the branch currents through voltage-defining elements. The term f (x, t)
describes the impact of the static elements, gr (x, t) denotes the vector of
noise intensities for the r-th noise source, and ξ is an m-dimensional vector of
independent Gaussian white noise sources (see e.g. [4, 16]). One has to deal
with a large number of equations as well as of noise sources, where one can
and has to exploit the fact that compared to the other quantities the noise
intensities gr (x, t) are small.
Though the system (2) formally differs only by the additional noise term
from the deterministic system, a completely different mathematical framework
has to be applied. A serious mathematical description begins by introducing
the Brownian motionor the Wiener t process that is caused by integrating the
t
white noise “W (t) = 0 ξ(s)ds = 0 dW (s)” (see e.g. [1]). Problem (2) is then
understood as a stochastic integral equation
t t m t

A q(X(s)) + f (X(s), s)ds + gr (X(s), s)dWr (s) = 0, t ∈ [t0 , T ] ,
t0 t0 t0
r=1
(3)
where the second integral is an Itô-integral, and W denotes an m-dimensional
Wiener process (or Brownian motion) given on the probability space (Ω, F , P )
with a filtration (Ft )t≥t0 . The solution is a stochastic process depending on
the time t and on the random sample ω. The value at fixed time t is a random
variable X(t, ·) = X(t) whose argument ω is usually dropped. For a fixed
sample ω representing a fixed realization of the driving Wiener process, the
function X(·, ω) is called a realization or a path of the solution. Due to the
Efficient Transient Noise Analysis in Circuit Simulation 41
influence of the Gaussian white noise, typical paths of the solution are nowhere
differentiable.
The theory of stochastic differential equations distinguishes between the
concepts of strong, i.e., pathwise solutions and weak, i.e., the distribution law
of solutions. We decided to aim at the simulation of solution paths, i.e., strong
solutions that reveal the phase noise that is of particular interest in case of
oscillating solutions. From the solution paths statistical data of the phase as
well as moments of the solution can be computed in a post-processing step.
We therefore use the concept of strong solutions and strong (mean-square)
convergence of approximations.
By the implicitness of the systems (2) or (3) and the singularity of the
matrix A the model is not an SDE, but an SDAE. We refer to [15] for analytical
results as well as convergence results for certain drift-implicit methods.
In this paper we discuss adaptive linear multi-step methods, in particu-
lar stochastic analogues of the trapezoidal rule and the two-step backward
differentiation formula, see Sect. 2. The applied step-size control strategy is
described in Sect. 3. Here we extensively use the smallness of the noise. In
Sect. 4 new ideas for the control both of time and chance-discretization are
discussed. Test results including real-life problems that illustrate the perfor-
mance of the presented methods are given in Sect. 5.
2 Adaptive Numerical Methods

The key idea to design methods for SDAEs is to force the iterates to fulfill the
constraints of the SDAE at the current time-point. Here we consider stochas-
tic analogues of methods that have proven very useful in the deterministic
circuit simulation. Paying attention to the DAE structure, the discretization
of the deterministic part (drift) is implicit, whereas the discretization of the
stochastic part (diffusion) is explicit.
We consider stochastic analogues of the variable coefficient two-step back-
ward differentiation formula (BDF2 ) and the trapezoidal rule, where only the
increments of the driving Wiener process are used to discretize the diffusion
part. Analogously to the Euler-Maruyama scheme we call such methods multi-
step Maruyama methods. The variable step-size BDF2 Maruyama method for
the SDAE (3) has the form (see [11] and, for constant step-sizes, e.g. [2])
q(X ) + α1, q(X−1 ) + α2, q(X−2 )
A + β0, f (X , t )
h
m
ΔWr m
ΔWr−1
+ gr (X−1 , t−1 ) − α2, gr (X−2 , t−2 ) = 0, (4)
r=1
h r=1
h
= 2, . . . , N . Here, X denotes the approximation to X(t ), h = t − t−1 ,

and ΔWr = Wr (t ) − Wr (t−1 ) ∼ N (0, h ) on the grid 0 = t0 < t1 < · · · <
tN = T . The coefficients α1, , α2, , β0, depend on the step-size ratio κ =
42 G. Denk et al.
h /h−1 and satisfy the conditions for consistency of order one and two in the
deterministic case. By construction the scheme has order 1/2 in the stochastic
case (see [11]). A correct formulation of the stochastic trapezoidal rule for
SDAEs requires more structural information (see [12]). It should implicitly
realize the stochastic trapezoidal rule for the so called inherent regular SDE
of (3) that governs the dynamical components. Both the BDF2 Maruyama
method and the stochastic trapezoidal rule of Maruyama type have only an
asymptotic order of strong convergence of 1/2, i.e.,
X(t ) − X L2 (Ω) := max (E|X(t ) − X |2 )1/2 ≤ c · h1/2 , (5)
=1,...,N
where h := max=1,...,N h is the maximal step-size of the grid. For additive

noise the order may be 1. This holds true for all numerical schemes that
include only information on the increments of the Wiener process.
However, the noise densities given in Section 1 contain small parameters
and the error behaviour is much better. In fact, the errors are dominated by
the deterministic terms as long as the step-size is large enough [2, 11]. In more
detail, the error of the given methods behaves like O(h2+ εh + ε2 h1/2 ), when ε
is used to measure the smallness of the noise, i.e., gr (x, t) = εĝr (x, t), r = 1,...,m
where ε 1. Thus we can expect order 2 behaviour if h ε. Higher numerical
effort for higher deterministic order pays off only if the noise is very small.
3 Local Error Estimates

The smallness of the noise allows special estimates of the local error terms,
which can be used to control the step-size. We aim at an efficient estimate
of the mean-square of local errors by means of a number of simultaneously
computed solution paths. This leads to an adaptive step-size sequence that
is identical for all paths. For the drift-implicit Euler-Maruyama scheme this
step-size control has been presented in [10], see also [4, 16].
In [13, 14] the authors extended this strategy to stochastic linear multi-step
methods with deterministic order 2 and provided a reliable error estimate. Let
approximates the dominating local error in AX by
L

2κ 2κ2
L = c h · f (X , t ) − 2κ f (X−1 , t−1 ) + f (X−2 , t−2 ) , (6)
κ + 1 κ + 1
where c is the error constant of the related deterministic scheme. This esti-
mate is based on already computed values of the drift term. Recall that L is
a vector valued random variable as is the solution X . For the measurement
of errors we use the mean-square norm in L2 (Ω). In dependence on the small
parameter ε and the step-size h the L2 -norm of the local error behaves like
3/2
O(h3 + εh + ε2 h ). The term of order O(h3 ) dominates the local error be-
3/2
haviour as long as h3 is much larger than εh , i.e., ε2/3 h . Under this
L2 approximates the local error.
condition also the expression L
Depending on the available information we will monitor different quantities

to satisfy accuracy requirements,
L2 to match a given tolerance for X ,
i) control (A + h β0, J )−1 L

ii) control L L2 to match a given tolerance for AX , or
iii) control A− L L2 to match a given tolerance for P X .
Here J is the Jacobian of the drift function f w.r.t. the first variable, and
A− denotes the pseudo inverse of A, and P is an appropriate projector. Since
(A/h +β0, J ) = 1/h ·(A+h β0, J ) is the Jacobian of the discrete scheme (4)
this matrix (or a good approximation to it) and its factorization are usually
available. In case of M sampled paths, the L2 -norm in i)–iii) is estimated by
using the M values L i , i = 1, . . . , M . For example, in case i) we use

1
M 2 1/2
−1 i −1 i
(A + h β0, J ) L ≈ (A + h β0, J ) L . (7)
L2 M i=1
4 A Solution Path Tree Algorithm

In the analysis so far, the number M of sample paths has not been specified
yet. It influences the sampling error in the approximation of the L2 -norm in the
error estimate (7). We have L L2 = η̂ + ϑ , where η̂ is the approximation
of the dominating local error term based on M sample paths and ϑ is the
sampling error.
Our aim in tuning the number of paths is to balance the local error and
the sampling error. Let d be a given upper bound for the sampling error ϑ at
time t , e.g. calculated as an approximation of the higher deterministic error
term of order O(h4 ). We then derive the best number M of paths by

1 μ̂2 · σ̂2
M = 2 2 , (8)
d μ̂ + σ̂2
where μ̂ and σ̂2 are estimates of the mean and the standard deviation of the
error estimate at time-point t , respectively. Here x denotes the smallest
integer greater or equal to x.
The best number of paths M depends on the time-point t and is realized
by approximate solutions generated on a tree of paths that is extended, re-
duced or kept fixed adaptively. In [9] the authors describe the construction of
a solution path tree in detail. The method uses probabilities πi ( = 1, . . . , N ;
i = 1, . . . , M ) to weight the solution paths. Figure 1 gives an impression,
how a solution path tree looks like. At each time-step the optimal expansion
or reduction problem is formulated by means of combinatorial optimization
models. The path selection is modelled as a mass transportation problem in
terms of the L2 -Wasserstein metric (see [5] in context of scenario reduction in
stochastic programming). The algorithm has been implemented in practice.
The results presented in the next section show its performance.
44 G. Denk et al.
Fig. 1. A solution path tree: Variable time-points t , solution states xi and path
weights πi
5 Numerical Results
Here we present numerical experiments for the stochastic BDF2 applied to
two circuit examples. The first one is a small test problem, for which we have
used an implementation of the adaptive methods discussed in the previous
sections in Fortran code. To be able to handle real-life problems, a slightly
modified version of the schemes has been implemented in Qimonda’s in-house
analog circuit simulator TITAN. The second example shows the performance
of this industrial implementation.
A MOSFET Inverter Circuit
We consider a model of an inverter circuit with a MOSFET transistor, under

the influence of thermal noise. The related circuit diagram is given in Fig. 2.
The MOSFET is modelled as a current source from source to drain that is
controlled by the nodal potentials at gate, source and drain.
The thermal noise of the resistor and of the MOSFET is modelled by
additional white noise current sources that are shunt in parallel to the original,
Fig. 2. Thermal noise sources in a MOSFET inverter circuit

Fig. 3. Simulation results for the noisy inverter circuit: (left) 1 path 127(+29 re-
jected) steps; (right) 100 paths 134(+11 rejected) steps
noise-free elements. To highlight the effect of the noise, we scaled the diffusion
coefficient by a factor of 1000.
In Fig. 3 we present simulation results, where we plotted the input voltage
Uin and values of the output voltage e1 versus time. Moreover, the applied
step-sizes, suitably scaled, are shown by means of single crosses. We compare
the results for the computation of a single path (left picture) with those for
the computation of 100 simultaneously computed solution paths (right pic-
ture). The additional solid lines show two different solution paths, the dashed
line gives the mean of 100 paths and the outer thin lines the 3σ-confidence
interval for the output voltage e1 . We observe that using the information of an
ensemble of simultaneously computed solution paths smoothes the step-size
sequence and considerably reduces the number of rejected steps, when com-
pared to the simulation of a single path. The computational cost that is mainly
determined by the number of computed (accepted+ rejected) steps is reduced.
We have applied the solution path tree algorithm to this example. The
upper graph in Fig. 4 shows the computed solution path tree together with
the applied step-sizes. The lower graph shows the simulation error (solid line),
its error bound (dashed line) and the used number of paths (marked by ×),
vs. time. The maximal number of paths was set to 250.
The results indicate that there exists a region from nearly t = 1·10−8 up to
t = 1.5·10−8 where we have to use much more than 100 paths. This is exactly
the area in which the MOSFET is active and the input signal is inverted.
Outside this region the algorithm proposes approximately 70 simultaneously
computed solution paths.
A Voltage Controlled Oscillator
As an industrial test application we us a voltage controlled oscillator that is

a simplified version of a fully integrated 1.3 GHz VCO for GSM in 0.25 µm
standard CMOS (see [8]). For simulation, the oscillator is embedded in a test
environment. The VCO is tunable from about 1.2 GHz up to 1.4 GHz. The
46 G. Denk et al.
Fig. 4. Simulation results for the noisy inverter circuit: Solution path tree and
step-sizes (top), sampling error, its error bound and the number of paths (bottom)
Fig. 5. Noisy transient output signal of a VCO
unknowns of the VCO in the MNA system are the charges of the six capacities,
the fluxes of the four inductors, the 15 nodal potentials and the currents
through the voltage sources. This circuit contains 5 resistors and 6 MOSFETs,
which induce 53 sources of thermal or shot noise. To make the differences
Fig. 6. Boxplots of the phase noise, scaled by a factor of 500
between the solutions of the noisy and the noise-free model more visible, the
noise intensities had been scaled by a factor of 500.
Numerical results obtained with a combination of the BDF2 and the trape-
zoidal rule are shown in Fig. 5, where we plotted the difference of the nodal
potential V (7) − V (8) of node 7 and 8 versus time. The solution of the noise-
free system is given by a dashed line. Four sample paths (dark solid lines) are
shown. They cannot be considered as small perturbations of the deterministic
solution, phase noise is highly visible.
To analyze the phase noise we performed 10 simultaneous simulations with
different initializations of the pseudo-random numbers. In a postprocessing
step we computed the length of the first 50 periods for each solution path
and then from these the corresponding frequencies. In Fig. 6 the mean μ
of the frequencies (horizontal lines), the smallest and the largest frequencies
(boundaries of the vertical thin lines) and the boundaries of the confidence
interval μ ± σ (the plump lines) are presented, where σ was computed as the
empirical estimate of the standard deviation. The mean appears increased and
differs by about +0.25% from the noiseless, deterministic solution.
Further on, the frequencies vacillate from 1.18 GHz (−0.95%) up to 1.21
GHz (+1.55%). So the transient noise analysis shows that the voltage con-
trolled oscillator runs in a noisy environment with increased frequencies and
smaller phases, respectively.
48 G. Denk et al.
6 Conclusions
Quite similar to deterministic circuit simulation, it is essential to have the

application in mind while developing new efficient algorithms. We have pre-
sented variable step-size two-step schemes for SDAEs which require only the
increments of the driving Wiener process. Though these schemes possess only
convergence order 1/2 from a theoretical point of view, they show order 2 in
circuit simulation, as the deterministic terms dominate the errors. This can
be considered in the step-size control. Taking the stochastic properties of the
circuit into account leads to an increased efficiency of the methods.
An important application of transient noise analysis is to get insight into
the statistical properties of the solution paths. We showed that the number of
paths necessary for this purpose varies with the time-points. By implementing
a solution path tree algorithm, it is possible to save computing time or to get
more accurate outputs compared to a naive approach which would require to
calculate all paths over the complete integration interval.
These results allow an efficient transient noise simulation which helps the
designer to cope with challenges due to technology progress. Further improve-
ments may include parallelisation and the handling of flicker noise.
References
1. Arnold, L.: Stochastic differential equations: Theory and Applications. Wiley,
New York, 1974.
2. Buckwar, E., Winkler, R.: Multi-step methods for SDEs and their application
to problems with small noise. SIAM J. Num. Anal., 44(2), 779–803 (2006)
3. Denk, G.: Circuit simulation for nanoelectronics. In: M. Anile, G. Alı̀, G. Mascali
(Eds.): Scientific Computing in Electrical Engineering (Mathematics in Industry
9), 13–20. Springer, Berlin, 2006.
4. Denk, G., Winkler, R.: Modeling and simulation of transient noise in circuit
simulation. Math. and Comp. Modelling of Dyn. Systems, 13(4), 383–394 (2007)
5. Dupačová J., Gröwe-Kuska, N., Römisch, W.: Scenario reduction in stochastic
programming. Math. Program., Ser. A 95, 493–511 (2003)
6. Günther, F., Feldmann, U.: CAD-based electric-circuit modeling in industry I.
Mathematical structure and index of network equations. Surv. Math. Ind., 8,
97–129 (1999)
7. Higham, D.J.: An algorithmic introduction to numerical simulation of stochastic
differential equations. SIAM Review, 43, 525–546 (2001)
8. Tiebout, M.: A fully integrated 1.3 GHz VCO for GSM in 0.25 µm standard
CMOS with a phasenoise of −142 dBc/Hz at 3MHz offset. In: Proceedings 30th
European Microwave Conference, Paris (2000)
9. Römisch, W., Sickenberger, Th.: On generating a solution path tree for efficient
step-size control. In preparation.
10. Römisch, W., Winkler, R.: Stepsize control for mean-square numerical methods
for SDEs with small noise. SIAM J. Sci. Comp., 28(2), 604–625 (2006)
11. Sickenberger, T.: Mean-square convergence of stochastic multi-step methods

with variable step-size. J. Comput. Appl. Math., to appear
12. Sickenberger, T., Winkler, R.: Efficient transient noise analysis in circuit simu-
lation. In: Proceedings of the GAMM Annual Meeting 2006, Berlin, Proc. Appl.
Math. Mech. 6(1), 55–58 (2006)
13. Sickenberger, T., Weinmüller, E., Winkler, R.: Local error estimates for moder-
ately smooth problems: Part I – ODEs and DAEs. BIT Numerical Mathematics,
47(2), 157–187 (2007).
14. Sickenberger, T., Weinmüller, E., Winkler, R.: Local error estimates for moder-
ately smooth problems: Part II – SDEs and SDAEs. Preprint 07-07, Institut für
Mathematik, Humboldt-Universität zu Berlin (2007). Submitted for publication.
15. Winkler, R.: Stochastic differential algebraic equations of index 1 and applica-
tions in circuit simulation. J. Comput. Appl. Math., 157(2), 477–505 (2003)
16. Winkler, R.: Stochastic differential algebraic equations in transient noise ana-
lysis. In: M. Anile, G. Alı̀, G. Mascali (Eds.): Scientific Computing in Electrical
Engineering (Mathematics in Industry 9), 151–158. Springer, Berlin, 2006.
Part II
Thin Films
Numerical Methods for the Simulation
of Epitaxial Growth and Their Application
in the Study of a Meander Instability
Frank Haußer1 , Felix Otto2 , Patrick Penzler2 , and Axel Voigt3

1
FB II, Technische Fachhochschule Berlin, Luxemburger Str. 10, 13353 Berlin,
Germany, hausser@tfh-berlin.de
2
IAM, Universität Bonn, Wegelerstr. 10, 53115 Bonn, Germany
{otto,penzler}@iam.uni-bonn.de
3
IWR, Technische Universität Dresden, Zellescher Weg 12–14, 01062 Dresden,
Germany, axel.voigt@tu-dresden.de
Summary. The surface morphology of thin crystalline films grown by molecular

beam epitaxy (MBE), a technique to produce high-quality, almost defect-free crys-
tals, is strongly influenced by kinetic processes on an atomistic scale. To incorporate
these effects in a continuum model requires some care. Here we use a step flow
model, which is a free boundary problem for the position of atomic height steps on
the crystalline surface. We present two complementary approaches to derive a numer-
ical method for solving this problem: a front tracking ansatz and a diffuse interface
approximation. The numerical methods are used to study the nonlinear regime of
a step meandering instability
1 Epitaxial Growth
1.1 Introduction
Epitaxial growth is a modern technology of growing crystalline films that

inherit atomic structures from substrates. It produces almost defect-free, high
quality materials that have a wide range of device applications. Microscopic
processes in epitaxial growth include the deposition of atoms or molecules,
atom adsorption and desorption, adatom (adsorbed atom) diffusion, adatom
island nucleation, the attachment and detachment of adatoms to and from
island boundaries or terrace steps, and island coalescence.
There are various models for epitaxial growth of thin films that are distin-
guished by different scales in time and space, see Voigt [2005] for an overview.
These models range from full atomistic descriptions using molecular dynam-
54 F. Haußer et al.
Fig. 1. Epitaxial growth of Si(001). The figure shows atomistically flat terraces,
which are separated by steps of atomic height. Courtesy of Polop, Bleikamp, and
Michely
ics (MD) and kinetic Monte Carlo (KMC) methods over discrete-continuous
models in which only the growth direction is resolved on an atomistic scale
and the lateral direction is coarse-grained, to fully continuous models which
describe the thin film as a smooth hypersurface.
An initially atomistically flat surface typically does not remain flat dur-
ing growth, but is subject to various instabilities. There are essentially three
types of instabilities which influence the film morphology during growth:
step bunching, step meandering and mound formation, see e.g. Politi et al.
[2000], Krug [2005]. They all have their origin on the atomistic scale and re-
sult from asymmetries in the energy barriers for individual hops of atoms
on the surface. However, a fully atomistic description of the film is lim-
ited to sample sizes of several nm and thus far off from any feature size
in semiconductor devices. In order to predict the surface morphology on
larger length scales, continuum models are required which incorporate the
instabilities generated on the atomistic scale. Fully continuous models with
these properties still have to be derived. On a mesoscopic scale, discrete-
continuum models – so called step flow models – are promising candidates.
These models are discrete in the growth direction but continuous in the lat-
eral directions. Figure 1 shows a scanning tunneling microscopy (STM) im-
age of a Si(001) surface, consisting of large terraces separated by atomistic
height steps, which motivates this modeling approach. Atomistic hops on ter-
races are modeled by a continuum diffusion equation. The atomistic processes
of attachment and detachment at the atomic height steps are incorporated
by appropriate boundary conditions. Moreover, the atomistically rough steps
are treated as smooth curves and the local geometry enters via the curva-
ture.
Numerical Methods for the Simulation of Epitaxial Growth 55
Fig. 2. Linear instability: Some wave numbers of an initially small perturbation of

a straight step grow and the step starts to meander
1.2 Step Meandering Instability
In this work we are concerned with a meandering instability of a step. The

seminal work of Bales and Zangwill [1990] showed the influence of a (ter-
race) Ehrlich–Schwoebel barrier on the morphology of a step. The Ehrlich–
Schwoebel barrier is an atomistic energy barrier atoms have to overcome
if they attach to a step from the upper terrace [Ehrlich and Hudda, 1966,
Schwoebel and Shipsey, 1966]. This barrier leads to an effective uphill current
resulting in a meander instability. This means that given an initially straight
step with a perturbation of small magnitude, there are some wave numbers
of the perturbation which will grow, see Fig. 2 for a sketch. Based on a linear
stability analysis, an explicit formula for the growth rate ω(k) depending on
the wave number k has been given in Bales and Zangwill [1990].
Experimentally observed step meanders do not fall into the linear regime,
as they show large amplitudes. Thus, they cannot be described within the
analytic treatment. Our goal is therefore to use numerical tools to study step
meandering in the nonlinear regime, where the step flow might even break
down.
1.3 Step Flow Model
The layer-by-layer growth of the crystalline surface is described by the classical

Burton–Cabrera–Frank (BCF) model [Burton et al., 1951], which is a semicon-
tinous model – discrete in the height, but continuous in the lateral directions.
Here, the surface is assumed to consist of flat terraces separated by atom-high
steps; the steps are modeled as continuous curves, see Fig. 3.
Vapor atoms arriving at the surface become adatoms (ad-sorbed atoms)
and diffuse on the flat terraces. The diffusive fluxes at the steps lead to growth
via propagating steps. If nucleation of new islands or steps on the terraces can
be neglected, the growth dynamics are essentially described by the attachment
kinetics at the steps, i.e., the boundary conditions of the adatom density
at the terrace boundaries. This leads to a free boundary problem for the
adatom densities (concentrations) on the terraces with free boundaries given
Fig. 3. Step flow model: Vapour atoms are deposited on a surface, where they be-
come ad(sorbed)atoms and diffuse on atomistically flat terraces; eventually adatoms
attach/detach at steps
by the step position. On each terrace, the adatom density c obeys the following
diffusion equation:
∂t c + ∇ · j = F, j = −D∇c , (1a)
where D is the diffusion constant and F is the deposition flux rate. Desorption
of adatoms has been neglected, which is valid in typical MBE experiments
[Maroutian et al., 1999]. The fluxes of adatoms to a step are given by
j± := ±(c± v − j± · n) , (1b)
where subscripts “+” and “−” denote quantities at a step up (i.e. on the lower
terrace) and a step down, respectively, n denotes the normal pointing from
upper to lower terrace and v is the normal velocity of the step. Assuming
first order kinetics for the attachment/detachment of adatoms at the steps,
the diffusive fluxes at the step (terrace boundary) are proportional to the
deviation of the adatom density from equilibrium ceq , i.e., the adatom density
satisfies the following kinetic boundary conditions at a step:
j± = k± (c± − ceq ) . (1c)
With this notation, asymmetric attachment rates 0 < k− < k+ model the
(terrace) Ehrlich-Schwoebel (ES) effect. The equilibrium density ceq at a step
is given by the linearized Gibbs–Thomson relation
ceq = c∗ (1 + Γ κ), Γ = a2 (γ + γ )/(kB T ) ,
where c∗ is the constant equilibrium density for a straight step, a2 is the atomic
area with a being the lattice spacing and κ denotes the curvature of the step
(we define the curvature of a circular island as positive); kB is Boltzmann’s
constant, T the temperature and γ = γ(θ) denotes the step free energy per
unit length, which may depend on the local orientation θ = θ(n).
Finally, the normal velocity of a step is given by

v = a2 (j+ + j− ) − a∂s js , js = −Dst ∂s (Γ κ) , (1d)
where ∂s denotes the tangential derivative along the step and Dst is the diffu-
sion constant of atoms along the step. The second term in the velocity law (1d)
represents edge diffusion of edge atoms along the step, whereas the first term
ensures mass conservation. For a more detailed description of the step flow
model see e.g. Krug [2005].
2 Modeling and Numerical Methods

In this section we present two complementary powerful numerical methods to
solve the free boundary problem as described in Sect. 1.3.
In the first part we describe a front-tracking-type method, where the free
boundaries (the steps) and the adatom concentrations are discretized on two
independent meshes. Using an operator splitting ansatz and a careful dis-
cretization based on linear finite elements leads to an accurate and efficient
semi-implicit scheme. However, topological changes are not naturally included
in this approach and would require awkward remeshing.
To overcome this limitation, in the second part, we present a diffuse-
interface approximation. Via this approximation, we transform the (predom-
inantly) third-order free boundary problem into a fourth-order problem –
without boundaries – for an auxiliary quantity φ mimicking the shape of the
crystal. The step positions are then given by level sets of φ. Both φ and the
adatom concentration use the same mesh.
2.1 Front Tracking Method

We will review the front tracking type finite element method for discretizing
the moving boundary problem (1a)–(1d) that has been developed in Bänsch
et al. [2004]. First note that the moving boundary problem may be naturally
divided into two subproblems: the adatom diffusion and the step evolution.
Thus we proceed as follows:
(a) We derive a weak formulation for the time-dependent diffusion equation.
To avoid the complexity in the spatial discretization near boundaries, in
each time step, we extend the diffusion equation from terraces of same
height to the whole computational domain. The extended equation is dis-
cretized using the linear finite element method. This leads to two diffusion
equations to be solved in each time step: one for the adatom densities on
terraces of even height and one for the adatom densities on terraces of odd
height. Note that this properly reflects the fact, that the kinetic boundary
conditions (1c) in general lead to discontinuities of the adatom densities
at the steps, which are resolved by having two degrees of freedom at the
terrace boundaries.
Fig. 4. (a) Finite element mesh (1d) for terrace boundaries (b) Locally refined
Finite element mesh (2d) for adatom density
(b) The geometric motion of the island boundaries includes both the mean cur-
vature flow originating from the Gibbs-Thomson relation and the line diffu-
sion. It is treated in a variational formulation utilizing the curvature vector,
and discretized by a semi-implicit front-tracking method using parametric
finite elements. This method is adapted with modification from Bänsch
et al. [2005b] and is generalized to allow for anisotropic line energies.
This operator splitting approach results in the following numerical scheme: in
each time
(i) solve the geometric evolution equation for the terrace boundaries (steps)
using the step position and the adatom densities of the last time step
(ii) solve the diffusion equation for the adatom density using the updated step
position.
We remark that the two-dimensional (2d) and the one-dimensional (1d) finite
element meshes are essentially independent from each other. To obtain sat-
isfactory computational results, meshes with sufficiently fine resolutions are
needed for both the adatom diffusion equation and the boundary evolution
equation. Thus, it is indispensable to use adaptivity in order for the method
to be efficient. We use simple error indicators within an h-adaptive method to
locally increase the spatial resolution. In Fig. 4 we give a prototype example
of a 1d mesh representing two steps and the corresponding locally refined 2d
mesh for the adatom densities.
Adatom Diffusion
Let ci (x, t) denote the adatom density on the domain (terraces) Ωi (t) of atomic
height i, with boundaries Γ− (t) and Γ+ (t) representing the downward and
upward steps, respectively. Multiplying Equation (1a) with a smooth time-
independent test function φ and integrating over the domain Ωi (t) yields

∂t ci φ + D∇ci · ∇φ − D∇ci · nφ + D∇ci · nφ = Fφ
Ωi Ωi Γ− Γ+ Ωi

Using the identity dt d
c φ = Ωi (t) ∂t ci φ − Γ+ (t) ci vφ + Γ− (t) ci vφ and
Ωi (t) i
plugging in the boundary conditions (1c) leads to

d
ci φ+ D∇ci · ∇φ + k+ ci φ + k− ci φ
dt Ωi (t) Ωi (t) Γ+ (t) Γ− (t)

= Fφ + k+ c∗ (1 + Γ κ)φ + k− c∗ (1 + Γ κ)φ (2)
Ωi (t) Γ+ (t) Γ− (t)
Now we use the first-order implicit scheme to discretize the time derivative:
Consider discrete time instants t0 < t1 < · · · with time steps Δtm = tm+1 −tm
and denote Ωi (tm ) = Ωim , ci (tm ) = cm
i . Substituting

d 1
ci φ −→ c m+1
φ− m
ci φ
dt Ωi (t) tm+1 − tm Ωim+1 i Ωim
and using extended variables

i (x), D, F ) : x ∈ Ω̄i
(cm m
cm m m
i (x), Di (x), Fi (x) =
(0, 0, 0) : x ∈ Ω \ Ωim
leads to a weak formulation of the time discretized eq. (2) on the whole (time
independent) domain Ω:

cm+1 − cm
i i
φ+ Dim+1 ∇cm+1
i · ∇φ + k+ c m+1
i φ + k− cm+1
i φ
Ω Δtm Ω Γ+ Γ−

= k+ c∗ (1 + Γ κ)φ + k− c∗ (1 + Γ κ))φ + Fim+1 φ (3)
Γ+ Γ− Ω
Equation (3) is discretized in space using linear finite elements leading to

a symmetric positive definite system to be solved in each time step. However,
on elements where coefficients are discontinuous, a careful integration is nec-
essary when assembling the matrices. Also note, that the boundaries Γ+ , Γ−
are the solutions Γ+m+1 , Γ−m+1 of the boundary evolution. For a more detailed
description see Bänsch et al. [2004].
Boundary Evolution
To solve the evolution equation (1d) the boundary conditions (1c) are used to
¯ ∩ Ω̄i )
arrive at a geometric evolution equation (for each boundary Γi := Ωi+1
of the form
v = f − βΓ κ + α∂ss Γ κ , (4)
where β = a2 (k+ + k− )c∗ and α = aDst are positive constants, Γ = Γ (θ) is
a possibly orientation dependent positive function and the adatom densities of
the upper and lower terrace, c+ , c− enter via f = a2 k+ (c+ −c∗ )+a2 k− (c− −c∗ ).
Equation (4) may be interpreted as an equation for (1d) anisotropic surface

diffusion with lower order terms and is solved following the ideas of Bänsch
et al. [2005b], for an adaption to the anisotropic case see Bänsch et al. [2005a].
First a second order splitting of the fourth order equation is performed by in-
troducing the curvature vector κ := κn and using the relation −∂ss x = κ
with the position vector x. Then a semi-implicit time discretization is intro-
duced by representing the next boundary Γ (tm+1 ) in terms of the current
boundary Γ (tm ) by updating the position vector xm+1 = xm + Δtm vm+1 .
The resulting system of second order equations is discretized in space using
linear parametric finite elements.
2.2 Diffuse Interface Approximation
We now discuss the second approach: a diffuse-interface approximation (DIA).

Let us first show how to imagine such an approximation.
The BCF model is two-dimensional, but every domain Ωi is labeled with
a discrete height, so one can imagine it as a three-dimensional landscape with
sharp jumps, see Fig. 5a. The DIA can now be thought of as a smeared-out
version of this landscape, where the sharp jumps are replaced by a smooth
transition region of width ε, see Fig. 5b.
Diffuse-interface models have been used for various applications, e.g. spin-
odal decomposition. The connection between diffuse-interface models and the
nonlocal sharp-interface models of the type considered here were first shown
by Pego [1989] and Caginalp [1989] using formal asymptotic analysis. Alikakos
et al. [1994] calculated the error between the true solution of the Cahn–Hilliard
equation and the formal expansion, thereby gaining a rigorous proof of the
convergence ε → 0, as long as the underlying solution stays smooth. For
a general phase-field equation, the rigorous proof for a convergence to a sharp
interface model was given by Caginalp and Chen [1998].
Diffuse-interface approximations for step flow growth have already been
introduced in Liu and Metiu [1997] and Karma and Plapp [1998], but none of
these models allowed the incorporation of the Ehrlich–Schwoebel-barrier.
Fig. 5. (a) BCF model: each domain is associated with a discrete height, thus form-
ing a three-dimensional landscape with sharp interfaces. (b) Diffuse-interface ap-
proximation: the sharp interfaces are “smeared out”, resulting in a smooth function
Nondimensionalization
Before we proceed to the diffuse-interface approximation, we rewrite equations

(1). We will focus on a regime where we have no edge diffusion and the adatom
density has enough time to relax to its quasi-stationary equilibrium on the
terraces. We state the equations in terms of the excess adatom density w :=
c − c∗ . Rescaling time and space as in Otto et al. [2004], we get a Mullins–
Sekerka-type free boundary problem:
−Δw = f in Ωi (5a)
± ± ±
w = κ ± ζ ∇w · n on Γ (5b)
−
v = ∇(w − w ) · n
+
on Γ (5c)
The dimensionless parameters ζ ± are antiproportional to the attachment rate

k± . In the following, we will only consider the case of unlimited attachment
to a step up, i.e. ζ + = 0.
A Cahn–Hilliard-type Equation
Consider the equation

δEε
∂t φ + ∇ · −M (φ)∇ (φ) = f . (6)
δφ
The energy functional Eε (φ) is the Ginzburg–Landau free energy with a double-
well potential G

ε
Eε (φ) = |∇φ|2 + ε−1 G(φ) , G(φ) = 18φ2 (1 − φ)2 (7)
Ω 2
and M (φ) is a mobility function modeling the Ehrlich–Schwoebel barrier:
M (φ) = (1 + ε−1 ζ − σ(φ))−1 ,
see Fig. 6. Here σ(φ) is an asymmetric function in φ. The potential G is

restricted to the interval [0, 1] and then periodically continued, so that we get
a multiwell potential with equally deep wells at the integers.
In Otto, Penzler, Rätz, Rump, and Voigt [2004] it was shown by formal
asymptotic expansion that the above equation yields for ε → 0 the BCF-model
(5) including the Ehrlich–Schwoebel barrier.
The position of the boundary is given by the level sets
{φ = Z + 12 }
and the L2 –differential δE

δφ (φ) =: w is the approximation of the excess density
ε
from equations (5), see Fig. 7.

Fig. 6. The potential G (dashed line) and the mobility function M (solid line):
coming from an upper terrace (“φ = 2”), the atoms experience reduced mobility
while attaching to the step (“φ = 1.5”). On the other hand, coming from a lower
terrace to the step, there is no reduction in mobility. This models the Ehrlich–
Schwoebel barrier
Fig. 7. Excess density w for data as in Fig. 5b. Clearly visible are the boundary
values w = κ at a step up, the jump due to the ES-barrier and the smooth solution
of −Δw = 1 on the terraces
Discretization
In order to formulate a finite element approximation, we rewrite equation (6)

by introducing the flux J:
∂t φ + ∇ · J = f (8a)
1 δEε
J = −∇ (φ) . (8b)
M (φ) δφ
As we shall see, this splitting yields a conformal spatial discretization and
a time discretization which leads to a symmetric positive definite system. As
boundary conditions we use
• equilibrium and no-flux boundary conditions: ∇φ · ν = 0, J · ν = 0 or
• periodic boundary conditions on a square domain Ω = [0, Lx ] × [0, Ly ].
The weak form in either case is

d
φζ + (∇ · J)ζ = fζ ∀ζ ∈ L2 (Ω) (9a)
dt Ω
Ω Ω
1 δEε
J · J˜ − (φ)(∇ · J)
˜ =0 ∀J˜ ∈ H(∇·, Ω) . (9b)
Ω M (φ) Ω δφ
Here, the space H(∇·, Ω) is defined as

H(∇·, Ω) := F ∈ L2 (Ω)2 : ||∇ · F ||L2 (Ω) < ∞ .
Time Discretization
The guideline for our discretization is a gradient flow formulation: consider

the manifold

M = φ φ = const , Tφ M = v v=0
Ω Ω
with the metric

gφ (u, v) = u, vH −1 (Ω) := ∇w · M (φ)∇w̃
M
Ω
where w and w̃ are the solutions of
∇ · (M (φ)∇w) = u and ∇ · (M (φ)∇w̃) = v ,
respectively, using Neumann boundary conditions. The gradient flow is then

given by ∂t φ = −∇E. By definition of the gradient, this is equivalent to
gφ (∂t φ, v) + diff Eε (φ)v = 0 ∀v ∈ Tφ (M ) ,
where we use the notation

d
diff Eε (φ)v := Eε (φ + δv) .
dδ δ=0
Inserting the definition of the metric yields

δEε
∇w · M (φ)∇w̃ + (φ)v = 0 ∀v ∈ Tφ M (10)
Ω Ω δφ
with
∂t φ + ∇ · (−M (φ)∇w) = 0 and v = ∇ · (M (φ)∇w̃) .
Defining J := −M (φ)∇w and J˜ := −M (φ)∇w̃ finally gives
∂t φ + ∇ · J = 0

1 δEε
J · J˜ = (φ)(∇ · J)
˜ ∀J˜ ∈ H(∇·, Ω)
Ω M (φ) Ω δφ
which is equivalent to equation (9) with f = 0. Partial integration in (10)

reveals w = δE
δφ (φ).
ε
The idea is now to use a semi-implicit time discretization, where the only
explicit quantity is the base point of the metric.
To allow for bigger time steps, we use a second-order timestep method.
Our first choice was the Crank–Nicolson method, but since perturbations of
high frequency are only damped very weakly (amplification factor → 1), this
method is not suitable for Cahn–Hilliard-type equations. Following Weikard
[2002], we chose a method by Bristeau et al. [1987] (BGP-scheme). To solve
an equation ∂t φ = F (φ), it uses two intermediate steps per time step τ :
1 ∗
φ − φk = αF (φ∗ ) + βF φk (11a)
θτ
1
(φ∗∗ − φ∗ ) = βF (φ∗∗ ) + αF (φ∗ ) (11b)
(1 − 2θ)τ
1 k+1
φ − φ∗∗ = αF φk+1 + βF (φ∗∗ ) (11c)
θτ
with
1 √ √
θ = 1 − √ , α = 2 − 2, β = 2 − 1 .
2
To keep notation simple, we present the main ideas for time discretization
using the backward Euler scheme.
To handle the nonlinearity, we use Newton’s method. Again to simplify
notation, we show only one Newton–step, i.e. a linearization of the nonlinear-
ity, and understand the following equations to be valid for all J˜ ∈ H(∇·, Ω).
This yields

1 k+1
φ − φk ζ + (∇ · J)ζ = fζ
τ Ω Ω Ω

1 δEε k δ 2 Eε k k+1
k
J ·
k+1 ˜
J = φ + 2
φ φ − φk
∇ · J˜
Ω M (φ ) Ω δφ δφ
for equations (9). Using the first equation, we can eliminate φk+1 completely
from the second equation:

1 δ 2 Eε k
J k+1 · J˜ + φ ∇ · J k+1 ∇ · J˜
Ω τ M (φk ) Ω δφ
2

1 δEε k δ 2 Eε k
= φ + 2
φ f ∇ · J˜ . (12)
Ω τ δφ δφ
So starting with φ0 , we solve equtaion (12) to get J 1 and then use

φk+1 ζ = φk ζ + τ f ζ − (∇ · J)ζ ∀ζ ∈ L2 (Ω)
Ω Ω Ω Ω
1
to get φ .
Spatial Discretization
Inserting the energy (7) into equation (12) yields

1
J k+1 · J˜ + ε ∇∇ · J k+1 · ∇∇ · J˜
τ M (φk )
Ω Ω

−1

+ε G φk ∇ · J k+1 ∇ · J˜ = rhs . (13)
Ω
This is the equation we will now discretize using finite elements.

Given a triangulation of Ω, we denote by Eh the set of edges and by Th
the set of triangles. As finite-dimensional subspaces, we choose
• {piecewise constant functions} =: L0 (Th ) ⊂ L2 (Ω) and
• {lowest order Raviart–Thomas elements} =: RT 0 (Eh ) ⊂ H(∇·, Ω).
Raviart–Thomas Elements
The Raviart–Thomas elements are linear on each triangle – more specifi-

cally they are of the form V (x) = ax + b (note that a is scalar, so this
is not the full set of linear vector fields) – and their normal component is
continuous across the edges. This makes the space RT 0 (Eh ) a subspace of
H(∇·, Ω).
A general property of such vector fields is that their normal component is
constant along straight lines, therefore prescribing the normal component on
the edges yields a basis for RT 0 (Eh ).
The basis elements are
|Ei |
Ψi (x) := ± x − PTi , i = 1, . . . , #edges ,
2|T |
with |Ei | denoting the length of the edge, |T | denoting

the area of the triangle and PTi is the point in triangle
T which does not belong to Ei .
Weak Gradient
Since the divergence of a vector field in RT 0 is piecewise constant and the

operator ∇∇· appears in equation (13), we have to define a discrete gradient
∇h : L0 → RT 0 . It is natural to define ∇h of a function f ∈ L0 (Th ) by
duality:
∇h f · Jh = −
˜ f ∇ · J˜h ∀J˜h ∈ RT 0 (Eh ) .
Ω Ω
This leads to a splitting of equation (13) into

Kh · J˜h = ∇ · Jhk+1 ∇ · J˜h ∀J˜h ∈ RT 0 (Eh )
Ω Ω
and

1
k Jhk+1 · J˜h + ε (∇ · Kh ) · (∇ · J˜h )
Ω τ M φh Ω

+ ε−1 G φkh ∇ · Jhk+1 ∇ · J˜h = rhsh ∀J˜h ∈ RT 0 (Eh ) .
Ω
Matrix Representation
Taking the above basis for RT 0 (Eh ), we get
B0 K = A0 J k+1 (14a)
1 k k+1
B J + εA0 K + ε−1 Ak1 = r , (14b)
τ 1
where B0 is the mass matrix, Bk1 is the mass matrix weighted with the mobility
and A0 , Ak1 are the constant fourth order and non-constant second order
stiffness matrices. The underlined quantities are coefficient vectors.
One might be tempted to lump masses and insert the first equation into
the second. Unfortunately, this is not possible for Raviart–Thomas elements.
Therefore, we have to find other ways to solve the system (14). After investi-
gating in a number of possible approaches, we decided to follow Arnold and
Brezzi [1985]: the idea is to search for vector fields not in RT 0 (Eh ), but in
a bigger space “RT −1 (Th )” and enforce the solution to be in RT 0 (Eh ) via
a constraint. Then the mass matrices are block diagonal (with 3 × 3 blocks)
and can be easily inverted. On the minus side, we get a Lagrange multiplier
making the system to be solved bigger.
The resulting linear system is positive definite and is solved using the
conjugate gradient method.
3 Results
In this section we give some examples of numerical simulations using the
described numerical methods. In particular we will study step meandering in
a nonlinear regime. Here, being the computationally more efficient method, the
front tracking approach is used to explore a wide range of parameters in order
to find interesting nonlinear behavior. If topological changes are encountered,
the front-tracking simulations have to stop. Using the same parameters and
initial conditions, the diffuse-interface approximation is used to go beyond the
topological change.
3.1 Front Tracking
The numerical scheme resulting from the front tracking method as described
in Sect. 2.1 has been implemented in the FEM-Package Alberta [Schmidt and
Siebert, 2005]. It has been shown to be quite accurate and efficient when
simulating the growth of islands, see Bänsch et al. [2004]. In particular, the
influence of capillary forces (strength of the line tension) and of the presence
of edge diffusion as well the importance of anisotropy can be explored. Since
neither topological changes nor a nucleation model are included so far, simu-
lations of island dynamics are essentially restricted to the monolayer regime.
In this context, an important application is Ostwald ripening of monolayer
islands on a crystalline surface, where – as long as the coverage is not to
large – only trivial topological changes (disappearing of islands) occur. The
front tracking method has been used to simulate Ostwald ripening with a cou-
ple of hundreds of islands in Haußer and Voigt [2005].
Here we will present some results for the growth of vicinal surfaces. Using
“skew periodic” boundary conditions for index of the step height an endless
step train can be modeled. In this case, the growth of hundreds of atomic
layers can be simulated.
We start with presenting a simulation of the linear instability caused by
the ES-Effect as introduced in Sect. 1. This will allow to check the overall
accuracy of the full numerical scheme by comparing theoretically obtained
growth-rates with the numerical results. As a second example we investigate
the nonlinear regime of the meander instability. Finally we present an example
where anisotropic edge energy does lead to coarsening of the meander wave
length in the nonlinear regime.
Linear Instability
We will use the linear instability and in particular the dispersion relation to
validate the numerical scheme. To this end we consider a periodic step train
modeled as two down steps with terrace width l = 10 on a periodic domain
of size 100 × 20. Using the parameters D = 102 , c∗ = 10−3 , k− = 1, k+ = 10,
Γ = 10, Dst = 0 and F = 2 · 10−3 , the predicted most unstable wavelength
is λmax ≈ 102.7. In the numerical simulations, the randomly perturbed steps
synchronize very fast and then develop the predicted meander with a growth
rate coinciding very well with the theoretical dispersion relation, see Fig. 8.
We also note, that in all numerical tests with a larger number of equally
spaced steps, the step meander synchronized at an early stage of the evolution.
Thus it is sufficient to simulate the evolution of two steps on a periodic domain
to investigate the meandering instability in the nonlinear regime.
Nonlinear Regime
For practical purposes, the nonlinear regime of the instability is of much more
importance, because meandering patterns observed during growth show large
amplitudes. We used numerical simulations to explore the nonlinear behavior
in various parameter regimes, for a detailed discussion we refer to Haußer and
Voigt [2007].
Fig. 8. Time evolution of two equidistant, initially straight steps with small-
amplitude random perturbation on a periodic domain. (Top left) Profile of one
of the two steps at different times (given in units of ML, i.e. monolayer) shows the
emergence of a meander instability with a wavelength λm = 100a corresponding to
the most unstable wavelength λmax = 102.7 of the linear instability. (Top right)
The Fourier spectrum of the step profile clearly shows, that only the most unstable
mode survives. (Bottom left) The growth rate as given in Bales and Zangwill [1990]
is depicted for the parameters used in the simulation. As can be seen, there is only
one unstable mode in the chosen domain size of length λm = 100a. (Bottom right)
The predicted growth rate ω(λm ) = 0.0516(M L)−1 compares very well with the
numerically obtained value ω = 0.0512(M L)−1
As has been predicted by Pierre-Louis et al. [1998], Danker et al. [2003],

Pierre-Louis et al. [2005] using a local amplitude equation, we observe endless
growth of the meander amplitude in parameter regimes, where the meander
wavelengths is large compared to the terrace width. In this regime the steps
are strongly coupled. Passing to shorter meander wavelength being of similar
size as the terrace width, the step profile starts to develop overhangs, which
eventually lead to a self-crossing of the steps and thus to the formation of
a closed loop, i.e., a vacancy island – a void of the depths of one atomic
height. If the steps become even more isolated, i.e., if the meander wavelength
is considerably smaller than the terrace width, we observe stationary step
profiles with a fixed amplitude, see Fig. 9.
As the formation of the vacancy island and the subsequent evolution can
not be simulated within this numerical approach, the parameters are passed
Fig. 9a–d. Evolution of a step meander; crossover to large l/λm . For intermediate
l/λm ≈ 1 the step profile starts to develop overhangs, but we still observe endless
growth of the amplitude, see (a), (b). Further increase of l/λm leads to a pinch-off,
i.e. the formation of a vacancy island as shown in (c). For even larger l/λm the
step profile evolves to a steady state with a finite amplitude, see (d). The following
parameters have been used in the simulations: k− = 1, k+ = 100, F = 10−3 ,
D = 102 , c∗ = 10−3 , Dst = 0, l = 10. The most unstable wavelength λm (and
therefore the ratio l/λm ) is varied by changing the stiffness Γ from Γ = 1.4 to
Γ = 10−3
to the diffuse interface approximation and used there to study the pinch-off,
see Fig. 13.
Anisotropic Edge Energy
We finally give an example, where anisotropy does play role. Performing

simulation on a larger domain, it appears, that no coarsening appears for
isotropic edge energies, whereas for anisotropic edge energies interrupted
coarsening can be observed, see Fig. 10, in agreement with Danker et al.
[2003].
3.2 Diffuse-Interface-Approximation
We have developed a Finite Element software implementing the ideas of

Sect. 2.2. We use adaptivity in space and time, see Fig. 11a. Another very
Fig. 10. Meandering on a larger domain of size 10λm , λm being the most unstable
wavelength in the isotropic case. (Left) Isotropic edge energy: The wavelength is fixed
in the initial stage followed by endless growth of the amplitude. (Right) Anisotropic
edge energy: after selection of the most unstable wavelength (being smaller as in the
isotropic case, since Γ (θ = 0) < Γ0 ), one observes coarsening in the intermediate
stage. In the late stage, the coarsening stops
Fig. 11. (a) Grid belonging to Fig. 5b. Around the steps, the mesh is fine enough
to resolve the diffuse interface and gets geometrically coarser with distance from the
steps. Away from the steps, it is still fine enough to resolve the solution of the Poisson
problem. (b) A step train. The computational domain is marked with the orange box
useful feature is the ability to use “skew–periodic” boundary conditions,

i.e.
φ(Lx , y) = φ(0, y) + n, n ∈ Z, φ(x, Ly) = φ(x, 0) .
This enables us to simulate step trains, see Fig. 11b, which are a common
setup for studying the morphology of steps, e.g. the meander instability.
Coalescence
In this example, the Ehrlich–Schwoebel barrier is turned off (ζ − = 0 ⇒

M (φ) ≡ 1) and there is no deposition (f = 0), so equation (6) becomes the
Cahn–Hilliard equation, which is an approximation of the classical Mullis–
Sekerka problem.
In that case, the Γ –limit of the energy Eε for ε → 0 is the total length
of the interface. Therefore, the solutions will try to minimize the interface
length.
Fig. 12. Coalescence of two islands
We take two ellipses as initial datum. Since both ellipses change their
shape to become a circle, they will touch after some time and then coalesce,
see Fig. 12. The final shape will be a single circle.
Fig. 13. Pinch-off of a vacancy. The leftmost picture corresponds to the last po-
sition in Fig. 9c. As the simulation continues, vacancies keep pinching off and then
disappear due to deposition, see the rightmost picture
Pinch-off
In case (c) of the parameter study carried out with the front tracking method,
see Fig. 9, a pinch-off became apparent. To study the behavior after the pinch-
off, we use the same parameters. In the nondimensinoal form, they translate
into
ζ − ≈ 94.1, ˆl ≈ 9.41, and λ̂ ≈ 3.64 .
As expected, we see the same instability and can continue the simulation after
the pinch-off, see Fig. 13.
References
N. Alikakos, P. Bates, and X. Chen. Convergence of the Cahn–Hilliard equation to
the Hele–Shaw model. Arch. Ration. Mech. An., 128:165–205, 1994.
D. N. Arnold and F. Brezzi. Mixed and nonconforming finite element methods:
implementation, postprocessing and error estimates. RAIRO Modél. Math. Anal.
Numér., 19(1):7–32, 1985.
G. Bales and A. Zangwill. Morphological instability of a terrace edge during step-
flow growth. Phys. Rev. B, 41:5500, 1990.
E. Bänsch, F. Haußer, O. Lakkis, B. Li, and A. Voigt. Finite element method
for epitaxial growth with attachment-detachment kinetics. J. Comp. Phys., 194:
409–434, 2004.
E. Bänsch, F. Haußer, and A. Voigt. Finite element method for epitaxial growth
with thermodynamic boundary conditions. SIAM J. Sci. Comp., 26:2029–2046,
2005a.
E. Bänsch, P. Morin, and R. H. Nochetto. A finite element method for surface
diffusion: the parametric case. J. Comput. Phys., 203:321–343, 2005b.
M. O. Bristeau, R. Glowinski, and J. Périaux. Numerical methods for the Navier-
Stokes equations. Applications to the simulation of compressible and incompress-
ible viscous flows. In Finite elements in physics (Lausanne, 1986), pages 73–187.
North-Holand, Amsterdam, 1987.
W. K. Burton, N. Cabrera, and F. C. Frank. The growth of crystals and the equilib-
rium of their surfaces. Phil. Trans. Roy. Soc. London Ser. A, 243(866):299–358,
1951.
G. Caginalp. Stefan and Hele-Shaw type models as asymptotic limits of the phase-
field equations. Phys. Rev. A, 39(11):5887–5896, Jun 1989.
G. Caginalp and X. Chen. Convergence of the phase field model to its sharp interface
limits. Eur. J. Appl. Math., 9:417–445, Aug 1998.
G. Danker, O. Pierre-Louis, K. Kassner, and C. Misbah. Interrupted coarsening of
anisotropic step meander. Phys. Rev. E, 68:020601, 2003.
G. Ehrlich and F. G. Hudda. Atomic view of surface diffusion: tungsten on tungsten.
J. Chem. Phys., 44:1036–1099, 1966.
F. Haußer and A. Voigt. Ostwald ripening of two-dimensional homoepitaxial islands.
Phys. Rev. B, 72:035437, 2005.
F. Haußer and A. Voigt. Step meandering in epitaxial growth. J. Cryst. Growth,
303(1):80–84, 2007.
A. Karma and M. Plapp. Spiral surface growth without desorption. Phys. Rev.
Lett., 81:4444–4447, 1998.
J. Krug. Introduction to step dynamics and step instabilities. In A. Voigt, editor,
Multiscale modeling of epitaxial growth, volume 149 of ISNM. Birkhäuser, 2005.
F. Liu and H. Metiu. Stability and kinetics of step motion on crystal surfaces. Phys.
Rev. E, 49:2601–2616, 1997.
T. Maroutian, L. Douillard, and H. Ernst. Wavelength selection in unstable ho-
moepitaxial step flow growth. Phys. Rev. Lett., 83:4353, 1999.
F. Otto, P. Penzler, A. Rätz, T. Rump, and A. Voigt. A diffusive-interface aproxi-
mation for step flow in epitaxial growth. Nonlinearity, 17(2):477–491, 2004.
R. Pego. Front migration in the nonlinear Cahn–Hilliard equation. P. Roy. Soc. A,
422:261–278, Apr 1989.
O. Pierre-Louis, C. Misbah, Y. Saito, J. Krug, and P. Politi. New nonlinear evolution
equation for steps during molecular beam epitaxy. Phys. Rev. Lett., 80:4221, 1998.
O. Pierre-Louis, G. Danker, J. Chang, K. Kassner, and C. Misbah. Nonlinear dy-
namics of vicinal surfaces. J. Cryst. Growth, 275:56, 2005.
P. Politi, G. Grenet, A. Marty, A. Ponchet, and J. Villain. Instabilities in crystal
growth by atomic or molecular beams. Phys. Rep., 324:271, 2000.
C. Polop, S. Bleikamp, and T. Michely. I. Phys. Institut, RWTH Aachen.
A. Schmidt and K. Siebert. Design of adaptive finite element software: The finite
element toolbox ALBERTA. Number 42 in LNCSE. Springer, 2005.
R. L. Schwoebel and E. J. Shipsey. Step motion on crystal surfaces. J. Appl. Phys.,
37:3682–3686, 1966.
A. Voigt, editor. Multiscale modeling in epitaxial growth, volume 149 of Interna-
tional Series of Numerical Mathematics. Birkhäuser Verlag, Basel, 2005. ISBN
978-3-7643-7208-8; 3-7643-7208-7. Selected papers from the workshop held in
Oberwolfach, January 18–24, 2004.
U. Weikard. Numerische Lösungen der Cahn-Hilliard-Gleichung und der Cahn-
Larché-Gleichung. Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn,
Oct 2002.
Micro Structures in Thin Coating Layers:
Micro Structure Evolution and Macroscopic
Contact Angle
Julia Dohmen1 , Natalie Grunewald2 , Felix Otto2 , and Martin Rumpf1

1
Institut für Numerische Simulation, Universität Bonn Nussallee 15, 53115 Bonn,
Germany, {julia.dohmen,martin.rumpf}@ins.uni-bonn.de
2
Institut für Angewandte Mathematik, Universität Bonn, Wegelerstr. 10,
53115 Bonn, Germany, {grunewald,otto}@iam.uni-bonn.de
Summary. Micro structures of coating surfaces lead to new industrial applica-

tions. They allow to steer the wetting and dewetting behaviour of surfaces and
in particular to enhance hydrophobicity. Here, we discuss the formation of mi-
cro structures in the drying process of a coating. Furthermore, for a given mi-
cro structured surface we show how to predict the effective contact angle of
drops on the surface. At first, we derive a new approach for the simulation of
micro structure evolution based on a gradient flow perspective for thin liquid
films. This formulation includes a solvent dependent surface tension, viscosity and
evaporation rate. In each time step of the resulting algorithm a semi implicit
Rayleigh functional is minimized. The functional itself depends on the solution
of a transport problem. We apply a finite difference discretization both for the
functional and the transport process. As in PDE optimization a duality argu-
ment allows the efficient computation of descent directions. Next, given a certain
micro structured coating we mathematically describe effective contact angles in
different configurations and their impact on the macroscopic hydrophilic or hy-
drophobic surface properties. On periodic surfaces we aim at the computation of
effective contact angles. This involves a geometric free boundary problem on the
fundamental cell. Its solution describes vapor inclusions on the wetted surface.
The free boundary problem is solved by a suitable composite finite element ap-
proach. Furthermore, we introduce a new model for the influence of micro struc-
tures on contact angle hysteresis. This model is adapted from elasto–plasticity
and dry friction. It identifies stable contact angles not only as global or local en-
ergy minimizers but as configurations at which the energy landscape is not too
steep.
1 Introduction
Micro structures in coatings are of great industrial relevance. They can be

desirable and undesirable. On the one hand they might lead to rupture of
76 J. Dohmen et al.
a paint. On the other hand they can enhance hydrophobicity of the surface.
Here we discuss two different aspects of these phenomena.
In Sect. 2 we consider a model for the formation of micro structures in
a drying coating. These strucutures can for instance evolve from a non ho-
mogeneous solvent distribution in an originally flat coating. We model the
coating by an adapted thin film model. It is based on a gradient flow model
with solvent dependent viscosity, surface tension and evaporation rate, see
Sect. 2.1. This introduces Marangoni effects to the film which can lead to
a structured film height but also counteract rupture. It also takes into ac-
count the solvent evaporation in a coating, which is fast at low film heights,
due to a faster heating up. A third effect considered is the hardening, i.e.
the temporal change of the viscosity of the coating. In Sect. 2.2 and 2.3 we
introduce a numerical algorithm based on a semi implicit time discretization,
which takes advantage of the gradient flow structure. In each time step a cor-
responding Rayleigh functional is minimized in Sect. 2.5 we show numerical
results.
In the second part in Sect. 3 we discuss the implications of a structured
surface to contact angles of macroscopic drops sitting on the surface. The
micro structures highly influence the contact angle and thereby the sticking
of the drop to the surface. One governing effect is the formation of vapor
inclusions on the surface at a micro scale. This reduces the contact of the
drop to the surface – hence, it rolls off easily. We introduce an algorithm
in Sect. 3.1, which simulates the vapor inclusions in a periodic setup. The
corresponding liquid vapor interface is a minimal surface with prescribed
microscopic contact angle of the triple contact line. In the limit of small
scale periodicity of the surface this enables the calculation of effective contact
angles.
Finally, in Sect. 3.2 we consider the stability of drop configurations on
the micro structured surface. A new model is introduced which determines
the stability of effective contact angles. Their stability depends on the micro
configuration of the drop, i.e. on the possible vapor inclusions. The model
allows for intervals of stable contact angles (contact angle hysteresis). It is
adapted from elasto–plasticity and dry friction, and assumes a configuration
not only to be stable if it minimizes (locally) the relevant surface energy
but also if the energy landscape at this configuration is not too steep. This
leads to different hysteresis intervals for configurations with and without va-
por inclusions. A change in the vapor configuration at the surface can ex-
plain the highly non monotone dependence of the hysteresis on the surface
roughness, known since the sixties, [JD64], as well as more recent experi-
ments.
Micro Structures in Thin Coatings 77
2 Modeling and Simulation of the Micro Structure

Formation in Thin Coatings
2.1 Modeling Thin Coatings as a Gradient Flow
We propose a simple model for coatings similar to the one considered in

[HMO97], which in spite of its simplicity reproduces many of the interest-
ing features known for a drying paint. We assume the paint to consist of two
components, the non–volatile resin and the volatile solvent, whose concentra-
tion is given by s. Together they form a well-mixed fluid with height h. In the
simulations we plot both the height (on the left) and the solvent concentra-
tion (on the right), see Fig. 1. These are the two parameters describing the
physical properties of the fluid:
Fig. 1. A time evolution (back to front) of a coating is described by its height (on
the left) and solvent concentration (on the right). Here the trivial case with constant
Solvent Concentration is depicted
The solvent concentration influences the viscosity μ (the drying coating

becomes harder with descreasing solvent concentration) as well as the sur-
face tension σ (the surface tension increases with decreasing solvent concen-
tration) and the evaporation rate e. The evaporation rate also depends on
solvent concentration and on the height of the film, as a thin film dries fast
due to its closeness to the warm substrate. We assume a well–mixed coating,
where both components are transported by the same horizontal fluid veloc-
ity u.
This model can introduce micro structures even on an initially flat coating.
Indeed, they may be originated in a inhomogeneous distribution of solvent.
Local areas on the coating where the solvent concentration is high have less
surface tension. This induces a Marangoni flow in the direction from high
to low solvent concentration. This flow reduces the surface energy as the
interface with less surface tension is strechted in comparison to the interface
with high surface tension, which is condensed. Hence, fluctuation in the solvent
concentration lead to a structured film height. On the other hand, surface
tension primarily induces a flow which reduces the area of the interface. It
therefore drives the fluid to a flat film. These two forces can in the absence of
evaporation compensate each other leading to an inhomogeneous structured
78 J. Dohmen et al.
but stable film, c.f.[W93]. Figure 2 shows a Marangoni induced stable micro
structure.
Furthermore, the combination of a height dependent evaporation rate e
of the solvent and of Marangoni effects (i.e. the solvent dependent surface
tension) counteracts film rupture at points, where the height of the film tends
to zero. In fact, due to their closeness to the warm surface the film dries
quickly at low film heights. This reduces the solvent concentration at these
points, which again induces a Marangoni flow to the valleys on the film surface
due to a higher surface tension in case of a low solvent concentration. This
flow counteracts rupture. Indeed our simulations (Figs. 5 and 4) do not show
a critical deepening of the film leading to rupture.
Gradient Flow Structure. For our model we firstly assume a balance of vis-
cous and capillary forces but neglect the momentum of the fluid. We assume
an over-damped limit in which the quasi stationary Stokes equations for an
incompressible fluid are appropriate. By the well known lubrication approx-
imation [BDO97] they can be reduced to the thin film equations, which are
of gradient flow structure (cf. [GO03]). The height of the film h performs
a steepest descent of an energy functional E:

ḣ = −gradE h . (1)
To make sense of the gradient of the energy one has to identify the metric
structure of the manifold M on which the gradient flow takes place. In this
case, this is the manifold of all heights of the film with prescribed volume.
The metric is described by its metric tensor gh (δh, δh) on the tangent spaces,
which consist of the infinitesimal height variations δh. Denoting diffE h .δh =
limε→0 1ε (E(h + εδh) − E(h)) turns (1) into

gh (ḣ, δh) = − diff E h .δh ∀ δh ∈ Th M. (2)
Equation (2) can be seen as the Euler–Lagrange equation of

1
F (δh) = gh (δh, δh) + diff E h .δh (3)
2
with respect to δh. Indeed, the actual rate of change ḣ minimizes F under all
possible infinitesimal variations δh. We will use such a gradient flow structure
to model thin coatings, inspired by the gradient flow model for thin films,
which we will explained first.
Thin Films as a Gradient Flow. Thin fluid films are described by the well
known thin film equation
σ
ḣ = − div(h3 ∇Δh), (4)
3μ
for the height of the film [BDO97]. Here, we might impose either periodic or
natural boundary conditions. This evolution is a gradient flow, as introduced
in [O01]. The relevant energy is the linearized surface energy:

1 2
E(h) := σ 1 + |∇h| dx.
Ω 2
The metric tensor is given by the minimal energy dissipated by viscous friction,
i. e.
3μ 2
gh (δh, δh) = inf u dx ,
u Ω h
where Ω is the underlying domain. Note that the metric tensor is base point
dependent. The infimum is taken over any velocity profile u that realizes the
given change in film height δh described by the transport equation
δh + div (h u) = 0. (5)
On the first sight the metric tensor seems to be a complicated object, as it in-
volves the minimization of the viscous friction. Therefore finding the minimizer
of the functional F in (3) requires to solve a nested minimization problem.
This can be avoided, if one describes the tangent space, i.e. all infinitesimal
changes in film height h, directly by an admissible velocity fields u via (5) (of
course the same δh may be described by many u’s). In this sense the metric
tensor can be lifted onto the space of admissible velocities u:

3μ 2
gh (u, u) = u dx. (6)
Ω h
Rewriting (3) leads to a formulation of the gradient flow as the evolution
ḣ + div (h u∗ ) = 0, (7)
∗
where u minimizes the Rayleigh functional
1
F (u) = gh (u, u) + diff E h .u (8)
2

over all fluid velocities u. Here diff E h .u is defined as diff E h .δh with δh
satisfying (5). It is now easy to see that the gradient flow given by (6)–(8)
coincides with the evolution of the thin film equation (4). Indeed, we observe
that u∗ solves the Euler–Lagrange equation corresponding to the Rayleigh
functional (8):

3μ ∗
0 = gh (u∗ , u) + diff E h .u = u · u dx − σ∇h∇div(h u) dx
Ω h Ω
for all test velocities u. For periodic or natural boundary conditions this im-
mediately implies
σh2
u∗ = ∇Δh.
3μ
Finally, plugging u∗ into (7) yields the thin film equation (4). The thin film
is a special case of a thin coating, i.e. the one with constant solvent con-
centration. Numerical results for the spreading of a thin film are shown in
Fig. 1.
80 J. Dohmen et al.
Thin Coatings as a Gradient Flow. The model for thin coatings is more dif-
ficult, as the state of the paint is not only described by its film height h but
also by the solvent concentration s in the film. We assume a thin film model,
which is inspired by the gradient flow described above. Here, we adopt a point
of view developed in [GP05]: The gradient flow evolves on the manifold of all
possible film heights. The solvent will be transported along with the fluid and
is taken into account as a vector bundle on the manifold. At any given film
height, there is a vector space of possible solvent concentrations, the fiber.
They are not part of the manifold. The tangent spaces therefore consist only
of the infinitesimal changes in film height δh. These are induced by a velocity
u (as explained above):
δh + div (h u) = 0 (9)
The solvent concentration is transported by parallel transport. That is, we
assume a mixed fluid, where the solvent is transported by the same velocity.
As s is the concentration of solvent, the actual amount of solvent is given by
h s. Therefore
δ(hs) + div (hs u) = 0. (10)
This vector bundle construction to model an extra component slaved to the
transport of the fluid was introduced in [GP05] for a thin film with surfactant.
The gradient flow is now given by the reduced energy and the metric on
the manifold. As in the thin film case, the relevant energy is the linearized
surface energy:
1 2
E(h, s) := σ(s) 1 + |∇h| dx. (11)
Ω 2
The surface tension σ depends on the solvent concentration s. This introduces
Marangoni effects to the model, which we see in a drying coating. The metric is
given by the minimal energy dissipated by viscous friction, where the viscosity
μ depends on the solvent concentration. The drying coating becomes hard.
One has the metric tensor

3μ(s) 2
gh,s (u, u) = u dx. (12)
Ω h
The gradient flow is (9) and (10) with the velocity field u = u∗ , where u∗
minimizes the Rayleigh functional
1
F (u) = gh,s (u, u) + diff E h,s .u (13)
2
over all velocities u. This model is similar to the thin film model, but has
included the solvent features of a thin coating. On the one hand it tries to
minimize the (linearized) surface energy (11) by mean surface tension and
Marangoni flows. They reduce the energy by elongating the surface with low
surface tension. One the other hand the flow is hindered by viscous friction
(12). The viscous friction increases as the evaporation continues (as μ(s) is an
increasing function). The only effect not yet modeled is the evaporation. On
a continuous level this would include the modeling of the full vapor phase. On
the discrete level the evaporation is included as a second step in an operator
splitting method, see below.
2.2 Natural Time Discretization
Any gradient flow has a natural time discretization. It involves the natural
distance function dist on the manifold M defined via
1 ! 2 "
dist 2 (h0 , h1 ) := inf gγ(t) (γ̇, γ̇) dt ,
γ 0
with γ any smooth curve with γ(0) = h0 and γ(1) = h1 . If M is actually

Euclidean instead of genuinely Riemannian as in our case
dist 2 (h0 , h1 ) = |h0 − h1 |2 . (14)
If τ denotes the time step size, the solution hk+1 at step k + 1 can be inferred
from the state hk at step k via the variational problem:

k+1 1 2
k

h = argminh dist h, h + E(h) . (15)
2τ
As a motivation consider the Euclidean case (14). Here the Euler–Lagrange
equation for (15) turns into the implicit Euler scheme
1 k+1
h − hk = −∇E hk+1 .
τ
We want to use (15) as a starting point to construct a natural and stable
discretization. The drawback of (15) is, it is fully nonlinear and it involves
two nested minimizations.
One natural idea to overcome this drawback, which is also used for epitax-
ial growth, see the corresponding chapter in this book, is the following: We
approximate the functional by its quadratic at hk and then lift the variational
problem on the level of possible velocities u in the spirit of (7) and (8). We
first turn to the quadratic approximation: Writing h = hk + τ δh, we have
1
dist 2 h, hk + E(h) ≈
2τ
(16)
τ τ2
k
g k (δh, δh) + E(h ) + τ diff E hk .δh +
g k δh, Hess E hk δh ,
2 h 2 h

where Hess E hk denotes the Hessian of E in hk . Hence we can solve

1 τ #
∗
δh = argminδh ghk (δh, δh) + diff E hk .δh + ghk δh, Hess E hk δh
2 2
(17)
82 J. Dohmen et al.
and then set hk+1 = hk + τ δh∗ , cf. (3). However, as in (3), (17) still involves
two nested minimizations. Therefore, using (5) we may lift (17) on the level
of possible velocities u as before. This yields

1 τ
uk+1 = argminu ghk (u, u) + diff E hk .u + ghk u, Hess E hk u (18)
2 2

and then set hk+1 = hk + τ div hk uk+1 . Compare (18) to (7) and (8). This
is the basis for the gradient flow algorithm used for epitaxial growth.
For our algorithm we use an alternative approach. We consider a semi
implicit time discretization. For this we only approximate the squared distance
dist 2 in (15) by its metric based approximation and keep E fully nonlinear.
We use the following notation: For given velocity field u varying in space and
fixed in time define the transport operator h(·, ·), which maps a height field hk
at time tk onto a height field h(hk , u) = h(tk+1 ), where h solves the transport
equation ∂t h + div(h u) = 0 with initial data h(tk ) = hk . Given this operator,
we again apply a linearization of the distance map dist in (15) and evaluate
the energy on h[hk , u]. This energy is again implicitly defined via the velocity
field u, which minimizes a corresponding functional. Thus, we define

k+1 τ k
u = argminu g k (u, u) + E h h , u , (19)
2 h
which can be considered as a semi-implicit alternative to the time discretiza-
tion in (18). The new height field is then given by hk+1 = h(hk , uk+1 ). Here,
we still use the metric for the linearization of the distance map and evaluate
this at the height field hk at the old time tk .
This gradient flow model for the thin film equation can easily be general-
ized for the thin coating model. To simplify the presentation let us introduce
the vector q = (h, hs) consisting of the two conservative quantities film height
h and amount of solvent hs. Furthermore, we again define a transport oper-
ator q(·, ·), which maps q k = (hk , hk sk ) at time tk onto q(q k , u) = q(tk+1 ),
where q is a the solution of the system of transport equations
∂t h + div(h u) = 0 (20)
∂t (hs) + div(hs u) = 0 (21)
with initial data q(tk ) = q k = (hk , hk sk ). In analogy to (19), we consider an
implicit variational definition of the motion field

k+1 τ k
u = argminu g k (u, u) + E q h , u , (22)
2 q
where E[q] is given by (11). Hence, in every time step we ask for the minimizer
of a functional whose integrand depends on the solution of a hyperbolic initial
value problem. Indeed this is a PDE constrained optimization problem. In the
next section we will solve this problem numerically based on a suitable space
discretization and duality techniques.
2.3 Space Discretization for the Gradient Flow
Let us consider a discretization of (22) in one and two space dimensions and
for simplicity restrict to a domain Ω = [0, 1]d , where d ∈ {1, 2}, and impose
periodic boundary conditions. We suppose Ω to be regularly subdivided into
N interval of width Δ := N1 (d = 1) or squares of edge length Δ (d = 2).
By Q = (Qi )i∈I = (Hi , Hi Si )i∈I and U = (Ui )i∈I we denote nodal vectors of
discrete q and u quantities, respectively, where the ith component corresponds
to a grid nodes xi . Here I is supposed to be the lexicographically ordered index
set of nodes (for d = 2 these indices are 2-valued, i. e. i = (i1 , i2 ), where the
two components indicate the integer coordinates on the grid lattice). Spatial
periodicity can be expressed by the notational assumption Qi = Qi+N e and
Vi = Vi+N e , where e = 1 for d = 1 and e = (1, 0) or (0, 1) for d = 2. Now,
we define in a straightforward way a discrete energy value E[Q] on R2I and
a discrete metric GQ [U, U ] on RdI × RdI :

1 2
E[Q] = Δd σ(S̃i ) 1 + (∇i H) , (23)
2
i∈I
3μ(Si )
GQ (U, U ) = Δd |Ui |2 , (24)
Hi
i∈I
where S̃ = 12 (Si + Si+1 ) (d = 1) or S̃ = 14 (Si + Si+(0,1) + Si+(1,0) + Si+(1,1) )

(d = 2) are interpolated values for the solvent concentration at cell centers,
and ∇i H = Δ 1
(Hi+1 − Hi ) (d = 1) or ∇i H = 2Δ 1
(Hi+(1,0) + Hi+(1,1) − Hi −
Hi+(0,1) , Hi+(0,1) + Hi+(1,1) − Hi − Hi+(1,0) ) (d = 2) is the difference quotient
approximation of the gradient of the height field. Next, we define an operator
Q, which computes Q(Qk , U ) = Qk+1 = (Hik , Hik Sik )i∈I as the solution of an
implicit Lax–Friedrich scheme for the associated transport problem for given
data Qk at time tk and a discrete velocity vector U . Let us detail this here in
the one dimensional case, where we obtain the following system of equations
i+1 − Ui−1 Qi−1

Ui+1 Qk+1 Qk+1 − 2Qk+1
k+1
Qk+1 − Qki + Qk+1
i
= + i+1 i i−1
τ 2Δ Δ2
for all i ∈ I and a small positive constant . The two dimensional case is
completely analogous. This scheme can be rewritten in matrix vector notation
Qk = A(U )Q(Qk , U ) (25)
where A(U ) ∈ R 2I×2I
is a matrix depending on the discrete vector field U ,
which can easily be extracted from the Lax-Friedrich scheme. For > 0 this
matrix is invertible. Thus, we obtain the explicit representation Q(Qk , U ) =
A(U )−1 Qk for the discrete transport operator. With these ingredients at hand,
one obtains a discrete counterpart of the variational problem (22)

k+1 τ k
U = argminU∈RdI G k (U, U ) + E Q Q , U . (26)
2 Q
84 J. Dohmen et al.
Finally, we define Qk+1 = Q(Qk , U k+1 ). In each time step we aim at comput-
ing the discrete minimizer U k+1 via a gradient descent scheme on RdI . Hence,
besides the energy on the right hand side of (26) we have to compute the gradi-
ent vector on RdI . For the variation of the energy E(Q(Qk , U )) in a direction
W ∈ RdI we get ∂U E(Q(Qk , U ))(W ) = ∂Q E(Q(Qk , U ))(∂U Q(Qk , U )(W )).
A direct application of this formula for the evaluation of the gradient of the
energy E would require the computation of
∂U Q(Qk , U )(W ) = −A−1 (U )(∂U A(U )(W ))A−1 (U )Qk
for every nodal vector W in RdI . To avoid this, let us introduce the dual
solution P = P (Qk , U ) ∈ R2I which solves
A(U )T P = − ∂Q E(Q(Qk , U )).
Computing the variation of the linear system (25) with respect to U we achieve

0 = (∂U A(U )(W ))Q Qk , U + A(U ) ∂U Q Qk , U (W ) ,
from which we then derive

∂U E Q Qk , U (W ) = ∂Q E Q Qk , U ∂U Q Qk , U (W )

= −A(U )T P (Qk , U ) · ∂U Q Qk , U (W )

= −P Qk , U · A(U ) ∂U Q Qk , U (W )

= P Qk , U · (∂U A(U )(W ))Q Qk , U .
This representation of the variation of the energy can be evaluated without
solving dI linear systems of equations. In our implementation we consider
the Armijo rule as a step size control in the descent algorithm on RdI .
2.4 Evolution of Thin Coatings with Solvent Evaporation
So far the model for the evolution of a thin film consisting of resin and solvent
is considered as a closed system and formulated as a gradient flow. Evaporation
of the solvent from the liquid into the gas phase – the major effect in the drying
of the coating – still has to be taken into account. As already mentioned,
incorporating this in a gradient flow formulation would require to model the
gas phase as well. To avoid this we use an operator splitting approach and
consider the evaporation separately as a right hand side in the transport
equations. Thus, we consider the modified transport equations
∂t h + div(h u) = e(h, s) ,
∂t (hs) + div(hs u) = e(h, s) ,
where e(h, s) = − c+hs
C
is the usual model for the evaporation [BDO97], where
C, c > 0 are evaporation parameters. In the time discretization we now al-
ternate the descent step of the gradient flow and an explicit time integration
of the evaporation. In the first step, the velocity uk+1 is computed based on
(22) . Solving the corresponding transport equations (20) and (21) we obtain
updated solutions for the height and the solvent concentration at time tk+1 ,
which we denote by h̃k+1 and s̃k+1 , respectively. In the second step, applying
an explicit integration scheme for the evaporation we finally compute

hk+1 = h̃k+1 + τ e h̃k+1 , s̃k+1 ,
−1 k+1 k+1
sk+1 = hk+1 h̃ s̃ + τ e h̃k+1 , s̃k+1 .
For the fully discrete scheme, we proceed analogously and update the nodal
values Qk+1 in each time step. In fact, given U k+1 as the minimizer of (26) we
compute Q̃k+1 = (H̃ k+1 , S̃ k+1 ) = A(U k+1 )−1 Qk and then update pointwise
Qk+1
i = Q̃k+1
i + τ e(H̃ik+1 , S̃ik+1 ).
2.5 Numerical Results
The numerical results show the features of thin coatings introduced by

Marangoni and surface tension effects combined with evaporation and hard-
ening. We will discuss them separately. A first test of our algorithm was to
run it with constant solvent concentration, which turns the model for thin
coatings into the simpler thin film model described above. Numerical results
are already shown in Fig. 1. They are numerically consistent with results ob-
tained by a finite volume scheme for the thin film equation [GLR02], where
thin films with (and without) surfactant are simulated. Figure 2 shows the
effects introduced by Marangoni forces. In particular an inhomogeneous sol-
vent concentration can lead to a structure formation in the film height. In the
absence of evaporation this structure becomes stable as the Marangoni forces
are opposed by mean surface tension forces, which want to reduce the length
of the film surface.
An inhomogeneous solvent concentration also introduces a structured film
height via evaporation, Fig. 3. This leads – as only solvent evaporates – to
valleys in the film located at positions with a high amount of solvent. Still
the coating is by no means close to rupture, as this is opposed by Marangoni
forces. Figure 5 shows that the combination of these effects leads to a micro
Fig. 2. Evolution of a coating with a marangoni flow introduced by an inhomoge-

neous solvent concentration
86 J. Dohmen et al.
Fig. 3. Evolution of a coating with evaporating solvent
Fig. 4. The drying of a coating with (artifically) constant viscosity with a vanishing
of micro structures
Fig. 5. The evolution of a coating with hardening, where micro structures persist
structure. This micro structures turns into a stable pattern of the dry coating.
This is due to a solvent dependent viscosity, which leads to hardening during
the drying process. Figure 4 shows that in a coating with constant viscosity
the mean surface tension forces dominate the evolution at later times. This
finally leads to a flat coating similar to the thin film case. Micro structures
occur only at intermediate times.
3 Micro Structured Coatings

and Effective Macroscopic Contact Angle
Micro structures in thin coatings are not only an unwanted feature, like the
rupture of a coating. They also can be desirable, as micro structures enhance
water repellent properties of a surface. This feature is known as the lotus
effect. Among other plants, the lotus plant makes use of this [BN97], to let
water roll off their leaves. One can also spot it at the back of a duck. The duck
will stay dry while the water rolls off in pearls, as the feathers have a micro
structure whose cavities are not filled with the water. To analyze this effect
one has to understand how the form of the drops especially the contact angles
are determined by the surface energy, which is the relevant energy in the quasi
static case we are considering here.
The surface energy E is the sum of the energies of the three different inter-
faces in our problem. That is, the liquid/vapor interface ΣLV , the solid/liquid
interface ΣSL and the solid/vapor interface ΣSV . Each of these interfaces is
weigthened with its surface tension:
E = |ΣSL | · σsl + |ΣLV | · σlv + |ΣSV | · σsv .
The shape of the drop is the one with the least energy given the volume
of the drop. This also determines the contact angle, which is important to
understand the lotus effect. Drops with large contact angles take a nearly
pearl like form and roll of easily. Drops with small contact angles are flatter
and stick more to the surface.
For a flat surface the contact angle θY can be calculated using Young’s law,
which can be derived from minimizing property with respect to the surface
energy (see below):
σsv − σsl
cos θY = . (27)
σlv
Drops on surfaces with micro structures are more complicated. They can either
fill the micro structure with water, a situation described by Wenzel in [W36]
(Fig. 6), or they can sit on air bubbles situated in the surface cavities, as
considered by Cassie and Baxter in [CB44], see Fig. 7. For a nice review on
this effect see either [Q02] or the book [GBQ04].
On a periodic surface it is possible to calculate effective contact angles.
These are contact angles that would be attained in the limit of small scale
periodicity. These contact angles determine the shape of the drop, see Figs. 6
and 7. The micro structure is much smaller than the size of the drop. It
therefore makes sense to think of an effective surface tension of the micro
Fig. 6. A Wenzel type drop
Fig. 7. A Cassie–Baxter type drop

88 J. Dohmen et al.
structured surface. The justification for this is given in [AS05], where it is

shown that the energy minimizing drops behave in the limit of small surface
periodicity like the drops with the corresponding effective surface tensions.
This is a mathematically rigorous argument using the Γ -convergence of the
energies.
The effective surface tensions are the ones assigned to a macroscopically
flat surface with a small scale micro structure. In the Wenzel situation the
solid surface and thereby the solid/liquid interface as well as the solid/vapor
interface are enlarged by the roughness r. (r equals the area of the surface on
∗ ∗
the unit square.) The effective surface tensions σsl and σsv are:
∗ ∗
σsl = r · σsl and σsv = r · σsv ,
The effective contact angle θW is then determined by an adapted Young’s law,

cf. (27):
∗ ∗
σsv − σsl σsv − σsl
cos θW = = r· .
σlv σlv
Therefore a Wenzel type situation enlarges large contact angles and shrinks
small ones in comparison to the flat surface case. Thus it enhances water re-
pellent properties of a surface (with pearl like drops and large contact angles),
as well as hydrophilic properties (with flat drops and low contact angles).
In the Cassie–Baxter situation the calculation of the effective surface ten-
sion is more difficult as it involves a determination of the size of the vapor
bubbles at the micro scale, see Fig. 7. In a periodic set up this leads to a free
boundary problem to be solved on the periodicity cell. The solution may be
a configuration with or without vapor inclusions. At the triple line the con-
tact angle for a flat surface θY is attained. Below, we developed an algorithm
which solves the free boundary problem and thereby determines the shape of
the vapor inclusions.
The solution of the cell problem provides the area α of the liquid/vapor
interface in one periodicity cell, the area β of the solid/liquid interface and
the area of the solid/vapor interface, which is r − β, see Fig. 8. The effective
Fig. 8. The Configuration of a cell problem in the Cassie–Baxter regime


surface tension σsl is the sum of the surface tensions of the interfaces:

σsl = α · σlv + β · σsl + (r − β) · σsv .
We obtain a modified Young’s law (cf. (27)) for the effective solid/vapor sur-

face tension σsv = r · σsv and thereby determine the effective Cassie–Baxter
contact angle:
σ − σsl

cos θCB = sv = −α + β · cos θY .
σlv
For α → 1 and β → 0 the Cassie–Baxter contact angle tends to 180◦. This
is the situation when the drop hardly touches the surface but rests mostly
on the air pockets. The drop takes a nearly spherical shape and rolls off
easily.
The effective contact angles calculated above are derived under the as-
sumption of periodicity of the surface. An assumption typically not satisfied
by natural surfaces. Theses surfaces show a highly inhomogeneous structure
with both sizes and shape of the micro structure varying over several orders
of magnitude, see Fig. 9.
A future perspective is to derive a mathematical model which captures
these inhomogeneities. It should be based on a stochastical model where one
asks for the expectation of the effective contact angle.
There is a second drawback of Young’s law which describes the the absolut
minimizer of the energy. In fact, drops on surface can have many different
stable contact angles. Rain drops on a window sheet demonstrate this in our
daily life. They stick to the window and do not roll off, in spite of the window
being inclined. These drops are not spherical caps but take an non symmetric
shape, see Fig. 10.
Fig. 9. Natural surfaces with micro structure (copyright: Bayer Material Science)
Fig. 10. A drop sticking to a tilted plane

90 J. Dohmen et al.
The contact angles at the upward point part of the contact line are much
smaller than those at the downward pointing part. Nevertheless all contact
angles are stable, as the drop does not move. We developed a new model
to understand which drops are stable, [SGO07], see Sect. 3.2. This model
is adapted from models used in dry friction and elasto–plasticity. It mainly
states that a drop should by stable, if the energy landscape is not to steep at
its configuration.
3.1 Computing the Effective Contact Angle
In this section we will discuss how to compute the effective contact angle on
a rough coating surface in the regime of the Cassie–Baxter model. Thus, we
consider a periodic surface micro structure described by a graph on a rectangu-
lar fundamental cell Ω (cf. Fig. 11). The surface itself is supposed to be given
as a graph f : Ω → R, whereas the graph of a second function u : Ω → R rep-
resents the gas/liquid interface between a vapor inclusion on the surface and
the covering liquid. In fact, we suppose {(x, y) ∈ Ω × R | f (x) < y < u(x)} to
be the enclosed gas volume. Following [SGO07] we take into acount the total
(linearized) surface energy on the cell Ω given by

E(u, f ) = σsv 1 + |∇f | + σlv 1 + |∇u| dx +
2 2 σsl 1 + |∇f |2 dx
[u>f ] [u<f ]

= (σsv − σsl ) 1 + |∇f |2 + σlv 1 + |∇u|2 dx
[u>f ]

+ σsl 1 + |∇f |2 dx
Ω
Here, [u > f ] = {x ∈ Ω | f (x) < u(x)} represents the non wetted domain
of the vapor inclusion, also denoted by Ωsv , and [u < f ] = {x ∈ Ω | f (x) >
u(x)} the wetted domain, respectively (cf. Figs. 7, 11). Let us emphasize
that for fixed f the energy effectively depends only on u|[u>f ] . In the energy
minimization we have to compensate for this by a suitable extension of u
outside [u > f ]. The variation of the energy E with respect to u in a direction
w is given by

∂u E(u, f )(w) = (v · ν) (σsv − σsl ) 1 + |∇f |2 + σlv 1 + |∇u|2 dH1
∂[u>f ]

∇u · ∇w
+ σlv dx ,
1 + |∇u|2
[u>f ]
where ν denotes the outer normal at the triple line ∂[u > f ] and v is the normal
velocity field of this interface induced by the variation w of the height function
u. The relation between v·ν and w is given by (v·ν)(∇f ·ν−∇u·ν) = w. A min-

imizer u of E(·, f ) describing the local vapor inclusion attached to the surface
is described by the necessary condition ∂u E(u, f )(w) = 0 for all smooth varia-
tions w. Applying integration by parts we deduce the minimal surface equation
−div √ ∇u 2 = 0 for u on [u > f ] and the boundary condition
1+|∇u|

(σsv − σsl ) 1 + |∇f |2 + σlv 1 + |∇u|2 σlv ∇u · ν
0= +
∇f · ν − ∇u · ν 1 + |∇u|2
on ∂[u > f ]. The energy is invariant under rigid body motions. Hence, for
∇u(x)
a point x on ∂[u > f ] we may assume ∇f (x) = 0. In this case ν(x) = − |∇u(x)|
2
and thus σlsσ−σ sv
= 1 + |∇u(x)|2 − √ |∇u(x)| 2 = √ 1 2
= cos(θ),
lv 1+|∇u(x)| 1+|∇u(x)|
where θ is the contact angle between the solid–liquid and the liquid vapor
interface. Hence, we have recovered Young’s law on the micro scale of the cell
problem.
Finally we end up with the following free boundary problem to be solved:
Find a domain Ωsv and a function u, such that the graph of u on Ωsv is
a minimal surface with Dirichlet boundary condition u = f and prescribed
Fig. 11. The effective contact angle on a rough surface is calculated based on the
numerical solution of a free boundary problem on a fundamental cell. The liquid
vapor interface of the vapor inclusion on the surface forms a minimal surface with
a contact angle on the surface of the solid determined by Young’s law
Fig. 12. Each row shows on the periodic cell a family of coating surfaces together
with the liquid vapor interfaces of the corresponding vapor inclusions in the wetting
regime of the Cassie–Baxter model. In the first row the transition in the surface
configuration from a wavelike pattern in one axial direction to more spike type
structures is depicted from left to right, whereas in the second row the transition
from the same wave pattern to elongated troughs is shown
92 J. Dohmen et al.
contact angle θ on ∂Ωsv , and this graph should be periodically extendable as

a continuous graph on R2 (cf. Fig. 11 and Fig. 12).
The numerical solution of this free boundary problem is based on a time
discrete gradient descent approach for a suitable spatially discrete version of
the above variational problem. Let us denote by Vh the space of piecewise
affine, continuous functions (with a continuous periodic extension on R2 on
some underlying simplicial mesh of grid size h covering the rectangular fun-
damental cell Ω. For a discrete graph F ∈ Vh of the coating surface we start
from some initial guess U 0 ∈ Vh for the (extended) discrete graph of the liq-
uid vapor interface on top of the vapor inclusions and successively compute
a family (U k )k≥0 with decreasing Energy E(·, F ). For given U k we first solve
k
the discrete Dirichlet problem for a minimal surface on Ωsv := [U k > F ] in
a composite finite element space Vh [HS97, HS98] and based on that compute
k
the next iterate U k+1 . In fact, following [HS97a] we define Vhk as a suitable
subspace of functions W ∈ Vh with W = 0 on ∂Ωsv k
. Thereby, the degrees of
k
freedom are nodal values on the original grid contained in Ωsv whose distance
k
from ∂Ωsv is larger than some = (h) > 0. Then, a constructive extension
k
operation defines nodal values on all grid nodes of cells intersec ted by Ωsv
k+1
(for details we refer to [HS97a]). Hence, we compute a solution Ũ with
Ũ k+1 − F ∈ Vhk , such that

∇Ũ k+1 · ∇Φ
0= dx
k
Ωsv 1 + |∇U k |2
for all test functions Φ ∈ Vhk . Next, based on Ũ k+1 data on ∂Ωsv
k
we compute
a discrete descent direction V ∈ Vh as the solution of
k

Gk V k+1 , Φ = −∂u E Ũ k , F (Φ)
for all Φ ∈ Vh . Here, with the intention of a proper preconditioning

of the gra-
dient descent, we take into account the metric Gk (Ψ, Φ) = σlv Ω k √ ∇Ψ ·∇Φk 2 .
sv 1+|∇U |
Given V k+1 we finally determine the actual descent step applying Amijo? step
size control rule and compute U k+1 = Ũ k+1 +τ k+1 V k+1 for a suitable timestep
τ k+1 . Here, we implicitly assume that the built–in extension of Ũ k+1 on whole
Ω is sufficiently smooth.
3.2 A New Model for Contact Angle Hysteresis
We consider a drop on a micro structured plane. Experiments show that there

is an hysteresis interval [θr , θa ] of stable contact angles. It is bounded by the
receding contact angle θr and the advancing contact angle θa . The dependence
of this interval on the surface roughness is badly understood. We introduced
a new model for contact angle hysteresis [SGO07] to understand the experi-
mental evidence of a complicated dependence of the hysteresis interval on the
roughness:
Well known experiments from the sixties [JD64] show that the width as
well as the position of the hysteresis interval depend in a nonlinear way on
the surface roughness, see Fig. 13.
Especially the receding contact angle shows a jump like behavior at a cer-
tain surface roughness.
Furthermore, recent experiments [QL03] show that the receding contact
angle not only depends on the surface roughness, but also on the way the
drop is put on the surface. In Fig. 14 we show how the receding contact
angles depends on a pressure applied to press the drop into surface cavities.
Fig. 13. Experimental Dependence of Advancing and Receding Contact Angles on

the Surface Roughness. Reprinted with Permission from [JD64]. Copyright (1964)
American Chemical Society
Fig. 14. Experimental Dependence of Receding Contact Angles on the Pressure

Pushing the Drop onto the Surface. Reprinted from [QL03] with Permission
94 J. Dohmen et al.
The pressure is then released and the contact angle is measured. Figure 14
again shows a jump like behavior of the receding contact angle.
We introduce a new model to capture these phenomena. It is similar to
models used in dry friction [MT04] and elasto–plasticity [MSM06]. The main
idea of our model is that stability of drops is primarily not related to global or
local minimality of its interfacial energy, but rather to the fact that the local
energy-landscape seen by the drop should not be too steep such that dissipa-
tion energy pays off the modify the configuration. To be be more precise, if
the energy that would be gained moving the drop (i.e. controlled up to first
order by the slope of the energy landscape) is smaller than the energy that
would be dissipated while moving, then the drop will not move. In order to
implement these concept, we use the derivative-free framework proposed in
[MM05] (see also the review [M05]).
That is, we assume a drop L0 (with its contact angle) to be stable if
E(L0 ) − E(L̃) ≤ dist(L0 , L̃)
for all L̃ with the same volume. Here we have modeled the distance of two
drops to be the area of the coating surface wetted by only one of them. This
seems reasonable, as we know that the most energy is dissipated around the
moving triple line. Therefore a drop which has significantly changed its bot-
tom interface on the coating surface is far apart from its initial configura-
tion.
Our new model implies two different diagrams of stable contact angles,
depending on the type of drop (Wenzel or Cassie–Baxter type). These are
shown in Figs. 15 resp. 16 in the case of a surface with flat plateau and
vallees, separated by steep edges. The roughness of this type of surface can be
increased by deepening the asperities without changing the size of the wetted
surface plateau.
The hysteresis interval for Cassie–Baxter drops is much narrower than
the one for Wenzel drops. This can explain qualitatively both the downward
Fig. 15. Stable contact angles for Wenzel type drops

Fig. 16. Stable contact angles for Cassie–Baxter type drops
jump at large pressures of the receding contact angles in Fig. 14, and the jump
behavior in Fig. 13.
The latter can be understood as a superposition of the two stability dia-
grams. The jump in the width of the hysteresis interval results from a tran-
sition from Wenzel type drops to Cassie–Baxter type drops. At low surface
roughnesses Wenzel type drops are stable. They exhibit a wide hysteresis in-
terval. At higher roughness, the stable configurations in the experiment are
instead Cassie–Baxter. They display a much narrower hysteresis interval. The
stable contact angles resulting from the transition from Wenzel to Cassie–
Baxter drops are shown schematically in Fig. 17, where they are superposed on
the experimental results of Johnson and Dettre. The comparison is only quali-
tative, because experimentally roughness is measured only indirectly, through
the number of heat treatments undergone by the solid surface in the sample
Fig. 17. A schematic sketch of the stable contact angles is given according to our
model. The shaded regions represents the set of stable angles for varying surface
roughness, superposed on experimental data from Fig. 13
96 J. Dohmen et al.
preparation. The figure shows a transition from a regime in which the differ-
ence between advancing and receding contact angles increases monotonically
with roughness, to one in which such a difference is smaller, and nonsensitive
to roughness.
Figure 14 reflects the fact that the stability interval depends on the type
of drop. Assuming that the corresponding surface has sufficiently large rough-
ness, we see from Figs. 15 and 16 that forcing a transition from a Cassie–
Baxter to a Wenzel type drop (by applying a large enough pressure) may
reduce the lower end of the stability interval (i.e., the receding contact angle)
from well above to well below 90◦ .
References
[AS05] Alberti, G., DeSimone, A.: Wetting of rough surfaces: a homogenization
approach. Proc. R. Soc. London Ser. A Math. Phys. Eng. Sci., 461,
79–97, (2005)
[BN97] Barthlott, W., Neinhuis, C.: Characterization and Distribution of
Water-repellent, Self-cleaning Plant Surfaces. Annals of Botany, 79,
667–677, (1997)
[BGLR02] Becker, J., Grün, G., Lenz, M., Rumpf, M.: Numerical Methods for
Fourth Order Nonlinear Degenerate Diffusion Problems. Applications
of Mathematics, 47, 517–543, (2002)
[CB44] Cassie, A.B.D., Baxter, S.: Wettability of Porous Surfaces. Trans. Fara-
day Soc., 40 , 546–551, (1944)
[MSM06] Dal Maso, G., DeSimone, A., Mora, M.G.: Quasistatic evolution prob-
lems for linearly elastic-perfectly plastic materials. Arch. Rat. Mech.
Anal., 180, 237–291, (2006)
[GBQ04] de Gennes, P.G., Brochard-Wyart, F., Quéré, D.: Capillarity and Wet-
ting Phenomena. Springer (2004)
[SGO07] DeSimone, A., Grunewald, N., Otto, F.: A new model for contact angle
hysteresis. Networks and Heterogeneous Media, 2, 2, 211–225, (2007)
[JD64] Dettre, R.H., Johnson, R.E.: Contact Angle Hysteresis II. Contact An-
gle Measurements on Rough Surfaces. Contact Angle, Wettability and
Adhesion, Advances in Chemistry Series, 43, 136–144, (1964)
[GO03] Giacomelli, L., Otto, F.: Rigorous lubrication approximation. Interfaces
and Free boundaries, 5, No. 4, 481–529, (2003)
[GP05] Günther, M., Prokert, G.: On a Hele–Shaw–type Domain Evolution with
convected Surface Energy Density. SIAM J. Math. Anal., 37, No. 2, 372–
410, (2005)
[GLR02] Grün, G.,Lenz, M., Rumpf: A Finite Volume Scheme for Surfactant
Driven Thin Film Flow. in: Finite volumes for complex applications III,
M. Herbin, R., Kröner, D. (ed.), Hermes Penton Sciences, 2002, 567–574
[HS97] Hackbusch, W. and Sauter, S. A.: Composite Finite Elements for Prob-
lems Containing Small Geometric Details. Part II: Implementation and
Numerical Results. Comput. Visual. Sci., 1 15–25, (1997)
[HS97a] Hackbusch, W. and Sauter, S. A.: Composite Finite Elements for prob-
lems with complicated boundary. Part III: Essential boundary condi-
tions. technical report, Universität Kiel, (1997)
[HS98] Hackbusch, W. and Sauter, S. A.: A New Finite Element Approach

for Problems Containing Small Geometric Details. Archivum Mathe-
maticum, 34 105-117, (1998)
[HMO97] Howison, S.D., Moriarty, J.A., Ockendon, J.R., Terril, E.L., Wilson,
S.K.: A mathematical model for drying paint layers. J. of Eng. Math.,
32, 377–394, (1997)
[QL03] Lafuma, A., Quéré, D.: Superhydrophobic States. Nature Materials, 2,
457–460, (2003)
[MM05] Mainik, A., Mielke, M.: Existence results for energetic models for rate-
independent systems. Calc.Var., 22, 73–99, (2005)
[M05] Mielke, M.: Evolution of rate-independent systems. Handbook of Differ-
ential Equations. Evolutionary Equations, 2, 461–559, (2005)
[MT04] Mielke, M., Theil, F.: On rate independent hysteresis models. NoDEA
11, 151–189, (2004)
[BDO97] Bankoff, S.G, Davis, S.H., Oron, A.: Long-scale evolution of thin liquid
films. Rev. of modern Physics, 69, No. 3, 931–980, (1997)
[O01] Otto, F.: The geometry of dissipative evolution equations: the porous
medium equation. Comm. Partial Differential Equations, 26, No.1–2,
101–174, (2001)
[Q02] Quéré, D.: Rough ideas on wetting. Physica A, 313, 32–46, (2002)
[W36] Wenzel, R.N.: Resistance of Solid Surfaces to Wetting by Water. Ind.
Eng. Chem., 28, 988–994, (1936)
[W93] Wilson, S.K.: The levelling of paint films. Journal of Applied Mathe-
matics, 50, 149–166, (1993)
Part III
Biochemical Reactions and Transport

Modeling and Simulation
of Hairy Root Growth
Peter Bastian1 , Jenny Bauer4 , Andrés Chavarrı́a-Krauser2,

Christian Engwer1 , Willi Jäger2, Sven Marnach3 , Mariya Ptashnyk2 ,
and Bernhard Wetterauer4
1
IWR, Universität Heidelberg, INF 368, 69120 Heidelberg, Germany (now at3 )
{peter.bastian,christian.engwer}@ipvs.uni-stuttgart.de
2
IAM, Universität Heidelberg, INF 294, 69120 Heidelberg, Germany
{andres.chavarria.krauser,jaeger,
mariya.ptashnyk}@iwr.uni-heidelberg.de
3
IPVS, Universität Stuttgart, Universitätsstraße 38, 70569 Stuttgart, Germany
sven.marnach@ipvs.uni-stuttgart.de
4
IPMB, Universität Heidelberg, INF 364, 69120 Heidelberg, Germany
jenny bauerbiol@yahoo.de, bernhard.wetterauer@urz.uni-heidelberg.de
Summary. A multiscale approach is presented to model growth of hairy roots.

On the macroscopic scale, a continuous model is derived, which includes growth
and nutrient transport. Water transport is considered on the microscopic scale.
A Discontinuous Galerkin scheme for complex geometries is used to compute the
permeability of root bulks. This permeability constitutes the linkage between micro-
and macroscopic scale. The models are applied then to describe shaker cultures of
hairy roots and simulations are compared to measurements.
1 Introduction
Plants remain a major source of pharmaceuticals and biochemicals. Many of
these commercially valuable phytochemicals are secondary metabolites that
are not essential to plant growth. Hairy roots, obtained from plants through
transformation by Agrobacterium rhizogenes, produce many of the same im-
portant secondary metabolites and can be grown in relatively cheap hormone-
free medium. Thus they may provide an alternative to agricultural processes
to produce phytochemicals on a large scale [9, 11]. Hairy roots can be culti-
vated under sterile conditions either in a bioreactor or in shake flasks. The fast
growing hairy roots are unique in their genetic and biosynthetic stability and
are able to regenerate whole viable plants for further subculturing [6]. The
yield of secondary metabolites is determined by biomass accumulation and
by the level of secondary metabolite produced per unit biomass. Therefore
102 P. Bastian et al.
Fig. 1. Ophiorrhiza mungos and hairy root of O. mungos
a number of biological studies have focused on the growth process, growth

dynamic, and production of secondary metabolites in bioreactors of different
design [17, 10, 9].
Hairy roots of Ophiorrhiza mungos Linn., the Chinese camptotheca tree,
are currently gaining the interest of pharmacologists, since a secondary
metabolite, camptothecin, can be used to treat cancer diseases [26]. Camptho-
thecin is a modified monoterpene indole alcaloid produced by Camptotheca
acuminata, Nothapodytes foetida, some species of the genus Ophiorrhiza, Er-
vatamia heyneana, and Merrilliodendron megacarpum [24, 27]. In order to
produce camptothecin efficiently, it is necessary to optimize the biological
processes behind its biosynthesis (either in bioreactors or shaker cultures).
However, to achieve this, it is essential to understand metabolism, growth
and transport processes of and in root networks.
The aim of the project was to derive a mathematical model which de-
scribes growth of root networks and nutrient transport through root tissues.
To describe the biological system a multiscale approach was used. The proc-
esses on macroscopic and microscopic scale are linked. Numerical solutions
were compared to experimental data obtained from O. mungos hairy roots
grown as shaker cultures. The model and numerical algorithms are general
enough to describe growth and transport processes in bioreactors.
2 Biological Processes
The processes observed in a bioreactor are water transport, diffusion and
transport of nutrients in medium and roots, and growth of roots. These proc-
esses are taking place on different scales, each of which contributes to the
global system.
On the macroscopic level roots form a dense bulk which resembles a porous
medium. This allows to use well known modeling approaches to describe
porous media. The root bulk is hence treated as a continuous medium of
Modeling and Simulation of Hairy Root Growth 103
varying porosity, and all processes are defined on this continuum. Growth
and nutrient transport are observed on the macroscopic scale and described
through distributions. Growth is assumed to depend on nutrient concentra-
tion in the medium and inside the roots [20, 10, 21]. Three processes are
responsible for changes in the mediums nutrient concentration: uptake on
the root surface, convection due to pressure gradients and diffusion arising
from concentration gradients. The macroscopic diffusion coefficient and the
uptake kinetics depend on the density of the root network and are defined
phenomenologically.
On the microscopic scale the root structure influences flow and trans-
port processes around the root network, which has a complex highly ramified
structure. The surface of a single root is covered with fine hairs, reducing
conductivity [16]. Here it becomes clear that the microscopic structure de-
termines substantially the macroscopic properties, in particular porosity and
permeability.
Nutrient transport inside the roots is also a microscopic process. Since
transport inside the root network is substantially faster in comparison to
growth and branching, it is legitimate to consider only the average internal
nutrient concentration and use a macroscopic internal nutrient concentra-
tion.
3 Macroscopic Model
Two densities are used to describe growth of hairy root networks: the root
volume per unit volume ρ (0 ≤ ρ(x, t) ≤ 1) and the cross section area of
tips per unit volume n (n(x, t) ≥ 0). Growth can then be assumed to occur
due to tip movement (elongation), tip formation (branching), and secondary
thickening. Thus the change of density n is defined by a transport equation
with growth velocity v and a branching term f . A similar approach has been
used to model growth of fungi mycelia [7, 4]. The change of root density ρ
is determined by the root volume produced due to tip movement. Secondary
thickening is defined phenomenologically as a production term in the equation
for ρ. Growth velocity and branching kinetics depend on the concentration of
nutrients in the medium (denoted by c(x, t)) and within the roots (denoted
by s(x, t)).
The transport of nutrients in the medium is defined by a convection-
diffusion equation with a reaction term describing the active and passive
nutrient uptake on the roots surface. Active uptake is assumed to be unidi-
rectional (into the root network) and dependent only on the local medium
nutrient concentration c. Passive uptake depends on the nutrient gradi-
ent between medium and roots, given by the difference c − s. Four pro-
cesses which change the total internal nutrients S = s Vr , where Vr (t)
is the root volume, are considered here: uptake, growth, ramification and
metabolism.
The precise formulation of the macroscopic model of hairy root growth

reads
∂t n + ∇ · (n v) = f in (0, T ) × Ω,
∂t ρ = n v + q in (0, T ) × Ω,
(1)
∂t ((1 − ρ)c) + ∇ · (u c) − ∇ · (Dc (1 − ρ)∇c) = −g in (0, T ) × Ω,
R R R
d
dt
S = g dx − γg (n v + q) dx − γr f dx − γm S in (0, T ),
Ω Ω Ω
with
v = R s (ρmax − ρ) (∇μ + ατ τ ) ,
∇μ = αc ∇c − αρ ∇ρ − αn ∇n ,
q = χ s ρ (ρmax − ρ) ,
f = β c s ρ (ρmax − ρ) ,
2λ n
g= ρ ( Km c + P (c − s)) ,
r
where R is a growth rate, ρmax is the maximal root density, χ is a secondary
thickening rate, β is a branching rate, λ is the characteristic length of the
uptake-active tissue around a tip, Km is a constant describing active uptake
rate, P is a permeability characterizing passive uptake, u is the flow velocity
of the medium, Dc is a diffusion constant, and γg , γr and γm are constants
describing the proportion of metabolites used for growth, ramification and
metabolism, respectively. Since hairy roots are agravitropic [14] growth ve-
locity can be assumed to be independent of gravity. Growth can then be
presumed to occur along nutrient gradients and away from dense tissue. Pure
densification of the root system is modeled by the local rotation τ , which is
a unit vector orthogonal to ∇μ and ∇n. It does not affect the density distri-
bution of tips, although mass is still produced and ρ changes. Here αc , αρ ,
αn , and ατ are phenomenological constants, which relate the growth driving
gradients to the resulting growth velocity.
Initial density distributions and nutrient concentrations are prescribed.
For both the bioreactor and the shake flasks the side walls of the reactor
vessel (Γsw ) are impermeable to the medium. In the bioreactor we have in-
flow (Γin ) and outflow (Γout ) boundaries. On Γin the nutrient concentration is
given and Dirichlet boundary condition can be posed. On Γout we have out-
flow boundary condition. In the case of the shaker cultures no-flux boundary
condition can be posed (i.e. ∂Ω = Γsw and Γin = Γout = ∅). Since roots can-
not extend beyond the vessel, the tip density fulfills also the no-flux boundary
conditions.
n(0, x) = n0 (x), ρ(0, x) = ρ0 (x) in Ω,

c(0, x) = c0 , S(0) = S0 , in Ω,
nv ·ν = 0 on (0, T ) × ∂Ω,
c = cD on (0, T ) × Γin ,
Dc (1 − ρ) ∇c − u c) · ν = 0 on (0, T ) × Γout ,
∇c · ν = 0 on Γsw .
The water velocity u is defined by Darcy’s law. We distinguish between Dirich-

let condition given by pressure values on the boundary ΓD and Neuman con-
ditions defined by flux through the boundary ΓN . Depending on the experi-
mental setup, either Neumann or Dirichlet conditions are posed on Γin and
Γout . For a shake flask and on the sidewalls (Γsw ) of a bioreactor, no-flux (i.e.
homogeneous Neumann) conditions need to be posed.
∇·u= 0 in Ω,
u = −K∇p in Ω,
(2)
p = p0 on ΓD ⊂ ∂Ω,
u·ν = j on ΓN = ∂Ω \ ΓD .
Here K is the permeability function of the root network. On the macroscopic

scale K changes with density, but derivation of the relation is cumbersome if
not impossible for general geometries. It must be computed hence for a certain
structure on the microscopic scale.
The effective permeability K is assumed to be of the form
K = K0 · Krel , (3)
where K0 is the average coefficient relating the flow velocity u to the pressure
gradient ∇p in an empty reactor (ρ = 0). Krel (ρ) a dimensionless relative
permeability which reflects the local root structure. K0 can be obtained by
determining the Hagen–Poiseuille flow in the reactor, while Krel is computed
using simulations of the microscopic model.
4 Microscopic Model
On the microscopic scale we consider water flow between single root branches.
To simplify the problem, we assume an incompressible potential flow:
∇·u= 0 in Ω,
u = −∇p in Ω,
(4)
p = p0 on ΓD ⊂ ∂Ω,
∇p · ν = j on ΓN = ∂Ω \ ΓD .
The domain Ω has a complex geometry, given by the root structure. Water
uptake by the roots (growth) is small compared to the water flow, therefore
Neumann boundary conditions with j = 0 can be assumed on the root sur-
faces.
4.1 Numerical Methods for Microscale Simulations
The root structure on the microscopic scale exhibits a complex shape. Classical
numerical methods require a grid resolving such complex geometries. Creat-
ing these grids is a very sophisticated process and generates a high number
of unknowns. We developed a discretization scheme for complex geometries,
based on a Discontinuous Galerkin (DG) discretization on a structured grid
and a structured grid for the construction of trial and test functions [8]. This
method offers a discretization where the number of unknowns is not directly
determined by the possibly very complicated geometrical shapes, but still al-
lows the provision of fine structures, even if their size is significantly smaller
than the grid cell size.
Let Ω ⊆ Rd be a domain. On a sub-domain Ω ∗ ⊆ Ω we want to
solve Eqn. (4). The shape of Ω ∗ is usually based on geometrical prop-
erties retrieved from experiments, like micro-CT images, or from compu-
tations. T (Ω) = {E0 , . . . , EM−1 } is a partitioning, where the mesh size
h = min { diam(Ei ) | Ei ∈ T } is not directly determined by the geometri-
cal properties. Nevertheless error control on solution of the partial differential
equation might require a smaller h due to the shape of ∂Ω. For Ω ∗ a triangu-
lation based on T (Ω) is defined T (Ω ∗ ) = {En∗ | En∗ = Ω ∗ ∩ En ∧ En∗ = ∅}, see
Fig. 2. As En∗ is always a subset of En we will call En fundamental element of
En∗ . The internal skeleton Γint and external skeleton Γext of the partitioning
are denoted by
$ #

Γint = γe,f = ∂Ee∗ ∩ ∂Ef∗ Ee∗ , Ef∗ ⊂ Ω ∗ and Ee∗ = Ef∗ and |γe,f | > 0 ,
Γext = {γe = ∂Ee∗ ∩ ∂Ω ∗ | Ee∗ ⊂ Ω ∗ and |γe,f | > 0 } .
In the finite element mesh T (Ω ∗ ) each element En∗ can be shaped arbitrarily.
Using DG, unlike conforming methods, the shape functions can be chosen in-
Fig. 2. Construction of the partitions T (Ω ∗ ) from the partitions G and T of the

domain Ω and of E ∗ from its fundamental element and Ω ∗ . The local triangulation
of Ei∗ and ∂Ei∗
dependently from the shape of the element. Note that certain DG formulations
are especially attractive, because they are element wise mass conservative and
therefore able to accurately describe fluxes over element boundaries. We use
a DG formulation with local base functions ϕ∗n,j being polynomial functions
ϕn,j ∈ Pk defined on the fundamental element En , with their support re-
stricted to En∗ :
ϕn,j inside of En∗
ϕ∗n,j = , (5)
0 outside of En∗
%
Pk = {ϕ : Rd → R | ϕ(x) = |α|≤k cα xα } is the space of polynomial functions
of degree k and α a multi–index. The resulting finite element space is defined
by
Vk∗ = v ∈ L2 (Ω ∗ ) v|En∗ ∈ Pk (6)
and is discontinuous on the internal skeleton Γint . With each γe,f ∈ Γint we
associate a unit normal n. The orientation can be chosen arbitrarily, in this
implementation we have chosen n oriented outwards Ee∗ for e > f and inwards
otherwise. With every γe ∈ Γext we associate n oriented outwards Ω ∗ . The
jump [ . ] and the average . of a function v ∈ Vk∗ at x ∈ γ ∈ Γint are defined
as
1
[ v ] = v|Ee∗ − v|Ef∗ and v = v|Ee∗ + v|Ef∗ .
2
Assembling the local stiffness matrix in DG requires integration over the vol-
ume of each element En∗ and its surface ∂En∗ . Integration is done using a local
triangulation of En∗ . En∗ is subdivided into a disjoint set {En,k
∗
} of simple ge-
∗
& ∗
ometric objects, i.e. simplices and hypercubes, with Ēn = Ēn,k , see Fig. 2.
k
The integral over a function f on En∗ can then be evaluated as

f (x) dx = f (x) dx,
∗
En k E∗
n,k

where ∗
En,k f dx is evaluated using standard quadrature rules.
4.2 Numerical Estimation of Macroscopic Parameters

from Microscopic Simulations
Following the approach described in the previous subsection and applying

this method to the microscopic problem in (4), the relative permeability, as
introduced in (3), can be computed from direct simulation of flow through
a root bulk: ⎛ ⎞ ⎛ ⎞−1

p − p
Krel = ⎝ udx⎠ · ⎝ dx⎠ ,
2 1
(7)
h
Γin Γin
Fig. 3. The Marching Cube algorithm in R2 distinguishes six basic cases, depending
on the value of a function Φ in the corners. The pictures show these six different
cases, together with their key in the look-up table
Fig. 4. (a) shows the scalar function describing the geometry, the marching simplex
algorithm yields the geometry visible in (b). Pressure and velocity are computed on
the given domain, using direct simulations with a Discontinuous Galerkin scheme.
The resulting velocity can be seen in (b). (c) shows a closeup
where Γin ⊂ ΓD describes the inflow boundary. Dirichlet boundary conditions

are posed both on the inflow and the outflow boundary.
The domain Ω ∗ is implicitly given by a scalar function Φ. This scalar
function will usually be obtained through post processing of image data, i.e.
from CT images. In these calculations we use artificially generated structures
(Fig. 4.2a), based on structural parameters (using the PlantVR software [5]).
The sub-domain boundary ∂Ω ∗ is given as an iso-surface Φ = 0.
The local triangulation is based on the Marching Cube/Simplex Algo-
rithm [13]. These algorithms give a surface reconstruction for an iso-surface.
Each vertex of an element can be below or above the value of the iso-
surface, read inside or outside the sub-domain. For a cube element in R2
this gives 16 different cases. Each of these cases corresponds to one of six
basic cases and can be transformed using simple geometric operations (see
Fig. 3). A look-up table maps each case to the appropriate surface recon-
struction. The key for the look-up table is given by assigning the state
of each corner (inside → 1, outside → 0) to one bit of an integer. The
look-up table was extended to provide a surface and a volume reconstruc-
tion.
Using the formulation described in [18, 19], (4) reads: Find p ∈ Vk∗ such
that
a (p, v) + Jσβ (p, v) = lσβ (v) ∀v ∈ Vk∗ .
Fig. 5. Estimation of macroscale parameters from microscale simulations. Direct

simulation of (4) for different root bulks with different densities yields correlation
of Krel and ρ. Model assumed: Krel = 1 − ρb , where b = 0.82 ± 0.02
The bilinear form

a (p, v) = (K∇p) · ∇v dV
Ee∗ ∈T ∗ E ∗

e

+ (K∇v) · n [ p ] − (K∇p) · n [ v ] ds
γef ∈Γint γ

ef

+ (K∇v) · n p − (K∇p) · n v ds
γe ∈ΓD γ
e
is parametrized by = ±1. Choosing = 1 we get a non–symmetric scheme

introduced by Oden, Babušky and Baumann in [15]. For = −1 we obtain the
Symmetric Interior Penalty method which needs an additional stabilization
term added to the bilinear form:
σ
σ
Jσβ (p, v) = [ p ][ v ]ds + pv ds
|γef |β |γe |β
γef ∈Γint γef γe ∈ΓD γe
with σ > 0 and β, where β depends on the dimension of Ω (β = 1 when

dim = 2). Choosing = 1 and σ > 0 results in the Non–Symmetric Interior
Penalty method.
The right hand side is a linear form

lσβ (v) = f v dV + J v ds
Ee∗ ∈T ∗ E ∗ γe ∈ΓN γ
e

e

σ
+ (K∇v) · n g ds + v g ds.
|γef |β
γe ∈ΓD γ γe ∈ΓD γe
e
Direct simulation of Eq. (4) for different realizations of root bulks, yields
Krel and ρ. Krel is assumed to be of the form Krel (ρ) = 1 − ρb and b is
determined using a mean square fit.
Fig. 5 shows the dependence of Krel on ρ and the fitted function Krel (ρ),
where b = 0.82 ± 0.02.
5 Application of the Macroscopic Model
There are two common ways of cultivating hairy roots, either in shake flasks
or in bioreactors. Shaker cultures are used more often because of their
simple assembly and usage in experiments and biological research. How-
ever, experiments in shaker cultures do not provide information about the
spatial structure and distribution of roots. Bioreactors have rather indus-
trial applications and are more complex to operate and to use as experi-
mental set-ups. In the work presented here both cultivation methods were
considered. In fact, equations (1) are able to describe both situations, as
these differ only slightly in the method used to guarantee nutrient supply.
While the principle of shake flask is based on permanent shaking of medium
and culture, bioreactors use medium fluxes to ensure nutrient and oxygen
supply.
5.1 Simulation of Shaker Cultures
For numerical simulation Eqs. (1) can be simplified to reduce the amount of
free parameters. Uptake of nutrients can be considered to be purely of active
nature, neglecting the passive transport (P = 0). Moreover, the energy cost for
branching of new tips can be neglected (γr = 0). Since the root branches are
very thin and variation in radius is small, root thickening can be neglected
as well (χ = 0). The main purpose of shaking is to supply oxygen and to
ensure a homogeneous distribution of nutrients. This means that transport
in the medium is non-limiting to uptake and growth. In the simulation this
homogeneous distribution can be achieved via large diffusion. Active water
transport is neglected.
A personal computer was used to simulate the macroscopic model (1),
using a implementation of the numerical schemes based on the DUNE frame-
work [3, 1]. For spatial discretization of the first and third equation in (1) a cell
centered finite volume scheme on a structured grid was used [12]. The diffusive
and convective/reactive part of the third equation in (1) were decoupled for
discretization in time (second order operator splitting [23]). To prevent both
instabilities in the transport term and effects from strong numerical diffusion,
the convection equation was solved using an explicit second order Godunov
upwind scheme with a minmod slope limiter [25, 12]. The diffusive part of
the equation was solved implicitly. The ordinary differential equations for the
Fig. 6a,b. Comparison of simulation and experimental data from hairy roots grown
as shaker cultures. The evolution in time of root mass (a) and concentration of
sucrose in the medium (b) are compared to measurements. At 336 hours cultures
were transferred into fresh medium
root density and the inner nutrient concentration S [second and fourth Eq.
in (1)] were solved with Euler’s method [22].
The parameters in the model are chosen such that the numerical results fit
experimental data obtained from O. mungos hairy roots (Fig. 6). Gradients
of nutrients and tip density, with moderate tissue compaction (local rota-
tion), were chosen here as the driving force of growth. Very good agreement is
found between measurements and simulation (mass increase: R2 = 0.85, nu-
trient uptake: R2 = 0.93). The model delivers spatial information on growth
patterns as well. A simulation of a two dimensional flask is found in Fig. 7.
The distributions are assumed to be constant in one of the the three dimen-
sions. Simulation in three dimensions could however be easily implemented.
Measurements deliver at the moment only data describing the kinetics of mass
increase and nutrient uptake (compare Fig. 6), which do not include quanti-
tative information on spatial patterns. Therefore, verification of the patterns
in Fig. 7 is at the moment not possible. It is hence not clear which growth
process dominates in the growth force ∇μ + ατ τ . Do hairy roots follow rather
nutrient gradients than space gradients, or is diffusion of root tips more im-
portant? Or is mass increase a consequence of tissue compaction (local rota-
tion)? It will probably be a mixture of all and other processes not accounted
for. A detailed discussion regarding this issue can be found in Bastian et al.
(2007b).
5.2 Hairy Roots Bioreactor

Root growth in bioreactors can be described by the full macroscopic model
(1). Active water transport has to be considered. The flow velocity u can be
calculated using Darcy’s-Law (2), which requires the permeability K(ρ). The
relation between K and the root volume density ρ is determined by the mi-
croscopic model (compare Fig. 5). The model and numerical algorithms are
Fig. 7a–d. Simulated two dimensional spatial growth patterns of hairy roots grown
as shaker cultures. (a) root volume density, (b) root tip density, (c) nutrient con-
centration in medium, and (d) local mass increase (all after 380 h of growth)
elaborated sufficiently to calculate and optimize the transport and growth

processes in bioreactors. For meaningful simulation, parameters related to
the three dimensional structure are required for (1). However these are not
available from shaker culture experiments.
In order to study three dimensional growth patterns, a small experimen-
tal bioreactor (WEWA I ) for root cultures was constructed in the group of
Prof. Wink (IPMB, Universität Heidelberg) to deliver the information needed.
“WEWA I”s construction is based on the Low Cost Mist Bioreactor (LCMB)
of the company ROOTec GmbH (Heidelberg, Germany; Patentnumber: US
2003/0129743A1 / EP 1268743B1). The bioreactor system consists of a 3 l
reactor reservoir (Fig. 8), a medium reservoir, a gas reservoir, and a mist
chamber for the distribution of medium over the fixed root bed. Every 15
min the root inoculum is sprayed with 20 ml medium. Due to spraying,
a pressure difference of 0.3 bar arises, which is used to return the surplus
medium into the medium reservoir. The pressure in the reactor is controlled
by a pneumatic relief valve (0.5 bar opening pressure). Instead of the B5
medium (1% sucrose; ROOTec GmbH, Heidelberg, Germany) used in shaker
cultures, a modified version with 0.5% sucrose is used to obtain an optimal
medium osmolarity. In order to prevent contaminations in the reactors a bac-
teriostatic (0.5 g/l Claforan) and a fungicide (40 mg/l Nystain) were added
to the medium.
Fig. 8a–c. Scheme of the WEWA I mist chamber: injection valve of medium (a),
mesh on which cultures are fixed (b), drain for surplus medium (c)
“WEWA I” is being put into operation at the moment. Several test runs
have been done so far. However, these test measurements have not supplied
reliable enough experimental data to identify the model parameters. Therefore
meaningful simulations of the processes in bioreactors were not possible until
now.
6 Conclusion
The work presented here has a general interest for modeling and simulation
of complex growing networks like root systems and growth processes in biore-
actors. A macroscopic model describing water and nutrient transport through
a growing root network was derived. The growth of roots in a water solu-
tion and the dense growth habit of hairy roots give the possibility to define
growth as a change of tissue volume density. This allows further to expand
a one dimensional growth model, namely pure elongation of single roots, to
a continuous three dimensional model, which delivers information on spatial
growth patterns. Linkage of the microscopic and the macroscopic scale is ac-
complished by the elaborate and novel numerical algorithms developed for
solution of elliptic equations on complex shaped domains. Comparison of nu-
merical solutions and measurements of hairy roots shaker cultures, showed
that the model is able to describe very well the kinetics of growth and nutri-
ent uptake. Moreover, model and numerical algorithms are general enough to
describe growth in bioreactors. Optimization of camptothecin production is
still an open task due to the lack of experimental data. The model presented
here is a good step forward towards achieving this and represents a sound base
for further research. The numerical methods developed for microscale simula-
tions are of interest for a wide range of applications, ranging from pore scale
problems for water resources to cell biology. In future work these methods will
be extended to time-dependent problems.
References
[1] P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, R. Kornhuber,
M. Ohlberger, and O. Sander. A generic grid interface for parallel and adaptive
scientific computing. part ii: Implementation and tests in dune. Preprint, 2007.
Submitted to Computing.
[2] P. Bastian, A. Chavarrı́a-Krauser, C. Engwer, W. Jäger, S. Marnach, and
M. Ptashnyk. Modelling in vitro growth of dense root networks. In prepa-
ration, 2007.
[3] P. Bastian, M. Droske, C. Engwer, R. Klöfkorn, T. Neubauer, M. Ohlberger,
and M. Rumpf. Towards a unified framework for scientific computing. In
R. Kornhuber, R.H.W. Hoppe, D.E. Keyes, J. Périaux, O. Pironneau, and
J. Xu, editors, Domain Decomposition Methods in Science and Engineering,
number 40 in Lecture Notes in Computational Science and Engineering, pages
167–174. Springer-Verlag, 2004.
[4] G. P. Boswell, H. Jacobs, F. A. Davidson, G. M. Gadd, and K. Ritz. Growth
and function of fungal mycelia in heterogeneous environments. Bulletin of
Mathematical Biology, 65:447–477, 2003.
[5] S. Chuai-Aree, S. Siripant, W. Jäger, and H.G. Bock. PlantVR: Software
for Simulation and Visualization of Plant Growth Model. In Proceedings of
PMA06: The Second International Symposium on Plant growth Modeling, Sim-
ulation, Visualization and Applications. Beijing, 2006.
[6] P. M. Doran. Hairy Roots: Culture and Applications. Harwood Academic
Publishers, Amsterdam, 1997.
[7] L. Edelstein-Keshet. Mathematical models in biology. The Random House/
Birkhäuser Mathematics Series, New York, 1988.
[8] C. Engwer and P. Bastian. A Discontinuous Galerkin Method for Simulations
in Complex Domains. Technical Report 5707, IWR , Universität Heidelberg,
http://www.ub.uni-heidelberg.de/archiv/5707/, 2005.
[9] Y. J. Kim, P. J. Weathers, and B. E. Wyslouzil. Growth of Artemisia annua
hairy roots in liquid and gas-phase reactors. Biotechnology and Bioengineering,
80:454–464, 2002.
[10] Y. J. Kim, P. J. Weathers, and B. E. Wyslouzil. Growth dynamics of Artemisia
annua hairy roots in three culture systems. Journal of Theoretical Biology,
83:428–443, 2003.
[11] Y. J. Kim, B. E. Wyslouzil, and P. J. Weathers. Invited review: Secondary
metabolism of hairy root cultures in bioreactors. In Vitro Cell. Dev. Biol.
Plant, 38:1–10, 2002.
[12] Randall J. LeVeque. Finite Volume Methods for Hyperbolic Problems. Cam-
bridge University Press, 2002.
[13] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution
3d surface construction algorithm. In SIGGRAPH ’87: Proceedings of the 14th
annual conference on Computer graphics and interactive techniques, pages 163–
169, New York, NY, USA, 1987. ACM Press.
[14] E. Odegaard, K. M. Nielsen, T. Beisvag, K. Evjen, A. Johnsson, O. Rasmussen,
and T. H. Iversen. Agravitropic behaviour of roots of rapeseed (Brassica napus
L.) transformed by Agrobacterium rhizogenes. J Gravit Physiol., 4(3):5–14,
1997.
[15] J. T. Oden, I. Babuška, and C. E. Baumann. A discontinuous hp-finite element

method for diffusion problems. Journal of Computational Physics, 146:491–519,
1998.
[16] M. Ptashnyk. Derivation of a macroscopic model for nutrient uptake by a single
branch of hairy-roots. University Heidelberg, Preprint, 2007.
[17] D. Ramakrishnan and R. W. Curtis. Trickle-bed root culture bioreactor design
and scale-up: Growth, fluid-dynamics, and oxygen mass transfer. Biotechnology
and Bioengineering, 88(2):248–260, 2004.
[18] B. Rivière and M.F. Wheeler. Discontinuous galerkin methods for flow and
transport problems in porous media. Communications in Numerical Methods
in Engineering, 18:63–68, 2002.
[19] B. Rivière and M.F. Wheeler. A posteriori error estimates and mesh adapta-
tion strategy for discontinuous galerkin methods applied to diffusion problems.
Computers & Mathematics with Applications, 46(1):141–163, 2003.
[20] D. Robinson. Tansley review no. 73. The responses of plants to non-uniform
supplies of nutrients. The New Phytologist, 127:635–674, 1994.
[21] S. R. Schnapp, W. R. Curtis, R. A. Bressan, and P. M. Hasegawa. Growth
yields and maintenance coefficients of unadapted and NaCI-adapted tobacco
cells grown in semicontinuous culture. Plant Physiol, 96:1289–1293, 1991.
[22] J. Stoer and R. Burlisch. Numerische Mathematik 2. Springer, 4 edition, 2000.
[23] Gilbert Strang. On the construction and comparison of difference schemes.
SIAM J. Numer. Anal., 5:506–517, 1968.
[24] H. Sudo, T. Yamakawa, M. Yamazaki, N. Aimi, and K. Saito. Bioreactor pro-
duction of camptothecin by hairy root cultures of Ophiorrhiza pumila. Biotech-
nology Letters, 24:359–363, 2002.
[25] P. K. Sweby. High resolution schemes using flux-limiters for hyperbolic conser-
vation laws. SIAM J. Num. Anal., 21:995–1011, 1984.
[26] C. H. Takimoto, J. Wright, and S. G. Arbuck. Clinical applications of the
camptothecins. Biochim Biophys Acta, 1400:107–119, 1998.
[27] M. Wink, A. W. Alfermann, R. Franke, B. Wetterauer, M. Distl, J. Windhövel,
O. Krohn, E. Fuss, H. Garden, A. Mohagheghzadeh, E. Wildi, and P. Rip-
plinger. Sustainable bioproduction of phytochemicals by plant in vitro cultures:
anticancer agents. Plant Genetic Resources. NIAB., 3(2):90–100, 2005.
Simulation and Optimization
of Bio-Chemical Microreactors
Rolf Rannacher and Michael Schmich
Institut für Angewandte Mathematik, Universität Heidelberg,

Im Neuenheimer Feld 293/294, 69120 Heidelberg, Germany
{rolf.rannacher,michael.schmich}@iwr.uni-heidelberg.de
Summary. We describe the main steps in the development of a tool for the nu-
merical simulation and optimization of a bio-chemical microreactor. Our approach
comprises a 2D/3D grid generator for the prototypical Lilliput R
chip, a stabilized
finite element discretization in space coupled with an implicit time-stepping scheme
and level-set techniques.
1 Description of the Project

Aim of this project is the development of a numerical tool for the simulation
and optimization of so-called microfluidic diagnosis chips which are used for
instance in clinical microbiology. In their channels and cavities, fluid dynami-
cal and chemical processes take place. Because of the large surface-to-volume
ratio, microfluidic aspects like capillary pressure, adhesion and viscosity, in
contrast to macroscopic flow problems, dominate the overall behaviour of
such chips.
1.1 The Lilliput

R
Analysis Laboratory
An example of such a “lab-on-a-chip” is the Lilliput

R
analysis laboratory, see
Fig. 1, which was developed by Boehringer Ingelheim microParts GmbH for
Merlin Diagnostika GmbH. This 20 × 37 × 3 mm plastic chip has 96 reaction
arrays with a volume of only 1.8 μl each. This corresponds to one hundredth
of the volume of conventional titer plates. Hence, much more tests can be
performed with the same amount of sample.
1.2 The Mathematical Model
The flow on the diagnosis chip is modelled as a two phase flow. Here the two
phases are the liquid and the air that is replaced by it. The flow is governed
118 R. Rannacher, M. Schmich
Fig. 1. “Lab-on-a-chip” Lilliput

R
by the incompressible Navier–Stokes equations. It is driven by a geometry

dependent pressure difference – as a simplified model for capillary pressure –
as well as other volume forces like gravitation. The whole system of equations
written in primitive variables reads
ρ∂t v − μΔv + ρ(v · ∇)v + ∇p = ρf,

∇ · v = 0,
∂t ρ + v · ∇ρ = 0.
This set of equations is completed by appropriate initial and boundary con-

ditions. For the tracking of the interface between both phases we use a level
set approach, see e. g. [OF03]. To this end, we introduce the so-called level set
function φ which is positive in the fluid and negative in the air. We then have
∂t φ + v · ∇φ = 0.
Let ρfl and ρgas denote the densities of the fluid and the air respectively and let
μfl and μgas denote the corresponding viscosities. Introducing the Heaviside
function
0 x < 0,
H(x) =
1 x > 0,
we may write
ρ = ρgas + (ρfl − ρgas )H(φ) and μ = μgas + (μfl − μgas )H(φ).
1.3 Goal of the Numerical Simulation
Goal of the numerical simulation is to analyse the process of the filling of

the chip with liquid. This is essential for the subsequent optimization of the
geometry of the chip such that an almost uniform filling of the reaction arrays
under all relevant operating conditions is achieved. This requires a careful
balance between driving forces like capillary pressure and antagonistic forces
like gravitation or friction.
Simulation and Optimization of Bio-Chemical Microreactors 119
To achieve this goal, many difficulties have to be overcome:

• very complex three dimensional geometry (see Fig. 1),
• highly anisotropic meshes,
• small time steps to resolve the whole filling process,
• locally refined meshes for capturing the interface which strongly influences
the velocity of propagation.
These complications result in high computational costs.
2 The Solution Approach
Our solution approach is based on a Galerkin finite element method in space

coupled, in the first implementation, with the backward Euler time stepping
scheme for the temporal discretization. The finite element method is based on
a variational formulation of the governing equations.
Let (·, ·) denote the inner product of L2 on Ω. Furthermore, H 1 (Ω) de-
notes the space of L2 functions with generalized (distributional) first-order
derivatives in L2 (Ω).
2.1 Temporal and Spatial Discretization
Choosing discrete time points 0 = t0 < · · · < tm < · · · < tM = T and applying
the backward Euler scheme to the system of equations yields
−1 m
ρkm (v − v m−1 ) − μΔv m + ρ(v m · ∇)v m + ∇pm = ρf m , (1)
∇ · v m = 0, (2)
−1 m
km (φ − φm−1 ) + v m · ∇φm = 0, (3)
for m = 1, . . . , M , where km := tm − tm−1 . This system has to be com-

plemented by appropriate initial and boundary conditions. Furthermore, we
replace H by a regularized version Hε defined as
⎧
⎪
⎪
⎨0 if x < −ε,
sin( πx
ε )
1 x
Hε (x) := 2 1 + ε + π if |x| ≤ ε,
⎪
⎪
⎩1 if x > ε,
and set
ρε := ρgas + (ρfl − ρgas )Hε (φ), με := μgas + (μfl − μgas )Hε (φ).
After introducing the spatial discretization, we will choose ε = O(h) varying

with the local mesh size h.
The variational formulation of the problem is obtained by multiplying

equations (1)–(3) by appropriate test functions ϕ := (ψ, ξ, χ) and integrating
over the domain Ω. This leads us to define the semi-linear form a(·)(·):

a(um )(ϕ) := ρε v m + km ρε (v m · ∇)v m , ψ + km (με ∇v m , ∇ψ)
− km (pm , ∇ · ψ) + km (∇ · v m , ξ) + (φm + km v m · ∇φm , χ)
with um := (v m , pm , φm ). The variational form of system (1)–(3) then reads:

For m = 1, . . . , M find um ∈ V + û such that
a(um )(ϕ) = F (ϕ) ∀ϕ ∈ V,
where

F (ϕ) = (ρ v m−1 + km ρ f m , ψ) + (φm−1 , χ) − km P ψ · n ds
Γin
and û represents prescribed Dirichlet boundary conditions while P is the pre-

scribed mean value of the pressure on the inflow boundary which drives the
flow. The function space V is the tensor product of certain subspaces of H 1 (Ω)
for the velocity and the level set function while the space for the pressure is
L2 (Ω).
The spatial discretization via the Galerkin finite element method is de-
fined on quadrilateral/hexahedral meshes Th = {K} covering the domain Ω.
The trial and test spaces Vh ⊂ V consist of continuous, piecewise polynomial
functions. The polynomial space is Q1 , the space of (isoparametric) tensor-
product polynomials of degree 1. For a detailed description of this standard
construction process, we refer to [BS02]. In order to facilitate local mesh re-
finement, we allow the cells to have nodes which lie on midpoints of faces of
neighboring cells, see Fig. 2. There are no degrees of freedom associated to
such “hanging nodes”. The value of a finite element function at such points
is determined by interpolation to enforce global continuity.
Since we use equal-order trial functions for v and p, our scheme requires
pressure stabilization because the Babuška–Brezzi inf-sup-stability condition
is not fulfilled. In addition, the convective terms need stabilization. This is
done via the so-called Local Projection Stabilization (LPS), see e. g. [BB06].
To this end, we introduce the interpolation operator i2h : Vh → V2h from
the current grid to a coarser grid in which four/eight adjacent cells of Th
form a macro cell in T2h . This can always be done if the mesh has a patch
Fig. 2. Mesh with hanging nodes

Fig. 3. Mesh with patch structure and corresponding coarsened mesh
structure, see Fig. 3. Let then the fluctuation operator π : Vh → Vh be defined

as π := id − i2h . The stabilization process comprises in the modification of
the original semi-linear form by a mesh-dependent semi-linear form:
ah (um )(ϕ) := a(um )(ϕ) + sh (um )(ϕ)
where sh (um )(ϕ) is given by

$
sh (um )(ϕ) := km αK (∇πpm , ∇πξ)K
K∈Th
(v)
+ δK (ρε (v m · ∇)πv m , ρε (v m · ∇)πψ)K
#
(φ)
+ δK (v m · ∇πφm , v m · ∇πχ)K .
The first term accounts for the pressure stabilization whereas the second and
third term serve as stabilization of the convective term for the velocity and
(v) (φ)
the level set function respectively. The cellwise parameters αK , δK , and δK
are chosen as
h2K
αK := α0 ,
6με + hK ρε v m K
(v) (v) h2K
δK := δ0 −1 2 ,
6με + hK ρε v m K + km hK
(φ) (φ) h2K
δK := δ0 −1 2
hK v m K + km hK
(v) (φ)
with α0 , δ0 , δ0 ≈ 0.3. Hence, the discrete system of equations to be solved
h ∈ Vh + ûh such that
reads: For m = 1, . . . , M find um
ah (um
h )(ϕh ) = F (ϕh ) ∀ϕh ∈ Vh . (4)
2.2 Solution of the Algebraic Systems
In each time step, we have to solve a nonlinear system of equations. This is

done via Newton’s method. If we denote by un the n-th iterate in computing
the discrete solution um h , one Newton step for (4) consists of solving the linear
system
ah (un )(δun , ϕh ) = F (ϕh ) − ah (un )(ϕh ) ∀ϕh ∈ Vh , (5)
for the correction δun := un+1 − un . The directional derivative of the semi-
linear form ah (·)(·) at un in direction δun , given by
1$ #
ah (un )(δun , ϕh ) := lim ah (un + sδun )(ϕh ) − ah (un )(ϕh ) ,
s→0 s
is derived analytically on the continuous level and then discretized. The linear
subproblems (5) are solved by the GMRES method with preconditioning by
a multigrid iteration. The multigrid component uses a block ILU smoother
in which those unknowns corresponding to the same node of the mesh are
blocked. For a detailed description of this approach, we refer to [BR99].
3 Numerical Results
The first step was to implement a grid generator for (a part of) the geometry
of the chip. The specific features of this geometry lead to anisotropic cells
with aspect ratios of approximately 1:10. A picture of the resulting coarse
grid which we used in our computations can be seen in Fig. 4.
Fig. 4. Three dimensional coarse grid used for computation (bottom: zoom into
coarse grid)
3.1 Simulation
In the first computation we simulated a flow with two identical phases. At

the beginning, the whole chip is filled with fluid 1 while fluid 2 is entering at
the lower left part of the computational domain. This computation was done
without respecting gravitational forces, i. e., we set f = 0. Figure 5 shows the
propagation of fluid 2 (red) at different times.
Fig. 5a–j. Computation 1: Filling process (phase 1 blue, phase 2 red )
The next computation, which was done on only a part of the chip’s domain,
simulated a real two phase flow under the influence of gravitation. Figure 6
shows the corresponding sequence of snapshots. One can nicely see how the
liquid flows through the channels down into the reaction arrays and fills them
before leaving them.
Fig. 6a–j. Computation 2: Filling process (phase 1 blue, phase 2 red )
For a better tracking of the interface between both phases, we like to ap-
ply adaptive mesh refinement. The selection of the cells to be refined here
is still a quite heuristic one: We mark a cell K to be refined if the level-set
function φ has positive as well as negative values in K which corresponds
to the fact that the interface which is given by points x with φ(x) = 0 lies
in this cell. This technique has been applied to a 2D version of the chip.
Figure 7 shows by a sequence of locally refined meshes that this method
works quite well. We see that the interface between both phases is always lo-
Fig. 7a–c. Computation 3: Adaptive mesh refinement for tracking of the interface
cated in the most refined zone. In the next step, the mesh refinement will be
done automatically, driven by residual and sensitivity information, in order
to obtain the correct propagation speed of the interface. For further infor-
mation on this so-called Dual Weighted Residual (DWR) method, we refer
to [BR01].
3.2 Optimization
On the 2D version of the chip, we also performed a first simple optimization.

By slightly modifying the chip’s geometry, we were able to achieve a much
more efficient filling of the reaction arrays. The result of this optimization
process can be seen in Fig. 8 where the corresponding filling states at different
times are compared.
4 Conclusion and Outlook

We have presented a tool for the numerical simulation of chemical microreac-
tors. The method is based on the incompressible Navier–Stokes equations with
non-homogeneous density and uses equal-order finite elements together with
local projection stabilization as well as level-set techniques for the tracking
Fig. 8a–h. Solution before (left) and after (right) geometry “optimization”
of the interface. At first, computations with two identical phases and with-
out gravitation have been performed. In a second step, we simulated a two
phase flow under the influence of gravitation. In this context, adaptive mesh
refinement for tracking the interface between both phases has been tested for
a 2D version of the chip. The next step will be to combine the 3D solver
components with adaptivity based on the DWR method for a better inter-
face tracking. Finally, we will incorporate effects by surface tension as well as
capillary forces.
References
[BS02] Brenner, S. C., Scott, L. R.: The Mathematical Theory of Finite Element
Methods. Springer, New York, Berlin, Heidelberg (2002)
[OF03] Osher, S., Fedkiw, R.: Level Set Methods and Dynamic Implicit Surfaces.
Springer, New York, Berlin, Heidelberg (2003)
[BR01] Becker, R., Rannacher, R.: An optimal control approach to a posteriori error
estimation in finite element methods. In: Iserles, A. (ed) Acta Numerica
2001. Cambridge University Press (2001)
[BR99] Braack, M., Rannacher, R.: Adaptive finite element methods for low-Mach-
number flows with chemical reactions. In: Deconick, H. (ed) 30th Compu-
tational Fluid Dynamics. The von Karman Institute for Fluid Dynamics,
Belgium (1999)
[BB06] Braack, M., Burman, E.: Local projection stabilization for the Oseen prob-
lem and its interpretation as a variational multiscale method. SIAM J.
Numer. Anal., 43(6), 2544–2566 (2006)
Part IV
Computeraided Medicine
Modeling and Optimization of Correction
Measures for Human Extremities
René Brandenberg, Tobias Gerken, Peter Gritzmann, and Lucia Roth
Zentrum Mathematik, Technische Universität München, Boltzmannstr. 3,

85747 Garching bei München, Germany
{brandenb,gerken,gritzman,roth}@ma.tum.de
Summary. Deformities of the lower extremities can be congenital or post-traumatic.

Recent progress in their treatment arises from a new technique based on fully im-
plantable intramedullary limb lengtheners. This allows patient friendly and effective
treatment of complex length and rotation deformities. The advancement in tech-
nology and growing demands on the precision of a correction require a revision of
common operation planning schemes. In this paper, we discuss the questions from
computational geometry that arise in semi-automated operation planning based on
three-dimensional CT data. We present algorithms for these problems that have
been implemented and integrated into a newly developed 3D planning device.
1 Introduction
Deformities of the lower extremities can be congenital or post-traumatic. They
can be treated according to the principle of callus distraction which allows for
bone growth among adults. This was first studied systematically by Siberian
orthopedist G. Ilizarov [33], [34]. Classical treatment of such deformities relies
on the Ilizarov apparatus, and standard operation planning schemes [44] are
based on 2D X-rays (see Fig. 1).
Recent progress arises from a new technique based on fully implantable
intramedullary limb lengtheners (see Fig. 1) that were designed by R. Baum-
gart and his team at the Limb Lengthening Center Munich [6]. Treatment
with the aid of this device is much gentler on the patient. Moreover, the risk
of infections is minimized as the apparatus, once implemented, is fully se-
cluded within the bone, with no parts penetrating the skin. In contrast to
the Ilizarov apparatus, however, the lengthening direction is completely de-
termined with the implantation of the device and non-invasive corrections
during the distraction phase are impossible. This fact results in increasing
demands on the precision of common operation planning procedures. There-
fore, in cooperation with the Limb Lengthening Center, a new planning
scheme based on three-dimensional computer tomography (CT) data was de-
veloped.
132 R. Brandenberg et al.
Fig. 1. The Ilizarov apparatus, classical operation planning in 2D, and an implanted
intramedullary nail
The paper is organized as follows: In Sect. 2, we briefly describe the treat-

ment of deformities. Additionally, we identify the basic tasks that are specific
to the planning procedure for the new technique. In Sect. 3, we introduce
the underlying mathematical problems that arise from the previous section.
Algorithmic solutions to these problems are discussed in Sect. 4. Finally, in
Sect. 5, we describe the resulting software tool.
2 Medical Treatment of Deformities

In this section, we briefly discuss the basic tasks of deformity correction.
For more detailed information on the medical background see [6], [44] and
the references there. The general procedure in deformity correction can be
characterized by the following steps (compare Fig. 2):
A. Analysis of the state of the patient’s musculoskeletal system.
B. Identification of the site of the osteotomy and the positioning of the limb
lengthening device in order to achieve an optimal post-treatment status.
C. Implantation of the intramedullary nail and callus distraction.
D. Consolidation of newly formed bone tissue and removal of the device.
The present paper is concerned with the planning procedure (Step B), where
mathematical models come into play. Based on the current state of the pa-
tient’s musculoskeletal system and relying on biomechanical considerations,
a goal state has to be defined. Then, the site of the osteotomy and the po-
sition of the limb lengthening device must be determined accordingly. The
planning process relies on critical points and axes. These characterize the
Modeling and Optimization for Extremity Correction 133
Fig. 2. Schematic illustration of a deformity correction procedure: initial state,

post-operational state, projected post-treatment state, and consolidated state
Fig. 3. Slight inaccuracies cause significant deviation after lengthening
anatomical and the (idealized) mechanical components of the musculoskele-

tal system. Deformities are identified by comparison of the axes and an-
gles to those of normal alignments. Ideally, hip, knee and ankle should be
collinear in the frontal view [44]. Naturally, the details of Step B depend
on the specific device. In particular, limb lengthening with intramedullary
nails has to be carried out along the anatomical axis of the corresponding
bone.
We stress again the critical fact that once placed inside the bone, no post-
operational corrections of the implanted device are possible. Lengthening with
a slightly improperly placed nail may result in serious malalignment that may
necessitate further operations (compare Fig. 3). Hence, the intramedullary
technique demands high precision in operation planning.
3 Basic Geometric Problems in Operation Planning
3.1 Medical Modeling
In this section, we discuss the mathematical questions that result from mod-
eling the planning procedure. Naturally, our mathematical model must reflect
the actual medical situation as exactly as necessary but must also meet the
demands on computation times for an interactive software tool executable on
standard devices. We do not aim at a fully automated planning process but
at a tool supporting the physician’s work, while still allowing comprehensive
use of expertise and experience.
We presume that a voxel based model of the patient’s skeleton has been
extracted from three-dimensional CT data via standard image processing tech-
niques. More precisely, the skeleton may be given as a finite point set P ⊂ R3 .
Registration of limb images in the context of orthopedic interventions is prac-
tical and, ‘since the bone contrast [in CT or X-ray images] is very high, most
methods, including segmentation tasks, can be automated’ [39, p. 20]. For
details about medical imaging, see e.g. [5], [47] and the references therein.
The primal task is now to identify different centers and axes of P , which
represent the geometry of the extremity, including the center of the femur
head or the knee joint, the site of the deformity, and the anatomical axis of
the thigh bone. In addition, the placement of the intramedullary nail has to
be determined.
The center of a the femur head may be found via approximating the con-
cerned part of the bone by a Euclidean ball. First models for the central part of
a long bone may be circular or elliptical cylinders, others may involve several
cylindrical segments. This coarse approximation is appropriate for the specific
application as the key feature for a successful operation is the anatomically
correct axis direction.
A special geometric approximation problem in operation planning is find-
ing the site for an optimal osteotomy. A deformity involving twisting and
bending of the bone axis can be corrected via a single cut, followed by a suit-
able rotation and translation of the two resulting bone segments [29]. The
affected part of the bone can be modeled by two cylindrical segments with
intersecting axes. The site of the osteotomy is therefore determined by the
point where the axes meet (the center of the deformity) and their directions.
When specifying the approximation strategy, one should be aware of the
fact that the data points are usually distributed non-uniformly on the bone
surface. In fact, in order to minimize radiation exposure, high-precision CT is
limited to the parts of the skeleton which are most relevant for the operation
(e.g. joints). A lower precision is used for less relevant parts. Since the planning
procedure relies in an essential way on the mechanical and anatomical axes,
we are interested in the geometric structure underlying the data points. The
planning procedure also includes the device positioning. This step is carried
out as follows in the 3D model: the part of the bone cut by the osteotomy is
virtually set to its position in the goal state and the extracted nail is placed
so as to appropriately connect the two parts. The post-operational state is
now determined by simply contracting the lengthening device and translating
the cut part accordingly (compare Fig. 2).
The intramedullary nail can be seen as a (circular) finite cylinder (com-
pare Fig. 1). In order to locate a proper placement, we need an appropriate
geometric model for the interior of the bone. A simple set of data points does
not provide this directly. Therefore, we represent the bone by a structure built
from blocks that are the convex hulls of two ‘adjacent’ circles or ellipses re-
spectively. The ellipses are approximations of the original CT-slices of bone
layers. Note that these layers need not be parallel any more, since, in the goal
state, bone parts have been cut, rotated and translated.
Consequently, we are lead to a set of basic geometric problems, all involving
the containment of objects, which can be classified within the more general
settings of computational geometry or computational convexity. While our
main application ‘lives’ in 3D, we will formally introduce the relevant prob-
lems in general dimensions in order to give a more complete account of their
structure. Later, we will specialize again to the specific medical application.
3.2 Mathematical Description
The first geometric problem occurs when approximating parts of the thigh
bone with basic geometric shapes (Fig. 4). These shapes are obtained from
simple geometric objects via homothety or similarity: Two sets C, C ∗ are
similar if there exists a rotation Φ : Rd → Rd , c ∈ Rd and ρ ≥ 0, such that
C ∗ = ρΦ(C) + c. If Φ can be chosen as the identity, C and C ∗ are homothetic.
& point set P ⊂ R and

d
Problem 1 (Basic k-Containment). For a finite
∗
closed convex sets C1 , . . . , Ck ⊂ R , find C = i ρi Φi (Ci ) + ci such that
d
P ⊂ C ∗ , minimizing maxi ρi .
Important examples for basic 1-containment are the computation of the
smallest enclosing Euclidean ball (C1 = Bd , the Euclidean unit ball) or a small-
est enclosing cylinder (C1 = l + Bd , where l denotes a 1-dimensional linear
subspace). Of course, as we have seen, more complicated container shapes and
higher numbers of different containers occur naturally. Other objective func-
tions may be chosen in the above problem definition, such as the sum of the
dilatation factors. The above choice is justified in our application since the
proportions of the bone parts (and the geometric objects representing them)
are roughly constant.
Some of our medical tasks involve more general containments. For instance,
the problem of finding the center of deformity leads to a restricted version of
the double-ray center problem, a 2-containment problem where the transfor-
mations of the two container sets are not independent (which is the reason
for the ‘double’ instead of a 2 in the name). The double-ray center problem
Fig. 4. Approximating parts of the thigh bone with basic geometric shapes: complete
leg and detail
is a well known mathematical problem. For our medical application, an addi-

tional constraint about the position of the center c of the deformity needs to
be enforced (see Sect. 4.3 for details).
Problem 2 (Restricted Double-Ray Center). For a finite set of points
P ⊂ Rd , an approximate center c ∈ Rd , and a maximal distance δ, determine
two rays r1 = {c + λv1 : λ ≥ 0} and r2 = {c + λv2 : λ ≥ 0}, with v1 , v2 ∈
Rd \{0}, emanating from the same center c ∈ Rd such that P ⊂ (r1 ∪r2 )+ρBd ,
c − c ≤ δ, minimizing ρ.
As we saw above, the device positioning problem involves non-convex con-
tainers.
Problem 3 (Cylinder Inclusion). For given length λ, radius ρ, and a se-
quence E1 , . . . , Em of (d − 1)-dimensional ellipsoids, find a line segment L of
length λ such that for the cylinder Z(L, ρ) = L + (ρB ∩ L⊥ )
,
m−1
Z(L, ρ) ⊂ conv (Ei ∪ Ei+1 )
i=1
holds or decide that no such cylinder exists.
4 Mathematical Treatment
In this section, we give an overview of the state of the art for the mathematical
problems raised in Sect. 3 and their relatives as well as our basic algorithmic
approaches. First, a general framework is provided through the notion of Op-

timal Containment . Second, the specific problems from Sect. 3 are discussed
in detail.
4.1 Optimal Containment
The notion of containment problems generically summarizes all problems of

computing, approximating or measuring the sets which among a given class
are the largest contained in or the smallest containing a given set [26].
Optimal Containment. Let B and C be families of closed sets in Rd and

∗ ∗
0 ∪ {∞} a functional. The task is to find B ∈ B and C ∈ C
ω : B × C → R+
∗ ∗ ∗ ∗
such that B ⊂ C and ω(B , C ) is minimal or to decide that no such pair
exists.
In order to cope with optimal containment problems algorithmically, spe-

cialized problem classes have to be considered. Since set inclusion provides
a partial order on the objects in B and C, it is natural to ask for ω to be
monotonically decreasing in the first argument and increasing in the second
with respect to this order.
In case of the outer containment problems considered here, the inner ob-
ject is a specific point set and C is a family of sets obtained from one or
more reference containers. The function ω therefore depends only on the con-
tainers. In an instance of Problem 1, k basic convex containers Ci are fixed
and C represents & all feasible transformations of the containers; i.e., C is the
set of all unions ki=1 ρi Φi (Ci ) + ci with ρi ≥ 0, Φi a rotation, and ci ∈ Rn
for all i = 1, . . . , k. In the double-ray center problem, the transformations
of the two rays are not independent. Hence, C consists of the Minkowski
sum of unions of two rays with common origin c and a ball of radius ρ
there.
In Problem 3, an instance is defined by a specific outer container formed
by the given ellipsoids and B is the set of all cylinders of a given length and
radius. All feasible inner objects can therefore be obtained by rotations and
translations of one representative cylinder. (Of course, in theory, this kind of
problem can also be formulated as an outer containment problem since all
involved transformations are invertible.)
One should recognize that in all three problems discussed here, ω(C ∗ ) is
obtained from the dilatation factors of C ∗ . Other objectives may be considered
in many variants, some of which are relevant to the medical tasks described
here. Ongoing research is concerned with containment under affinity, where
the objective may for instance involve the volume of C ∗ .
In the survey [26], the complexity of some convex containment problems
in general dimensions is addressed, classifying the problems according to the
representation of the container, the enclosed object and the functional ω. Due
to the present application, our focus is the three-dimensional case. However,
we will point out NP-hardness results for specific problems in varying dimen-
sions as they can, to a certain extent, explain jumps in the computational
costs from dimension two to three.
4.2 Problem 1: Basic k-Containment
Theoretical background: First, let us consider the task of covering a point set
with a (Euclidean) ball of minimal radius as it occurs when determining the
center of the femur head. Since all Euclidean balls are translations and dilata-
tions of the unit ball, this is a case of basic 1-containment under homothety
with C1 = Bd , see e.g. [14], [20], [25], [37], [52], or [46] for an overview. The
computational complexity of 1-containment under homothety in general is
addressed in [17] and [26].
While the smallest enclosing ball can be regarded as containment under
homothety, the smallest enclosing cylinder is an example of containment under
similarity and is known to be NP-complete in general dimensions [40] (a result
derived from the close relation to line
stabbing problems, see also [1], [27], [31], and compare Sect. 4.4). In fixed
dimensions polynomial time solvability is shown in [19], but neither this re-
sult nor the polynomial time approximation schemes in [30] and [31] provide
practical algorithms in 3D.
Both smallest enclosing balls and cylinders are special cases of outer radii
of convex bodies [24], [25], [50]. Apart from shape approximation, the problem
of computing inner and outer radii arises in various applications in computer
science, optimization, statistics, and other fields [25].
In general, containment under similarity is substantially harder than con-
tainment under homothety. Here is another example: while deciding whether
a point set can be covered by an axis aligned unit cube is trivial, the com-
plexity of the corresponding task allowing rotations is unknown, though con-
jectured to be NP-complete in general dimensions [40]. Note that a regu-
lar simplex in dimension√ d can be covered by a rotated unit cube scaled by
a dilatation factor of d if and only if a Hadamard matrix of order d + 1
exists.
The medical application requires also approximations by several objects.
Regarding basic k-containment problems under homothety with k ≥ 2, as
in the case of k = 1, most attention has been placed on the Euclidean k-
center problem which asks for the minimal radius to cover a given point set
with k balls. For k ≥ 2, it is NP-complete in general dimensions [40]. The
k-center problem for axis aligned cubes can still be solved in polynomial time
for k = 2 but is NP-complete for k ≥ 3 [40]. Many practical approaches focus
on the planar 2-center problem, e.g. [15], [18]. However, if k is part of the
input, approximation to a certain accuracy is NP-complete for both, circles
and squares, even in the plane [41].
Very few papers deal with k-containment problems under similarity as in
higher dimensions it combines the difficulties of both rotation with k = 1
and homothety for k > 1. A special case is the k-line center problem, that is
covering a point set with k cylinders of equal radius. While in [42] it is shown
that already the planar case is NP-complete if k is part of the input, a linear
time approximation scheme for every fixed k in 2D is given in [3].
Algorithms: 1-Containment under homothety is fundamental here since it oc-

curs not only in many applications itself but also in subroutines for even harder
containment problems. Already three-dimensional instances of 1-containment
under similarity or basic k-containment under homothety require considerable
computational effort. Usually, these problems are solved within branch-and-
bound schemes that rely on the solution of many instances of 1-containment
under homothety [10], [14], [16], [36]. Of course, basic 1-containment under
homothety is a convex programming problem and can in principle (assum-
ing well-boundedness) be solved by the ellipsoid method if a suitable sep-
aration oracle is known for C1 . Typically, however, the ellipsoid method is
not competitive when more information about the container is available. For
the special case of the smallest enclosing ball problem, for instance, one can
choose from a considerable number of fast algorithms, see e.g. [20], [37] and
[52] for implementations and experimental studies. In [11], different solution
strategies for 1-containment under homothety are formulated and tested. De-
pending on the container shape, some instances can be formulated as Lin-
ear or Second-Order-Cone Programs. A cutting plane algorithm also based
on a separation oracle but with good practical performance is provided as
well.
So, while basic 1-containment under homothety is usually solvable in the-
ory and in practice, a suitable approach to 1-containment under similarity is
less straightforward. Even in the case of a smallest enclosing cylinder, the com-
plexity increases significantly from planar to 3D instances. In [12], algebraic
methods for smallest enclosing and circumscribing cylinders of simplices are
addressed and a polynomial formulation with significantly reduced algebraic
complexity is provided. Exact algorithms for smallest enclosing cylinders in
3D are proposed in [1] and [48]. In [31], it is shown that the radius can be
approximated in linear time in both the dimension and the number of input
points.
When k ≥ 2, basic k-containment can be reduced to k basic 1-containment
problems by finding an optimal partition of the point set P . A core set ap-
proach yields a linear time approximation scheme for the Euclidean k-center
problem in general dimensions [14]. The algorithm uses the fact that the core
set sizes depend neither on the dimension nor on the number of input points,
but only on the error and therefore, for a fixed error bound, enumeration
of all possible partitions of the core sets is feasible (in principle). A more
practical algorithm based on a simple branch-and-bound scheme is proposed
in [36]; however, computations for 3D point sets from practical applications
are impossible when k > 4 or ε < 0.01. In [10], algorithms for the general
basic k-containment problem under homothety are developed. One should be
aware that these problems do not admit core sets of sizes independent of the
dimension in general [11]. Moreover, the algorithms that do compute small

core sets in the Euclidean case may construct dimension dependent core sets
when the container is a cube, even though every diametral pair of points is
a core set of size 2 in this case. See [7] for a fast algorithm when C is the cube
and k = 2.
Treatment in practice: For the smallest enclosing ball, we use an approx-
imation algorithm based on Second-Order-Cone Programming [37]. Basic
1-containment problems under homothety with other container shapes are
solved by an approach suggested in [11].
The smallest enclosing cylinder computation is based on non-orthogonal
projections along a-priorily fixed directions given by an adaptive discretiza-
tion of the unit ball [16]. Though this algorithm has exponential running time
in general dimensions, it is more practical for three-dimensional input than
the polynomial time approximation scheme in [31] aiming at high dimensions.
Nonetheless, a direct implementation of the algorithm is slow and additional
effort has to be made in order to reduce running times in practice [45]. In gen-
eral, computing the radius of the smallest enclosing cylinder is a non-convex
problem which can have many local minima. However, the bone data is not
a worst-case input but in fact rather cylindrical and therefore large regions for
the cylinder axis can be discarded in advance. A first step towards practical
computations is to derive a-priori bounds for the radius, e.g. via ellipsoidal ap-
proximation [45], semidefinite programming [50], or diameter approximation
[31]. If one allows cylinders with elliptical (rather than just circular) cross-
sections, ellipsoidal approximation even allows good a-priori bounds on the
minimal volume of the elliptical cross-sections and almost optimal a-posteriori
results in our application.
We use regular discretizations of the 3D sphere instead of techniques de-
veloped with a view towards general dimensions. The discretization is refined
adaptively, the direction vectors are ranked according to their likelihood to
improve the currently best solution and processed in a branch-and-bound
scheme. Incrementally generated core sets [51] are used to reduce the process-
ing time caused by the size of P .
In order to solve basic k-containment problems when k ≥ 2, we use practi-
cal techniques described in [10]. A-priori upper and lower bounds (with prov-
able quality even for different, not necessarily symmetric containers) are com-
puted via methods based on diameter partitioning. Moreover, relaxations of
a mixed integer Second-Order-Cone formulation of the problem are solved
at each node of the branch-and-bound-tree instead of simply computing the
dilatation factors for the current partition. Besides being capable of solving
containment problems with general and possibly different containers, the de-
scribed approach even improves the performance for the Euclidean k-center
problem [10].
Besides the general methods discussed here, using a fast local search heuris-
tic can be profitable when significant additional information about a basic
containment problem is provided by the application.
4.3 Problem 2: Restricted Double-Ray Center
Theoretical background: The planar double-ray center problem is discussed

in [23]. Here we focus on the 3D case, though in principle, the algorithm
we formulate generalizes to higher dimensions. Related problems include the
smallest enclosing cylinder (compare Problem 1), the anchored ray problem
[21], and the 2-line center problem [2], [3]. Note that for finite point sets
covering with two cylinders is the same as covering with two half-cylinders.
However, the additional constraint in the double-ray center problem is forcing
the two axes to have a common point (see Sect. 4.2) which adds further
difficulties. In fact, a partition of the point set into two parts approximated by
a single ray each does not yield a reduction to independent subproblems. Even
when the partition of the point set is known, the position of one ray always
depends on the other as the rays must issue from a common center. This
makes approaches based on projections and algebraic formulations difficult in
spite of the fact that the problem has fewer degrees of freedom than the 2-line
center problem [22].
In case of the general double-ray center, a slight shift of just one of the
input points may change the optimal configuration completely [22]. The same
may happen in case of the smallest enclosing cylinder, but only for ‘non-
cylindrical’ input data. In case of the double-ray center problem, however,
even well-shaped input may yield unwanted optimal solutions. For instance,
the optimal solution can degenerate to parallel rays meeting at infinity. In
order to get meaningful solutions for the application, the additional constraint
on the center c of the deformity has to be respected.
Algorithms: In [22] an approach based on testing all relevant partitions of the
point set and solving the resulting optimization problems is proposed. At first,
the fraction of the points close to the center is determined and afterwards the
other points are split by separating hyperplanes into two subsets belonging to
one ray each. The final optimization problem is still non-convex, but can be
handled by discretizations of the direction space (similar to those considered
for smallest enclosing cylinders in Sect. 4.2). The algorithm is polynomial in
the number of input points in fixed dimensions, but a full computation is still
impracticable for point sets of sizes relevant for our application.
Of course, when the specified ball containing c is very small, the prob-
lem character changes and it can be advantageous to consider alternate ap-
proaches, such as testing for the center position instead of the ray directions.
Treatment in practice: In the medical application, the ray directions are lim-
ited and possible sites for the center are in fact constrained to a small area,
as, obviously, it should be located inside the bone (see Fig. 5). Constraints
on the ray directions, e.g., on the angle between a ray and a given axis or
between the two rays, can easily be taken into account in the algorithm. Since
many possible ray directions can be disregarded for the medical application,
the restricted version now actually allows for practical computations.
Fig. 5. The double-ray center problem in 3-space
An initial approximation of the double-ray center problem can easily be

found directly allowing a local search strategy in the tool. The algorithm then
starts from an approximate solution which is used to constrain the possible
partitions of the point set, the position of the center and the directions of
the emanating rays and optimizes the site of the cut for the restricted prob-
lem.
4.4 Problem 3: Cylinder Inclusion

Problem modification: For an optimal positioning of the nail, medical require-
ments have to be met. Criteria for a ‘good’ nail position are the stability of
the arrangement and the expected healing process of the bone tissue. There-
fore, the nail should leave enough margin on all sides and the cut surfaces of
the bone parts divided by the osteotomy should have a large overlap in the
post-operational state. In addition, the nail has to penetrate both bone parts
by a sufficient depth to guarantee stability.
Our approach for the nail positioning is therefore to center the nail in
a cylindrical corridor of maximal radius through the bone. This meets the
demands from the application since the minimal margin between nail and
bone boundary is maximized. Moreover, it guarantees that the cut surfaces in
the post-operational state (or, more precisely, their projections in the plane
perpendicular to the lengthening direction) overlap at least by the area of
a disk of the computed maximal radius. When the cylindrical corridor tra-
verses all the blocks forming the bone model, the nail can be placed in a way
that maximizes the penetration of the two bone segments.
Hence, in the modified problem we are looking for optimal traversing cylin-
ders.
Optimal Traversing Cylinder. Let E = {E1 , . . . , Em } be a finite set of

(d − 1)-dimensional ellipsoids in Rd . The task is to find a line l ⊂ Rd and
a maximal ρ ≥ 0 such that ∅ = (l + ρBd ) ∩ aff (E) ⊂ E for all E ∈ E, or
decide that no such line exists.
Theoretical background: The associated feasibility problem reduces to the

question whether a given arrangement of ellipsoids allows (at least) for
a traversing (or stabbing) line. When the dimension is part of the input,
NP-completeness of this task can be shown using a modification of a proof
from [40]. The original theorem shows the hardness of line stabbing for full-
dimensional balls, yet the construction can be adopted to work also for care-
fully chosen (d − 1)-dimensional balls (compare [27]). Stabbing problems are
strongly related to visibility computations [28]. Applications from computer
graphics are common to both. The structure of the solution space can be quite
complex in general. In 3D for degenerate ellipses the results for transversals of
line segments apply and the solution space may consist of up to m connected
components [13]. Visually, two lines are in the same connected component if
one can continuously be transformed into the other without leaving the solu-
tion space. Lines in the same connected component define the same geometric
permutation of the sets to be traversed, i.e. the induced orders are either
the same or reversed. However, when additional information about the input
data is available, the structure of the solution space is often much simpler.
For instance, when all ellipsoids are parallel, the solution space can actually
be seen to form a convex set in R2(d−1) [4]. For disjoint full-dimensional balls
and a fixed geometric permutation, the set of directions of traversing lines is
also convex [8].
In our application, there is only one reasonable ordering of the ellipses,
yet this information is not explicitly included in the modified formulation.
However, the information has not been lost, since for inputs from the appli-
cation, aff (Ei ) has dimension d − 1 and aff (Ei ) ∩ Ej = ∅ for all Ei = Ej ∈ E.
This suffices to show that the arrangement admits at most one geometric
permutation [35] as any of the hyperplanes aff (Ei ) determines a partition of
E \ {Ei } into two disjoint subsets. It also implies l ⊂ aff (Ei ) for any i and,
consequently, ensures that the modified formulation allows no more than the
desired solutions.
Algorithms and treatment in practice: For the algorithmic solution of the
problem, we are only interested in directions permitting traversing cylinders
of (not too small) positive radius. Assuming we know the direction of the axis,
we can project the ellipsoids in E along the axis direction to obtain ellipsoids
in Rd−1 and then compute the largest ball in the intersection of the projected
ellipsoids (again a containment problem under homothety) via semidefinite
programming; see [9, pp. 45, 46]. An approximately optimal axis direction
can be found by a discretization procedure of the sphere similar to that of
Sect. 4.2 for computing smallest enclosing cylinders.
The size of the discretization can be kept suitably small. For any two
ellipsoids Ei , Ej from E, all feasible direction vectors are normalizations of
points from the difference set Ei −Ej . The most promising strategy is therefore
to pick the lowermost and uppermost ellipsoids E1 and Em to generate the
directions.
Figure 6 shows an approximately optimal cylinder as computed by an
implementation of the described algorithm.
Fig. 6. Example of a traversing cylinder and its projection
5 Software
The algorithms described in Sect. 4 were implemented in Matlabc [49].
A new software tool (see Fig. 7, [32]) for operation planning based on three-
dimensional CT data extending the amirac visualization software [43] was
developed. Additional customized components were created in the C++ pro-
gramming language.
The tool features a simulation of the osteotomy, the repositioning of the
bone and the implantation of the intramedullary nail. It allows both for
straight and hand-drawn cuts. The cut part of the bone can be positioned ac-
cording to specified axes and points. When the position of the intramedullary
nail is specified, an animation of the postoperative limb lengthening is pos-
sible. Semi-automated planning aids are provided for the determination of
relevant axes and points of the musculoskeletal system via geometric approx-
imations. Also, a test for the feasibility of a chosen nail position is included.
Alternatively, the maximal corridor and an optimal nail positioning can be
computed. Usually, additional medical factors like the local state of the bone
tissue and the stability of the resulting bone parts affect the placement of the
osteotomy and the implantation of the device. Therefore, at any stage of the
Fig. 7. 3D–Operation Planning Tool
planning process, the user is able to modify the geometric structure manually,
save it, or reload it.
Acknowledgements
The project was carried out in cooperation with the Limb Lengthening Center
Munich (ZEM), Germany [38]. It is a great pleasure to acknowledge stimu-
lating discussions with Prof. Baumgart, Dr. Thaller and other members of
ZEM. Funding from the German Federal Ministry of Education and Research
(BMBF), grant number 03GRNGM1, is gratefully acknowledged.
References
1. P.K. Agarwal, B. Aronov, and M.Sharir. Line transversals of balls and smallest
enclosing cylinders in three dimensions. Discrete Comput. Geom., 21:473–388,
1999.
2. P.K. Agarwal, C.M. Procopiuc, and K.R. Varadarajan. A (1 + )-approximation
algorithm for 2-line-center. Comput. Geom. Theory Appl., 26(2):119–128, 2003.
3. P.K. Agarwal, C.M. Procopiuc, and K.R. Varadarajan. Approximation algo-
rithms for a k-line center. Algorithmica, 42:221–230, 2005.
4. N. Amenta. K-transversals of parallel convex sets. In Proc. 8th Canadian Conf.
Comput. Geom., pages 80–86, 1996.
5. I.N. Bankman, editor. Handbook of Medical Imaging: Processing and Analysis.
Academic Press, Inc., Orlando, FL, USA, 2000.
6. R. Baumgart, P. Thaller, S. Hinterwimmer, M. Krammer, T. Hierl, and
W. Mutschler. A fully implantable, programmable distraction nail (fitbone) –
new perspectives for corrective and reconstructive limb surgery. In K.S. Leung,
G. Taglang, and R. Schnettler, editors, Practice of Intramedullary Locked Nails.
New developments in Techniques and Applications, pages 189–198. Springer Ver-
lag Heidelberg, New York, 2006.
7. S. Bespamyatnikh and D. Kirkpatrick. Rectilinear 2-center problems. In Proc.
11th Canadian Conf. Comput. Geom., pages 68–71, 1999.
8. C. Borcea, X. Goaoc, and S. Petitjean. Line transversals to disjoint balls. To

appear in Discrete Comput. Geom.
9. S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear matrix inequalities
in system and control theory, volume 15 of SIAM studies in applied mathematics.
SIAM, Philadelphia, 1994.
10. R. Brandenberg and L. Roth. k-Containment. Manuscript.
11. R. Brandenberg and L. Roth. Minimal Containment under Homothetics – So-
lution Methods in Practice. Manuscript.
12. R. Brandenberg and T. Theobald. Algebraic methods for computing smallest
enclosing and circumscribing cylinders of simplices. Appl. Algebra Eng. Com-
mun. Comput., 14(6):439–460, 2004.
13. H. Brönnimann, H. Everett, S. Lazard, F. Sottile, and S. Whitesides. Transver-
sals to line segments in R3 . Discrete. Comput. Geom., 34(3):381–390, 2005.
14. M. Bădoiu, S. Har-Peled, and P. Indyk. Approximate clustering via core-sets.
In Proc. 34th Annu. ACM Sympos. Theory of Computing, pages 250–257. ACM
Press, 2002.
15. T.M. Chan. More planar two-center algorithms. Comput. Geom., 13(3):189–198,
1999.
16. T.M. Chan. Approximating the diameter, width, smallest enclosing cylinder,
and minimum-width annulus. In Proc. 16th Annu. Sympos. Comput. Geom.,
pages 300–309, 2000.
17. B.C. Eaves and R.M. Freund. Optimal scaling of balls and polyhedra. Math.
Prog., 23:138–147, 1981.
18. D. Eppstein. Faster construction of planar two-centers. In Proc. 8th Symp.
Discrete Algorithms, pages 131–138. ACM and SIAM, January 1997.
19. U. Faigle, W. Kern, and M. Streng. Note on the computational complexity of
j-radii of polytopes in Rn . Math. Prog., 73(1):1–5, 1996.
20. K. Fischer, B. Gärtner, and M. Kutz. Fast smallest-enclosing-ball computation
in high dimensions. In Proc. 11th Annu. European Symp. Algorithms, pages
630–641, 2003.
21. F. Follert. Maxmin location of an anchored ray in 3-space and related problems.
In Proc. 7th Canadian Conf. Comp. Geom., pages 7–12, 1995.
22. T. Gerken. On the Double-Ray Center Problem in 3-Space with an Applica-
tion to Surgical Operation Planning. Diplomarbeit, Zentrum Mathematik, TU
München, 2003.
23. A. Glozman, K. Kedem, and G. Shpitalnik. Computing a double-ray center for
a planar point set. J. Comput. Geom., 9(2):103–123, 1999.
24. P. Gritzmann and V. Klee. Inner and outer j-radii of convex bodies in finite-
dimensional normed spaces. Discrete Comput. Geom., 7:255–280, 1992.
25. P. Gritzmann and V. Klee. Computational complexity of inner and outer j-radii
of polytopes in finite-dimensional normed spaces. Math. Program., 59:163–213,
1993.
26. P. Gritzmann and V. Klee. On the complexity of some basic problems in com-
putational convexity I: Containment problems. Discrete Math., 136:129–174,
1994.
27. P. Gritzmann and T. Theobald. On algorithmic stabbing problems for polytopes.
Manuscript.
28. P. Gritzmann and T. Theobald. On the computational complexity of visibility
problems with moving viewpoints. In J.E. Goodman, J. Pach, and E. Welzl,
editors, Combinatorial and Computational Geometry, volume 52 of MSRI Pub-

lications, pages 377–397. Cambridge University Press, 2005.
29. L. Gürke, W. Strecker, and S. Martinoli. Korrektur mehrdimensionaler Deforma-
tionen durch eine einzige Osteotomie: graphische Analyse und Operationstech-
nik. Der Unfallchirurg, 102:684–690, 1999.
30. S. Har-Peled and K. Varadarajan. Projective clustering in high dimensions
using core-sets. In Proc. 18th Annu. Symp. Comput. Geom., pages 312–318.
ACM Press, 2002.
31. S. Har-Peled and K.R. Varadarajan. High-dimensional shape fitting in linear
time. Discrete Comput. Geom., 32:269–288, 2004.
32. D. Häublein. Entwicklung von Planungssoftware für die Extremitätenchirurgie.
Diplomarbeit, Fakultät für Informatik, TU München, 2007.
33. G. Ilizarov. Clinical application of the tension-stress effect for limb lengthening.
Cli. Orthop., 250:8–26, 1990.
34. G. Ilizarov. Transosseous osteosynthesis: Theoretical and clinical aspects of the
regeneration and growth of tissue. Springer Verlag, Berlin, Heidelberg, New
York, 1992.
35. M. Katchalski. Thin sets and common transversals. J. Geometry, 14(2):103–107,
1980.
36. P. Kumar. Clustering and reconstructing large data sets. PhD thesis, Depart-
ment of Computer Science, Stony Brook University, 2004.
37. P. Kumar, J.S.B. Mitchell, and E.A. Yıldırım. Approximate minimum enclosing
balls in high dimensions using core-sets. J. Exp. Algorithmics, 8, 2003.
38. Limb Lengthening Center Munich (ZEM), www.beinverlaengerung.de.
39. J. Maintz and M. Viergever. A survey of medical image registration. Medical
Image Analysis, 2(1):1–36, 1998.
40. N. Megiddo. On the complexity of some geometric problems in unbounded
dimension. J. Symbolic Comp., 10:327–334, 1990.
41. N. Megiddo and K.J. Supowit. On the complexity of some common geometric
location problems. Siam J. Comput., 10:182–196, 1990.
42. N. Megiddo and A. Tamir. On the complexity of locating linear facilities in the
plane. Oper. Res. Lett., 1:194–197, 1982.
43. Mercury Computer Systems, Inc., www.mc.com.
44. D. Paley. Principles of Deformity Correction. Springer, 2002.
45. S. Rittsteiger. Shape Fitting Algorithmen–Theorie, Implementation und Anwen-
dung in der chirurgischen Operationsplanung. Diplomarbeit, Zentrum Mathe-
matik, TU München, 2006.
46. L. Roth. Exakte und -approximative Algorithmen zur Umkugelberechnung.
Diplomarbeit, Zentrum Mathematik, TU München, 2005.
47. J.C. Russ. The Image Processing Handbook. CRC Press, 5th edition, 2006.
48. E. Schömer, J. Sellen, M. Teichmann, and C.-K. Yap. Smallest enclosing cylin-
ders. Algorithmica, 27:170–186, 2000.
49. The MathWorks, www.mathworks.com.
50. K.R. Varadarajan, S. Venkatesh, Y. Ye, and J. Zhang. Approximating the radii
of point sets. SIAM J. Comput., 36(6):1764–1776, 2007. Preliminary versions
in Proc. 43th IEEE Symp. Foundations of Computing, IEEE Comp. Soc., 2002,
561–569 and in APPROX 2003 + RANDOM 2003, LNCS, vol. 2764, Springer-
Verlag, 2003, 178–187.
51. H. Yu, P.K. Agarwal, R. Poreddy, and K.R. Varadarajan. Practical methods for
shape fitting and kinetic data structures using core sets. In Proc. 20th Annu.
ACM Sympos. Comput. Geom., pages 263–272, 2004.
52. G.L. Zhou, K.C. Toh, and J. Sun. Efficient algorithms for the smallest enclosing
ball problem. Comput. Optim. Appl., 30(2):147–160, 2005.
Image Segmentation for the Investigation
of Scattered-Light Images
when Laser-Optically Diagnosing
Rheumatoid Arthritis
Herbert Gajewski1 , Jens A. Griepentrog1 , Alexander Mielke1 ,

Jürgen Beuthan2 , Urszula Zabarylo2, and Olaf Minet2
1
Forschungsverbund Berlin e.V., Weierstraß-Institut für Angewandte Analysis
und Stochastik, Mohrenstraße 39, 10117 Berlin, Germany
{gajewski,griepent,mielke}@wias-berlin.de
2
Charité Universitätsmedizin Berlin, Campus Benjamin Franklin,
Institut für Medizinische Physik und Lasermedizin, Fabeckstraße 60–62,
14195 Berlin, Germany
{juergen.beuthan,urszula.zabarylo,olaf.minet}@charite.de
1 Introduction
With 1–2 % of the population affected, rheumatoid arthritis (RA) is the most
frequent inflammatory arthropathy. In most cases, RA initially affects the
small joints, only, especially the finger joints. The inflammations of the joints
caused by this disease usually start with a synovitis. At the same time, there
is a change in the filtration properties of the synovialis, which increases the
enzyme rate within the synovia thus accelerating the progress of inflammation.
In a later stage, granulation and neovascularisation occur in the synovia
(Figs. 1 and 2), which may finally lead to the destruction of cartilage and bone
structures [1]. So, it is rather unsurprising that the optical parameters [2, 3]
(Table 1) change in these early stages of the disease.
The examination of scattered-light images in laser-optical diagnostics
opens up new possibilities in medicine on the basis of tissue optics [4, 5].
For quantitative prognoses on the successful application of lasers for diag-
nostics and therapy in humans, one has to understand how light propagates
in biological tissues. Various scientific studies in this field used Monte-Carlo
simulations (MCS) for this purpose [6].
Any information gained through the light from the interior of the body
differs decisively from that acquired by radiological procedures like radiograms
or NMR. The light is strongly scattered and reflected on the boundaries in
almost any biological tissue which makes conventional projecting impossible.
Compared to optical imaging in geometrical optics, both shape and size of
150 H. Gajewski et al.
Fig. 1. Schematic of a small joint (left) healthy; (right) rheumatoid arthritis; com-
pare [1]
Fig. 2. Detail of a joint in an early stage of rheumatoid arthritis (according to

Fig. 1)
Table 1. Optical parameters (ex vivo) of healthy and rheumatic human finger joints
at 685 nm [3]; mean values of 14 samples
Tissue healthy finger joint rheumatic finger joint

μa (1/mm) μs (1/mm) μa (1/mm) μs (1/mm)
skin 0,02 1,95 0,02 1,95

bone 0,08 2,1 0,08 2,1
cartilage 0,17 1,8 0,17 1,8
articular capsule 0,15 0,6 0,24 1,1
synovia 0,004 0,006 0,011 0,012
the image details are decisively influenced by scattering and absorption in

the tissue. A medical application example is the excitation and detection of
fluorescent light in superficial tumours for an exact evaluation and resection
of the tumour at some safety margin for healthy tissue.
We investigate the finger joint screening by red laser light in order to
diagnose early rheumatic inflammations on finger joints. Thanks to their small
size, the joints are well suited for this non-invasive examination method [2, 7,
8, 9]. A first clinical study [10] showed that laser diaphanoscopy of finger joints,
Image Segmentation of Scattered-Light Images 151
if complemented by adapted image processing, may be a valuable additional

tool for follow-ups of inflamed joints. In this process dominant gray shades
are assigned to the significant image details (bones, cartilage and synovia)
which are interpreted as phases. Consequently, the scattered light appears as
a mixture of particles of different phases which are subjected to a non-local
separation with the aim of being segmented. On this basis a new algorithm
for image segmentation and reconstruction was developed [11, 12].
2 Optical Imaging
2.1 Tissue Optics
Any material is optically characterized by the absorption coefficient μa , the

scattering coefficient μs , and the anisotropy factor g (mean cosine of the scat-
tering angle) at different wavelengths. Due to the low absorption and high
scattering in the tissue, the light in the near infrared (NIR) can deeply pene-
trate into the tissue. The light propagation in homogeneous scattering media
is determined by the radiative transfer equation [13]:

dI(r, ŝ)
= −μt I(r, ŝ) + μs p(ŝ, ŝ )I(r, ŝ ) dΩ ,
ds S
where dΩ is the differential solid angle in ŝ direction, while S stands for the
unit sphere. The left term of the transfer equation determines the change rate
of the intensity at point r in ŝ direction. The right term describes the intensity
loss due to the total interaction μt = μa + μs and the intensity gained by
light scattering from all other directions in the direction ŝ, whereas μt merely
indicates the attenuation of the so-called ballistic photons of the incident
beam. The phase function p(ŝ, ŝ ) describes the scattering of the light, which
comes from the direction ŝ and is diverged to ŝ . The approximation of the
phase function often used in tissue optics was proposed in another context by
Henyey and Greenstein back in 1941 [14]:
1 1 − g2
pHG (cos θ) = .
2 (1 + g 2 − 2g cos θ)3/2
The anisotropy factor g is within the interval [−1, 1] and indicates how
strongly the scattering deviates from isotropic conditions (g = 0). In the case
of biological applications the strongly forward-directed scattering predomi-
nates, i.e. g > 0. In most practical problems the transport equation cannot
be analytically solved. The stochastic method of Monte-Carlo simulation has
been used since 1983 [15] to model the interaction between light and tissue.
Meanwhile, the MCS has become a standard method to calculate the laser
light distribution in tissue, with a large number of photon trajectories be-
ing calculated based on the probability density functions for scattering and
absorption; refer [16].
Fig. 3. Double integrating sphere for measuring the reflectance and transmittance
of a tissue layer
The light distribution inside the material, the diffuse reflectance as well
as the diffuse and the collimated transmittance of a substance can be cal-
culated from these parameters using the MCS. In order to determine these
microscopic coefficients, however, the inverse method is employed. The diffuse
and collimated intensities are measured for optically thin samples (Fig. 3) and
the parameters are adapted to the measured values using the MCS and the
gradient method [17]. The data for various kinds of tissue affected by RA are
listed in Table 1.
2.2 Experimental Set-Up
The experimental set-up for scattered-light measurements on finger joints is

shown in Fig. 4; refer [8]. A laser diode (Lasiris, 670 nm, max. 5 mW cw)
Fig. 4. Experimental set-up for measuring the scattered-light distribution on finger

joints [8]: The system consists of a laser diode, a camera and an ergonomically
adapted hand rest with finger holder
Fig. 5. Scattered-light distributions on a healthy (I ) and a rheumatoid (II ) finger

joint
screens the finger joint. The scattered-light image is taken by a CCD camera
(Hitachi KP-160 B/W) on the opposite side.
Figure 5 shows two examples for the gradient of the scattered-light intensi-
ties measured on a healthy finger joint (curve I) and a diseased one (curve II).
Scheel et al. [10] have shown that laser diaphanoscopy of finger joints may
be a valuable contribution to sensitive follow-ups of inflamed joints. For the
classification of one-dimensional scattered-light distributions a neuronal net
was used; refer Fig. 5.
The potential for medical diagnostics is, however, limited by the light
scatter in the skin. This scatter caused by the skin does not contain any use-
ful information and can be eliminated by deconvolution, which increases the
diagnostic value of this non-invasive optical procedure. For known optical pa-
rameters the kernel of the deconvolution operator can be exactly constructed
by means of the MCS [9].
3 Mathematical Algorithms for Image Segmentation

Contrary to segmentation procedures based on solving evolution equations
(refer [18, 19, 20]), our algorithm for non-local image segmentation is based
on a direct descent method for the free energy of the image, refer [11, 12].
We acquire the (already normalized) grayscale picture, the pixels of which
cover a domain Ω ⊂ Rn , by a measurable function c : Ω → [0, 1]. In order to
segment this picture c in respect of the given gray tones
a0 , . . . , am ∈ [0, 1], 0 = a0 < · · · < am = 1,
the following algorithm is used.

3.1 Decomposition into Phases

The aim is to decompose the picture c into several phases
u00 , . . . , u0m : Ω → [0, 1],
whereby the i-th component u0i , each, corresponds to the gray tone ai ∈ [0, 1]:

m
0 ≤ u00 , . . . , u0m ≤ 1, u0i = 1. (1)
i=0
For this purpose we choose a continuous partition ζ0 , . . . , ζm : [0, 1] → [0, 1]

of unity and weights b0 , . . . , bm ∈ R in such a way that

m
0 ≤ ζi ≤ 1, ζi (ai ) = 1, ζi = 1, (2)
i=0

m 1
0 < bi < 1, bi = 1, ζi (s) ds = bi (3)
i=0 0
hold for each i ∈ {0, . . . , m} and define the transformation

c → u0 = u00 , . . . , u0m = ζ0 (c), . . . , ζm (c) . (4)
Using the following partition of unity for concrete simulations, we choose
weights b0 , . . . , bm ∈ (0, 1) with the property

m
i−1
bj = 1, b∗i = bj ∈ (ai−1 , ai ) for all i ∈ {1, . . . , m},
j=0 j=0
and define exponents ωi > 0 and continuous functions hi : [ai−1 , ai ] → R by

ωi
ai − b∗i ai − s
ωi = ∗ , hi (s) = for s ∈ [ai−1 , ai ].
bi − ai−1 ai − ai−1
Then, we get a partition ζ0 , . . . , ζm : [0, 1] → [0, 1] of unity with the proper-
ties (2) and (3) by setting
h1 (s) if s ∈ [a0 , a1 ],
ζ0 (s) =
0 otherwise,
and
1 − hm (s) if s ∈ [am−1 , am ],
ζm (s) =
0 otherwise,
for the cases i = 0 and i = m as well as
⎧
⎪
⎨1 − hi (s) if s ∈ [ai−1 , ai ],
ζi (s) = hi+1 (s) if s ∈ [ai , ai+1 ],
⎪
⎩
0 otherwise,
in the case i ∈ {1, . . . , m − 1}.
3.2 Non-Local Phase Separation

The multi-component distribution u0 = u00 , . . . , u0m defined by (4) is sub-
jected to a non-local phase separation process, which minimizes the free en-
ergy F of the system under the constraint of mass conservation. For this
purpose, we developed a descent method delivering a sequence uk of inter-
mediate states and converging to an equilibrium state u∗ , refer [11].
Equilibrium distributions u∗ = (u∗0 , . . . , u∗m ) : Ω → [0, 1]m+1 of the multi-
component system are such states u = (u0 , . . . , um ), in which the free en-
ergy F (u) = Φ(u) + Ψ (u), defined as the sum of the segmentation entropy

m
Φ(u) = ui log(ui ) dx, (5)
Ω i=0
and the non-local interaction energy

1
m

Ψ (u) = ui Kij uj + ui − u0i Lij uj − u0j dx, (6)
2 Ω i,j=0
has a critical point under the constraint of mass conservation

u dx = u0 dx. (7)
Ω Ω
To describe the non-local interactions between the type i and j ∈ {0, . . . , m}

particles, we introduce compact linear operators Kij and Lij in (6), selecting
the maps Kij in such a way that particles of similar type drag on while parti-
cles of different type are repellent, which leads to the desired phase separation.
At the same time, the suitable selection of the operators Lij ensures that the
final state u∗ is kept close to the initial value u0 .
In our concrete applications the interaction operators Kij and Lij are
defined as solution maps to elliptic boundary value problems with Neumann
boundary conditions:
−#2ij Δ(Kij ψ) + Kij ψ = σij ψ in Ω, ν · ∇(Kij ψ)= 0 on ∂Ω, (8)

−rij
2
Δ(Lij ψ) + Lij ψ = sij ψ in Ω, ν · ∇(Lij ψ) = 0 on ∂Ω. (9)
Here, we give the respective effective ranges #ij , rij > 0, and intensities σij ,
sij ∈ R of the interactions between the type i and j ∈ {0, . . . , m} parti-
cles, with both matrices being assumed to be symmetric. The cases σij > 0
and σij < 0, respectively, correlate with the repellent and dragging interac-
tions.
We minimize the free energy F = Φ + Ψ of the multi-component system
under the constraint (7) by solving the appropriate Euler–Lagrange equations,
with the following descent method being employed: Assuming τ ∈ (0, 1] is
a suitably chosen relaxation parameter and u0 is the initial distribution and
knowing the old state uk , the intermediate state v k and the corresponding
Lagrange multiplier λk ∈ R are uniquely determined by solving the auxiliary
problem

λk = DΦ v k + DΨ uk , v k dx = u0 dx, (10)
Ω Ω
with a strongly monotone operator. As the new state, we define
uk+1 = τ v k + (1 − τ )uk . (11)
Our analytical
investigations
in [11] have shown that for k → ∞ both the
sequence uk , λk converges to a solution (u∗ , λ∗ ) of the Euler–Lagrange equa-
tions
λ∗ = DΦ(u∗ ) + DΨ (u∗ ), u∗ dx = u0 dx, (12)
Ω Ω

and the sequence F uk of the free energies converges monotonically de-
creasing to the limit value F (u∗ ).
3.3 Image Composition of the Segmented Phases
In the last step we calculate the segmented image c∗ : Ω → [0, 1] as convex

combination of the gray values a0 , . . . , am ∈ [0, 1] in terms of the weight
functions u∗0 , . . . , u∗m :
m
u∗ → c∗ = u∗i ai .
i=0
3.4 Numerical Implementation
Basing on our analytical results we are in a position to carry out numerical

simulations, which consequently implement right into the sphere of machine
accuracy the descent method (5)–(11) earlier described herein, in particular
its analytical convergence. For this purpose, we have developed a stand-alone,
extendable and portable programme for the segmentation of gray tone pictures
in respect of an arbitrary number of gray tones. Our programme is based on
two widely used libraries, the sources of which are freely accessible. It can be
used as a stand-alone programme, but also be integrated into existing image
processing programmes. Data transfer, input and output processes are orga-
nized using the netpbm library for image files in portable gray map format.
The descent method parameters can be imported through a separate control
and configuration file. The rectangular geometry of the image file permits the
discrete Fourier transformation to be used for effectively solving the linear
elliptic boundary value problems (8) and (9) with the library fftw being uti-
lized for that purpose. Due to the strong monotonicity of the related operator
in the auxiliary problem (10) the Lagrange multipliers can be quickly and
precisely calculated by means of the Newton method.
4 Simulation Results
Our image segmentation algorithm is simulated based on the reconstruction
of a grainy test pattern as well as on the segmentation and evaluation of
scattered-light images of finger joints, with two-dimensional images being rep-
resented naturally as rectangular domains Ω ⊂ R2 . The interaction range is
given in the unit of length related to the problem, i.e., related to the edge
length of a square pixel. Generally speaking, our method can also be used
for the segmentation of image data which are available in multi-dimensional
domains Ω ⊂ Rn .
4.1 Reconstruction of a Grainy Test Pattern
Subjecting our new method to a performance test, we start with reconstruct-

ing an artificially generated grainy test pattern in respect of three gray val-
ues (m = 2)
49 39 22 39
a0 = 0, a1 = , a2 = 1, of the weights b0 = , b1 = , b2 =
100 100 100 100
and the parameters
⎛ ⎞ ⎛ ⎞
−10 +10 +10 +12 −12 −12
#ij = rij = 2, (σij ) = ⎝+10 −10 +10⎠ , (sij ) = ⎝−12 +12 −12⎠
+10 +10 −10 −12 −12 +12
according to (6), (8) and (9). Due to the selected matrix (σij ) particles of dif-
ferent type are strongly repellent while particles of similar type are dragging
on at equal force. At the same time, the inverted sign structure of the ma-
trix (sij ) keeps the reconstructed image close to the initial value u0 . Figure 6
shows the numerical results of the grainy test pattern with 200 × 200 pixels.
A satisfactory result is delivered, of course, not before all three components
will have been reconstructed.
Fig. 6a–c. Image reconstruction: (a) Initial value (grainy pattern); (b) reconstruc-
tion result referring to the components black and white (gray region remains grainy);
(c) reconstruction result referring to all three components, i.e. black, gray and white
4.2 Segmentation and Evaluation

of Medical Scattered-Light Images
In our joint project we use the non-local image segmentation method for the
investigation of scattered-light images for laser-optically diagnosing rheuma-
toid arthritis. This disease is the most frequent inflammatory arthropathy
in humans. Initially, it typically affects the small joints, especially the fin-
ger joints. The disease starts with an inflammation of the capsular ligaments
as a consequence of which the synovia opacifies. With the disease progress-
ing granulation and neovascularisation develop in the capsular ligaments, and
both cartilage and bone structures are destroyed.
We demonstrate our new method exemplarily by means of two finger joints
of the right hand of a patient suffering from rheumatoid arthritis. So doing,
we took one each scattered image of 385 × 209 pixels, of both joints at two
times at an interval of circa six months. Figures 7 und 8 present not only the
originals, but also the results of the image segmentation referring to the four
gray values (m = 3) black, dark gray, light gray and white,
1 2 1
a0 = 0, a1 = , a2 = , a3 = 1, the weights b0 = b1 = b2 = b3 =
3 3 4
and the parameters
⎛ ⎞ ⎛ ⎞
−4 +4 +4 +4 +4 −4 −4 −4
⎜+4 −4 +4 +4⎟ ⎜−4 +4 −4 −4⎟
#ij = rij = 4, (σij ) = ⎜ ⎟
⎝+4 +4 −4 +4⎠ , (sij ) = ⎜
⎝−4 −4 +4 −4⎠
⎟
+4 +4 +4 −4 −4 −4 −4 +4
according to (6), (8) and (9). In both cases, the progress of the disease is
already clearly visible by the increasing opacity of the synovia. To offer the
rheumatologist an additional diagnostic option, our method calculates more-
over a relative distance of the original scattered-light image u0 to the result u∗
of the respective image segmentation by means of squared expressions like

m
∗ m
∗
u − u0 2 dx or ui − u0i Lij u∗j − u0j dx,
i i
Ω i=0 Ω i,j=0
refer (6). This distance gets larger for both the two finger joints in the course
of time which we interpret as a progress of the disease. This is based on
the observation that the high-definition and high-contrast image of a healthy
joint seems to change less during image segmentation than the obliterated
low-contrast image of a diseased joint.
Our figures merely show the results of two selected finger joints of the
diseased patient. If all the investigated finger joints are considered, our assess-
ment of the progress of the disease corresponds with the radiologist’s diagnosis
to 70–80 %.
Fig. 7a–d. Progressing rheumatoid arthritis of a finger joint on the index of the
patient’s right hand: (a) Scattered-light image of an earlier date; (b) respective
image segmentation result; (c) scattered-light image of a later date; (d) respective
image segmentation result
Fig. 8a–d. Progressing rheumatoid arthritis of a finger joint on the little finger of
the patient’s right hand: (a) Scattered-light image of an earlier date; (b) respective
image segmentation result; (c) scattered-light image of a later date; (d) respective
image segmentation result
Acknowledgements
We would like to express our thanks to the Federal Ministry of Education and
Research (BMBF) for supporting our joint project (grant no. FKZ 03BENGB6
and 03GANGB5). In addition, our thanks are expressed to Dr. Scheel, of
Georg-August-Universität Göttingen, Medizinische Klinik für Nephrologie
und Rheumatologie, for the anonymized patient images provided to us.
References
1. Steinbrocker, O., Traeger, C.H., Batermann, R.C.: Therapeutic criteria in
rheumatoid arthritis. J. Amer. Med. Ass. 140, 659–662 (1949).
2. Beuthan, J., Minet, O., Müller, G., Prapavat, V.: IR-Diaphanoscopy in
Medicine. SPIE Institute Series 11, 263–282 (1993).
3. Prapavat, V., Runge, W., Mans, J., et al.: Development of a finger joint phan-
tom for the optical simulation of early stages of rheumatoid arthritis. Biomed.
Tech. 42, 319–326 (1997).
4. Tuchin, V.: Tissue Optics: Light Scattering Methods and Instruments for Medical
Diagnosis. Bellingham: SPIE Press, 2000.
5. Minet, O., Beuthan, J.: Investigating laser-induced fluorescence of metabolic
changes in Guinea pig livers during Nd:YAG laser irradiation. Las. Phys. Lett. 2,
39–42 (2005).
6. Minet, O., Dörschel, K., Müller, G.: Lasers in Biology and Medicine. In:
Poprawe, R., Weber, H., Herziger, G. (eds.): Laser Applications. Landolt–
Börnstein Vol. VIII/1c, 279–310. Heidelberg, New York: Springer, 2004.
7. Beuthan, J., Prapavat, V., Naber, R., Minet, O., Müller, G.: Diagnostic of In-
flammatory Rheumatic Diseases with Photon Density Waves. Proc. SPIE 2676,
43–53 (1996).
8. Beuthan, J., Netz, U., Minet, O., Klose, A., Hielscher, A., Scheel, A., Hen-
niger, J., Müller, G.: Light scattering study of rheumatoid arthritis. Quantum
Electronics 32, 945–952 (2002).
9. Minet, O., Zabarylo, U., Beuthan, J.: Deconvolution of laser based images for
monitoring rheumatoid arthritis. Las. Phys. Lett. 2, 556–563 (2005).
10. Scheel, A.K., Krause, A., Mesecke-von Rheinbaben, I., Metzger, G., Rost, H.,
Tresp, V., Mayer, P., Reuss-Borst, M., Müller, G.A.: Assessment of proximal
finger joint inflammation in patients with rheumatoid arthritis using a novel
laser-based imaging technique. Arthritis Rheum. 46, 1177–1184 (2002).
11. Gajewski, H., Griepentrog, J.A.: A descent method for the free energy of mul-
ticomponent systems. Discrete Contin. Dynam. Systems 15, 505–528 (2006).
12. Minet, O., Gajewski, H., Griepentrog, J.A., Beuthan, J.: The analysis of laser
light scattering during rheumatoid arthritis by image segmentation. Las. Phys.
Lett. 4, 604–610 (2007).
13. Ishimaru, A.: Wave propagation and scattering in random media. New York:
Academic Press, 1978.
14. Henyey, L.G., Greenstein, J.L.: Diffuse radiation in the galaxy. Astrophys. J. 93,
70–83 (1942).
15. Wilson, B.C., Adam, G.: Monte Carlo model for the absorption and flux distri-
butions of light in tissue. Med. Phys. 10, 824–830 (1983).
16. Henniger, J., Minet, O., Dang, H.T., Beuthan, J.: Monte Carlo Simulations in
Complex Geometries: Modeling Laser Transport in Real Anatomy of Rheuma-
toid Arthritis. Las. Phys. 13, 796–803 (2003).
17. Roggan, A., Minet, O., Schröder, C., Müller, G.: Measurements of optical tissue
properties using integrating sphere technique. SPIE Institute Series 11, 149–165
(1993).
18. Morel, J.-M., Solimini, S.: Variational Methods in Image Segmentation. Boston,
Basel, Berlin: Birkhäuser, 1994.
19. Gajewski, H., Gärtner, K.: On a nonlocal model of image segmentation. Z.
Angew. Math. Phys. 56, 572–591 (2005).
20. Broser, P.J., Schulte, R., Lang, S., Roth, A., Helmchen, F., Waters, J.,
Sakmann, B., Wittum, G.: Nonlinear anisotropic diffusion filtering of three-
dimensional image data from two-photon microscopy. J. Biomed. Optics 9, 1253–
1264 (2004).
Part V
Transport, Traffic, Energy

Dynamic Routing of Automated Guided
Vehicles in Real-Time
Ewgenij Gawrilow1 , Ekkehard Köhler2 , Rolf H. Möhring1 , and

Björn Stenzel1
1
Technische Universität Berlin, Institut für Mathematik, MA 6–1,
Straße des 17. Juni 136, 10623 Berlin, Germany
{gawrilow,moehring,stenzel}@math.tu-berlin.de
2
Brandenburgische Technische Universität Cottbus, Institut für Mathematik,
Postfach 10 13 44, 03013 Cottbus, Germany
ekoehler@math.tu-cottbus.de
Summary. Automated Guided Vehicles (AGVs) are state-of-the-art technology for

optimizing large scale production systems and are used in a wide range of applica-
tion areas. A standard task in this context is to find efficient routing schemes, i.e.,
algorithms that route these vehicles through the particular environment. The pro-
ductivity of the AGVs is highly dependent on the used routing scheme.
In this work we study a particular routing algorithm for AGVs in an automated
logistic system. For the evaluation of our algorithm we focus on Container Terminal
Altenwerder (CTA) at Hamburg Harbor. However, our model is appropriate for an
arbitrary graph. The key feature of this algorithm is that it avoids collisions, dead-
locks and livelocks already at the time of route computation (conflict-free routing),
whereas standard approaches deal with these problems only at the execution time
of the routes. In addition, the algorithm considers physical properties of the AGVs
and certain safety aspects implied by the particular application.
1 Introduction
Automation of large scale logistic systems is an important method for im-
proving productivity. Often, in such automated logistic systems Automated
Guided Vehicles (AGVs) are used for transportation tasks. Especially, so
called free-ranging AGVs are more and more used since they add a high
flexibility to the system. The control of these AGVs is the key to an efficient
transportation system that aims at maximizing its throughput.
In this work we focus on the problem of routing AGVs. This means we
study how to compute good routes on the one hand and how to avoid collisions
on the other hand. Note that dispatching of AGVs, i.e., the assignment of
transportation tasks to AGVs, is not part of the routing and therefore not
considered in this paper.
166 E. Gawrilow et al.
Fig. 1. The HHLA Container Terminal Altenwerder (CTA). HHLA

c
Our application is the Container Terminal Altenwerder (see Fig. 1), which
is operated by our industrial partner, the Hamburger Hafen und Logistik AG
(HHLA).
We represent the AGV network by a particular grid-like graph that consists
of roughly 10,000 arcs. and models the underlying street network of a traffic
system consisting of a fleet of AGVs. The task of the AGVs is to transport
containers between large container bridges for loading and unloading ships
and a number of container storage areas. The AGVs navigate through the
harbor area using a transponder system and the routes are sent to them from
a central control unit. AGVs are symmetric, i.e., they can travel in both
of the two driving directions equally well and can also change directions on
a route.
Previous Work
First ideas for free-ranging AGV systems were introduced by Broadbent
et al. [2]. Since then, several papers concerning this topic have been pub-
lished [18]. In this paper we focus on routing approaches in the case where
dispatching of AGVs is already made.
In so-called offline approaches all requests (transportation tasks) are
known right from the beginning. Krishnamatury, Batta and Karwan [11] as
well as Qui and Hsu [13] discuss the AGV routing problem in this case. While
Krishnamatury, Batta and Karwan present a heuristic solution for general
graphs (where this routing problem is N P-hard [16]), Qui and Hsu consider
a very simple graph and present a polynomial time algorithm.
Dynamic Routing of Automated Guided Vehicles in Real-Time 167
In contrast, online approaches assume that requests appear sequentially.

The local approach by Taghaboni-Dutta and Tanchoco [17] is such an on-
line method. Here, a decentralized algorithm decides about the routes, based
only on local information. In particular, it does not determine the whole path
for an AGV, but iteratively computes sub-paths from checkpoint to check-
point.
The static approach uses the full information about the already routed
AGVs. Algorithms that are based on these approaches usually compute ge-
ographically shortest paths with optional additional penalty costs on con-
gested arcs from the source to the destination of the current request [12].
This static formulation needs an additional collision avoidance system to make
the routes collision-free, since time dependences are not represented in that
model.
Our Contribution
In this work we study a dynamic online AGV routing model for an arbitrary
graph. This approach is motivated by dynamic flow theory (see [6, 7, 10])
and several papers on the Shortest Path Problem with Time-Windows [3,
4, 5, 14]. The main advantage of our model and algorithm over the known
online methods is that the time-dependent behavior of AGVs is fully modeled,
such that both conflicts and deadlock situations can be prevented already at
the time of route computation. The newly designed model is not only very
accurate in the mapping of properties of the actual application but, as we
show in our computational experiments, it is also well suited for being used
in a real-world production system.
The paper is organized as follows. In Sect. 2 we describe how we model the
AGV network. In Sect. 3 we introduce our algorithm and show that it runs
in polynomial time. Section 4 explains how subtle technical characteristics of
the particular application can be represented in our model to get it ready for
being used in practice. The computational results are presented in Sect. 5.
2 The Model
We model the automated transportation system by a directed graph G rep-
resenting the feasible lanes of the system. These lanes are given by certain
transponder positions. In the application, this graph has about 10,000 arcs.
Initially, we assume that every arc a has a fixed, constant transit time
τ (a).
Transportation tasks are consecutively arriving over time and are modeled
by a sequence σ = r1 , . . . , rn of requests. Each request rj = (sj , tj , θj ) consists
of a start node sj , an end node tj , and a desired starting time θj . The aim is
to minimize the overall transit times, that is the sum of transit times over all
requests. We approach this goal by iteratively computing a shortest path for

each request, which is a natural method in this online setting.
The physical properties of the AGVs demand for a variety of special fea-
tures of the model. Although each route of an AGV can be represented in
the given graph, not every route in this graph can in fact be conducted by
an AGV. The reason for this difficulty is the complicated turning behavior,
which makes it necessary to start turning the wheel already long before the
particular intersection is reached. As a consequence, an AGV needs a suffi-
ciently long straight route segment between two consecutive curves. To cope
with this rather complicated turning behavior we introduce in a preprocess-
ing step a set of artificial arcs to the network, each representing a possible
turn (see Fig. 2). In addition, at each node of the graph we introduce turning
rules defining which out-going arcs of a node can be used from a particular
in-going arc. As a result we get a much larger network (about 45,000 arcs)
that captures all possible movements of an AGV — each feasible route in this
network can be executed by an AGV.
Another complicating property is the size of the AGVs compared to the
rather closely meshed network of lanes. If an AGV traverses or stands on an
arc a, it affects a much larger portion of the network than only arc a, which
is then blocked for other AGVs (see Fig. 3). In Sect. 2.2, we describe how we
take this into account in our conflict-free (dynamic) approach.
Fig. 2. The figure illustrates an artifical arc (blue dotted arrow ) that models a curve.
This is done for all permitted curves
Fig. 3. The figure illustrates the polygons that are claimed by an AGV that moves
in the indicated direction. Polygon A, B and D pairwise intersect each other while
polygons C and E do not intersect any of the others, respectively
2.1 Static Routing and its Drawbacks
A standard approach for routing of AGVs in an online setting is the so-called

static routing. In this case one only computes static routes in the network,
ignoring their time dependent nature. More precisely, one computes a stan-
dard shortest path, e.g., using Dijkstra’s algorithm, with respect to arc costs
consisting of the transit times τa plus a load dependent penalty cost which is
a function of the number of routes that are already using this arc (see [12]).
The computed routes are, of course, not collision-free. Hence, one needs an
additional conflict avoidance system that, at execution time of the routes,
guarantees that there are no collisions. One way to do this is to iteratively
allocate to an AGV the next part of its route (the “claim”) and block it for
all other AGVs (“claiming”).
The advantage of the static approach is clear. It is easy to implement and
allows very fast route computation. However, various drawbacks are caused by
the collision avoidance at execution time. In particular, the claiming rules can
cause deadlocks and have a deteriorating effect on the system performance.
Deadlocks (see Fig. 4) appear if a group of AGVs wish to claim an area
which is already occupied by another AGV in this group. None of them is able
to continue its route and thus the system is blocked. Algorithmic solutions for
that problem are only suitable for a very small number of AGVs [9, 12].
In addition to deadlocks, a variety of other drawbacks like detours and high
congestion occur in the static setting; again, since time-dependent behavior is
not considered. This results in traveling times that can be far away from the
shortest possible traveling time. Moreover, actual arrival times of the AGVs
at their destinations are completely random and cannot be predicted at the
time of route computation. This is a major drawback for other planning steps
in the logistic chain that depend on the knowledge of these arrival times.
Fig. 4. Simplified deadlock situation. Both AGVs are trying to occupy the same
arc of the network, thereby blocking each other
2.2 Dynamic Routing of AGVs
In order to avoid the problems of the simple model given in Sect. 2.1, we follow
a completely different approach that computes shortest (w.r.t. traveling time)
and conflict-free routes simultaneously.
There are two key ingredients which must be considered in our approach.
On the one hand, one has to deal with the physical dimensions of the AGVs
because they usually have to claim several arcs in the directed graph at the
same time. On the other hand, the approach has to be time-dependent (dy-
namic).
Every arc can be seen as a set of time intervals, each representing a different
AGV that is routed over this arc or, at least, blocks this arc during some time
interval. Note that these intervals have to be mutually disjoint since an overlap
would mean that the corresponding AGVs collide on this arc at the time of
the overlap. In fact, in our algorithm, we will not maintain the set of intervals
in which an arc is blocked, but the complementary set of free time-intervals
(time-windows).
Maintaining these sets of intervals may be seen as a compact representa-
tion of the standard time-expanded graph, in which there is a copy of each
vertex/arc for each point in time (with respect to some time discretization).
In contrast, the set of time-windows of an arc a only models those times,
in which there actually is no AGV on a. Similar compact representations
of a time-expanded graph by time intervals have been studied before, see
e.g. [3, 4, 5, 14] and Sect. 3.2.
For dealing with the physical dimensions of the AGVs we use polygons
P (a) for each arc a, which describe the blocked area when an AGV (the
center of an AGV) is located on arc a (Fig. 3). Thus, it is prohibited to use
two arcs at the same time if the corresponding polygons intersect. For each
arc a, this leads to a set confl(a) of so called geographically dependent arcs
which must not be used at the same time. If an AGV travels along an arc a
during the interval [θ1 , θ2 ], all geographically dependent arcs are blocked from
θ1 to θ2 . Note that in this approach there is no need to model traveling on
nodes since each edges contains its end nodes.
After routing a request one has to readjust the time-windows according
to the arc usage of the newly found route and their geographically dependent
arcs. Note that this implies that one does not have to take care of the physical
dimensions of the AGVs during route computation, since it is already fully
represented by readjusting the time-windows on all affected arcs.
As mentioned before, the advantage of this approach is the fact that the
problems of Sect. 2.1 are avoided because in a conflict-free approach there
is no need for an additional collision avoidance since the routes are planned
conflict-free in advance.
Additionally, as a welcome side effect, the completion time of a request is
known immediately after route computation since the time-dependent behav-
ior is fully modeled. This is a great advantage for a higher-level management
system which plans the requests.
3 The Algorithm
The algorithm consists of two parts. The first part is a preprocessing step;
during the second part all requests are routed iteratively in a real-time route
Fig. 5a–c. Illustration of the real-time computation on three consecutive arcs with
transit time 1. (a) shows the situation before the new request arrives. There is
a graph with some blockings (red ) and some time-windows (green) on the time axis
(y axis). The task is to compute a quickest path that respects the time-windows.
This is illustrated in (b). The chosen path is blocked afterwards (see (c))
computation and for each computed route the time-windows of the affected
arcs are adjusted.
The structure of the real-time computation (route computation and read-
justment of time-windows) is illustrated in Fig. 5.
3.1 Preprocessing
The preprocessing step determines the conflict sets and the turning rules for
each arc a. First, all polygons P (a) (see Sect. 2.2) are compared pairwise. If
the polygons P (a) and P (b) of arcs a, b ∈ A(G) intersect, then a is added to
confl(b) and b is added to confl(a). Second, one computes for each arc a a list
OU T (a) of arcs containing those arcs b that are permitted to be used after
arc a on a feasible route respecting the physical properties of an AGV. This
is done only once for a given layout (harbor).
3.2 Route Computation: Quickest Paths with Time-Windows
As pointed out in Sect. 2.2, the route computation can be done in an ide-
alized model where the dimension of the AGV need no longer to be consid-
ered, since the conflict sets take care of this. Instead, one just has to com-
pute a route for an infinitesimal mass point representing the center of the
AGV.
This simplified problem is related to the Shortest Path Problem with Time-
Windows (SPPTW) [3, 4, 5, 14] and can be formulated as follows: Given
a graph G, a source node s, a destination node t, a start time θ, tran-
sit times τ (a), costs c(a) and a set of time-windows F (a) on each arc a;
compute a shortest path (w.r.t. arc costs c(a)) that respects the given time-
windows.
Since AGVs are allowed to stop during their route, waiting is allowed on
such a path. ‘Respecting’ the time-windows means that AGVs wait on an
arc or traverse an arc a only during one of its “free” time-windows given by
F (a).
The SPPTW and also our variant is N P-hard. The hardness can be shown
by reduction of the Constrained Shortest Path Problem (CSPP [1]).3
Our algorithm for this problem is a generalized arc-based label setting
algorithm resembling Dijkstra’s algorithm. A label L = (aL , cL , IL , predL )
on an arc aL consists of a cost value cL , a predecessor predL and a time
interval IL . Each label L represents a path from start node s to the tail
of aL , whereas cL contains the cost value of the path up to the tail of aL ;
the label interval IL = (AL , BL ) represents an interval of possible arrival
times at arc aL (at the tail of aL ); predL is the predecessor of aL on that
path.
We define an ordering for these labels. We say that a label L dominates
a label L if and only if
cL ≤ cL and IL ⊆ IL .
The labels are stored in a priority queue H, e.g., a binary heap. The
generalized arc-based Dijkstra algorithm works as follows.
• Initialization
Create a label L = (a, 0, (θ, ∞), nil) for all out-going arcs a of s and add
them to the priority queue H.
• Loop
Take the label L with lowest cost value cL from H. If there is no label left
in the queue, output the information that there is no feasible path from s
to t. If t is the tail of aL , output the corresponding path.
– For each time-window on arc aL .
• Label Expansion
Try to expand the label interval along aL through the time-window
of the arc aL (new label interval should be as large as possible, see
Fig. 6), add the costs c(aL ) to the cost value cL and determine the
new predecessor. If there is no possible expansion, consider the next
time-window of arc aL .
• Dominance Test
For each out-going arc a in OU T (aL ), add the new label to the heap
if it is not dominated by any other label on a. Delete the labels in
the heap that are dominated by the new label.
Since the SPPTW is N P-hard the algorithm cannot be executed in poly-
nomial time, unless P = N P. However, the AGV routing problem differs
from the SPPTW in a subtle point. In AGV routing, the cost of a path
is the sum of the transit times of the arcs on the path plus waiting times
on arcs which is the crucial property that makes the problem polynomial.
3
The instance of the SPPTW is constructed by placing time-windows [0, R] at each
arc while R denotes the resource constraint in the CSPP instance.
Fig. 6a–d. Label Expansion on three consecutive arcs. The label intervals are rep-
resented by blue bars and are placed above the nodes. The blockings are colored
red (arcs). The green intervals between these blockings are the time-windows. The
figures (a) to (d) show the successive expansion of the label intervals
Thus, the costs on an arc a (the transit time of a plus possible waiting
time on a) depend no longer only on the arcs itself, but also on the rout-
ing history given by the label interval and the current time-window. We
call the resulting problem the Quickest Path Problem with Time-Windows
(QPPTW).
For the QPPTW we can obtain a polynomial time algorithm, since here
the costs correlate to the lower bounds of the label intervals.
Theorem 1. The described generalized arc-based Dijkstra algorithm solves the

QPPTW in polynomial time (in the number of time-windows).
Proof. The algorithm computes all required paths since the expansion of the
label intervals is maximal and no optimal path (label) will be dominated.
Therefore, on termination the algorithm has computed an optimal path re-
specting the time-windows. That it terminates follows from the complexity
analysis given below.
Consider the correlation between costs and traveling time (including wait-
ing times): they differ only by an additive constant, namely the starting time.
Thus, for any two labels that are expanded w.r.t. the same time-window the
cost value controls the dominance relation. Hence, one label dominates an-
other if and only if it has a lower transit time.
Therefore, the number of possible labels on an arc a is bounded by the
number of time-windows on all in-going arcs (in-going time-windows F − (a)).
As a consequence, the number of iterations in each loop (the number of labels
taken from the priority queue) is bounded from above by the product of the
number% of arcs and the maximum number of in-going time-windows over all
arcs ( a∈A |F − (a)|). In each iteration a label is expanded along at most the
number of time-windows
% |F(a)| at a and each of the resulting labels is com-
pared with at most b∈OUT (a) |F − (b)| existing labels. If the priority queue is
%
implemented as a heap, updating can be done in O(log( a∈A |F − (a)|)). This
leads to the following run time:
0 !„ 0 8 91 !1
X « < X = X
O@ |F − (a)| · max |Fa | · @max |F − (b)| A ·log |F − (a)| A .
a∈A a∈A : ;
a∈A b∈OU T (a) a∈A
Hence, the algorithm terminates in polynomial time with an optimal path or

the notification that there is no feasible path at all.

For an additional acceleration of the algorithm we use goal-oriented

search [8, 15].
After each route computation we verify for each arc a of the computed
path, whether the time-windows on the arcs in confl(a) have to be readjusted.
4 Additional Practical Requirements

To make the algorithm practical, additional ingredients have to be taken into
account.
4.1 Container Orientation
Since containers are not completely symmetric, it may be necessary to give an

AGV an explicit target orientation. We model this orientation constraint by
a flag, indicating whether the AGV is in the right driving direction to reach
the target in the correct orientation without a turn.
To this end, we maintain labels at an arc for each of the two possible
directions and define the domination rule accordingly. Using the observations
of Theorem 1 we get the following result.
Theorem 2. The described generalized arc-based Dijkstra algorithm solves the

QPPTW in polynomial time (in the number of time-windows) even if the
orientation of AGVs (containers) is taken into consideration.
4.2 Safety Tubes and Re-Routing
In spite of the fact that the computed routes are conflict-free, additional safety
is required in practice because the AGVs possibly deviate from the computed
routes in time. Also technical problems may occur while traveling through
the network. We have implemented two different safety tubes, a distance-
dependent and a time-dependent one, and re-routing techniques to cope with
this difficulty.
The distance-dependent tube blocks an area in front of the AGV. The
length depends on the speed of the AGV and is at least the distance needed
to come to a complete stop (braking distance). This allows the AGV to stop
if something unexpected happens (for example an unexpected stop of another
AGV) without causing a collision.
The time-dependent tube allows a little deviation from the computed time,
i.e., the expected arrival time at a specific point. This is necessary because
there will always be small differences between computed times in the model
and the times when the AGVs reaches a point in reality.
In order to cope with more challenging perturbations as large deviations
from the expected starting time, lower driving speeds than expected or vehicle
breakdowns we also implemented re-routing strategies based on the described
algorithm.
4.3 Non-Constant Transit Times
Instead of constant transit times as described in Sect. 2, we take the variable

speed of the AGVs into account, i.e., we model the acceleration behavior and
the different possible maximum speeds.
The maximum speed depends on the kind of movement (curve or straight
section), the weather conditions, and the status of the AGV. The acceleration
value depends on the current speed.
5 Computational Results
We now address two important questions with our approach.
• Is the approach better than the static one?
• Is the algorithm suitable for real-time computation?
Both questions can be answered in the affirmative. The comparison of both
approaches shows that the conflict-free approach is superior to the static one
(exact numbers at CTA have to be kept confidential). Additionally, the pre-
sented algorithm is able to provide fast answers. On average, the computation
in all scenarios does not require more than a few hundredth of a second. And
also the maximum values of less than half a second are small enough to ensure
fast real-time computation in practice.
Acknowledgements
We are grateful to Andreas Parra and Kai-Uwe Riedemann from the Ham-
burger Hafen und Logistik AG (HHLA) and Boris Wulff from Container Ter-
minal Altenwerder (CTA) for their valuable remarks and suggestions and for
providing us with real world data.
References
1. Beasley, J. E., Christofides, N. (1989) An algorithm for the resource constrained
shortest path problem. Networks 19, 379–394
2. Broadbent, A. J. et al. (1987) Free-ranging AGV and Scheduling System. In
Automated Guided Vehicle Systems, 301–309
3. Desrosiers et al. (1986) Methods for routing with time windows. European Jour-
nal of Operations Research 23, 236–245
4. Desrosiers, M., Soumis, F. (1988) A generalized permanent labelling algorithm
for the Shortest Path Problem with Time Windows, INFOR 26(3), 191–212
5. Desrosiers, J., Solomon, M. (1988) Time window constrained routing and
scheduling problems, Transportation Science 22, 1–13
6. Ford, L. R., Fulkerson, D. R. (1959) Constructing maximal dynamic flows from
static flows. Operations Research 6, 419–433
7. Ford, L. R., Fulkerson, D. R. (1962) Flows in Networks. Princeton University
Press, Princeton, NJ
8. Hart, P., Nilsson, N., Raphael, B. (1968) A formal basis for the heuristic de-
termination of minimum cost paths. In IEEE Transactions on Systems, Science
and Cybernetics SCC-4, 100–107
9. Kim, K. H., Jeon, S. M., Ryu, K. R. (2006) Deadlock prevention for automated
guided vehicles in automated container terminals. OR Spectrum 28 (4), 659–679
10. Köhler, E., Möhring, R. H., Skutella, M. (2002) Traffic networks and
flows over time. In Jürg Kramer, Special Volume Dedicated to the
DFG Research Center “Mathematics for Key Technologies Berlin”, pub-
lished by Berliner Mathematische Gesellschaft, 49–70, http://www.math.tu-
berlin.de/coga/publications/techreports/2002/Report-752-2002.html
11. Krishnamurthy, N., Batta, R., Karwan, M. (1993) Developing conflict-free routes
for automated guided vehicles. Operations Research 41, 1077–1090
12. Moorthy, K. M. R. L., Guan, W. H. (2000) Deadlock Prediction and Avoidance
in an AGV System. SMA Thesis
13. Qui, J., Hsu, W.-J. (2000) Conflict-free AGV routing in a bi-directional path
layout. In Proceedings of the 5th International Conference on Computer Inte-
grated Manufacturing (ICCIM 2000), volume 1, 392–403
14. Sancho, N. G. F. (1994) Shortest path problems with time windows on nodes
and arcs. Journal of mathematical analysis and applications 186, 643–648
15. Sedgewick, R., Vitter, J.S. (1986) Shortest paths in Euclidian graphs. Algorith-
mica 1, 31–48
16. Spenke, I. (2006) Complexity and Approximation of Static k-Splittable Flows
and Dynamic Grid Flows. PhD Thesis Technische Universität Berlin
17. Taghaboni-Dutta, F., Tanchoco, J. M. A. (1995) Comparison of dynamic route-

ing techniques for automated guided vehicle system. International Journal of
Production Research 33(10), 2653–2669
18. Vis, I. F. A. (2006) Survey of research in the design and control of automated
guided vehicle systems. European Journal of Operational Research 170, 677–709
Optimization of Signalized Traffic Networks
Ekkehard Köhler1 , Rolf H. Möhring2 , Klaus Nökel3 , and Gregor Wünsch2∗

1
Technische Universität Cottbus, Institut für Mathematik, Postfach 10 13 44,
03013 Cottbus, Germany, ekoehler@math.tu-cottbus.de
2
Technische Universität Berlin, Institut für Mathematik, MA 6–1,
Straße des 17. Juni 136, 10623 Berlin, Germany
{moehring,wuensch}@math.tu-berlin.de
3
PTV AG, Stumpfstraße 1, 76131 Karlsruhe, Germany
klaus.noekel@ptv.de
Summary. The coordination of signal controls in a traffic network has a significant

impact on throughput and journey times of the vehicles in the network. We present
a mixed-integer linear programming model optimizing the coordination, i.e., a model
that minimizes waiting times of individual traffic by adjusting the offsets between
signals. Compared to existing approaches, our model supports non-uniform cycle-
lengths of signals and yields optimal solutions (w.r.t. the objective function of our
model) in less computation time. We use microsimulation which shows that high
quality solutions can be obtained with little computational effort.
1 Introduction
As traffic in urban street networks continues to increase, traffic engineers strive
to manage capacity intelligently rather than merely add road space which may
not be an option in many built-up areas. Signalized intersections are a critical
element in such networks and care must be taken in the definition of the
signal phasing. Not only do the cycle length and splits influence capacity, but
so does signal coordination. In order to avoid (both objective and perceived)
delays through frequent stops and to reduce queue waiting time and length,
the signal timings at several intersections should be offset from each other
in such a way, that platoons of cars can progress through the network with
minimum impedance.
Optimization approaches that consider offsets between signals have been
considered by Little [9], by Gartner et al. [2] and by Improta [7], among others.
A genetic algorithm approach for network coordination has been developed
∗
Supported by the DFG Research Training Group GK-621 “Stochastic Mod-
elling and Quantitative Analysis of Complex Systems in Engineering” (MAGSI).
03/2004−02/2007.
180 E. Köhler et al.
by Robertson [10] and successfully implemented in the still state-of-the-art

software tool TRANSYT [10]. However, both, the absence of guarantees on the
quality of solutions and the sometimes high running times are disadvantages
of this tool.
In this paper we develop a mixed-integer linear program that minimizes
the total delay occurring in a traffic network by adjusting optimal offsets and
split modes. It is based on an approach by Gartner et al. [4], also [2] and [3]
and produces high quality solutions with little computational effort.
In Sect. 2 we give a detailed description the model. In Sect. 3 we show via
real-world instances that the model produces good solutions in very reason-
able running times. Finally, in Sect. 3.1, we discuss different aspects of our
optimization model in practice.
2 Our Approach
In our approach we consider an inner-city traffic network with fixed-time
signal control at the intersections. Non-uniform cycle lengths are permitted
in the network. Moreover, we assume the network to operate at near sat-
urated condition. Hence, as already motivated by Wormleighton [13], it is
justified to model the traffic macroscopically via platoons. With doing so
we further assume the traffic volumes to be given link-wise. So, we set up
a mathematical model that minimizes the sum of delays on all links in the
network with offsets between the signals and split modes as decision vari-
ables. In Sects. 2.1 to 2.4 we discuss the meaning of the variables, the con-
straints that are formulated and how the evaluation of the waiting time on
a link is estimated. However, it is not the scope of this paper to provide all
details.
2.1 Preliminaries and the Offset Variables
We model the traffic network by a directed multi-graph G = (V, A), where

the vertices v ∈ V represent the signalized intersections and the edges a ∈
A stand for the links between the intersections. We will also use the terms
node and link, respectively. The reason for allowing parallel edges is that we
distinguish the traffic flow on a link w.r.t. the signal groups of the adjacent
intersections that these vehicles come from and go to, respectively. However,
in the following we omit the indices for different copies of a link to simplify
notation.
Furthermore, we assume that each signal control in the network has a cycle
length of either 60, 80, or 120 seconds, which is exogenously given. While not
being overly restrictive, this constraint allows us to handle non-uniform cycle
lengths.
As already mentioned above, there are two main characteristics of our
model. First, it is assumed that vehicles move in platoons, see Figs. 2 and 3.
Optimization of Signalized Traffic Networks 181
Fig. 1. An illustration of the so-called intra-node offset Ψ v,i . It quantifies the shifting
of the beginning of a green phase of a particular signal group of the signal relative
to the beginning of the green of signal group 1 at that signal. In any mode Ψ v,1 = 0
for all v ∈ V
Within a platoon, non-uniform flow rates may exist. For expository reasons,
this is not displayed in the figures. The model’s second characteristic is that
it does not use any time discretization. The offset at the signal is modelled as
a continuous variable. Although in the application the offset has to be given
for each intersection, it is defined here for each link. On link e = (vi , vj ) the
arrival time of the platoon’s head at intersection vj is denoted by γij . Clearly,
γij ∈ [0, cj ], where cj is the cycle length at intersection vj . Whenever it is
obvious from the context, we omit the indices of the parameters. Fig. 1 shows
the intra-node offset Ψ , which is defined for each signal group at an inter-
section. In addition, this figure shows the two possible offset interpretations.
In our approach, the offset φij of link e = (vi , vj ) is defined as the distance
in time between the start of the platoon at node vi and the beginning of
the green phase that defines the relevant interval for the arrival time of that
platoon.
The connection between arrival time γ and offset φ and all other con-
straints are formalized in the next section.
2.2 The Constraints
Now we describe the constraints for our mixed-integer program formulation.

The first constraint specifies the dependency between the arrival time of
the platoon and the associated offset. For each link e = (vi , vj ),
τij − γij + rij = φij , (1)
where τij denotes the travel time on the link and rij is the length of the red
phase at intersection vj . Note that we use the notation rij instead of rj in
order to stress the assignment of a red phase to a link. However, the simple
linear dependency of γ and φ enables us to use the arrival times γ instead of
the offsets φ as decision variables for the model. This will be useful when we
calculate the delay on a link.
Before we define the next group of constraints, we need to define the set
of circuits of G. The set of circuits, denoted by L, contains all cycles of the
Fig. 2. The relation between parameters offset and arrival time
underlying undirected graph. Notice that a pair of parallel or anti-parallel

edges form a circuit as well.
For a particular circuit ∈ L that is traversed clockwise, F () denotes the
set of forward edges and R() the set of reverse edges, respectively. Further,
we define the circuit’s cycle time c as the greatest common divisor of the cycle
lengths at the vertices, i.e., signals, of . So, each circuit ∈ L is required to
fulfill
φe − φe + Ψ = n · c , (2)
e∈F () e∈R()
with n ∈ Z. The periodicities expressed in (2) are physical constraints that

are necessary to convert the link offsets into signal offsets.
Note that the intra-node offsets have to be included as well, since the
normal offsets do not respect “changes” of signal groups at an intersection.
This can be seen as follows. Consider a circuit that contains two links that
share a node v. Further, assume that these edges have the same orientation
with respect to the clockwise orientation of the circuit. If the edges do not
share the same signal group at node v, the shifting between the two signal
groups has to be taken into account.
At this point we already mention that providing good bounds for the
integer variables in (2) is essential for a fast solving of the MIP. We will
discuss this in more detail in Sect. 2.4.
In contrast to Gartner’s approach [4], our model offers the possibility to
choose between different red-green split modes. However, they need to be
specified in advance. Let Rv be the set of signal groups at node v. Then,
v,p
for a vertex v ∈ V and a signal group p ∈ Rv , the parameter Ψ m denotes
the intra-node offset at intersection v of signal group p in mode m, where
m = 1, . . . , αv . The parameter αv stands for the number of predetermined
modes at signal v. The binary variables dv,m then select a mode. Hence, we
obtain the following two groups of constraints

αv
v,p
dv,m · Ψ m = Ψ v,p ∀v ∈ A, p ∈ Rv , (3)
m=1

αv
dv,m = 1 ∀v ∈ A, (4)
m=1
where Equalities 4 ensure that exactly one mode is adjusted. In any of the
predetermined modes the signal group 1 has intra-node offset 0.
2.3 The Objective
Our objective is to minimize the total network traffic delay, that is due to
missing or bad coordination within the network. We still need to specify how
this delay is actually determined from the arrival pattern of the platoons.
Since non-uniform cycle lengths are allowed at an intersection we have to
consider a time span equal to the least common multiple of the cycle lengths
at the signals incident to the particular link.
Evaluating the arrival pattern means the following. Depending on the ar-
rival time of a platoon – which itself depends on the offset of the link – vehicles
may have to stop during the red phase and form a queue. Then, in the green
phase they are released. Note that we require that waiting queues must be
empty at the end of each circulation of the cycle length at the link’s destina-
tion signal.
This mechanism is sketched in Fig. 3. So, the link’s average waiting time z
equals the size of these queues accumulated over time and divided by the num-
ber of vehicles. Unfortunately, this term is not linear in our decision variable
γ, the arrival time of the platoon. We overcome this by piece-wise linearizing
the delay function which we evaluate at characteristic intermediate points.
We chose only a few intermediate points – between three and five – such that
the linearization becomes convex. Unlike in MITROP [4], this can be done
consistently and effectively, since we are considering a function of only one
variable. So, for a link e = (vi , vj ), let gke denote the line segments in a piece-
wise linear approximation of the delay function, where k = 1, . . . , βe , and βe
stands for the number of line segments. Thus,
zij ≥ gke (γij ), (5)
together with the convexity of the linearization ensures a consistent setting
Fig. 3. The arrival pattern on a link e = (vi , vj ) with cycle lengths at the intersec-
tions of 60 and 80 seconds, respectively
for the delay variable z. Hence, the objective is formulated simply by

fij · zij , (6)
(i,j)∈A
with fi,j denoting the traffic volume on a link.

Although it is not explicitly stated in our MIP several other optimization
criteria can be included. For example, the average number of stops can be
incorporated in a very similar way.
2.4 The MIP

Now, we define the problem as a mixed-integer linear program with offsets
and red-green split modes as decision variables. Again, for simplicity, we omit
indices of different copies of link e = (vi , vj ) in the MIP formulation. However,
the interaction between different traffic flows on parallel and anti-parallel links
is considered in the linearization of the delay function.

minimize fij zij
(i,j)∈A

h
φe − φe + Ψ vi ,p = n · c ∀ ∈ C,
e∈F () e∈R() r=1

αk
k,p
ck,m Ψ m = Ψ k,p ∀k ∈ K, p ∈ R,
m=1

αk
ck,m = 1 ∀k ∈ K,
m=1
zij ≥ gre (γij ) ∀e, r,
τij − γij + rij = φij ∀(i, j),
n ≤ n ≤ n ∀,
n ∈ Z, γij ∈ [0, Tij ], ck,m ∈ {0, 1}.
Here vi , i = 1, . . . , h denote the vertices on circuit ∈ L. Note that the cycle
equations, (2), do not have to be considered for all circuits of the underlying
graph G. Rather, it suffices to require them to hold for all cycles of an integral
cycle basis [8]. Hence, we have a polynomial number of constraints in the MIP.
For solving this mixed-integer linear program either standard tools like
ILOG CPLEX [6] or academic software such as SCIP [1] can be used. However,
any solver’s performance strongly depends on the bounds (n and n ) that
are provided for the integer variables n . Although simple bounds for the
n are quite immediate – just take advantage of the bounds for the offsets –
optimizing them over all integral cycle bases is often neglected. The advantages
of strengthening bounds are well studied for the related problem of cyclic
timetabling [8].
3 Application of the Model
We have tested our MIP model on several real world instances. Two promi-
nent examples are inner-city area networks of Portland and Denver, as they
have a regular structure and their fixed-timed signals correspond well to our
model.
First, we applied the model to a section of the network of Portland with 16
fixed-time controlled signals and route volumes adapted to peak hour volumes.
We observed results which are promising in two respects. First, by simulat-
ing the network and its signal control in VISSIM [12], we observed that the
calculated offsets show a consistent character. For those paths through the
network, where the optimizer had achieved compatible offsets, smooth pro-
gression was indeed reproduced within the simulation. Then we reconstructed
the grid-like network in TRANSYT, ran its genetic algorithm to find good
offsets and compared the results. Afterwards, we again used VISSIM to com-
pare simulation results – one hour of simulation with an appropriate start-up
phase – of both our model’s offsets and the ones obtained by TRANSYT. The
results are displayed in Table 1(a).
As a second real-world instance we chose the inner-city area network
of Denver, see Fig. 4, with an original morning peak hour traffic. On this
network with 146 fixed-time signals the MIP could be solved within 4 min-
utes (CPLEX) leaving an optimality gap of only 4%. The results in Table 1(b)
show that this solution is comparable to the present coordination.
We are currently preparing more large real-world instances that allow us
comparisons with TRANSYT without reconstructing the network.
Table 1. Comparison of delays from different coordinations.

(a) Portland (section), 16 signals
Coordination Delay CPU time Relative
in s/veh difference
best random 27.0 - 63%
present 16.6 - 0%
optimization 16.1 < 1s −3%
TRANSYT 15.9 (22.2) 800s (10s) −4% (34%)
(b) Denver, 146 signals

Coordination Delay Relative
in s/veh difference
random 192.08 49.2%
best of 20 random 171.75 33.4%
present 128.70 0
optimization 132.85 3.2%
Fig. 4. A screenshot from the microsimulation VISSIM showing the Denver network
3.1 Different Aspects of the Optimization Algorithm in Practice
The optimization algorithm mainly has two different practical applications.

First, the optimization of the network coordination can be used as a stand-
alone optimizer. Here, the optimization is done once and interactions between
signal-control and demand are neglected.
On the other hand, the model, i.e., the network coordination can be em-
bedded within an optimization scheme that reflects the dependencies between
signal-control and demand. We describe the latter of the two applications in
more detail below.
Optimizers, such as TRANSYT [10] and SYNCHRO [11], accept as input
a description of the supply, i.e. the street network (mainly links with travel
times, permitted turns at intersections), and of the demand, i.e. traffic flows
through the network. The demand can be either observed or derived from
a demand model. Interestingly, demand models are themselves extending in
the direction of traffic engineering, supporting models for signal control at
intersections and offering analysis methods for them, e.g. by incorporating
capacity analysis according to the HCM [5]. These models aim to provide
a more realistic node impedance for route choice by incorporating capacity
analysis into the assignment step. So far, most packages are limited to the
optimization of cycle lengths c and green time fractions (g/c), as run time
requirements have precluded network coordination. This is due to the na-
ture of the solution methods used in practice (genetic algorithms or quasi-
exhaustive search). A faster solution method would enable a closer integra-
tion between macroscopic demand modeling and network coordination. The
network coordination could then be part of the assignment process, allow-
ing to update route flows in response to changed offsets, and thus reach an
equilibrium between demand and supply that includes all aspects of signal
control.
This overall optimization process is illustrated by Fig. 5. Given OD flows
(from previous model steps) are assigned to a network with signal control,
Fig. 5. The organization of the overall optimization process with components as-
signment, local optimization and network coordination. We focus on the latter one
stressing the requirements in the context of the optimization scheme
using an assignment method which calculates node impedance from sig-

nal timings and offsets between different intersections. The assignment pro-
duces node flows (traffic volumes per maneuver at each intersection) which
are input to the local optimization of g/c. Path-based assignment proce-
dures also produce route flows on the sub-paths between successive signal-
ized intersections, or more specifically, between successive signal groups or
stages. A network coordination optimizer takes these sub-path flows and
globally optimizes the offsets. Then, the set of offsets are fed back to the
assignment. The loop is executed at least once and continues until route
impedances (including loss times experienced at signal controls) and flows
converge.
4 Conclusion
In this paper we focused on network coordination with minimizing delay
as objective. Based on a macroscopic approach used in the software tool
MITROP[4], we developed a mixed-integer linear program, thus enabling
guarantees on the solution quality, which is a step forward compared with
heuristic methods such as genetic algorithms.
Our mixed-integer linear program can handle non-uniform cycle lengths
at intersections and a selection between different red-green split modes. Last,
but not least, we suggested the application of new graph-theoretical insights,
i.e. tightening bounds of variables via the use of good integral cycle bases to
accelerate the running time of the MIP. Empirical studies showed promising
results concerning both the time needed to find a good solution and the solu-
tion quality. All this encourages the use of our network coordination approach
as a component in the overall optimization procedure sketched in Fig. 5.
References
1. Achterberg, Tobias (2004). SCIP – a framework to integrate constraint and
mixed integer programming. Technical Report 04-19. Zuse Institute. Berlin.
2. Gartner, N.H., J.D.C. Little and H. Gabbay (1975). Optimization of traffic signal
settings by mixed-integer linear programming, part i: The network coordination
problem. Transportation Science 9, 321–343.
3. Gartner, N.H., J.D.C. Little and H. Gabbay (1975). Optimization of traffic
signal settings by mixed-integer linear programming, part ii: The network syn-
chronization problem. Transportation Science 9, 344–363.
4. Gartner, N.H., J.D.C. Little and H. Gabbay (1976). Mitrop: a computer program
for simultaneous optimisation of offsets, splits and cycle time. Traffic Engineer-
ing and Control 17, 355–359.
5. Highway Capacity Manual (2000). TRB, Washington, USA.
6. ILOG CPLEX 10.1. (2006) User’s Manual.
7. Improta, G., A. Sforza (1982) Optimal Offsets for Traffic Signal Systems in
Urban Networks. Transportation Research Board 16B, No. 2, 143–161.
8. Liebchen, Christian (2003). Finding short integral cycle bases for cyclic
timetabling.. In: ESA. Vol. 2832 of LNCS. Springer. pp. 715–726.
9. Little, J. D. C. (1966) The Synchronization of Traffic Signals by Mixed-Integer
Linear Programming. Operations Research 14, 568–594.
10. Robertson, D.I. (1969). Transyt, a traffic network study tool. Technical Report
LR 253. Transport and Road Research Laboratory.
11. SYNCHRO, User’s Guide (2000). Trafficware, Sugar Land, USA.
12. VISSIM 4.10, User’s Guide (2005). ptv AG, Karlsruhe, Germany.
13. Wormleighton, R. (1965). Queues at a fixed time traffic signal with periodic
random input. CORS-Journal 3, 129–141.
Optimal Sorting of Rolling Stock
at Hump Yards
Ronny S. Hansmann and Uwe T. Zimmermann
Institute of Mathematical Optimization, Pockelstraße 14, 38106 Braunschweig,

Germany, {r.hansmann,u.zimmermann}@tu-bs.de
Summary. In this paper we provide a quite general description of a class of prob-

lems called Sorting of Rolling Stock Problem(s). An SRSP consists in finding an
optimal schedule for rearranging units of rolling stock (railcars, trams, trains, . . . )
at shunting yards, covering a broad range of specially structured applications. Here,
we focus on versions of SRSP at particular shunting yards featuring a hump. We
analyze the use of such a hump yard in our research project Zeitkritische Ablauf-
bergoptimierung in Rangierbahnhöfen 1 in cooperation with BASF, The Chemical
Company, in Ludwigshafen. Among other results we present a remarkably efficient
algorithm with linear running time for solving the practical SRSP at the BASF
hump yard.
1 Practical Problem Description

In railroad shunting yards, see Fig. 1, incoming freight or passenger trains
are split, parked and rearranged according to destinations or according to
construction type of railcars. Uncertain arrival times, ad hoc changing orders
of incoming railcars, the increasing number of rolling stock, sparse capacities
and financial constraints complicate the process and offer large potential for
optimization.
In the above context, we provide a description of a quite general class of
problems called Sorting of Rolling Stock Problems, covering a broad range of
special applications. In general, an SRSP consists of three processes: arrival,
parking, and departure. At the beginning an ordered input sequence of units
of rolling stock (railcars, trams, complete trains, . . . ) arrives at the shunting
yard. Then the parking process starts and the units enter the tracks of the
shunting yard. Here, incoming units have to be parked in such a way that at
departure the parked units can leave the shunting yard in a structured output
sequence. The difficulty of SRSP depends on the structural differences of the
1
Funded by the German Federal Ministry of Education and Research (BMBF),
grant no. 03ZINJBS
190 R. S. Hansmann, U. T. Zimmermann
Fig. 1. One of the shunting yards at the site of our practical partner: BASF, The
Chemical Company, Ludwigshafen
input sequence and the requested output sequence as well as on the structure
and flexibility of the shunting yard.
As most practitioners will confirm the following rule of thumb holds in
general: the less tracks used during the sorting procedure the lower the oper-
ational costs. Thus, a main goal of SRSP is to use as few tracks as possible.
In other words, an optimal solution is a schedule, i.e., an assignment of tracks
to units, using only a minimal number of tracks for the required sorting.
Structure of Output Sequence
As usual the incoming units are classified by a particular distinctive criterion,

e. g., their destination or their construction type. As common in practice, we
say that units satisfying the same criterion form a group.
We distinguish the following different structures of “labeled output se-
quences”. At first, all positions of the output sequence are labeled (e.g. with
integers). Secondly, all actually assigned units departing at positions with
identical label have to be members of the same group and, vice versa, all
members from a group have to be assigned to positions with identical labels.
In particular, the number g of different groups is the same as the number of
different labels. The labels of the positions of the output sequence form cer-
tain patterns. If the positions of the output sequence are labeled in a block-
wise manner, i.e., the labeled output sequence contains no subsequence of
positions labeled (u, v, u) for distinct labels u = v, we say that the labeled
output sequence consists of blocks. In view of the number of different groups
in the input, we call the labeled output sequence a g-pattern sequence or
Optimal Sorting of Rolling Stock at Hump Yards 191
Fig. 2. Sorting of Rolling Stock at a shunting yard
Table 1. Parameters for SRSP (b ≥ 1, h ≥ 0, s ≥ 0, g ≥ 1)

track topology sorting mode structure of
design length shunting timing splitting output sequence
stacks unbounded no shunting sequential s-split free g-blocks
queues b-bounded h-hump-shunting concurrent split ordered g-pattern
stacks/queues time windows chain-split
sido
diso
dido
a g-blocks sequence. For free output sequences, there is no fixed assignment

between groups and labels. Otherwise units departing at some position of the
output sequence are members of a pre-defined group, see Fig. 2, and we speak
of ordered output sequences.
Summarizing, we consider four types of structures of the output sequence:
free g-pattern and ordered g-pattern and their special cases free g-blocks
and ordered g-blocks. The departing order of units within a group is not
fixed and offers potential for optimization.
Of course, the choice of tracks is affected by several other parameters of
the shunting yard specifying either the track topology or the required sorting
mode, see Table 1.
Track Topology
There are various track topologies depending on the design and the length of
the tracks in the shunting yard.
Design. If the tracks may be accessed only from one side, that is, entrance and
exit are on the same side of the tracks, such that the other end is a dead-lock,
we speak of stacks. Note, that any two units parked on the same stack will
change their order from arrival to departure if not additionally rearranged or
shunted. Under the same assumption any two units preserve their order from
arrival to departure when placed on a queue (queues), which is a one way
track where the units arrive at one end and leave at the opposite side. In
the case denoted as stacks/queues one may freely decide whether a track is
used as queue or stack. In the above three cases the entrance as well as the
exit are only on one (possibly differing) side of the track, which is known in
literature as siso (single in single out). In the following further track designs
units may arrive at or depart from both sides of the tracks: sido (single in
double out), i. e., entrance is on one side, exit is on both sides; diso (double
in single out), i. e., entrance is on both sides, exit is on one side; dido (double
in double out), i. e., entrance and exit are on both sides.
Length. Of course, in real shunting yards, tracks are bounded in length. In
the case b-bounded, at most b units may be placed on each track. Though
unbounded track lengths are seemingly a rather theoretical issue, they may
well be reasonable from a practical point of view. Firstly, in general, it is
much harder and much more time-consuming to determine optimal schedules
complying with the real track lengths. Secondly, solutions being optimal with
respect to unbounded tracks, in practice seem to be easily transformable into
practically good schedules on bounded tracks. For example, there are two well
known approaches for handling “overfilled” tracks during actual operations.
Immediately after a track gets filled to capacity, one may either empty it using
an additional buffer, i. e., tracks beyond the shunting yard, or redirect units
initially planned to go on the filled track to another track in the shunting
yard.
Sorting Mode
The technical and organizational infrastructure of the yard leads to different

constraints on the feasible movements of units. We mainly capture these differ-
ences in the distinctive feasible movements available in a sorting mode: a unit
may (be) move(d) from the input(-track) to a track (i-t-move), from a track
to another or the same track (t-t-move or shunting-move), from a track to the
output(-track) (t-o-move) or directly from the input(-track) to the output(-
track) (i-o-move). If the structures of the input sequence and the output
sequence differ, at least some i-t-moves and t-o-moves are necessary.
Shunting. We consider special cases for which i-o-moves and/or different kinds
of shunting-moves are permitted or not. For instance, in the no shunting case,
no shunting-moves and no i-o-moves are permitted. Otherwise, depending on
the actual infrastructure of the shunting yard, there are various constraints on
the type of feasible shunting-moves. If the shunting yard features a so called
hump for splitting trains, we speak of a hump yard. At the site of our project
partner, the BASF AG in Ludwigshafen, the sorting and shunting operations
are mainly performed using a large and expensive hump yard facility. Sorting
or shunting over a hump is a common strategy to rearrange trains of units
without own power units. At arrival such units are pushed over the hump
before rolling down one by one into either appropriately chosen tracks (i-t-
moves) or the output-track (i-o-moves). As a consequence, instead of many
time-consuming pushing/pulling operations of units by locomotives on tracks
we only need one pushing operation of the complete input sequence at arrival.
In the same convenient manner one may use the hump for t-t-moves and t-
o-moves. For example, at BASF t-t-moves and t-o-moves are performed in
the following fashion. Per humping step all units placed on one track are
pulled back over the hump. Then these units are again pushed over the hump
either to the output-track (t-o-move) or to other tracks (t-t-move). For h-
hump-shunting we allow at most h such humping steps. The only difference
between no shunting and 0-hump-shunting is that i-o-moves are infeasible
for no shunting but feasible for 0-hump-shunting. If shunting-moves are
allowed, it is necessary to additionally describe the performance of the required
shunting-moves within the schedules.
Timing. We distinguish cases in which i-t-moves and t-o-moves appear com-
pletely separated or mixed on the time line. For example, at night depots it is
quite common that the first outgoing unit departs in the morning, long after
the parking process is completely finished at night. In this case (first t-o-move
after last i-t-move) we say that arrival and departure are sequential. Other-
wise, if we allow that departure and arrival are concurrent (i-t-moves and
t-o-moves are not consecutively), we may freely choose the departure time
of any track-leaving unit. However, since we want to minimize the number of
tracks used, a unit or group should obviously leave a track as soon as possible,
since then the chance of blockades of departures of other units or groups is
reduced. The input information for the sequential as well as the concurrent
SRSP is the input sequence, i.e., one only knows in which order the units
arrive. Contrary to these sequence SRSP, in the more general time win-
dows SRSP more input information has to be taken into account. Here, the
arrival time and departure time for each unit are exactly fixed in advance. We
assume w.l.o.g. that units with identical departure time belong to the same
group.
Splitting. Finally, the sorting mode is influenced by the way units may depart
from tracks which is closely related to the type of the units. Suppose our
problem consists in sorting units with own power units such as trams. Such
units are able to leave the shunting yard without the help of other devices like
locomotives. Thus, it is possible to split the units of one group arbitrarily over
the tracks (split). However, this splitting might not be reasonable if we want
to sort railcars without own power units into blocks since then one or more
locomotives would have to collect the units of one group from several tracks.
It would be much less time-consuming for the locomotive to pick all units of
one group as a block from a single track. As a consequence, we also consider
s-split SRSP where the units of one group may only be split up over at most
s+ 1 tracks, s ≥ 0. The particularly restrictive splitting condition chain-split
is reasonable only for sequential and no shunting. In this case, units may
be distributed over all tracks in such a way that collecting the units track by
track, i.e., all units placed on a track depart completely before all units of the
next track depart etc., leads to the required output sequence.
Similar to the notation of scheduling problems we propose an α|β|γ =: V
notation for the description of the many different versions V of SRSP
which result from the specification of the various parameters. Here, α spec-
ifies the track topology, β the sorting mode, and γ the structure of the
output sequence. A complete detailed list using the abbreviations defined
in Table 1 reads as follows: α ∈ {{st, qu, sq, sd, ds, dd} × {ub, b-bd}},
β ∈ {{nsh, h-hsh} × {se, co, tw} × {s-sp, sp, csp}}, and γ ∈ {{fr, or} ×
{g-bl, g-pa}}. Note, that the combination of free and time windows is not
reasonable, since a priori known departure times of the units obviously result
in a particular departure order as in the ordered case. In view of complexity
results we will distinguish optimization and decision problems. By α|β|γ, we
denote the corresponding optimization problem, i.e., finding a correspondingly
feasible schedule with minimal number of tracks. By k − α|β|γ we denote the
corresponding decision problem, i.e., answering the question wether there is
a correspondingly feasible schedule using at most k tracks.
In Sect. 2, we remind some notation for integer sequences and we give
a short description of the corresponding sequence versions of SRSP. In
Sect. 3, we briefly survey known results on SRSP. In Sect. 4, we present new
results with direct practical relevance for our joint project with BASF. In
particular, we discuss the versions st,ub|h-hsh,se,sp|or,g-bl and st,ub|h-
hsh,se,sp|fr,g-bl. Finally, in Sect. 5, we summarize some computational and
practical results.
2 Sequence Versions of SRSP
We denote an integer sequence, i. e., a sequence of integers, by S = (si1 , . . . , sin )

where s is a surjective function mapping each integer i ∈ I = {i1 , . . . , in |
i1 < i2 < · · · < in } to an integer si ∈ G. In particular each integer g ∈ G
occurs in S. A permutation Π = (π1 , . . . , πn ) of the integers 1, . . . , n corre-
sponds to an integer sequence with G = {1, . . . , n}. A subsequence of S is
a sequence S = (sij1 , sij2 , . . . , sijm ) such that 1 ≤ jk < jl ≤ n for each k < l.
We say that S contains a subsequence (u, v) (for example (1, 2)), if there exists
a subsequence (si , sj ) of S with si = u, sj = v (or si = 1, sj = 2). A set of
subsequences PS of S is called partition of S if each element of S belongs to
exactly one subsequence in PS . With respect to a certain property of subse-
quences (for example monotonicity), we will call a subsequence of S feasible,
if it has the property, or infeasible otherwise. Then, a minimum (feasible)
partition of S is a partition of S containing a minimum number of feasible
subsequences of S.
Two sequences S̄ = (s̄i1 , s̄i2 , . . . , s̄in ) and S̃ = (s̃j1 , s̃j2 , . . . , s̃jn ) are said
to be equal if s̄il = s̃jl for all l = 1, . . . , n. Otherwise, the two sequences are
called different. The concatenation S ⊕ S̄ = (s̃l1 , . . . , s̃ln+n̄ ) of two sequences
S = (si1 , si2 , . . . , sin ) and S̄ = (s̄j1 , s̄j2 , . . . , s̄jn̄ ) is a binary operation defined
by s̃lk := sik for k = 1, . . . , n and s̃lk := s̄jn−k for k = n + 1, . . . , n + n̄.
For sequence versions of SRSP an input sequence of n units (elements) is
described by an integer sequence S = (s1 , . . . , sn ). Here, the i-th unit (unit at
position i in S) belongs to the si -th group (integer). In particular in ordered
versions the groups are numbered according to their departing order. For
example S = (2, 3, 1, 2, 1, 2, 3) implies that the third and the fifth incoming
unit form the group containing the first outgoing unit. Otherwise, in free
versions, the groups may be arbitrarily numbered and these numbers do not
carry any information on the ordering of the outgoing units. The optimal
value, i. e., the minimal number of tracks used in a feasible schedule, will be
denoted by zV∗ (S) for the sequence version V, or shortly by z ∗ .
The sequence versions correspond to minimum partition problems, i. e.,
find a partition of S into a minimal number of feasible subsequences S1 ,. . . ,Sz∗
where the definition of feasibility depends on the version. Each subsequence
Sk corresponds to the units placed on track k.
There are several versions of SRSP where we can characterize whether
any two units (groups) may be placed on the same track or not. Then, the
SRSP corresponds to a minimum coloring problem of a corresponding graph
whose vertices are the units (groups). Two vertices in this graph are adjacent
if and only if the two units (groups) may not be placed on the same track. In
any feasible coloring, the vertices colored with color k correspond to the units
(groups) placed on track k.
3 Overview of Results in Literature

In the following we will focus on publications containing results for versions
listed in Table 1. Of course, there is a broad range of literature dealing with
similar or related problems.
In [4], Di Stefano and Koči present results for some n-blocks versions. For
the equivalent cases n-pattern and n-blocks the input sequence is a permuta-
tion and the splitting conditions coincide. It is shown that the equivalent ver-
sions st,ub|nsh,se,sp|or,n-bl, qu,ub|nsh,se,sp|or,n-bl, qu,ub|nsh,tw,sp|
or,n-bl, and qu,ub|nsh,co,sp|or,n-bl correspond to minimum coloring of

permutation graphs and consequently are solvable in O(n log n) time. It is
mentioned that st,ub|nsh,tw,sp|or,n-bl is equivalent to minimum coloring
of circle graphs and therefore N P-hard. Furthermore, minimum coloring for-
mulations in different hypergraphs for the versions sd,ub|nsh,se,sp|or,n-bl,
ds,ub|nsh,se,sp|or,n-bl, and dd,ub|nsh,se,sp|or,n-bl are introduced. It
is shown that version k−sd,ub|nsh,se,sp|or,n-bl is equivalent to deciding
whether there exists a partition of the given permutation (input sequence)
into at most k unimodal subsequences, which is shown to be N P-complete in
[5]. Since the versions sd,ub|nsh,se,sp|or,n-bl and ds,ub|nsh,se,sp|or,n-bl
are equivalent (see [4]), they are both N P-hard and both 3.42-approximable
in polynomial time, see [6]. The version sq,ub|nsh,se,sp|or,n-bl is equiv-
alent to minimum cocoloring of permutation graphs, which is proven to be
N P-hard in [13], and which is 1.71-approximable in polynomial time, see [7].
The version st,ub|nsh,se,csp|or,g-bl (st,ub|nsh,se,csp|fr,g-bl) is obvi-
ously equivalent to the problem of finding a minimum partition of the (input)
sequence S into subsequences S1 , . . . , Sz∗ such that the (output) sequence
S1 ⊕ S2 ⊕ · · · ⊕ Sz∗ has the structure ordered g-blocks (free g-blocks).
Dahlhaus et al. provide an algorithm in [3] which solves st,ub|nsh,se,csp|or,
g-bl in O(n) time and in [2] the version st,ub|nsh,se,csp|fr,g-bl (called Train
Marshalling Problem) is shown to be N P-hard.
In [9], we derive results for further versions. Here, we summarize some of
these results in the following table.
Equivalence to Theoretical
Version Min Coloring of Complexity
st,ub|nsh,se,sp|or,g-bl permutation graphs O(n log n)
st,ub|nsh,se,0-sp|gr,g-bl interval graphs O(n log n)
st,ub|nsh,se,0-sp|or,g-bl PI-graphs O(n log n)
st,ub|nsh,tw,sp|or,g-bl circle graphs N P-hard
st,ub|nsh,tw,0-sp|or,g-bl circle-polygon graphs N P-hard
st,ub|nsh,co,0-sp|or,g-bl circle-polygon graphs N P-hard
st,ub|nsh,co,0-sp|gr,g-bl circle-polygon graphs N P-hard
st,ub|nsh,co,sp|or,g-bl subclass of circle graphs N P-hard
Moreover, in [9] we introduce an Integer Programming model for mini-

mum coloring of circle-polygon graphs based on a particular network flow
model. Computational results show that the model can effectively be solved.
In this way, optimal colorings of circle-polygon graphs and therefore optimal
schedules for the above N P-hard versions can be determined with surprisingly
small computational effort.
Of course, if an unbounded version is N P-hard then the respective
b-bounded version is also N P-hard. Using results for Mutual Exclusion
Scheduling presented in [12] and [1], in [9] we show N P-hardness of the
versions st,b-bd|nsh,se,sp|or,g-bl and st,b-bd|nsh,se,0-sp|gr,g-bl. In [14]
it is shown that st,b-bd|nsh,se,sp|or,g-pa is N P-hard. A more compre-

hensive version of the proof is published in [15]. In the unpublished paper
[11], Jacob proves N P-hardness of the version st,b-bd|h-hsh,se,sp|or,n-bl
which extends to the more general versions st,b-bd|h-hsh,se,sp|or,g-bl and
st,b-bd|h-hsh,se,sp|or,g-pa.
Further results will be published in [8].
4 Versions of SRSP Applied at Hump Yards

Internal logistics at the site of our practical partner, BASF, The Chemi-
cal Company, in Ludwigshafen is mainly based on rail transport. The aim
of our project was to provide our practical partner with effective sched-
ules for rearranging trains. Though there are a few shunting yards at the
site, nearly all rearrangements are performed using the main hump yard, see
Fig. 3.
The tracks in the hump yard are stacks, railcars may arbitrarily split up
over tracks, and arrival and departure of trains are sequential.
For each incoming railcar we know in which train and block within the train
it has to leave. Each train serves several factory buildings and the railcars re-
quested by one factory form a block. To assure a secure and efficient handling,
Fig. 3. Railcars being pushed over the hump (left) such that they roll on appro-
priately chosen tracks (right). Pictures of the hump yard at the site of BASF, The
Chemical Company, Ludwigshafen
trains consist in suitably ordered blocks matching the served buildings;

hence the railcars of the respective block can simply be decoupled at the rear of
the train, preventing shunting in the streets. The trains leave the hump yard in
a known order. Both ordering informations can easily be encoded into one inte-
ger input sequence. For example, the input information about (train numbers,
block numbers) in ((2, 1), (1, 3), (1, 2), (2, 2), (1, 1), (2, 2), (1, 3), (1, 1)), i. e., the
first incoming railcar has to leave in the first block of the second outgoing train
is encoded as (4, 3, 2, 5, 1, 5, 3, 1).
For h-hump-shunting the movement of railcars as already described in
Sect. 1 proceeds as follows: per humping step all railcars placed on one track
are pulled back over the hump and then each of these railcars is again pushed
over the hump rolling either to the output-track (t-o-move) or to tracks in the
shunting yard (t-t-moves).
Obviously, the practical problem corresponds to the version st,b-bd|h-
hsh, se,sp|or,g-bl. However, dispatchers prefer the optimal schedules of the
respective unbounded version, see Sect. 5 for some reasoning. In the fol-
lowing, we develop a computationally very fast, linear time algorithm for
solving the version st,ub|h-hsh,se,sp|or,g-bl. We previously presented this
algorithm at OR 2006, Karlsruhe, September 2006. In Jacob [11], seem-
ingly the same algorithm is described in different terminology for the spe-
cial case with permutations as input. Furthermore, we show that the version
st,b-bd|h-hsh,se,sp|fr,g-bl is N P-hard.
For bookkeeping purposes, we arbitrarily number the tracks and, then, we
consider the track execution order E = (e1 , . . . , eh ). At humping step i, the
ei -th track is executed. For each railcar, we consider its path through the tracks,
i. e., the integer sequence of the track numbers of visited tracks. In particular,
the path of a railcar moving directly from the input to the output corresponds
to the empty sequence (), denoted by ∅. Two easily derived observations (cf.
the example shown in Fig. 4) are:
Observation 1. Railcars taking the same path correspond to a subsequence

of the input sequence.
Observation 2. A path corresponds to a subsequence of the track execution

order.
For fixed number of tracks and humping steps, the number of different paths,
along which a railcar can move, depends on the choice of the track execution
order. We call such a path realizable. For example for 2 tracks and 3 humping
steps, we consider the two track execution orders E1 = (1, 2, 2) and E2 =
(1, 2, 1). As shown in Fig. 4, E2 admits seven different realizable paths; on
the other hand, E1 only admits the six paths ∅, (1), (2), (1, 2), (2, 2), (1, 2, 2).
The following theorem contains an explicit formula for the maximum number
of different realizable paths and shows that it is achieved for cyclic track
execution.
Fig. 4. Optimal schedule, shunting moves over the hump, and the paths of the rail-
cars for the instance (7, 6, 5, 4, 3, 4, 1, 2, 1, 2) for version st,ub|h-hsh,se,sp|or,g-bl
Theorem 1. For k ≥ 1 tracks in the hump yard and h ≥ 0 humping steps

the maximal number f (k, h) of different realizable paths is achieved by per-
forming the humping steps according to the cyclic track execution or-
der (1, 2, . . . , k, 1, 2, . . .) of length h. A recursion for f (k, h) is f (k, h + 1) =
2 · f (k, h) − f (k, h − k) for h > k with starting values f (k, h) = 2h for h ≤ k.
An explicit formula reads

k+1
h

h−j·k
h
f (k, h) = 2 + (−1) ·
j
· 2h−j(k+1) .
j=1
j
Theorem 1 is an immediate consequence of Observation 2 and of the observa-

tions made in the proof of the following Corollary.
Corollary 1. Among all integer sequences Sk,n of length n containing integers

from {1, . . . , k} the cyclic sequence Sk,n
c
= (1, 2, . . . , k, 1, 2, . . .) contains the
maximum number of different subsequences.
Proof. We denote the number of different subsequences of an arbitrary se-
quence S by f¯(S), the number of different subsequences of S ending with
integer t by g(S, t), and the position of the rightmost integer t in S by r(S, t).
It suffices to show f¯(Sk,n ) ≤ f¯(Sk,n c
) for all subsequences Sk,n with n > k
containing all integers 1, . . . , k.
For an inductive proof, let Sk,0 := ∅, and c let
Sk,i := (s 1 , . . . , si ) for
i = 1, . . . , n − 1. Thus, we have to show that f¯ Sk,i ≥ f¯ Sk,i ① for some i
with k ≤ i < n implies the corresponding inequality for i + 1.
We observe the following properties of Sk,i :

② f¯ Sk,i+1 = 2 · f¯ Sk,i − g Sk,i , si+1 ,
%k
③ ¯
f Sk,i = 1 + t=1 g Sk,i , t ,
④ g Sk,i , t = f¯ Sk,r(Sk,i ,t)−1 , t = 1, . . . , k.
c
For easier comparison with Sk,i+1 we renumber the integers in Sk,i+1 and de-

note the result by S̃
k,i+1 := (s̃ 1 . , s̃i+1 ). Let r Sk,i , sj be the l-th greatest
, . .
index in r Sk,i , t | t = 1, . . . , k . Then, s̃j := si+1−l for j = 1, . . . , i + 1. For
c
example, S̃3,6 = (3, 3, 1, 2, 2, 1) for S3,6 = (2, 2, 3, 1, 1, 3) due to S3,5 = (1, 2, 3, 1, 2).
As a consequence, we get r(S̃k,i , sci+1 ) ≤ r(S̃k,i , s̃i+1 ) which leads to

f¯ S̃k,r(S̃k,i ,sc )−1 ≤ f¯ S̃k,r(S̃k,i ,s̃i+1 )−1 and with ④ to
i+1

⑤ g S̃k,i , sci+1 ≤ g S̃k,i , s̃i+1 .
Moreover, we observe
⑥ r(S̃k,i , t) ≤ r(Sk,i
c
, t), t = 1, . . . , k,

⑦ ¯ ¯
f S̃k,j = f Sk,j , j = 1, . . . , i + 1.
Finally, the following inequalities complete the proof.

⑦ ②
f¯ Sk,i+1 = f¯ S̃k,i+1 = 2 · f¯ S̃k,i − g S̃k,i , s̃i+1
⑤
≤ 2 · f¯ S̃k,i − g S̃k,i , sci+1
③ %k
= 2· 1+ g S̃k,i , t − g S̃k,i , sci+1
t=1
④ %k
= 2+2 f¯ S̃k,r(S̃k,i ,t)−1 + f¯ S̃k,r(S̃k,i ,sc )−1
i+1
t=1
t=sci+1
⑥ %
k
≤ 2+2 f¯ S̃k,r(Sk,i
c
,t)−1 + ¯
f S̃k,r(S c c
)−1
k,i ,si+1
t=1
t=sci+1
⑦ %
k
= 2+2 f¯ Sk,r(S c + f¯ Sk,r(S c
k,i ,t)−1 k,i ,si+1 )−1
c
t=1
t=sci+1
①
%
k
¯ ¯
≤ 2+2 c
f Sk,r c ,t −1
c
+ f Sk,r
(Sk,i ) (Sk,i i+1 )−1
c ,sc
t=1
t=sci+1
④ % k c c
= 2· 1+ g Sk,i , t − g Sk,i , sci+1
t=1
③ c c ② c
= 2 · f¯ Sk,i − g Sk,i , sci+1 = f¯ Sk,i+1
In [10], Hirschberg and Régnier prove that among all integer sequences Sk,n of
length n containing integers from {1, . . . , k} the cyclic sequence Sk,n
c
contains
c
¯
the maximum number ft Sk,n of different subsequences with cardinality n−t
for t = 0, . . . , n. Obviously this implies the above Corollary 1. They also
provide recursions for computing each of these maximum numbers in O(t +
(k − 1)n2 ) time. In our direct proof with respect to subsequences
c of arbitrary
cardinality, we obtain a recursion for the total number f¯ Sk,n of all different
c
subsequences of Sk,n and we derive an explicit formula for this maximum
number,
c cf. Theorem 1. Applying our recursion, we can compute f (k, n) =
f¯ Sk,n significantly faster in O(n) time.
Observations 1 and 2 together with Theorem 1 imply the validity of
the following Algorithm 1 for solving the versions st,ub|h-hsh,se,sp|or,g-bl
(st,ub|h-hsh,se,sp|fr,g-bl).
Algorithm 1:
begin
Step 1: Find a minimum partition of the input sequence S into
subsequences S1 , . . . , Sz ∗ such that S1 ⊕ S2 ⊕ · · · ⊕ Sz ∗
has the structure ordered g-blocks (free g-blocks)
Step 2: For given number h of humping steps determine the min-
imal number k∗ of tracks with at least z ∗ different real-
izable paths, i.e., with f (k∗ , h) ≥ z ∗
Step 3: For all i = 1, . . . , z ∗ assign a suitable path to subsequence
Si (railcars moving along path i)
end
For fixed number h of humping steps, Algorithm 1 computes an optimal sched-

ule for st,ub|h-hsh,se,sp|or,g-bl in O(n) time: Step 1 can be solved in O(n),
see [3], in Step 2 the minimal number k ∗ of tracks can be computed recur-
sively, see Theorem 1, in O(log n · h), and in Step 3 the assignment of paths
to subsequences again can be done in O(n). For a fixed number k of tracks,
a schedule with minimal number of humping steps can be computed in linear
time applying Algorithm 1 with a suitably modified Step 2.
Theorem 2. The version st,ub|h-hsh,se,sp|fr,g-bl is N P-hard.

Proof. If 1-st,ub|(h − 1)-hsh,se,sp|fr,g-bl is decidable in polynomial time,
then, due to previous results and due to the fact that f (1, h − 1) = h, we can
decide in polynomial time whether the input sequence S can be partitioned
into subsequences S1 , . . . , Sh such that S1 ⊕S2 ⊕· · ·⊕Sh has the structure free
g-blocks. However, the latter decision problem is known to be N P-complete,
see [2]. Thus, 1-st,ub|h-hsh,se,sp|fr,g-bl is N P-complete.
5 Practical Results
The task of our project at BASF, i. e., to provide an optimal schedule real-
izing the required rearrangements of trains at the hump yard, corresponds to
the version st,b-bd|h-hsh,se,sp|or,g-bl and is therefore a hard optimization
problem (N P-hard, see [11]). However, for practical data from real instances
per day with up to 600 incoming railcars we can compute optimal schedules
via solving our Integer Programming model with the commercial software
Cplex 10.0 within half an hour, see [8] for detailed computational results.
Although these solutions comply with the modeled fixed track lengths, they
are mostly not directly applicable since in practice the number of railcars that
can be placed on a track depends on several “soft” parameters. For example,
railcars in fact vary in length and railcars rolling down to tracks are slowed
down by automatic brakes which may lead to gaps between the railcars on the
track. Thus, again, it seems to be reasonable to compute optimal solutions of
the respective unbounded version st,ub|h-hsh,se,sp|or,g-bl and leave it to
the dispatcher to handle “full” tracks adequately, see also paragraph on Track
Topology in Sect. 1. Another practical advantage of the unbounded version is
that the running time of Algorithm 1 for practical data is less than one second
allowing quick reactions on real time changes in the predicted input sequence.
Although we implemented and demonstrated a prototype, the approach
will not be put into daily action before the dispatcher can use it directly
within the Monitoring and Control System VICOS, developed by SIEMENS
and installed at BASF. Following the proposal of our practical partners at
BASF, we contacted SIEMENS Braunschweig and we hope that the method
may be added as tool within VICOS in due course.
Acknowledgements
We are very grateful for the many motivating and supporting practical dis-
cussions as well as the real world data made available by Holger Schmiers,
Matthias Ostmann, and Rainer Scholz (BASF, Service Center Railway). We
would also like to express our sincere thanks to Dr. Anna Schreieck and Prof.
Dr. Josef Kallrath (BASF, Scientific Computing) for their steady support and
encouragement from the very first conceptual ideas until the not yet reached
successful end of the project.
References
1. H.L. Bodlaender and K. Jansen. Restrictions of graph partition problems. Part I.
Theoretical Computer Science, 148(1):93–109, 1995.
2. E. Dahlhaus, P. Horak, M. Miller, and J.F. Ryan. The train marshalling prob-
lem. Discrete Applied Mathematics, 103(1–3):41–54, 2000.
3. E. Dahlhaus, F. Manne, M. Miller, and J.F. Ryan. Algorithms for combinatorial
problems related to train marshalling. In Proceedings of AWOCA 2000, pages
7–16, Hunter Valley, 2000.
4. G. Di Stefano and M.L. Koči. A graph theoretical approach to the shunting
problem. Electronic Notes in Theoretical Computer Science, 92:16–33, 2004.
5. G. Di Stefano, St. Krause, M.E. Lübbecke, and U.T. Zimmermann. On minimum
k-modal partitions of permutations. In J.R. Correa, A. Hevia, and M. Kiwi, edi-
tors, Latin American Theoretical Informatics (LATIN2006), volume 3887 of Lec-
ture Notes in Computer Science, pages 374–385. Springer-Verlag, Berlin, 2006.
6. G. Di Stefano and U.T. Zimmermann. Short note on complexity and approx-
imability of unimodal partitions of permutations. Technical report, Inst. Math.
Opt., Braunschweig University of Technology, 2005.
7. F.V. Fomin, D. Kratsch, and J.-Ch. Novelle. Approximating minimum cocolour-
ings. Information Processing Letters, 84(5):285–290, 2002.
8. R.S. Hansmann. Optimal sorting of rolling stock (in preparation). PhD thesis,
Inst. Math. Opt., Braunschweig Technical University Carolo-Wilhelmina.
9. R.S. Hansmann and U.T. Zimmermann. The sorting of rolling stock problem
(in preparation).
10. D.S. Hirschberg and M. Régnier. Tight bounds on the number of string subse-
quences. Journal of Discrete Algorithms, 1(1):123–132, 2000.
11. R. Jacob. On shunting over a hump. unpublished. 2007.
12. K. Jansen. The mutual exclusion scheduling problem for permutation and com-
parability graphs. Information and Computation, 180(2):71–81, 2003.
13. K. Wagner. Monotonic coverings of finite sets. Elektronische Informationsver-
arbeitung und Kybernetik, 20(12):633–639, 1984.
14. T. Winter. Online and real-time dispatching problems. PhD thesis, Inst. Math.
Opt., Technical University Carolo-Wilhelmina, 2000.
15. T. Winter and U.T. Zimmermann. Real-time dispatch of trams in storage yards.
Annals of Operations Research, 96:287–315, 2000.
Stochastic Models and Algorithms
for the Optimal Operation of a Dispersed
Generation System Under Uncertainty
Edmund Handschin1 , Frederike Neise2 , Hendrik Neumann1 ,

and Rüdiger Schultz2
1
University of Dortmund, Institute of Energy Systems and Energy Economics,
Emil-Figge-Str. 70, 44227 Dortmund, Germany
{edmund.handschin,hendrik.neumann}@edu.udo
2
University of Duisburg-Essen, Department of Mathematics, Forsthausweg 2,
47048 Duisburg, Germany, {schultz,neise}@math.uni-duisburg.de
Summary. Due to the impending renewal of generation capacities and present deci-
sions concerning energy policy, dispersed generation systems become more and more
important. The optimal operation of such a system and corresponding trading activ-
ities are substantially influenced by uncertainty and require powerful optimization
techniques. We present expectation-based as well as risk-averse stochastic mixed-
integer linear optimization models using risk measures and dominance constraints.
Two case studies show the benefit of stochastic optimization in power generation
and the superiority of tailored solution methods over standard solvers.
1 Introduction
Technical, economical and also political developments have lead to substantial
changes in the field of the power industry in recent years. Politically motivated
decisions such as the nuclear power phase-out, the support of renewable en-
ergies as well as the amendment of the energy industry law (ENWG) and
the internationally declared Kyoto protocol will significantly influence the
realignment of the future energy supply. Technical aspects as the obsoles-
cence of German and also European power plants and the intention of saving
CO2 without use of nuclear energy pose big challenges for the future energy
supply.
Owing to the planned and partly started new building of conventional
power plants one can assume that the predominant part of our electrical de-
mand will be supplied by these conventional power plants. Due to the perma-
nent further technical development the dispersed generation (DG) close to the
consumer will admittedly play an important role in the future and therefore
meet a significant part of the generation capacity to be substituted. Especially
206 E. Handschin et al.
the combined heat and power production (CHP) becomes very important be-
cause of high overall efficiencies of these units.
The profitability of a single DG-unit is exclusively determined by the
amount of electricity and heat it produces. Contrary to a single DG-unit
the coordinated operation of several units offers an additional potential for
maximizing their profitability. Such an interconnection of DG-units partly
combined with storage devices is called a Virtual Power Plant (VPP). This
VPP is integrated into the existing power economical structures and can be
considered as a participant at the electricity market. Hence this VPP offers
a broad potential of optimization to its operator.
The paper is organized as follows: In Sect. 2 we describe the current de-
velopments of the electric power industry which offer many opportunities for
an economical operation of VPP’s. Further, Sect. 3 deals with an appropriate
mixed-integer linear modeling of all technical and economical constraints of
a VPP. In Sect. 4 we discuss uncertainties in the forecasted input data and
how they can be handled in the spirit of two-stage stochastic optimization.
Section 5 contains mathematical details of the different applied stochastic
optimization models and deterministic equivalents which are derived if we as-
sume discrete probability distributions of the uncertain input data. In Sect. 6
algorithms are developed which exploit the deterministic equivalents’ special
structure. We finally present two case studies showing the benefits of stochas-
tic optimization for the operation of VPP’s and the superiority of tailored
algorithms over standard solvers.
2 Current Developments of the Electric Power Industry

Investigating the current aging structure of the German generation system it
becomes clear that a big need for action exists. More than 40 % of the thermal
power plants will be elder than 35 years in the year 2010. Many of these power
plants will reach the end of their technical life time within the next few years.
Besides this technical aspect the politically motivated decision of the nuclear
power phase-out causes an additional deficit of installed generation capacity
of 20 GW. The increasing demand of replacing generation capacity offers a big
opportunity for a supply structure by use of dispersed generation.
The present electrical energy supply system is based on a central structure
with conventional power plants feeding into the high and extra high voltage
levels. This present situation causes a vertical power flow from the transmis-
sion networks via the distribution networks to the consumer. A significant rise
in DG of electricity and heat is prognosticated for the future which reduces
this vertical power flow [5].
DG is characterized by the following properties:
• Location of the DG-unit close to the consumer (at least in the same dis-
tribution network and/or local heat network)
Optimal Operation of Dispersed Generation Under Uncertainty 207
• Nominal capacity of the DG-unit smaller than 5 MW (Integration into

distribution networks only)
• Frequently coupled generation of electricity and heat
By use of CHP the primary energy is simultaneously converted into mechan-
ical or electrical energy and useful heat. Energy conversion in CHP-units has
clearly higher efficiency factors than the comparable separated production of
electricity and heat. That? why a better use of the loaded primary energy and
lower CO2 emissions can be achieved.
A basic advantage of DG is the reduction of electrical network losses caused
by a reduced power flow over the network. In certain situations with low local
load an inversion of the load flow from the DG-unit into the network is even
possible [2]. Compared to conventional power plants DG-units require low
investment costs. Therefore the associated entrepreneurial risk of the invest-
ment is relatively low. Further advantages of DG are the opportunity to delay
investments in the networks and the possibility to deliver ancillary services.
The deficit power in case of an outage of a DG-unit is relatively small because
of the low capacity of the DG-unit. These advantages come along with techni-
cal problems concerning the connection to the network [17] and disturbances
during network operation [33].
2.1 Structure of the Electric Energy Supply System

Whereas a big part of the activities in the field of DG deals with the technical
aspects of a area-wide integration of DG-units, important tasks concerning the
coordinated operation under consideration of economical aspects are only in-
sufficiently investigated. That is the reason why in this project an optimization
model is developed which allows the short and very short term optimization of
a VPP with respect to typical trade relations and contracts in the context of
a liberalized electricity market. The uncertainties in this time range are of big
importance for the development of the model and the selection of a suitable
optimization method.
2.2 Virtual Power Plant

Current publications show that only a few projects have been realized in the
field of Virtual Power Plants [1]. One of the first projects named VPP dealt
with the network integration and the required communication technology of
31 fuel cells in the range of some kW in the year 2001. Besides the field test of
the fuel cell, it was the aim of this project to prove the central management
and control of all facilities by use of modern technology [21, 20]. A coordinated
operation of the fuel cells in terms of an economical optimization was not as-
pired and would not have been sensible because of the long distance between
the DG-units. Another project analyzed and developed technical concepts
to enable a safe and reliable energy supply with a high penetration of re-
newable energies [6]. These investigations concentrate on the integration of
DG-units into the public low voltage networks. The aim of the optimization
was to minimize the operating costs from the network operator perspective
[10]. A project initiated in 2002 considered the operation of DG-units in con-
nection with gas, heat and electricity networks. The DG-units in this project
comprise two boilers, a district heating power station operated with biomass
and a battery station as well as two stochastic infeeds from wind and photo-
voltaic. The objective of this project was the calculation of optimal operation
schedules of the DG-units and controllable loads to reach the supply of the
electrical and thermal loads in terms of households, industry and trade at
minimal costs. A survey of the units considered in these projects shows that
the number of DG-units used for an optimal operation is restricted to only
a low quantity [4].
The aforementioned projects had a research character exclusively and were
used to try out the technical feasibility. In contrast to that in the two sub-
sequently described projects the operator aspires an increase of economic ef-
ficiency during practical operation. A VPP operated by a municipal utility
comprises 20 DG-units with 5.1 MW total electrical and 39.2 MW total ther-
mal capacity. The installed CHP-units include nine gas motors and one micro
turbine. These CHP-units are installed in five local heat networks. Each of
them is equipped with at least one thermal storage device [19]. The day-ahead
forecasts of the electrical and thermal load as well as of the gas demand pro-
vide the basis for the calculation of optimal schedules of the DG-units. By use
of the thermal storages the CHP-units are used to avoid expensive peaks in
the electrical import.
A project called Virtual Control Power Plant has been realized in Septem-
ber 2003. It comprises the pooling of different power plants and big customers
spread over the whole country with the objective to provide minute reserve.
The bundled capacity has increased to about 1400 MW within the last four
years. A bundled capacity of at least 30 MW allows the participation in the
market of ancillary services.
A VPP is defined as an interconnection of DG-units and storage devices
whose operation is optimized by a superior entity. Figure 1 illustrates this
definition.
Besides always existing non influenceable electrical and thermal loads also
controllable loads can be part of the VPP. The VPP is integrated into the
existing technical and economical structures. Therefore an interaction with
energy markets, e.g. in terms of energy contracts, trade at the spot market
and the provision of ancillary services can be used to achieve an optimal
operation of the VPP. The required processes are controlled by an Energy
Management System (EMS). This EMS contains the forecast of all required
data as the electrical and thermal demand of the customers belonging to the
VPP and the stochastic infeed from wind turbines and photovoltaic as well
as the spot market prices. The core of the EMS is the optimization module.
Within this module the short term optimization of the VPP is performed
based on models of the DG-units, storages and other available instruments.
Fig. 1. Illustration of the definition of a VPP
Furthermore the communication technology which is needed to submit the

optimal schedules to the different units and to control their operation is an
important part of the EMS [31].
VPP as a Market Participant
Among the operation of his DG-units the operator of the VPP is also able
to trade at different markets to maximize his profit resp. minimize his costs.
He is at least a buyer at the market of electrical energy because the total
demand within the VPP is generally not exclusively supplied by the DG-
units (see Definition VPP in Sect. 2.2). This demand can be met by bilateral
contracts at the OTC-market and the spot market as well as the intraday
trading. In addition to that the operator is also able to sell ancillary services
to the transmission system operator (TSO). All in all the following trading
platforms are available:
• OTC-market (bilateral contracts)
• Stock exchange (spot market, intraday trading)
• Market of ancillary services (regulation and reserve power)
DG-Units and Storage Devices as Components of a VPP
Principally DG-units with different technologies are available on the market.

The following listing gives a rough survey about the most important technolo-
gies. Among gas or oil driven motors also gas turbines are used as CHP-units.
These technologies are already available whereas fuel cells are in general not
ready for the market yet. One exception is a phosphor acid fuel cell (PAFC)
which is commercially available with an electrical capacity of 200 kW and

a thermal capacity of 220 kW. To meet the thermal peak demand gas or oil
fired peak load boilers are used. Additionally hot water storages are applied
as thermal storage devices. The much more difficult storage of electricity is
normally realized by use of lead-acid batteries.
3 Dispersed Generation Model
An integral component for an economical optimization of the operation of

technical equipment for energy supply comprises the mathematical modeling.
The modeling has to cover all specific technical properties of the units and
all hereby associated basic conditions. As mentioned before the economical
boundary conditions are of great importance for the operation of a VPP and
therefore have to be considered also within the model.
3.1 Modeling of the Technical Basic Conditions
The modeling of the technical basic conditions comprises the following aspects:
• Technical limits and gradients of the power output of DG-units and storage
devices
• Efficiency characteristics of fuel cells with different technology, gas turbines
and gas motors
• Technical specifications concerning operation times, down times and switch-
ing frequencies of DG-units
• Start-up and warm-up behavior of the different DG-units
• Diverse effects in conjunction with storage losses of electrical and thermal
storage devices (self discharge losses, charge and discharge losses)
• Controllable electric loads
As an example of modeling technical aspects within a mixed-integer linear
programming model (MILP) an efficiency characteristic model of a polymer
electrolyte membrane (PEM) fuel cell is subsequently described. Most DG-
units have a non-constant efficiency behavior which depends on the current
power output of the unit. These characteristics can be described with efficiency
characteristic curves. The developed model features by exclusively linear de-
pendencies and is therefore very well applicable for the implementation within
a linear program. If the achieved accuracy of the model is not sufficient in indi-
vidual cases an extended model based on a segmented efficiency characteristic
curve can be applied [25].
The principal approach for modeling typical efficiency characteristic curves
becomes clear by equation (1), where Pit denotes the power ouptut and Gti
the fuel input of unit i. The variable sti distinguishes between an on- and off-
status of unit i. The model can be adapted to the real efficiency characteristic
curve derived from measurements by adjusting the parameters ki,η1 and ki,η2 .

Pit = ηi,max · Gti − ki,η2 · sti · Pi,max − ki,η1 · Pit (1)
This approach can be applied to DG-units with exclusively thermal or elec-

trical output as well as for CHP-units. To model the entire efficiency charac-
teristic of a CHP-unit the modeling of the electrical and the total efficiency
characteristic is sufficient. The resulting expression of the electrical efficiency
characteristic in (2) can be achieved by transforming equation (1).
t
t
Pi,el ηi,el max
ηi,el (Pi,el ) = t = Pi,el max
Gi 1 − ki,elη1 · ηi,el max + ki,elη2 · ηi,el max · sti · t
Pi,el
(2)
By use of equation (2) the power from fuel of a DG-unit i can be expressed
with the following linear equality constraint:

1
Gti = Pi,el
t
· − ki,elη1 + ki,elη2 · sti · Pi,el max
ηi,el max
The thermal power output Q̇ti of a CHP-unit arises according to equation (3)
as the difference of the total power output and the electric power output.
t t
Pi,total = Pi,el + Q̇ti (3)
The equality constraint for a CHP-unit can be formulated by means of the

expression for the entire efficiency (analogous to (2)) and equation (3) in the
following way:
Q̇ti = Gti · ηi,total (Pi,total

t
) − Pi,el
t
ηi,total max ( Gti − ki,totalη2 · sti · Pi,total max )

= − Pi,el
t
1 − ki,totalη1 · ηi,total max
This procedure allows the modeling of output-variable CHP-coefficients. In
case of constant electrical, thermal or entire efficiency factors this can be
considered as a special case with ki,elη1 = ki,elη2 = 0, ki,thη1 = ki,thη2 = 0 or
ki,gesη1 = ki,gesη2 = 0, respectively, within the presented approach.
Figure 2 shows the real and the modeled efficiency characteristic curve of
a PEM fuel cell with a nominal electrical power Pmax of 212 kW and a nominal
thermal power Q̇max of 237 kW. The fuel cell has a minimum electrical output
of 10 kW. The parameters ki,elη1 and ki,elη2 have been calculated to 0.565
and 0.502 by means of the least square errors method. The dashed line in
Fig. 2 shows the error of the linear model compared to the values taken from
the manufacturer’s data. Due to a maximum error of −1.6 % and an average
absolute error of 0.6 % the model offers a very good accuracy.
Fig. 2. Real and modeled electrical efficiency characteristic curve of the PEM fuel
cell
3.2 Modeling of the Economical Basic Conditions
As economical basic conditions the following aspects have to be considered:

• Network usage fees of the gas and electric network (price per kWh and
price per kW)
• Network usage fees avoided by infeed of DG-units
• Legally fixed feed-in tariffs according to the Renewable Energy Law (EEG)
and CHP Law (KWK-G)
• Bilateral contracts for electricity and gas
• Trading at the German spot market (EEX)
• Costs of positive and negative balance energy
4 Uncertain Input Data

For optimizing the operation of a VPP the forecast of all input data is needed
which influence the operation of the VPP during the time range of the opti-
mization. The following analysis of the input data and especially their forecast
errors is the basis for the selection of a capable optimization method.
4.1 Analysis of the Forecasted Input Data
The definition of a VPP explained in Sect. 2.2 directly reveals the data which
are required in form of their forecasts. Some decisions which significantly
influence the operation of a VPP, like the bids at the spot market or the ad-
justment of flexible energy contracts, have to be taken day-ahead. Therefore
the forecasts have to be duly available. The optimization time range has to
span at least a period up to the end of the next working day. Hence the results
of the following forecasts are relevant [14]:
• Forecast of the electrical load
• Forecast of the thermal load in each local heat network
• Forecast of the stochastic infeed from Wind energy and Photovoltaic
• Forecast of the spot market prices
Subsequently the analysis of the electrical load forecast is exemplarily ex-
plained. Forecasted and instantaneous values provided by a municipal utility
over a time range of two years form the basis for this investigation. The
forecasts for the following working day are calculated each morning at 7
a.m. on the basis of the actual weather forecast. In general they comprise
a time range of 41 hours (short term forecast). The distribution of the stan-
dard deviation σ of the relative forecast error over all following days was
calculated and is depicted in Fig. 3. The dashed marked best fit straight
line shows a nearly constant distribution over the day with an average stan-
dard deviation of approximately 3.1 %. Furthermore the very short term fore-
casts have been calculated by means of the Box–Jenkins-method at 0 and
6 a.m.. These results are also shown in Fig. 3. The analysis of Fig. 3 re-
veals that the forecast quality can be increased over the first five to six
hours compared to the short term forecast by use of the simple Box-Jenkins-
method.
Fig. 3. Standard deviations of the forecast error of the short term forecast and very
short term forecast
4.2 Handling of the Uncertain Input Data
Although well developed forecast tools are available nowadays the performed
analysis of the forecast data shows that forecast errors of different quantities
appear which can not be ignored. In addition to that some decisions have to
be taken day-ahead in a liberalized energy market. In a VPP these decisions
comprise e.g. the bids for the spot market and the adjustment of flexible en-
ergy contracts. Principally the influence of uncertainties in this system can
be described like this: Some decisions have to be taken at a time at which
the consequences of these decisions as well as their assessment can not be
exactly determined because of uncertainties at the time of realization. This
phenomenon can easily be described for trading at the spot market. To get
a physical delivery of energy at day d + 1 the operator has to send his bids
up to 12 a.m. of day d latest to the European Energy Exchange (EEX). At
that time neither the spot market prices for the energy nor the electrical and
thermal load and the stochastic infeed of day d + 1 are known precisely. Due
to the different accuracies of the short and very short term forecast and the
temporally limited validity of the spot market prices, which are only known
up to the end of the actual day, a two-stage structure arises. This information
structure can be handled mathematically with a two-stage stochastic pro-
gramming model. It distinguishes between first-stage decisions that have to
be made without anticipation of the future and second-stage decisions that
can observe the realization of uncertainties. For the first-stage the input data
is assumed to be certainly known, while the data concerning the second-stage
is modeled by scenarios.
Input Data of the First Stage
The analysis of the forecast errors has shown that performing a very short
term forecast of the electrical and thermal load and the stochastic infeed
from wind leads to a highly reduced forecast error within the following five
to nine hours. Therefore these input data are assumed to be known when
the first-stage decisions are made. The time range of certain spot market
prices depends on the starting time of the optimization. In general it does not
match the time range where the loads and the stochastic infeed are known
with certainty. Thus, the spot market prices up to the end of the actual day
are supposed to be deterministic.
Scenario Generation for the Second Stage
The probability density functions of the forecast errors form the basis of the
development of data scenarios. In the following a normal distribution of the
forecast errors is assumed. We aim at the transformation of the continuous
probability density function of each forecast error in each time interval t into
a predefined number of scenarios with minimal loss of information. As an ex-
Fig. 4. Continuous (left) and discretized normal probability density function (right)
ample the discretization of a normal distribution into three strata is demon-

strated in Fig. 4.
The conditional expected value μ̃tj and its occurrence probability π(μ̃tj )
of stratum j are used as the substitute value of each scenario. The optimal
strata limits are calculated by means of statistical methods in dependence
on the predefined number of scenarios [25]. By calculating the optimal strata
limits and the resulting conditional expected values μ̃tj it is ensured that all
appearing forecast errors minimally diverge from their substitute values μ̃tj .
The number of scenarios in practice has to be determined individually in each
single case. This selection will always be a compromise between accuracy of
modelling and computational manageability. With this knowledge the scenar-
ios of the second-stage can be generated under consideration of the different
analysed forecast errors. According to the afore explained method the elec-
trical load forecast modelled with three scenarios is shown in Fig. 5 for an
optimization starting at 9 a.m.. Here the development of three scenarios by
superposition of the day-ahead forecast and the conditional expected values
μ̃tj is illustrated at 1 p.m. of the second day.
5 Stochastic Optimization
In this section, we introduce the mathematical optimization models which are
able to reflect the above described real-world situation. We start out from the
random optimization problem

min c x + q y : T x + W y = z(ω), x ∈ X, y ∈ Y , (4)
together with the information constraint that x has to be selected without

anticipation of the realization of the random data z(ω). This yields a two-
stage scheme of decision and observation: The first-stage decision x is made
before the observation of z(ω) which is followed by the second-stage decision y,

thus depending on x and z. Since X ⊆ Rm and Y ⊆ Rm̄+m are polyhedra,
possibly involving integer requirements to x and y, (4) is a mixed-integer
linear program.
Fig. 5. Exemplary illustration of modelling the first and second-stage of the elec-
trical load with three scenarios
The two-stage dynamics of (4) becomes explicit by a reformulation:

$ #
min c x + min q y : W y = z(ω) − T x, y ∈ Y x ∈ X
x y
$ #
= min c x + Φ(x, z) : x ∈ X . (5)
x
This formulation is known as the above mentioned two-stage stochastic op-

timization problem. Obviously, each first-stage decision x induces a random
variable and the selection of a decision x corresponds to the selection of a mem-
ber of the family
f (x, z) := c x + Φ(x, z) (6)
x∈X
of random variables.
Also the input data c, q, T, W could be assumed to be influenced by un-
certainty. In view of the optimal management of dispersed generation, the
uncertain spot market prices and fuel costs are reflected by c and q, respec-
tively, the forecasted infeed from renewable resources is included in W and
probabilistic load profiles go into z. For simplicity, we restrict ourselves to
a random right hand side.
It remains to define the criterion for the selection of the random variable.
A risk neutral approach is the application of the expectation E. This leads to
the expectation-based stochastic optimization problem
$ #
min E f (x, z) : x ∈ X , (7)
which is very common in stochastic programming, see [27, 28].

If one aims at risk averse decisions, mean-risk models as well as dominance

constrained problem formulations come to the fore, which shall be described
in detail below.
5.1 Mean-Risk Models

If the ranking of the random variables in (6) is done via a weighted sum of
the expectation E and some risk measure R, we arrive at the basic mean-risk
model $ #
min E f (x, z) + ρ · R f (x, z) : x ∈ X (8)
with a fixed parameter ρ > 0. The specification of the risk measure R should
be done according to the decision maker’s preferences, but also with respect
to the mathematical structure it induces to (8) and the algorithmic possi-
bilities for solving (8). As the initial random problem (5) is mixed-integer
linear, we should choose specifications that maintain the mixed-integer linear
characteristics in the case of finite, discrete distributions of z(ω). Moreover,
for dimensionality reasons, our algorithms will be able to exploit a special
block-structure of these mixed-integer linear programs which arises for many
specifications of R, but not for all. The following risk measures, for instance,
are in this sense mathematically and structurally sound:
Expected Excess:
QDη (f (x, z)) := E max f (x, z) − η, 0 ,
which computes the excess of the random variable f (x, z) over a prefixed
threshold η ∈ R.
Excess Probability:
QPη (f (x, z)) := P ω : f (x, z) > η ,
which is the probability that the random variable f (x, z) exceeds a given
threshold η ∈ R.
Conditional Value-at-Risk:
QαCVaR (f (x, z)) := min g(η, f (x, z)),
η∈R
where
1
g(η, f (x, z)) := η + E max f (x, z) − η, 0 ,
1−α
which reflects the expectation of the (1 − α) · 100 % worst outcomes of the
random variable f (x, z) for a fixed probability level α ∈ (0, 1).
Counterexamples for sound specifications of the risk measure are for example
the semideviation or the variance. For more details concerning the mathemat-
ical structure of different mean-risk models and algorithms associated with
(8) including different risk measures R see [23, 29, 30], for instance.
5.2 Deterministic Equivalents for Mean-Risk Models
For computational reasons, we assume the random input data z(ω) to be

discretely distributed with finitely many scenarios zl and probabilities πl , l =
1, . . . , L. The risk neutral model (7) can be equivalently restated as
$ %L "
min c x + l=1 πl q yl : T x + W yl = zl ∀l
, (9)
x ∈ X, yl ∈ Y ∀l
which again is mixed-integer linear. Furthermore, the constraint matrix of (9)

shows the block-structure depicted in Fig. 6.
The constraint matrix has an L-shaped form, which results from the fact,
that there are no constraints including second-stage variables yl1 and yl2 be-
longing to different scenarios l1 , l2 . Only the implicit coupling due to the
nonanticipativity (NA) of x prevents (9) from decomposition into L scenario-
specific subproblems. Similar results can be obtained for the above introduced
risk measures:
For instance, let R = QDη (Expected Excess), then (8) is equivalent to
$ %L %L ⎫
min c x + l=1 πl q yl + ρ · l=1 πl vl : ⎪
⎪
⎪
⎪
⎪
⎪
⎪
T x + W yl = zl ∀l ⎪⎪
⎪
⎪
⎬

c x + q yl ≤ vl ∀l ⎪ . (10)
⎪
⎪
⎪
yl ∈ Y, vl ∈ R+ ∀l ⎪⎪
⎪
⎪
⎪
⎪
⎪
⎭
x∈X
For all three risk measures deterministic equivalents can be derived that main-
tain the structure depicted in Fig. 6. This is important in view of the algo-
rithmic methods developed in Sect. 6.
Fig. 6. Block-structure of the constraint matrix of (9)

In general, the deterministic equivalents of all mentioned mean-risk mod-

els could be solved by standard mixed-integer linear programming solvers
(MILP). But, since their dimension increases substantially with an increasing
number of scenarios L, the solution by a decomposition of the equivalents
seems reasonable. The idea behind this decomposition is to exploit the above
described structure of the constraint matrix.
5.3 Dominance Constrained Model
A more recent approach to tackle (5) is the application of stochastic domi-

nance constraints. In contrast to mean-risk modeling, this approach aims at
the identification of acceptable members among the family of random variables
(6). Therefore, a reference benchmark a(ω) is introduced, which mathemati-
cally is again a random variable. It represents a random profile of the total
costs that is just acceptable in context of the application at hand and can be
derived from former optimization or experience of the decision maker, for ex-
ample. This approach is a more appropriate reflection of the stochastic nature
inherent in (4) and offers more flexibility to the decision maker to define what
is an acceptable solution to him. Mathematically the acceptability of a ran-
dom variable can be expressed by different notions of stochastic dominance,
for details see for instance [11, 15, 24].
Let μ and ν denote the Borel probability measures induced by z(ω) and
a(ω), respectively. Then f (x, z), with a fixed x ∈ X, is said to dominate the
reference benchmark a(ω) to first-order, denoted by f (x, z) (1) a, if

μ { z : f (x, z) ≤ η } ≥ ν { a : a ≤ η } ∀η ∈ R.
This means, that f (x, z) takes smaller values with a higher probability than a.
Note that the first-order dominance relation is suitable to reflect the behavior
of a rational decision maker. The underlying theory of utility functions is
established in [32].
Applying the first-order dominance to the two-stage random optimization
problem (5) leads to the following stochastic optimization problem with first-
order dominance constraints induced by mixed-integer linear recourse:

min g x : f (x, z) (1) a, x ∈ X (11)
where g x is supposed a new linear objective function that can be chosen

appropriate to the decision makers prerequisites. For example, in the context
of optimal operation of an energy generation system it could count the start-
ups of the single units and hence aim at minimizing abrasion.
Dominance constraints for stochastic optimization have already been es-
tablished in [7, 8, 9]. There, more general random variables are involved,
whereas in our case, f (x, z) results from the specific two-stage stochastic pro-
gramming context.
5.4 Deterministic Equivalents for the First-Order Dominance

Constrained Model
As done in case of mean-risk models, we assume the random variables z, a to

be discretely distributed with finitely many realizations zl with probabilities
πl , l = 1, . . . , L and ak with probabilities pk , k = 1, . . . , K. The first-order
dominance constraint then simplifies:

f (x, z) (1) a ⇔ μ { z : f (x, z) ≤ ak } ≥ ν { a : a ≤ ak } ∀k. (12)
For a proof see [12, 13, 26].

In view of (12) and assuming the discretization of the distribution of z, we
obtain a deterministic equivalent of (11):
There exists a constant M > 0 such that the dominance constrained pro-
gram (11) can be equivalently restated as
⎫
min g x : c x + q ylk − ak ≤ Mθlk ∀l ∀k ⎪ ⎪
⎪
⎪
%L ⎪
⎪
≤ āk ∀k ⎪
⎬
l = 1 πl θlk
, (13)
⎪
T x + W ylk = zl ∀l ∀k ⎪
⎪
⎪
⎪
⎪
⎪
⎭
x ∈ X, ylk ∈ Y, θlk ∈ {0, 1} ∀l ∀k
where āk := 1 − ν({ a : a ≤ ak }), k = 1, . . . , K.

The constraint matrix of this equivalent shows a similar desirable block-
structure, as the deterministic equivalents for mean-risk models. The structure
of the constraint matrix is displayed in Fig. 7.
As for the mean-risk models introduced above, we have the implicit cou-
pling of different scenarios by the nonanticipativity (NA) of x. Additionally,
Fig. 7. Block-structure of the constraint matrix of (13)

there are K constraints, represented by the block called “Dominance Con-

straint” in Fig. 7, which couple second-stage variables belonging to different
scenarios:
L
πl θlk ≤ āk ∀k (14)
l=1
This structure gives rise to a relaxation of coupling constraints and the estab-
lishment of lower bounds by some scenario-wise decomposition of the arising
problems. This is described in detail both, for the mean-risk models and the
first-order dominance constrained model, in the following section. As compu-
tational results presented in Sect. 7 will show, such a decomposition is superior
to the application of standard MILP solvers to (13).
6 Algorithmic Issues
To reasonably reduce the presentation of algorithmic issues, we restrict our-
selves to the mean-risk model with the Expected Excess as risk measure (10)
and the first-order dominance constrained model (13). The results can then
be transferred to the other stochastic models introduced above.
The overall idea to solve the above established deterministic equivalents is
to calculate lower and upper bounds iteratively and to embed this procedure
into a branch-and-bound scheme in the spirit of global optimization. The lower
bounds are derived by different relaxations of coupling constraints, which is
presented in the subsequent section. The next but one subsection deals with
the selection of upper bounds and a final subsection describes the applied
branch-and-bound method.
6.1 Lower Bounds
A first step towards lower bounds is to introduce copies x1 , x2 , . . . , xL of x

and to add %the explicit NA-constraint x1 = x2 = . . . = xL , which can be
restated by L l=1 Hl xl = 0 with suitable matrices Hl , l = 1, . . . , L. For the
Fig. 8. Block-structure of the constraint matrix of (10) with explicit NA

Fig. 9. Block-structure of the constraint matrix of (13) with explicit NA
deterministic equivalents (10) and (13) this yields the constraint matrices
displayed in Figs. 8 and 9, respectively.
Obviously, (10) readily decomposes into scenario-specific subproblems, if
the NA-constraint is relaxed, whereas for a decomposition of (13) the scenario-
coupling dominance constraints have to be relaxed, too.
Lower Bounds for Mean-Risk Models

The Lagrangean relaxation of the NA-constraint in (10) yields the Lagrangean
function

L

L(x, y, v, λ) := πl c xl + q yl + vl + λ Hl xl
l=1

L
= Ll (xl , yl , vl , λ)
l=1
We obtain the Lagrangean dual

$ #
max D(λ) : λ ∈ RL·(m−1) , (15)
with
$ ⎫
D(λ) := min L(x, y, v, λ) : T xl + W yl = zl ⎪
∀l ⎪
⎪
⎪
⎬
c xl + q yl ≤ vl ∀l
⎪
⎪
⎪
⎪
⎭
xl ∈ X, yl ∈ Y, vl ∈ R+ ∀l
%L $ ⎫
= min Ll (xl , yl , vl , λ) : T xl + W yl = zl ⎪
⎪
l=1 ⎪
⎪
⎬
c xl + q yl ≤ vl ⎪
. (16)
⎪
⎪
⎪
xl ∈ X, yl ∈ Y, vl ∈ R+ ⎭
The Lagrangean dual (15) is a nonsmooth concave maximization problem

which is tackled with bundle subgradient methods from nondifferentiable op-
timization, see [18]. In each step of such a method, we have to compute the
objective function of (15) for a fixed λ and one subgradient. At this point,
the above presented block-structure of (10) plays an important role: Together
with the Lagrangean relaxation of the NA-constraint it allows the computa-
tion of the objective value D(λ) by a decomposition into small subproblems.
This is reflected by (16) which shows that D(λ) can be restated as the sum
of L independent subproblems corresponding to the different scenarios.
Lower Bounds for the First-Order Dominance

Constrained Problem
The complexity of the Lagrangean relaxation mainly results from the num-
ber of relaxed constraints. Since usually the number of reference profiles K
in (13) is much smaller than the number of scenarios L, we apply the La-
grangean relaxation only to the linking dominance constraints and ignore the
NA-constraint. Moreover, the NA-constraint can be recovered much easier
than the dominance constraints.
We obtain the following Lagrangean function

L
K
L
L(x, θ, λ) := πl · g xl + λk πl θlk − āk
l=1 k=1 l =1

L
= Ll (xl , θl , λ).
l=1
The Lagrangean dual then reads

$ #
max D(λ) : λ ∈ RK
+ , (17)
with
$ ⎫
D(λ) := min L(x, θ, λ) : T xl + W ylk ⎪
∀l, ∀k ⎪
= zl
⎪
⎪
⎬

c xl + q ylk − ak ≤ Mθlk ∀l, ∀k
⎪
⎪
⎪
⎪
⎭
xl ∈ X, ylk ∈ Y, θlk ∈ {0, 1} ∀l, ∀k
%L $ ⎫
= min Ll (xl , θl , λ) : T xl + W ylk = zl ∀k ⎪
⎪
⎪
l=1 ⎪
⎬

c xl + q ylk − ak ≤ Mθlk ∀k ⎪ . (18)
⎪
⎪
⎪
xl ∈ X, yl ∈ Y, θlk ∈ {0, 1} ∀k ⎭
Again, the Lagrangean dual decomposes and can be computed by the solution
of scenario-specific subproblems.
6.2 Upper Bounds
Both described lower bounding procedures provide x̄1 , . . . , x̄L which can be
understood as proposals for a feasible first-stage solution x̄ of (10) and (13),
respectively. There are different possibilities to derive x̄ which means to re-
cover the relaxed nonanticipativity. For example, we can compute the mean
of the x̄1 , . . . , x̄L and round the result to the next integer if required, we can
choose the proposal x̄l belonging to the scenario with the highest probability
πl or we choose the proposal xl if scenario l incurs the highest costs. Having
selected x̄ we check its feasibility.
In case of the dominance constrained model, this is included into a pro-
cedure that tries to find θ̄lk which – together with the x̄ – fulfill the relaxed
dominance constraints. In a first step, we solve for all l = 1, . . . , L the problem
$% ⎫
min
K
θ : c
x̄ + q
y − a ≤ Mθ ∀k ⎪
⎪
k=1 lk lk k lk ⎪
⎪
⎬
T x̄ + W ylk = zl ∀k
⎪
⎪
⎪
⎪
ylk ∈ Y, θlk ∈ {0, 1} ∀k ⎭
which aims at pushing as many θlk as possible to zero.
The second step is to check for all k = 1, . . . , K, if the found θ̄lk fulfill

L
πl θ̄lk ≤ āk .
l=1
If they do, (x̄, θ̄lk ) is a feasible point for (13).
6.3 A Branch-and-Bound Algorithm
The iterative calculation of lower and upper bounds is embedded into a branch-
and-bound scheme which establishes partitions of the feasible set X with in-
creasing granularity. This is done via additional linear inequalities to maintain
the mixed-integer linear description of the constraint set. In the procedure, el-
ements of the partition of X, which correspond to nodes of the arising branch-
ing tree, can be pruned because of infeasibility, inferiority, or optimality. In
each node the gap between the currently best solution value and the currently
best lower bound is computed and serves as a stopping criterion. This results
in the following algorithm for the dominance constrained problem, which can
simply be transferred to mean-risk models.
Let P denote a list of problems and ϕLB (P ) a lower bound for the optimal
value of P ∈ P. Furthermore, ϕ̄ denotes the currently best upper bound to the
optimal value of (13) and X(P ) is the element in the partition of X belonging
to P .
Branch-and-Bound Algorithm:
Step 1 (Initialization):
Let P := {(13))} and ϕ̄ := +∞.
Step 2 (Termination):
If P = ∅ then the feasible point x̄ that yielded ϕ̄ = g x̄ is optimal.
Step 3 (Bounding):
Select and delete a problem P from P. Compute a lower bound ϕLB (P )
solving the Lagrangean dual (17) and find a feasible point x̄ of P with the
above described upper bounding procedure.
Step 4 (Pruning):
If ϕLB (P ) = +∞ (infeasibility of a subproblem in (18)) or ϕLB (P ) > ϕ̄
(inferiority of P ), then go to Step 2.
If ϕLB (P ) = g x̄ (optimality of P ), then check whether g x̄ < ϕ̄. If yes,
then ϕ̄ := g x̄. Go to Step 2.
If g x̄ < ϕ̄, then ϕ̄ := g x̄.
Step 5 (Branching):
Create two new subproblems by partitioning the set X(P ). Add these
subproblems to P and go to Step 2.
7 Computational Results – Case Study 1

To analyze the optimization results a VPP consisting of two local heat net-
works is taken as a basis (see Fig. 10). The thermal demand is supplied by
four CHP-units and two peak load boilers as well as two hot water storages.
To exhaust excessive heat in case of an emergency two cooling devices are
installed. The electrical demand is covered by the four CHP-units, a battery
storage and a flexible energy contract (10 % of fixed schedule adaptable day-
ahead) as well as by energy bought from the spot market. The complete block
diagram is depicted in Fig. 10.
To determine the optimal adaption of the flexible energy contract and
the optimal bids at the spot market for the following day the results of an
Fig. 10. VPP Configuration
optimization started at 9 a.m. are analyzed. The electrical demand (see also
Fig. 11), the thermal demand and the spot market prices are modeled by four
scenarios each. Thus 64 scenarios are considered in total.
Figure 11 shows a selection of optimal first-stage decisions under consider-
ation of the uncertain electrical and thermal demand as well as the uncertain
spot market prices. Among the optimal adaption of the flexible contract the
optimal spot market bids are displayed for a period of two consecutive working
days in winter. The power delivered by the contract is during the whole time
at its lowest limit excepting the period between 9 a.m. and 3 p.m.. Due to
the high spot market price to be expected during this period only very little
energy is bought at the spot market instead it is taken out of the contract.
Between 11 and 12 a.m. the high spot market price is used to gain profit by
trading at the spot market (sale of energy).
7.1 Benefit of the Stochastic Optimization
When solving optimization problems with stochastic input data in practice

the uncertainties are often substituted by their expected value, i.e. the fore-
cast [3]. Hence, in the following the benefit is identified which can be achieved
performing the stochastic optimization for the above configuration. There-
fore, the objective function value of the stochastic optimization (Value of the
recourse problem, VRP) is compared to the value of the deterministic objec-
Fig. 11. Optimization results for trading at the spot market and the day-ahead
adaption of the flexible contract
tive function when all uncertainties are substituted by their expected values
(Expected result of the expected value solution, EEV). The difference of both
values (Value of stochastic solution, VSS) can be distinguished as a benchmark
of the stochastic optimization [3].
Table 1 shows an excerpt of the results achieved for different constellations
in which the volatility of the spot market prices, the price of balance energy
and the forecast accuracy have been varied. The so called gap quantifies the
maximum decrease of the objective function value if longer computation time
would be admitted.
The results indicate that in each of the constellations investigated the
stochastic optimization leads to a better result (lower costs) than the conven-
tional deterministic alternative. It becomes apparent that the overall operat-
ing costs over a time period of 39 hours can be reduced by 1–2 % applying
the stochastic optimization. Especially the increasing benefit of the stochas-
tic optimization in context with the decreasing quality of the electrical load
forecast becomes apparent (see instances 2 and 5 in Table 1). In addition to
that the price of balance energy also has a big influence on the benefit of the
stochastic optimization. In this case an increase of 5 ct/kWh to 30 ct/kWh
leads to a VSS which is 295 e higher (see instances 2 and 4 in Table 1).
8 Computational Results – Case Study 2

We consider the optimal operation of a dispersed generation system with
respect to the supply of uncertain heat and power demand. The system,
which is run by a German utility, consists of five engine-based cogeneration
(CG) stations which produce thermal and electrical energy simultaneously.
The CG stations altogether include nine gas motors, eight boilers and one
gas turbine. They are each equipped with a thermal storage and a cool-
Table 1. Results for different parameters and input data

Instance No. Constellation VRP (e) VSS (e) VSS (%) Gap (%)
1 winter, low volatility 45939 409 0.88 3.20
2 winter, high volatility 45259 538 1.17 3.21

cbal Energy, + = 20 ct/kWh
cbal Energy, + = 30 ct/kWh
σel, error +1 %
σth, error +1 %
7 winter high volatility 45164 602 1.32 3.03
σSpot, error +1 %
Table 2. Increasing dimensions of the expectation-based model

Scenarios Boolean variables Continuous variables Constraints
1 8.959 8.453 22.196
5 38.719 36.613 96.084
10 75.919 71.813 188.444
50 373.519 353.413 927.324
ing device, such that excessive heat can be stored or exhausted. The sys-
tem is completed by twelve wind turbines and one hydroelectric power plant.
Whereas the thermal energy is distributed locally around each CG station,
the electrical energy is fed into the distribution network to supply the total
demand.
As shown in Sect. 3 the operation of such a system with all its technical
characteristics can be described as a mixed-integer linear model. Assuming
a time horizon of 24 hours, divided into 96 quarter-hourly time intervals, and
that all input data is deterministic, we obtain a problem formulation with
about 9000 boolean and 8500 continuous variables and 22000 constraints.
Though already large-scale, such a deterministic problem can be handled with
standard MILP solvers like Ilog-Cplex ([22]). Computational tests for differ-
ent deterministic load profiles showed that we are able to find operational
schedules inducing minimal costs with an optimality gap below 0.1 % for all
test instances in less than 20 seconds on a Linux-PC with a 3GHz processor
and 2GB ram. For details see [16].
The situation changes, if we include uncertain load profiles and use a mean-
risk formulation of the optimization problem. Dimensions for an expectation
based model, which can be compared to the dimensions of the mentioned
mean-risk formulations, are displayed in Table 2.
Table 3. Computation times for the expectation-based model

Number of Ilog-Cplex Decomposition
scenarios Objective Gap (%) Time (sec.) Objective Gap (%) Time (sec.)
5 7.150.945 0,0072 11 7.150.903 0,0067 8
10 5.987.642 0,0099 253 5.987.578 0,0087 52
20 5.837.584 0,0093 1.769 5.837.559 0,0088 172
30 5.928.500 0,0091 9.486 5.928.542 0,0093 364
40 5.833.156 0,0107 26.286 5.833.191 0,0093 487
50 5.772.335 0,0129 22.979 5.772.268 0,0094 611
Table 4. Results for the pure risk model with Conditional Value-at-Risk
Number of Ilog-Cplex Decomposition
scenarios R Gap (%) Time (sec.) R Gap (%) Time (sec.)
5 9.256.610 0,0007 23 9.256.642 0,0010 37
10 9.060.910 0,0033 67 9.060.737 0,0013 79
20 8.522.190 0,0040 504 8.522.057 0,0023 436
30 8.996.950 0,0049 5.395 8.996.822 0,0033 594
40 8.795.120 0,0049 7.366 8.795.064 0,0039 1.038
50 8.557.813 0,0050 9.685 8.557.755 0,0039 1.286
For the solution of the expectation-based problem the usage of the dual
decomposition algorithm, described in Sect. 6, already is preferable over the
solution of the corresponding deterministic equivalent by Cplex. Table 3 com-
pares computational results for expectation-based instances with 5 up to 50
scenarios with a stopping criterion of a gap smaller than 0.1 % gained from
Cplex and the decomposition method, respectively. The computation times of
Cplex are up to 35 times of the computation times of the decomposition.
Similar results can be observed for the models including risk measures.
Table 4 displays computations for a pure risk model with the α-Conditional
Value-at-Risk and with an optimality gap falling below 0.005 % as stopping
criterion. Obviously, Cplex is superior only in cases with 5 or 10 scenarios and
the decomposition is preferable if more scenarios are included.
As an example for computations including the expected costs and a risk
measure, Table 5 presents results for the mean-risk model with the Excess
Probability. Here, we apply a timelimit of 5 up to 25 minutes and can there-
fore use the reached gap as a criterion to compare the two solution meth-
ods. While for the instance with 5 scenarios Cplex beats the decomposition
method, in all instances with more scenarios the decomposition reaches bet-
ter solutions. In cases with 20 scenarios or more, Cplex even fails to find one
feasible solution.
Moreover, computations for a first-order dominance constrained problem
were made. As objective function g x we considered the sum of all start-ups
of the generation units in the first-stage which means, that the program aims
Table 5. Results for the mean-risk model with Excess Probability

No. Time Ilog-Cplex Decomposition
scen. (sec.) ρ E R Gap (%) E R Gap (%)
5 300 0 7.150.850 – 0,0057 7.150.866 – 0,0062
0,0001 7.150.900 0,400 0,0064 7.150.913 0,400 0,0066
10.000 7.150.880 0,400 0,0342 7.150.930 0,400 0,0348
10 450 0 5.987.670 – 0,0104 5.987.526 – 0,0078
0,0001 5.987.630 0,370 0,0098 5.987.554 0,370 0,0084
10.000 5.987.630 0,370 0,0315 5.987.556 0,370 0,0302
20 600 0 infeasible infinity 5.837.500 – 0,0078
0,0001 infeasible infinity 5.837.560 0,205 0,0089
10.000 infeasible infinity 5.837.556 0,205 0,0183
50 1500 0 infeasible infinity 5.772.226 – 0,0087
0,0001 infeasible infinity 5.772.271 0,205 0,0096
10.000 infeasible infinity 5.772.287 0,205 0,0151
Table 6. Dimensions of the equivalent of the dominance constrained problem

Number of 10 scenarios 20 scenarios 30 scenarios 50 scenarios
Boolean variables 299159 596799 894439 1489719
continuous variables 283013 564613 846213 1409413
constraints 742648 1481568 2220488 3698328
at the minimization of abrasion of the units. The benchmark profile is derived

from the solution of the expectation-based problem. We pick some of the so-
lution’s single scenario costs, cluster all other scenarios around the chosen
scenarios and assign as benchmark probabilities the sum of the scenario prob-
abilities. The dimensions of the arising deterministic equivalents for K = 4
benchmark profiles and L = 10 − 50 scenarios are displayed in Table 6.
Tables 7 and 8 picture the solution process of five problem instances with
different benchmark profiles and 10 or 30 data scenarios, respectively. The
status of the solution – represented by the currently best upper and lower
bounds – is given for those points in time, when one of the solvers reaches
a feasible solution, when a solver proves optimality of a currently best solution,
or when a solver runs out of memory. The timelimit for all computations is
eight hours.
In case of 10 scenarios the decomposition finds the first feasible solution
faster than Cplex and for instances 1, 4 and 5 it also proves optimality before
Cplex does. In contrast, for instances 2 and 3 Cplex is able to prove optimality
faster than the decomposition. Hence, for 10 scenarios the decomposition is
preferable if we aim at feasibility, but there is no obvious superiority.
For instances including 30 data scenarios the superiority of the decompo-
sition gets evident. In these cases the decomposition is still able to solve all
instances up to optimality, while Cplex reaches the limits of available memory
before it finds a first feasible solution, which is marked by ‘mem.’ in Table 8.
Table 7. Results for instances with 10 data scenarios and 4 benchmark scenarios
Number of Instance Benchmarks Time (sec.) Ilog-Cplex ddsip.vSD
scenarios Probability Benchmark Value Upper Lower Upper Lower
Bound Bound Bound Bound
1 0.12 2895000 430.43 – 29 29 15
0.21 4851000 899.16 – 29 29 29
0.52 7789000 15325.75 29 29 29 29
0.15 10728000
2 0.12 2900000 192.48 – 27 28 15
0.21 4860000 418.90 28 28 28 15
0.52 7800000 802.94 28 28 28 28
0.15 10740000
3 0.12 3000000 144.63 – 21 21 12
10 0.21 5000000 428.61 21 21 21 18
0.52 8000000 678.79 21 21 21 21
0.15 11000000
4 0.12 3500000 164.34 – 11 13 10
0.21 5500000 818.26 – 12 13 13
0.52 8500000 28800.00 13 12 13 13
0.15 11500000
5 0.12 4000000 171.52 – 7 8 8
0.21 6000000 3304.02 8 8 8 8
0.52 9000000
0.15 12000000
Table 8. Results for instances with 30 data scenarios and 4 benchmark scenarios
Number of Instance Benchmarks Time (sec.) Ilog-Cplex ddsip.vSD
scenarios Probability Benchmark Value Upper Lower Upper Lower
Bound Bound Bound Bound
1 0.085 2895000 473.27 – 28 29 12
0.14 4851000 1658.02 – 29 29 29
0.635 7789000 3255.99 – 29 mem. 29 29
0.14 10728000
2 0.085 2900000 1001.53 – 26 28 18
0.14 4860000 2694.93 – 27 28 28
0.635 7800000 3372.24 – 27 mem. 28 28
0.14 10740000
3 0.085 3000000 469.93 – 17 23 10
30 0.14 5000000 3681.15 – 18 mem. 21 20
0.635 8000000 28800.00 – – 21 20
0.14 11000000
4 0.085 3500000 618.21 – 10 14 8
0.14 5500000 3095.02 – 11 mem. 14 10
0.635 8500000 28800.00 – – 14 13
0.14 11500000
5 0.085 4000000 672.73 – 7 8 8
0.14 6000000 8504.88 – 8 mem. 8 8
0.635 9000000
0.14 12000000
If one tries to solve instances with 50 scenarios, the situation for Cplex
gets even worse, because then the available memory is not sufficient to
build up the (lp-)model file which is needed as input for Cplex. In contrast,
for the decomposition only a (lp-)model file including one single-scenario is
needed.
All in all, the decomposition is obviously preferable over standard solvers
when it comes to computations with a high number of data scenarios and
benchmark profiles.
Acknowledgement
This research has been funded by the German Federal Ministry of Education
and Research under grants 03HANIVG and 03SCNIVG.
References
1. U. Arndt, S. von Roon, and U. Wagner. Virtuelle Kraftwerke: Theorie oder
Realität? BWK, das Energie-Fachmagazin, 58(6), Juni 2006.
2. R. Becker, E. Handschin, E. Hauptmeier, and F. Uphaus. Heat-controlled com-
bined cycle units in distribution networks. CIRED 2003, Barcelona, 2003.
3. J. R. Birge and F. Louveaux. Introduction to Stochastic Programming. Springer,
New York, 1997.
4. B. Buchholz, N. Hatziargyriou, I. Furones, and U. Schlücking. Lessons Learned:
European Pilot Installations for Distributed Generation – An Overview by the
IRED Cluster. Cigré Session, Paris, August 2006.
5. European Commission. EU-15 Energy and Transport – Outlook to 2030, part
II, October 2006.
6. T. Degner, J. Schmid, and P. Strauss, editors. Final Public Report, Distributed
Generation with high Penetration of Renewable Energy Sources (DISPOWER),
Juni 2006.
7. D. Dentcheva, R. Henrion, and A. Ruszczyński. Stability and sensitivity of
optimization problems with first order stochastic dominance constraints. SIAM
Journal on Optimization, 18(1):322–337, 2007.
8. D. Dentcheva and A. Ruszczyński. Optimization with stochastic dominance
constraints. SIAM Journal on Optimization, 14(2):548–566, 2003.
9. D. Dentcheva and A. Ruszczyński. Portfolio optimization with stochastic dom-
inance constraints. Journal of Banking and Finance, 30:433–451, 2006.
10. T. Erge et. al. Reports on Improved Power Management in Low Voltage Grids
by the Application of the PoMS System, December 2005.
11. P. C. Fishburn. Utility Theory for Decision Making. Wiley, New York, 1970.
12. R. Gollmer, U. Gotzes, and R. Schultz. Second-order stochastic dominance con-
straints induced by mixed-integer linear recourse. Preprint Series, Department
of Mathematics, University of Duisburg-Essen, 644–2007, 2007.
13. R. Gollmer, F. Neise, and R. Schultz. Stochastic programs with first-order dom-
inance constraints induced by mixed-integer linear recourse. Preprint Series,
Department of Mathematics, University of Duisburg-Essen, 641–2006, 2006.
14. G. Gross and F. Galiana. Short-Term Load Forecasting. In Proceedings of the
IEEE, volume 75, December 1987.
15. J. Hadar and W. R. Russell. Rules for ordering uncertain prospects. The
American Economic Review, 59:25–34, 1969.
16. E. Handschin, F. Neise, H. Neumann, and R. Schultz. Optimal operation of
dispersed generation under uncertainty using mathematical programming. In-
ternational Journal of Electrical Power and Energy Systems, 28:618–626, 2006.
17. E. Hauptmeier. KWK-Erzeugungsanlagen in zukünftigen Verteilungsnetzen –
Potenzial und Analysen –. Dissertation, University of Dortmund, 2007.
18. C. Helmberg and K. C. Kiwiel. A spectral bundle method with bounds. Math-
ematical Programming, 93:173–194, 2002.
19. E. Hennig. Betriebserfahrungen mit dem virtuellen Kraftwerk Unna. BWK,

7/8:28–30, 2006.
20. http://ec.europa.eu/research/energy/pdf/efchp fuelcell16.pdf. Europe’s first
virtual fuel cell power plant, October 2006.
21. http://www.cogen.org/projects/vfcpp.htm. The virtual fuel cell power plant,
October 2006.
22. ILOG. CPLEX Callable Library 9.1.3, 2005.
23. A. Märkert and R. Schultz. On deviation measures in stochastic integer pro-
gramming. Operations Research Letters, 55:441–449, 2005.
24. A. Müller and D. Stoyan. Comparison Methods for Stochastic Models and Risks.
John Wiley and Sons, Chichester, UK, 2002.
25. H. Neumann. Zweistufige stochastische Betriebsoptimierung eines Virtuellen
Kraftwerks. Dissertation, University of Dortmund, July 2007.
26. N. Noyan, G. Rudolf, and A. Ruszczyński. Relaxations of linear programming
problems with first order stochastic dominance constraints. Operations Research
Letters, 34:653–659, 2006.
27. A. Ruszczyński and A. Shapiro, editors. Stochastic Programming, volume 10 of
Handbooks in Operations Research and Management Science. Elsevier, Amster-
dam, 2003.
28. R. Schultz. Stochastic programming with integer variables. Mathematical Pro-
gramming, 97:285–309, 2003.
29. R. Schultz and S. Tiedemann. Risk aversion via excess probabilities in stochas-
tic programs with mixed-integer recourse. SIAM Journal on Optimization,
14(1):115–138, 2003.
30. R. Schultz and S. Tiedemann. Conditional value-at-risk in stochastic programs
with mixed-integer recourse. Mathematical Programming, 105(2/3):365–386,
2006.
31. F. Uphaus. Objektorientiertes Betriebsführungssystem zur Koordinierung dezen-
traler Energieumwandlungsanlagen. Dissertation, University of Dortmund, May
2006.
32. J. von Neumann and O. Morgenstern. Theory of Games and Economic Be-
haviour. Princeton University Press, Princeton, 1947.
33. Th. Wiesner. Technische Aspekte einer großflächigen Integration dezentraler
Energieversorgungsanlagen in elektrische Verteilungsnetze. Fortschritt Berichte
VDI, Reihe 21, 313, 2001.
Parallel Adaptive Simulation
of PEM Fuel Cells
Robert Klöfkorn1∗, Dietmar Kröner1 , Mario Ohlberger2

1
Department for Applied Mathematics, University of Freiburg,
Hermann-Herder-Str. 10, 79104 Freiburg i. Br., Germany
{robertk,dietmar}@mathematik.uni-freiburg.de,
http://www.mathematik.uni-freiburg.de/IAM/
2
Institute for Numerical and Applied Mathematics, University of Münster,
Einsteinstr. 62, 48149 Münster, Germany
mario.ohlberger@uni-muenster.de,
http://wwwmath1.uni-muenster.de/u/ohlberger/
1 Introduction
Polymer electrolyte membrane (PEM) fuel cells are currently being developed
for production of electricity in stationary and portable applications. They ben-
efit from pollution free operation and a potential for high energy conversion
efficiency. As PEM fuel cells are currently operated within low temperature
and pressure ranges, water management is one of the critical issues in perfor-
mance optimization.
In this paper we present numerical simulations for liquid water and gas flow
in the cathodic gas diffusion layer of a PEM fuel cell. Hereby, we focus on re-
solved three dimensional simulations of the two phase flow regime using mod-
ern numerical techniques like higher order discontinuous Galerkin discretiza-
tions, local grid adaptivity, and parallelization with dynamic load-balancing.
A detailed model for the simulation of PEM fuel cells, including two wa-
ter transport modes in the membrane was given in [16]. Here, we restrict to
the transport mechanisms within the cathodic gas diffusion layer and extract
a suitable sub-model, including two-phase flow and species transport in the
gas phase. Details of this simplified model problem are presented in Sect. 2.
In Sect. 3 we comment on the discretization schemes that were used for the
simulation including remarks on adaptation and parallelization. In Sect. 4 nu-
merical results for an instationary parallel adaptive simulation in three space
dimensions are presented.
∗
Robert Klöfkorn was supported by the German Ministry of Education and Re-
search under contract 03KRNCFR.
236 R. Klöfkorn, D. Kröner, M. Ohlberger
2 The Reduced Model Problem

The three-dimensional, coupled, adaptive, and parallel simulation software is
tested on a reduced model problem which describes the fluid flow within the
cathodic gas diffusion layer (GDL). The reduced model considers the following
physical processes:
• Two-phase flow with phase transition in a porous medium.
• Transport of the gas species O2 and H2 O with reaction.
2.1 Two-Phase Flow with Phase Transition
A PEM fuel cell is operating at relatively low temperatures (around 60–80

degrees Celsius). Therefore, at the cathode side of the fuel cell liquid water
is produced. As a consequence, the flow in the gas diffusion layers (GDL) is
modeled via two-phase flow in porous media taking into account the phases,
liquid water and gas mixture, consisting of the species oxygen, hydrogen, water
vapor, and some rest mostly consisting of nitrogen. In the following the index
g denotes the gaseous phase whereas the index w denotes the liquid water
phase.
From the balance of the volume saturations sw , sg of the two phases we
get the two-phase flow in porous media (see [14, 11]):
∂t (Φρi si ) + ∇ · (ρi vi ) = qi , (1)

kri
vi = −K (∇pi − ρi g), i = w, g. (2)
μi
Here ρi denotes the density, si the saturation, vi the Darcy velocity, pi the
pressure of the phase i = w, g respectively, and g the gravity vector. Further-
more, K denotes the absolute permeability tensor, μi the viscosity of phase
i, kri the relative permeability of phase i, and Φ the porosity of the porous
medium. Additionally, qi denotes the source term modeling the phase transi-
tion of the phase i which is defined as follows (see [14, 11]):
qw := rphase , qg := −rphase , (3)

M Φs cH2 O
rphase := kc g RTg (pg,H2 O − psat
w ), if pg,H2 O ≥ pw .
sat
(4)
kv Φ sw ρw (pg,H2 O − pw ),
sat
else
Thereby, pg,H2 O denotes the partial pressure of the water vapor. For the de-
scription of the other physical parameters see Table 1.
The following constitutive conditions close the two-phase flow system
sw + sg = 1, pg − pw = pc (sw ). (5)
Here pc (sw ) denotes the capillary pressure. Whereas in the liquid phase there
exists only the species H2 O, in the gaseous phase we have the species k =
Parallel Adaptive Simulation of PEM Fuel Cells 237
O2 , H2 O, and R. Here, R denotes all other existing species (mostly nitrogen).

As a consequence, in the gaseous phase (i = g) the transport of species has
to be taken into account separately. Following [11, 14] we get from the mass
balances of the species in the gaseous phase the equations

∂t Φρg sg ck + ∇ · ρg vg ck − ∇ · ρg Def f (Φ, sg ) ∇ck = qgk . (6)
Here ck denotes the mass concentration of the k-th species, Def f the effective
diffusion coefficient, and qgk the source term of the k-th species in the gaseous
phase. As all the species together form the hole gaseous phase, we get the
following constitutive condition
cO2 + cH2 O + cR = 1. (7)
Therefore, the transport equation for cR will be dropped as the concentration

can be calculated using the constitutive condition from equation (7). The
source terms qgH2 O and qgO2 are modeled as follows
qgH2 O := qg , qgO2 := −rreac cH2 O . (8)
The equations (1) to (8) describe the two-phase flow with species transport
including phase transition and reactions in the gas diffusion layer (GDL) of
a fuel cell. See Sect. 4 for a detailed description of the domain for the PDEs
as well as the description of initial and boundary conditions.
2.2 Physical Parameters
The physical parameters in the two-phase flow equations and the transport
equations are chosen from Table 1.
2.3 Global Pressure Formulation and Resulting Equations
First, the two-phase flow system in (1) and (2) is reformulated with s := sw
and p := pg − πw (s) as independent variables. Therefore, we introduce the
following notation
global velocity: u := vw + vg ,
phase mobility: λi := kri /μi ,
total mobility: λ := λw + λg ,
fractional flow: fi := λi /λ,
phase velocity water: vw := fw u + λg fw K∇pc ,
phase velocity gas: vg := fg u − λg fw K∇pc ,
s
global pressure: p := pg − πw (s), πw (s) := 0 fw (z) pc (z) dz + pc (0),
Furthermore, the densities ρw and ρg are assumed to be constant and the
influence of the gravity is neglected. The equation for the global pressure is
Table 1. Physical parameters for the model problem
parameter symbol value unit

porosity Φ 0.7
abs. permeability K 5 · 10−11 [m2 ]
√
rel. permeability of gas krg (s) 1 − s (1 − (1 − (1 − s)1/m )m )2
√
rel. permeability of water krw (s) s (1 − (1 − s1/m )m )2
Van Genuchten coefficient m 0.95
−1
capillary pressure pc pc (s) = 5300(1 − s) 2.3 [Pa]
viscosity water μw 1.002 10−3 [Pa s]
viscosity gas μg 1.720 10−5 [Pa s]
density water ρw 998.2 [kg/m3 ]
density air ρg ρg (pg , T ) = Mg pg /RT [kg/m3 ]
temperature T 343.15 [K]
molar mass gas Mg Mg (c) = 1/(cH2 O /MH2 O + cO2 /Mair ) [kg/mol]
molar mass water MH2 O 0.018 [kg/mol]
molar mass of dry air Mair 0.02897 [kg/mol]
diffusions coefficient of
water vapor in air DgH2 O 0.345 · 10−4 [m2 /s]
effective diffusion Def f Def f (Φ, sg ) = (Φsg )3 DgH2 O [m2 /s]
coefficient
gas constant R 8.3144 [J/mol K]
condensation rate kc 1 106 [1/s]
vaporization rate kv 1 10−2 [1/Pa s]
saturation vapor pressure psat
w (T ) a exp( Tb + c − dT + eT 2 + f ln(T )) [Pa]
coefficient for psat
w a 1.00519
w b −6094.4642
w c 21.1249952
w d 2.724552 10−2
w e 1.6853396 10−5
w f 2.4575506
obtained by summing up the equation (1) for i = w, g, applying the consti-

tutive conditions (5), and inserting the above notations. Finally, we obtain the
pressure equation
−∇ · (Kλ(sw )∇p) = 0, (9)
and the
velocity equation
u = −Kλ(sw )∇p. (10)
Assuming that the saturation sw is given, equation (9) can be used to calculate
the global pressure, and finally equation (10) to compute the global velocity.
Assuming the global pressure, the global velocity, and the species concen-
trations are given, then, since the density is assumed to be constant, from
equation (1) including the definition of vw we obtain for sw the
saturation equation
∂t (Φsw ) + ∇ · (fw (sw )(u(sw ) + λg (sw )K∇pc (sw ))) = qw . (11)
With the pressure p, the velocity of the gaseous phase vg , and the saturation
of the gaseous phase sg the transport of species can be described by inserting
these three values into equation (6). For k = O2 , H2 O we obtain the
transport equations

∂t Φsg ck + ∇ · vg ck − ∇ · Def f (Φ, sg ) ∇ck = qgk . (12)
Now the considered model problem consists of the equations (9)–(12). Suit-
able boundary and initial conditions will be presented in the description of
the simulated test problem in Sect. 4.2. Throughout the rest of this paper
numerical simulations using this model problem will be presented.
3 Discretization of the Model Problem

and Implementation
The discretization of the model problem (9)–(12) uses Discontinuous Galerkin
methods. With these methods on one hand, higher order discretizations can be
achieved without increasing the stencil of the methods which is an appealing
feature when one wants to do parallel computations. Furthermore, there arise
no special difficulties when dealing with non-conform grids. Non-conform grids
on the other hand have the very nice feature, that the refinement zone stays
local unlike for example when conformal bisection refinement is applied. This
is again very useful for parallel computations. On the other hand higher order
Discontinuous Galerkin methods have a higher number of unknowns than for
example continuous Finite Element methods.
The implementation of the discretizations uses the software package DUNE
[2, 7, 6]. DUNE has a modular structure and in the following the DUNE
modules DUNE-Common, DUNE-Grid, DUNE-Istl, and DUNE-Fem have
been used. A detailed description of the modules can be found on the Dune
homepage [2]. DUNE-Common provides basic classes. DUNE-Grid defines
the abstract grid interface and provides its implementations for several dif-
ferent grids* . The following numerical simulations use ALUCubeGrid which
is the grid interface implementation of ALUGrid (see [1]) using hexahe-
dral elements. DUNE-Istl (see [5]) is an Iterative Solver Template Library.
∗
For a complete list we refer to [2].
It provides classes for matrix – vector handling and solvers. For the solu-
tion of the pressure equations the BCRSMatrix and the BiCG-Stab solver
from DUNE-Istl have been used. DUNE-Fem provides several implementa-
tions of discrete functions spaces such as Lagrange spaces or Discontinuous
Galerkin spaces. The Discontinuous Galerkin space, i.e. the base functions
and a mapping from local number of degrees of freedom to global number
which is needed to store the data in vectors, have been used for the dis-
cretization of the model problem. Furthermore, DUNE-Fem provides mech-
anisms for projections of data and re-arrangement of memory during adap-
tion. For solving the considered time depended problems the ODE Solvers
implemented in DUNE-Fem or in the software package ParDG (see [13]) are
available.
3.1 Discretization of the Pressure Equation
The equation (9) is discretized by using the Discontinuous Galerkin method.

In [4] a variety of Discontinuous Galerkin methods for elliptic problems of
the form −"u = f are presented and analyzed. For the following numerical
simulation the Oden–Baumann method with polynomial degree 2 has been
chosen. This method has been applied to two-phase flow in porous media for
example in [8] and led to good results. The resulting linear system is stored
using the block wise compressed row storage matrices (BCRSMatrix) from
DUNE-Istl [5]. For preconditioning a block diagonal preconditioner is applied
and the system is solved using the BiCG-Stab solver, both implemented in
DUNE-Istl [5].
3.2 Discretization of the Velocity Equation
The equation (10) is discretized by using the Local Discontinuous Galerkin

method [10]. Consider ϕj with j = 1, ..., N the vectorial basis functions of
the discrete functions space consisting of piecewise polynomial functions of
degree q ∈ {1, 2, 3}. Multiplying (10) with ϕj , integrating equation (10) over
the domain Ω, integration by parts and taking into account that the integral
over Ω can be split into integrals over all grid cells, on a single cell T we
get

−1
λ(s) u · ϕj = −Kp ∇ · ϕj + K p2 n · ϕj , ∀ j = 1, ..., N. (13)
T T ∂T
Here we used that the saturation sw is constant on a single cell and p2 is

a numerical flux as described in [4, Table 3.1]. Here n denotes the out-
ward normal with respect to ∂T . In our computation we have chosen the
LDG flux p2 := {p} − β · [[p]], where {.} denotes the mean value of p on ∂T
and [[p]] the jump in normal direction across ∂T which is a vector paral-
lel to the normal (see [4]). Thereby, β was chosen as β := |∂Tk |10 n where
∂Tk is a face segment of ∂T . In these definitions we follow the notation

in [4]. As the β · [[p]] in the LDG flux works as a penalty term for dis-
continuities, the velocity field has almost continuous normal components
across element intersections. This is a very appealing feature for higher
order discretizations of the saturation and transport equation. In the fol-
lowing numerical examples the polynomial degree for the discretization of
the velocity equation was chosen to be 2. The implementation is based on
the general framework for discretizing evolution equations presented in [9].
The implementation of this framework is part of the DUNE-Fem module
(see [3]).
3.3 Discretization of the Saturation Equation
The saturation equation (11) is also discretized by using the Local Discontin-
uous Galerkin approach described in [10]. Here, as conservative numerical flux
the Engquist-Osher flux is taken. Although in DUNE-Fem higher order LDG
discretizations are implemented up to order 3, this equation is discretized by
using piecewise constant base functions, i.e. polynomial degree 0. The rea-
son is, as the non-linearity of the flux function is self-compressive, also the
first order method produces a satisfactory result. For the time discretization
an explicit or implicit Runge-Kutta solver of order q + 1, is applied, where
q is the polynomial degree of the DG base functions. As for the velocity
equation the discretization of the saturation equation is also implemented us-
ing the general framework for discretizing evolution equations presented in
[9]. The used implicit ODE solvers are part of the DUNE module DUNE-
Fem.
3.4 Discretization of the Transport Equation
The transport equation is discretized in the same way as the saturation equa-
tion. The only difference is that in this case a simple linear upwind flux can
be chosen as the numerical flux. Due to the fact that we have a linear flux
function the polynomial degree of the base functions can be chosen larger
than 0. Although higher order LDG discretizations for this type of equa-
tion are stable, one can get oscillations which are even higher if the dis-
continuities of the velocity field are strong. To overcome this problem here
the factor β is chosen sufficiently large which then works as a penalty term
for jumps of the velocity u in normal direction across cell boundaries. In
the following numerical example the polynomial order of the DG space for
the transport equations is also 0. For the time discretization an explicit or
implicit Runge-Kutta solver of order q + 1 is applied, where q is the poly-
nomial degree of the DG base functions. Again the discretization uses the
general framework for discretizing evolution equations presented as in [9].
Also the same implicit ODE solvers which are implemented in DUNE-Fem
are used.
3.5 Operator Splitting
Since the equations (9) to (12) are non-linearly coupled, we apply an operator
splitting to decouple the equations and solve each equation separately. This
is done as follows. Assume that the unknowns sw and ck , k = O2 , H2 O are
given. Then one time step is solved as follows
1. for given sw the pressure p can be calculated using equation (9),
2. for given saturation sw and given pressure p, the velocity u can be calcu-
lated using equation (10),
3. for given saturation sw , given pressure p, given velocity u, and given con-
centration c the new saturation can be calculated using equa-
tion (11),
4. for given saturation sw , given pressure p, given velocity u, and given
concentration c the new concentration of species can be calculated using
equation (12).
3.6 Parallelization
The parallelization follows the concept of single program multiple data, mean-
ing that one and the same program is executed on multiple processors. The
computational domain is distributed via domain decomposition. Here the
graph partitioner METIS has been applied to calculate these distributions.
Due to the distribution of the data to multiple processes communication
between processes sharing data is necessary during computation of the un-
knowns. Here the DG methods have the nice property to be local meth-
ods, which means that during calculation only neighbor elements have to
be available. For parallelization this is a very appealing feature. Furthermore,
due to the discontinuity of the methods only element data have to be com-
municated to neighboring cells located on other processes. All communica-
tions during the solution process therefore are interior – ghost communica-
tions.
Pressure Equation:
One communication for each iteration step of the BiCG-stab solver is neces-
sary. Furthermore, each evaluation of a scalar product within the solver needs
a global sum operation.
Velocity Equation:
One communication after calculation of the velocity is applied.
Saturation and Transport Equation:

One communication during each iteration step of the ODE solver is needed.
3.7 Adaptivity and Load Balancing
During the computation of the unknowns p, u, sw , and c an error indicator

is evaluated to monitor the local errors introduced by a too coarse grid.
Error Indicators:
For the pressure and velocity together this indicator consists of the jump
of the DG velocity in normal direction across cell boundaries in the nor-
mal direction. For the saturation equation and the transport equations we
use the error indicators described in [15]. In [15] an a-posteriori error esti-
mator is developed for advection-diffusion problems with source terms dis-
cretized by an implicit finite volume scheme. Although, the theoretical proof
only holds for this type of discretization for weakly coupled systems, the de-
scribed local error indicators work very well also for similar discretization
schemes.
Marking Strategy:
Given a global adaptation tolerance the local cell tolerance is obtained by an
equi-distribution strategy. Cells where the local indicator violates the local
tolerance are marked for refinement. Cells, where the local indicator is 100
times smaller then the local tolerance are marked for coarsening.
Adaptation and Load Balancing:

After marking of elements is done a grid adaptation is performed. Elements
that are marked for refinement are refined. Elements that were marked for
coarsening are coarsened if possible. After each adaptation step the load of
each processor is checked. For each macro grid cell the load consists mainly
of the number of leaf cells that have its source in the macro cell. If the given
imbalance tolerance is violated, a re-balancing, i.e. re-construction of a new
well balanced distribution of the cells to the processors, is performed. This
process is described in detail in [12].
4 Numerical Results
In this section simulation results for a model problem are presented. The
model problem consists of water and gas transport with phase transition and
reaction within the cathodic gas diffusion layer of a PEM fuel cell.
4.1 Geometry of the Model Problem

3 −4
4 3 −4
4
As
3 the computational
4 3 domain we consider Ω := 0, 2 10 × 0, 6 10 ×
−4
0, 2 10 m in three space dimensions.
Fig. 1. Sketch of the computational domain
The boundaries are defined as follows
Γ1 := ∂Ω \ (Γ2 ∪ Γ3 ∪ Γ4 ),
3 4 3 4
Γ2 := {0} × 0, 2 · 10−4 × 0, 2 · 10−4 m2 ,
3 4 3 4
Γ3 := {0} × 4 · 10−4 , 6 · 10−4 × 0, 2 · 10−4 m2 ,
3 4 3 4
Γ4 := 2 · 10−4 × 4 · 10−4 , 6 · 10−4 × 0, 2 · 10−4 m2 .
The domain Ω with boundaries Γ1 , ..., Γ4 represents the GDL of a PEM fuel
cell. Figure 1 shows a sketch of the computational domain.
4.2 Boundary Conditions and Initial Values
On the boundaries Γ1 , ..., Γ4 the following boundary conditions were defined

Γ1 : (No-flow boundary – bipolar plate)
Two-phase flow: no-flux boundary condition.
Transport equation: no-flow boundary condition.
Γ2 : (Inflow boundary – channel)
Two-phase flow: sw = 0.0, p = p1cha .
Transport equation: cH2 O = 0.2, cO2 = 0.8.
Γ3 : (Outflow boundary – channel)
Two-phase flow: sw = 0.0, p = p2cha .
Transport equation: Outflow boundary condition.
Γ4 : (Catalyst layer)
Two-phase flow: sw = 1.0, no-flow boundary condition for p.
Transport equation no-flow boundary condition.
The following initial values have been chosen
sw (x, 0) = s0 (x) = 0.1 in Ω.

c(x, 0) = c0 (x) = (cH
0
2O
, cO2 T
0 ) in Ω with cH
0
2O
= 0.2, cO2
0 = 0.8.
4.3 Technical Data of the Simulation
The following numerical simulations where performed on the XC4000 Linux

Cluster of the Scientific Supercomputing Center of the University Karlsruhe
with 32 processors. Table 2 shows the discretization method and polynomial
order for equations (9)–(12). In addition, the resulting number of degrees of
freedom (DOF) per element is shown.
To solve the resulting linear system from equation (9), a BiCG-Stab solver
and a block diagonal pre-conditioner were applied. To solve the saturation
equation (11) and the transport equation (12), an implicit ODE solver of or-
der 1 has been used. The macro grid contains 98.304 hexahedrons, 32 in x and
z-direction and 96 in y-direction. The edge length of one macro hexahedron
is 6.25μm. The resulting dynamically adapted mesh contains 543.448 hexa-
hedrons. Thus, 43 DOFs per element lead to an overall number of 23.368.364
degrees of freedom per time step. To store the resulting matrix of the lin-
ear system from equation (9) a 10 × 10 block matrix has to be stored per
element. This leads to an additional memory requirement for 54.344.800
DOFs.
The simulation took approximately 2 days and 4613 time steps were cal-
culated. To compute one time step approximately 25 seconds were needed.
Part of the time is spend in setting up the matrix and in the linear solver
which took about 8.5 seconds each time step. The solution of the saturation
equation took approximately 4.5 seconds. The transport equations needed 11
seconds and the adaptation step took 0.2 seconds. The missing difference was
spend in other parts of the program or in waiting for other processes. The
load balancing process costs between 10 and 20 seconds but is applied very
rarely.
Table 2. Discretization method, polynomial order and degrees of freedom per ele-
ment for the model problem.
Equation Discr. Method pol. deg. q DOFs

pressure equation (9) Oden-Baumann 2 10
velocity equation (10) LDG 2 30
saturation equation (11) LDG 0 1
transport equation (12) LDG 0 2
4.4 Simulation Results
Figures 2, 3 show snapshots of the solution of the model problem taken at time
step 4613 which corresponds to simulation time T = 0.00022. In addition,
Figs. 4 and 5 show the time evolution of the saturation distribution from
T = 0.00011 to T = 0.00022.
In Fig. 2, on the left hand side the pressure distribution is shown. As
expected due the choice of the boundary values there is a continuous pressure
drop from Γ2 to Γ3 . The refinement level of the adapted mesh is shown in the
Fig. 2. Level of refinement (left), partitioning of the grid (middle), and pressure
distribution (right) at computational time T = 0.00022
Fig. 3. Global velocity u. Components ux (left), uy (middle), and uz (right) at

computational time T = 0.00022
middle. One can see that the higher levels (red color) are located in the area
where Γ2 and Γ3 are connected to Γ1 on the left hand side of the geometry.
Finally, on the right hand side of Fig. 2 the partitioning of the grid for the 32
processors is shown.
In Fig. 3 the components of the global velocity u are shown. One can see
that in the areas where the velocity has strong variations, the grid is refined
(see Fig. 2, middle). We state that the error indicator applied to monitor the
velocity errors works well. In the middle of Fig. 3 also the influence of the water
saturation to the y-component uy of the velocity field near the membrane can
be seen. The velocity is reduced in this area due to the presence of both
phases, liquid water and gas. On the right side of Fig. 3 the z-component of
the velocity is shown which is as expected close to zero.
The saturation sw at time T = 0.00011 is illustrated on the left side of
Fig. 4. The initial value was 0.1. One can see that phase transition has taken
place in the area near the gas channels, as only dry air is entering the cell
at Γ2 . Water is entering from the side of the membrane (Γ4 ) . The water is
slowly moving towards the gas channels. Also for the saturation and transport
equation the applied error indicator is able to monitor the zones of higher
activity (see Fig. 2, left side). On the middle part of Fig. 4 the concentration
of water cH2 O in the gaseous phase is shown. In the area where saturation is
non-zero, phase transition takes place. The mass concentration of water in the
gas phase is increased due to vaporization. This, on the other hand decreases
the concentration of oxygen, as both concentrations should sum up to 1. The
middle and right picture also demonstrate that the concentrations sum up
nicely to 1 as required.
Fig. 4. Saturation sw (left), concentration of water cH2 O (middle), and concentra-

tion of oxygen cO2 (right) at computational time T = 0.00011
Fig. 5. Saturation sw (left), concentration of water cH2 O (middle), and concentra-

tion of oxygen cO2 (right) at computational time T = 0.00022
Figure 5 shows the same variables as Fig. 4 but at a later time, i.e. T =
0.00022. One can see that more of the water has been vaporized but also that
more and more dry air is covering the left side of the cathodic gas diffusion
layer. From Γ4 more and more water is entering the gas diffusion layer and
moving towards the gas channel.
5 Conclusion and Outlook

We could show that our developed software is able to handle complex time-
dependent models in three space dimension. Simulations done in 3d usually
lead to a large number of unknowns. To cope with such large systems, the
simulation tool can be used in parallel. In order to increase the efficiency while
keeping the accuracy of the simulation, local grid adaptivity and dynamic load
balancing is included.
The implementation of the detailed fuel cell model described in [16] will
be subject of future work. Furthermore, the use of higher order methods will
be extended to the transport equations.
References
1. ALUGrid.
http://www.mathematik.uni-freiburg.de/IAM/Research/alugrid/.
2. DUNE. http://www.dune-project.org.
3. DUNE-Fem.
http://www.mathematik.uni-freiburg.de/IAM/Research/projectskr/
dune/feminfo.html.
4. D.N. Arnold, F. Brezzi, B. Cockburn, and L.D. Marini. Unified analysis of
discontinuous galerkin methods for elliptic problems. SIAM J. Num. Anal,
39:1749–1779, 2002.
5. P. Bastian and M. Blatt. The Iterative Template Solver Library. In Proc. of
the Workshop on State-of-the-Art in Scientific and Parallel Computing, PARA
’06. Springer, 2006.
6. P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, R. Kornhuber,
M. Ohlberger, and O. Sander. A Generic Grid Interface for Parallel and Adaptive
Scientific Computing. Part II: Implementation and Tests in DUNE. Preprint
404, DFG Research Center MATHEON, 2007.
7. P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, M. Ohlberger, and
O. Sander. A Generic Grid Interface for Parallel and Adaptive Scientific Com-
puting. Part I: Abstract Framework. Preprint 403, DFG Research Center
MATHEON, 2007.
8. P. Bastian and B. Riviere. Discontinuous galerkin methods for two-phase flow
in porous media. Technical Report 28, IWR (SFB 359), Universität Heidelberg,
2004.
9. A. Burri, A. Dedner, D. Diehl, R. Klöfkorn, and M. Ohlberger. A general object
oriented framework for discretizing nonlinear evolution equations. In Proc. of
The 1st Kazakh-German Advanced Research Workshop on Computational Sci-
ence and High Performance Computing, 2005.
10. B. Cockburn and C.-W. Shu. The Local Discontinuous Galerkin Method
for Time-Dependent Convection-Diffusion Systems. SIAM J. Numer. Anal.,
35(6):2440–2463, 1998.
11. M.Y. Corapcioglu and A. Baehr. A compositional multiphase model for ground-
water contamination by petroleum products: 1. theoretical considerations. Wa-
ter Resource Research, 1987.
12. A. Dedner, C. Rohde, B. Schupp, and M. Wesenberg. A parallel, load bal-
anced mhd code on locally adapted, unstructured grids in 3d. Computing and
Visualization in Science, 7:79–96, 2004.
13. D. Diehl. Higher Order Schemes for Simulation of Compressible Liquid – Vapor
Flows with Phase Change. Dissertation, Institute of Mathematics, University
Freiburg, 2007.
14. R. Helmig. Multiphase Flow and Transport Processes in the Subsurface: A con-
tribution to the modeling of hydrosystems. Springer, Berlin, Heidelberg, 1997.
15. M. Ohlberger and C. Rohde. Adaptive finite volume approximations for
weakly coupled convection dominated parabolic systems. IMA J. Numer. Anal.,
22(2):253–280, 2002.
16. K. Steinkamp, J. Schumacher, F. Goldsmith, M. Ohlberger, and C. Ziegler.
A non-isothermal pem fuel cell model including two water transport mechanisms
in the membrane. Preprint 4, Mathematisches Institut, Universität Freiburg,
2007.
Part VI
Risk Management in Finance and Insurance

Advanced Credit Portfolio Modeling
and CDO Pricing
Ernst Eberlein1 , Rüdiger Frey2 , and Ernst August von Hammerstein1

1
Department of Mathematical Stochastics, University of Freiburg, Eckerstraße 1,
79104 Freiburg, Germany
{eberlein,hammerstein}@stochastik.uni-freiburg.de
2
Department of Mathematics, University of Leipzig, 04081 Leipzig, Germany
ruediger.frey@math.uni-leipzig.de
1 Introduction
Credit risk represents by far the biggest risk in the activities of a traditional
bank. In particular, during recession periods financial institutions loose enor-
mous amounts as a consequence of bad loans and default events. Traditionally
the risk arising from a loan contract could not be transferred and remained
in the books of the lending institution until maturity. This has changed com-
pletely since the introduction of credit derivatives such as credit default swaps
(CDSs) and collaterized debt obligations (CDOs) roughly fifteen years ago.
The volume in trading these products at the exchanges and directly between
individual parties (OTC) has increased enormously. This success is due to the
fact that credit derivatives allow the transfer of credit risk to a larger commu-
nity of investors. The risk profile of a bank can now be shaped according to
specified limits, and concentrations of risk caused by geographic and industry
sector factors can be reduced.
However, credit derivatives are complex products, and a sound risk-man-
agement methodology based on appropriate quantitative models is needed
to judge and control the risks involved in a portfolio of such instruments.
Quantitative approaches are particularly important in order to understand the
risks involved in portfolio products such as CDOs. Here we need mathematical
models which allow to derive the statistical distribution of portfolio losses.
This distribution is influenced by the default probabilities of the individual
instruments in the portfolio, and, more importantly, by the joint behaviour
of the components of the portfolio. Therefore the probabilistic dependence
structure of default events has to be modeled appropriately.
In this paper we use two different approaches for modeling dependence.
To begin with, we extend the factor model approach of Vasiček [32, 33] by
using more sophisticated distributions for the factors. Due to their greater
flexibility these distributions have been successfully used in several areas of
finance (see e.g. [9, 10, 11]). As shown in the present paper, this approach
254 E. Eberlein, R. Frey, E. A. von Hammerstein
leads to a substantial improvement of performance in the pricing of synthetic

CDO tranches. Moreover, in the last section we introduce a dynamic Markov
chain model for the default state of a credit portfolio and discuss the pricing
of CDO tranches for this model.
2 CDOs: Basic Concepts and Modeling Approaches

A collateralized debt obligation (CDO) is a structured product based on an un-
derlying portfolio of reference entities subject to credit risk, such as corporate
bonds, mortgages, loans or credit derivatives. Although several types of CDOs
are traded in the market which mainly differ in the content of the portfolio
and the cash flows between counterparties, the basic structure is the same.
The originator (usually a bank) sells the assets of the portfolio to a so-called
special purpose vehicle (SPV), a company which is set up only for the purpose
of carrying out the securitization and the necessary transactions. The SPV
does not need capital itself, instead it issues notes to finance the acquisition
of the assets. Each note belongs to a certain loss piece or tranche after the
portfolio has been divided into a number of them. Consequently the portfolio
is no longer regarded as an asset pool but as a collateral pool. The tranches
have different seniorities; the first loss piece or equity tranche has the low-
est, followed by junior mezzanine, mezzanine, senior and finally super-senior
tranches. The interest payments the SPV has to make to the buyer of a CDO
tranche are financed from the cash flow generated by the collateral pool.
Therefore the performance or the default risk of the portfolio is taken over by
the investors. Since all liabilities of the SPV as a tranche seller are funded by
proceeds from the portfolio, CDOs can be regarded as a subclass of so-called
asset-backed securities. If the assets consist mainly of bonds resp. loans, the
CDO is also called collateralized bond obligation (CBO) resp. collateralized
loan obligation (CLO). For a synthetic CDO which we shall discuss in more
detail below, the portfolio contains only credit default swaps. The motivation
to build a CDO is given by economic reasons:
• By selling the assets to the SPV, the originator removes them from his
balance sheet and therefore he is able to reduce his regulatory capital. The
capital which is set free can then be used for new business opportunities.
• The proceeds from the sale of the CDO tranches are typically higher than
the initial value of the asset portfolio because the risk-return profile of the
tranches is more attractive for investors. This is both the result from and
the reason for slicing the portfolio into tranches and the implicit collation
and rebalancing hereby. Arbitrage CDOs are mainly set up to exploit this
difference.
In general, CDO contracts can be quite sophisticated because there are no
regulations for the compilation of the reference portfolio and its tranching or
the payments to be made between the parties. The originator and the SPV can
Advanced Credit Portfolio Modeling and CDO Pricing 255
design the contract in a taylormade way, depending on the purposes they want
to achieve. To avoid unnecessary complications, we concentrate in the follow-
ing on synthetic CDOs which are based on a portfolio of credit default swaps.
2.1 Structure and Payoffs of CDSs and Synthetic CDOs

As mentioned before, the reference portfolio of a synthetic CDO consists en-
tirely of credit default swaps (CDSs). These are insurance contracts protecting
from losses caused by default of defaultable assets. The protection buyer A pe-
riodically pays a fixed premium to the protection seller B until a prespecified
credit event occurs or the contract terminates. In turn, B makes a payment
to A that covers his losses if the credit event has happened during the life-
time of the contract. Since there are many possibilities to specify the default
event as well as the default payment, different types of CDSs are traded in
the market, depending on the terms the counterparties have agreed on. The
basic structure is shown in Fig. 1.
Throughout this article we will make the following assumptions: The ref-
erence entity of the CDS is a defaultable bond with nominal value L, and the
credit event is the default of the bond issuer. If default has happened, B pays
(1 − R)L to A where R denotes the recovery rate. On the other side A pays
quarterly a fixed premium of 0.25rCDS L where rCDS is the annualized fair
CDS rate. To determine this rate explicitly, we fix some notation:
r is the riskless interest rate, assumed to be constant over the lifetime
[0, T ] of the CDS,
u(t) is the discounted value of all premiums paid up to time t when the
annualized premium is standardized to 1,
G1 (t) is the distribution function of the default time T1 with corresponding
density g1 (t) (its existence will be justified by the assumptions in
subsequent sections).
The expected value of the discounted premiums (premium leg) can then be
written as
T
P L(rCDS ) = rCDS L u(t)g1 (t) dt + rCDS L u(T )(1 − G1 (T )) .
0
The expected discounted default payment (default leg) is given by
T
D = (1 − R)L g1 (t)e−rt dt .
0
Fig. 1. Basic structure of a CDS

Fig. 2. Schematic representation of the payments in a synthetic CDO. The choice

of the attachment points corresponds to DJ iTraxx Europe standard tranches
The no-arbitrage condition P L(rCDS ) = D then implies

T
(1 − R) 0 g1 (t)e−rt dt D
rCDS = T = . (1)
0 u(t)g 1 (t) dt + u(T )(1 − G1 (T )) P L(1)
To explain the structure and the cash flows of a synthetic CDO assume that
its reference portfolio consists of N different CDSs with the same notional
value L. We divide this portfolio in subsequent tranches. Each tranche covers
a certain range of percentage losses of the total portfolio value N L defined
by lower and upper attachment points Kl , Ku ≤ 1. The buyer of a tranche
compensates as protection seller for all losses that exceed the amount of Kl N L
up to a maximum of Ku N L. On the other hand the SPV as protection buyer
has to make quarterly payments of 0.25rc Vt , where Vt is the notional value of
the tranche at payment date t. Note that Vt starts with N L(Ku − Kl ) and is
reduced by every default that hits the tranche. rc is the fair tranche rate. See
also Fig. 2.
In recent years a new and simplified way of buying and selling CDO
tranches has become very popular, the trading of single index tranches. For
this purpose standardized portfolios and tranches are defined. Two counter-
parties can agree to buy and sell protection on an individual tranche and
exchange the cash flows shown in the right half of Fig. 2. The underlying
CDS portfolio however is never physically created, it is merely a reference
portfolio from which the cash flows are derived. So the left hand side of Fig. 2
vanishes in this case, and the SPV is replaced by the protection buyer. The
portfolios for the two most traded indices, the Dow Jones CDX NA IG and
the Dow Jones iTraxx Europe, are composed of 125 investment grade US
and European firms respectively. The index itself is nothing but the weighted
credit default swap spread of the reference portfolio. In Sects. 2.2 and 3.1 we
shall derive the corresponding default probabilities. We will use market quotes
for different iTraxx tranches and maturities to calibrate our models later in
Sects. 3.2 and 4.2.
In the following we denote the attachment points by 0 = K0 < K1 < · · · <
Km ≤ 1 such that the lower and upper attachment points of tranche i are
Ki−1 and Ki respectively. Suppose for example that (1 − R)j = Ki−1 N and
(1 − R)k = Ki N for some j < k, j, k ∈ N. Then the protection seller B of
tranche i pays (1−R)L if the (j +1)st reference entity in the portfolio defaults.
For each of the following possible k−j−1 defaults the protection buyer receives
the same amount from B. After the k th default occurred the outstanding
notional of the tranche is zero and the contract terminates. However, the losses
will usually not match the attachment points. In general, some of them are
divided up between subsequent tranches: If (j−1)(1−R) < Ki < j(1−R) for some
(j−1)(1−R)
N N
j ∈ N, then tranche i bears a loss of N L Ki − N (and is exhausted
thereafter) if the j th default occurs. The overshoot is absorbed by the following

tranche whose outstanding notional is reduced by N L j(1−R) N − Ki . We use
the following notation:
Ki−1 , Ki are the lower/upper attachment points of tranche i,
Zt is the relative amount of CDSs which have defaulted up to time t,
expressed as a fraction of the total number N ,
Lit = min[(1 − R)Zt , Ki ] − min[(1 − R)Zt , Ki−1 ] is the loss of tranche i
up to time t, expressed as a fraction of the total notional value N L,
ri is the fair spread rate of tranche i,
0 = t0 < · · · < tn are the payment dates of protection buyer and seller,
β(t0 , tk ) is the discount factor for time tk .
Remark 1. Under the assumption of a constant riskless interest rate r we

would have β(t0 , tk ) = e−rtk . Since this assumption is too restrictive one
uses zero coupon bond prices for discounting instead. Therefore β(t0 , tk ) will
denote the price at time t0 of a zero coupon bond with maturity tk .
The assumption that all CDSs have the same notional value may seem
somewhat artificial, but it is fulfilled for CDOs on standardized portfolios like
the Dow Jones CDX or the iTraxx Europe.
With this notation the premium as well as the default leg of tranche i can be
expressed as

n
3 4
P Li (ri ) = (tk − tk−1 )β(t0 , tk ) ri E Ki − Ki−1 − Litk N L ,
k=1
(2)

n
3 4
Di = β(t0 , tk )E Litk − Litk−1 N L ,
k=1
where E[·] denotes expectation. For the fair spread rate one obtains
%n 3 i 4 3 i 4
k=1 β(t0 , tk ) E Ltk − E Ltk−1
ri = %n 3 i 4 . (3)
k=1 (tk − tk−1 )β(t0 , tk ) Ki − Ki−1 − E Ltk
Remark 2. To get arbitrage-free prices, all expectations above have to be taken

under a risk neutral probability measure, which is assumed implicitly. One
should be aware that risk neutral probabilities cannot be estimated from his-
torical default data.
Since payment dates and attachment points are specified in the CDO contract
and discount factors can be obtained from the market, the remaining task is
to develop a realistic portfolio model from which the risk neutral distribution
of Zt can be derived, i.e. we need to model the joint distribution of the default
times T1 , . . . , TN of the reference entities.
2.2 Factor Models with Normal Distibutions
To construct this joint distribution, the first step is to define the marginal
distributions Qi (t) = P (Ti ≤ t). The standard approach, which was proposed
in [21], is to assume that the default times Ti are exponentially distributed,
that is, Qi (t) = 1 − e−λi t . The default intensities λi can be estimated from the
i
clean spreads rCDS /(1 − R) where rCDSi
is the fair CDS spread of firm i which
can be derived using formula (1). In fact, the relationship λi ≈ rCDS i
/(1 − R)
is obtained directly from (1) by inserting the default density g1 (t) = λi e−λi t
(see [22, section 9.3.3]).
As mentioned before, the CDX and iTraxx indices quote an average CDS
spread for the whole portfolio in basis points (100bp = 1 %), therefore the
market convention is to set
sa
λi ≡ λa = (4)
(1 − R)10000
where sa is the average CDX or iTraxx spread in basis points. This implies that
all firms in the portfolio have the same default probability. One can criticize
this assumption from a theoretical point of view, but it simplifies and fastens
the calculation of the loss distribution considerably as we will see below. Since
λa is obtained from data of derivative markets, it can be considered as a risk
neutral parameter and therefore the Qi (t) can be considered as risk neutral
probability distributions.
The second step to obtain the joint distribution of the default times is
to impose a suitable coupling between the marginals. Since all firms are sub-
ject to the same economic environment and many of them are linked by di-
rect business relations, the assumption of independence of defaults between
different firms obviously is not realistic. The empirically observed occur-
rence of disproportionally many defaults in certain time periods also con-
tradicts the independence assumption. Therefore the main task in credit
portfolio modeling is to implement a realistic dependence structure which
generates loss distributions that are consistent with market observations.
The following approach goes back to [32] and was motivated by the Mer-
ton model [25].
For each CDS in the CDO portfolio we define a random variable Xi as

follows:
√
Xi := ρ M + 1 − ρ Zi , 0 ≤ ρ < 1, i = 1, . . . , N, (5)
where M, Z1 , . . . , ZN are independent and standard normally distributed. Ob-

viously Xi ∼ N (0, 1) and Corr(Xi , Xj ) = ρ, i = j. Xi can be interpreted as
state variable for the firm that issued the bond which CDS i secures. The
state is driven by two factors: the systematic factor M represents the macro-
economic environment to which all firms are exposed, whereas the idiosyn-
cratic factor Zi incorporates firm specific strengths or weaknesses.
To model the individual defaults, we define time-dependent thresholds by
di (t) := Φ−1 (Qi (t))
where Φ−1 (x) denotes the inverse of the standard normal distribution function
or quantile function of N (0, 1). Observe that the di (t) are increasing because so
are Φ−1 and Qi . Therefore we can define each default time Ti as the first time
point at which the corresponding variable Xi is smaller than the threshold
di (t), that is
Ti := inf{t ≥ 0 | Xi ≤ di (t)}, i = 1, . . . , N. (6)
This also ensures that the Ti have the desired distribution, because

P (Ti ≤ t) = P Xi ≤ Φ−1 (Qi (t)) = P Φ(Xi ) ≤ Qi (t) = Qi (t),
where the last equation follows from the fact that the random variable Φ(Xi )
is uniformly distributed on the interval [0, 1]. Moreover, the leftmost equation
shows that Ti = Q−1
d
i (Φ(Xi )), so the default times inherit the dependence
structure of the Xi . Since the latter are not observable, but serve only as
auxiliary variables to construct dependence, such models are termed ‘latent
variable’ models. Note that by (4) we have Qi (t) ≡ Q(t) and thus di (t) ≡ d(t),
therefore we omit the index i in the following.
Remark 3. Instead of inducing dependence by latent variables that are linked
by the factor equation (5), one can also define the dependence structure of the
default times more directly by inserting the marginal distribution functions
into an appropriately chosen copula. We do not discuss this approach here
further, but give some references at the end of Sect. 2.3.
To derive the loss distribution let Atk be the event that exactly k defaults have
happened up to time t. From (6) and (5) we get
√
d(t) − ρM
P (Ti < t |M ) = P (Xi < d(t) |M ) = Φ √ .
1−ρ
Since the Xi are independent conditional on M , the conditional probability

P (Atk |M ) equals the probability of a binomial distribution with parameters
N and p = P (Ti < t |M ):
√ k √ N −k
N d(t) − ρM d(t) − ρM
P (Atk |M ) = Φ √ 1−Φ √ .
k 1−ρ 1−ρ
The probability that at time t the relative number of defaults Zt does not
exceed q is

[N q]
FZt (q) = P (Atk )
k=0
N d(t) − √ρu k
∞ [N q] √
d(t) − ρ u N −k
= Φ √ 1−Φ √ dPM (u) .
k 1−ρ 1−ρ
−∞ k=0
If the portfolio is very large, one can simplify FZt further using the following
approximation which was introduced in [33] and which is known as√large
d(t)− ρ M
homogeneous portfolio (LHP) approximation. Let pt (M ) := Φ √
1−ρ
and Gpt be the corresponding distribution function, then we can rewrite FZt
in the following way:
1 [N q]
N
FZt (q) = sk (1 − s)N −k dGpt (s). (7)
k
0 k=0
Applying the LHP approximation means that we have to determine the be-
haviour of the integrand for N → ∞. For this purpose suppose that Yi
are independent and identically distributed (iid) Bernoulli variables with
P (Yi = 1) = s = 1 − P (Yi = 0). Then the strong law of large numbers
%N
states that ȲN = N1 i=1 Yi → s almost surely which implies convergence of
the distribution functions FȲN (x) → [0,x] (s) pointwise on R \ {s}. For all
q = s we thus have
[N q]

N
N

sk (1 − s)N −k = P Yi ≤ N q = P ȲN ≤ q −→ [0,q] (s).
k N →∞
k=0 i=1
Since the sum on the left hand side is bounded by 1, by Lebesgue’s theorem
we get from (7)
1 √
1 − ρ Φ−1 (q) − d(t)
FZt (q) ≈ [0,q] (s) dGpt (s) = Gpt (q) = P − √ ≤ M
0 ρ
√ −1

1 − ρ Φ (q) − d(t)
= Φ √ (8)
ρ
where in the last equation the symmetry relation 1 − Φ(x) = Φ(−x) has
been used. This distribution is, together with the above assumptions, the
current market standard for the calculation of CDO spreads according to
equation (3). Since 3the relative
4 portfolio loss up to time t is given by (1−R)Zt ,
the expectations E Litk contained in (3) can be written as follows:
Ki
1−R ∧1
3 4 3 Ki 4
E Litk = Ki−1
(1−R) q− K
1−R dFZtk (q)+(Ki −Ki−1 ) 1−FZtk 1−R ∧1 .
i−1
1−R ∧1
(9)
2.3 Deficiencies and Extensions

The pricing formula obtained from (3), (8) and (9) contains one unknown
quantity: the correlation parameter ρ. This parameter has to be estimated
before one can derive the fair rate of a CDO tranche. A priori it is not clear
which data and which estimation procedure one could use to get ρ. In the Mer-
ton approach, defaults are driven by the evolution of the asset value of a firm.
Consequently the dependence between defaults is derived from the depen-
dence between asset values. The latter cannot be observed directly, therefore
some practitioners have used equity correlations, which can be estimated from
stock price data. A more direct and plausible alternative would be to infer
correlations from historical default data, but since default is a rare event, this
would require data sets over very long time periods which are usually not
available.
With the development of a liquid market for single index tranches in the
last years, a new source of correlation information has arisen: the implied
correlations from index tranche prices. Similar to the determination of implied
volatilities from option prices by inverting the Black–Scholes formula, one can
invert the above pricing formula and solve numerically for the correlation
parameter ρ which reproduces the quoted market price. This provides also
a method to examine if the model and its assumptions are appropriate. If this
is the case, the correlations derived from market prices of different tranches of
the same index should coincide. However, in reality one observes a so-called
correlation smile: the implied correlations of the equity and (super-)senior
tranches are typically much higher than those of the mezzanine tranches.
See Fig. 3 for an example. The smile indicates that the classical model is
not flexible enough to generate realistic dependence structures. This is only
partly due to the simplifications made by using the LHP approach. The deeper
reason for this phenomenon lies in the fact that the model with normal factors
strongly underestimates the probabilities of joint defaults. This has led to
severe mispricings and inadequate risk forecasts in the past. The problem
became evident in the so-called correlation crisis in May 2005: the factor
model based on normal distributions was unable to follow the movement of
market quotes occuring in reaction to the downgrading of Ford and General
Motors to non-investment grade.
Fig. 3. Implied correlations calculated from the prices of DJ iTraxx Europe standard
tranches at November 13, 2006, for different maturities T
A number of different approaches for dealing with this problem have been
investigated. A rather intuitive extension to remedy the deficiencies of the
normal factor model which we shall exploit in Sect. 3, is to allow for factor
distributions which are much more flexible than the standard normal ones.
Different factor distributions do not only change the shape of FZt , but also
have a great influence on the so-called factor copula implicitly contained in the
joint distribution of the latent variables. In fact, the replacement of the normal
distribution leads to a fundamental modification of the dependence structure
which becomes much more complex and can even exhibit tail-dependence.
A necessary condition for the latter to hold is that the distribution of the sys-
tematic factor M is heavy tailed. This fact was proven in [24]. The first paper
in which alternative factor distributions are used is [17] where both factors
are assumed to follow a Student t-distribution with 5 degrees of freedom. In
[19], Normal Inverse Gaussian distributions are applied for pricing synthetic
CDOs, and in [1] several models based on Gamma, Inverse Gaussian, Variance
Gamma, Normal Inverse Gaussian and Meixner distributions are presented.
In the last paper the systematic and idiosyncratic factors are represented by
the values of a suitably scaled and shifted Lévy process at times ρ and 1 − ρ.
Another way to extend the classical model is to implement stochastic cor-
relations and random factor loadings. In the first approach which was de-
veloped in [15], the constant correlation parameter ρ in (5) is replaced by
a random variable taking values in [0, 1]. The cumulative default distribu-
tion can then be derived similarly as before, but one has to condition on
both, the systematic factor and the correlation variable. The concept of ran-
dom factor loadings was first published in [2]. There the Xi are defined by
Xi := mi (M ) + σi (M )Zi with some deterministic functions mi and σi . In the

simplest case Xi = m + (l{M<e} + h{M≥e} )M + νZi where l, h, e ∈ R are
additional parameters and m, ν are constants chosen such that E[Xi ] = 0 and
Var[Xi ] = 1. Further information and numerical details for the calibration of
such models to market data can be found in [8].
As already mentioned in Remark 3, other approaches use copula models to
define the dependence between the default times Ti . The concept of copulas
was introduced in probability theory by Sklar [31]. A very useful and illustra-
tive introduction to copulas and their application in risk management can be
found in [22, chapter 5], for a thorough theoretical treatment, we refer to [26].
The first papers where copulas were used in credit risk models are [21] and
[30]. A recent approach based on Archimedean copulas can be found in [5]. The
pricing performance of models with Clayton and Marshall–Olkin copulas was
investigated and compared with some other popular approaches in [7]. There
the prices calculated from the Clayton copula model showed a slightly better
fit to the market quotes, but they were still relatively close to those generated
by the Gaussian model. The Marshall–Olkin copulas performed worse, since
the deviations from market prices were greater than those of other models
considered.
Alternatively, it is possible to come up with stochastic models for the
dynamic evolution of the default state of the portfolio (instead of modeling
just the distribution of the default times as seen from a given point in time
t) and to look for dynamic models that can generate correlation skews. An
example of this line of research is discussed in Sect. 4.
3 Calibration with Advanced Distributions

The factor distributions we implement to overcome the deficiencies mentioned
above belong to the class of generalized hyperbolic distributions (GH) which
was introduced in [4]. In the general case, their densities are given by
(λ− 12 )/2 β(x−μ)
dGH(λ,α,β,δ,μ) (x) = a(λ, α, β, δ, μ) δ 2 + (x − μ)2 e
(10)
× Kλ− 12 α δ 2 + (x − μ)2
with the norming constant

λ
(α2 − β 2 ) 2
a(λ, α, β, δ, μ) = √ 1
.
2π αλ− 2 δ λ Kλ (δ α2 − β 2 )
Kν denotes the modified Bessel function of the third kind with index ν and
GH(λ, α, β, δ, μ) the corresponding probability distribution. The influence of
the parameters is as follows: α > 0 determines the shape, 0 ≤ |β| < α the
skewness, μ ∈ R is a location parameter and δ > 0 serves for scaling. λ ∈ R
characterizes certain subclasses and has considerable influence on the size of
Fig. 4. Influence of the GH parameters β (left) and λ (right), where on the right
hand side log densities are plotted
mass contained in the tails which can be seen from the asymptotic behaviour of
the densities: dGH(λ,α,β,δ,μ) (x) ∼ |x|λ−1 e−α|x|+βx for |x| → ∞. See also Fig. 4.
Generalized hyperbolic distributions have already been shown to be a very
useful tool in various fields of mathematical finance. An overview over different
applications can be found in [9]. Let us mention some special subclasses and
limiting cases which are of particular interest and which we will use later to
calibrate the iTraxx data:
For λ = −0.5 one obtains the subclass of Normal Inverse Gaussian distri-
butions (NIG) with densities

αδ K1 α δ 2 + (x − μ)2 δ√α2 −β 2 +β(x−μ)
dNIG(α,β,δ,μ) (x) = e ,
π δ 2 + (x − μ)2
whereas λ = 1 characterizes the subclass of hyperbolic distributions (HYP)
which was the first to be applied in finance in [10] and
√2
α2 − β 2 2
dHYP(α,β,δ,μ) (x) = e−α δ +(x−μ) +β(x−μ) .
2αδK1 (δ α − β )
2 2
For positive λ, Variance Gamma distributions (VG), which were introduced

in full generality in [23], can be obtained as weak limits of GH distributions.
If λ > 0 and δ → 0, then the density (10) converges pointwise to
1
(α2 − β 2 )λ |x − μ|λ− 2
dVG(λ,α,β,μ) (x) = √ 1 Kλ− 12 (α|x − μ|) eβ(x−μ) .
π(2α)λ− 2 Γ (λ)
However, if λ < 0 and α, β → 0, then (10) converges pointwise to the density
of a scaled and shifted t-distribution with f = −2λ degrees of freedom:
λ− 12 ∞
Γ (−λ + 12 ) (x − μ)2
dt(λ,δ,μ) (x) = √ 1+ , where Γ (λ) = xλ−1 e−x dx.
δ πΓ (−λ) δ2 0
For a detailed derivation of these limits and their characteristic functions, we

refer to [11].
Both the skewness and especially the heavier tails increase significantly
the probability of joint defaults in the factor model
√
Xi = ρ M + 1 − ρ Zi . (11)
In the following we assume that M, Z1 , . . . , ZN are independent as before, but

M ∼ GH(λM , αM , βM , δM , μM ) and all Zi are iid ∼ GH(λZ , αZ , βZ , δZ , μZ )
(including the above limiting cases). Thus the distribution functions of all Xi
coincide. Denote the latter by FX and the distribution functions of M and
of the Zi by FM and FZ , then one can derive the corresponding cumulative
default distribution FZt analogously as described in Sect. 2.2 and obtains
−1 √
FX (Q(t)) − 1 − ρ FZ−1 (q)
FZt (q) ≈ 1 − FM √ . (12)
ρ
Note that this expression cannot be simplified further as in equation (8) since
the distribution of M is in general not symmetric.
Remark 4. As mentioned above, almost all densities of GH distributions pos-
sess exponentially decreasing tails, only the Student t limit distributions have
a power tail. According to the results of [24], the joint distribution of the Xi
will therefore show tail dependence if and only if the systematic factor M is
Student t-distributed.
L
Further GH(λ, α, β, δ, μ) −→ N (μ + βσ 2 , σ 2 ) if α, δ → ∞ and δ/α → σ 2 ,
so the normal factor model is included as a limit in our setting.
3.1 Factor Scaling and Calculation of Quantiles
To preserve the role of ρ as a correlation parameter, we have to standardize

the factor distributions such that they have zero mean and unit variance. In
the general case of GH distributions we fix shape, skewness and tail behaviour
by specifying α, β, λ and then calculate δ̄ and μ̄ that scale and shift the density
appropriately. For this purpose we first solve the equation
2
δ 2 Kλ+1 (ζ) δ4 Kλ+2 (ζ) Kλ+1 (ζ)
1 = Var[GH(λ, α, β, δ, μ)] = + β2 2 −
ζ Kλ (ζ) ζ Kλ (ζ) Kλ2 (ζ)

with ζ := δ α2 − β 2 numerically to obtain δ̄ and then choose μ̄ such that
β δ̄ 2 Kλ+1 (ζ̄)
0 = E[GH(λ, α, β, δ̄, μ̄)] = μ̄ + , ζ̄ = δ̄ α2 − β 2 .
ζ̄ Kλ (ζ̄)
Since the Bessel functions Kn+1/2 , n ≥ 0, can be expressed explicitly in closed

forms, the calculations simplify considerably for the NIG subclass. We have
δα2 βδ
Var[NIG(α, β, δ, μ)] = 3 , E[NIG(α, β, δ, μ)] = μ + ,
(α2 − β2) 2 α2 − β 2
3
so the distribution can be standardized by choosing δ̄ = (α2 − β 2 ) 2 /α2 and
μ̄ = −β(α2 − β 2 )/α2 .
In the VG limiting case the variance is given by
2λ 4λβ 2 2
Var[VG(λ, α, β, μ)] = + =: σVG ,
α2 − β 2 (α2 − β 2 )2
so it would be tempting to use λ as a scaling parameter, but this would mean

to change the tail behaviour which we want to keep fixed. Observing the fact
that a VG distributed random variable XVG equals in distribution the shifted
sum of two Gamma variables, that is,
d σ λ λ−1 −σx
XVG = Γλ,α−β − Γλ,α+β + μ, where dΓλ,σ (x) = x e [0,∞) (x),
Γ (λ)
the correct scaling that preserves the shape is ᾱ = σVG α, β̄ = σVG β. Then
μ̄ has to fulfill
2λβ̄
0 = E[VG(λ, ᾱ, β̄, μ̄)] = μ̄ + 2 .
ᾱ − β̄ 2
The second moment of a Student t-distribution exists only if the number of
degrees of freedom satisfies f > 2, so we have to impose the restriction λ < −1
in this case. Mean and variance are given by
δ2
Var[t(λ, δ, μ)] = and E[t(λ, δ, μ)] = μ,
−2λ − 2
√
therefore one has to choose δ̄ = −2λ − 2 and μ̄ = 0.
We thus have a minimum number of three free parameters in our general-
ized factor model, namely λM , λZ and ρ if both M and Zi are t-distributed,
up to a maximum number of seven (λM , αM , βM , λZ , αZ , βZ , ρ) if both factors
are GH or VG distibuted. If we restrict M and Zi to certain GH subclasses
by fixing λM and λZ , five free parameters are remaining.
Having scaled the factor distributions, the remaining problem is to com-
−1
pute the quantiles FX (Q(t)) which enter the default distribution FZt by equa-
tion (12). Since the class of GH distributions is in general not closed under
convolutions, the distribution function FX is not known explicitly. Therefore
one central task of the project was to develop a fast and stable algorithm for
the numerical calculation of the quantiles of Xi , because simulation techniques
had to be ruled out from the very beginning for two reasons: The default prob-
abilities Q(t) are very small, so one would have to generate a very large data
set to get reasonable quantile estimates, and the simulation would have to
be restarted whenever at least one model parameter has been modified. Since
the pricing formula is evaluated thousands of times with different parameters
during calibration, this procedure would be too time-consuming. Further, the
routine used to calibrate the models tries to find an extremal point by search-
ing the direction of the steepest ascend within the parameter space in each
optimization step. This can be done successfully only if the model prices de-
pend exclusively on the parameters and not additionally on random effects.
In the latter case the optimizer may behave erratically and will never reach
an extremum.
We obtain the quantiles of Xi by Fourier inversion. Let P̂X , P̂M and P̂Z
denote the characteristic functions of Xi , M and Zi , then equation (11) and
the independence of the factors yield
√
P̂X (t) = P̂M ρ t · P̂Z 1 − ρt .
With the help of the inversion formula we get a quite accurate approximation
−1
of FX from which the quantiles FX (Q(t)) can be derived. For all possible
factor distributions mentioned above, the characteristic functions P̂M and P̂Z
are well known; see [11] for a derivation and explicit formulas.
In contrast to this approach there are two special settings in which the
quantiles of Xi can be calculated directly. The first one relies on the following
convolution property of the NIG subclass,
NIG(α, β, δ1 , μ1 ) ∗ NIG(α, β, δ2 , μ2 ) = NIG(α, β, δ1 + δ2 , μ1 + μ2 ),
α β
and the fact that if Y ∼ NIG(α, β, δ, μ), then aY ∼ NIG |a| , a , δ|a|, μa .
Thus if both M and Zi are NIG distributed √ and the distribution√parameters
√ √
of the latter are defined by αZ := αM 1 − ρ/ ρ and βZ = βM 1 − ρ/ ρ,
αM βM δ̄M μ̄M
then it follows together with equation (11) that Xi ∼ NIG √ρ , √ρ , √ρ , √ρ ,
where δ̄M and μ̄M are the parameters of the standardized distribution of M
as described before.
In the VG limiting case the parameters α, β and μ behave as above under
scaling, and the corresponding convolution property is
VG(λ1 , α, β, μ1 ) ∗ VG(λ2 , α, β, μ2 ) = VG(λ1 + λ2 , α, β, μ1 + μ2 ).
Consequently if both factors are VG distributed and the free parameters of
the idiosyncratic factor are chosen as follows, λZ = λM (1 − ρ)/ρ, αZ = αM ,

βZ = βM , then Xi ∼ VG λρM , ᾱ√Mρ , β̄√Mρ , μ̄√Mρ ).
This stability under convolutions, together with the appropriate parameter
choices for the idiosyncratic factor, was used in [19] and all models considered
in [1]. We do not use this approach here because it reduces the number of free
parameters and therefore the flexibility of the factor model. Moreover, in such
a setting the distribution of the idiosyncratic factor is uniquely determined by
the systematic factor, which contradicts the intuitive idea behind the factor
model and lacks an economic interpretation.
3.2 Calibration Results for the Dj iTraxx Europe

We calibrate our generalized factor model with market quotes of DJ iTraxx
Europe standard tranches. As mentioned before, the iTraxx Europe index is
based on a reference portfolio of 125 European investment grade firms and
quotes its average credit spread which can be used to estimate the default
intensity of all constituents according to equation (4). The diversification of
the portfolio always remains the same. It contains CDSs of 10 firms from au-
tomotive industry, 30 consumers, 20 energy firms, 20 industrials, 20 TMTs
(technology, media and telecommunication companies) and 25 financials. In
each sector, the firms with the highest liquidity and volume of trade with
respect to their defaultable assets (bonds and CDSs) are selected. The iTraxx
portfolio is reviewed and updated quarterly. Not only companies that have de-
faulted in between are replaced by new ones, but also those which no longer
fulfill the liquidity and trading demands. Of course, the recomposition affects
future deals only. Once two counterparties have agreed to buy and sell protec-
tion on a certain iTraxx tranche, the current portfolio is kept fixed for them
in order to determine the corresponding cash flows described in Sect. 2.1. The
names and attachment points of the five iTraxx standard tranches are given
in Figs. 2 and 3. For each of them four contracts with different maturities (3,
5, 7 and 10 years) are available.
The settlement date of the sixth iTraxx series was December 20, 2006, so
the 5, 7, and 10 year contracts mature on December 20, 2011 resp. 2013 and
2016. We consider the market prices of the latter on all standard tranches at
November 13, 2006. For the mezzanine and senior tranches, these equal the
annualized fair spreads ri which can be obtained from equation (3) and are
also termed running spreads. However, the market convention for pricing the
equity tranche is somewhat different: In this case the protection buyer has to
pay a certain percentage s1 of the notional value K1 N L as an up-front fee
at the starting time t0 of the contract and a fixed spread of 500bp on the
outstanding notional at t1 , . . . , tn . Therefore the premium leg for the equity
tranche is given by

n
3 4
P L1 (s1 ) = s1 K1 N L + 0.05 (tk − tk−1 )β(t0 , tk )E K1 − L1tk N L ,
k=1
and the no-arbitrage condition P L1 (s1 ) = D1 then implies

%n 3 4 3 1 4 3 1 4
k=1 β(t ,
0 kt ) E L 1
tk − E L tk−1 − 0.05(t k −t k−1 ) K 1 − E Ltk
s1 = . (13)
K1
Since the running spread is set to a constant of 500 bp, the varying market
price quoted for the equity tranche is the percentage s1 defining the magnitude
of the up-front fee.
We calibrate our generalized factor model by least squares optimization,
that is, we first specify to which subclass of the GH family the distribu-
tions FM and FZ belong and then determine the correlation and distribution
parameters numerically which minimize the sum of the squared differences be-
tween model and market prices over all tranches. Although our algorithm for
−1
computing the quantiles FX (Q(t)) allows us to combine factor distributions
of different GH subclasses, we restrict both factors to the same subclass for
simplicity reasons. Therefore in the following table and figures the expression
VG, for example, denotes a factor model where M and the Zi are variance
gamma distributed. The model prices are calculated from equations (3) and
(13), using the cumulative default distribution (12) resp. (8) for the normal
factor model which serves as a benchmark. The recovery rate R which has
a great influence on the expected losses E[Litk ] according to equation (9) is
always set to 40 %; this is the common market assumption for the iTraxx
portfolio.
One should observe that the prices of the equity tranches are usually given
in percent, whereas the spreads of all other tranches are quoted in basis points.
In order to use the same units for all tranches in the objective function to
be minimized, the equity prices are transformed into basis points within the
optimization algorithm. Thus they are much higher than the mezzanine and
senior spreads and therefore react to parameter changes in a more sensitive
way, which amounts to an increased weighting of the equity tranche in the
calibration procedure. This is also desirable from an economical point of view
since the costs for mispricing the equity tranche are typically greater than for
all other tranches.
Remark 5. For the same reason, the normal factor model is usually calibrated
by determining the implied correlation of the equity tranche first and then
using this to calculate the fair spreads of the other tranches. This ensures
that at least the equity price is matched perfectly. To provide a better com-
parison with our model, we give up this convention and also use least squares
estimation in this case. Therefore the fit of the equity tranche is sometimes
less accurate, but the distance between model and market prices is smaller for
the higher tranches instead.
Our calibration results for the 5 and 7 year iTraxx tranches are shown in
Figs. 5 and 6. The normal benchmark model performs worst in all cases. The
performance of the t model is comparable with the NIG and HYP models,
whereas the VG model provides the best fit for both maturities. Since the
t model is the only one exhibiting tail dependence (confer Remark 4) but does
not outperform the NIG, HYP and VG models, one may conclude that this
property is negligible in the presence of more flexible factor distributions. This
may also be confirmed by the fact that all estimated GH parameters βM and
βZ are different from zero which implies skewness of the factor distributions.
Furthermore the parameter ρ is usually higher in the GH factor models than
in the normal benchmak model, indicating that correlation is still of some
importance, but has a different impact on the pricing formula because of the
more complex dependence structure.
Fig. 5. Comparison of calibrated model prices and market prices of the 5 year
iTraxx contracts
Fig. 6. Comparison of calibrated model prices and market prices of the 7 year
iTraxx contracts
The VG model even has the potential to fit the market prices of all tranches
and maturities simultaneously with high accuracy, which we shall show below.
However, before that we want to point out that the calibration over differ-
ent maturities requires some additional care to avoid inconsistencies when
calculating the default probabilities. As can be seen from Fig. 7, the aver-
age iTraxx spreads sa are increasing in maturity and by equation (4) so do
the default intensities λa . This means that the estimated default probabilities
Q(t) = 1 − e−λa t of a CDO with a longer lifetime are always greater than
those of a CDO with a shorter maturity. While this can be neglected when
concentrating on just one maturity, this fact has to be taken into account
Fig. 7. Constant iTraxx spreads of November 13, 2006, and fitted Nelson–Siegel
curve r̂NS with parameters β̂0 = 0.0072, β̂1 = −0.0072, β̂2 = −0.0069, τ̂1 = 2.0950
when considering iTraxx CDOs of different maturities together. Since the un-
derlying portfolio is the same, the default probabilities should coincide during
the common lifetime.
To avoid these problems we now assume that the average spreads sa =
s(t) are time-dependent and follow a Nelson–Siegel curve. This parametric
family of functions has been introduced in [27] and has become very popular
in interest rate theory for the modeling of yield curves where the task is
the following: Let β(0, tk ) denote today’s price of a zero coupon bond with
maturity tk as before, then one has to find a function f (instantaneous forward
t
rates) such that the model prices β(0, tk ) = exp − 0 k f (t) dt approximate
the market prices reasonably well for all maturities tk . Since instantaneous
forward rates cannot be observed directly in the market, one often uses an
equivalent expression in terms of spot rates: β(0, tk ) = exp(−r(tk )tk ), where
t
the spot rate is given by r(tk ) = t1k 0 k f (t) dt. Nelson and Siegel suggested
to model the forward rates by
− τt t − τt
fNS(β0 ,β1 ,β2 ,τ1 ) (t) = β0 + β1 e 1 + β2 e 1.
τ1
The corresponding spot rates are given by
τ1 − t

− t
1 − e τ1 − β2 e τ1 .
rNS(β0 ,β1 ,β2 ,τ1 ) (t) = β0 + (β1 + β2 ) (14)
t
In order to obtain time-consistent default probabilities resp. intensities we
replace sa in equation (4) by a Nelson–Siegel spot rate curve (14) that has
been fitted to the four quoted average iTraxx spreads, that is,
r̂NS (t)
λa = λ(t) = , (15)
(1 − R)10000
and Q(t) := 1 − e−λ(t) t . The Nelson–Siegel curve estimated from the iTraxx
spreads of November 13, 2006, is shown in Fig. 7. At first glance the differences
between constant and time-varying spreads seem to be fairly large, but one
should observe that these are the absolute values which have already been
divided by 10000 and therefore range from 0 to 0.004338, so the differences in
the default probabilities are almost negligible.
Under the additional assumption (15), we have calibrated a model with
VG distributed factors to the tranche prices of all maturities simultaneously.
The results are summarized in Table 1 and visualized in Fig. 8. The fit is ex-
Table 1. Results of the VG model calibration simultaneously over all maturities.

Estimated parameters are as follows: λM = 0.920, αM = 5.553, βM = 1.157, λZ =
2.080, αZ = 2.306, βZ = −0.753, ρ = 0.321.
Tranches Market VG Market VG Market VG
5Y 7Y 10Y
0–3 % 13.60 % 13.60 % 28.71 % 28.72 % 42.67 % 42.67 %
3–6 % 57.16bp 53.30bp 140.27bp 132.27bp 360.34bp 357.60bp
6–9 % 16.31bp 17.19bp 41.64bp 41.83bp 105.08bp 111.17bp
9–12 % 6.65bp 8.23bp 21.05bp 19.90bp 43.33bp 52.00bp
12–22 % 2.67bp 3.05bp 7.43bp 7.34bp 13.52bp 18.97bp
Fig. 8. Graphical representation of the differences between model and market prices
obtained from the simultaneous VG calibration
cellent. The maximal absolute pricing error is less than 9bp, and for the 5 and
7 year maturities the errors are, apart from the junior mezzanine tranches,
almost as small as in the previous calibrations. The junior mezzanine tranche
is underpriced for all maturities, but it is difficult to say whether this is caused
by model or by market imperfections. Nevertheless the overall pricing perfor-
mance of the extended VG model is comparable or better than the perfor-
mance of the models considered in [1, 7, 19], although the latter were only
calibrated to tranche quotes of a single maturity.
Also note that this model admits a flat correlation structure not only over
all tranches, but also over different maturities: all model prices contained in
Table 1 were calculated using the same parameter ρ. Thus the correlation
smiles shown in Fig. 3 which in some sense question the factor equation (5)
are completely eliminated. Therefore the intuitive idea of the factor approach
is preserved, but one should keep in mind that in the case of GH distributed
factors the dependence structure of the joint distribution of the Xi is more
complex and cannot be described by correlation alone.
4 A Dynamic Markov Chain Model

In this section we discuss an entirely different approach to explain observed
CDO spreads, rooted more in the theory of stochastic processes. Our exposi-
tion summarizes results from [13].
4.1 The Model

We begin with some notation. Given some probability space (Ω, F , Q), Q the
risk-neutral measure used for pricing, we define the default indicator of firm
i at time t by Yt,i = {Ti ≤t} . Note that the default indicator process Yi =
(Yt,i )t≥0 is a right continuous process which jumps from 0 to 1 at the default
of firm i. The evolution of the default state of the portfolio is then described
by the process Y = (Yt,1 , . . . , Yt,N )t≥0 ; obviously, Yt ∈ S Y := {0, 1}N . We use
the following notation for flipping the ith coordinate of a default state: given
y ∈ S Y we define y i ∈ S Y by
yii := 1 − yi and yji := yj , j ∈ {1, . . . , N } \ {i} . (16)
The default history (the internal filtration of the process Y ) is denoted by
(Ht ), i.e. Ht = σ(Ys : s ≤ t). An (Ht )-adapted process (λt,i ) is called the
default intensity of Ti (with respect to (Ht )) if
Ti ∧t
Yt,i − λs,i ds is an (Ht )-martingale.
0
Intuitively, λt,i gives the instantaneous chance of default of a non-defaulted
firm i given the default history up to time t. It is well-known that the default
intensities determine the law of the marked point process (Yt ); see for instance
[6] for a detailed account of the mathematics of marked point processes.
Modeling the Dynamics of Y
We assume that the default intensity of a non-defaulted firm i at time t is

given by a function λi (t, Yt ) of time and of the current default state Yt . Hence
the default intensity of a firm may change if there is a change in the default
state of other firms in the portfolio; in this way dependence between default
events can be modeled explicitly. Formally, we model the default indicator
process by a time-inhomogeneous Markov chain with state space S Y . The
next assumption summarizes the mathematical properties of Y .
Assumption 1 (Markov family). Consider bounded and measurable func-
tions λi : [0, ∞) × S Y → R+ , 1 ≤ i ≤ N . There is a family Q(t,y) ,
(t, y) ∈ [0, ∞) × S Y , of probability measures on (Ω, F , (Ht )) such that
Q(t,y) (Yt = y) = 1 and such that (Ys )s≥t is a finite-state Markov chain with
state space S Y and transition rates λ(s, y1 , y2 ) given by
(1 − y1,i ) λi (s, y1 ), if y2 = y1i for some i ∈ {1, . . . , N },

λ(s, y1 , y2 ) = (17)
0 else.
Relation (17) has the following interpretation: In t the chain can jump only to
the set of neighbors of the current state Yt that differ from Yt by exactly one
default; in particular there are no joint defaults. The probability that firm i
defaults in the small time interval [t, t + h) thus corresponds to the probabil-
ity that the chain jumps to the neighboring state (Yt )i in this time period.
Since such a transition occurs with rate λi (t, Yt ), it is intuitively obvious that
λi (t, Yt ) is the default intensity of firm i at time t; a formal argument is given
in [13].
The numerical treatment of the model can be based on Monte Carlo simu-
lation or on the Kolmogorov forward and backward equation for the transition
probabilities; see again [13] for further information. An excellent introduction
to continuous-time Markov chains is given in [28].
Modeling Default Intensities
The default intensities λi (t, Yt ) are crucial ingredients of the model. If the port-
folio size N is large – such as in the pricing of typical synthetic CDO tranches –
it is natural to assume that the portfolio has a homogeneous group structure.
This assumption gives rise to intuitive parameterizations for the default in-
tensities; moreover, the homogeneous group structure leads to a substantial
reduction in the size of the state space of the model. Here we concentrate on
the extreme case where the entire portfolio forms a single homogeneous group
so that the processes Yi are exchangeable; this simplifying assumption is made
in most CDO pricing models; see also Sect. 2. Denote the number of defaulted
firms at time t by
N
Mt := Yt,i .
i=1
As discussed in [13], in a homogeneous model default intensities are necessarily

of the form
λi (t, Yt ) = h(t, Mt ) for some h : [0, ∞) × {0, . . . , N } → R+ . (18)
Note that the assumption that default intensities depend on Yt via the number
of defaulted firms Mt makes sense also from an economic viewpoint, as un-
usually many defaults might have a negative impact on the liquidity of credit
markets or on the business climate in general. This point is discussed further
in [12] and [16].
The simplest exchangeable model is the linear counterparty risk model.
Here
h(t, l) = λ0 + λ1 l , λ0 > 0, λ1 ≥ 0, l ∈ {0, . . . , N } . (19)
The interpretation of (19) is straightforward: upon default of some firm the
default intensity of the surviving firms increases by the constant amount λ1
so that default dependence increases with λ1 ; for λ1 = 0 defaults are indepen-
dent. Model (19) is the homogeneous version of the so-called looping-defaults
model of [18].
The next model generalizes the linear counterparty risk model in two im-
portant ways: first, we introduce time-dependence and assume that a default
event at time t increases the default intensity of surviving firms only if Mt
exceeds some deterministic threshold μ(t) measuring the expected number of
defaulted firms up to time t; second, we assume that on {l > μ(t)} the func-
tion h(t, ·) is strictly convex. Convexity of h implies that large values of Mt
lead to very high values of the default intensities, thus triggering a cascade of
further defaults. This will be important in explaining properties of observed
CDO prices below. The following specific model with the above features will
be particularly useful:

λ1 (l − μ(t))+
h(t, l) = λ0 + exp λ2 − 1 , λ0 > 0, λ1 , λ2 ≥ 0 ; (20)
λ2 N
in the sequel we call (20) convex counterparty risk model. In (20) λ0 is a level
parameter that mainly influences credit quality. λ1 gives the slope of h(t, l) at
μ(t); intuitively this parameter models the strength of default interaction for
“normal” realisations of Mt . The parameter λ2 controls the degree of convexity
of h and hence the tendency of the model to generate default cascades; note
that for λ2 → 0 (and μ(t) ≡ 0) (20) reduces to the linear model (19).
The Markov Property of M
It is straightforward that for default intensities of the form (18) the process
M = (Mt )t≥0 is itself a Markov chain with generator given by

[t] f (l) = (N − l)h(t, l) f (l + 1) − f (l) .
GM (21)
In fact, since Assumption 1 excludes joint defaults, Mt can jump only to Mt +1.
The intensity of such a transition is proportional to the number N − Mt of
surviving firms at time t as well as to their default intensity h(t, Mt ). This
is important: since the portfolio loss satisfies Lt = (1 − R)Mt /N , the loss
processes Li of the individual tranches (and of course the overall portfolio
loss) are given by functions of Mt , so that we may concentrate on the process
M . As shown in [13] this considerably simplifies the numerical analysis of the
model. Similar modeling ideas have independently been put forward in [3].
4.2 Analysis of CDO Tranches
Next we turn to an analysis of synthetic CDO tranches in the context of the

Markov chain model; in particular, we are interested in modeling the well-
known implied correlation skew described in Sect. 2.3. Recall that according
to equation (3), the computation of fair tranche spreads ri boils down to
evaluating the distribution of Lit – and hence the distribution of Mt – at the
premium payment dates. The latter can be computed efficiently using the
Kolmogorov forward equations or by simulation; see [13] for details.
The basic idea for generating correlation skews in the context of the convex
counterparty risk model (20) is simple: by increasing λ2 we can generate occa-
sional large clusters of defaults without affecting the left tail of the distribution
of Lt too much; in this way we can reproduce the spread of the mezzanine and
senior CDO tranches in a way which is consistent with the observed spread of
the equity tranche. In order to confirm this intuition we consider a numerical
example with spread data from [17]. In Table 2 we give the CDO spreads if the
convexity parameter λ2 is varied; λ0 and λ1 were calibrated to the index level
Table 2. CDO spreads in the convex counterparty risk model (20) for varying λ2 .
λ0 and λ1 were calibrated to the index level of 42bp and the market quote for the
equity tranche, assuming δ = 0.6. For λ2 ∈ [8, 10] the qualitative properties of the
model-generated CDO spreads resemble closely the behaviour of the market spreads;
with state-dependent LGD the fit is almost perfect
tranches [0,3] [3,6] [6,9] [9,12] [12,22]

market spreads 27.6 % 168.0bp 70.0bp 43.0bp 20.0bp
P
model spreads abs. err.
λ2 = 0 27.6 % 223.1bp 114.5bp 61.1bp 16.9bp 120.8bp
λ2 = 5 27.6 % 194.2bp 95.7bp 54.9bp 23.3bp 67.1bp
λ2 = 8 27.6 % 172.1bp 80.0bp 46.7bp 23.7bp 21.5bp
λ2 = 8.54 27.6 % 168.0bp 77.1bp 45.1bp 23.5bp 12.7bp
λ2 = 10 27.6 % 156.9bp 69.4bp 40.7bp 22.7bp 16.7bp
state-dependent LGD 27.6 % 168.0bp 71.2bp 39.3bp 19.6bp 5.3bp
δ0 = 0.5; δ1 = 7.5
and the observed market quote of the equity tranche. The results show that
for appropriate values of λ2 the model can reproduce the qualitative behavior
of the observed tranche spreads in a very satisfactory way. This observation
is interesting, as it provides an explanation of correlation skews of CDOs in
terms of the dynamics of the default indicator process. Similarly as in [2], the
model fit can be improved further by considering a state-dependent loss given
default of the form δt = δ0 + δ1 Mt with δ0 , δ1 > 0; see again Table 2.
Comments
Implied correlations for CDO tranches on the iTraxx Europe have changed
substantially since August 2004. More importantly, the analysis presented in
Table 2 presents only a “snapshot” of the CDO market at a single day. For
these reasons in [13] the convex counterparty risk model (20) was recalibrated
to 6 months of observed 5 year tranche spreads on the iTraxx Europe in the
period 23.9.2005–03.03.2006. It turned out that the resulting parameters were
quite stable over time.
In this paper we have calibrated a parametric version of the model (20)
to observed CDO spreads. For an interesting nonparametric calibration pro-
cedure based on the Kolmogorov forward equation we refer to [3, 20, 29].
Dynamic Markov chain models are very useful tools for studying the prac-
tically relevant risk management of CDO tranches via dynamic hedging; we
refer to the recent papers [14] and [20] for details.
References
1. H. Albrecher, S. A. Ladoucette, and W. Schoutens, A generic one-factor Lévy
model for pricing synthetic CDOs. In: M. C. Fu, R. A. Jarrow, J.-Y. Yen, and
R. J. Elliott (eds.), Advances in Mathematical Finance, Birkhäuser, Boston,
259–277 (2007)
2. L. Andersen, J. Sidenius, Extensions to the Gaussian copula: random recovery
and random factor loadings. Journal of Credit Risk 1, 29–70 (2005)
3. M. Arnsdorf, I. Halperin, BSLP: Markovian bivariate spread-loss model for port-
folio credit derivatives. Working paper, JP Morgan (2007)
4. O. E. Barndorff-Nielsen, Exponentially decreasing distributions for the logarithm
of particle size. Proceedings of the Royal Society London, Series A 353, 401–419
(1977)
5. T. Berrada, D. Dupuis, E. Jacquier, N. Papageorgiou, and B. Rémillard, Credit
migration and basket derivatives pricing with copulas. Journal of Computational
Finance 10, 43–68 (2006)
6. P. Brémaud, Point Processes and Queues: Martingale Dynamics. Springer, New
York (1981)
7. X. Burtschell, J. Gregory, and J.-P. Laurent, A comparative analysis of CDO
pricing models. Working paper, BNP Paribas (2005)
8. X. Burtschell, J. Gregory, and J.-P. Laurent, Beyond the Gaussian copula:
stochastic and local correlation. Journal of Credit Risk 3, 31–62 (2007)
9. E. Eberlein, Application of generalized hyperbolic Lévy motions to finance. In:

O. E. Barndorff-Nielsen, T. Mikosch, and S. Resnick (eds.), Lévy processes:
Theory and Applications, Birkhäuser, Boston, 319–337 (2001)
10. E. Eberlein, U. Keller, Hyperbolic distributions in finance. Bernoulli 1, 281–299
(1995)
11. E. Eberlein, E. A. von Hammerstein, Generalized hyperbolic and inverse Gaus-
sian distributions: limiting cases and approximation of processes. In: R. C.
Dalang, M. Dozzi, and F. Russo (eds.), Seminar on Stochastic Analysis, Ran-
dom Fields and Applications IV, Progress in Probability 58, Birkhäuser, Basel,
221–264 (2004)
12. R. Frey, A mean-field model for interacting defaults and counterparty risk. Bul-
letin of the International Statistical Institute (2003)
13. R. Frey, J. Backhaus, Credit derivatives in models with interacting default in-
tensities: a Markovian approach. Preprint, Universität Leipzig (2006), available
from www.math.uni-leipzig.de/~frey/publications-frey.html.
14. R. Frey, J. Backhaus, Dynamic hedging of synthetic CDO tranches with
spread- and contagion risk. Preprint, Universität Leipzig (2007), available from
www.math.uni-leipzig.de/~frey/publications-frey.html.
15. J. Gregory, J.-P. Laurent, In the core of correlation. Risk, October, 87–91 (2004)
16. U. Horst, Stochastic cascades, credit contagion and large portfolio losses. Journal
of Economic Behavior & Organization 63, 25–54 (2007)
17. J. Hull, A. White, Valuation of a CDO and an nth to default CDS without Monte
Carlo simulation. Journal of Derivatives 12, 8–23 (2004)
18. R. Jarrow, F. Yu, Counterparty risk and the pricing of defaultable securities.
Journal of Finance 53, 2225–2243 (2001)
19. A. Kalemanova, B. Schmid, and R. Werner, The normal inverse Gaussian dis-
tribution for synthetic CDO pricing. Working paper, risklab germany (2005)
20. J.-P. Laurent, A. Cousin, and J. D. Fermanian, Hedging default risk of CDOs in
Markovian contagion models. Workinp paper, ISFA Actuarial School, Université
de Lyon (2007)
21. D. Li, On default correlation: a copula function approach. Journal of Fixed In-
come 9, 43–54 (2000)
22. A. J. McNeil, R. Frey, and P. Embrechts, Quantitative Risk Management. Prince-
ton Series in Finance, Princeton University Press, Princeton (2005)
23. D. B. Madan, P. Carr, and E. C. Chang, The variance gamma process and option
pricing. European Financial Review 2, 79–105 (1998)
24. Y. Malevergne, D. Sornette, How to account for extreme co-movements between
individual stocks and the market. Journal of Risk 6, 71–116 (2004)
25. R. C. Merton, On the pricing of corporate debt: the risk structure of interest
rates. Journal of Finance 29, 449–470 (1974)
26. R. B. Nelsen, An Introduction to Copulas. Lecture Notes in Statistics 139,
Springer, New York (1999)
27. C. R. Nelson, A. F. Siegel, Parsimonious modeling of yield curves. Journal of
Business 60, 473–489 (1986)
28. J. R. Norris, Markov Chains. Cambridge University Press, Cambridge (1997)
29. P. J. Schönbucher, Portfolio loss and the term-structure of loss transition rates:
a new methodology for the pricing of portfolio credit derivatives. Working paper,
ETH Zürich (2006)
30. P. J. Schönbucher, D. Schubert, Copula-dependent default risk in intensity mod-
els. Preprint, ETH Zürich and Universität Bonn (2001)
31. A. Sklar, Fonctions de répartition à n dimensions et leurs marges. Publications

de l’Institute de Statistique de l’Université de Paris 8, 229–231 (1959)
32. O. A. Vasiček, Probability of loss on loan portfolio. Mimeo, KMV Corporation
(1987)
33. O. A. Vasiček, Limiting loan loss probability distribution. Mimeo, KMV Corpo-
ration (1991)
Contributions to Multivariate Structural
Approaches in Credit Risk Modeling
Swantje Becker∗, Stefanie Kammer∗ , and Ludger Overbeck
Universität Gießen, Arndtstraße 2, 35392 Gießen, Germany

{swantje.becker,stefanie.kammer,ludger.overbeck}@math.uni-giessen.de
Summary. Credit risk models are usually differentiated into reduced form models
and structural models. The latter are usually more powerful if many credits are
to be modelled, more precisely if the focus stays with the dependency structure of
credits, whereas reduced form models are more adequate if single credits, like term
structure of credit spreads, are considered. This paper has two objectives the first
one is to analyze the credit spread dynamcis of a wide class of structural models
and the second one to understand the dependency structure if the multivariate asset
value model is assumed to be a shot-noise process.
1 Introduction
Credit risk and the capital markets products transferring and structuring
credit risk have evolved dramatically over the last decade. The first liquidly
traded credit derivative was the simple credit default swap, which transfer
the credit risk of an underlying bond synthetically to an investor, without the
necessity that the investor buys the bond.
If the risk out of entire portfolios of bonds, loans or other credit risky
instruments is transferred to the capital market, one usually employs a se-
curitization or other structured credit products. As an example: Synthetic
transaction based on credit default swaps (CSO).
Fig. 1. Credit Default Swap (CDS)
∗
Supported by BMBF-grant 03OVNHF
282 S. Becker, S. Kammer, L. Overbeck
Fig. 2. Collaterized Swap Obligation (CSO)
Losses in the underlying portfolio, are first covered by the lowest tranche,
i.e. in a synthetic structure, if the first loss in the portfolio on the left hand
side happens, of an amount of l1 , then the holder of the lowest tranche has to
pay the amount of l1 the swap counterparty.
This increasing market made it indispensible to model the underlying
risk properly. There are two classes of models. The first one are made up
by the reduced form models, which are reduced to the modeling of the
default probabilities themselves, without asking how defaults happens, cf.
e.g. [10].
In contrast, the basic feature of a structural model lies in the attempt to
model a causal structure, which finally leads to default. In the simplest case
the structure consists in a stochastic process Y = (Yt ) determining, whether
a default event has happened up to a given time T > 0. In that case
DT = {YT ≤ K} (1)
with a default threshold KT . A more dynamic approach, the exit or hit-

ting time model, constructs the default time τ as the first hitting time
of K,
τ = inf{s ≥ 0 : Ys < K} . (2)
In Sect. 2 we will present the one-dimensional dynamics of the exit time model
in the case where X is a Brownian motion, and a Brownian motion with de-
terministic or stochastic time change. The deterministic time change provides
an exit time model matching any term structure of default probabilities. In
the context of time changes, we will present for any distribution function p(t)
Multivariate Structural Credit Risk Modeling 283
on [0, ∞), the distribution of the default time implied by the term struc-
ture of default probabilities, a time change Ỹ of a Brownian motion Y such
that
P [inf{s|Ỹs ≤ K} ≤ t] = p(t).
This model yields then a stochastic process s(t) for spreads, basically the dy-
namics of future default probabilities. More generally also process Ỹs = Ygs
with a stochastic time change (gs ) are introduced and their implied spread
dynamics in an exit time model.
In Sect. 3 we will extend the exit time models based on Brownian motion
towards a shot-noise process. Roughly, a shot-noise process is a Brownian
motion which additional Poisson distributed jump times and arbitrary jump
distribution. The “shot” character comes from the fast mean reversion of the
process, i.e. the reversion to the state just before the jump is done along
a determinstic exponential function.
We mainly focus on the impact of the shots to the joint default probabilities
and in a simulation study we are going to present some results on the valuation
of CDOs.
2 Exit Times of Time Changed Processes

In a structural first-passage approach a default event happens the first time
the asset-value or ability-to-pay process falls below some pre-specified default
boundary K depending on the firm’s debt. The structural model thus links
equity and debt of a firm’s capital structure. The default time of an entity was
defined in (2). We assume that we observe the underlying asset-value process
and know whether a default happened. Mathematically spoken we assume the
information is given by the filtration
FtY = σ(Ys : s ≤ t) ,
such that τ is a IFY -stopping time. A default event can e.g. indicate a credit-
rating downgrade or a total default; for us this is not important. We call
IP(τ ≤ t) the default probability, and for 0 ≤ t ≤ T define the survival
probability conditional on the information FtY as follows:

Q(t, T ) := IP τ > T | FtY .
We now describe the credit default swap (CDS, see Fig. 1) and the CDS credit
spread:
A CDS is a contract between a protection seller and a protection buyer. The
protection seller offers protection against default of a reference entity during
a certain time period, say between t and maturity T . Therefore the protection
buyer regularly pays an insurance fee, the credit spread s(t, T ), but only as
long as the entity is not defaulted. In case of a default before maturity of
the CDS the protection seller pays the claim amount 1 − R to the protec-
tion buyer, where R denotes the recovery rate of the entity. We assume that
there is no counterparty default risk, and for simplicity zero interest rates.
Under no default until t, the fair continuously-paid credit spread is given
by
(1 − R)(1 − Q(t, T ))
s(t, T ) = T . (3)
t Q(t, u) du
Usually a CDS spread is considered in terms of time to maturity M , then we
have to insert T = t + M into formula (3). We want to describe the dynam-
ics of the credit-spread via a stochastic differential equation (SDE), a local
volatility model:
dst = s(t) [μ(t, st ) dt + σ(t, st ) dWt ] , (4)
where W denotes a Brownian motion, μ can be interpreted as drift and σ as

volatility of credit spread. We are especially interested in the credit-spread
volatility. Fig. 3 shows credit-spread volatility of IBM versus time and credit
spread, respectively. Certainly credit-spread volatility changes with time and
seems to depend on the spread level.
Before we introduce the deterministic and the stochastic time-change
model we shortly summarize the history of the time change and give its defi-
nition:
Fig. 3. Credit-spread volatility estimates versus time (first plot) resp. versus credit
spread (second plot)
Bochner [5] first introduced a time-changed Brownian motion. Feller [12]

first presented subordinators as a time change to Markov processes. Clark
[8] introduced Brownian motion with an independent time change as a price
process in finance. Monroe [18] showed that a very general semi-martingale
can be embedded in Brownian motion via a time change. Ikeda & Watan-
abe [15] studied time-change models for solving SDEs. Øksendal [19] studied
when a stochastic integral can be represented as a time change of a diffusion.
Geman, Madan & Yor [13] and Carr & Wu [7] introduced subordinated
Lévy process. Schoutens [22] used these in derivative pricing.
Definition 2.1 (Time change). A deterministic time change is a posi-
tive, non-decreasing function and a stochastic time change a positive, non-
decreasing stochastic process.
The non-decreasing property of the time transformation can be understood
such that information that has been obtained once, will never be lost.
A stochastic time change adds stochastic volatility to a process. The original
clock will sometimes be called normal clock and the new clock will be called
business clock. The time change may be interpreted as experienced time, that
runs faster when the information flow is bigger (or speeds up). In other words,
experienced time is a measure of the amount of information arrival.
2.1 Deterministic Time-Change Model
The difficulty of the structural model is its calibration to a market-given term

structure of default probabilities F . The deterministic time-change model by
Overbeck & Schmidt [20],
⎡ ⎤2
K
τ = inf{s ≥ 0 : WTs < K} , Tt = ⎣ ⎦ ,
F (t)
Φ−1 2
perfectly matches any default probability curve: F (t) = IP (τ ≤ t) ∀ t ≥ 0.

The model leaves no degrees of freedom to influence the credit-spread dy-
namics: for a given asset values Yt = WTt and a corresponding credit spread
the credit-spread volatility is fixed. Figure 4 shows survival probabilities cor-
responding to various asset values WTt , and credit-spread volatilities corre-
sponding to various spread values st .
In order to be able to influence the credit-spead dynamics we introduce
a stochastic time-change model in the next subsection.
2.2 Stochastic Time-Change Model
Our general stochastic time-change process is given by
Yt = σWGt + μ Gt , (5)
Fig. 4. Survival probabilities corresponding to various asset values WTt , and credit-
spread volatilities corresponding to various spread values st
with drift parameter μ and volatility parameter σi . We make the following

assumption about the time change:
Assumption 2.2 (Continuous time change). G be a non-decreasing proc-
ess with continuous paths and starting value G0 = g ≥ 0, furthermore G be
independent of the Wiener process W .
Herewith the survival probability can be determined:
Q(t, T ) = P [τ > t|Ft ] (6)
∞
K − Yt μ√
= 1− Φ √ − z
0 σ z σ

2 σμ2 (K−Yt ) K − Yt μ√
+ e Φ √ + z IP GT − Gt ∈ dz| FtY ,
σ z σ
and thus, with (3), also the credit spread as a function of t and Yt , i.e.
s(t) = f (t, Yt ). (7)
The formulas are analytical whenever the conditional densitiy of the time-
change increment, IP(GT − Gt ∈ dz| FtY ), is analytical. This is for example
the case for the so-called simple time change:
t
Gt = g + σ̂ 2 Bs2 ds , (8)
0
where B is a Brownian motion independent of W .
2.3 Credit-Spread Dynamics
Applying Itô’s rule on the spread formula (3) in the representation indicated
by (7) yields the following spread dynamics:
dst = st [μs (t, Yt ) dt − σ s (t, Yt ) dYt ] ,

ft + 12 fyy f
with drift μs := s and σ s := − sy , where ft , fy , fyy are the partial
derivatives of f w.r.t. t and y.
We want to determine credit-spread dynamics, under the model without
drift Yt = WGt , that follow a SDE (4). Therefor we make an additional as-
sumption:
Assumption 2.3 (Absolute continuity). There is a stochastic process (gt )
t
with E[ 0 gs2 ds] < ∞ for all t ≥ 0 such that
t
Gt = g02 + gs2 ds ,
0
that is G is absolutely continuous.
Then Yt = WGt is equivalent in distribution to the stochastic integral,
L t
Yt = Y0 + 0 gs dWs .
This leads to the following credit-spread dynamics:
L
dst = st [μs (t, Yt ) dt − σ s (t, Yt )gt dWt ] ,
with credit-spread volatility (for T = t + M )
2(1 − R)
σ s · gt = gt ·
st
∞ 1 K−Yt
0
√ φ
z
√
z
IP Gt+M − Gt ∈ dz | FtY
αt
∞ "
0 Φ K−Y
√ t
z
IP Gt+M − Gt ∈ dz | FtY
+2 βt ,
α2t
and abbreviations
∞
t+M
K − Yt
αt = M − 2 Φ √ IP Gu − Gt ∈ dz | FtY
t 0 z
∞
t+M
1 K − Yt
βt = √ φ √ IP(Gu − Gt ∈ dz | FtY ) du .
t 0 z z
Again credit-spread dynamics are analytical whenever this holds for IP(GT −
Gt ∈ dz| FtY ).
As an example we consider the simple time change (8). It has three de-
μσ̂
grees of freedom for calibration: in K−Y 0
σσ̂ , σ , and g. The first two can be
used for calibration to the default-probability cirve and one to influence the
credit spread volatility. Figure 5 shows possible calibrations to an average
CCC default-probability curve (from the years 1981 to 2002) of Standard &
Poor’s, and Fig. 6 illustrates the influence of G0 = g on default-probability
and credit-spread curves. Especially g > 0 leads to non-zero instantaneous
credit spreads.
Fig. 5. Possible calibrations to an average CCC default-probability curve of S&P
Fig. 6. Influence of the starting value G0 = g on default-probability and credit-

spread curves
2.4 The Multivariate Model
The extension of (5) to a multivariate model is straightforward:
Yti = σi WGi t + μi Gt , i = 1, . . . , m ,
where each asset-value process has individual drift and volatility parameter,
Brownian motion and if wanted also an individual time change Gi . We fo-
cus on the two-dimensional model under a joint time change G, given by the
simple time change (8), and allow for correaltion between the Brownian mo-
tions, ρ = Corr(W 1 , W 2 ). Figure 7 shows joint survival probabilities (JSP)
curves, for fixed parameters g = 0, Y0 = 0, σ = 1, when varying the time-
change parameter σ̂ the threshold level K, and the Brownian correlation ρ.
A higher correlation yields higher JSPs, and a higher time change volatil-
ity (influenced by σ̂) yields steeper JSP curves. Note that the time change
and the Brownian correlation parameter ρ lead to different dependence struc-
tures.
Fig. 7. Joint survival-probability curves for ρ = 0.9 (uppermost curves), 0.5, 0.1
and −0.5 (lowest curves)
3 Asset-Value Processes with Shot-Noise

We define the asset-value process with shot-noise via
„ « “ √ ” N
σ2 Pt
μi − 2i t+σi ρBtM + 1−ρ2 Bti + (Yj −sgn(Yj )·min(|α·(t−τj )|,|Yj |))
Vti = V0i e j=1
, (9)
where μi is a constant drift and σi a positiv volatility. BtM is the brownian

motion of the market and Bti the asset specific brownian motion. All assets
have the same jump parameters. We assume that the jump amplitude Yj is
%
i.i.d. N (μY , σY2 ). (Nt )t≥0 is a poisson process with intensity λ. τj = jk=1 Ek
are the jump times, with Ek i.i.d. Exp(λ). α > 0 is our shot-noise parameter.
In the classical approach by Merton [17] a firm defaults if
VTi ≤ Ki .
In the first passage approach a default appears if
min Vsi ≤ Ki .
0≤s≤T
3.1 Some Parameter Sensitivities for the Joint Default Probability
We consider two firms and simulate the default probabilities in the classical
approach. We fix the single firm first passage default probability at 2% and
Fig. 8. (a) Impact of α on the JDP (blue: JDP in T , red: first passage JDP).
(b) Impact of λ on the first passage JDP (blue: α = 0, red: α = 1)
Fig. 9a,b. JDP for different ρ (blue: JDP in T , black: first passage JDP)
1% (therefor we adapt the default thresholds K1 and K2 ). Figure 8a shows

the impact of α on the joint default probability (JDP). The blue curve is the
JDP in maturity T , the red curve is the first passage JDP. The bigger the
α the smaller the JDP. Figure 8b shows the impact of the intensity λ for
different α. The blue curve is a jump diffusion model (α = 0) the red curve is
a shot-noise model (α = 1). The curves are increasing in λ (as expected) and
the shot-noise curve is lower than the jump diffusion curve.
Figure 9 shows the effect of different correlations. The JDP is increasing
in ρ.
3.2 Valuation Of CDOs
We simulate 125 firms and evaluate the spreads for CDO tranches with at-
tachement points at 0%, 3%, 6%, 9%, 12% and 22%. For definitions, terminol-
ogy and valuations for CDOs we refer e.g. to [3] or [11]. The single firm default
probabilities are fixed, we consider two kinds of firms.
First, we check the influence of ρ. As we can see in Fig. 10a the effect is not
the same for all tranches. For the equity tranche a higher correlation implies
a higher probability for no defaults, less risk and a lower spread. For the senior
Fig. 10a–d. blue: equity, red: first mezzanine, green: second mezzanine, orange:
third mezzanine, pink: senior
tranche a higher correlation implies a higher probability for many defaults,

more risk and a higher spread. For the mezzanine tranches both behaviours
are possible.
For increasing α the behaviour of the tranches differs, too (see Fig. 10b).
The JDP is decreasing in α. A smaller JDP is like a smaller correlation, it
implies a higher equity spread and a lower senior spread.
Figure 10c shows, that the effect of different times to maturity is as ex-
pected (long time → big risk → high spread). The first tranches react more
sensitive to time.
Increasing λ causes an increasing JDP-curve which provokes different
changes of the spreads of the different tranches (see Fig. 10d).
Calibration to the Market
We calibrate our model to the market data (5-year-iTraxx-Data, October

2006). Table 1 shows the results for normal distributed jumps and for ex-
ponential distributed jumps. The sum of the differences is 0.538% in the nor-
mal case and 0.1923% in the exponential case. In both cases we have a good
fit, but the exponential case fits a bit better. In the normal case, the first
Table 1. 5-year-CDO
tranche market price normal jumps exponential jumps
0%–3% 13.0% 13.24% 12.97%
3%–6% 59.25 bp 82.09 bp 47.42 bp
6%–9% 13.25 bp 12.03 bp 10.92 bp
9%–12% 5.25 bp 1.96 bp 4.22 bp
12%–22% 2.6 bp 0.16 bp 1.56 bp
mezzanine spread ist overestimated, the higher tranches are underestimated.

In the exponential case, the spreads for the tranches higher than equity are
underestimated.
Implied Correlation
Comparable to the implied volatility for options in the model of Black and
Scholes [2] we can evaluate the implied correlation for CDOs. Generally the
market observes different implied correlations for the different tranches, the so
called correlation smile (cf. volatility smile). We show that our model produces
a correlation smile, too.
We calculate the spreads for different fictive portfolios, assume that the
spreads are the real ones and evaluate the implied correlations via the Vasicek-
model (see [23]). The first two portfolios are the resulting models of the calibra-
tion section, model 1 is the one with normal jumps, model 2 with exponential
jumps. Model 3 is a pure diffusion model (λ = 0). The other portfolios have
normal distributed jumps. Model 4 is a jump diffusion model (λ = 4, α = 0)
Table 2. Models with normal and exponential distributed jumps

Model 1 Model 2
Tranche CDO ρ CDO ρ
0%–3% 13.24% 0.657 12.97% 0.726
3%–6% 82.09 bp 0.361 47.42 bp 0.272
6%–9% 12.03 bp 0.373 10.92 bp 0.348
9%–12% 1.96 bp 0.369 4.22 bp 0.389
12%–22% 0.16 bp 0.373 1.56 bp 0.449
Table 3. Diffusion-, Jump-Diffusion- and Shot-Noise-Model

Model 3 Model 4 Model 5
Tranche CDO ρ CDO ρ CDO ρ
0%–3% 3.05% 0.822 32.28% 0.880 16.18% 0.573
3%–6% 6.93 bp 0.269 860.07 bp 0.814 59.51 bp 0.321
6%–9% 0.089 bp 0.250 646.74 bp 0.813 6.50 bp 0.334
9%–12% n.a. n.a. 512.68 bp 0.343 0.81 bp 0.335
12%–22% n.a. n.a. 341.83 bp 0.754 0.06 bp 0.342
and model 5 a shot-noise model (λ = 4, α = 3). The results are shown in

Tables 2 and 3.
The implied correlation is higher for the equity and the senior tranche.
Table 3 shows that adding jumps causes higher implied correlations, a shot-
noise effect reduces the implied correlations.
4 Summary
The asset-value model is extended in two directions.

First time changed process, which can be viewed as time dependent or
stochastic volatility models, provide a flexible tool. They provide the possibil-
ity to calibrate the model to several points in the term structure of defaults or
in the CDS curve. Additionally, features of the spread dynamic could be repro-
duced. But, currently no special features of the spread dynamics are available,
to test the model on real data. Also in terms of dependency modeling they
deliver an additional degree of freedom going beyond the standard modeling
based on correlated Brownian motions. The common time change, or the de-
pendency structure in the multivariate time changes enables the introduction
of these additional dependency features.
This is also true for shot-noise process. Here common jumps in the asset
value process yield to an additional joint behavior of default going beyond
correlations. The sensitivity of the Joint Default Probabilities to the different
kind of dependency parameters, like correlation and common jump intensity is
analysed in detail. This additional degree of freedom enables also a consistent
way to price all tranches in a CDO structure properly. Therefore shot-noise
processes are able to reproduce the so-called correlation skew observed in the
liquidly traded market of standard single tranche CDOs.
References
1. F. Black and J.C. Cox (1976). Valuing corporate securities: Some effects of bond
indenture provisions, Journal of Finance 31, 351–367.
2. F. Black and M. Scholes (1973). The pricing of options and corporate liabilities.
Journal of Political Economy 81, 637–653.
3. C. Bluhm and L. Overbeck (2006). Structured Credit Portfolio Analysis, Baskets
& CDOs. Chapman & Hall/CRC Financial Mathematics Series; CRC Press.
4. C. Bluhm, L. Overbeck and C. Wagner (2003). An Introduction to Credit Risk
Modeling. Chapman & Hall/CRC Financial Mathematics Series; 2nd Reprint;
CRC Press.
5. S. Bochner (1949). Diffusion equation and stochastic processes. Journal of Math-
ematics 35, 368–370.
6. A. Borodin and P. Salminen (2002). Handbook of Brownian Motion – Facts and
Formulae, 2nd edition. Birkhäuser.
7. P. Carr and L. Wu (2003). Time-changed Lévy processes and option pricing.

Journal of Financial Economics 71, 113–141.
8. P.K. Clark (1973). A subordinated stochastic process model with finite variance
for speculative prices. Econometrica 41, 135–155.
9. J.C. Cox, J. Ingersoll and S. Ross (1985). A theory of the term structure of
interest rates. Econometrica 53, 385–407.
10. D. Duffie & K. Singleton. Princeton (2003).
11. D. Duffie and N. Garleanu (2001). Risk and valuation of collateralized debt
obligations, Financial Analysts Journal 57 (1), 41–59.
12. W. Feller (1966). Infinitely divisible distributions and Bessel functions associated
with random walks. SIAM Journal of Applied Mathematics 14, 864–875.
13. H. Geman, D. Madan and M. Yor (2000). Time changes for Lévy processes.
Mathematical Finance 11, 79–96.
14. J.M. Harrison (1985). Brownian Motion and Stochastic Flow Systems. Wiley.
15. N. Ikeda and S. Watanabe (1981). Stochastic Differential Equations and Diffu-
sion Processes. North-Holland and Kodansha.
16. R. A. Jarrow, D. Lando and S. M. Turnbull (1997). A Markov Model for the
Term Structure of Credit Risk Spreads. Review of Financial Studies 10, 481–523.
17. R.C. Merton (1974). On the pricing of corporate debt: the risk structure of
interest rates. Journal of Finance 29, 449–470.
18. I. Monroe (1978). Processes that can be embedded in Brownian Motion, The
Annals of Probability 6(1), 42–56.
19. B. Øksendal (1990). When is a stochastic integral a time change of a diffusion?
Journal of Theoretical Probability 3(2), 207–226.
20. L. Overbeck and W. Schmidt (2005). Modeling default dependence with thresh-
old models. Journal of Derivatives 12(4), 10–19.
21. J.A. Rebholz (1994). Planar diffusions with applications to Mathematical fi-
nance. Thesis for Ph.D., University of California, Berkeley.
22. W. Schoutens (2003). Lévy processes in Finance. Wiley.
23. O. Vasicek (1987). Probability of loss on loan portfolio, KMV Corporation,
[http://www.moodyskmv.com/research/whitepaper/Probability of Loss on
Loan Portfolio.pdf].
24. C. Zhou (2001). An analysis of default correlations and multiple defaults. Review
of Financial Studies 14(2), 555–576.
Economic Capital Modelling and Basel II
Compliance in the Banking Industry
Klaus Böcker1 and Claudia Klüppelberg2

1
Risk Integration, Reporting & Policies – Risk Analytics and Methods –
UniCredit Group, Munich Branch, klaus.boecker@hvb.de
2
Center for Mathematical Sciences, Munich University of Technology,
85747 Garching bei München, Germany, cklu@ma.tum.de
1 Introduction
It would be a mistake to conclude that the only way to succeed in banking
is through ever-greater size and diversity. Indeed, better risk management
may be the only truly necessary element of success in banking.
Alan Greenspan, Speach to the American Bankers Association, 10/5/2004.
Risk is an inevitable part of every financial institution, above all banks and
insurance companies. Risks are implicitly accepted when such institutions
provide their financial services to customers and explicitly when they take
risk positions that offer profitable, above-average returns. There is no unique
view on risk and usually it is considered in certain sub-classes such as market
risk, credit risk and operational risk, also interest rate risk and liquidity risk.
Market risk is associated with trading activities; it is defined as the poten-
tial loss arising from adverse price changes of a bank’s positions in financial
markets and encompasses interest rate, foreign exchange, equity and credit-
spread risk. Credit risk is defined as potential losses arising from a customer’s
default or loss of credit rating. Such risks usually include loan default risk,
counterparty risk, issuer risk and country risk. Finally, operational risk is due
to losses resulting from inadequate or failed internal processes, human errors,
technological breakdowns, or from external events.
Moreover, risk can be distinguished by the negative effects and poten-
tial hazards it has on different kinds of stakeholders, e.g risks may seriously
threaten the firm’s market value (shareholders’ perspective), create losses to
their lenders (debtholders’ perspective), or jeopardizing the stability of the
financial system (regulators’ perspective). Though the individual interests of
these groups may be rather diverse, all parties are interested in an continued
existence of the institution. Hence, a bank needs a certain amount of capi-
tal relative to its risk as a buffer against future potential losses. This capital
296 K. Böcker, C. Klüppelberg
base must be sufficient so that also very unlikely losses, measured at a high
confidence level, can be absorbed.
The growing awareness of risk inherent in banking industry is partially
owing to spectacular crunches like the Saving & Loans crisis in the 1970s or
the Japanese banking crisis in the 1990s and led to an increasing demand for
banking supervision at the international level, finally resulting in the Basel
Committee of Banking Supervision under the auspices of the Bank for Interna-
tional Settlement (BIS) in Basel. The basic idea underlying modern banking
regulation is pretty simple, namely that banks should quantify their risks
and then are required to keep a certain amount of equity capital (the so-
called “capital charge”) as a buffer against it. For instance, the minimum
capital ratio according to the “Basel Accord” should be 8 % of the so-called
“risk-weighted assets”, although some regulators set different target levels for
individual banks, which may be substantially higher than 8 %.
The first important proposal of the Committee was the “1988 Accord”, and
even though it was primarily dealing with rather crude methods for assessing
credit risk, “Basel I” was a major step towards a common framework for
calculating minimum capital standards for international banks. In 1996 the
Committee then released an amendment to the Basel I Accord where banks
were allowed to build sophisticated internal models for calculating capital
charges for their market risk exposures.
The new Basel Accord “Basel II” [BII04], which should be fully imple-
mented by year-end 2007, describes a more comprehensive risk measure and
minimum standard for capital adequacy and is structured in three Pillars.
Pillar I imposes new methodologies of calculating regulatory capital, thereby
mainly focusing on credit risk and operational risk. For the latter, banks can
then use – similar as it is already the case for market risk – their own internal
modelling techniques (commonly referred to as advanced measurement ap-
proaches (AMA)) to determine capital charges, and we consider this subject
again in Sect. 2.
Pillar II then introduces the so-called Internal Capital Adequacy Assess-
ment Process (ICAAP) and contains guidance to supervisors on how they
should review an institution’s ICAAP. Besides the treatment of so-called
“other” risks that are not covered under Pillar I such as interest rate risk
or credit concentration risk, it deals with an institution’s overall risk expo-
sure. According to the Committee of European Banking Supervisors [CEBS],
banks should calculate an “overall capital number” as an integral part of their
ICAAP. This single-number metric should encompass all risks related to dif-
ferent businesses and risk types. Above all, regulators want to understand
the extent to which the institution has introduced diversification and corre-
lation effects when aggregating different risk types. A particularly important
example of this issue is considered in Sect. 3 where the inter-risk correlation
between credit and market risk is investigated.
A milestone in mathematical finance was the idea of dynamic replica-
tion introduced in 1973 by Fischer Black, Myron Scholes and Robert C.
Mathematical Modelling of Economic Capital in the Banking Industry 297
Merton [BS73], revolutionizing the theory of pricing and hedging of finan-

cial derivatives completely. Then, since the introduction of internal market
risk models in 1996, quantitative risk management has become an interesting
and fruitful research area for mathematicians and statisticians; cf. Föllmer &
Klüppelberg [FK02].
Although our project focussed at the beginning on credit risk problems
alone with results documented in Hillebrand [H06], our industry partner was
interested in further collaboration in operational risk and aggregation of dif-
ferent risk types, more precisely in aggregation of market and credit risk.
As these are new areas with many interesting open problems, we henceforth
concentrate on these cutting-edge topics.
Our paper is organised as follows. In Sect. 2 we suggest a novel method
for calculating operational risk at a high confidence level by using the new
concept of Lévy copulas. Our results can be used as an approximation for
operational Value-at-Risk and deliver important insights into extremal de-
pendence modelling in general. In Sect. 3 we then investigate the interaction
between a credit portfolio and another risk type, which can be thought of
as market risk. Combining Merton-like factor models for credit risk with lin-
ear factor models for market risk, we analytically calculate their inter-risk
correlation and show how inter-risk correlation bounds can be derived. For
known inter-risk correlation the total aggregated credit and market risk can
be approximated (cf. (20) below). We conclude with a discussion of possible
overlapping risk and indicate the assignment problem of a simple financial
instrument to one specific risk like operational, credit or market risk.
2 Analytical Approximation of Operational Risk

One of the determinants of Basel II is Operational Risk, defined as losses
resulting from inadequate or failed internal processes, human errors, techno-
logical breakdowns, or from external events. Risk in all categories of Basel
II is defined as Value-at-Risk (VAR) of the total loss (per year) at a certain
confidence level κ near 1. If we denote by S this total loss, then VAR(κ) is
the capital amount such that total losses remain below VAR with at least
probability κ. This is a rather simplistic risk measure; it only becomes non-
trivial because the total loss S is not a straightforward quantity. Below we
concentrate on the advanced measurement approach (AMA) and indicate the
problems involved for obtaining VAR(κ). It is important to note that the Basel
Committee specifies as quantitative standards a confidence level of κ = 0.999
and only models, which capture potentially severe tail loss events.
2.1 The Loss Distribution Approach
A required feature of AMA for measuring operational risk in the context of

Pillar II is that it allows for explicit correlations between different operational
risks, usually classified according to an event type/business line matrix con-

sisting of eight business lines and seven loss event types. The core problem
here is the multivariate modelling and how the dependence structure between
different matrix cells affects a bank’s total operational risk. The prototypical
loss distribution approach (LDA) assumes that, for each cell i = 1, . . . , d, the
cumulated operational loss Si (t) up to time t is described by an aggregate loss
process

Ni (t)
Si (t) = Xki , t ≥ 0, (1)
k=1
where for each i the sequence (Xki )k∈N are independent and identically dis-
tributed (iid) positive random variables with distribution function Fi describ-
ing the magnitude of each loss event (loss severity), and (Ni (t))t≥0 counts
the number of losses in the time interval [0, t] (called frequency), independent
of (Xki )k∈N . For regulatory capital and economic capital purposes, the time
horizon is usually fixed to t = 1 year. The bank’s total operational risk is then
given as
S + (t) := S1 (t) + S2 (t) + · · · + Sd (t) , t ≥ 0. (2)
The present literature suggests to model dependence between different op-

erational risk cells by means of different concepts, which basically split into
models for frequency dependence on the one hand and for severity dependence
on the other hand.
Here we suggest a model based on the new concept of Lévy copulas (see e.g.
Cont & Tankov [CT04]), which models dependence in frequency and severity
simultaneously, yielding a model with comparably few parameters. Moreover,
our model has the same advantage as a distributional copula: the dependence
structure between different cells can be separated from the marginal processes
Si for i = 1, . . . , d. This approach allows for closed-form approximations for
operational VAR (OpVAR).
2.2 Dependent Operational Risks and Lévy Copulas
In accordance with a recent survey of the Basel Committee on Banking Su-

pervision about AMA practices at financial services firms, we assume that
the loss frequency processes Ni in (1) follows a homogeneous Poisson process
with rate λi > 0. Then the aggregate loss (1) constitutes a compound Poisson
process and is therefore a Lévy process .
A key element in the theory of Lévy processes is the notion of the so-
called Lévy measure. A Lévy measure controls the jump behaviour of a Lévy
process and, therefore, has an intuitive interpretation, in particular in the
context of operational risk. The Lévy measure of a single operational risk cell
measures the expected number of losses per unit time with a loss amount in
a prespecified interval. For our compound Poisson model, the Lévy measure
Πi of the cell process Si is completely determined by the frequency parameter
λi > 0 and the distribution function Fi of the cell’s severity: Πi ([0, x)) :=
λi P (X i ≤ x) = λi Fi (x) for x ∈ [0, ∞). The corresponding one-dimensional
tail integral is defined as
Π i (x) := Πi ([x, ∞)) = λi P (X i > x) = λi F i (x) . (3)
Our goal is modelling multivariate operational risk. Hence, the question is

%Ni (·) i
how different one-dimensional compound Poisson processes Si (·) = k=1 Xk
can be used to construct a d-dimensional compound Poisson process S =
(S1 , S2 , . . . , Sd ) with in general dependent components. It is worthwhile to
recall the similar situation in the case of the more restrictive setting of static
random variables. It is well-known that the dependence structure of a random
vector can be disentangled from its marginals by introducing a distributional
copula. Similarly, a multivariate tail integral
Π(x1 , . . . , xd ) = Π([x1 , ∞) × · · · × [xd , ∞)) , x ∈ [0, ∞]d , (4)
can be constructed from the marginal tail integrals (3) by means of a Lévy
copula. This representation is the content of Sklar’s theorem for Lévy proc-
esses with positive jumps, which basically says that every multivariate tail
integral Π can be decomposed into its marginal tail integrals and a Lévy
copula C2 according to
2 1 (x1 ), . . . , Π d (xd )) ,
Π(x1 , . . . , xd ) = C(Π x ∈ [0, ∞]d . (5)
For a precise formulation of this Theorem we refer to Cont & Tankov

[CT04], Theorem 5.6. Now we can define the following prototypical LDA
model.
Definition 1 (Multivariate Compound Poisson Model).

(1) All aggregate loss processes Si for i = 1, . . . , d are compound Poisson
processes with tail integral Π i (·) = λi Fi (·).
(2) The dependence between different cells is modelled by a Lévy copula
C2 : [0, ∞)d → [0, ∞), i.e. the tail integral of the d-dimensional compound
Poisson process S = (S1 , . . . , Sd ) is defined by
2 1 (x1 ), . . . , Π d (xd )).

Π(x1 , . . . , xd ) = C(Π
2.3 The Bivariate Clayton Model
A bivariate model is particularly useful to illustrate how dependence modelling

via Lévy copulas works. Therefore, we now focus on two operational risk cells
as in Definition 1(1). The dependence structure is modelled by a Clayton
Lévy copula, which is similar to the well-known Clayton copula for distri-
+
Fig. 1. Decomposition of the domain of the tail integral Π (z) for z = 6 into
+
a simultaneous loss part Π (z) (orange area) and independent parts Π ⊥1 (z) (solid
black line) and Π ⊥2 (z) (dashed black line)
bution functions and parameterized by ϑ > 0 (see Cont & Tankov [CT04],
Example 5.5):
2ϑ (u, v) = (u−ϑ + v −ϑ )−1/ϑ ,

C u, v ≥ 0 .
This copula covers the whole range of positive dependence. For ϑ → 0 we ob-
tain independence and then, as we will see below, losses in different cells never
occur at the same time. For ϑ → ∞ we get the complete positive dependence
Lévy copula given by C2 (u, v) = min(u, v). We now decompose the two cells’
aggregate loss processes into different components (where the time parameter
t is dropped for simplicity),
N

N ⊥1
1 1
S1 = S⊥1 + S1 = X⊥k + Xl ,
k=1 l=1 (6)
N

N ⊥2
2 2
S2 = S⊥2 + S2 = X⊥m + Xl ,
m=1 l=1
where S1 and S2 describe the aggregate losses of cell 1 and 2 that is gen-
erated by “common shocks”, and S⊥1 and S⊥2 describe aggregate losses of
one cell only. Note that apart from S1 and S2 , all compound Poisson pro-
cesses on the right-hand side of (6) are mutually independent. The frequency
of simultaneous losses is given by

2ϑ (λ1 , λ2 ) = lim Π 2 (x) = lim Π 1 (x) = λ−ϑ + λ−ϑ −1/ϑ =: λ ,
C 1 2
x↓0 x↓0
which shows that the number of simultaneous loss events is controlled by

the Lévy copula. Obviously, 0 ≤ λ ≤ min(λ1 , λ2 ), where the left and right
Fig. 2. Two-dimensional LDA Clayton Pareto model (with Pareto tail index α =
1/2) for different parameter values. Left column: compound processes, right column:
frequencies and severities. Upper row: δ = 0.3 (low dependence), middle row: δ = 2
(medium dependence), lower row: δ = 10 (high dependence)
bounds refer to ϑ → 0 and ϑ → ∞, respectively. Consequently, in the case of

independence, losses never happen at the same instant of time.
Also the severity distributions of X1 and X2 as well as their dependence
structure are determined by the Lévy copula. To see this, define the joint
survival function as
1 2
F (x1 , x2 ) := P X1 > x1 , X2 > x2 = Cϑ Π 1 (x1 ), Π 2 (x2 ) (7)
λ
Fig. 3. Visualisation of the cells’ loss frequencies controlled by the Clayton Lévy
copula for λ1 = 10 000 and λ2 = 100. Left blue axis: frequency λ of the simultaneous
loss processes S1 and S2 as a function of the Lévy Clayton copula parameter ϑ
(blue, dashed line). Right orange axis: frequency λ⊥1 of the independent loss process
S⊥1 of the first cell as a function of the Lévy Clayton copula parameter ϑ (orange,
solid line)
with marginals
1 2
F 1 (x1 ) = lim F (x1 , x2 ) = Cϑ (Π 1 (x1 ), λ2 ) (8)
x2 ↓0 λ
1 2
F 2 (x2 ) = lim F (x1 , x2 ) = Cϑ (λ1 , Π 2 (x2 )) . (9)
x1 ↓0 λ
In particular, it follows that F1 and F2 are different from F1 and F2 , respec-
tively. To explicitly extract the dependence structure between the severities of
simultaneous losses X1 and X2 we use the concept of a distributional survival
copula. Using (7)–(9) we see that the survival copula Sϑ for the tail sever-
ity distributions F 1 (·) and F 2 (·) is the well-known distributional Clayton
copula; i.e.
Sϑ (u, v) = (u−ϑ + v −ϑ − 1)−1/ϑ , 0 ≤ u, v ≤ 1 .
For the tail integrals of the independent loss processes S⊥1 and S⊥2 . we obtain
for x1 , x2 ≥ 0
2ϑ (Π 1 (x1 ), λ2 ) ,
Π ⊥1 (x1 ) = Π 1 (x1 ) − Π 1 (x1 ) = Π 1 (x1 ) − C
2ϑ (λ1 , Π 2 (x2 )) ,
Π ⊥2 (x2 ) = Π 2 (x2 ) − Π 2 (x2 ) = Π 2 (x2 ) − C
so that λ⊥1 = λ1 − λ , λ⊥2 = λ2 − λ .
2.4 Analytical Approximations for Operational VAR
In this section we turn to the quantification of total operational loss encom-

passing all operational risk cells and, therefore, we focus on the total aggregate
loss process S + defined in (2). Our goal is to provide some general insight to
multivariate operational risk and to find out, how different dependence struc-
tures (modelled by Lévy copulas) affect OpVAR, which is the standard metric
in operational risk measurement. We need some notation to define it properly.
The tail integral associated with S + is given by
"
+
d
Π (z) = Π (x1 , . . . , xd ) ∈ [0, ∞)d : xi ≥ z , z ≥ 0. (10)
i=1
For d = 2 we can write

+ +
Π (z) = Π ⊥1 (z) + Π ⊥2 (z) + Π (z) , z ≥ 0, (11)
where Π ⊥1 (·) and Π ⊥2 (·) are the independent jump parts and
+
Π (z) = Π (x1 , x2 ) ∈ (0, ∞)2 : x1 + x2 ≥ z , z ≥ 0,
describes the dependent part due to simultaneous loss events.

Since for every compound Poisson process with intensity λ > 0 and positive
jumps with distribution function F , the tail integral is given by Π(·) = λF (·),
it follows from (11) that the total aggregate loss process S + is again compound
Poisson with frequency parameter and severity distribution
+
+ Π (z)
+
λ = lim Π (z) and F (z) = 1 − F (z) = 1 −
+ +
, z ≥ 0 . (12)
z↓0 λ+
This result proves now useful to determine a bank’s total operational risk
consisting of several cells. Before doing that, recall the definition of OpVAR
for a single operational risk cell (henceforth called stand-alone OpVAR.) For
each cell, stand-alone OpVAR at confidence level κ ∈ (0, 1) and time horizon
t is the κ-quantile of the aggregate loss distribution, i.e.
VARt (κ) = G←
t (κ) = inf{x ∈ R : P (S(t) ≤ x) ≥ κ} . (13)
In Böcker & Klüppelberg [BK05, BK06, BK07a, BK07b] it was shown that
OpVAR at high confidence level can be approximated by a closed-form expres-
sion, if the loss severity is subexponential, i.e. heavy-tailed. As this is common
believe we consider in the sequel this approximation, which can be written as

1−κ
VARt (κ) ∼ F ← 1 − , κ ↑ 1, (14)
EN (t)
where the symbol “∼” means that the ratio of left and right hand side con-
verges to 1. Moreover, EN (t) is the cell’s expected number of losses in the
time interval [0, t]. Important examples for subexponential distributions are
lognormal, Weibull, and Pareto. We want to emphasize already here that such
first order asymptotics work extremely well for heavy-tailed Pareto-like tails,
which are realistic in operational risk. Since the loss frequencies only enter as
their mean EN (t), any sophisticated modelling of the loss number process is
superfluous, see Böcker & Klüppelberg [BK06] for more details. Instead all
effort should be directed into a more accurate modelling of the loss severity
distribution.
Here, we extend the idea of an asymptotic OpVAR approximation to the
multivariate problem. In doing so, we exploit the fact that S + is a compound
Poisson process with parameters as in (12). In particular, if F + is subex-
ponential, we can apply (14) to estimate total OpVAR. Consequently, if we
+
are able to specify the asymptotic behaviour of F (x) as x → ∞ we have
automatically an approximation of VARt (κ) as κ ↑ 1.
To make more precise statements about OpVAR, we focus our analysis on
Pareto distributed severities with distribution function
x −α
F (x) = 1 + , x > 0,
θ
with shape parameters θ > 0 and tail parameter α > 0. Pareto’s law is the
prototypical parametric example for a heavy-tailed distribution and suitable
for operational risk modelling. As a simple consequence of (14), in the case
of a compound Poisson model with Pareto severities (Pareto–Poisson model)
analytic OpVAR is given by
1/α 1/α
λt λt
VARt (κ) ∼ θ −1 ∼θ , κ ↑ 1. (15)
1−κ 1−κ
To demonstrate the kind of results we obtain by such approximation methods

we consider a Pareto–Poisson model, where the severity distributions Fi of
the first (say) b ≤ d cells are tail equivalent with tail parameter α > 0 and
dominant to all other cells, i.e.
α
F i (x) θi F i (x)
lim = , i = 1, . . . , b , lim = 0 , i = b + 1, . . . , d .
x→∞ F 1 (x) θ1 x→∞ F 1 (x)
(16)
In the important cases of complete positive dependence and independence,

closed-form results can be found and may serve as extreme cases concerning
the dependence structure of the model.
Theorem 1. Consider a compound Poisson model with cell processes
S1 , . . . , Sd with Pareto distributed severities satisfying (16). Let VARit (·) be
the stand-alone OpVAR of cell i.
(1) If all cells are completely dependent with the same frequency λ for all cells,
then S + is compound Poisson with parameters
b α
+
+
λ =λ and F (z) ∼ θi z −α , z → ∞ ,
i=1
and total OpVAR is asymptotically given by

b
t (κ) ∼
VAR+ VARit (κ), κ ↑ 1. (17)
i=1
(2) If all cells are independent, then S + is compound Poisson with parameters
b α
+ 1 θi
λ = λ1 + · · · + λd and F (z) ∼ +
+
λi , z → ∞ , (18)
λ i=1 z
and total OpVAR is asymptotically given by
b 1/α
α
VAR⊥t (κ) ∼
+ i
VARt (κ) , κ ↑ 1. (19)
i=1
Theorem 1 states that for the completely dependent Pareto-Poisson model,

total asymptotic OpVAR is simply the sum of the dominating cell’s asymp-
totic stand-alone OpVARs. Recall that this is similar to the new proposals
of Basel II, where the standard procedure for calculating capital charges for
operational risk is just the simple-sum VAR. To put it another way, regulators
implicitly assume complete dependence between different cells, meaning that
losses within different business lines or risk categories always happen at the
same instants of time.
Very often, the simple-sum OpVAR (17) is considered to be the worst
case scenario and, hence, as an upper bound for total OpVAR in general,
which in the heavy-tailed case can be grossly misleading. To see this, assume
the same frequency λ in all cells also for the independent model, and de-
note by VAR+ +
(κ) and VAR⊥ (κ) completely dependent and independent total
OpVAR, respectively.
Then, as explained in detail in [BK06] for heavy-tailed severity data with
F i (xi ) ∼ (xi /θi )−α as xi → ∞, subadditivity of OpVAR is violated because
the sum of stand-alone OpVARs is smaller than independent total OpVAR.
The following table, taken from [RK99], illustrates this.
More general dependence structures can be investigated within the frame-
work of multivariate regular variation. For homogeneous models, in particular
for the Clayton Lévy copula, precise results have been derived in Klüppelberg
and Resnick [KR07] and applied to find OpVAR approximations in Böcker
and Klüppelberg [BK06].
3 Inter-Risk Correlation of Market and Credit Risk

3.1 The Necessity for Risk Aggregation
A core element of modern risk control is the calculation of an aggregated
group-wide risk figure, which is used to evaluate the capital adequacy of a fi-
nancial institution. Until now no standard procedure for risk aggregation has
Table 1. Comparison of total OpVaR for two operational risk cells (each with stand
alone VaR of 100 million) in the case of complete dependence () and independence
(⊥) for different values of α
α VAR+
VAR+ ⊥
1.2 178.2
1.1 187.8
1.0 200.0
200.0
0.9 216.0
0.8 237.8
0.7 269.2
emerged, but a widespread approach in the banking industry is “aggregation

across risk types”, where in a first step marginal, institution-wide loss distri-
butions for all relevant risk types are calculated. These marginal risk figures
describe the group-wide, pre-aggregated risk of a given risk type encompassing
different legal entities, divisions, regions etc. Then, in a second step, the de-
pendence structure between these pre-aggregated risk-type figures is modelled
and finally the total risk can be calculated.
The easiest way of aggregating risks is simply to add up all pre-aggregated
risk-type figures (cf. Theorem 1(1) in the case of different operational risk
figures). Problems with this procedure have been indicated after Theorem 1
and made transparent in Table 1). Consequently, this yields only a very rough
estimate of the bank-wide total risk. Furthermore, banks usually try to reduce
overall risk by accounting for diversification between different risk types –
measured by correlation – because this allows them to reduce expensive equity
capital. Hence, advanced approaches for risk aggregation begin with an ana-
lysis of the dependence structure between different risk types.
Important measures of dependence in the context of risk-type aggregation
are correlation (which models linear dependence); possible non-linear depen-
dence is often modeled by means of copulas. In practise, a widespread ap-
proach for aggregating different risk types is the so-called square-root-formula
approach or variance-covariance approach. Though mathematically justified
only in the case of elliptically distributed risk types (with the multivariate
normal or t distributions as prominent examples), this approach is very often
used as a first approximation because total aggregated capital can then be cal-
culated explicitly without expensive simulations. If X T = (X1 , . . . , Xm ) is the
vector of pre-aggregated risk figures (e.g. economic capital Xi for risk-types
i = 1, . . . , m), and R the inter-risk correlation matrix, then total aggregated
risk Xtot is for elliptically distributed X given by
√
Xtot = X T R X . (20)
Hence, a typical problem of risk aggregation is the estimation of the inter-risk
correlation matrix R.
In the sequel we concentrate on the two-dimensional problem consisting

of credit risk together with another risk type, which henceforth is referred to
as market risk. Credit risk can be more than six times as large as the classical
market risk associated with trading activities, and it is clear that in this case
total risk (20) is mainly dominated by credit risk alone and, in particular,
it is only little affected by inter-risk correlation. However, the exposures of
other market-like risk types like financial investment risk, real estate risk, or
business risk, which are often measured by banks in the context of economic
capital and Basel II compliance, are comparable in volume to overall credit
risk, and the question regarding correct modelling of inter-risk correlation
again becomes important.
We combine a Merton-like factor model for credit risk with a linear factor
model for market risk. Both models are driven by a set of (macroeconomic)
factors Y = (Y1 , . . . , YK ) where the factor weights are allowed to be zero so
that a risk type may only depend on a subset of Y . This section is based
on [BH07].
3.2 Modelling Credit and Market Risk
Normal Factor Model for Credit Risk
To describe credit portfolio loss, we choose a classical structural model as

it can be found e.g. in Bluhm, Overbeck & Wagner [BOW02]. Within these
models, a borrower’s credit quality is driven by a so-called “ability-to-pay”
process. Consider a portfolio of n loans. Then, default of an individual obligor
i ∈ {1, . . . , n} is described by a Bernoulli random variable Li with P(Li = 1) =
pi = 1 − P(Li = 0) where pi is the obligor’s probability of default within time
period [0, T ] for fixed T > 0. Following Merton’s idea, counterparty i defaults if
its asset value log-return Ai falls below some threshold Di , sometimes referred
to as default point, i.e.
Li = 11{Ai <Di } , i = 1, . . . , n . (21)
If we denote the exposure at default (perhaps enriched by discounting factors

and/or net of recovery rates) of an individual obligor by ei , portfolio loss is
given by

n
L(n) = ei L i . (22)
i=1
In a factor-model approach, the asset values Ai are linked to a set of macroe-

conomic factors Y1 , . . . , YK , which are assumed to be normally distributed and
the vector (Y1 , . . . , YK ) has been transformed to standard normal.
Definition 2 (Normal factor model for credit risk). Let Y = (Y1 , . . . , YK )

be a random vector of (macroeconomic) factors with multivariate standard
normal distribution. We assume that each of the asset value log-returns Ai

for i = 1, . . . , n linearly depends on Y as well as on a standard normally
distributed idiosyncratic factor εi (which models the performance of firm i)
independent of Y , i.e.
9
:
K : K
Ai = βik Yk + ;1 − 2 ε ,
βik i i = 1, . . . , n , (23)
k=1 k=1
%K
with factor loadings βik satisfying Ri2 := k=1 βik2
∈ [0, 1], which is that part
of the variance of Ai which can be explained by the systematic factor vector Y .
Then L(n) as given in (22) is called normal factor model for credit risk.
Equation (23) implies that log-returns A1 , . . . , An are standard normally dis-

tributed, but dependent with correlations

K
ρij := corr(Ai , Aj ) = βik βjk , i, j = 1, . . . , n . (24)
k=1
Owing to the normal factor structure of the model, the default point Di of
every obligor is related to its default probability pi by
Di = Φ−1 (pi ) , i = 1, . . . , n , (25)
where Φ is the standard normal distribution function. Moreover, the joint

default probability of two obligors is given by
Φρij (Di , Dj ) , i = j ,
pij := P(Ai ≤ Di , Aj ≤ Dj ) = (26)
pi , i=j,
where Φρij denotes the bivariate normal distribution function with standard-
ized marginals and correlation ρij given by (24). Finally, the default correla-
tion between two different obligors is given by
pij − pi pj
corr(Li , Lj ) = , i, j = 1, . . . , n . (27)
pi (1 − pi ) pj (1 − pj )
Factor Models for Market Risk
We assume that market risk is already pre-aggregated and can be approxi-

mated by a one-dimensional random variable Z, representing the aggregated
profit and loss (P/L) distribution due to changes in some market variables,
such as interest rates or equity prices.
As in the credit risk model of Definition 2, we explain fluctuations of
the P/L random variable Z by means of (macroeconomic) factors Y =
(Y1 , . . . , YK ). We use the same macroeconomic factors for credit and market
risk, where independence of risk from such a factor is indicated by a loading

factor 0.
As we want to add market and credit risk quantities, we use the convention
that losses correspond to positive values of Z. One can think of Y as a vector
describing the healthiness of the economy in the sense that positive (negative)
values of the Yk correspond to a good (bad) economy, implying a decreasing
(increasing) market risk.
Definition 3 (Normal factor model for market risk). Let Y = (Y1 , . . . ,
YK ) be a random vector of (macroeconomic) factors with multivariate standard
normal distribution. Then, the normal factor model for the pre-aggregated
market risk P/L is given by
⎛ 9 ⎞
:

K : K
Z = −σ ⎝ γk Yk + ;1 − γk2 η ⎠ (28)
k=1 k=1
%K
with factor loadings satisfying k=1 γk2 ∈ [0, 1], which is that part of the vari-
ance of Z which can be explained by the systematic factor Y . Furthermore,
η is a standard normally distributed idiosyncratic factor, independent of Y .
Finally, σ is the standard deviation of Z.
Definition 4 (Normal factor model for credit and market risk). Let
Y = (Y1 , . . . , YK ) be a random vector of (macroeconomic) factors with multi-
variate standard normal distribution. Let the credit portfolio loss L(n) be given
by (22) and the asset value log-returns Ai for i = 1, . . . , n are modeled by the
normal factor model (23). Let Z be the pre-aggregated market risk P/L mod-
eled by the normal factor model (28). When the credit model’s idiosyncratic
factors εi for i = 1, . . . , n are independent of η, then we call (L(n) , Z) the
normal factor model for credit and market risk.
In order to account for possible heavy tails for Z we introduce the following
global shock approach.
Definition 5 (Shock model for market risk). Let Y = (Y1 , . . . , YK ) be
a random vector of (macroeconomic) factors with multivariate standard nor-
mal distribution and let η be the standard normally distributed idiosyncratic
factor, independent of Y . Further, let W be a positive random variable, inde-
pendent of Y and η. Then the shock model for the pre-aggregated market risk
P/L is given by the normal mixture model
⎛ 9 ⎞
:
K : K
= −σW ⎝
Z γk Yk + ;1 − γk2 η ⎠ , (29)
k=1 k=1

where σ is a scaling factor. If W = ν/Sν and Sν is a χ2ν distributed random
a tν -model for the pre-
variable with ν degrees of freedom, then we call Z
aggregated market risk P/L.
The mixing variable W can be interpreted as a “global shock” driving the

variance of all factors. Such an overarching shock may occur from political
distress, severe economic recession or some natural disaster.
3.3 Inter-Risk Correlation
We now investigate the correlation between credit risk L(n) and market risk
Z, which is defined as

cov L(n) , Z
corr L , Z =
(n)
. (30)
var(L(n) ) var(Z)
Within our modelling framework, we are able to analytically investigate inter-

risk correlation yielding closed-form results.
First we assume that both market and credit risk have a normally dis-
tributed factor structure.
Theorem 2 (Inter-risk correlation for the normal factor model). Sup-

pose that credit portfolio loss L(n) and market risk Z are described by the
normal factor model of Definition 4. Then correlation between L(n) and Z is
given by
%n r e exp − 1 D2
i i
corr L(n) , Z = i=1 2 i
, (31)
(n)
2π var(L )
where Di is the default point (25)

K
ri := corr(Ai , Z) = βik γk , i = 1, . . . , n, (32)
k=1
and

n
var L(n) = ei ej (pij − pi pj ), (33)
i,j=1
where pij the joint default probability (26).
Proof. Using E(Z) = 0 and that η in (28) is independent of Y (and thus

of Li ), the covariance between L(n) and Z is

n
K
cov L(n) , Z = E ZL(n) = −σ ei γk E(Yk Li ) . (34)
i=1 k=1
Recall the definition of Li in (21) with Ai as in (23), and define for k ∈

{1, . . . , K} 9
:
K :
K
βil Yl + ;1 −
(−k)
A i = β 2 εi . ij
l=1 j=1
l=k
Conditioning on Yk yields for the expectation

E(Yk Li ) = E Yk E 11{Ai <Di } | Yk

(−k)
= E Yk P Ai ≤ Di − βik Yk

Di − βik Yk
= E Yk Φ ,
1 − βik
2
(−k)
where we have used that Ai is normally distributed with variance 1 − βik
2
.
By partial integration and the fact that for the density ϕ of the standard
normal distribution y ϕ(y) has antiderivative ϕ(y), we obtain

∞
Di − βik y
E(Yk Li ) = yΦ ϕ(y) dy
−∞ 1 − βik
2

∞
βik Di − βik y
= − ϕ ϕ(y) dy .
1 − βik
2
−∞ 1 − βik
2
right-hand side is −βik times the density of a random variable U =

The
1 − βik
2 X + β Y for standard normal iid X, Y at point D . Since U is then
ik i
again standard normal, we obtain
2
βik Di
E(Yk Li ) = −βik ϕ(Di ) = − √ e− 2 . (35)
2π
Plugging this into (34) with ri as in (32) this yields

σ
n
Di2
cov L(n) , Z = √ ei ri e− 2 .
2π i=1
Furthermore, from (22) we calculate
n

var L(n) = ei ej E(Li Lj ) − E(Li )E(Lj )
i,j=1
n
= ei ej (pij − pi pj ) ,
i,j=1
where pij is the joint default probability (26).
Note that ri may become negative if (some) factor weights βik and γk have
different signs. Therefore, in principal, also negative inter-risk correlations can
occur between the credit and market portfolio. Typical values for the inter-risk
correlation lie in a range between 10 % and 60 % and vary significantly within
the banking sector. A similar result can be obtained for the shock model of
Definition 5.
Theorem 3 (Inter-risk correlation for the tν factor model). Suppose
that credit portfolio loss L(n) is described by the normal factor model of Def-
inition 2. Denote by Z and Z the market risk described by the normal factor
and by the shock model of Definition 3 and Definition 5, respectively. If W
has finite second moment, then

= E(W ) corr L(n) , Z .
corr L(n) , Z (36)
E(W 2 )
For the tν model with ν > 2 we get

= f (ν) corr L(n) , Z
corr L(n) , Z (37)
with

ν − 2 Γ ν−1
f (ν) := ν2 . (38)
2 Γ 2
Proof. Since E(Z) = 0, we obtain with

= E(W ) cov L(n) , Z
cov L(n) , Z = E W 2 var(Z)
and var Z
that
= E(W ) corr L(n) , Z .
corr L(n) , Z
E(W 2 )

For the tν model with ν > 0 we have W = ν/S, where S is χ2ν distributed
with density
2−ν/2
fν (s) = ν e−s/2 sν/2−1 , s ≥ 0.
Γ 2
It follows for ν > 1 that

1 2−ν/2 ∞ −s/2 ν/2−3/2 Γ ν−1
E √ = ν e s ds = √ ν .
2
S Γ 2 0 2Γ 2
1
Analogously, for ν > 2 we calculate E S = 1
ν−2 . Plugging this into (36)
gives formula (37).
Remark 1. Since E(W ) > 0, by the Cauchy–Schwarz inequality,
E(W )
0< ≤ 1.
E(W 2 )
As a consequence thereof, given a positive inter-risk correlation corr(L(n) , Z) ∈

(0, 1] for normally distributed market risk, introducing a shock into the model
results in a smaller inter-risk correlation (36). For the tν model this situation
is depicted in Fig. 4.
The fact that corr(L(n) , Z) linearly depends on the correlations ri and thus
on the factor loadings γk implies the following Proposition, which can be
used to estimate upper bounds for the inter-risk correlation, when no specific
information about market risk is available.
Proposition 1 (Inter-risk correlation bounds). Suppose that credit port-

folio loss L(n) and market risk Z are described by the normal factor model
of Definition 4. Assume that the market model factor loadings γk for k =
1, . . . , K are unknown. Then correlation between L(n) and Z is bounded by
%n !% 1 2
K
k=1 βik exp − 2 Di
e 2
i=1 i
|corr L , Z | ≤
(n)
≤ 1, (39)
2π var(L(n) )

where var L(n) is given in (33).
Proof. Since the obligor’s exposures ei are assumed to be positive, it follows

from (31) that
% 1 2
i ei |ri | exp − 2 Di
|corr L , Z | ≤ ! %
(n)
.
2π ij ei ej (pij − pi pj )
%K
From k=1 γk2 ≤ 1 it follows by the Cauchy–Schwartz inequality that
K K 1/2 K 1/2 K 1/2

|ri | = βik γk ≤ 2
βik 2
γk ≤ 2
βik .

k=1 k=1 k=1 k=1
%K
The right-hand side is bounded by one, since k=1 γk2 = 1 corresponds to the
correlation of the degenerate case of model (28).
Therefore, solely based on the parametrization of the normal credit factor
model and the assumption of a normally distributed, pre-aggregated market
risk, bounds for the inter-risk correlation can be derived. Moreover, from the
explicit form of (37) in Theorem 3 it is clear that a similar result holds also
for the tν distributed market risk.
Fig. 4. LHP approximations of the inter-risk correlation bound as a function of the

average portfolio rating according to (41). The solid line corresponds to the normal
factor model (formally ν → ∞) and the dashed line to the shock model with ν = 5.
The uniform asset correlation is assumed to be ρ = 10 %
One-Factor Approximations
Instructive examples regarding the inter-risk correlation and its bounds can
be obtained for one-factor models and they are useful to explain general char-
acteristics of inter-risk correlation. As shown in Böcker & Hillebrand [BH07],
Sect. 4.1, such a common one-factor framework for both credit and market
risk can be defined consistently, and in the sequel we want to summarize some
of their results.
Within the one-factor framework, the credit portfolio is assumed to be
homogenous; i.e. for i = 1, . . . , n exposure ei = e, default probability pi = p,
and factor loadings βik = βk for k = 1, . . . , K, i.e. these quantities are the
same for all credits of the portfolio. and both market and %Kcredit risk are
systematically explained only by one single factor Ỹ := √1ρ k=1 βk Yk , which
%K
is a compound of all Yk for k = 1, . . . , K, where ρ := k=1 βk2 is the uniform
asset correlation of the credit portfolio; i.e. for any two asset value log-returns
Ai , Aj the correlation is equal to ρ. The situation simplifies further in the case
of a sufficiently large portfolio, where we consider n → ∞, resulting in the
so-called large homogenous portfolio (LHP) approximation (see also Bluhm,
Overbeck & Wagner [BOW02], Sect. 2.5.1.)
√
L(n) a.s. D − ρ Ỹ
→ Φ √ =: L, n → ∞,
ne 1−ρ
where D = Φ−1 (p) and ne is the total exposure of the credit portfolio. The
LHP approximation plays an important role in the context of credit portfolio
modelling; e.g. it is the underlying assumption in the calculation formula for
regulatory capital charges in the internal-ratings-based (IRB) approach of
Basel II.
Adopting the LHP approximation for the tν market model with the normal
model as formal limit model with limν→∞ f (ν) = 1, inter-risk correlation
simplifies considerably. From (26) we get the joint default probability p12 =
Φρ (D, D) for two arbitray firms in the portfolio, and from (32) we see that
%K
r = k=1 βk γk . Then
−D2 /2
= f (ν) r e
corr(L, Z) , (40)
2π(p12 − p2 )
which is corr(L, Z) for the normal model with f (ν) = 1. The bound (39)
simplifies to
√ −D2 /2
ρe
|corr(L, Z)| ≤ f (ν) . (41)
2π(p12 − p2 )
According to equations (40) and (41), inter-risk correlation and its bound are
functions of the homogeneous asset correlation ρ and the average default prob-
ability p and thus on the average rating structure of the credit portfolio. This
is depicted in Fig. 4 where LHP approximations of the inter-risk correlation
bound are plotted as a function of the average portfolio rating.
A crucial point in the above approximation is the homogeneity of the credit
portfolio. Even if actual credit portfolios are rarely exactly homogenous, the
derived LHP approximations is a useful approximation in practice for the up-
per inter-risk correlation bound. Let us consider the normal factor model and
so equation (41). For a loss distribution of a general credit portfolio (obtained
for instance by Monte Carlo simulation) with expected loss μ, standard devi-
ation ς, and total exposure etot , estimators p̂ and ρ̂ for p and ρ, respectively,
can be found by moment matching; i.e. by comparing the expected loss and
the variance of the simulated portfolio with those of an LHP:
μ̂ = etot p̂ (42)
3 4
ςˆ2 = e2tot p̂12 − p̂2 = e2tot Φρ̂ Φ−1 (p̂), Φ−1 (p̂) − p̂2 . (43)
From (41) we then obtain the following moment estimator for the upper inter-
risk correlation bound
√ < 2 =
−1
ρ̂ exp − 1
Φ (p̂)
2LHP (p̂, ρ̂) = f (ν̂) etot
B √
2
. (44)
ςˆ 2π
4 Conclusion
In this paper we suggested separate models for operational risk, credit risk
and market risk, aiming at an integrated model quantifying the overall risk of
a financial institution. In doing so, we adopted the common idea that “risk”
of a financial position or even an entire bank can be separated into different

risk types.
In general, however, such a silo approach often causes problems when
risk-type definitions are overlapping, or the classification into risk types is
unrealistic or even not possible. We want to present a simple but convincing
example.
Consider a fixed-rate corporate bond, where the investor receives fixed,
regular interest payments (with a rate set at the time the bond is issued) until
the bond matures, called the coupon rate. On one hand, such an investment
bears market risk, in particular interest rate risk: If market interest rates rise,
then the market price of the bond will fall, because new bonds are expected
to be issued with higher coupon rates, making old bonds less attractive. On
the other hand, the bond also has credit risk, since the coupon rate of a bond
also depends on the financial health of the issuer; i.e. on the credit rating of
the company. The higher the company’s default probability is, the less likely
is that it will be able to pay the interest on the bond and to pay-off the bond
at maturity. In this example (and of course also for more complex financial
instruments) it does not make sense to distinguish market from credit risk,
the only threat for the trader is a decrease in the market value of the bond.
Furthermore, professional trading of financial instruments requires a com-
plex IT-infrastructure, and so also bears a significant fraction of operational
risk. However, even in our simple example of the coupon bond, the question
regarding its operational VAR remains unsolved. Similar problems arise in the
context of other Pillar II risk types such as business and strategic risk, which
are currently only poorly considered within a firm’s enterprise risk manage-
ment process. For a novel approach to this particular risk see Böcker [B07].
In accordance with Alan Greenspan we belief that a reliable and function-
ing risk management system is the basis for success in banking. Therefore,
future research has to tackle the problem of how total risk (beyond that of
market, credit and operational risk) can be measured and managed properly.
To achieve such a “grand unified theory” of risk, a more holistic view on risk
instead of the widespread silo approach is called for.
About the Authors/Disclaimer

Klaus Böcker is senior risk controller at UniCredit Group, Munich Branch at
HypoVereinsbank AG. Claudia Klüppelberg holds the chair of Mathematical
Statistics at the Center for Mathematical Sciences of the Munich University
of Technology. The opinions expressed in this article are those of the authors
and do not reflect the views of UniCredit Group or HypoVereinsbank AG.
References
[BII04] Basel Committee on Banking Supervision: International Convergence of
Capital Measurement and Capital Standards, Basel (2004)
[BS73] Black, Scholes Black, F., Scholes, M.: The Pricing of Options and Corpo-
rate Liabilities, Journal of Political Economy 81, 637–654 (1973)
[BOW02] Bluhm, C., Overbeck, L., Wagner, L. An Introduction to Credit Risk
Modeling. Chapman & Hall/CRC, Baco Raton (2003)
[B07] Böcker, K.: Modelling business risk. In Preparation (2007)
[BH07] Böcker, K., Hillebrand, M.: Interaction of market and credit risk: an analy-
sis of inter-risk correlation and risk aggregation. Submitted for publication
(2007)
[BK05] Böcker, K., Klüppelberg, C.: Operational VaR: a closed-form approxima-
tion. RISK, December, 90–93 (2005)
[BK06] Böcker, K., Klüppelberg, C.: Multivariate models for operational risk.
PRMIA “2007 Enterprise Risk Management Symposium Award for Best
Paper: New Frontiers in Risk Management Award”. Submitted for publi-
cation (2006).
[BK07a] Böcker, K., Klüppelberg, C.: Modelling and measuring multivariate oper-
ational risk with Lévy copulas. Submitted for publication (2007)
[BK07b] Böcker, K., Klüppelberg, C.: Multivariate operational risk: dependence
modelling with L copulas. 2007 ERM Symposium Online Monograph,
Society of Actuaries, and Joint Risk Management section newsletter. To
appear (2007).
[CEBS] Committee of European Banking Supervisors (CEBS): Application of the
Supervisory Review Process under Pillar 2 (CP03 revised), Consultation
Paper (2005)
[CT04] Cont, R., Tankov, P.: Financial Modelling with Jump Processes. Chapman
& Hall/CRC, Baco Raton (2004)
[EKM97] Embrechts, P., Klüppelberg, C., Mikosch, T.: Modelling Extremal Events
for Insurance and Finance. Springer, Berlin (1997)
[FK02] Föllmer, H., Klüppelberg, C.: Finanzmathematik. In: Deutsche For-
schungsgemeinschaft (Hrsg.) Perpektiven der Forschung und ihrer För-
derung. Aufgaben und Finanzierung 2002–2006. Wiley-VHC, Weinheim
(2002)
[H06] Hillebrand, M.: Modelling and estimating dependent loss given default.
RISK, September, 120–125 (2006)
[KR07] Klüppelberg, C., Resnick, S.I.: The Pareto copula, aggregation of risks
and the Emperor’s socks. Submitted (2007)
[RK99] Rootzén, H. and Klüppelberg, C. (1999) A single number can’t hedge
against economic catastrophes. Ambio 28, No 6, 550–555. Royal Swedish
Academy of Sciences.
Numerical Simulation for Asset-Liability
Management in Life Insurance
Thomas Gerstner1 , Michael Griebel1 , Markus Holtz1 , Ralf Goschnick2 ,

and Marcus Haep2
1
Institut für Numerische Simulation, Universität Bonn, Nussallee 15, 53115 Bonn,
Germany, griebel@ins.uni-bonn.de
2
Zürich Gruppe Deutschland, Poppelsdorfer Allee 25–33, 53115 Bonn, Germany
marcus.haep@zurich.com
Summary. New regulations and stronger competitions have increased the demand
for stochastic asset-liability management (ALM) models for insurance companies in
recent years. In this article, we propose a discrete time ALM model for the simula-
tion of simplified balance sheets of life insurance products. The model incorporates
the most important life insurance product characteristics, the surrender of contracts,
a reserve-dependent bonus declaration, a dynamic asset allocation and a two-factor
stochastic capital market. All terms arising in the model can be calculated recur-
sively which allows an easy implementation and efficient evaluation of the model
equations. The modular design of the model permits straightforward modifications
and extensions to handle specific requirements. In practise, the simulation of stochas-
tic ALM models is usually performed by Monte Carlo methods which suffer from
relatively low convergence rates and often very long run times, though. As alterna-
tives to Monte Carlo simulation, we here propose deterministic integration schemes,
such as quasi-Monte Carlo and sparse grid methods for the numerical simulation of
such models. Their efficiency is demonstrated by numerical examples which show
that the deterministic methods often perform much better than Monte Carlo simu-
lation as well as by theoretical considerations which show that ALM problems are
often of low effective dimension.
1 Introduction
The scope of asset-liability management is the responsible administration of
the assets and liabilities of insurance contracts. Here, the insurance company
has to attain two goals simultaneously. On the one hand, the available capital
has to be invested as profitably as possible (asset management), on the other
hand, the obligations against policyholders have to be met (liability man-
agement). Depending on the specific insurance policies these obligations are
often quite complex and can include a wide range of guarantees and option-
like features, like interest rate guarantees, surrender options (with or without
surrender fees) and variable reversionary bonus payments. Such bonus pay-
ments are typically linked to the investment returns of the company. Thereby,
320 T. Gerstner et al.
the insurance company has to declare in each year which part of the invest-
ment returns is given to the policyholders as reversionary bonus, which part is
saved in a reserve account for future bonus payments and which part is kept by
the shareholders of the company. These management decisions depend on the
financial situation of the company as well as on strategic considerations and
legal requirements. A maximisation of the shareholders’ benefits has to be bal-
anced with a competitive bonus declaration for the policyholders. Moreover,
the exposure of the company to financial, mortality and surrender risks has to
be taken into account. These complex problems are investigated with the help
of ALM analyses. In this context, it is necessary to estimate the medium- and
long-term development of all assets and liabilities as well as the interactions
between them and to determine their sensitivity to the different types of risks.
This can either be achieved by the computation of particular scenarios (stress
tests) which are based on historical data, subjective expectations, and guide-
lines of regulatory authorities or by a stochastic modelling and simulation.
In the latter case, numerical methods are used to simulate a large number
of scenarios according to given distribution assumptions which describe the
possible future developments of all important variables, e.g. of the interest
rates. The results are then analysed using statistical figures which illustrate
the expected performance or the risk profile of the company.
In recent years, such stochastic ALM models for life insurance policies
are becoming more and more important as they take financial uncertainties
more realistically into account than an analysis of a small number of deter-
ministically given scenarios. Additional importance arises due to new regula-
tory requirements as Solvency II and the International Financial Reporting
Standard (IFRS). Consequently, much effort has been spent on the develop-
ment of these models for life insurance policies in the last years, see, e.g.,
[2, 4, 7, 13, 19, 24, 33] and the references therein. However, most of the ALM
models described in the existing literature are based on very simplifying as-
sumptions in order to focus on special components and effects or to obtain
analytical solutions. In this article, we develop a general model framework for
the ALM of life insurance products. The complexity of the model is chosen
such that most of the models previously proposed in the literature and the
most important features of life insurance product management are included.
All terms arising in the model can be calculated recursively which allows an
straightforward implementation and efficient evaluation of the model equa-
tions. Furthermore, the model is designed to have a modular organisation
which permits straightforward modifications and extensions to handle specific
requirements.
In practise, usually Monte Carlo methods are used for the stochastic sim-
ulation of ALM models. These methods are robust and easy to implement but
suffer from their relatively low convergence rates. To obtain one more digit ac-
curacy, Monte Carlo methods need the simulation of a hundred times as many
scenarios. As the simulation of each scenario requires a run over all time points
and all policies in the portfolio of the company, often very long run times are
Numerical Simulation for Asset-Liability Management in Life Insurance 321
needed to obtain approximations of satisfactory accuracy. As a consequence,

a more frequent and more comprehensive risk management, extensive sensitiv-
ity investigations or the optimisation of product parameters and management
rules are often not possible. In this article we propose deterministic numerical
integration schemes, such as quasi-Monte Carlo methods (see e.g. [37, 22, 40])
and sparse grid methods (see, e.g., [9, 15, 16, 23, 38, 42]) for the numerical
simulation of ALM models. These methods are alternatives to to Monte Carlo
simulation, which have a faster rate of convergence, exploit the smoothness
and the anisotropy of the integrand and have deterministic upper bounds on
the error. In this way, they often can significantly reduce the number of re-
quired scenarios and computing times as we show by numerical experiments.
The performance of these numerical methods is closely related to the effec-
tive dimension and the smoothness of the problem under consideration. Here,
we show that ALM problems are often of very low effective dimension (in
the sense that the problem can well be approximated by sums of very low-
dimensional functions) which can, to some extent, explain the efficiency of
the deterministic methods. Numerical results based on a general ALM model
framework for participating life insurance products demonstrate that these
deterministic methods in fact often perform much better than Monte Carlo
simulation even for complex ALM models with many time steps. Quasi-Monte
Carlo methods based on Sobol sequences and dimension-adaptive sparse grids
based on one-dimensional Gauss–Hermite quadrature formulae turn out to be
the most efficient representatives of several quasi-Monte Carlo and sparse grid
variants, respectively. For further details, see [17, 18, 19].
The remainder of this article is as follows: In Sect. 2, we describe the
model framework. In Sect. 3, we then discuss how this model can be efficiently
simulated by numerical methods for multivariate integration. In Sect. 4, we
present numerical results which illustrate possible application of the ALM
model and analyse the efficiency of different numerical approaches. The article
finally closes in Sect. 5 with concluding remarks.
2 The ALM Model

In this section, we closely follow [19] and describe an ALM model framework
for the simulation of the future development of a life insurance company. We
first indicate the overall structure of the model and introduce a simplified
balance sheet which represents the assets and liabilities of the company. The
different modules (capital market model, liability model, management model)
and the evolution of the balance sheet items are then specified in the following
sections.
2.1 Overall Model Structure

The main focus of our model is to simulate the future development of all assets
and liabilities of a life insurance company. To this end, the future develop-
Fig. 1. Overall structure of the ALM model
ment of the capital markets, the policyholder behaviour and the company’s
management has to be modelled. We use a stochastic capital market model,
a deterministic liability model which describes the policyholder behaviour and
a deterministic management model which is specified by a set of management
rules which may depend on the stochastic capital markets. The results of the
simulation are measured by statistical performance and risk figures which are
based on the company’s most important balance sheet items. They are used
by the company to optimise management rules, like the capital allocation, or
product parameters, like the surrender fee. The overall structure of the model
is illustrated in Fig. 1.
We model all terms in discrete time. Here, we denote the start of the
simulation by time t = 0 and the end of the simulation by t = T (in years).
The interval [0, T ] is decomposed into K periods [tk−1 , tk ] with tk = k Δt,
k = 1, . . . , K and a period length Δt = T /K of one month.
The asset side consists of the market value Ck of the company’s assets at
time tk . On the liability side, the first item is the book value of the actuarial
reserve Dk , i.e., the guaranteed savings part of the policyholders after deduc-
tion of risk premiums and administrative costs. The second item is the book
value of the allocated bonuses Bk which constitute the part of the surpluses
that have been credited to the policyholders via the profit participation. The
free reserve Fk is a buffer account for future bonus payments. It consists of
surpluses which have not yet been credited to the individual policyholder
accounts, and is used to smooth capital market oscillations and to achieve
a stable and low-volatile return participation of the policyholders. The last
item, the equity or company account Qk , consists of the part of the surpluses
which is kept by the shareholders of the company and is defined by
Qk = Ck − Dk − Bk − Fk
such that the sum of the assets equals the sum of the liabilities. Similar to
the bonus reserve in [24], Qk is a hybrid determined as the difference between
a market value Ck and the three book values Dk , Bk and Fk . It may be
interpreted as hidden reserve of the company as discussed in [29]. The balance
sheet items at time tk , k = 0, . . . , K, used in our model are shown in Table 1.
In a sensitivity analysis for sample parameters and portfolios it is shown in [19]
Table 1. Simplified balance sheet of the life insurance company

Assets Liabilities
Capital Ck Actuarial reserve Dk
Allocated bonus Bk
Free reserve Fk
Equity Qk
that this model captures the most important behaviour patterns of the balance
sheet development of life insurance products. Similar balance sheet models
have already been considered in, e.g., [2, 3, 24, 33, 29].
2.2 Capital Market Model
We assume that the insurance company invests its capital either in fixed in-
terest assets, i.e., bonds, or in a variable return asset, i.e., a stock or a bas-
ket of stocks. For the modelling of the interest rate environment we use the
Cox-Ingersoll-Ross (CIR) model [11]. The CIR model is a one-factor mean-
reversion model which specifies the dynamics of the short interest rate r(t) at
time t by the stochastic differential equation

dr(t) = κ(θ − r(t))dt + r(t)σr dWr (t), (1)
where Wr (t) is a standard Brownian motion, θ > 0 denotes the mean reversion
level, κ > 0 denotes the reversion rate and σr ≥ 0 denotes the volatility
of the short rate dynamic. In the CIR model, the price b(t, τ ) at time t of
a zero coupon bond with a duration of τ periods and with maturity at time
T = t + τ Δt can be derived in closed form by
b(t, τ ) = A(τ ) e−B(τ ) r(t) (2)
as an exponential affine function of the prevailing short interest rate r(t) with
2κθ/σr2
2he(κ̂+h)τ Δt/2 2(ehτ Δt − 1)
A(τ ) = , B(τ ) = ,
2h + (κ̂ + h)(ehτ Δt − 1) 2h + (κ̂ + h)(ehτ Δt − 1)

and h = κ̂2 + 2σr2 . To model the stock price uncertainty, we assume that the
stock price s(t) at time t evolves according to a geometric Brownian motion
ds(t) = μs(t)dt + σs s(t)dWs (t), (3)
where μ ∈ R denotes the drift rate and σs ≥ 0 denotes the volatility of the
stock return. By Itô’s lemma, the explicit solution of this stochastic differential
equation is given by
2
s(t) = s(0) e(μ−σs /2)t+σs Ws (t) . (4)
Usually, stock and bond returns are correlated. We thus assume that the two
Brownian motions satisfy dWs (t)dWr (t) = ρdt with a constant correlation
coefficient ρ ∈ [−1, 1]. These and other models which can be used to simulate
the bond and stock prices are discussed in detail, e.g., in [6, 25, 28].
In the discrete time case, the short interest rate, the stock prices and the
bond prices are defined by rk = r(tk ), sk = s(tk ) and bk (τ ) = b(tk , τ ). For the
solution of equation (1), we use an Euler-Maruyama discretization3 with step
size Δt, which yields
√
rk = rk−1 + κ(θ − rk−1 )Δt + σr |rk−1 | Δt ξr,k , (5)
where ξr,k is a N (0, 1)-distributed random variable. For the stock prices one
obtains √ “ √ ”
μ−σs2 /2)Δt+σs Δt ρξr,k + 1−ρ2 ξs,k
sk = sk−1 e( , (6)
where ξs,k is a N (0, 1)-distributed random variable independent of ξr,k . Since

Cov ρξr,k + 1 − ρ2 ξs,k , ξr,k = ρ,
the correlation between the two Wiener processes Ws (t) and Wr (t) is re-
spected. More information on the numerical solution of stochastic differential
equations can be found, e.g., in [22, 30].
2.3 Management Model
In this section, we discuss the capital allocation, the bonus declaration mech-
anism and the shareholder participation.
Capital Allocation
We assume that the company rebalances its assets at the beginning of each
period. Thereby, the company aims to have a fixed portion β ∈ [0, 1] of its
assets invested in stocks, while the remaining capital is invested in zero coupon
bonds with a fixed duration of τ periods. We assume that no bonds are sold
before their maturity. Let Pk be the premium income at the beginning of
period k and let Ck−1 be the total capital at the end of the previous period.
The part Nk of Ck−1 + Pk which is available for a new investment at the
beginning of period k is then given by

τ −1
Nk = Ck−1 + Pk − nk−i bk−1 (τ −i),
i=1
3
An alternative to the Euler-Maruyama scheme, which is more time consuming
but avoids time discretization errors, is to sample from a noncentral chi-squared
distribution, see [22]. In addition, several newer approaches exist to improve the
balancing of time and space discretization errors, see, e.g., [21]. This and the time
discretization error are not the focus of this article, though.
where nj denotes the number of zero coupon bonds which were bought at
the beginning of period j. The capital Ak which is invested in stocks at the
beginning of period k is then determined by
Ak = max{min{Nk , β(Ck−1 + Pk )}, 0} (7)
so that the side conditions 0 ≤ Ak ≤ β(Ck−1 +Pk ) are satisfied. The remaining
money Nk − Ak is used to buy nk = (Nk − Ak )/bk−1 (τ ) zero coupon bonds
with duration τ Δt.4 The portfolio return rate pk in period k resulting from
the above allocation procedure is then determined by
−1

τ
pk = ΔAk + nk−i Δbk,i /(Ck−1 + Pk ), (8)
i=0
where ΔAk = Ak (sk /sk−1 −1) and Δbk,i = b(tk , τ −i−1)−b(tk−1 , τ −i) denote
the changes of the market values of the stock and of the bond investments
from the beginning to the end of period k, respectively.
Bonus Declaration
In addition to the fixed guaranteed interest, a variable reversionary bonus is

annually added to the policyholder’s account, which allows the policyholder
to participate in the investment returns of the company (contribution prin-
ciple). The bonus is declared by the company at the beginning of each year
(principle of advance declaration) with the goal to provide a low-volatile, sta-
ble and competitive return participation (average interest principle). Various
mathematical models for the declaration mechanism are discussed in the lit-
erature. In this article, we follow the approach of [24] where the declaration
is based on the current reserve rate γk−1 of the company, which is defined in
our framework by the ratio of the free reserve to the allocated liabilities, i.e.,
Fk−1
γk−1 = .
Dk−1 + Bk−1
The annual interest rate is then defined by
ẑk = max{ẑ, ω(γk−1 − γ)}.
Here, ẑ denotes the annual guaranteed interest rate, γ ≥ 0 the target reserve
rate of the company and ω ∈ [0, 1] the distribution ratio or participation
coefficient which determines how fast excessive reserves are reduced. This way,
a fixed fraction of the excessive reserve is distributed to the policyholders if the
reserve rate γk−1 is above the target reserve rate γ while only the guaranteed
4
Note that due to long-term investments in bonds it may happen that Nk < 0.
This case of insufficient liquidity leads to nk < 0 and thus to a short selling of
bonds.
interest is paid in the other case. In our model this annual bonus has to be
converted into a monthly interest

(1 + ẑk )1/12 − 1 if k mod 12 = 1
zk =
zk−1 otherwise
which is given to the policyholders in each period k of this year.
Shareholder Participation
Excess returns pk − zk , conservative biometry and cost assumptions as well as
surrender fees lead to a surplus Gk in each period k which has to be divided
among the free reserve Fk and the equity Qk . In case of a positive surplus,
we assume that a fixed percentage α ∈ [0, 1] is saved in the free reserve while
the remaining part is added to the equity account. Here, a typical assumption
is a distribution according to the 90/10-rule which corresponds to the case
α = 0.9. If the surplus is negative, we assume that the required capital is
taken from the free reserve. If the free reserves do not suffice, the company
account has to cover the remaining deficit. The free reserve is then defined by
Fk = max{Fk−1 + min{Gk , α Gk }, 0}. (9)
The exact specification of the surplus Gk and the development of the equity
Qk is derived in Sect. 2.5.
2.4 Liability Model

In this section, we discuss the modelling of the decrement of policies due to
mortality and surrender and the development of the policyholder’s accounts.
Decrement Model
For efficiency, the portfolio of all insurance contracts is often represented
by a reduced number m of model points. Each model point then represents
a group of policyholders which are similar with respect to cash flows and tech-
nical reserves, see, e.g., [27]. By pooling, all contracts of a model point expire
at the same time which is obtained as the average of the individual maturity
times.
We assume that the development of mortality and surrender is given de-
terministically and modelled using experience-based decrement tables. Let qki
and uik denote the probabilities that a policyholder of model point i dies or
surrenders in the k-th period, respectively. The probabilities qki typically de-
pend on the age, the year of birth and the gender of the policyholder while uik
often depends on the elapsed contract time. Let δki denote the expected num-
ber of contracts in model point i at the end of period k. Then, this number
evolves over time according to
i
δki = 1 − qki − uik δk−1 . (10)
We assume that no new contracts evolve during the simulation.
Insurance Products
In the following, we assume that premiums are paid at the beginning of a pe-
riod while benefits are paid at the end of the period. Furthermore, we assume
that all administrative costs are already included in the premium. For each
model point i = 1, . . . , m, the guaranteed part of the insurance product is
defined by the specification of the following four characteristics:
• premium characteristic: (P1i , . . . , PKi
) where Pki denotes the premium of
an insurance holder in model point i at the beginning of period k if he is
still alive at that time.
• survival benefit characteristic: E1i,G , . . . , EK i,G
where Eki,G denotes the
guaranteed payments to an insurance holder in model point i at the end
of period k if he survives period k.
• death benefit characteristic: T1i,G , . . . , TK i,G
where Tki,G denotes the
guaranteed payment to an insurance holder in model point i at the end
of period k if he dies in period
k.
• surrender characteristic: S1i,G , . . . , SKi,G
where Ski,G denotes the guaran-
teed payment to an insurance holder in model point i at the end of period
k if he surrenders in period k.
The bonus payments of the insurance product to an insurance holder in model
point i at the end of period k in case of survival, death and surrender, are
denoted by Eki,B , Tki,B and Ski,B , respectively. The total payments Eki , Tki and
Ski to a policyholder of model point i at the end of period k in case of survival,
death and surrender are then given by
Eki = Eki,G + Eki,B , Tki = Tki,G + Tki,B and Ski = Ski,G + Ski,B . (11)
The capital of a policyholder of model point i at the end of period k is col-

lected in two accounts: the actuarial reserve Dki for the guaranteed part and
the bonus account Bki for the bonus part. Both accounts can efficiently be
computed in our framework using the recursions
1+z i,G qki
Dki = (D i
+ P i
) − E − T i,G (12)
1 − qki k−1 k k
1 − qki k
and
1 + zk i zk − z i i,B qki
Bki = B + (D + P i
) − E − T i,B (13)
1 − qki k−1
1 − qki k−1 k k
1 − qki k
which results from the deterministic mortality assumptions, see, e.g., [2, 46].
Example 1. As a sample insurance product, an endowment insurance with

death benefit, constant premium payments and surrender option is considered.
Let P i denote the constant premium which is paid by each of the policyholders
in model point i in every period. If they are still alive, the policyholders receive
a guaranteed benefit E i,G and the value of the bonus account at maturity di .
In case of death prior to maturity, the sum of all premium payments and the
value of the bonus account is returned. In case of surrender, the policyholder
capital and the bonus is reduced by a surrender factor ϑ = 0.9. The guaranteed
components of the four characteristics are then defined by
Pki = P i , Eki,G = χk (di ) E i,G , Tki,G = k P i and Ski,G = ϑDki ,
where χk (di ) denotes the indicator function which is one if k = di and zero
otherwise. The bonus payments at the end of period k are given by
Eki,B = χk (di )Bki , Tki,B = Bki and Ski,B = ϑ Bki .
We will return to this example in Sect. 3.
2.5 Balance Sheet Model
In this section, we derive the recursive development of all items in the simpli-
fied balance sheet introduced in Sect. 2.1.
Projection of the Assets
In order to define the capital Ck at the end of period k, we first determine

the cash flows which are occurring to and from the policyholders in our model
framework. The premium Pk , which is obtained by the company at the be-
ginning of period k, and the survival payments Ek , the death payments Tk ,
and the surrender payments Sk to policyholders, which take place at the end
of period k, are obtained by summation of the individual cash flows (11), i.e.,

m
m
m
m
i
Pk = δk−1 Pki , Ek = δki Eki , Tk = qki δk−1
i
Tki , Sk = uik δk−1
i
Ski ,
i=1 i=1 i=1 i=1
(14)
where the numbers δki are given by (10). The capital Ck is then recursively
given by
Ck = (Ck−1 + Pk ) (1 + pk ) − Ek − Tk − Sk (15)
where pk is the portfolio return rate defined in equation (8).
Projection of the Liabilities
The actuarial reserve Dk and the allocated bonus Bk are derived by summa-
tion of the individual policyholder accounts (12) and (13), i.e.,

m
m
Dk = δki Dki and Bk = δki Bki .
i=1 i=1
In order to define the free reserve Fk , we next determine the gross surplus Gk in
period k which consists in our model of interest surplus and surrender surplus.
The interest surplus is given by the difference between the total capital market
return pk (Fk−1 + Dk−1 + Bk−1 + Pk ) on policyholder capital and the interest
payments zk (Dk−1 + Bk−1 + Pk ) to policyholders. The surrender surplus is
given by Sk /ϑ − Sk . The gross surplus in period k is thus given by
Gk = pk Fk−1 + (pk − zk ) (Dk−1 + Bk−1 + Pk ) + (1/ϑ − 1)Sk .
The free reserve Fk is then derived using equation (9). Altogether, the com-
pany account Qk is determined by
Qk = Ck − Dk − Bk − Fk .
Note that the cash flows and all balance sheet items are expected values
with respect to our deterministic mortality and surrender assumptions from
Sect. 2.4, but random numbers with respect to our stochastic capital market
model from Sect. 2.2.
Performance Figures
To analyse the results of a stochastic simulation, statistical measures are con-

sidered which result from an averaging over all scenarios. Here, we consider
the path-dependent cumulative probability of default

PDk = P min Qj < 0
j=1,...,k
as a measure for the risk while we use the expected future value E[Qk ] of
the equity as a measure for the investment returns of the shareholders in the
time interval [0, tk ]. Due to the wide range of path-dependencies, guarantees
and option-like features of the insurance products and management rules,
closed-form representations for these statistical measures are in general not
available so that one has to resort to numerical methods. It is straightforward
to include the computation of further performance and risk measures like the
variance, the value-at-risk, the expected shortfall or the return on risk capital.
To determine the sensitivity f (v) = ∂f (v)/∂v of a given performance figure
f to one of the model parameters v, finite difference approximations or more
recent approaches, like, e.g., smoking adjoints [20], can be employed.
3 Numerical Simulation
In this section, we discuss the efficient numerical simulation of the ALM model
described in Sect. 2. The number of operations for the simulation of a single
scenario of the model is of order O(m · K) and takes about 0.04 seconds
on a dual Intel(R) Xeon(TM) CPU 3.06GH workstation for a representative
portfolio with m = 500 model points and a time horizon of K = 120 periods.
The number of scenarios which have to be generated depends on the accu-

racy requirements, on the model parameters5 and on the employed numerical
method. In the following, we first rewrite the performance figures of the model
as high-dimensional integrals. Then, we survey numerical methods which can
be applied to their computation, discuss their dependence on the effective di-
mension and review techniques which can reduce the effective dimension in
certain cases.
3.1 Representation as High-Dimensional Integrals
It is helpful to represent the performance figures of the ALM simulation as

high-dimensional integrals to see how more sophisticated methods than Monte
Carlo simulation can be used for their numerical computation. To derive
such a representation, recall that the simulation of one scenario of the ALM
model is based on 2 K independent normally distributed random numbers
y = (y1 , . . . , y2K ) = (ξs,1 , . . . , ξs,K , ξr,1 , . . . , ξr,K ) ∼ N (0, 1). These numbers
specify the stock price process (6) and the short rate process (5). Then, the
term structure, the asset allocation, the bonus declaration, the shareholder
participation and the development of all involved accounts can be derived us-
ing the recursive equations of the previous sections. Altogether, the balance
sheet items CK , BK , FK and QK at the end of period K can be regarded
as (usually very complicated) deterministic functions CK (y), BK (y), FK (y),
QK (y) depending on the normally distributed vector y ∈ IR2K . As a conse-
quence, the expected values of the balance sheet items at the end of period
K can be represented as 2 K-dimensional integrals, e.g.,

e−y y/2
T
E[QK ] = QK (y) dy (16)

IR2K (2π)K
for the equity account. Often, monthly discretizations of the capital market
processes are used. Then, typical values for the dimension 2K range from
60−600 depending on the time horizon of the simulation.
Transformation
The integral (16) can be transformed into an integral over the 2K-dimensional
unit cube which is often necessary to apply numerical integration methods.
By the substitution yi = Φ−1 (xi ) for i = 1, . . . , 2K, where Φ−1 denotes the
inverse cumulative normal distribution function, we obtain

e−y y/2
T
E[QK ] = QK (y) dy = f (x) dx (17)

IR2K (2π)K [0,1]d
5
The model parameters affect important numerical properties of the model, e.g.
the effective dimension (see Sect. 3.3) or the smoothness.
with d = 2K and f (x) = Qk (Φ−1 (x)). For the fast computation of Φ−1 (xi ),
we use Moro’s method [35]. Note that the integrand (17) is unbounded on
the boundary of the unit cube, which is undesirable from a numerical as well
as theoretical point of view. Note further that different transformations to
the unit cube exist (e.g. using the logistic distribution or polar coordinates)
and that also numerical methods exist which can directly be applied to the
untransformed integral (16) (e.g. Gauss–Hermite rules).
3.2 Numerical Methods for High-Dimensional Integrals
There is a wide range of methods (see, e.g., [12]) available for numerical mul-
tivariate integration. Mostly, the integral (17) is approximated by a weighted
sum of n function evaluations
n
f (x) dx ≈ wi f (xi ) (18)
[0,1]d i=1
with weights wi ∈ IR and nodes xi ∈ IRd . The number n of nodes corresponds

to the number of simulation runs. Depending on the choice of the weights
and nodes, different methods with varying properties are obtained. Here, the
dimension as well as the smoothness class of the function f should be taken
into account.
Monte Carlo
In practise, the model is usually simulated by the Monte Carlo (MC) method.
Here, all weights equal wi = 1/n and uniformly distributed sequences of
pseudo-random numbers xi ∈ (0, 1)2K are used as nodes. This method is
independent of the dimension, robust and easy to implement but suffers from
a relative low probabilistic convergence rate of order O(n−1/2 ). This often
leads to very long simulation times in order to obtain approximations of sat-
isfactory accuracy. Extensive sensitivity investigations or the optimisation of
product or management parameters, which require a large number of simula-
tion runs, are therefore often not possible.
Quasi-Monte Carlo
Quasi-Monte Carlo (QMC) methods are equal-weight rules like Monte Carlo.
Instead of pseudo-random numbers, however, deterministic low-discrepancy
sequences (see, e.g., [37, 22]) or lattices (see, e.g., [40]) are used as point
sets which are chosen to yield better uniformity than random samples. Some
popular choices are Halton, Faure, Sobol and Niederreiter–Xing sequences and
extensible shifted rank-1 lattice rules based on Korobov or fast component-by-
component constructions. From the Koksma–Hlawka inequality it follows that
convergence rate of QMC methods is of order O(n−1 (log n)d ) for integrands
of bounded variation which is asymptotically better than the O(n−1/2 ) rate of
MC. For periodic integrands, lattice rules can achieve convergence of higher
order depending on the decay of the Fourier coefficients of f , see [40]. Using
novel digital net constructions (see [14]), QMC methods can also be obtained
for non-periodic integrands which exhibit convergence rates larger than one if
the integrands are sufficiently smooth.
Product Methods
Product methods for the computation of (17) are easily obtained by using the
tensor products of the weights and nodes of one-dimensional quadrature rules,
like, e.g., Gauss rules (see, e.g., [12]). These methods can exploit the smooth-
ness of the function f and converge with order O(n−s/d ) for f ∈ C s ([0, 1]d ).
This shows, however, that product methods suffer from the curse of dimension,
meaning that the computing cost grows exponentially with the dimension d of
the problem, which prevents their efficient applications for high-dimensional
(d > 5) applications like ALM simulations.
Sparse Grids
Sparse grid (SG) quadrature formulas are constructed using certain com-
binations of tensor products of one-dimensional quadrature rules, see, e.g.,
[9, 15, 23, 38, 42]. In this way, sparse grids can, like product methods, ex-
ploit the smoothness of f and also obtain convergence rates larger than one.
In contrast to product methods, they can, however, also overcome the curse
of dimension like QMC methods to a certain extent. They converge with or-
der O(n−s (log n)(d−1)(s−1) ) if the integrand belongs to the space of functions
which have bounded mixed derivatives of order s. Sparse grid quadrature
formula come in various types depending on the one-dimensional basis in-
tegration routine, like the trapezoidal, the Clenshaw-Curtis, the Patterson,
the Gauss-Legendre or the Gauss–Hermite rule. In many cases, the perfor-
mance of sparse grids can be enhanced by local adaptivity, see [5, 8], or by
a dimension-adaptive grid refinement, see [16].
3.3 Impact of the Dimension
In this section, we discuss the dependence of MC, QMC and SG methods on

the nominal and the effective dimension of the integral (17).
Tractability
In contrast to MC, the convergence rate of QMC and SG methods still exhibit
a logarithmic dependence on the dimension. Furthermore, also the constants
in the O-notation depend on the dimension of the integral. In many cases
(particularly within the SG method) these constants increase exponentially
with the dimension. Therefore, for problems with high nominal dimension d,
such as the ALM of life insurance products, the classical error bounds of the
previous section are no longer of any practical use to control the numerical
error of the approximation. For instance, even for a moderate dimension of
d = 20 and for a computationally unfeasibly high number n = 1090 of function
evaluations, n−1 (log n)d > n−1/2 still holds in the QMC and the MC error
bounds. For classical Sobolov spaces with bounded derivatives up to a certain
order, it can even be proved (see [39, 41]) that integration is intractable,
meaning that for these function classes deterministic methods of the form
(18) can never completely avoid the curse of dimension. For weighted Sobolov
spaces, however, it is shown in [39, 41] that integration is tractable if the
weights decay sufficiently fast. In the next paragraph and in Sect. 4.3 we will
give some indications that ALM problems indeed belong to such weighted
function spaces.
ANOVA Decomposition and Effective Dimension
Numerical experiments show that QMC and SG methods often produce much
more precise results than MC methods for certain integrands even in hundreds
of dimensions. One explanation of this success is that QMC and SG methods
can, in contrast to MC, take advantage of low effective dimensions. QMC
methods profit from low effective dimensions by the fact that their nodes
are usually more uniformly distributed in smaller dimensions than in higher
ones. SG methods can exploit different weightings of different dimensions by
a dimension-adaptive grid refinement, see [16]. The effective dimension of the
integral (17) is defined by the ANOVA decomposition, see, e.g., [10]. Here,
a function f : IRd → IR is decomposed by

f (x) = fu (xu ) with fu (xu ) = f (x)dx{1,...,d}\u − fv (xv )
u⊆{1,...,d} [0,1]d−|u| v⊂u
into 2d sub-terms fu with u ⊆ {1, . . . , d} which only depend on variables

xj with j ∈ u. Thereby, the sub-terms fu describe the dependence of the
function f on the dimensions j ∈ u. The effective dimension in the trunca-
tion sense of a function f : IRd → IR %with variance σ 2 (f ) is then defined
as the smallest integer dt , such that v⊆{1,...,dt } σv (f ) ≥ 0.99 σ 2 (f ) where
2
σu2 (f ) denotes the variances of fu . The effective dimension dt roughly de-

scribes the number of important variables of the function f . The effective
dimension % in the superposition sense is defined as the smallest integer ds ,
|v|≤ds σv (f ) ≥ 0.99 σ (f ) where |v| denotes the cardinality of
2 2
such that
the index set v. It roughly describes the highest order of important interac-
tions between variables in the ANOVA decomposition. For the simple function
f (x1 , x2 , x3 ) = x1 ex2 +x2 with d = 3, we obtain dt = 2 and ds = 2 for instance.
For large d, it is no longer possible to compute all 2d ANOVA sub-terms. The
effective dimensions can still be computed in many cases, though. For details
and an efficient algorithm for the computation of the effective dimension in
the truncation sense we refer to [45]. For the more difficult problem to com-
pute the effective dimension in the superposition sense, we use the recursive
method described in [44].
Dimension Reduction
Typically, the underlying multivariate Gaussian process is approximated by

a random walk discretization. In many cases, a substantial reduction of the
effective dimension in the truncation sense and an improved performance of
the deterministic integration schemes can be achieved if the Brownian bridge
or the principal component (PCA) decompositions of the covariance matrix of
the underlying Brownian motion is used instead as it was proposed in [1, 36]
for option pricing problems. The Brownian bridge construction differs from
the standard random walk construction in that rather than constructing the
increments sequentially, the path of the Gaussian process is constructed in
a hierarchical way which has the effect that more importance is placed on the
earlier variables than on the later ones. The PCA decomposition, which is
based on the eigenvalues and -vectors of the covariance matrix of the Brow-
nian motion, maximises the concentration of the total variance of the Brow-
nian motion in the first few dimensions.6 Its construction requires, however,
O(d2 ) operations instead of O(d) operations which are needed for the ran-
dom walk or for the Brownian bridge discretization. For large d, this often
increases the run times of the simulation and limits the practical use of the
PCA construction.
4 Numerical Results
We now describe the basic setting for our numerical experiments and in-
vestigate the sensitivities of the performance figures from Sect. 2.5 to the
input parameters of the model. Then, the risks and returns of two differ-
ent asset allocation strategies are compared. Finally, we compute the effec-
tive dimensions of the integral (17) in the truncation and superposition sense
and compare the efficiency of different numerical approaches for its computa-
tion.
4.1 Setting
We consider a representative model portfolio with 50, 000 contracts which have
been condensed into 500 equal-sized model points. The data of each model
6
Note that without further assumptions on f it is not clear which construction
leads to the minimal effective dimension due to possibly non-linear dependencies
of f on the underlying Brownian motion. As a remedy, also more complicated
covariance matrix decompositions can be employed which take into account the
function f as explained in [26].
Table 2. Capital market parameters p used in the simulation and their partial
derivatives f (p)/f (p) for f ∈ {PDK , E[QK ], E[FK ]}
stock price model interest rate model correlation
μ = 8 % σs = 20 % κ = 0.1 θ = 4 % σr = 5 % r0 = 3 % λ0 = −5 % ρ = −0.1
E[QK ] 0.028 0.035 0.007 0.085 −0.001 0.156 −0.001 −0.0008
E[FK ] 0.039 −0.008 0.009 0.136 −0.0014 0.212 −0.0014 −0.0002
PDK −0.431 0.219 −0.172 −0.884 0.729 −2.122 0.005 0.04
Table 3. Solvency rate, management and product parameters p used in the simu-
lation and their partial derivatives f (p)/f (p) for f ∈ {PDK , E[QK ], E[FK ]}
asset allocation bonus declaration shareholder product parameters solv. rate
β = 10 % τ = 3 ω = 25 % γ = 15 % α = 90 % ϑ = 90 % z = 3 % γ0 = 10 %
E[QK ] 0.083 0.004 −0.002 0.009 −0.101 −0.006 −0.086 0.011
E[FK ] 0.002 0.002 −0.009 0.03 0.013 −0.01 −0.22 0.034
PDK 0.265 −0.054 0 −0.002 0.001 0.08 2.706 −0.504
point i is generated according to the following distribution assumptions: en-

try age xi ∼ N (36, 10), exit age xi ∼ N (62, 4), current age xi0 ∼ U (xi , xi )
and monthly premium P i ∼ U (50, 500) where N (μ, σ) denotes the normal
distribution with mean μ and variance σ, and U (a, b) denotes a uniform dis-
tribution in the interval [a, b]. In addition, the side conditions 15 ≤ xi ≤ 55
and 55 ≤ xi ≤ 70 are respected. The probability that the contracts of a model
point belong to female policyholders is assumed to be 55 %. From the dif-
ference of exit age and current age the maturity time di = xi − xi of the
contracts is computed. As sample insurance product, an endowment insur-
ance with death benefit, constant premium payments and surrender option
is considered as described in Example 1. For simplicity, we assume that the
policies have not received any bonus payments before the start of the simu-
lation, i.e., B0i = 0 for all i = 1, . . . , m. We take the probabilities qki of death
from the DAV 2004R mortality table and choose exponential distributed sur-
render probabilities uik = 1 − e−0.03Δt . At time t0 , we assume a uniform bond
%τ −1
allocation, i.e., nj = (1 − β)C0 / i=0 b0 (i) for j = 1 − τ, . . . , 0. We assume
Q0 = 0 which means that the shareholders will not make additional pay-
ments to the company to avoid a ruin. This way, E[Qk ] serves as a direct
measure for the investment returns of the shareholders in the time interval
[0, tk ]. The total initial reserves of the company are then given by F0 = γ0 D0 .
In the following, we choose a simulation horizon of T = 10 years and a pe-
riod length of Δt = 1/12 years, i.e., K = 120. In our numerical tests we
use the capital market, product and management parameters as displayed in
the second rows of Table 2 and 3 unless stated otherwise. In Table 2 and 3
also the sensitivities f (v)/f (v) (see Sect. 2.5) are displayed for different func-
tions f ∈ {PDK , E[QK ], E[FK ]} and different model input parameter v, e.g.,
∂PDK /(∂μ PDK ) = −0.431.
Fig. 2. Risk-return profiles of the different capital allocation strategies
4.2 Capital Allocation
To illustrate possible applications of the ALM model, we compare the constant-

mix capital allocation strategy of Sect. 2.3 with an CPPI (constant proportion
portfolio insurance) capital allocation strategy (see, e.g., [34]) with respect to
the resulting default risk PDK and returns E[QK ]. Within the CPPI strat-
egy, the proportion of funds invested in (risky) stocks is linked to the current
amount of reserves. The strategy is realised in our model framework by re-
placing β(Ck−1 + Pk ) in equation (7) by β Fk−1 with β ∈ R+ . The resulting
risk-return profiles of the constant-mix strategy and of the CPPI strategy are
displayed in Fig. 2 for different choices of β.
We see that the slightly negative correlation ρ = −0.1 results in a diver-
sification effect such that the lowest default risk is not attained at β = 0 but
at about β = 2.5 % in the constant-mix case and at about β = 40 % in the
CPPI case. Higher values of β lead to higher returns but also to higher risks.
As an interesting result we further see that the CPPI strategy almost always
leads to portfolios with much higher returns at the same risk and is therefore
clearly superior to the constant-mix strategy almost independently of the risk
aversion of the company. The only exception is a constant-mix portfolio with
a stock ratio β of 2.5−4 %, which could be an interesting option for a very
risk averse company.
4.3 Effective Dimension
For the setting of Sect. 4.1, we determine in this section the effective dimen-
sions dt and ds of the integral (17) in the truncation and superposition sense,
respectively, see Sect. 3.3. The effective dimensions depend on the nominal
Table 4. Truncation dimensions dt of the ALM integrand (17) for different nominal
dimensions d and different covariance matrix decompositions
d Random walk Brownian bridge Principal comp.
32 32 7 12
64 64 7 14
128 124 13 12
256 248 15 8
512 496 16 8
dimension d, on the discretization of the underlying Gaussian process and

on all other model parameters. In Table 4, the effective dimensions dt are
displayed which arise by the methods described in [45] for different nominal
dimensions d if the random walk, the Brownian bridge and the principal com-
ponent (PCA) path construction is employed, respectively. One can see that
the Brownian bridge and PCA path construction lead to a large reduction
of the effective dimension dt compared to the random walk discretization. In
the latter case, the effective dimension dt is almost as large as the nominal
dimension d while in the former cases the effective dimensions are almost in-
sensitive to the nominal dimensions and are bounded by only dt = 16 even
for very large dimensions as d = 512. In case of the PCA construction, dt
is even slightly decreasing for large d which is related to the so-called con-
centration of measure phenomenon, see [31]. Further numerical computations
using the method described in [44] show that the ALM problem is also of very
low effective dimension ds in the superposition sense. Here, we only consider
moderately high nominal dimensions due to the computational costs which
increase with d. For d ≤ 32, we obtain that the integral (17) is ‘nearly’ addi-
tive, i.e. ds = 1, independent of d and independent of the covariance matrix
decomposition. Note that the effective dimensions are affected by several pa-
rameters of the ALM model. More results which illustrate how the effective
dimensions in the truncation sense vary in dependence of the capital market
model and of other parameters can be found in [18].
4.4 Convergence Rates
In this section, we compare the following methods for the computation of the
expected value (17) with the model parameters specified in Sect. 4.1:
• MC Simulation,
• QMC integration based on Sobol point sets (see [32, 43]),
• dimension-adaptive SG based on the Gauss–Hermite rule (see [16]).
In various numerical experiments, the Sobol QMC method and the dimension-
adaptive Gauss–Hermite SG method turned out to be the most efficient rep-
resentatives of several QMC variants (we compared Halton, Faure, Sobol low
discrepancy point sets and three different lattice rules with and without ran-
Fig. 3. Errors and required number of function evaluations of the different numerical
approaches to compute the expected value (17) with d = 32 (left) and with d = 512
(right) for the model parameters specified in Sect. 4.1
domisation) and of several SG variants (we compared trapezoidal, Clenshaw–

Curtis, Patterson, Gauss–Legendre and Gauss–Hermite rule and different grid
refinement strategies), respectively. The results for d = 32 and d = 512 are
summarised in Fig. 3 where the number n of function evaluations is displayed
which is needed to obtain a given accuracy. In both cases we used the Brown-
ian bridge path construction for the stock prices and short interest rates. One
can see that the QMC method clearly outperforms MC simulation in both
examples. The QMC convergence rate is close to one and nearly indepen-
dently of the dimension. Moderate accuracy requirements of about 10−3 −10−4
are obtained by the QMC method about 100-times faster as by MC simula-
tion. For higher accuracy requirements, the advantage of the QMC method
is even more pronounced. Recall that these results can not be explained by
the Koksma–Hlawka inequality but by the very low effective dimension of
the ALM problem, see Sect. 4.3. The performance of the SG method deteri-
orates for very high dimensions. In the high dimensional case d = 512, the
SG method is not competitive to QMC. For the moderately high dimension
d = 32, sparse grids are the most efficient method with a very high conver-
gence rate of almost three. With 129 function evaluation already an accuracy
of 10−6 is achieved. Further numerical experiments indicate that the perfor-
mance of the SG method is more sensitive than (Q)MC to different choices
of model parameters which affect the smoothness of the integrand, like more
aggressive bonus declaration schemes and more volatile financial markets.
5 Concluding Remarks
In this article, we first described a discrete time model framework for the
asset-liability management of life insurance products. The model incorporates
fairly general product characteristics, a surrender option, a reserve-dependent
bonus declaration, a dynamic capital allocation and a two-factor stochastic
capital market model. The recursive formulation of the model allows for an
efficient computation of the model equations. Furthermore, the model struc-

ture is modular and allows to be extended easily. Numerical experiments il-
lustrate that the model captures the main behaviour patterns of the balance
sheet development of life insurance products. In the second part of this arti-
cle, we investigated the application of deterministic integration schemes, such
as quasi-Monte Carlo and sparse grid methods for the numerical simulation
of ALM models in life insurance. Numerical results demonstrate that quasi-
Monte Carlo and sparse grid methods can often outperform Monte Carlo
simulation for the ALM of participating life insurance products. Furthermore,
quasi-Monte Carlo methods converge nearly independently of the dimension
and produce even for high dimensions d = 512 more precise results than MC.
Sparse grids are the most efficient method for moderately high dimensions, but
their performance is more sensitive to different choices of model parameters
which affect the smoothness of the integrand and deteriorates for very high
dimensions. In these cases additional transformations are required to improve
the smoothness of the integrand. To explain the efficiency of the determinis-
tic methods we computed the effective dimension of the ALM problem with
and without dimension reduction techniques and showed that ALM problem
are often of very low effective dimension in the truncation and also in the
superposition sense.
Acknowledgement
This research was supported by the program “Mathematics for innovations

in industry and service” of the German Ministry for Education and Research
(BMBF).
References
1. P. Acworth, M. Broadie, and P. Glasserman. A comparison of some Monte Carlo
and quasi-Monte Carlo methods for option pricing. In Monte Carlo and Quasi-
Monte Carlo Methods in Scientific Computing, P. Hellekallek, H. Niederreiter,
(eds.), pages 1–18. Springer, 1998.
2. A. Bacinello. Fair pricing of life insurance participating contracts with a mini-
mum interest rate guaranteed. Astin Bulletin, 31(2):275–297, 2001.
3. A. Bacinello. Pricing guaranteed life insurance participating policies with annual
premiums and surrender option. Astin Bulletin, 7(3):1–17, 2003.
4. L. Ballotta, S. Haberman, and N. Wang. Guarantees in with-profit and unitized
with-profit life insurance contracts: Fair valuation problem in presence of the
default option. Insurance: Math. Ecomonics, 73(1):97–121, 2006.
5. T. Bonk. A new algorithm for multi-dimensional adaptive numerical quadra-
ture. In W. Hackbusch and G. Wittum, editors, Adaptive Methods – Algorithms,
Theory, and Applications, NNFM 46, pages 54–68. Vieweg, 1994.
6. D. Brigo and F. Mercurio. Interest Rate Models – Theory and Practice. Springer,
2001.
7. E. Briys and F. Varenne. On the risk of life insurance liabilities: Debunking

some common pitfalls. J. Risk and Insurance, 64:37–57, 1997.
8. Hans-Joachim Bungartz and Stefan Dirnstorfer. Multivariate quadrature on
adaptive sparse grids. Computing, 71(1):89–114, 2003.
9. Hans-Joachim Bungartz and Michael Griebel. Sparse grids. Acta Numerica,
13:1–123, 2004.
10. R. E. Caflisch, W. J. Morokoff, and A. B. Owen. Valuation of mortgage backed
securities using Brownian bridges to reduce effective dimension. J. Comp. Fi-
nance, 1(1):27–46, 1997.
11. J. Cox, J. Ingersoll, and S. Ross. A theory of the term structure of interest
rates. Econometrica, 53:385–407, 1985.
12. P. Davis and P. Rabinowitz. Methods of Numerical Integration. Academic Press,
1984.
13. M. De Felice and F. Moriconi. Market based tools for managing the life insurance
company. Astin Bulletin, 1(35):79–111, 2005.
14. J. Dick. Walsh spaces containing smooth functions and quasi-monte
carlo rules of arbitrary high order. Availabe at http://www2.maths.
unsw.edu.au/Contacts/profile.php?logname=josi, 2007.
15. T. Gerstner and M. Griebel. Numerical integration using sparse grids. Numerical
Algorithms, 18:209–232, 1998.
16. T. Gerstner and M. Griebel. Dimension–adaptive tensor–product quadrature.
Computing, 71(1):65–87, 2003.
17. T. Gerstner, M. Griebel, and M. Holtz. Numerical simulation methods for asset-
liability management problems in life insurance. In preparation, 2007.
18. T. Gerstner, M. Griebel, and M. Holtz. The Effective Dimension of
Asset-Liability Management Problems in Life Insurance. In C. Fernan-
des, H. Schmidli, and N. Kolev, editors, Proc. Third Brazilian Confer-
ence on Statistical Modelling in Insurance and Finance, pages 148–153,
2007.
19. T. Gerstner, M. Griebel, M. Holtz, R. Goschnick, and M. Haep. A general
asset-liability management model for the efficient simulation of portfolios of life
insurance policies. Insurance: Math. Ecomonics, 2007. In press.
20. M. Giles and P. Glasserman. Smoking adjoints: Fast Monte Carlo greeks. Risk,
19:88–92, 2006.
21. M.B. Giles. Multi-level Monte Carlo path simulation. To appear in Operations
Research, 2007.
22. Paul Glasserman. Monte Carlo Methods in Financial Engineering. Springer,
2003.
23. M. Griebel. Sparse grids and related approximation schemes for higher dimen-
sional problems. In Proceedings of the conference on Foundations of Computa-
tional Mathematics (FoCM05), Santander, 2005.
24. A. Grosen and P. Jorgensen. Fair valuation of life insurance liabilities: The im-
pact of interest rate guarantees, surrender options and bonus policies. Insurance:
Math. Economics, 26(1):37–57, 2000.
25. J Hull. Options, Futures and other Derivative Securities. Prentice Hall, Upper
Saddle River, 2000.
26. Junichi Imai and Ken Seng Tan. Minimizing effective dimension using linear
transformation. In H. Niederreiter, editor, Monte Carlo and Quasi-Monte Carlo
Methods 2002, pages 275–292. Springer, 2004.
27. Reinhold Jaquemod. Abschlussbericht der DAV-Arbeitsgruppe Stochastisches

Unternehmensmodell für deutsche Lebensversicherungen. Schriftenreihe ange-
wandte Versicherungsmathematik. Versicherungswirtschaft, 2005.
28. Ioannis Karatzas and Steven E. Shreve. Methods of Mathematical Finance.
Springer, New York, 1998.
29. A. Kling, A. Richter, and J. Russ. The interaction of guarantees, surplus dis-
tribution, and asset allocation in with-profit life insurance policies. Insurance:
Math. Economics, 40(1):164–178, 2007.
30. Peter E. Kloeden and Eckhard Platen. Numerical Solution of Stochastic Differ-
ential Equations. Springer, 1992.
31. Michel Ledoux. The Concentration of Measure Phenomenon. Mathematical
Surveys and Monographs. American Mathematical Society, 2001.
32. C. Lemieux, M. Cieslak, and K. Luttmer. RandQMC user’s guide:
A package for randomized quasi-Monte Carlo methods in C. Available at
http://www.math.ucalgary.ca/∼lemieux, 2004.
33. K. Miltersen and S. Persson. Guaranteed investment contracts: Distributed and
undistributed excess returns. Scand. Actuarial J., 23:257–279, 2003.
34. Thomas Moller and Mogens Steffensen. Market-Valuation Methods in Life and
Pension Insurance. International Series on Actual Science. Cambridge Univer-
sity Press, 2007.
35. B. Moro. The full Monte. RISK, 8(2), 1995.
36. B. Moskowitz and R. Calfisch. Smoothness and dimension reduction in quasi-
Monte Carlo methods. J. Math. Comp. Modeling, 23:37–54, 1996.
37. H. Niederreiter. Random Number Generation and Quasi-Monte Carlo Methods.
SIAM, Philadelphia, 1992.
38. E. Novak and K. Ritter. High dimensional integration of smooth functions over
cubes. Numer. Math., 75:79–97, 1996.
39. E. Novak and H. Wozniakowski. Intractability results for integration and dis-
crepancy. J. Complexity, 17:388–441, 2001.
40. H. Sloan and S. Joe. Lattice Methods for Multiple Integration. Oxford University
Press, New York, 1994.
41. I.H. Sloan and H. Wozniakowski. When are quasi-Monte Carlo algorithms effi-
cient for high-dimensional integrals? J. Complexity, 14:1–33, 1998.
42. S.A. Smolyak. Quadrature and interpolation formulas for tensor products of
certain classes of functions. Dokl. Akad. Nauk SSSR, 4:240–243, 1963.
43. I. M. Sobol. The distribution of points in a cube and the approximate evaluation
of integrals. J. Comp. Mathematics and Math. Physics, 7:86–112, 1967.
44. X. Wang and I. Sloan. Why are high-dimensional finance problems often of low
effective dimension? SIAM J. Sci. Comput., 27(1):159–183, 2005.
45. Xiaoqun Wang and Kai-Tai Fang. The effective dimension and quasi-Monte
Carlo integration. J. Complexity, 19(2):101–124, 2003.
46. Kurt Wolfsdorf. Versicherungsmathematik Teil 1 Personenversicherung. Teub-
ner, 1986.
On the Dynamics of the Forward Interest Rate
Curve and the Evaluation of Interest Rate
Derivatives and Their Sensitivities
Cristian Croitoru1 , Christian Fries2,3 , Willi Jäger1 , Jörg Kampen1,4 ,

and Dirk-Jens Nonnenmacher2,5
1
Interdisciplinary Center for Scientific Computing (IWR) Heidelberg, INF 368,
69120 Heidelberg, Germany
{cristian.croitoru,jaeger}@iwr.uni-heidelberg.de
2
Dresdner Bank AG, Jürgen-Ponto-Platz 1, 60301 Frankfurt am Main, Germany
3
now at DZBank AG, Platz der Republik, 60265 Frankfurt am Main, Germany
mail@christian-fries.de
4
now at Weierstrass Institute for Applied Analysis and Stochastics,
Mohrenstr. 39, 10117 Berlin, Germany, kampen@wias-berlin.de
5
now at HSH Nordbank AG, Gerhart-Hauptmann-Platz 50, 20095 Hamburg,
Germany
Summary. We present an overview of in project developed new techniques in com-

puting key quantities of financial markets. Our approach is generic in the sense that
the techniques apply essentially in the general frame work of markets which are
described by systems of stochastic differential equations. We exemplify our methods
in the LIBOR market model which is a standard interest rate market model widely
used in practice and has its name from the daily quoted London interbank offered
rates. The LIBOR market model has been developed in recent years beyond the
classical framework in the direction of incomplete market models (with stochastic
volatility and with jumps). Particular challenges are the high dimensionality (up to
20–40 factors), the calibration, and related problems of derivative prices evaluation
and computation of sensitivities. We show how advanced Monte-Carlo techniques
can be combined with analytic results about transition densities in order to ob-
tain highly efficient and accurate numerical schemes for computing some of the key
quantities in financial markets, especially hedging parameters.
1 Introduction
Evaluation of derivatives and their sensitivities leads to mathematical prob-
lems of determining expected values of initial value systems of stochastic dif-
ferential equations and stochastic derivatives described in the framework of
1
Project supported by BMBF Program ‘Mathematics for innovations in industry
and services’ and in cooperation with Dresdner Bank AG
344 C. Croitoru et al.
Malliavin calculus or, equivalently, to Cauchy problems of (systems) of partial

differential equations. Evaluation of high-dimensional interest rate derivatives
and their sensitivities are based typically on Monte Carlo simulation meth-
ods. An alternative (and usually incompatible) method are sparse grids which
were investigated in a preceding project. In praxis one uses Monte-Carlo tech-
niques because they are well-tested for higher dimensionial advanced models,
and because they are practically the most robust with respect to the curse of
dimension and the curse of low regularity. The latter problem occurs mainly
for two types of reasons in finance. The first are inherent low regularity of
solutions of mathematical models such as free boundary problems describing
exotic products, e.g. American derivatives, where the holder of the option has
the right to exercise at any time where the contract is valid. Lower regularity
of derivatives may also occur apart from exotics in market models based on
systems of Feller processes where jump measures depend on the state of un-
derlyings. The second type of reason is low regularity which may not appear
on the mathematical level but appears at least on the numerical or computa-
tional level. It may often be the case that degenerate diffusions can be avoided
on the mathematical level, but come back on the numerical level when we have
to deal with low volatilities. It may even be that we are forced to deal with
degenerate equations for purely computational reasons of dimension reduc-
tion. Indeed in the case of the analytical WKB-representation of transition
densities, which we use here, it turns out that higher order approximations
(orders beyond quadratic precision with respect to time) are computable in
a reasonbale amount of time within the current technical possibilities if some
dimension reduction techniques are applied. There are, however, some sub-
tleties which have to be taken into account when applying the techniques of
analytic approximation of transition densities in order to obtain numerical
schemes based on Monte-Carlo techniques. The first one is that we cannot
simply sample from the analytical WKB-approximation. Therefore a simple
prior scheme is introduced from which we sample. The analytic approximation
is used then to obtain the weights in the Monte-Carlo scheme. Mathemati-
cally, derivatives can simply be obtained by derivation of the analytic WKB-
approximation. The resulting scheme (both for evaluation of derivatives and
sensitivities) was called the proxy scheme in [FrKa]. However, it turned out
that the resulting Monte-Carlo scheme for sensitivities has unbounded vari-
ance if the product of the squared volatility and time is small. Therefore,
a refined Monte-Carlo scheme was proposed in [KKS] which ensures that the
variance is small.
In this paper we shall give an overview about the techniques involved
starting with the description of the basic framework in Sect. 2 and of the
general results for analytic WKB-expansion obtained in [Ka] in Sect. 3. In
Sect. 4 we recall and improve the basic proxy scheme of [FrKa] for evaluation of
financial derivatives using higher order WKB-approximations for the Monte-
Carlo schemes. We also point out how dimension reduction can be put on
a solid mathematical basis (while the details are to complex to describe it
Dynamics of the Forward Interest Rate Curve 345
here and will be dealt within an upcoming paper). In Sect. 5 we describe the
niceties of exploding variance and how this affects the construction of robust
schemes for sensitivities. In Sect. 6 we apply the method to the LIBOR market
model and discuss some numerical results.
2 The Mathematical Framework of Market Models,

Derivatives and Sensitivities
We consider a Markovian system (S, B) of underlying asset process in Rn+ ×R+
(R+ := {x : x > 0}) on a filtered probability space (Ω, F , (Ft )t∈[t0 ,T ] , P ),
consisting of n risky assets S = (S 1 , ..., S n ), and a numeraire B, where the
filtration (Ft ) satisfies the usual conditions and that the system (t, St ) is
Markovian with respect to this filtration. Moreover we assume that S has
an absolute continuous transition kernel with density p(t, x, s, y), which has
derivatives of any order in 0 ≤ t < s, x, y ∈ Rn+ . Further we assume that Bt ,
t > 0, is adapted to (St , 0 ≤ t ≤ s) and is of finite variation. The dynamics
of the system (S, B) is given by
dS i n
dB
= r(t, S)dt + σ ij (t, S)dW j , = r(t, S)dt, 1 ≤ i, j ≤ n, (1)
Si j=1
B
in the (risk-neutral) measure P. In (1) W = (W 1 , ..., W n ) is an adapted

n-dimensional standard Wiener process. Here, (Ft ) is the P -augmentation of
the filtration generated by W . We could include additional jump terms but
since the theory of WKB-expansions of these extended models has not been
published yet we stick to the case of systems which have models with contin-
uous paths. We assume that the process has a transition density p(t, x, s, y)
which is differentiable with respect to x, y ∈ Rn+ , s, t ∈ [t0 , T ], t > s, up to
any order. For an (F· )-stopping time τ and payoff f (Sτ )Bτ contingent claims
are priced at time t0 (assuming B(0) = b0 = 1) by a formula of type
v(t0 , x0 ) = EQ f (Sτt0 ,x0 )
where Q is some equivalent measure (e.g. [Du]). In case of complete markets
this so-called risk-neutral measure is unique. In general, however, markets
are incomplete, and additional criteria and methods have to be introduced
in order to select a risk-neutral measure. Finding the criteria of selection
(modeling) as well as analyzing, computing, and estimating parameters of the
related equations (typically Hamilton–Jacoby equations) is a topic of ongoing
research which will be also considered in the next BMBF-project. Here we
assume that a measure has been chosen and computed or is given (as in the
case of a complete market). For deterministic τ, say τ ≡ T, the value process
of the European claim for t0 ≤ t ≤ T is
Rt
vt := v(t, St , Bt ) := Bt E Ft f (ST ) = e
r(s,Ss )ds
t0
Et f (ST ),
where Et denotes conditional expectation with respect to Ft , and the dis-

counted price process ut := vt /Bt can be expressed in terms of the transition
density via

Ft
ut := u(t, St ) := E f (ST ) = p(t, St , T, y)f (y)dy,
Rn
and where
u(t, s) = p(t, s, T, y)f (y)dy (2)
is the unique solution of the Cauchy problem

⎧
∂u 1 i j ij
n n
⎪
⎪ ∂2u ∂u
⎪
⎨ + s s σσ (t, s) i j + si r(t, s) i = 0,
∂t 2 i,j=1 ∂s ∂s ∂s
i=1 (3)
⎪
⎪
⎪
⎩
u(T, s) = f (s).
It is clear then that p(·, ·, T, y) is the fundamental solution of the latter equa-
tion (3) with p(T, s, T, y) = δ(s − y), where δ the Dirac distribution. The
optimal stopping problems for the pricing of American and Bermudean op-
tions lead to related Cauchy free boundary problems.
3 WKB-Expansions
Next we turn to some general results on analytic expansions of the fundamen-
tal solution (i.e. the transition density) p (cf. [Ka] for details). If we write the
Cauchy problem (3) in logarithmic coordinates xi = ln(si ), then it becomes
a Cauchy problem on the domain [0, T ] × Rn → R where we denote the cor-
responding diffusion coefficients by aij (t, x) and bi (t, x). In order to simplify
the notation we consider the time-homogenous case and drop the dependence
on time t for the moment (we come back to the time-dependent case later in
the context of the LIBOR market model). Pointwise valid analytic expansions
of the fundamental solution p exist if
(A) the matrix norm of (aij (x)) is bounded below and above by 0 < λ < Λ <
∞ uniformly in x,
(B) the smooth functions x → aij (x) and x → bi (x) and all their derivatives
are bounded.
For more subtle (and partially weaker conditions) we refer to [Ka]. Further,
we denote the additional condition
(C) there exists a constant c such that for each multiindex α and for all
1 ≤ i, j, k ≤ n,
∂a ∂b
jk i
α , α ≤ c exp c|x|2 . (4)
∂x ∂x
Then
Theorem 1. If the hypotheses (A), (B) are satisfied, then the fundamental
solution p has the representation
⎛ ⎞
1 d2
(x, y)
p(δt, x, y) = √ ⎝
n exp − + ck (x, y)δtk ⎠ , (5)
2πδt 2δt
k≥0
where d and ck are smooth functions, which are unique global solutions of the
first order differential equations (6), (7) and (9) below. Especially,
n d2
(δt, x, y) → δt ln p(δt, x, y) = − δt ln 2πδt − + ck (x, y)δtk+1
2 2
k≥0
2
is a smooth function which converges to − d2 as δt % 0, where d is the Rie-
%
mannian distance induced by the line element ds2 = ij a−1 ij dxi dxj , where
with a slight abuse of notation (a−1
ij ) denotes the inverse matrix of (aij ). If
the hypotheses (A), (B) and (C) are satisfied, then, in addition, the functions
d and ck , k ≥ 0, equal their Taylor expansion around y globally.
The recursion formulas for d and ck , k ≥ 0, are
1 2
d2 = d aij d2xj , (6)
4 ij xi
where d2xk denotes the derivative of the function d2 with respect to the variable
xk , with the boundary condition d(x, y) = 0 for x = y,
⎛ ⎞
n 1 2 1 ⎝ d2xj ∂c0
⎠
− + Ld + (aij (x) + aji (x)) (x, y) = 0, (7)
2 2 2 i j
2 ∂xi
where !
1
c0 (x, y) = − ln det (aij (y)), (8)
2
and for k + 1 ≥ 1 we obtain

1 d2xi ∂ck+1 d2xj ∂ck+1
(k + 1)ck+1 (x, y) + aij (x) +
2 ij 2 ∂xj 2 ∂xi
1 ∂cl ∂ck−l 1
k
∂ 2 ck ∂ck
= aij (x) + aij (x) + bi (x) ,
2 ij ∂xi ∂xj 2 ij ∂xi ∂xj i
∂xi
l=0
with boundary conditions
ck+1 (x, y) = Rk (y, y) if x = y, (9)

Rk being the right side of (9). We see that the Riemannian distance d has to be
approximated in regular norm in order to get an accurate WKB-expansion in
general. How this can be accomplished is shown in [Ka2]. Designing numerical
schemes we work with approximations both with respect to time and with
respect to spatial variables, of course. In order to analyze the time truncation
error we consider WKB-approximations of the fundamental solution p of the
form

d2 (x, y)
l
1
pl (t, x, T, y) = √ n exp −
k
+ ck (x, y)δt , (10)
2πδt 2δt
k=0
i.e. we assume that the coefficients d2 and ck , 0 ≤ k ≤ l have been computed

up to order l. Let (A)t , (B)t , (C)t denote the analogies to assumptions (A),
(B), (C) for time- and space-dependent coefficients, and let us denote the do-
main of the Cauchy problem by D = (0, T ) × Rn . For integers n ≥ 0 and
real numbers δ ∈ (0, 1) let C m+δ/2,n+δ (D) be the space of m (n) times differ-
entiable functions such that the m-th (n-th) derivative with respect to time
(space) is Hölder continuous with exponent 2δ (δ). Furthermore, |.|m+δ/2,n+δ
denote the natural norms associated with these function spaces. Then a con-
sequence of Safanof’s theorem (cf. [Kr]) is
Theorem 2. Assume that (A)t , (B)t and (C)t are satisfied and let g ∈ C 2+δ
(Rn ) and f ∈ C δ/2,δ (D). If
c ≤ −λ for some λ > 0, (11)
then the Cauchy problem
⎧
⎪
⎪ ∂w 1 ∂2w ∂w
⎪
⎨ ∂t + a ij (t, x) + bi (t, x) + c(t, x)w = f (δt, x) in D
2 ij ∂xi ∂xj i
∂xi
⎪
⎪
⎪
⎩ w(T, x) = g(x) for x ∈ Rn
(12)
has a unique solution w, and there exists a constant c depending only on δ,
n, λ, Λ and K = max{|a|δ , |b|δ , |c|δ } such that
3 4
|w|1+δ/2,2+δ ≤ c |f |δ/2,δ + |g|2+δ . (13)
It can be shown (cf. [Ka, KKS]) that the truncation error
uΔ (t, x) = u(t, x) − ul (t, x), (14)
where
u(t, x) = g(y)p(t, x, T, y)dy, (15)
Rn
and
ul (t, x) = g(y)pl (t, x, T, y)dy, (16)
Rn
satisfies the following
Theorem 3. Assume that conditions (A), (B), and (C) hold and that g ∈
C0δ (Rn ). Then
δ

|u(t, x, y) − ul (t, x, y)|1+δ/2,2+δ = O δtl− 2 (17)
Finer error estimates based on Taylor expansions of WKB-functions ck , k ≥ 0,

will be presented in [FrKa2]. A refined analysis in the special case of reducible
diffusion will be presented in [CrKa]. For the classical LIBOR market model
some simplifications are possible by a global transform to the Laplacian. We
recall (cf. [Ka])
Proposition 1. There is a global coordinate transformation for the operator
in (3) such that the second order part of the transformed operator equals the
Laplacian, if aij = (σσ )ij for a (square) matrix function σ which satisfies

n
∂σik (x)
n
∂σij (x)
σlj (x) = σlk (x), x ∈ Rn . (18)
∂xl ∂xl
l=1 l=1
If the condition of Proposition 1 is satisfied, then coordinate transformation

leads to second order coefficients of the form aij ≡ δij , so that the solution of
(6) becomes
d2 (x, y) = (xi − yi )2 . (19)
i
If conditions (A), (B), (C) and (18) hold, then in the transformed coordinates,
explicit formulas for the coefficient functions ck , k ≥ 0 can be computed via
the formulas
1
c0 (x, y) = (yi − xi ) bi (y + s(x − y))ds, (20)
i 0
and 1
ck+1 (x, y) = RkL (y + s(x − y), y)sk ds, (21)
0
where
1 ∂cl ∂ck−l
k
1 ∂ck
RkL (x, y) = + Δck + bi (x) . (22)
2 i ∂xi ∂xi 2 i
∂xi
l=0
Assuming that bi has a power series representation (here Δx := (x − y))

yγ
byi (x) = bi Δxγ (23)
γ
we obtain the recursion formulas

1 yγ 1
c0 (x, y) = − Δxi bi (y + sΔx)ds = − bi Δxγ+1i (24)
i 0 i γ
1 + |γ|
and
⎧
⎨1
k
ck+1 (x, y) = (βi + 1)(γi + 1)cyl(β+1i ) cy(k−l)(γ+1i )
⎩2
i δ≤α l=0 β+γ=δ
"
1
+ (δi + 2)(δi + 1)ck(δ+2i ) + bi (δi + 1)ck(δ+1i ) pyα
yδ δ
kδ Δx (25)
2

n
where we use, with δP := δi (for more details consider [Ka] and [CrKa]),
i=1
n
1
α
1 > αi !
(y + s(x − y)) s
α k−1
ds = y (α−δ)
Δxδ
0 δP + k i=1 δi !(αi − δi )!
δ=0

α
=: pyα δ
kδ Δx .
δ=0
4 The Proxy Scheme for Evaluation

of Financial Derivatives
We reconsider the proxy scheme of [FrKa]. There exist several refinements of
this scheme (cf. [Fr]), but the basic idea is the following. An option pricing
formula with target scheme S ∗ is reinterpretated in the framework of weighted
Monte-Carlo schemes by an auxiliary proxy scheme S p with weights w:
EQ (f (S ∗ (T ))|Ft ) = EQ (f (S p (T ))w|Ft ) (26)
We consider (conditioned on F0 and at S(0) = 0, for simplification of notation)

p∗ (S )
EQ (f (S )) = f (S )pT (S )dS = f (S ) Tp ppT (S )dS
∗ ∗
(27)
pT (S )
1
n
p∗ (S p (T, ωi ))
≈ f (S p (T, ωi )) Tp p (28)
n i=1 pT (S (T, ωi ))
? @A B
wi weights
where we put the time parameter of the fundamental solution as a subscript

(since time dependence is not essential here), and drop the dependence on
S(0) = 0 (since this reference is always the same and can be understood
implicitly). In principle this idea can be applied for sensitivities. We have
∂ ∗
∂ ∂ p (S )
EQ (f (S ∗ )) = f (S ) (p∗ (S )) dS = f (S ) ∂Xp T ppT (S )dS
∂X ∂X T pT (S )
⎛ ⎞
⎜1 n
p∗ (S p (T, ωi )) ⎟
∂ ⎜ ⎟
≈ ⎜ f (S p (T, ωi )) Tp p ⎟ (29)
∂X ⎝ n i=1 pT (S (T, ωi )) ⎠
? @A B
wi weights
with a parameter or an underlying X (higher derivatives work analogously, of

course). The weighted Monte-Carlo scheme gives us the opportunity to sam-
ple from a simple prior scheme (e.g. an Euler scheme) and then use a more
accurate scheme to get the correct weights. This is to avoid sampling from
a higher order WKB-approximation which is practically impossible. This was
the idea in [FrKa], where numerical schemes of slightly higher order than the
Euler scheme were used as target scheme. In an upcoming paper [FrKa] we
shall use higher order approximation of the LIBOR kernel. However, it turns
out higher order WKB-approximation are computationally expensive for the
full factor model. It is therefore useful to perform a dimension reduction by
principal component analysis at the reference starting point of the scheme.
But there are additional problems here in order to put everything on a solid
mathematical basis. First of all, the transition densities are transition densities
of degenerate diffusions, and the WKB-theory is derived for non-degenerate
diffusions. However, degenerate diffusions behave differently with respect to
uniqueness, existence and regularity. It can be shown that there are solutions
to the degenerate equations which are closest to the non-dengenerate transi-
tion densities (cf. [Ka] for the basic ideas and [FrKa2] for a detailed study).
The second problem is the adaptation of the error estimate. This will be con-
sidereed in [FrKa2]. A third problem was already mentioned in [FrKa]. The
change of measure in the reinterpretation requires that
pp (Ti , S, Ti+1 , S ) = 0 ⇒ p∗ (Ti , S, Ti+1 , S ) = 0.
This can be ensured by an Euler subdiscretization (cf. [Fr]). In equation (29)

we put the derivative ‘outside’ which indicates that the sensitivities can be
computed by numerical differentiation of the whole scheme. In case of the
WKB-approximation of the target scheme we can do the derivative directly
and can avoid numerical differentiation. But it turned out that estimators have
to be refined in order to have bounded variance estimators if the product of
the squared volatility and time is small as was discovered in [KKS], and we
shall describe it in the next section.
5 Refinement of the Proxy Scheme for Computation

of Sensitivities
A detailed analysis (which were originally discovered by numerical observa-
tions of exploding variance for the estimator in (29) for small time) shows
that a simple choice of the prior can lead to exploding variance for small time
or small volatilities. Let us consider this phenomenon (cf [KKS] for more de-
tails): consider a smooth function f : Rn+ → R+ and the transition function
p : Rn+ × Rn+ → R+ , where we drop the parameter T as a subscript of the
transition function because it is of no importance for the following discussion.
We want to estimate probabilistic representations for the integral

I(x) := p(x, y)f (y)dy,
and derivatives such as its gradient

∇x I(x) = ∇x p(x, y)f (y)dy.
Let Y be some random variable with density φ (our prior) on Rn+ , φ > 0. with
samples m Y for m = 1, ..., M, where M is some large integer. Then,
f (Y )
I(x) = E p(x, ζ) (30)
φ(Y )
may be estimated by the unbiased Monte Carlo estimator
1
M
2 f (m Y )
I(x) := p(x,m Y ) . (31)
M m=1 φ(m Y )
Hence, in accordance with the previous section, the estimator corresponding

to
∂ f (Y )
∇x I(x) = E p(x, Y ) (32)
∂x φ(Y )
is
1 ∂
M
C f (m Y )
∇ x I(x) := p(x,m Y ) . (33)
M m=1 ∂x φ(m Y )
However even the natural choice of a prior φ(·) := p(s, σ; x0 , ·), where p is the
lognormal distribution, i.e.
< i
=
>n exp − 1 ln2 y
1 2σ2 s xi0
p(s, σ; x0 , y) := (34)
(2πσ 2 s)n/2 i=1 yi
√ 1
such √that p(s, σ; x0 , ·) is the density of the random variable (x10 eσ sξ , ...,
n
xn0 eσ sξ ), with ξ i , i = 1, ..., d, are i.i.d. standard normal random variables,
leads for the simple choice of f ≡ ||x0 || (a constant of order x0 in magnitude)
to an variance
C∂I ||x0 /xj0 ||2 1
Var j
(x0 ) = (35)
∂x M σ2 s
which explodes when σ 2 s goes to zero. Therefore we choose estimators in
accordance to the following theorem (cf. [KKS] for a proof).
Theorem 4. Let λ be a reference density on Rn with λ(z) = 0 for all z (for

example, the standard normal density). Let ξ be an Rn -valued random variable
with density λ and g : Rn+ × Rn → Rn+ be a smooth map with ∇z g(x, z) = 0,
such that for each x ∈ Rn+ the random variable Y x := g(x, ξ) has a density
φ(x, ·) on Rn+ . Then, we have the probabilistic representation
p(x, Y x )f (Y x ) p(x, g(x, ξ))f (g(x, ξ))

∇x I(x) = E ∇x x
= E ∇x , (36)
φ(x, Y ) φ(x, g(x, ξ))
with corresponding Monte Carlo estimator
1
M
C p(x, g(x,m ξ))f (g(x,m ξ))
∇x I(x) = ∇x . (37)
M m=1 φ(x, g(x,m ξ))
Let | · | denote either a vector norm or a compatible matrix norm. Then it

holds

p(x, g(x, ξ))f (g(x, ξ)) 2
E ∇x ≤ 2M 2 M 2 M 2 + 4M 2 M 2 + 4M 2 M 2 M 2 ,
4 2 3 1 5 1 3 6
φ(x, g(x, ξ))
(38)
hence the second moments of the Monte Carlo samplers for the components
of ∂I/∂x are bounded by the right-hand-side of (38), if for fixed x ∈ Rn+ , there
are α1 , ..., α6 > 1 with
1 1 1 1 1 1 1 1 1 1
+ + = 1, + + = 1, + + + = 1,
α4 α1 α5 α4 α2 α3 α4 α1 α6 α3
such that,

E f 2α1 (g(x, ξ)) = f 2α1 (y)φ(x, y)dy ≤ M12α1 ,

E |∇y f (g(x, ξ))|2α2 = |∇y f (y)|2α2 φ(x, y)dy ≤ M22α2 ,
2α3
∂g
E (x, ξ) ≤ M32α3 ,
∂x
2α4 2α4
p(x, g(x, ξ)) p(x, y)
E = φ(x, y)dy ≤ M42α4 ,
φ(x, g(x, ξ)) φ(x, y)

px (x, g(x, ξ)) φx (x, g(x, ξ)) 2α5
E − =
p(x, g(x, ξ)) φ(x, g(x, ξ))

px (x, y) φx (x, y) 2α5

p(x, y) − φ(x, y) φ(x, y)dy ≤ M52α5 ,
and

py (x, g(x, ξ)) φy (x, g(x, ξ)) 2α6
E − =
p(x, g(x, ξ)) φ(x, g(x, ξ))

py (x, y) φy (x, y) 2α6
− φ(x, y)dy ≤ M62α6 ,
p(x, y) φ(x, y)
with shorthands px := ∇x p, etc.

Similar results can be obtained for higher order derivatives and for derivatives
with respect to parameters.
6 Applications to the LIBOR Market Model

We recall a LIBOR market model with respect to a tenor structure 0 <
T1 . . . < Tn+1 in the terminal measure Pn+1 (induced by the terminal zero
coupon bond Bn+1 (t)). The dynamics of the forward LIBORs Li (t), defined
in the interval [0, Ti ] for 1 ≤ i ≤ n, are governed by the following system of
SDE’s (e.g., see [Sc]),
n
δj Li Lj γi γj
dLi = − dt+ Li γi dWn+1 =: μi (t, L)+ Li γi dWn+1 , (39)
j=i+1
1 + δ j L j
where δi = Ti+1 − Ti are day count fractions and t → γi (t) = (γi,1 (t), . . . ,
γi,d (t)), (γiT γj )ni,j=1 =: ρ are deterministic volatility vector functions defined
in [0, Ti ]. We denote the matrix with rows γiT by Γ and assume that Γ is
invertible. In (39), (Wn+1 (t) | 0 ≤ t ≤ Tn ) is a standard d-dimensional Wiener
process under the measure Pn+1 with d, 1 ≤ d ≤ n, being the number of
driving factors. In what follows, we consider the full-factor LIBOR model
with d = n in the time interval [0; T1 ).
6.1 WKB Approximations for the LIBOR Kernel
In order to apply the recursion formulas for the coefficient functions c0

(cf. (24)) and ck+1 (cf. (25)) we first observe that the transition density
pL (s, u, t, v) of the classical LIBOR process (39) can be transformed to the
equation of form
dYi = μYi (t, Y )dt + dWn+1

i
, 1 ≤ i ≤ n. (40)
For the transitional density pY (s, x, t, y), we can compute ck , k = 0, 1, . . . in (5)

by (20)-(21) with bi (y) = μYi (t, y). After that, we find pL (s, u, t, v) by density
transformation formula. Note that the LIBOR-drift is an analytic function.
Nevertheless, it can be written in form of a power series only locally. In [Ka3]
techniques of regular polynomial interpolation are developed as a byproduct.

Current research in [CrKa] experiments with higher order approximations
of the transition density based on the formulas (24) and (25) and regular
polynomial interpolation. Currently WKB-approximations have been tested
numerically in a full factor LIBOR market model up to c1 (cf. [KKS]). We
finally report on that numerical experiments.
6.2 Case Study: European Swaptions

The estimators (31) and (37) are applied to pricing European swaptions
and computing Deltas in the classical LIBOR market model. For more de-
tails we refer to [KKS]. Simulations for more complex products are currently
done and show an excellent accuraccy of the WKB-expansion. A (payer)
swaption contract payoff with maturity Ti and strike θ with principal $1 is
given by
⎛ ⎞+
Bn+1 (0) ⎝
n
u(L) = Bj+1 (T1 ) (δj Lj (T1 ) − θ)⎠ . (41)
Bn+1 (T1 ) j=i
In the LIBOR market model (39) we take δi ≡ 0.5 when i ≥ 1, flat 3.5 %
initial LIBOR curve and constant volatility loadings
γi (t) ≡ 0.2ei , (ei unit vector of standard basis in Rn )
and with input correlation matrix ρ with components,

|j − i|
ρij = exp ln ρ∞ , 1 ≤ i, j ≤ n (42)
n−1
with n > 2 and ρ∞ = 0.3 (for more general correlation structures we refer to
[Sc]). We consider at-the-money (θ = 3.5 %) swaption over a period [T1 , T19 ].
In our experiments, we take as ϕ a canonical lognormal approximation of
transitional kernel pL
ln (s, x, t, y)

1 >n
Γii−1 (Γ −1 ((log uv11 . . . log uvnn ) − μln (s, t, x))T )2
n exp −
2π(t − s) i=1 vi 2(t − s)
with
⎛ ⎞
n
⎝ |γi | − |γi ||γj |ρij δj xj ⎠
2
i (s, t, x) = (t − s)
μln , 1 ≤ i ≤ n.
2 j=i+1
1 + δj xj
The bias achieves 5 % for European swaptions and 3 % for Deltas, see Table 1
and Table 2. Next we consider the estimators (31) and (37) with payoff (41)
Table 1. (the values are in basis points)

T1 Ibex (SD) Ibln (SD) Ib0 (SD) Ib1 (SD)
0.5 129.6(0.4) 128.9(0.4) 129.1(0.4) 128.4(0.4)
1.0 179.1(0.5) 179.4(0.5) 180.6(0.6) 178.7(0.5)
2.0 243.8(0.8) 246.0(0.8) 251.4(0.8) 245.1(0.8)
5.0 351.2(1.3) 357.8(1.3) 376.3(1.4) 349.4(1.3)
10.0 430.3(2.0) 453.3(2.2) 499.4(2.1) 430.6(1.8)
Table 2. (the values are in basis points)

(h) (h) (h) (h)
T1 d
∂Iex
∂x1
(SD) ∂dIln
∂x1
(SD) d
∂I0
∂x1
(SD) d
∂I1
∂x1
(SD)
0.5 2475.3(5.6) 2470.3(6.0) 2485.4(6.0) 2470.5(6.0)
1.0 2450.6(6.2) 2451.7(6.2) 2480.0(6.6) 2450.1(6.1)
2.0 2401.4(6.4) 2405.2(6.4) 2460.3(6.6) 2400.4(6.4)
5.0 2257.2(7.1) 2261.2(7.2) 2386.7(7.4) 2239.1(6.9)
10.0 2017.9(8.3) 2077.3(8.8) 2299.8(9.0) 2010.2(7.7)
at x = L(0), where
ϕ(x, ·) = pL
ln (0, x, T1 , ·),
0 (0, x, T1 , ·) and p(x, ·) = p1 (0, x, T1 , ·).

p(x, ·) = pL L
denoted by I20 and I21 , correspondingly. A comparison is done with “exact” val-
ues which are obtained by simulating M LIBOR trajectories (39) by log-Euler
scheme with very small time step, Δt = δi /10, and take
1 0,x
M
2
Iex = u m LT1 , (43)
M m=1

0,x+Δi x 0,x+Δi x
C
∂Iex 1 M u
m L T1
− u m L T1
= , 1 ≤ i ≤ n,
∂xi M m=1 2Δi x
Ibln
where Δi x = (Δδij )nj=0 . Analogously to (43), we compute I2ln and ∂∂x i
on
LIBORs simulated according to lognormal approximation of transition kernel
ln (0, x, T1 , ·). Table 1 and Table 2, we show 0-values of European swaptions
pL
and the Deltas, computed via estimators I2ex , I2ln , I20 , I21 and ∂C , C
Iex ∂ Iln C
, ∂I0 ,
∂x1 ∂x1 ∂x1
C∂I1
∂x1 , correspondingly, for different maturities T1 . To compute the values in
the tables, M is taken equal to 3 × 106 and 2 × 106 correspondingly, to keep
standard deviations within 0.5 % relative to the values. The WKB approxi-
mation is computed up to first order, i.e. we have computed c0 and c1 . This
leads to a very close estimate of the European swaptions and Deltas, even for
large maturities. This experiments have been confirmed recently in the case
of more complex products, even Bermudean options. This will be published
in a revised form of [KKS]. In Table 1 and Table 2 T1 denotes the maturity.
References
[CrKa] Croitoru, C., Kampen, J.: Accurate numerical schemes and computation
of a class of linear parabolic initial value problems. (in preparation)
[Du] Duffie, D.: Dynamic Asset Pricing Theory. Princeton, Princeton University
Press (2001)
[FrKa] Fries, C., Kampen, J.: Proxy Simulation Schemes for generic robust Monte-
Carlo sensitivities, process oriented importance sampling and high accu-
racy drift approximation (with applications to the LIBOR market model).
Journal of Computational Finance, Vol. 10, Nr. 2
[FrKa2] Fries, C., Kampen, J.: Proxy Simulation Schemes for generic robust Monte-
Carlo sensitivities based on dimension reduced higher order analytic ex-
pansions of transition densities. (in preparation)
[Fr] Fries, C.: Mathematical Finance. Theory, Modeling, Implementa-
tion. Wiley, Hoboken (2007), http://www.christian-fries.de/finmath/
book
[Ka] Kampen, J.: The WKB-Expansion of the fundamental solution of linear
parabolic equations and its applications. Book, submitted to Memoirs of
the American Mathematical Society (electronically published at SSRN
2006)
[Ka2] Kampen, J.: How to compute the length of a geodesic on a Riemannian
manifold with small error in arbitrarily regular norms. WIAS preprint (to
appear)
[Ka3] Kampen, J.: Regular polynomial interpolation and global approximation
of global solutions of linear partial differential equations. WIAS preprint
(2007)
[KKS] Kampen, J., Kolodko, A., Schoenmakers, J.: Monte Carlo Greeks for
callable products via approximative Greenian Kernels. WIAS preprint, re-
vised version to appear in SIAM Journal of computation (2007)
[Kr] Krylov, N.V.: Lectures on Elliptic and Parabolic Equations in Hölder
Spaces. Graduate Studies in Mathematics, Vol. 12, American Mathemati-
cal Society (1996)
[Sc] Schoenmakers, J.: Robust Libor Modelling and Pricing of Derivative Prod-
ucts. Financial Mathematics. Chapman & Hall/CRC (2005)

Mathematics Key

Uploaded by

Copyright:

Mathematics Key

Uploaded by

Document Information

Copyright

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Mathematics Key

Uploaded by

Copyright:

Mathematics – Key Technology for the Future

Willi Jäger · Hans-Joachim Krebs

ISBN 978-3-540-77202-6 e-ISBN 978-3-540-77203-3

Library of Congress Control Number: 2008923103

Cover design: WMXDesign, Heidelberg

Printed on acid-free paper

tial components of the funded projects. In order to strengthen and document

Heidelberg, Jülich Willi Jäger

Numerical Simulation of Multiscale Models

Numerical Simulation of High-Frequency Circuits

Part II Thin Films

Numerical Methods for the Simulation of Epitaxial Growth

Part III Biochemical Reactions and Transport

Modeling and Simulation of Hairy Root Growth

Part IV Computeraided Medicine

Modeling and Optimization of Correction Measures

Part V Transport, Traﬃc, Energy

Dynamic Routing of Automated Guided Vehicles in

Part VI Risk Management in Finance and Insurance

Advanced Credit Portfolio Modeling and CDO Pricing

Peter Bastian Martin Bodestedt

Klaus Böcker Cristian Croitoru

Georg Denk Herbert Gajewski

Christian Engwer Ralf Goschnik

Christian Fries Natalie Grunewald

Tobias Gerken Frank Haußer

Robert Klöfkorn Alexander Mielke

Stephanie Knorr Rolf Möhring

Dietmar Kröner Frederike Neise

Dirk-Jens Nonnenmacher Werner Römisch

Roland Pulch Thorsten Sickenberger

Björn Stenzel Gregor Wünsch

Caren Tischendorf Urszula Zabarylo

Qimonda AG, Am Campeon 1–12, 85579 Neubiberg, Germany

The challenging simulation problems mentioned above accumulate in RF

1.1 RF Transceiver Blocks

Fig. 2. PLL for generating RF clock

Fig. 3. PLL for clock synchronization

Fig. 4. A GSM transceiver circuit

1.2 New Mathematics for Simulating Transceivers

Traditionally, transceivers have been simulated with a standard circuit sim-

Martin Bodestedt and Caren Tischendorf

Universität zu Köln, Mathematisches Institut, Weyertal 86–90, 50931 Köln,

2 The Reﬁned Circuit Model

Electric Network Equations:

φ(jL ) − ATL e = 0, (1b)

Box 1. The reﬁned circuit model

The topology of the circuit is described by a network graph consisting of

current sources Is , controlled current sources Ic and semiconductor devices S.

3 Perturbation and Index Analysis

F (∂t wδ (t), wδ (t), t) = δ(t), F (∂t w(t), w(t), t) = 0.

F = 0 is said to have the perturbation index ν if it is the lowest integer such

holds for some constant c.

Example 1. Consider the PDAE

max ||u1δ (t) − u1 (t)||H01 ≤ c max (||δ1 (t)||H −1 + |∂t δ2 (t)|) ,

According to Deﬁnition 1 the perturbation index of (2) is two since a ﬁrst

Before we turn to the results, we brieﬂy summarize the assumptions needed

The domain Ω ∈ Rd for d ∈ {1, 2, 3} has a Lipschitz boundary. The Neu-

functions leading to the discrete deﬁned functions ψh , nh and ph . The Dirich-

4.1 The Software MECS

4.2 Flip Flop Circuitry

Fig. 1. Schematic diagram of a ﬂip ﬂop circuit

4.3 Voltage Controlled Oscillator (VCO)

As real benchmark we have tested a voltage controlled oscillator from the

= 2, . . . , N . Here, X denotes the approximation to X(t ), h = t − t−1 ,

where h := max=1,...,N h is the maximal step-size of the grid. For additive