The Concept of Stability in Numerical Mathematics

Springer Series in Computational Mathematics 45
Wolfgang Hackbusch
The Concept
of Stability
in Numerical
Mathematics
Springer Series in 45
Computational
Mathematics
Editorial Board
R. Bank
R.L. Graham
J. Stoer
R. Varga
H. Yserentant
For further volumes:

http://www.springer.com/series/797
Wolfgang Hackbusch
The Concept of Stability in

Numerical Mathematics
123
Wolfgang Hackbusch
MPI für Mathematik in den
Naturwissenschaften
Leipzig, Germany
ISSN 0179-3632
ISBN 978-3-642-39385-3 ISBN 978-3-642-39386-0 (eBook)
DOI 10.1007/978-3-642-39386-0
Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014931977
© Springer-Verlag Berlin Heidelberg 2014

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of
the work. Duplication of this publication or parts thereof is permitted only under the provisions of the
Copyright Law of the Publisher’s location, in its current version, and permission for use must always
be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright
Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Dedicated to Jana and Jörn
Preface
Usually, the subject of a book on numerical mathematics pertains to a certain field

of application or a certain numerical method. In this book we proceed in a different
way. Stability is a concept that appears in various fields of numerical mathematics
(as well as in other parts of mathematics). Although in each subfield stability is
defined differently, there is a common meaning for this term, roughly described
by the fact that perturbations are not amplifying the result in a dangerous way. In
examining different fields of numerical mathematics concerning stability, we have
the opportunity to recall some facts from numerical analysis. However, numerical
mathematics cannot control stability exclusively for its own purpose. It turns out
that stability is an ambiguous term, which also has strong connections to analysis
and functional analysis.
Although stability is an essential requirement for numerical methods, the particu-
lar stability conditions are often not as obvious as, e.g., consistency conditions. The
book may lead the reader to a better understanding of this term.
This book is an extension of a lecture held in the summer semester of 2003 at
the University of Kiel (Christian-Albrechts-Universität zu Kiel). The exposition is
self-contained, and the necessary facts from numerical analysis and analysis are
provided. Hence, the book is well suited, e.g., as material for seminars in numerical
mathematics.
The author wishes to express his gratitude to the publisher Springer for its
friendly cooperation. In particular, he thanks Ann Kostant, editorial consultant for
Springer, for polishing the English.
Leipzig, October 2013 Wolfgang Hackbusch
vii
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Stability of Finite Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 About Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 A Paradoxical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.1 First Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.2 Second Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.3 Explanation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Accuracy of Elementary Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Error Amplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.1 Cancellation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.2 Further Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1 Setting of the Problem and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.1 Quadrature Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.2 Interpolatory Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.3 Newton–Cotes Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.4 Gauss Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.1 Definitions and Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.2 Functionals, Dual Norm, and Dual Space . . . . . . . . . . . . . . . . 23
3.4 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4.1 Amplification of the Input Error . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4.2 Definition of Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4.3 Stability of Particular Quadrature Formulae . . . . . . . . . . . . . . 26
3.4.4 Romberg Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.5 Approximation Theorem of Weierstrass . . . . . . . . . . . . . . . . . 30
3.4.6 Convergence Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
ix
x Contents
3.4.7 Uniform Boundedness Theorem . . . . . . . . . . . . . . . . . . . . . . . . 36

3.4.8 Necessity of the Stability Condition, Equivalence Theorem . 38
3.4.9 Modified Definitions for Consistency and Convergence . . . . 39
3.5 Further Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5.1 General Intervals and Product Quadrature . . . . . . . . . . . . . . . . 41
3.5.2 Consistency Versus Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5.3 Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5.4 Arbitrary Slow Convergence Versus Quantitative
Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.1 Interpolation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Convergence and Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4 Equivalence Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.5 Instability of Polynomial Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.6 Is Stability Important for Practical Computations? . . . . . . . . . . . . . . . 53
4.7 Tensor Product Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.8 Stability of Piecewise Polynomial Interpolation . . . . . . . . . . . . . . . . . 56
4.8.1 Case of Local Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.8.2 Spline Interpolation as an Example for Global Support . . . . . 57
4.9 From Point-wise Convergence to Operator-Norm Convergence . . . . . 59
4.10 Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.1 Initial-Value Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.1.1 Setting of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.1.2 One-Step Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.1.3 Multistep Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Fixed-Point Theorem and Recursive Inequalities . . . . . . . . . . . . . . . . 66
5.3 Well-Conditioning of the Initial-Value Problem . . . . . . . . . . . . . . . . . 68
5.4 Analysis of One-Step Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4.1 Implicit Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4.2 Lipschitz Continuity of φ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4.3 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4.4 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4.5 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Analysis of Multistep Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5.1 Local Discretisation Error, Consistency . . . . . . . . . . . . . . . . . . 75
5.5.2 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.5.3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.5.4 Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.5.5 Stability and Convergence Theorems . . . . . . . . . . . . . . . . . . . . 83
Contents xi
5.5.6 Construction of Optimal Multistep Methods . . . . . . . . . . . . . . 85

5.5.7 Further Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.5.8 Other Stability Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6 Instationary Partial Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . 93

6.1 Introduction and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1.1 Notation, Problem Setting, Function Spaces . . . . . . . . . . . . . . 93
6.1.2 The Hyperbolic Case A = a∂/∂x . . . . . . . . . . . . . . . . . . . . . . 95
6.1.3 The Parabolic Case A = ∂ 2 /∂x2 . . . . . . . . . . . . . . . . . . . . . . . 96
6.2 Semigroup of Solution Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3 Discretisation of the Partial Differential Equation . . . . . . . . . . . . . . . . 100
6.3.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.3.2 Transfer Operators r, p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.3.3 Difference Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.4 Consistency, Convergence, and Stability . . . . . . . . . . . . . . . . . . . . . . . 105
6.4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.4.2 Convergence, Stability and Equivalence Theorems . . . . . . . . 106
6.4.3 Other Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.5 Sufficient and Necessary Conditions for Stability . . . . . . . . . . . . . . . . 109
6.5.1 First Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.5.2 Fourier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.5.3 Further Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.5.4 Implicit Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.5.5 Vector-Valued Grid Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.5.6 Generalisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.5.7 Dissipativity for Parabolic Discretisations . . . . . . . . . . . . . . . . 135
6.6 Consistency Versus Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7 Stability for Discretisations of Elliptic Problems . . . . . . . . . . . . . . . . . . . 139

7.1 Elliptic Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.2 Discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.3 General Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.3.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.3.2 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.3.3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
7.4 Application to Difference Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.4.1 Classical Choice of Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.4.2 Bijectivity of L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.5 Finite Element Discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.5.1 Variational Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.5.2 Galerkin Discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.5.3 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.5.4 Convergence and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
xii Contents
7.5.5 Quantitative Discretisation Error and Regularity . . . . . . . . . . 153

7.5.6 L2 Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.5.7 Stability of Saddle Point Problems . . . . . . . . . . . . . . . . . . . . . . 155
7.5.8 Further Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.5.9 Consistency Versus Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8 Stability for Discretisations of Integral Equations . . . . . . . . . . . . . . . . . . 167

8.1 Integral Equations and Their Discretisations . . . . . . . . . . . . . . . . . . . . 167
8.1.1 Integral Equation, Banach Space . . . . . . . . . . . . . . . . . . . . . . . 167
8.1.2 Discretisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.2 Stability Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.2.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.2.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.2.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.2.4 Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.3 Projection Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.4 Stability Theory for Nyström’s Method . . . . . . . . . . . . . . . . . . . . . . . . 176
8.5 Perturbation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
8.6 Application to Eigenvalue Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
List of Symbols
Symbols
k·k2 spectral norm of a matrix; cf. §2.4.2.2

k·k2 norm of functions in L2 (Ω)
k · k∞ maximum or supremum norm of vectors or functions; cf. §2.4.2.2
k · k∞ row-sum norm of matrices; cf. §2.4.2.2, §5.5.4.1
k·kX norm of the Banach space X
∗
k·kX dual norm; cf. §3.3.2
k·kY ←X operator norm; cf. (3.23)
h·, ·iX ∗ ×X , h·, ·iX×X ∗ dual pairing; cf. §3.3.2
Greek Letters
δa, δA, . . . perturbations of quantities a, A, . . .

∆t, ∆x step size in time and space; cf. §6.3.1
ε often, symbol for an error
εabs absolute error; cf. footnote 1 on page 6
εrel relative error; cf. footnote 1 on page 6
η(x, h), ηi discrete solution of ordinary differential equation at x or xi ; cf. §5.1.2
λ often, an eigenvalue
λ either ∆t/∆x or ∆t/∆x2 in §6; cf. (6.11)
λ parameter of the Fredholm integral operator of the second kind; cf.
(8.1)
Πn projection in §8.1.2.2
ρ(·) spectral radius; cf. (6.25)
σ(M ) spectrum of matrix M ; cf. §5.5.4.1
τ (x, y(x); h) consistency error of a one-step method; cf. (5.18)
Φi,n (·) Lagrange function; cf. (4.2)
xiii
xiv List of Symbols
φ(xi , ηi , [ηi+1 , ]h; f ) function defining an explicit [implicit] one-step method; cf.
§5.1.2, §5.4.1
φ(xj , ηj+r−1 , . . . , ηj , h; f ) function defining a multistep method; cf. (5.20a)
ψ(ζ) characteristic polynomial of a multistep method; cf. (5.21a)
Ωn grid for difference method; cf. §7.2
Latin Letters
ai,n quadrature weights in §3

B Banach space in §6
cond(·) condition of a matrix; cf. (8.10)
C set of complex numbers
C(λ, ∆t) difference operator in §6.3.3
C (D) space of continuous functions defined on D
C0 (D) space of continuous functions with zero boundary value on ∂D
C k (D) k ∈ N0 , space of functions with continuous derivatives of order ≤ k
C0∞ (R) infinitely differentiable functions with compact support
C λ (D) λ > 0, λ ∈ / N, Hölder space; cf. §4.9
Cn stability constant; cf. (3.13b), §4.3, §4.5
Cstab stability constant; cf. (3.14), (4.5), §8.2.2, (7.11)
D integration domain for the integral operator; cf. (8.2)
degree(·) degree of a polynomial
Ej shift operator; cf. (6.22)
eps machine precision; cf. §2.3
F Fourier map; cf. §6.5.2
fn right-hand side in finite difference equation (7.4)
fl(·) operation in the argument performed in floating point arithmetic; cf.
(2.13)
G(ξ) characteristic function; cf. (6.34)
G(x, ξ) characteristic function frozen at x; cf. (6.48)
h step size for ordinary or partial differential equations; cf. §5.1.2
H k (Ω), H t (Ω), H0k (Ω), H0t (Ω) Sobolev spaces; cf. pages 148, 149
In interpolation; cf. (4.4)
Jn Bessel function; cf. (2.2)
k(x, y) kernel of the integral operator; cf. (8.2)
kn (x, y) approximation of k(x, y); cf. (8.5)
K integral operator; cf. (8.2)
Kn discrete integral operator; cf. (8.4a)
`2 , `∞ Banach spaces of sequences in Z; cf. §6.3.1
L elliptic differential operator in §7
L(X, Y ) set of linear and continuous mappings from X to Y
Li,n (·) Lagrange polynomial; cf. (4.3)
Ln finite difference matrix corresponding to L; cf. (7.4)
List of Symbols xv
Ln (·) Legendre polynomial; cf. §3.1.4

M set of machine numbers; cf. §2.3
N natural numbers {1, 2, . . .}
N0 N ∪ {0} = {0, 1, 2, . . .}
N0 , N00 suitable infinite subsets of N; cf. §7.2
p prolongation in §6.3.2
P often, polynomial
PYn prolongation in §7.3.1
Qn (·) quadrature rule; cf. §3
r restriction in §6.3.2
r(·) numerical radius; cf. §6.5.5
rd(·) rounding to next machine number; cf. §2.3
R set of real numbers
n
RX , RYn restrictions in §7.3.1
T (·, h) trapezoidal rule of step size h; cf. §3.4.4
T (t) solution operator; cf. §6.2
un solution of finite difference equation (7.4)
Uνµ discretisation of solution u(t, x) in §6 at t = µ∆t, x = ν∆x
xi,n quadrature points in §3
X, Y often, Banach spaces
X, Xn , Y, Yn Banach spaces in §7.3.1
X∗ dual space L(X, R); cf. §3.3.2
X̂, X̂n Banach spaces in §7.3.2
Z entire numbers
Chapter 1
Introduction
In numerical mathematics we have to distinguish between two types of methods.

There are finite algorithms, which solve a given task with a finite number of arith-
metical operations. An example of such finite algorithms is the Gauss elimination
for solving a system of linear equations. On the other hand, there are problems P
that cannot be solved in finite time. Instead, there are (finite) approximation algo-
rithms that involve solving substituting problems Pn producing results xn , which,
hopefully, tend to the true solution x of the original problem. The increasing close-
ness of Pn to P for n → ∞ is the subject of the consistency condition. What really
matters is the convergence xn → x. Whether consistency implies convergence de-
pends on stability. It will turn out that under the assumption of consistency and
possibly some technical conditions, convergence and stability are equivalent.
The original German manuscript has been used for Diplom students in mathemat-
ics. The material in this book is intended for master and Ph.D. students. Besides the
discussion of the role of stability, a second goal is to review basic parts of numerical
analysis.
We start in Chapter 2 with finite algorithms. We recall the condition of a prob-
lem and the stability of an algorithm. The amplification of input and intermediate
floating point errors measures the quality of condition and stability. In this respect,
the terms remain vague, since the amplification factors are some positive real
numbers which may vary between ‘small’ (stable) and ‘large’ (unstable) without
a clear separation.
Chapter 3 is devoted to quadrature methods, more precisely, to families of
quadratures Qn , where, with increasing n, the quality should improve (‘consis-
tency’). ‘Stability’ is again connected with the amplification of input errors. In
contrast to Chapter 2, it is uniquely defined as to whether stability holds or not, since
the terms ‘small’ and ‘large’ are replaced by finiteness or infiniteness of sup Cn , the
supremum of condition numbers Cn .
Although the stability definition is inspired by numerical phenomena, it is also
suited to purely analytical purposes. Stability is almost equivalent
R to convergence
of the quadrature result Qn (f ) to the exact integral f dx. Correspondingly,
W. Hackbusch, The Concept of Stability in Numerical Mathematics, 1

Springer Series in Computational Mathematics 45, DOI 10.1007/978-3-642-39386-0_1,
2 1 Introduction
analytical tools from functional analysis are involved, namely Weierstrass’ approx-
imation theorem and the uniform boundedness theorem.
Interpolation treated in Chapter 4 follows the same pattern as in Chapter 3. In
both chapters one can pose the question of how important the stability statement
sup Cn < ∞ is, if one wants to perform only one quadrature or interpolation for a
fixed n. In fact, polynomial interpolation is unstable, but when applied to functions
of certain classes it behaves quite well.
This is different in Chapter 5, where one-step and multistep methods for the
solution of ordinary initial-value problems are treated. Computing approxima-
tions requires an increasing number of steps, when the step size approaches zero.
Often an instability leads to exponential growth of an error, eventually causing a
termination due to overflow.
For ordinary differential equations instability occurs only for proper multistep
methods, whereas one-step methods are always stable. This is different for partial
differential equations, which are investigated in Chapter 6. Here, difference methods
for hyperbolic and parabolic differential equations are treated. Stability describes the
uniform boundedness of powers of the difference operators.
Also in the case of elliptic differential equations discussed in Chapter 7, stability
is needed to prove convergence. In this context, stability describes the boundedness
of the inverse of the difference operator or the finite element matrix independently
of the step size.
The final chapter is devoted to Fredholm integral equations. Modern projection
methods lead to a very easy proof of stability, consistency, and convergence. How-
ever, the Nyström method—the first discretisation method based on quadrature—
requires a more involved analysis. We conclude the chapter with the analysis of the
corresponding eigenvalue problem.
Despite the general concept of stability, there are different aspects to consider
in the subfields. One aspect is the practical importance of stability (cf. §4.6),
another concerns a possible conflict between a higher order of consistency and
stability (cf. §3.5.2, Remark 4.15, Theorem 5.47, §6.6, §7.5.9).
Chapter 2
Stability of Finite Algorithms
2.1 About Algorithms
An algorithm is used to solve a (numerical) problem. For the mathematical formula-

tion of a problem (or task) we use a mapping Φ : X → Y , which is to be evaluated
numerically (cf. [9, Chap. 1]).
An algorithm is executable if the mapping Φ is composed of units that are
realisable in a computer program. We call these units elementary operations. In
the standard case, these elementary operations are the basic arithmetical opera-
tions +, −, ∗, / in the set of real or whole numbers. In addition, the√programming
languages offer the use of some special functions like sin, cos, exp, ·, . . .
An algorithm is the composition of elementary operations. A finite algorithm
is characterised as involving only a finite number of elementary operations. The
algorithm is a realisation of the mapping
Φ : (x1 , . . . , xn ) ∈ X = domain(Φ) 7→ (y1 , . . . , ym ) ∈ Y = range(Φ). (2.1)
If Φ is realisable by at least one algorithm, then there are even infinitely many
algorithms of this kind. Therefore, there is no one-to-one correspondence between
a task and an algorithm.
A finite algorithm can be described by a sequence of vectors
(j)
x(0) = (x1 , . . . , xn ) , x(1) , . . . , x(j) = (x1 , . . . , x(j)
nj ), . . . , x
(p)
= (y1 , . . . , ym ) ,
(j)
where the values xi from level j can be computed by elementary operations from
the components of x(j−1) .
Example 2.1. The scalar product y = xx12 , xx34 has the input vector (x1 , . . . , x4 ).

The algorithm uses the intermediate values x(1) = (z1 , z2 ) with z1 := x1 x2 ,

z2 := x3 x4 . Then the output value is obtained from y := z1 + z2 .
The opposite of a finite algorithm is an infinite one, which, e.g., computes a
sequence whose limit is the desired result of the problem. Since one has to terminate

4 2 Stability of Finite Algorithms
such an infinite process, one finally obtains a finite algorithm producing an approx-
imate result.
Here we only want to motivate the concept that algorithms should be constructed
carefully regarding stability. It remains to analyse various concrete numerical
methods (see, e.g., the monograph of Higham [5]).
2.2 A Paradoxical Example
2.2.1 First Algorithm
The following problem involves the family of Bessel functions. In such a case, one
is well advised to look into a handbook of special functions. One learns that the n-th
Bessel function (also called cylinder function) can be represented by a power series
or as an integral:
∞ k
X (−1) (x/2)n+2k
Jn (x) = for n ∈ N0 (2.2)
k! (n + k)!
k=0
n Z π
(−1)
= eix cos(ϕ) cos(nϕ)dϕ for n ∈ Z. (2.3)
π 0
The chosen task is the computation of J5 (0.6).

Assume that tabulated values of the first two Bessel functions J0 and J1 are
offered as, e.g., in [13, page 99],
J0 (0.6) = 0.9120, J1 (0.6) = 0.2867, (2.4)
but not the value of J5 (0.6). Furthermore, assume that the book contains the
recurrence relation
2n
Jn+1 (x) + Jn−1 (x) = Jn (x) for n ∈ Z (2.5)
x
as well as the property
∞
X
Jn (x) = 1 for all x ∈ R. (2.6)
n=−∞
Exercise 2.2. Prove convergence of the series (2.2) for all x ∈ C (i.e., Jn is an
entire function).
An obvious algorithm solving our problem uses the recursion (2.5) for n =
1, 2, 3, 4 together with the initial values (2.4):
2.2 A Paradoxical Example 5
2 2
J2 (0.6) = -J0 (0.6)+ 0.6 J1 (0.6) = -0.9120+ 0.6 0.2867 = 4.3666710 -2,
4 4
J3 (0.6) = -J1 (0.6)+ 0.6 J2 (0.6) = -0.2867+ 0.6 4.3666710 -2 = 4.4111110 -3,
6 6
(2.7)
J4 (0.6) = -J2 (0.6)+ 0.6 J3 (0.6) = -4.3666710 -2+ 0.6 4.4111110 -3 = 4.4444410 -4,
8 8
J5 (0.6) = -J3 (0.6)+ 0.6 J4 (0.6) = -4.4111110 -3+ 0.6 4.4444410 -4 = 1.5148110 -3.
The result is obtained using only eight elementary operations. The underlying equa-
tions are exact. Nevertheless, the computed result for J5 (0.6) is completely wrong,
even the order of magnitude is incorrect! The exact result is J5 (0.6) = 1.9948210 -5.
Why does the computation fail? Are the tabulated values (2.4) misprinted? No,
they are as correct as they can be. Is the (inexact) computer arithmetic, used in (2.5),
responsible for the deviation? No, even exact arithmetic yields the same results. For
those who are not acquainted with numerical effects, this might look like a paradox:
exact computations using exact formulae yield completely wrong results.
2.2.2 Second Algorithm
Before we give an explanation, we show a second ‘paradox’: an algorithm based on

inexact and even rather dubious formulae yields a perfect result. A numerical analyst
asking for advice would recommend that we use the recurrence relation (2.5) in the
opposite order; i.e., starting from Jm (0.6) and Jm+1 (0.6) for some m > 5 and
applying (2.5) in the order n = m − 1, m − 2, . . . . The drawback is that neither
Jm (0.6) nor Jm+1 (0.6) are available. The expert’s hint is to replace Jm+1 (0.6) by
zero (vague reasoning: that value does not matter). The unknown value Jm (0.6) will
be obtained from the additional property (2.6).
We denote the candidates for Jn (0.6) by jn . The plan is to start from jm+1 := 0
2n
and jm := Jm (0.6) and to apply (2.5): jn−1 = 0.6 jn − jn+1 . We observe that all
results jm−1 , jm−2 , . . . depend linearly on jm . Therefore, starting from
0 0
jm+1 := 0, jm := 1 (2.8)
and calculating
0 2n 0 0
jn−1 = j − jn+1 for n = m, m − 1, . . . , 0, (2.9)
0.6 n
we get quantities jn0 with the property jn = jm jn0 . Now the unknown value jm
can be determined from (2.6). From (2.2) or even from (2.5) we can derive that
n
J−n (x) = (−1) Jn (x). Hence, (2.6) is equivalent to
J0 (x) + 2J2 (x) + 2J4 (x) + . . . = 1. (2.10)

Pbm/2c
Replacing
the infinite sum by a finite one: j0 + 2 ν=1 j2ν = 1, we arrive at
Pbm/2c 0
jm · j00 + 2 ν=1 j2ν = 1 and, in particular,
 
bm/2c
X
J5 (0.6) ≈ jm j50 = j50 / j00 + 2 0 
j2ν .
ν=1
The choice m = 10 yields
J5 (0.6) = 1.994819537430010 -5,
where all 14 digits are correct.

Exercise 2.3. Prove (2.10). Hint: comparison of the coefficients shows that (2.10)
is equivalent to
` m
X (−1) 2, if m > 0
× =0 for ` > 0.
(` − m)! (` + m)! 1, if m = 0
m=0
For the latter identity use ex e−x = 1.
2.2.3 Explanation
As stated above, the tabulated values (2.4) are correct, since all four digits offered
in the table are correct. It turns out that we have a fortunate situation since even the
fifth digit is zero: the precise values with six digits are J0 (0.6) = 0.912005 and
J1 (0.6) = 0.286701. However, the sixth and seventh decimals cause the absolute
errors1
(0) (1)
εabs = 0.9120049 − 0.9120 = 4.910 -6, εabs = 0.28670099 − 0.2867 = 9.910 -7
(2.11)
and the relative errors
(0) 4.910 -6 (1) 9.910 -7
εrel = = 5.310 -6, εrel = = 3.410 -6 .
0.912005 0.286701
Both are relatively small. This leads to the delusive hope that the same accuracy
holds for J5 (0.6). Instead, we observe the absolute and relative errors
(5) (5)
εabs = 1.510 -3, εrel = 75. (2.12)
As we see, the absolute error has increased (from about 10−6 in (2.11) to 10−3 in
(2.12)). Additionally, the small value J5 (0.6) J1 (0.6) causes the large relative
error in (2.12).
In order to understand the behaviour of the absolute error, we consider the recur-
2n
sion jn+1 = 0.6 jn − jn−1 (cf. (2.5)) for general starting values j0 , j1 . Obviously,
1 x−x̃
Let x̃ be any approximation of x. Then εabs = x − x̃ is the absolute error, while εrel = x
is the relative error.
2.3 Accuracy of Elementary Operations 7
all jn depend linearly on j0 , j1 . In particular, we obtain

74 627 7820
j5 = j1 − j0 .
27 9
One verifies that the values j0 = 0.9120 and j1 = 0.2867 from (2.4) yield the value
j5 = 1.5148110 -3 from (2.7). The factor 7427 627
≈ 2764 is responsible for the strong
amplification of the absolute error of j1 .
We add that Bessel functions are particular examples of linear recurrence
relations that have been studied in the article of Oliver [6] (see also [3], [10]).
2.3 Accuracy of Elementary Operations
Floating point arithmetic used in computers is characterised by a fixed mantissa

length. For t digits corresponding to the basis b (t ≥ 1, b ≥ 2), the set M of
machine numbers consists of zero and all numbers of the form
t
X
E
x = ±0.d1 . . . dt ∗ b = ±b E
dk b−k , where
k=1
1 ≤ d1 < b and 0 ≤ dk < b for k > 1, E ∈ Z.
For simplicity, we ignore the boundedness of the exponent E which may lead to
overflow or underflow.
Exercise 2.4. (a) What is the best bound ε in supξ∈R minx∈M |x − ξ|/|x| ≤ ε ?
(b) What is ε in the special case of the dual basis b = 2?
The minimiser x of minx∈M |x − ξ| is the best ‘rounding’ of ξ ∈ R. Denoting

this mapping by rd, we get the following estimate with machine precision eps = ε
(ε from Exercise 2.4):
|x − rd(x)| ≤ eps|x|.
In the following we assume that eps is chosen such that all elementary operations
‘op’ fulfil the following inequality:
|fl (a op b) − (a op b)| ≤ eps |a op b| for all a, b ∈ M. (2.13)
Here, fl(. . .) indicates the evaluation of the operations indicated by ‘. . .’ in the sense
of floating-point computer arithmetic. A√similar estimate is assumed for elementary
operations with one argument (e.g., fl ( a)). Furthermore, we assume eps 1.
Note that (2.13) controls the relative error.
Remark 2.5. Inequality (2.13) can be regarded as a ‘consistency’ property.
More details about computer arithmetic can be found, e.g., in [7, §16] or [8, §2.5].
2.4 Error Amplification
The mathematical term ‘amplification’ will be used in a quite general way. If an

error ε is changed into cε, we say that ε is amplified by c, even if c ∈ (0, 1).
The standard case of interest is, of course, whether c > 1 or even c 1 may
occur.
2.4.1 Cancellation
The reason for the disastrous result for J5 (0.6) in the first algorithm can be discussed
by a single operation:
y = x1 − x2 . (2.14)
On the one hand, this operation is harmless, since the consistency (2.13) of the
subtraction guarantees that the floating point result ỹ = fl(x1 − x2 ) satisfies the
estimate2
|ỹ − y| ≤ eps |y|.
On the other hand, this estimate holds only for x1 , x2 ∈ M. The more realistic
problem occurs when replacing the exact difference η = ξ1 − ξ2 with the difference
x1 − x2 of machine numbers:
η = ξ1 − ξ2 , x1 = rd(ξ1 ), x2 = rd(ξ2 ), ỹ = fl(x1 − x2 ),
where ξ1 and ξ2 are general real numbers. The interesting error is
|η − ỹ| for ỹ = fl(x1 − x2 ).
The absolute error is bounded by
|η − ỹ| ≤ |x1 − x2 − fl(x1 − x2 )| + |∆x1 | + |∆x2 |

≤ eps (|x1 − x2 | + |∆x1 | + |∆x2 |)
with ∆xi := ξi − xi .
As long as |η| ∼ |xi | (i = 1, 2), the situation is under control, since also the relative
error has size eps. Dramatic cancellation appears when |η| |x1 | + |x2 |. Then the
relative error is . eps |x1 |+|x
|η|
2|
. In the worst case, |η| is of the size eps (|x1 | + |x2 |),
so that the relative error equals O(1).
Cancellation takes place for values ξ1 and ξ2 having the same sign and similar
size. On the other hand, the sum of two non-negative numbers is always safe, since
|η| = |ξ1 + ξ2 | ≥ |xi | and |η−ỹ| |η| ≤ 3eps. The factor 3 corresponds to the three
truncations ξ1 7→ x1 , ξ2 7→ x2 , x1 + x2 7→ fl(x1 + x2 ).
2
If x1 and x2 are close, often the exact difference is a machine number so that ỹ = y.
2.4 Error Amplification 9
2.4.1.1 Linear (Differential) Error Analysis
Given the mapping y = Φ(x) in (2.1), we have to check how errors ∆xi of the input
value x̃i = xi + ∆xi affect the output ỹ = Φ(x̃).
Assume that Φ is continuously differentiable. The Taylor expansion of ỹj =
∂Φ
Φj (x1 , . . . , xi + ∆xi , . . . , xn ) yields ỹj = Φj (x1 , . . . , xn ) + ∂xij ∆xi + o(∆xi ).
Hence,
∂Φj (x1 , . . . , xn )
(2.15)
∂xi
is the amplification factor of the input error ∆xi , provided that we consider the
absolute errors. The amplification of the relative error equals

∂Φj (x1 , . . . , xn ) |xi |
·
|yj | . (2.16)

∂xi
Which error is of interest (absolute or relative) depends on the particular problem.

Instead of individual components xi and yj , one may also consider certain norms
kxk and kyk.
2.4.1.2 Condition and Stability
A problem is called well-conditioned if the amplification of the input error is not

much larger than 1. On the other hand, if the error is strongly increased, the problem
is called ill conditioned. The maximal error amplification factor is called condition
or condition number.
Given a problem Φ, let A be any algorithm realising Φ. Let the algorithm use the
following intermediate results:
x(0) = (x1 , . . . , xn ) 7→ x(1) 7→ . . . 7→ x(p) = {y1 , . . . , ym }.

ϕ1 ϕ2 ϕp
We regain the mapping Φ of the problem as the composition Φ = ϕp ◦ϕp−1 ◦. . .◦ϕ1 .

The mappings
Φ(q) := ϕp ◦ ϕp−1 ◦ . . . ◦ ϕq+1 for 0 ≤ q ≤ p
describe the transfer from x(q) to x(p) = {y1 , . . . , ym } (for q = p, the empty
product is Φ(p) = id). While Φ(0) = Φ depends only on the underlying problem, the
(q) (q)
mappings Φ(q) for 1 ≤ q < p depend on the algorithm. The derivative ∂Φj /∂xi
(q)
corresponds to (2.15). It is the amplification factor describing ∆yj /∆xi , where
(q) (q)
∆xi is the error of xi and ∆yj is the induced error of yj . Since, by definition,
(q)
the intermediate result xi is obtained by an elementary operation, its relative error
is controlled by (2.13).
Let κ be the condition number of problem Φ. If all amplification factors
(q) (q)
|∂Φj /∂xi | are at most of size κ, the algorithm is called stable. Otherwise, the
algorithm is called unstable.
Note that the terms stability/instability are not related to the problem, but to the
algorithm. Since there are many algorithms for a given problem, one algorithm for
problem Φ may be unstable, while another one could be stable.
Furthermore, we emphasise the relative relation: if the condition κ is large, also
(q) (q)
the amplification factors |∂Φj /∂xi | of a stable algorithm may be large. In the
case of the example in §2.2, the algorithms are stable. The difficulty is the large
condition number.
We summarise: If a problem Φ possesses a large condition number, the disastrous
amplification of the input errors cannot be avoided3 by any algorithm realising Φ. If
the problem is well-conditioned, one has to take care to choose a stable algorithm.
On purpose, the definitions of condition and stability are vague4 . It is not fixed
whether we consider relative or absolute errors, single components or a norm of the
error. Furthermore, it remains open as to what amplification factors are considered
to be moderate, large, or very large.
2.4.2 Further Examples
2.4.2.1 Evaluation of the Exponential Function
We consider the simple (scalar) problem Φ(x) = exp(x) for x = ±20 under the
hypothetical assumption that exp does not belong to the elementary operations.
First, we consider the condition. Here, it is natural to ask for the relative errors.
According to (2.16), the condition number equals
d exp(x) |x|
· = |x|.
dx | exp(x)|
Hence, moderate-sized values of x (like x = ±20) lead to a well-conditioned

problem.
Any finite algorithm can only compute an approximate value of exp(x). We
prescribe an accuracy ε and ask for an approximation eg
xp(x) satisfying
xp(x)| ≤ ε exp(x) with ε := 10−6 .

|exp(x) − eg (2.17)
We fix x = 20 and choose a truncated Taylor expansion

n
X xν
En (x) = .
ν=0
ν!
3
The only exception occurs when the input values are exact machine numbers, i.e., ∆xi = 0.
4
A systematic definition is attempted by de Jong [2].
The obvious algorithm for the evaluation of En is Horner’s scheme consisting of a

sequence of elementary operators. We use the fact that n = 46 leads to a remainder
satisfying (2.17). Indeed, the computed value En (20) = 4.851 65110 8 shows six
correct leading digits, since the exact result is exp(20) = 4.851 652
P10∞8. ν
Next, we consider x = −20. Obviously, the remainder ν=n−1 xν! for
x = −20 is smaller than for +20. However, because of the relative error, the fac-
tor exp(x) on the right-hand side of (2.17) changes from exp(20) to exp(−20);
hence, a more accurate approximation is required. One can check that n = 79 is
sufficient. However, the computation of En (−20) yields the value 3.910 -9, which
deviates strongly from the exact value 2.06110 -9. Obviously, this algorithm is not
stable.
The reason for the latter failure is the cancellation of the positive terms in
P39 202ν P39 −202ν+1
ν=0 (2ν)! = 2.425810 8 with the negative ones in ν=0 (2ν+1)! = −2.424810 8.
A stable algorithm for approximating exp(−20) can be derived from the formula
exp(−20) = 1/ exp(20) ≈ 1/En (20). The choice n = 45 suffices to compute an
approximation satisfying (2.17).
This example elucidates that analytical estimates of approximations as in (2.17)
do not ensure at all that the evaluation of the approximating expression leads to
acceptable results.
2.4.2.2 Zeros of Polynomials
The following example shows that standard concepts, e.g., from linear algebra, may
lead to a completely unstable algorithm. We consider the eigenvalue problem for
symmetric matrices (also here an approximation is unavoidable, since, in general,
eigenvalues are not computable by a finite algorithm). Computing the eigenvalue of
symmetric matrices is well-conditioned, as the following theorem will show. Since
the notation of the spectral norm is needed in the theorem, we recall some facts
about matrix norms. Given some vectors norms k·k, the associated matrix norm is
defined by
kAk := max{kAxk / kxk : x 6= 0}
(cf. (3.23) for the operator case). If k·k = k·k∞ is the maximum norm, the associ-
ated matrix norm is denoted by the identical symbol k·k∞ and called the row-sum
norm (cf. §5.5.4.1). The choice k·k = k·k2 of the Euclidean vector norm leads us to
√ k·k2 of matrices. An explicit description of the spectral norm is
the spectral norm
kAk2 = max{ λ : λ eigenvalue of AT A}.
Theorem 2.6. Let A, ∆A ∈ Rn×n be matrices with A being symmetric (or normal),
and set Ã := A + ∆A. Then, for any eigenvalue λ̃ of Ã, there is an eigenvalue λ of
A such that |λ̃ − λ| ≤ k∆Ak2 .
Proof. A diagonalisation by a unitary transformation does not change the eigen-

values and norms. Therefore, without loss of generality, A = diag{λ1 , . . .} is
assumed to be diagonal. Let x̃ 6= 0 be an eigenvector of Ã corresponding to λ̃.
The identity
(λ̃I − A)−1 ∆A x̃ = (λ̃I − A)−1 (Ã − A) x̃ = (λ̃I − A)−1 (λ̃I − A) x̃ = x̃
proves k(λ̃I − A)−1 ∆Ak2 ≥ 1 for the spectral norm. We continue the inequality
by 1 ≤ k(λ̃I−A)−1 ∆Ak2 ≤ k(λ̃I−A)−1 k2 k∆Ak2 . Since A = diag{λ1 , . . .}, the
norm equals k(λ̃I − A)−1 k2 = 1/ min |λ̃ − λi |. Altogether, there is an eigenvalue
1≤i≤n
λ of A such that |λ̃ − λ| = min |λ̃ − λi | = 1/k(λ̃I − A)−1 k2 ≤ k∆Ak2 . t
u
i
Since in linear algebra eigenvalues are introduced via the characteristic poly-
nomial, one might get the following idea:
Pn
(a) Determine the characteristic polynomial P (x) = k=0 ak xk .
(b) Compute the eigenvalues as roots of P : P (x) = 0.
For simplicity we assume that the coefficients ak of P (x) = det(xI − A) can be
determined exactly. Then the second part remains to be investigated: How are the
zeros of P effected by perturbations of ak ? Here, the following famous example of
Wilkinson is very informative (cf. [11] or [12, p. 54ff]).
We prescribe the eigenvalues (roots) 1, 2, . . . , 20. They are the zeros of
20
Y
P (x) = (i − x) = a0 + . . . + a19 x19 + a20 x20
k=1
(a0 = 20! = 2 432 902 008 176 640 000, . . . , a19 = 190, a20 = 1). The large value
of a0 shows the danger of an overflow during computations with polynomials. The
determination of zeros of P seems to be rather easy, because P has only simple
zeros and these are clearly separated.
We perturb only the coefficient a19 into ã19 = a19 − 2−23 (2−23 = 1.19210 -7
corresponds to ‘single precision’). The zeros of the perturbed polynomial P̃ are
1, 2, 3, 4, 4.999 999 928, 6.000 006 9, 6.999 69, 8.0072, 8.917,

10.09 ± 0.64i, 11.79 ± 1.65i, 13.99 ± 2.5i, 16.73 ± 2.8i, 19.5 ± 1.9i, 20.84.
The (absolute and relative) errors are not only of size O(1), the appearance of five
complex pairs shows that even the structure of the real spectrum is destroyed.
For sufficiently small perturbations, conjugate complex pairs of zeros cannot
occur. Considering the perturbation ã19 = a19 − 2−55 (2−55 = 2.77610 -17), we
obtain the zeros
1, 2, . . . , 9, 10, 11 − 10−10 , 12 + 610 -9,

13 − 1710 -9, 14 + 3710 -9, 15 − 5910 -9, 16 + 4710 -9, ...
For instance, the perturbation of the 15-th zero shows an error amplification by
5910 -9/2−55 = 2.110 9 .
Stable algorithms for computing the eigenvalue directly are using the entries of
the matrix and avoid the detour via the characteristic polynomial (at least in the form
Pn
of P (x) = k=0 ak xk ; cf. Quarteroni et al. [8, §5]). In general, one should avoid
polynomials in the classical representation as sums of monomials.
2.4.2.3 Numerical Differentiation and Related Phenomena
The numerical computation of (partial) derivatives is often required if the derivative

is not implemented or if its evaluation is too costly. Assume f ∈ C 3 (I). The Taylor
expansion
f (x + h) − f (x − h) h2
Dh f (x) := = f 0 (x) + f 000 (ξ)
2h 6
with some x − h < ξ < x + h yields the estimate C3 h2 of the error, where C3
is a bound of f 000 . However, for h → 0 the truncation error spoils the computation
because of the cancellation error. A very optimistic assumption is that the numerical
evaluation of the function yields values f˜(x ± h) with the relative accuracy

˜
f (x ± h) − f (x ± h) ≤ |f (x ± h)| eps ≤ C0 eps,

where C0 is a bound of f. Therefore, the difference quotient Dh f is perturbed by

an error bounded by C0 eps/(2h), and the total error is
εh := C3 h2 + C0 eps/(2h).
Minimisation of this bound yields

r
C0 3 2 23 13 2
hopt = 3 eps and εhopt = 2 3 C0 C3 eps 3 .
4C3 4
If f is of reduced smoothness, say f ∈ C α (I), 1 < α < 3, the factor eps2/3

1
becomes eps1− α , which is very poor for α > 1 close to 1.
A similar phenomenon can appear in quite another situation. The following
example is due to Bini–Lotti–Romani [1]. Let
f : R2 × R2 → R2
be the bilinear function
f1 (x, y) = x1 y1 , (2.18)
f2 (x, y) = x2 y1 + x1 y2 .
We want to evaluate f (x, y) with a minimal number of multiplications.5 The

standard implementation requires three multiplications x1 · y1 , x2 · y1 , and x1 · y2 .
5
If we replace the scalars x1 , x2 , y1 , y2 by matrices, the matrix-matrix multiplication is the
dominant operation concerning cost.
Instead one can compute fε : R2 × R2 → R2 defined by
fε,1 (x, y) = x1 y1 + ε2 x2 y2 ,
fε,2 (x, y) = x2 y1 + x1 y2
for some ε > 0. Obviously, for ε small enough, we obtain results for any prescribed
accuracy. The second function can be evaluated by the algorithm

s := (x1 + εx2 ) · (y1 + εy2 ) (s1 + s2 )/2
(x, y) 7→ 1 7→ fε := ,
s2 := (x1 − εx2 ) · (y1 − εy2 ) (s1 − s2 )/(2ε)
which requires only two multiplications. The expression (s1 − s2 )/(2ε) shows the
analogy to the numerical differentiation, so that again cancellation occurs.
In fact, the numbers of multiplications (3 and 2, respectively) have their origin in
tensor properties. We may rewrite (2.18) as
X
fi = vijk xj yk for i = 1, 2.
1≤j,k≤2
The coefficients are
v111 = v221 = v212 = 1 and vijk = 0 otherwise.
v = (vijk ) is an element of the tensor space R2 ⊗ R2 ⊗ R2 . Its tensor rank is

rank(v) = 3; i.e., r = 3 is the minimal number of terms such that6
r
(µ) (µ) (µ)
X
vijk = ai bj ck
µ=1
for some vectors a(µ) , b(µ) , c(µ) ∈ R2 (cf. [4, Definition 3.32]). The solution for
r = 3 is
a(1) = b(1) = b(3) = c(1) = c(2) = 10 ,

a(2) = a(3) = b(2) = c(3) = 01 .

The corresponding coefficients vε = (vε,ijk ) of fε have tensor rank 2:

h i
1/2
a(1) = 1/(2ε) b(1) = c(1) = 1ε ,

,
h i h i
1/2 1
a(2) = −1/(2ε) , b(2) = c(2) = −ε .
It can be shown that, whenever rank(vε ) < rank(v) and vε → v as ε → 0, then

the computation of vε is unstable (cf. Hackbusch [4, Proposition 9.16]).
6
Pr (µ) (µ)
Note that the minimal r in vij = µ=1 ai bj is the usual matrix rank of (vij ).
References 15
References
1. Bini, D., Lotti, G., Romani, F.: Approximate solutions for the bilinear form computational
problem. SIAM J. Comput. 9, 692–697 (1980)
2. de Jong, L.S.: Towards a formal definition of numerical stability. Numer. Math. 28, 211–219
(1977)
3. Deuflhard, P.: A summation technique for minimal solutions of linear homogeneous difference
equations. Computing 18, 1–13 (1977)
4. Hackbusch, W.: Tensor spaces and numerical tensor calculus, Springer Series in Computa-
tional Mathematics, Vol. 42. Springer, Berlin (2012)
5. Higham, N.J.: Accuracy and stability of numerical algorithms. SIAM, Philadelphia (1996)
6. Oliver, J.: The numerical solution of linear recurrence relations. Numer. Math. 11, 349–360
(1968)
7. Plato, R.: Concise Numerical Mathematics. AMS, Providence (2003)
8. Quarteroni, A., Sacco, R., Saleri, F.: Numerical Mathematics, 2nd ed. Springer, Berlin (2007)
9. Stoer, J., Bulirsch, R.: Introduction to Numerical Analysis. North-Holland, Amsterdam (1980)
10. Van der Cruyssen, P.: A reformulation of Olver’s algorithm for the numerical solution of
second-order linear difference equations. Numer. Math. 32, 159–166 (1979)
11. Wilkinson, J.H.: Rounding errors in algebraic processes. Prentice-Hall, Englewood Cliffs
(1964). Reprinted by Dover Publications, New York, 1994
12. Wilkinson, J.H.: Rundungsfehler. Springer, Berlin (1969)
13. Zeidler, E. (ed.): Oxford Users’ Guide to Mathematics. Oxford University Press, Oxford
(2004)
Chapter 3
Quadrature
In this chapter some facts from interpolation are used. Therefore, the reader may
first have a look at §4.1 of the next chapter, which is devoted to interpolation. We
prefer to start with quadrature instead of interpolation, since a projection between
function spaces (interpolation) is more involved than a functional (quadrature).
3.1 Setting of the Problem and Examples
3.1.1 Quadrature Formulae
Any systematic approximation of an integral is called quadrature. We start with a

simple setting of the problem.
Problem 3.1. Assume that f ∈ C ([0, 1]). Compute an approximate value of the
integral
Z 1
f (x)dx.
0
Other intervals and additional weight functions can be treated, but these general-
isations are uninteresting for our purpose.
For all n ∈ N0 we define a quadrature formula
n
X
Qn (f ) = ai,n f (xi,n ). (3.1)
i=0
Here ai,n are the quadrature weights and xi,n are the (disjoint) quadrature points.
A sequence
{Qn : n ∈ N0 }
yields a family of quadrature formulae (sometimes also called a quadrature rule
meaning that this rule generates quadrature formulae for all n).

18 3 Quadrature
3.1.2 Interpolatory Quadrature
The usual way to derive (3.1) is the ‘interpolatory quadrature’ via a (family of)
interpolation methods (see §4.1). Consider an interpolation
n
X
f (x) ≈ fn (x) := f (xi,n )Φi,n (x)
i=0
with Lagrange functions Φi,n (i.e., Φi,n belongs to the desired function space and
satisfies Φi,n (xj,n ) = δij ; cf. (4.2)). Integrating fn instead of f , one obtains
Z 1 Z 1 n
X Z 1
f (x)dx ≈ fn (x)dx = f (xi,n ) Φi,n (x)dx.
0 0 i=0 0
| {z }
=:ai,n
The standard choice is polynomial interpolation. In this case, Φi,n = Li,n are the
Lagrange polynomials
Y x − xk,n
Li,n (x) := . (3.2)
xi,n − xk,n
k∈{1,...,n}\{i}
The advantage of interpolatory quadrature rules is their exactness for certain

functions.
Remark 3.2. Let the interpolatory quadrature be based on interpolation involving

the functions Φi,n . Then all functions
R1 from Un := span{Φi,n : 0 ≤ i ≤ n} are
integrated exactly: Qn (f ) = 0 f (x)dx for f ∈ Un . In the case of polynomial
interpolation, Un is the space of polynomials of degree ≤ n.
The latter statement does not exclude that even more functions (e.g., polynomials
of higher order) are also integrated exactly.
The next exercise refers to the case of polynomials, where Φi,n = Li,n are the
Lagrange polynomials (3.2).
R1
Exercise 3.3. The definition ai,n = 0 Li,n (x)dx is less helpful for its computation.
Show that instead one can obtain the values ai,n by solving the following system of
linear equations:
n
X 1
ai,n xki,n = for k = 0, 1, . . . , n.
i=0
k+1
R1
Hint: Verify that Qn (xk ) = 0
xk dx.
So far the quadrature points xi,n are not fixed. Their choice determines the
(family of) quadrature formulae.
3.2 Consistency 19
3.1.3 Newton–Cotes Quadrature
Choose equidistant quadrature points in [0, 1]:1
i
xi,n = for i = 0, 1, . . . , n.
n
The Newton–Cotes quadrature2 is the interpolatory quadrature based on the poly-
nomial interpolation at the nodes xi,n .
Exercise 3.4. Prove the symmetry ai,n = an−i,n (0 ≤ i ≤ n) of the Newton–Cotes

quadrature weights.
For later use we cite the following asymptotic statement about the Newton–Cotes
weights from Ouspensky [6] (also mentioned in [3, p. 79]):
i−1 n
(−1) n! 1 (−1) 1

ai,n = + 1+O (3.3)
i! (n − i)!n log2 n
2 i n−i log n
for 1 ≤ i ≤ n − 1.
3.1.4 Gauss Quadrature
The Legendre polynomial Ln is a polynomial of degree n with the orthogonality

R1
property −1 Ln (x)p(x)dx = 0 for all polynomials p of degree ≤ n − 1 (cf.
[7, §10.1.2]). This defines Ln uniquely up to a scalar factor, which is irrelevant
for our purpose. We transform Ln from [−1, 1] onto the integration interval [0, 1]:
L̂n (t) = Ln (2t − 1). It is well known that Ln has n distinct roots ξi in [−1, 1].
Hence L̂n has n distinct roots xi := (1 + ξi ) /2 in [0, 1]. In the sequel, we rename
L̂n by Ln and call xi the roots of the (transformed) Legendre polynomial.
Let xi,n (0 ≤ i ≤ n) be the zeros of the (transformed) Legendre polynomial
Ln+1 (not Ln !). The interpolatory quadrature with these quadrature points yields
the Gauss quadrature.3
3.2 Consistency
For the interpolatory quadrature defined via polynomial interpolation one uses the
following consistency definition.
1
The family of Newton–Cotes formulae is defined for n ∈ N only, not for n = 0. Formally, we
may add Q0 (f ) := f (0).
2
See Newton’s remarks in [11, pp. 73-74].
3
The original publication of Gauss [4] is from 1814. Christoffel [2] generalises the method to
integrals with a weight function. For a modern description, see Stroud–Secrest [10].
20 3 Quadrature
Definition 3.5. A family {Qn : n ∈ N0 } of quadratures is called consistent if there

is a function g : N0 → N with g(n) → ∞ for n → ∞, so that
Z 1
Qn (P ) = P (x)dx for all polynomials P with degree(P ) ≤ g(n). (3.4)
0
An immediate consequence is the next statement.
Corollary 3.6. Let {Qn : n ∈ N0 } be consistent. Then for any polynomial P we

have Z 1
lim Qn (P ) = P (x)dx.
n→∞ 0
Proof. Set γ := degree(P ). Because of g(n) → ∞ for n → ∞, there is an n0

R1
with g(n) ≥ γ for all n ≥ n0 . Hence, Qn (P ) = 0 P (x)dx for n ≥ n0 proves the
assertion. t
u
The quadrature families mentioned above satisfy (3.4) with the following values
of g(n).
Proposition 3.7. The maximal orders g(n) are

n n odd
g(n) = for the Newton–Cotes quadrature,
n + 1 n even
g(n) = 2n + 1 for the Gauss quadrature.
Proof. (i) Let P be a polynomial of degree n. The polynomial interpolation Pn of

P in n + 1 points yields Pn = P . Hence, by definition, any interpolatory quadrature
R1 R1
satisfies Qn (P ) = 0 Pn (x)dx = 0 P (x)dx, which proves g(n) ≥ n.
(ii) For even n, any polynomial of degree n + 1 has the form
Pn+1 = a(x − 1/2)n+1 + Pn ,
where a ∈ C and Pn has degree n. This proves
Qn (Pn+1 ) = aQn ((x − 1/2)n+1 ) + Qn (Pn ).
Qn (Pn ) = 0 follows from Part (i), while in the case of the Newton–Cotes quadra-
ture Qn ((x − 1/2)n+1 ) = 0 can be concluded from Exercise 3.4.
(iii) Let P be a polynomial of degree 2n + 1 and consider the Gauss quadra-
ture. The Euclidean algorithms allows us to divide P by the (transformed) Legendre
polynomial Ln+1 : P = pLn+1 + q. Both polynomials p and q are of degree ≤ n.
R1
The integral 0 p(x)Ln+1 (x)dx vanishes because of the orthogonality property of
Ln+1 . On the other hand, Qn (pLn+1 ) = 0 follows from the fact that Ln+1 vanishes
R1 R1
at the quadrature points. Hence, 0 P (x)dx = 0 q(x)dx =(i) Qn (q) = Qn (P )
proves the assertion. t u
3.2 Consistency 21
Later, we shall formulate an alternative (more general) consistency condition (cf.

§3.4.9).
We conclude this section with consistency considerations for sufficiently smooth
functions known as the Peano kernel representation of the quadrature error. Below
we use the maximum norm
kϕk∞ = max{|ϕ(x)| : x ∈ [0, 1]}. (3.5)
Remark 3.8 (Peano kernel). Let Qn have maximal consistency order gn . For some
m ∈ N with m ≤ gn suppose that f ∈ C m+1 ([0, 1]). Then the quadrature error
equals
Z 1 Z 1
f (x)dx − Qn (f ) = πm (x, y)f (m+1) (y)dy (πm defined in the proof).
0 0
R1 (3.6a)
The error εn := 0 f (x)dx − Qn (f ) is estimated by
εn ≤ α1,m kf (m+1) k∞ and εn ≤ α2,m kf (m+1) kL2 ([0,1]) (3.6b)

s
Z 1 Z 1
with α1,m := |πm (x, y)|dy, α2,m := |πm (x, y)|2 dy.
0 0
If πm (x, y) does not change sign, the following error equality holds for a suitable
intermediate value ξ ∈ [0, 1]:
Z 1 Z 1
f (x)dx − Qn (f ) = αm f (m+1) (ξ) with αm := πm (x, y)dy. (3.6c)
0 0
Proof. The Taylor representation with remainder term yields f (x) = Pm (x)+r(x),
where the polynomial Pm of degree ≤ m is irrelevant, since its quadrature error
vanishes. Hence the quadrature error of f is equal to that of r. The explicit form of
r is
Z x Z 1
1
r(x) = (x − y)m f (m+1) (y)dy = (x − y)m +f
(m+1)
(y)dy,
m! 0 0
where (t)+ = t for t ≥ 0 and (t)+ = 0, otherwise. The quadrature error of r equals
Z 1
r(x)dx − Qn (r)
0
Z 1 Z 1 n
X Z 1
= (x − y)m
+f
(m+1)
(y)dydx − ai,n (xi,n − y)m
+f
(m+1)
(y)dy
0 0 i=0 0
"Z n
#Z
Z 1 1 X 1
= (x − y)m
+ dx − ai,n (xi,n − y)m
+ f (m+1) (y)dy,
0 0 i=0 0
where the bracket defines the Peano kernel πm (x, y) in (3.6a). t

u
22 3 Quadrature
3.3 Convergence
3.3.1 Definitions and Estimates
One can pose different questions concerning convergence. Given a function f , we

define the error Z 1

εn (f ) := Qn (f ) − f (x)dx . (3.7)
0
The first question concerns pure convergence: does limn→∞ εn (f ) = 0 hold?

For numerical purposes, limn→∞ εn (f ) = 0 is a rather poor statement, since
convergence may be arbitrarily slow. Usually, we have in mind a certain accuracy
τ and try to perform the computation of Qn (f ) for some n such that εn (f ) ≤ τ .
Or, vice versa, we perform the computation for a fixed n and ask for an estimate of
εn (f ) for this very n.
In the case of interpolatory quadrature, an estimate of εn (f ) for a fixed n can be
obtained via the interpolation error f − fn , since
Z 1 Z 1 Z 1

εn (f ) = Qn (f ) − f (x)dx = fn (x)dx − f (x)dx
0 0 0
Z 1
≤ |fn (x) − f (x)| dx ≤ kfn − f k∞ ,
0
(for k·k∞ compare §3.4.7.1). However, this estimate may be too pessimistic. An
optimal error analysis can be based on the Peano kernel (cf. Remark 3.8). In any
case one obtains bounds of the form
εn (f ) ≤ cn kf (kn ) k∞ , (3.8)
involving derivatives of f of order kn ≤ g(n) + 1.

Obviously, the right-hand side in (3.8) depends on cn and f (including its deriva-
tives). The quadrature family {Qn } is only responsible for the constants cn , so that
the convergence cn kf (kn ) k∞ → 0 cannot be taken as a property of {Qn : n ∈ N0 }.
Moreover, the inequality (3.8) is applicable to f ∈ C k ([0, 1]) only for n satisfying
kn ≤ k, which excludes kn → ∞.
One may ask for error estimates (3.8) with kn = 0:
εn (f ) ≤ cn kf k∞ (or more generally, εn (f ) ≤ cn kf (k) k∞ with fixed k).
The answer is that in this case, cn → 0 cannot hold. For a proof, modify the constant
function f = 1 in η-neighbourhoods of the quadrature points xi,n (0 ≤ i ≤ n) such
that 0 ≤ f˜ ≤ f and f˜(xi,n ) = 0. Because of f˜(xi,n ) = 0, we conclude that
R1
Qn (f˜) = 0, while for sufficiently small η, the integral 0 f˜(x)dx is arbitrarily close
R1
to 0 f (x)dx = 1 (the difference is bounded by δ := 2nη). Since kf k∞ = kf˜k∞ = 1,
we obtain no better error estimate than
3.3 Convergence 23
Z 1
εn (f˜) = f˜(x)dx ≥ 1 − δ = 1 · kf˜k∞ − δ; i.e., cn ≥ 1.
0
This proves the following remark.
Remark 3.9. Estimates of the form εn (f ) ≤ cn kf k∞ require constants cn ≥ 1.
Hence the right-hand sides cn kf k∞ cannot be a zero sequence. Nevertheless,

εn (f ) → 0 may hold. This leads to the next definition.
Definition 3.10. A family {Qn : n ∈ N0 } of quadrature formulae is called conver-

gent if
Z 1
Qn (f ) → f (x)dx for all f ∈ C([0, 1]). (3.9)
0
Note that (3.9) is an equivalent formulation of εn (f ) → 0.
3.3.2 Functionals, Dual Norm, and Dual Space
Above we made use of the Banach space X = C([0, 1]) equipped with the maxi-
mum norm k·kX = k·k∞ from (3.5). The dual space of X consists of all linear and
continuous mappings φ from X into R. Continuity of φ is equivalent to bounded-
ness; i.e., the following dual norm must be finite:
∗
kφkX := sup{|φ(f )| : f ∈ X, kf kX = 1}
= sup{|φ(f )| / kf kX : 0 6= f ∈ X}.
Hence the dual space X ∗ is defined by

∗
X ∗ := {φ : X → R linear with kφkX < ∞}.
For the application of φ ∈ X ∗ onto an element f ∈ X, there are the following

equivalent notations:
φ(f ) = hφ, f iX ∗ ×X = hf, φiX×X ∗ . (3.10)
A consequence of the Hahn–Banach theorem (cf. Yosida [12, 1.§IV.4]) is the

following statement.
Corollary 3.11. For any f ∈ X, there is a functional φf ∈ X ∗ with

∗
kφf kX = 1 and φf (f ) = kf kX .
In §2.4.1.2 we introduced the condition of a mapping by the amplification of

perturbations. In the case of (linear) functionals there is a simple answer.
24 3 Quadrature
Remark 3.12. A perturbation of f ∈ X by δf ∈ X is amplified by the mapping

∗
φ ∈ X ∗ by at most the factor kφkX ; i.e.,
∗
|φ(f + δf ) − φ(f )| ≤ kφkX kδf kX .
∗
Proof. Use φ(f + δf ) − φ(f ) = φ(δf ) and the definition of kφkX . t
u
Exercise 3.13. The integral
Z 1
I(f ) := f (x)dx
0
∗
is a linear functional on X = C([0, 1]) with kIkX = 1.
Another functional is the Dirac function(al) δa ∈ X ∗ defined by δa (f ) := f (a).
∗
Here kδa kX = 1 holds.
The quadrature formula (3.1) is a functional which may be expressed in terms of
Dirac functionals:
n
X
Qn = ai,n δxi,n . (3.11)
i=0
Lemma 3.14. Let X = C([0, 1]). The quadrature formula (3.11) has the dual norm
n
∗
X
kQn kX = |ai,n | .
i=0
Proof. Let f ∈ X satisfy kf kX = 1. Then

n n
X X
|Qn (f )| = ai,n f (xi,n ) ≤ |ai,n |

i=0 i=0
∗ Pn
holds because of |f (xi,n )| ≤ kf kX = 1. This proves kQn kX ≤ i=0 |ai,n |. The
equality sign is obtained by choosing the following particular function g ∈ X. Set
g(xi,n ) := sign(ai,n ) and interpolate between the quadrature points and the end
points 0, 1 linearly. Then g is continuous, i.e., g ∈ X, and kgkX = 1. The definition
yields the reverse inequality
n
∗
X
kQn kX = sup{|Qn (f )| : f ∈ X, kf kX = 1} ≥ |Qn (g)| = |ai,n | . t
u
i=0
The convergence definition (3.9) now reads as En (f ) → 0 for all f ∈ X, where

En := Qn − I is the error functional.
In principle, Qn may consist of functionals R 1 other than point evaluations δxi,n .
However, nonlocal functionals like f 7→ 0 qi,n (x)f (x)dx are impractical, since
again they require an integration. One may take into account evaluations of deriva-
tives: f 7→ f 0 (xi,n ); e.g., interpolatory quadrature based on Hermite interpolation.
In this case, the Banach space X = C 1 ([0, 1]) must be chosen.
3.4 Stability 25
3.4 Stability
3.4.1 Amplification of the Input Error
First, we introduce stability by a numerical argument. Let f ∈ C([0, 1]) be the

function to be integrated. The input data for Qn (f ) are {f (xi,n ) : 0 ≤ i ≤ n}.
Let f˜ be the floating point realisation4 of f with absolute errors δfi,n :
f˜i,n := f˜(xi,n ) = f (xi,n ) + δfi,n (0 ≤ i ≤ n).
It follows that
|δfi,n | ≤ kf − f˜k∞ .
First, we investigate the conditioning of integration and quadrature.

R1
Remark 3.15. (a) The integration problem f 7→ I(f ) := 0 f (x)dx is well-
conditioned. The error estimate is
|I(f ) − I(f˜)| ≤ kf − f˜k∞ for all f, f˜ ∈ C([0, 1]). (3.12)
(b) The corresponding error estimate for the quadrature Qn is
|Qn (f˜) − Qn (f )| ≤ Cn kf − f˜k∞ (3.13a)

n
∗
X
with the condition number Cn := kQn kX = |ai,n | . (3.13b)
i=0
Proof. Use Remark 3.12, Exercise 3.13, and Lemma 3.14. t

u
3.4.2 Definition of Stability
The constant Cn from (3.13b) is the error amplification factor of the quadrature Qn .
To avoid an increasing error amplification, we have to require that Cn be uniformly
bounded. This leads us directly to the stability definition.
Definition 3.16 (stability). A quadrature family {Qn : n ∈ N0 } is called stable if

n
X
Cstab := sup Cn = sup |ai,n | < ∞. (3.14)
n∈N0 n∈N0 i=0
An equivalent formulation is given below.

4
To be quite precise, f˜ is only defined on the set M of machine numbers. However, for theoretical
purposes, we may extend the function continuously (use, e.g., piecewise interpolation).
26 3 Quadrature
Remark 3.17. (a) A family {Qn : n ∈ N0 } of quadrature formulae is stable if and

only if there is some C such that
|Qn (f )| ≤ Ckf k∞ for all f ∈ C([0, 1]) and all n ∈ N0 . (3.15)
(b) Cstab from (3.14) is the minimal constant C in (3.15).
From Part (b) of the remark and after replacing f by f − g, one obtains the
following estimate analogous to (3.13a).
Corollary 3.18. |Qn (f ) − Qn (g)| ≤ Cstab kf − gk∞ holds for all f, g ∈ C([0, 1])
and all n ∈ N0 .
Next we discuss the consequences of stability or instability in the numerical con-

text. By f˜ we denoted the numerically evaluated function, which may be considered
R1
as a floating point version of f . The total error Qn (f˜)− 0 f (x)dx can be estimated
by the triangle inequality:
Z 1 Z 1
Qn (f˜) − ˜

f (x)dx ≤ Qn (f ) − Qn (f ) + Qn (f ) − f (x)dx

0 0
≤ Cn kf˜ − f k∞ + εn (f )
(εn (f ) from (3.7)). If stability holds, the error is bounded by Cstab kf˜−f k∞ +εn (f )
(cf. Corollary 3.18). Provided that εn (f ) → 0, the total error approaches the level
of numerical noise Cstab kf˜ − f k∞ , which is unavoidable since it is caused by the
input error kf˜ − f k∞ .
In the case of instability, Cn → ∞ holds. While the term εn (f ) approaches zero,
Cn kf˜ − f k∞ tends to infinity. Hence, an enlargement of n does not guarantee a
better result. If one does not have further information about the behaviour of εn (f ),
it is difficult to find an n such that the total error is as small as possible.
In spite of what has been said about the negative consequences of instability,
we have to state that the quality of a quadrature rule for fixed n has no relation to
stability or instability. The sensitivity of Qn to input errors is given only by the
amplification factor Cn . Note that the size of Cn is not influenced by whether the
sequence (Cn )n∈N is divergent or convergent.
3.4.3 Stability of Particular Quadrature Formulae
Under the minimal condition that Qn be exact for constants (polynomials of order
R1 Pn
zero; i.e., g(n) ≥ 0 in (3.4)), one concludes that 1 = 0 dx = Qn (1) = i=0 ai,n :
n
X
ai,n = 1. (3.16)
i=0
3.4 Stability 27
Conclusion 3.19. Assume (3.16). (a) If the quadrature weights are non-negative
(i.e., ai,n ≥ 0 for all i and n), then the family {Qn } is stable with
Cstab = Cn = 1 for all n.
(b) In general, the stability constant is bounded from below by Cstab ≥ 1.

Pn Pn Pn
Proof. For (a) use |ai,n | = ai,n = 1. Part (b) follows from |ai,n | ≥
i=0 i=0 i=0
n
P
ai,n = |1| = 1. t u

i=0
The latter statement is based upon (3.16). A weaker formulation is provided next.
Exercise 3.20. Conclusion 3.19 remains valid, if (3.16) is replaced by the condition
lim Qn (1) = 1.
n→∞
A very important property of the Gauss quadrature is the following well-known

fact.
Lemma 3.21. The Gauss quadrature has non-negative weights: ai,n ≥ 0. Hence
the family of Gauss quadratures is stable with constant Cstab = 1.
Proof. Define the polynomial P2n (x) := 0≤k≤n,k6=i (x − xk,n )2 of degree 2n.
Q
R1
Obviously, 0 P2n (x)dx > 0. Since Qn is exact for polynomials of degree ≤ 2n+1,
also ai,n P2n (xi,n ) = Qn (P2n ) > 0. The assertion follows from P2n (xi,n ) > 0. t
u
The Newton–Cotes formulae satisfy ai,n ≥ 0 only for n ∈ {1, 2, 3, 4, 5, 6, 7, 9}.

On the other hand, the existence of some negative weights ai,n < 0 does not neces-
sarily imply instability. The following table shows the values of Cn from (3.13b):
n 1 to 7 8 9 10 11 12 14 16 18 20 22 24
Cn 1 1.45 1 3.065 1.589 7.532 20.34 58.46 175.5 544.2 1606 9923
Obviously, Cn increases exponentially to infinity; i.e., the Newton–Cotes formulae
seem to be unstable. An exact proof of instability can be based on the asymptotic
description (3.3) of ai,n . For even n, the following inequality holds:
n
X
Cn = |ai,n | ≥ a n2 ,n (ai,n from (3.3)).
i=0
Exercise 3.22. (a) Recall Stirling’s formula for the asymptotic representation of n!
(cf. [13, §1.14.16], [5, Anhang 1]).
(b) Using (a), study the behaviour of |a n2 ,n | and conclude that the family of
Newton–Cotes formulae is unstable.
A further example of a quadrature rule follows in the next section.

28 3 Quadrature
3.4.4 Romberg Quadrature
ThePexistence of negative weights ai,n is not yet a reason for instability, as long
n
as i=0 |ai,n | stays uniformly bounded. The following Romberg quadrature is an
example of a stable quadrature involving negative weights.
For h = 1/N with N ∈ N, the sum

1 1
T (f, h) := h f (0) + f (h) + f (2h) + . . . + f (1 − h) + f (1)
2 2
represents the compound trapezoidal rule. Under the assumption f ∈ C m ([0, 1]),
m even, one can prove the asymptotic expansion
Z 1
T (f, h) = f (x)dx+h2 e2 (f )+. . .+hm−2 em−2 (f )+O(hm kf (m) k∞ ) (3.17)
0
(cf. Bulirsch [1], [7, §9.6]). Hence, the Richardson extrapolation is applicable:
compute T (f, hi ) for different hi , i = 0, . . . , n, and extrapolate the values
2
hi , T (f, hi ) : i = 0, . . . , n
at h = 0. The result can be represented explicitly by the Lagrange polynomials (cf.

Exercise 4.2):
n n n
X Y h2ν X
Qn (f ) := T (f, hi ) = ci,n T (f, hi ). (3.18)
i=0
h2 − h2i
ν=0 ν i=0
ν6=i
| {z }
=:ci,n
We fix an infinite sequence of step sizes
h0 > h1 > h2 > . . . , hi = 1/Ni , Ni ∈ N,
with the property
hi+1 ≤ αhi with 0 < α < 1 for all i ≥ 0. (3.19)
The original quadrature of Romberg [8] is based on α = 1/2. Condition (3.19)

enforces hi → 0 as i → ∞. One infers from (3.19) that
hi /hj ≤ αi−j for j ≤ i hi /hj ≥ αi−j for j ≥ i.

and (3.20)
Pn
Lemma 3.23. There is a constant C < ∞, so that i=0 |ci,n | ≤ C for all n ∈ N0
(ci,n from (3.18)).
Proof. First, we recall simple inequalities, gathered in the next exercise.

3.4 Stability 29
1
Exercise 3.24. Prove (a) 1 + x ≤ ex for all real x and (b) 1−x ≤ 1 + ϑx with
1
ϑ = 1−x 0
for all 0 ≤ x ≤ x0 < 1.
1
(i) Part (b) with ϑ = 1−α2 yields
m m m ∞
Y 1 Y
2j
Y 2j
Y
exp ϑα2j =: A,

2j
≤ 1 + ϑα ≤ exp ϑα ≤
j=1
1−α j=1 j=1 j=1
P
∞ α2 ϑ

where A = exp j=1 ϑα2j = exp 1−α 2 2
2 = exp α ϑ . This implies that
m m
Y α2j Y
2j
≤ A α2j = Aαm(m+1) ≤ Aαm for all m ≥ 0.
j=1
1 − α j=1
Q Qi−1 Qn
(ii) Split the product ν6=i in (3.18) into the partial products ν=0 and ν=i+1 .
The first one is estimated by
i−1 i−1 i−1 i
Y h2ν Y 1 Y 1 Y 1
= ≤ = ≤ A,

h2 − h2 1 − h 2 /h2
1 − α 2(i−ν) 1 − α2j (i)

ν=0 ν i ν=0 i ν (3.20) ν=0 j=1
while the second one satisfies

n n n n−i
Y h2ν Y 1 Y 1 Y 1
= ≤ =

h2 − h2
h2 /h2 − 1
α2(i−ν) − 1 α −2j − 1
(3.20)
ν=i+1 ν i ν=i+1 i ν

ν=i+1 j=1
n−i
Y α2j
= 2j
≤ Aαn−i .
j=1
1 − α (ii)
(iii) The estimate

n n i−1 n n
X X Y h2ν Y h2ν 2
X
|ci,n | = 2 − h2
× 2 − h2
≤ A αn−i
h h

(iii)
i=0 ν=0 ν i ν=i+1 ν i

i=0 i=0
∞
X A2
< A2 αj = =: C
j=0
1−α
proves the assertion. t

u
Sn
The quadrature points {xj,n } used in Qn are i=0 {0, hi , 2hi , . . . , 1}. For
hi = 1/Ni such that Ni and Nj (0 ≤ i < j ≤ n) are relative prime,5 the weight
associated to the quadrature point xj,n = hn−1 is aj,n = 12 hn−1 cn−1,n . Because of
n−i
sign(ci,n ) = (−1) , we conclude that Qn contains negative weights.
5
Under this assumption, the set of interior grid points {νhi : 1 ≤ ν ≤ Ni − 1} are disjoint and
the weights do not add up. However, even without this assumptions one finds grid points xj,n with
negative weight.
30 3 Quadrature
Lemma 3.25. The family of the Romberg quadratures {Qn } in (3.18) is stable.
PNi
Proof. The compound trapezoidal rule T (f, hi ) = k=0 τk,i f (khi ) has the weights
τk,i = hi for 0 < k < Ni and τk,i = hi /2 for k = 0, Ni . In particular,
Ni
X Ni
X
|τk,i | = τk,i = 1
k=0 k=0
holds. The quadrature formula Qn is defined by
X Ni
X X X
Qn (f ) = ci,n τk,i f (khi ) = aj,n f (xj,n ) with aj,n := ci,n τk,i .
i k=0 j (i,k):khi =xj,n
P P PNi P
Now, j |aj,n | ≤ i |ci,n | k=0 |τk,i | = i |ci,n | ≤ C proves the sta-
Lemma 3.23
bility condition. t
u
Lemma 3.26. The family of the Romberg quadratures {Qn } in (3.18) is consistent.
Proof. Let P be a polynomial of degree ≤ g(n) := 2n + 1. In (3.17), the remainder
term for m := 2n + 2 vanishes, since P (m) = 0. This proves that T (f, h) is a
polynomial of degree ≤ n in the variable h2 . Extrapolation eliminates the terms
R1
hji ej (P ), j = 2, 4, . . . , 2n, so that Qn (P ) = 0 P (x)dx. Since g(n) = 2n + 1 → ∞
for n → ∞, consistency according to Definition 3.5 is shown. t u
The later Theorem 3.36 will prove convergence of the Romberg quadrature.
Exercise 3.27. Condition hi+1 ≤ αhi from (3.19) can be weakened. Prove: if an
` ∈ N and an α ∈ (0, 1) exist such that hi+` ≤ αhi for all i ≥ 0, then Lemma 3.23
remains valid.
3.4.5 Approximation Theorem of Weierstrass
For the next step of the proof we need the well-known approximation theorem of
Weierstrass.
Theorem 3.28. For all ε > 0 and all f ∈ C([0, 1]) there is a polynomial P = Pε,f
with kf − P k∞ ≤ ε.
An equivalent formulation is: the set P of all polynomials is a dense subset of
C([0, 1]).
In the following we prove a more general form (Stone–Weierstrass theorem). The
next theorem uses the point-wise maximum Max(f, g)(x) := max{f (x), g(x)}
and point-wise minimum Min(f, g)(x) := min{f (x), g(x)} of two functions. The
following condition (i) describes that F is closed under these mappings. Condition
(ii) characterises the approximability at two points (‘separation of points’).
3.4 Stability 31
Lemma 3.29. Let Q ⊂ Rd be compact. Suppose that a family F ⊂ C(Q) of

continuous functions satisfies the following two properties:
(i) Max(f1 , f2 ), Min(f1 , f2 ) ∈ F for all f1 , f2 ∈ F.
(ii) For all x0 , x00 ∈ Q, all ε > 0, and all g ∈ C(Q) there is a ϕ ∈ F such that
|ϕ(x0 ) − g(x0 )| < ε and |ϕ(x00 ) − g(x00 )| < ε.
Then for all ε > 0 and all g ∈ C(Q) there exists a function f ∈ F with
kf − gk∞ < ε (i.e., F is dense in C(Q)).
Proof. (a) Let ε > 0 and g ∈ C(Q) be given. Fix some x0 = x0 ∈ Q, while the
second point x00 = y ∈ Q will be variable. By assumption (ii) there is a function
h = h(· ; x0 , y, ε) with
|h(x0 ) − g(x0 )| < ε, |h(y) − g(y)| < ε.
The latter inequality yields in particular that g(y) − ε < h(y). By continuity of h
and g, this inequality holds in a whole neighbourhood U (y) of y:
g(x) − ε < h(x) for all x ∈ U (y).

S
Since U (y) covers the compact set Q, there is a finite subset of neighbour-
y∈Q
Sn
hoods {U (yi ) : i = 1, . . . , n} covering Q: i=1 U (yi ) ⊃ Q. Each yi is associated
to a function h(· ; x0 , yi , ε) ∈ F with
g(x) − ε < h(x; x0 , yi , ε) for all x ∈ U (yi ).
By assumption (i), h(· ; x, ε) := Maxi=1,...,n h(· ; x0 , yi , ε) again belongs to F and

satisfies
g(x) − ε < h(x; x0 , ε) for all x ∈ Q.
(b) Next, the parameter x0 becomes a variable in Q. Since all h(· ; x0 , yi , ε) approx-
imate the function g at x0 , the opposite inequality g(x0 ) + ε > h(x0 ; x0 , ε) holds.
Again, there is a neighbourhood V (x0 ), so that
g(x) + ε > h(x; x0 , ε) for all x ∈ V (x0 ).
As in Part (a), one finds a finite covering {V (xi ) : i = 1, . . . , m} of Q. The function

f := Mini=1,...,m h(· ; xi , ε)
belongs again to F and satisfies g + ε > f . Since each h(· ; xi , ε) satisfies the
inequality g − ε < h(· ; xi , ε) from Part (a), also g − ε < f follows. Together, one
obtains kf − gk∞ < ε; i.e., f ∈ F is the desired approximant. t u
Remark 3.30. Instead of the lattice operations Max and Min, one can equivalently
require that F be closed with respect to the absolute value:
f ∈F ⇒ |f | ∈ F,
where |f | is defined point-wise: |f |(x) := |f (x)| for all x ∈ Q.

32 3 Quadrature
Proof. Use Max(f, g) = 12 (f + g)+ 12 |f −g| and Min(f, g) = 1

2 (f + g)− 12 |f −g|
and in the reverse direction |f | = Max(f, −f ). t
u
The addition and multiplication of functions is defined point-wise: (f + g) (x) =

f (x) + g(x), (f · g) (x) = f (x)g(x), x ∈ Q. Correspondingly, multiplication by
scalars from the field K is defined point-wise: (λf ) (x) = λf (x).
Definition 3.31 (algebra of functions). A set A of functions is called an algebra, if

A (without multiplication) is a vector space, and, additionally, is equipped with the
multiplication satisfying the usual distributive law.
Example 3.32. Examples of algebras are all (a) continuous functions on Q ⊂ Rd

(no compactness of Q required), (b) bounded functions on Q ⊂ Rd , (c) polynomials,
(d) trigonometric functions.
In the case of (d) in the previous example one has to show that, e.g., the product
sin(nx) cos(mx) (n, m ∈ N0 ) is again a trigonometric function. This follows from
2 sin(nx) cos(mx) = sin ((n + m)x) + sin ((n − m)x).
If A ⊂ C(Q) is an algebra, the closure Ā (with respect to the maximum norm
k·k∞ ) is called the closed hull of the algebra A.
Exercise 3.33. If A is an algebra of continuous functions, also Ā is an algebra of

continuous functions; i.e., f, g ∈ Ā implies f + g ∈ Ā and f · g ∈ Ā.
Lemma 3.34 (Weierstrass). Let A ⊂ C(Q) be an algebra. Then |f | ∈ Ā for all

f ∈ A.
Proof. (i) A simple scaling argument shows that it suffices to show the assertion for
f ∈ A with kf k∞ ≤ 1. p
(ii) Let ε > 0 be given. The function T (ζ) := Pζ + ε2 is holomorphic in the
complex half-plane <e ζ > −ε2 . The Taylor series αν (x − 12 )ν of T (x) has the
convergence radius 12 + ε2 and converges uniformly in the interval [0, 1]. Hence
there is a finite Taylor polynomial Pn of degree n, so that
p
x + ε2 − Pn (x) ≤ ε for all 0 ≤ x ≤ 1.

(iii) Replacing x by x2 , we obtain

p
x2 + ε2 − Pn (x2 ) ≤ ε for all − 1 ≤ x ≤ 1.

The particular case x = 0 shows |ε − Pn (0)| ≤ ε and, therefore, |Pn (0)| ≤ 2ε. The
polynomial Q2n (x) := Pn (x2 ) − Pn (0) of degree 2n has a vanishing absolute term
and satisfies
p p
x2 + ε2 − Q2n (x) ≤ x2 + ε2 − Pn (x2 ) + |Pn (0)| ≤ 3ε

for all −1 ≤ x ≤ 1. Using

3.4 Stability 33
p ε2 ε2
x2 + ε2 − |x| = √ 2 ≤ =ε for all − 1 ≤ x ≤ 1,

x + ε2 + |x| ε
we obtain the inequality

|x| − Q2n (x) ≤ 4ε for all − 1 ≤ x ≤ 1.

(iv) For all f with kf k∞ ≤ 1, the values f (ξ) (ξ ∈ Q) satisfy the inequality
−1 ≤ f (ξ) ≤ 1, so that x = f (ξ) can be inserted into the last inequality:

|f (ξ)| − Q2n (f (ξ)) ≤ 4ε for all ξ ∈ Q.

Because of 6
n n
!
X 2ν
X
2ν
Q2n (f (ξ)) = qν (f (ξ)) = qν f (ξ),
ν=1 ν=1
Q2n (f ) belongs again to A and satisfies the estimate ||f | − Q2n (f )| ≤ 4ε. As
ε > 0 is arbitrary, |f | belongs to the closure of A. t
u
Now we prove the theorem of Weierstrass in the generalised form of Stone:
Theorem 3.35 (Stone–Weierstrass). Assume that

(i) Q ⊂ Rd is compact,
(ii) A is an algebra of continuous functions on Q (i.e., A ⊂ C(Q)),
(iii) A separates the points of Q; i.e., for any pair of points x0 , x00 ∈ Q with
x 6= x00 there is a function f ∈ A with f (x0 ) 6= f (x00 ).
0
Then the closed hull Ā satisfies either Ā = C(Q) or there is an x0 ∈ Q so that
Ā = {f ∈ C(Q) : f (x0 ) = 0} . (3.21)
Proof. (a) By Lemma 3.34 and Remark 3.30, F = Ā satisfies the requirement (i) of
Lemma 3.29. As soon as (ii) from Lemma 3.29 is shown, F̄ = C(Q) follows. Since
F = Ā is already closed, the first case Ā = C(Q) follows.
(b) For the proof of (ii) from Lemma 3.29, we consider the following alternative:
either for any x ∈ Q there is an f ∈ A with f (x) 6= 0 or there exists an x0 ∈ Q
with f (x0 ) = 0 for all f ∈ A. The first alternative will be investigated in (c), the
second one in (d).
(c) Assume the first alternative. First we prove the following:
Assertion: For points x0 , x00 ∈ Q with x0 6= x00 from Assumption (ii) in Lemma
3.29 there exists an f ∈ A with 0 6= f (x0 ) 6= f (x00 ) 6= 0.
For its proof we use the separability property (iii): f (x0 ) 6= f (x00 ) holds for a
suitable f . Assume f (x0 ) = 0, which implies f (x00 ) 6= 0 (the case f (x00 ) = 0 and
6
Here we use that q0 = 0, since f 0 = 1 may not necessarily belong to the algebra A.
34 3 Quadrature
f (x0 ) 6= 0 is completely analogous). The first alternative from Part (b) guarantees
the existence of an f0 ∈ A with f0 (x0 ) 6= 0. The function
fλ := f − λf0
has the properties
0 6= fλ (x00 ) for sufficiently small λ, (because of f (x00 ) 6= 0)

fλ (x00 ) 6= fλ (x0 ) for sufficiently small λ, (because of f (x0 ) 6= f (x00 ))
fλ (x0 ) = λf0 (x0 ) 6= 0 for all λ 6= 0.
Hence, for sufficiently small but positive λ, we have
0 6= fλ (x0 ) 6= fλ (x00 ) 6= 0, fλ ∈ A,
and fλ (renamed by f ) has the required properties. This proves the assertion.
Let g ∈ C(Q) be the function from assumption (ii) of Lemma 3.29. Concerning
the required ϕ, we make the ansatz ϕ = αf + βf 2 with f from the assertion. f ∈ A
implies that also f 2 , ϕ ∈ A. The conditions ϕ(x0 ) = g(x0 ) and ϕ(x00 ) = g(x00 ) lead
to a 2 × 2-system of linear equations for α and β. Since the determinant
f (x0 )f 2 (x00 ) − f (x00 )f 2 (x0 ) = f (x0 )f (x00 ) [f (x00 ) − f (x0 )]
does not vanish, solvability is ensured. Therefore, assumption (ii) of Lemma 3.29 is
satisfied even for ε = 0 and Part (a) shows Ā = C(Q).
(d) Assume the second alternative: there is an x0 ∈ Q with f (x0 ) = 0 for all
f ∈ A. Let 1I ∈ C(Q) be the function with constant value 1. Denote the algebra
generated from A and {1I} by A∗ ; i.e., A∗ = {f = g + λ1I: g ∈ A, λ ∈ K}.
Obviously, for any x ∈ Q there is an f ∈ A∗ with f (x) 6= 0 (namely f = 1I). Hence
the first alternative applies to A∗ . The previous proof shows A∗ = C(Q).
Let g ∈ C(Q) be an arbitrary function with g(x0 ) = 0, i.e., belonging to
the right-hand side of (3.21). Because of g ∈ C(Q) = A∗ , for all ε > 0 there
is an f ∗ ∈ A∗ with kg − f ∗ k∞ < ε. By definition of A∗ one may write f ∗ as
f ∗ = f + λ1I with f ∈ A. This shows that
kg − f − λ1Ik∞ < ε.
In particular, at x0 we have |g(x0 ) − f (x0 ) − λ| = |λ| < ε. Together, one obtains

kg − f k∞ < 2ε, where f ∈ A. This proves (3.21). t u
For the proof of Theorem 3.28 choose Q = [0, 1] (compact subset of R1 ) and
A as the algebra of all polynomials. For this algebra, assumption (iii) of Theorem
3.35 is satisfied with f (x) = x. Hence, one of the two alternatives Ā = C(Q) or
(3.21) holds. Since the constant function 1I belongs to A, (3.21) is excluded and
Ā = C(Q) is shown.
3.4 Stability 35
3.4.6 Convergence Theorem

R1
We recall Definition 3.10: Qn is convergent if Qn (f ) → 0 f (x)dx holds for all
f ∈ C([0, 1]). So far it has remained open as to whether this property would hold.
Note that all previous convergence results require more smoothness than continuity
of f . Now we can give a positive answer. The next theorem follows the pattern
consistency + stability =⇒ convergence. (3.22)
Theorem 3.36 (convergence theorem). If the family {Qn : n ∈ N0 } of quadrature

formulae is consistent and stable, then it is convergent.
Proof. Let ε > 0 be given. We have to show that for all f ∈ C([0, 1]) there is an n0
such that Z 1

Qn (f ) − f (x)dx ≤ε for n ≥ n0 .

0
Let P be an arbitrary polynomial. The triangle inequality yields

Z 1

Qn (f ) − f (x)dx

0
Z 1 Z 1 Z 1

= Qn (f ) − Qn (P ) + Qn (P ) −
P (x)dx + P (x)dx − f (x)dx
0 0 0
Z 1 Z 1 Z 1

≤ |Qn (f ) − Qn (P )| + Qn (P ) −
P (x)dx +
P (x)dx − f (x)dx .
0 0 0
We choose P according to Theorem 3.28 such that kf − P k∞ ≤ ε/ (1 + Cstab ),

where Cstab is the stability constant. Now, thanks to Corollary 3.18, the first term
|Qn (f ) − Qn (P )| can be estimated by
Cstab kf − P k∞ ≤ εCstab / (1 + Cstab ) .
The chosen P has a fixed degree(P ). Because of g(n) → ∞ there is an n0 such

that g(n) ≥ degree(P ) for all n ≥ n0 . Hence, consistency guarantees exactness of
R1
the quadrature: Qn (P ) − 0 P (x)dx = 0.
Remark 3.15 yields

Z 1 Z 1
P (x)dx − f (x)dx ≤ kf − P k∞ ≤ ε/ (1 + Cstab )
0 0
for the last term.

εCstab ε
Together, the sum of the three terms is bounded by 1+Cstab + 1+Cstab = ε. t
u
According to Theorem 3.36, stability is sufficient for convergence. Next we show

that the stability condition (3.14) is also necessary for convergence. As a tool, we
need a further theorem from functional analysis.
36 3 Quadrature
3.4.7 Uniform Boundedness Theorem
3.4.7.1 Banach Space Notations
X is called a normed (linear) space (the norm may be expressed by the explicit
notation (X, k·k)), if a norm k·k is defined on the vector space X. If necessary, we
write k·kX for the norm on X.
X is called a Banach space, if X is normed and complete (‘complete’ means that
all Cauchy sequences converge in X).
By L(X, Y ) we denote the set of linear and continuous mappings from X to Y .
Remark 3.37. (a) If X, Y are normed, also L(X, Y ) is normed. The associated
‘operator norm’ of T ∈ L(X, Y ) equals7
kT xkY
kT k := kT kY ←X := sup . (3.23)
x∈X\{0} kxkX
(b) By definition, a continuous map T : X → Y from L(X, Y ) leads always to a

finite supremum (3.23). Vice versa, if a linear operator T : X → Y yields a finite
value in (3.23) (i.e., T is bounded), then T is also continuous.
An example of a Banach space is the set X = C(D) of continuous functions
defined on D with bounded norm kf k∞ := supx∈D |f (x)| (for a compact D, the
supremum is even a maximum).
3.4.7.2 Theorem
The uniform boundedness theorem of Banach and Steinhaus connects point-wise

and normwise boundedness.
Theorem 3.38. Assume that
(a) X is a Banach space,
(b) Y is a normed space,
(c) T ⊂ L(X, Y ) is a (in general, infinite) subset of mappings,
(d) supT ∈T kT xkY < ∞ holds for all x ∈ X.
Then T is uniformly bounded; i.e., supT ∈T kT kY ←X < ∞.
First we add some remarks. Let K be the unit sphere K := {x ∈ X : kxk ≤ 1}.
Definition (3.23) states that
kT kY ←X = sup kT xkY .
x∈K
The statement of the theorem becomes supT ∈T supx∈K kT xk < ∞. Since suprema
commute, one may also write supx∈K supT ∈T kT xk < ∞. Assumption (d) of the
7
For the trivial case of X = {0}, we define the supremum over the empty set by zero.
3.4 Stability 37
theorem reads C(x) := supT ∈T kT xk < ∞; i.e., the function C(x) is point-wise
bounded. The astonishing8 property is that C(x) is even uniformly bounded on K.
In the later applications we often apply a particular variant of the theorem.
Corollary 3.39. Let X and Y be as in Theorem 3.38. Furthermore, assume that the
operators T, Tn ∈ L(X, Y ) (n ∈ N) satisfy either
(a) {Tn x} a Cauchy sequence for all x ∈ X, or
(b) there exists an operator T ∈ L(X, Y ) with Tn x → T x for all x ∈ X.
Then supn∈N kTn k < ∞ holds.
Proof. In this case T = {Tn ∈ L(X, Y ) : n ∈ N} is countably infinite.

Since any Cauchy sequence is bounded, the boundedness supn∈N kTn xkY < ∞
follows so that Theorem 3.38 is applicable. This proves part (a).
Assumption (b) implies (a). t u
3.4.7.3 Proof
The proof of Theorem 3.38 is based on two additional theorems. The first is called
Baire’s category theorem or the Baire–Hausdorff theorem.
6 ∅ be a complete metric space. Assume that X has a repre-

Theorem 3.40. Let X =
sentation [
X= Ak with closed sets Ak .
k∈N
Then there exists at least one k0 ∈ N, so that Åk0 6= ∅ (Åk0 denotes the interior of
Ak0 ).
Proof. (a) For an indirect proof assume Åk = ∅ for all k. We choose a non-empty,
open set U ⊂ X and some k ∈ N. Since Ak closed, U \Ak is again open and non-
empty (otherwise, Ak would contain the open set U ; i.e., Åk ⊃ U 6= ∅). Since
U \Ak is open, it contains a closed sphere Kε (x) with radius ε > 0 and midpoint x.
Without loss of generality, ε ≤ 1/k can be chosen.
(b) Starting with ε0 := 1 and x0 := 0, according to (a), we choose by induction
Kεk (xk ) ⊂ Kεk−1 (xk−1 )\Ak and εk ≤ 1/k.
Since x` ∈ Kεk (xk ) for ` ≥ k and εk → 0 (k → ∞), {xk } is a Cauchy sequence.

Because of completeness, it must converge to some x := lim xk ∈ X and belong
to KSεk (xk ) for all k. Since Kεk (xk ) ∩ Ak = ∅ by construction, it follows that
x∈/ k∈N Ak = X, which is a contradiction. t u
8
Only in the case of a finite-dimensional vector space X, is there a simple proof using
supT ∈T kT bi kY < ∞ for all basis vectors bi of X.
38 3 Quadrature
Theorem 3.41. Let X be a complete metric space and Y a normed space. For some
subset F ⊂ C 0 (X, Y ) of the continuous mappings, assume that sup kf (x)kY < ∞
for all x ∈ X. Then there exist x0 ∈ X and ε0 > 0 such that f ∈F
sup sup kf (x)kY < ∞. (3.24)

x∈Kε0 (x0 ) f ∈F
T
Proof. Set Ak := f ∈F {x ∈ X : kf (x)kY ≤ k} for k ∈ N and check that Ak
S According to the assumption, each x ∈ X must belong to some Ak ; i.e.,
is closed.
X = k∈N Ak . Hence the assumptions of Theorem 3.40 are satisfied. Correspond-
ingly, Åk0 6= ∅ holds for at least one k0 ∈ N. By the definition of Ak , we have
supx∈Ak supf ∈F kf (x)kY ≤ k0 . Choose a sphere with Kε0 (x0 ) ⊂ Ak0 . This
0
yields the desired inequality (3.24) with the bound ≤ k0 . t
u
For the proof of Theorem 3.38 note that a Banach space is also a complete metric
space and L(X, Y ) ⊂ C 0 (X, Y ), so that we may set F := T . The assumption
supf ∈F kf (x)kY < ∞ is equivalent to supT ∈T kT xkY < ∞. The result (3.24)
becomes supx∈Kε (x0 ) supT ∈T kT xkY < ∞. For an arbitrary ξ ∈ X\{0}, the
0
ε0
element xξ := x0 + kξkX ξ belongs to Kε0 (x0 ), so that
kT ξkY 1 1
= kT (xξ − x0 )kY ≤ kT xξ kY + kT x0 kY
kξkX ε0 ε0
is uniformly bounded for all T ∈ T and all ξ ∈ X\{0}. Hence the assertion of
kT ξk
Theorem 3.38 follows: sup sup kξk Y = sup kT k < ∞.
X
T ∈T ξ∈X\{0} T ∈T
3.4.8 Necessity of the Stability Condition, Equivalence Theorem
X = C([0, 1]) together with the maximum norm k·k∞ is a Banach space, and
Y := R is normed (its norm is the absolute value). The mappings
Z 1
f ∈ C([0, 1]) 7→ I(f ) := f (x)dx ∈ R and f ∈ C([0, 1]) 7→ Qn (f ) ∈ R
0
are linear and continuous; hence they belong to L(X, Y ). By Remark 3.37b, conti-
nuity is equivalent to boundedness, which is quantified in the following lemma.
Lemma 3.42. The operator norms of I and Qn ∈ L(X, Y ) are
n
X
kIk = 1, kQn k = Cn := |ai,n | . (3.25)
i=0
Proof. The estimates kIk ≤ 1 and kQn k ≤ Cn are equivalent to the estimates
(3.12) and (3.13a) from Remark 3.15. According to Remark 3.17b, Cn is the
3.4 Stability 39
minimal constant, implying kQn k = Cn . The example f = 1 shows that one is

the best bound for kIf k∞ / kf k∞ ; i.e., kIk = 1. t
u
Apply Corollary 3.39 to T := I and Tn := Qn . Convergence of the quadrature
formulae {Qn } is expressed by Qn (f ) → I(f ). From Corollary 3.39 we conclude
that supn∈N kTn k < ∞. According to Lemma 3.42, this means supn∈N Cn < ∞
and it is identical to the stability condition of Definition 3.16. Hence, the following
theorem is proved.
Theorem 3.43 (stability theorem). If the family {Qn } of quadrature formulae is
convergent, then {Qn } is stable.
We have already proved ‘consistency + stability =⇒ convergence’ (cf. (3.22)).
Theorem 3.43 yields ‘stability ⇐= convergence’. Together, we obtain the following
equivalence theorem.
Theorem 3.44 (equivalence theorem). Assume consistency of the family {Qn } of
quadrature formulae. Then stability and convergence are equivalent.
3.4.9 Modified Definitions for Consistency and Convergence
The terms ‘consistency’ and ‘convergence’ can be even better separated, without
weakening the previous statements.
The previous definition of convergence contains not only the statement that the
R1
sequence Qn (f ) is convergent, but also that it has the desired integral 0 f (x)dx
as the limit. The latter part can be omitted:
{Qn } is convergent, if lim Qn (f ) exists for all f ∈ C([0, 1]). (3.26)

n→∞
So far, the definition of consistency is connected with polynomials. Polynomials

come into play since we started from interpolatory quadrature based on polyno-
mial interpolation. An interpolatory quadrature, e.g., based on trigonometric inter-
polation, would not be consistent in the sense of Definition 3.5. According to the
sentence following Theorem 3.28, the decisive property of polynomials is that they
are dense in C([0, 1]). One may replace the polynomials by any other dense subset.
This leads us to the following generalisation of the term ‘consistency’:
{Qn } is consistent if there is a dense subset X0 ⊂ C([0, 1]) such that (3.27)
Z 1
Qn (f ) → f (x)dx for all f ∈ X0 .
0
R1
Note that, simultaneously, we have replaced the exactness Qn (f ) = 0 f (x)dx for
n ≥ n0 by the more general convergence definition (3.26). The stability property
remains unchanged.
Then the previous theorem can be reformulated as follows.
40 3 Quadrature
Theorem 3.45. (a) Let {Qn } be consistent in the more general sense of (3.27)
and stable. Then {Qn } is convergent in the sense of (3.26) and, furthermore,
R1
limn→∞ Qn (f ) = 0 f (x)dx is the desired value of the integral.
(b) Let {Qn } be convergent in the sense of (3.26). Then {Qn } is also stable.
(c) Under the assumption of consistency in the sense of (3.27), stability and conver-
gence (3.26) are equivalent.
R1
Proof. (i) It suffices to show that limn→∞ Qn (f ) = 0 f (x)dx, since this implies
(3.26). Let f ∈ C([0, 1]) and ε > 0 be given. Because X0 from (3.27) is dense,
there is a g ∈ X0 with
ε
kf − gk∞ ≤ (Cstab : stability constant).
2 (1 + Cstab )
R1
According to (3.27), there is an n0 such that Qn (g) − 0 g(x)dx ≤ 2ε for all

n ≥ n0 . The triangle inequality yields the desired estimate
Z 1

Qn (f ) − f (x)dx

0
Z 1 Z 1 Z 1

≤ |Qn (f ) − Qn (g)| + Qn (g) − g(x)dx + g(x)dx − f (x)dx
0 0 0
ε
≤ Cstab kf − gk∞ + + kf − gk∞ ≤ ε.
2 kf −gk∞ ≤ε/[2(1+Cstab )]
(ii) Convergence in the sense of (3.26) guarantees that {Qn (f )} has a limit for
all f ∈ C([0, 1]). Alternative (a) of Corollary 3.39 applies and yields
sup kQn k = sup Cn < ∞

n∈N0 n∈N0
proving stability.
(iii) Part (c) follows from Parts (a) and (b). t
u
Finally, we give a possible application of generalised consistency. To avoid
the difficulties arising from the instability of the Newton–Cotes formulae, one
often uses compound Newton–Cotes formulae. The best known example is the
compound trapezoidal rule, which uses the trapezoidal rule on each subinterval
[i/n, (i + 1) /n]. The compound trapezoidal rule defines again a family {Qn }. It
is not consistent in the sense of Definition 3.5, since except for constant and linear
functions, no further polynomials are integrated exactly. Instead, we return to the
formulation (3.8) of the quadrature error. The well-known estimate states that
Z 1
1
kf 00 k∞ → 0

Qn (f ) − f (x)dx ≤ (3.28)
12n 2
0
for all f ∈ C 2 ([0, 1]) (cf. [9, §3.1]). The subset C 2 ([0, 1]) is dense in C([0, 1])
(simplest proof: C 2 ([0, 1]) ⊃ {polynomials} and the latter set is already dense
3.5 Further Remarks 41
according to Theorem 3.28). Hence the compound trapezoidal rule {Qn } satisfies
the consistency condition (3.27) with X0 = C 2 ([0, 1]). The stability of {Qn }
R 1 are positive and Qn (1) = 1.
follows from Conclusion 3.19a, since all weights
From Theorem 3.45a we conclude that Qn (f ) → 0 f (x)dx for all continuous f .
The trapezoidal rule is the Newton–Cotes method for n = 1. We may fix any
nNC ∈ N and use the corresponding Newton–Cotes formula in each subinterval
[i/n, (i + 1) /n]. Again, this compound formula is stable, where the stability
constant is given by CnNC from (3.25).
3.5 Further Remarks
3.5.1 General Intervals and Product Quadrature
The restriction of the integral to [0, 1] is a kind of normalisation. If quadrature is

needed over an interval [a, b] of length L = b − a, use the affine mapping
φ : [0, 1] → [a, b] defined by φ(x) = a + xL.

Rb R1
For g ∈ C([a, b]), we use a g(t)dt = 0 f (x)dx with f (x) := Lg(φ(x)) and
apply the quadrature Qn from (3.1) to f :
Z b n
X
g(t)dt ≈ Qn (Lg(φ(·))) = L ai,n g(φ(xi,n )).
a i=0
Obviously, expressed in g evaluations, we obtain a new quadrature on [a, b] by

Z b n
X
g(t)dt ≈ Q[a,b]
n (g) := L ai,n g(ti,n ) with ti,n := φ(xi,n ) = a + xi,n L.
a i=0
Also the error estimate can be transferred from [0, 1] to a general interval [a, b].
Assume an error estimate (3.8) for f ∈ C kn ([0, 1]) by
Z 1
≤ cn kf (kn ) k∞ .

f (x)dx − Qn (f )
0
The transformation from above shows immediately that

Z
b
g(t)dt − Q[a,b]
n (g) ≤ cn L
kn +1
kg (kn ) k∞ .

a
The stability constant Cn is the minimal cn for kn = 0. One sees that Cn in [0, 1]
[a,b]
becomes Cn := LCn in [a, b]. This fact can be interpreted in the way that the
42 3 Quadrature
Rb [a,b]
relative quadrature error L1 a g(t)dt − Qn (g) possesses an unchanged stability
[a,b]
constant. Anyway, the stability properties of {Qn } and {Qn } are the same.
In applications it happens that the integrand is a product f (x)g(x), where one
factor—say g—is not well-suited for quadrature (it may be insufficiently smooth,
e.g., containingPa weak singularity or it may be highly oscillatory). Interpolation of
n
f by In (f ) = i=0 f (xi,n )Φi,n (x) (cf. §3.1.2) induces a quadrature of f g by
Z 1 n
X Z 1
f (x)g(x)dx ≈ ai,n f (xi,n ) with ai,n := Φi,n (x)g(x)dx,
0 i=0 0
R1
which requires that we have precomputed the (exact) integrals 0
Φi,n (x)g(x)dx.
3.5.2 Consistency Versus Stability
As we shall see consistency is often restricted by stability requirements (cf. Re-

mark 4.15, §5.5.6, §6.6). In this respect, quadrature is an exceptional case. Gauss
quadrature Qn is optimal with respect to stability (its stability constant Cstab has
the smallest possible value 1) and it possesses the largest possible consistency order
2n + 1.
Another astonishing observation is that the Gauss quadrature is stable, although
it is an interpolatory quadrature based on the unstable polynomial interpolation
(cf. §4.5).
3.5.3 Perturbations
In this subsection we consider perturbations of f as well as of Qn . Instead of a fixed

function f , consider a sequence fn → f in C([0, 1]) with the intention of replacing
Qn (f ) by Qn (fn ).
A possible application of this setting may be as follows. Let tn → ∞ be a
sequence of natural numbers and fn the computer realisation of f by a floating-
point arithmetic with mantissa length tn , so that kfn − f k∞ ≤ C2−tn . Then
Qn (fn ) means that parallel to the increase of the number of quadrature points the
arithmetical precision also improves.
The following theorem states that such a perturbation does not destroy conver-
gence.
Theorem 3.46. Let the family {Qn : n ∈ N0 } of quadrature formulae be consistent

and stable. Furthermore, assume fn → f for a sequence of fn ∈ C([0, 1]). Then
R1
also Qn (fn ) converges to 0 f (x)dx.
3.5 Further Remarks 43
R1
Proof. The previous considerations guarantee Qn (f ) → 0 f (x)dx. Thanks to
stability, the perturbation |Qn (fn ) − Qn (f )| is bounded by Cstab kfn − f k∞ → 0
and a zero sequence. t u
Similarly, we may perturb Qn . For instance, assume that the exact weights ai,n
are replaced by the truncation ag
i,n to mantissa length tn . Then Qn = Qn + δQn
f
holds, where |δQn (f )| ≤ C2−tn kf k∞ . For tn → ∞ we obtain the norm conver-
gence kδQn k → 0, where
kδQn k := sup{|δQn (f )| : f ∈ C([0, 1]) with kf k∞ = 1}.
Theorem 3.47. Let the family {Qn : n ∈ N0 } of quadrature formulae be consis-

tent and stable. Furthermore, assume kδQn k → 0 for a sequence of perturbations
R1
δQn . Then also (Qn + δQn ) (f ) converges to 0 f (x)dx.
R1
Proof. By definition δQn (f ) → 0, while Qn (f ) → 0 f (x)dx. tu
R1
Combining both theorems we even get (Qn + δQn ) (fn ) → 0 f (x)dx.
3.5.4 Arbitrary Slow Convergence Versus Quantitative Convergence
While estimates as in (3.28) describe the convergence in a quantitative form, the

R1
previous statement Qn (f ) → 0 f (x)dx says nothing about the speed of conver-
gence. Such non-quantitative convergence statements are not very helpful in numer-
ical applications. If one does not know whether an error ε = 0.01 has already been
10
obtained for n = 5, or for n = 106 , or even only for n = 1010 , one cannot rely on
such numerical methods.
Exercise 3.48. Given a family of quadrature formulae and any number N , construct
a continuous function f with the properties
Z 1
kf k∞ = 1, f (x)dx ≥ 1/2, but Qn (f ) = 0 for all n ≤ N.
0
R1
Can something be stated about the convergence speed of Qn (f ) → 0 f (x)dx
for a general f ∈ C([0, 1])? A quantified version of convergence can be described
in two equivalent ways. Either we prescribe an ε > 0 and ask for an n(ε) such
R1
that Qm (f ) − 0 f (x)dx ≤ ε kf k∞ for m ≥ n(ε). Or there is a monotone zero
R1
sequence εn such that Qn (f ) − 0 f (x)dx ≤ εn kf k∞ .
R1
Remark 3.9 yields a negative result: if Qn (f ) − 0 f (x)dx ≤ εn kf k∞ holds
for all f ∈ C([0, 1]), necessarily εn ≥ 1 must hold excluding any zero sequence εn .
R1
Consequently, the convergence Qn (f ) → 0 f (x)dx can be arbitrarily slow.
The quantitative convergence result of inequality (3.28) holds for f ∈ C 2 ([0, 1]);
R1
i.e., for smoother functions. In fact, we get a similar result for Qn (f )− 0 f (x)dx,
44 3 Quadrature
if we consider, say f ∈ C 1 ([0, 1]), instead of f ∈ C([0, 1]). Note that f ∈ C 1 ([0, 1])
comes with the norm kf kC 1 ([0,1]) = max{|f (x)| , |f 0 (x)| : 0 ≤ x ≤ 1}.
The next result is prepared by the following lemma. We remark that a subset B of
a Banach space is precompact if and only if the closure B̄ is compact, which means
that all sequences {fk } ⊂ B possess a convergent subsequence: limn→∞ fkn ∈ B̄.
The term ‘precompact’ is synonymous with ‘relatively compact’.
Lemma 3.49. Let M ⊂ X be a precompact subset of the Banach space X. Let the
operators An ∈ L(X, Y ) be point-wise convergent to A ∈ L(X, Y ) (i.e., An x → Ax
for all x ∈ X) . Then the sequences {An x} converge uniformly for all x ∈ M ; i.e.,
sup kAn x − AxkY → 0 for n → ∞. (3.29)
x∈M
Proof. (i) C := sup{kAn k : n ∈ N} is finite (‘stability’, cf. Corollary 3.39).

Furthermore, kAk ≤ C is a simple conclusion.
(ii) We disprove the negation of (3.29). Assume that there are an ε > 0 and
a subsequence N0 ⊂ N such that supx∈M kAn x − AxkZ ≥ ε for all n ∈ N0 .
Therefore, some xn ∈ M exists with
kAn xn − Axn kZ ≥ ε/2 for all n ∈ N0 .
Since M is precompact, there is a further subsequence N00 ⊂ N0 , so that the limit

limn∈N00 xn =: ξ ∈ M exists. Choose n ∈ N00 with kxn − ξkY ≤ ε/ (8C) and
kAn ξ − AξkZ < ε/4. For this n we obtain
kAn xn − Axn kZ ≤ k(An − A) (xn − ξ)kZ + k(An − A) ξk

< (kAn k + kAk) kyn − ξkY + ε/4 ≤ ε/2
| {z }
≤2C
in contradiction to the previous inequality. t

u
Theorem 3.50. There is a zero sequence εn → 0 such that 9
Z 1
f (x)dx ≤ εn kf kC 1 ([0,1]) for all f ∈ C 1 ([0, 1]).

Qn (f ) − (3.30)

0
Proof. Let X = C([0, 1]). The subset M := {f ∈ C 1 ([0, 1]) : kf kC 1 ([0,1]) } ⊂ X

is precompact due to the theorem of Arzelà–Ascoli recalled below. Apply Lemma
R1
3.49 with Y = R, An = Qn , and A(f ) = 0 f (x)dx (note that L(X, R) = X ∗ ).
Set
Z 1

εn := sup Qn (f ) − f (x)dx
f ∈M 0
A simple scaling argument shows that (3.30) holds. Furthermore, Lemma 3.49 states
that εn → 0. t u
9
We may replace C 1 ([0, 1]) in Theorem 3.50 by the Hölder space C δ ([0, 1]) for any exponent
δ > 0. Its norm is kf kC δ ([0,1]) = max{kf k∞ , supx6=y |f (x) − f (y)| / |x − y|δ }.
References 45
Theorem 3.51 (Arzelà–Ascoli). Let D be compact and M ⊂ C(D). Suppose that

M is uniformly bounded:
sup{kf kC(D) : f ∈ M },
and equicontinuous; i.e., for any ε > 0 and x ∈ D, there is some δ such that
|f (x) − f (y)| ≤ ε for all f ∈ M, and all x, y ∈ D with |x − y| ≤ δ.
Then M is precompact.
For a proof see Yosida [12, 1,§III.3].
References
1. Bulirsch, R.: Bemerkungen zur Romberg-Integration. Numer. Math. 6, 6–16 (1964)

2. Christoffel, E.B.: Über die Gaußische Quadratur und eine Verallgemeinerung derselben. J.
Reine Angew. Math. 55, 61–82 (1858)
3. Davis, P.J., Rabinowitz, P.: Methods of Numerical Integration, 2nd ed. Academic Press, New
York (1984)
4. Gauss, C.F.: Methodus nova integralium valores per approximationem inveniendi. In: Werke,
vol. 3, pp. 163–196. K. Gesellschaft Wissenschaft, Göttingen (1876). (reprint by Georg Olms,
Hildesheim, 1981)
5. Natanson, I.P.: Konstruktive Funktionentheorie. Akademie-Verlag, Berlin (1955)
6. Ouspensky, J.V.: Sur les valeurs asymptotiques des coefficients de Cotes. Bull. Amer. Math.
Soc. 31, 145–156 (1925)
8. Romberg, W.: Vereinfachte numerische Quadratur. Norske Vid. Selsk. Forh. [Proceedings of
the Royal Norwegian Society of Sciences and Letters], Trondheim 28, 30–36 (1955)
10. Stroud, A., Secrest, D.: Gaussian Quadrature Formulas. Prentice-Hall, Englewood Cliffs
(1966)
11. Whiteside, D.T. (ed.): The mathematical papers of Isaac Newton, Vol. 4. Cambridge University
Press, Cambridge (1971)
12. Yosida, K.: Functional Analysis. Springer, New York (1968)
(2004)
Chapter 4
Interpolation
Interpolation by polynomials is a field in which stability issues have been addressed

quite early. Section 4.5 will list a number of classical results.
4.1 Interpolation Problem
The usual linear 1 interpolation problem is characterised by a subspace Vn of the

Banach space C([0, 1]) (with norm k·k∞ ; cf. §3.4.7.1) and a set
{xi,n ∈ [0, 1] : 0 ≤ i ≤ n}
of n + 1 different2 interpolation points, also called nodal points. Given a tuple

{yi : 0 ≤ i ≤ n} of ‘function values’, an interpolant Φ ∈ Vn with the property
Φ(xi,n ) = yi (0 ≤ i ≤ n) (4.1)
has to be determined.
Exercise 4.1. (a) The interpolation problem is solvable for all tuple {yi : 0 ≤ i ≤ n},
if and only if the linear space
n
Vn := (Φ(xi,n ))i=0 ∈ Rn+1 : Φ ∈ Vn

has dimension n + 1.
(b) If dim Vn = n + 1, the interpolation problem is uniquely solvable.
The interpolation problem (4.1) can be reduced to a system of n + 1 linear equa-

tions. As is well known, there are two alternatives for linear systems:
1
The term ‘linear’ refers to the underlying linear space Vn , not to linear functions.
2
In the case of the more general Hermite interpolation, a p-fold interpolation point ξ corresponds
to prescribed values of the derivatives f (m) (ξ) for 0 ≤ m ≤ p − 1.

48 4 Interpolation
(a) either the interpolation problem is uniquely solvable for arbitrary values yi or
(b) the interpolant either does not exist for certain yi or is not unique.
The polynomial interpolation is characterised by
Vn = {polynomials of degree ≤ n}
and is always solvable. In the case of general vector spaces Vn , we always assume
that the interpolation problem is uniquely solvable.
For the special values yi = δij (j fixed, δij Kronecker symbol), one obtains an
interpolant Φj,n ∈ Vn , which we call the j-th Lagrange function (analogous to the
Lagrange polynomials in the special case of polynomial interpolation).
Exercise 4.2. (a) The interpolant for arbitrary yi (0 ≤ i ≤ n) is given by

n
X
Φ= yi Φi,n ∈ Vn . (4.2)
i=0
(b) In the case of polynomial interpolation, the Lagrange polynomial is defined by

Y x − xj
Li,n (x) := Φi,n (x) := . (4.3)
xi − xj
j∈{0,...n}\{i}
For continuous functions f we define

n
X
In (f ) := f (xi,n )Φi,n (4.4)
i=0
as interpolant of yi = f (xi,n ). Hence
In : C([0, 1]) → C([0, 1])
is a linear mapping from the continuous functions into itself.
Exercise 4.3. (a) The interpolation In : X = C([0, 1]) → C([0, 1]) is continuous
and linear; i.e., In ∈ L(X, X).
(b) In is a projection; i.e., In In = In .
The terms ‘convergence’, ‘consistency’ and ‘stability’ of the previous chapter

can easily be adapted to the interpolation problem. Note that we have not only one
interpolation In , but a family {In : n ∈ N0 } of interpolations.
The interval [0, 1] is chosen without loss of generality. The following results
can immediately be transferred to a general interval [a, b] by means of the affine
mapping φ(t) = (t − a)/(b − a). The Lagrange functions Φi,n ∈ C([0, 1]) become
Φ̂i,n := Φi,n ◦ φ ∈ C([a, b]). Note that in the case of polynomials, Φi,n and Φ̂i,n
have the same polynomial degree n. The norms kIn k and the stability constant Cstab
from §4.3 do not change! Also the error estimate (4.8) remains valid.
4.3 Stability 49
Another subject are interpolations on higher-dimensional domains D ⊂ Rd . The

general concept is still true, but the concrete one-dimensional interpolation methods
do not necessarily have a counterpart in d dimensions. An exception are domains
which are Cartesian products. Then one can apply the tensor product interpolation
discussed in §4.7.
4.2 Convergence and Consistency
Definition 4.4 (convergence). A family {In : n ∈ N0 } of interpolations is called

convergent if
lim In (f ) exists for all f ∈ C([0, 1]).
n→∞
Of course, we intend that In (f ) → f , but here convergence can be defined

without fixing the limit, since lim In (f ) = f will come for free due to consistency.
Concerning consistency, we follow the model of (3.27).
Definition 4.5 (consistency). A family {In : n ∈ N0 } of interpolations is called

consistent if there is a dense subset X0 ⊂ C([0, 1]) such that
In (g) → g for all g ∈ X0 .
Exercise 4.6. Let {In } be the interpolation by polynomials of degree ≤ n. Show

that a possible choice of the dense set in Definition 4.5 is X0 := {polynomials}.
4.3 Stability
First, we characterise the operator norm kIn k (cf. (3.23)).

Pn
Lemma 4.7. kIn k = k i=0 |Φi,n (·)|k∞ holds with Φi,n from (4.4).
Pn
Proof. (i) Set Cn := k i=0 |Φi,n (·)|k∞ . For arbitrary f ∈ C([0, 1]) we conclude
that

X n X n Xn
|In (f )(x)| = f (xi,n )Φi,n (x) ≤ |f (xi,n )| |Φi,n (x)| ≤ kf k∞ |Φi,n (x)|

| {z }
i=0 i=0 i=0
≤kf k∞
≤ kf k∞ Cn .
Since this estimate holds for all x ∈ [0, 1], it follows that kIn (f )k ≤ Cn kf k∞ .
Because f is arbitrary, kIn k ≤ Cn is proved.
Pn Pn
(ii) Let the function i=0 |Φi,n (·)| be maximal at x0 : i=0 |Φi,n (x0 )| = Cn .
Choose f ∈ C([0, 1]) with kf k∞ = 1 and f (xi,n ) = sign(Φi,n (x0 )). Then
50 4 Interpolation

Xn X n
|In (f )(x0 )| = f (xi,n )Φi,n (x0 ) = |Φi,n (x0 )| = Cn = Cn kf k∞

i=0 i=0
holds; i.e., kIn (f )k∞ = Cn kf k∞ for this f . Hence the operator norm
kIn k = sup {kIn (f )k∞ / kf k∞ : f ∈ C([0, 1])\{0}}
is bounded from below by kIn k ≥ Cn . Together with (i), the equality kIn k = Cn is
proved. tu
Again, stability expresses the boundedness of the sequence of norms kIn k.
Definition 4.8 (stability). A family {In : n ∈ N0 } of interpolations is called stable if

X n
Cstab := sup kIn k < ∞ for kIn k = |Φi,n (·)| . (4.5)

n∈N0
i=0

∞
In the context of interpolation, the stability constant Cstab is called Lebesgue

constant.
Polynomial interpolation is a particular way to approximate a continuous func-
tion by a polynomial. Note that the more general approximation due to Weierstrass
is convergent. The relation between the best possible polynomial approximation and
the polynomial interpolation is considered next.
Remark 4.9. Given f ∈ C([0, 1]), let p∗n be the best approximation to f by a poly-
nomial 3 of degree ≤ n, while pn is its interpolant. Then the following estimate
holds:
kf − pn k ≤ (1 + Cn ) kf − p∗n k with Cn = kIn k . (4.6)
Proof. Any polynomial of degree ≤ n is reproduced by interpolation, in particular,

In p∗n = p∗n . Hence,
f − pn = f − In f = f − [In (f − p∗n ) + In p∗n ] = f − p∗n + In (f − p∗n )
can be estimated as claimed above. t

u
Note that by the Weierstrass approximation theorem 3.28,
kf − p∗n k → 0
holds. An obvious conclusion from (4.6) is the following: If stability would hold
(i.e., Cn ≤ Cstab ), also kf − pn k → 0 follows. Instead, we shall show instabil-
ity, and the asymptotic behaviour on the right-hand side in (4.6) depends on which
process is faster: kf − p∗n k → 0 or Cn → ∞.
3
The space of polynomials can be replaced by any other interpolation subspace Vn .
4.5 Instability of Polynomial Interpolation 51
4.4 Equivalence Theorem
Following the scheme (3.22), we obtain the next statement.
Theorem 4.10 (convergence theorem). Assume that the family {In : n ∈ N0 } of

interpolations is consistent and stable. Then it is also convergent, and furthermore,
In (f ) → f holds.
Proof. Let f ∈ C([0, 1]) and ε > 0 be given. There is some g ∈ X0 with
ε
kf − gk∞ ≤ ,
2 (1 + Cstab )
where Cstab is the stability constant. According to Definition 4.5, there is an n0 such
that kIn (g) − gk∞ ≤ 2ε for all n ≥ n0 . The triangle inequality yields the desired
estimate:
kIn (f ) − f k∞ ≤ kIn (f ) − In (g)k∞ + kIn (g) − gk∞ + kg − f k∞

ε
≤ Cstab kf − gk∞ + + kf − gk∞ ≤ ε. t
u
2 kf −gk∞ ≤ε/[2(1+Cstab )]
Again, the stability condition turns out to be necessary.
Lemma 4.11. A convergent family {In : n ∈ N0 } of interpolations is stable.
Proof. Since {In (f )} converges, the In are uniformly bounded. Apply Corollary
3.39 with X = Y = C([0, 1]) and Tn := In ∈ L(X; Y ). t u
Theorem 4.10 and Lemma 4.11 yield the following equivalence theorem.
Theorem 4.12. Let the family {In : n ∈ N0 } of interpolations be consistent. Then

convergence and stability are equivalent.
4.5 Instability of Polynomial Interpolation
We choose the equidistant interpolation points xi,n = i/n and restrict ourselves
to even n. The Lagrange polynomial L n2 ,n is particularly large in the subinterval
(0, 1/n). In its midpoint we observe the value
n 1 j n 1
2n − n − j
1
Y Y 2
L n ,n = =
2n 1 j n
2
j=0 2
−n
j=0 2
−j
j6= n
2 j6= n
2
1 1 3 n
− 32 × n2 + 12 × . . . × n − 12

2 × 2 × 2 × ... × 2
= n 2 .
2 !
Exercise 4.13. Show that the expression from above diverges exponentially.
52 4 Interpolation
Pn 1

Because of Cn = k i=0 |Li,n (·)|k∞ ≥ kL n2 ,n k∞ ≥ L n2 ,n ( 2n ) , interpolation
(at equidistant interpolation points) cannot be stable. The true behaviour of Cn has
first4 been described by Turetskii [21]:
2n+1
Cn ≈ .
e n log n
The asymptotic is improved by Schönhage [18, Satz 2] to5
Cn ≈ 2n+1 /[e n (γ + log n)],
where γ is Euler’s constant.6 Even more asymptotic terms are determined in [11].
One may ask whether the situation improves for another choice of interpolation
points. In fact, an asymptotically optimal choice are the so-called Chebyshev points:
1
xi,n = 1 + cos i+1/2
n+1 π
2
(these are the zeros of the Chebyshev polynomial7 Tn+1 ◦ φ, where φ(ξ) = 2ξ + 1 is
the affine transformation from [0, 1] onto [−1, 1]). In this case, one can prove that8
2
kIn k ≤ 1 + log(n + 1) (4.7)
π
(cf. Rivlin [17, Theorem 1.2]), which is asymptotically the best bound, as the next
result shows.
Theorem 4.14. There is some c > 0 such that

2
kIn k > log(n + 1) − c
π
holds for any choice of interpolation points.
In 1914, Faber [6] proved

1
kIn k > log(n + 1),
12
while, in 1931, Bernstein [1] showed the asymptotic estimate
4
For historical comments see [20].
5
Pn
The function ϕ = i=0 |Li,n (·)| attains its maximum Cn in the first and last interval. As
pointed out by Schönhage [18, §4], ϕ is of similar size as in (4.7) for the middle interval.
6
The value γ = 0.5772 . . . is already given in Euler’s first article [5]. Later, Euler computed 15
exact decimals places of γ.
7
The Chebyshev polynomial Tn (x) := cos(n arccos(x)), n ∈ N0 , satisfies the three-term
recursion Tn+1 (x) = 2xTn (x) − Tn−1 (x) (n ≥ 1), starting from T0 (x) = 1 and T1 (x) = x.
2 2 8
8

A lower bound is kIn k > π log(n + 1) + π γ + log π = limn→∞ kIn k , where
2 8

π
γ + log π = 0.962 52 . . .
4.6 Is Stability Important for Practical Computations? 53
2−ε
kIn k > log(n + 1) for all ε > 0.
π
The estimate of Theorem 4.14 originates from Erdös [4]. The bound
1
kIn k > √ log(n + 1)
8 π
can be found in Natanson [12, p. 370f].

The idea of the proof is as follows. Given xi,n ∈ [0, 1], 0 ≤ i ≤ n, construct
a polynomial P of degree ≤ n (concrete construction, e.g., in [12, p. 370f], [13])
such that |P (xi,n )| ≤ 1, but P (ξ) > Mn for at least one point ξ ∈ [0, 1]. Since the
interpolation of P is exact, i.e., In (P ) = P , the evaluation at ξ yields

X n Xn
kIn k = |Li,n (·)| ≥ |Li,n (ξ)|

i=0 ∞ i=0
n
n
X X
≥ |P (xi,n )Li,n (ξ)| ≥ P (xi,n )Li,n (ξ) = |P (ξ)| > Mn ,

i=0 i=0
proving kIn k > Mn .

We conclude that any sequence of polynomial interpolations In is unstable.
4.6 Is Stability Important for Practical Computations?
Does the instability of polynomial interpolation mean that one should avoid polyno-
mial interpolation altogether? Practically, one may be interested in an interpolation
In∗ for a fixed n∗ . In this case, the theoretically correct answer is: the property of
In∗ has nothing to do with convergence and stability of {In }n∈N . The reason is that
convergence and stability are asymptotic properties of the sequence {In }n∈N and
are in no way related to the properties of a particular member In∗ of the sequence.
One can construct two different sequences {In0 }n∈N and {In00 }n∈N —one stable, the
other unstable—such that In0 ∗ = In00∗ belongs to both sequences. This argument also
holds for the quadrature discussed in the previous chapter.
On the other hand, we may expect that instability expressed by Cn → ∞ may
lead to large values of Cn , unless n is very small. We return to this aspect later.
The convergence statement from Definition 4.4 is, in practice, of no help. The
reason is that the convergence from Definition 4.4 can be arbitrarily slow, so that
for a fixed n, it yields no hint concerning the error In (f ) − f . Reasonable error
estimates can only be given if f has a certain smoothness, e.g., f ∈ C n+1 ([0, 1]).
Then the standard error estimate of polynomial interpolation states that
1
kf − In (f )k∞ ≤ Cω (In ) f (n+1) , (4.8)

(n + 1)! ∞
54 4 Interpolation
where
n
Y
Cω (In ) := kωk∞ for ω(x) := (x − xi,n )
i=0
(cf. [14, §1.5], [19], [15, §8.1.1], [8, §B.3]). The quantity Cω (In ) depends on the
location of the interpolation points. It is minimal for the Chebyshev points, where
Cω (In ) = 4−(n+1) .
In spite of the instability of polynomial interpolation, we conclude from estimate

(4.8) that convergence holds, provided that kf (n+1) k∞ does not grow too much
as n → ∞ (of course, this requires that f be analytic). However, in this analysis
we have overlooked the numerical rounding errors of the input data.9 When we
evaluate the function values f (xi,n ), a perturbed result f (xi,n ) + δi,n is returned
with anP error |δi,n | ≤ η kf k∞ . Therefore, the true interpolant is In (f ) + δIn with
n
δIn = i=0 δi,n Φi,n . An estimate of δIn is given by ηkIn k kf k∞ . This yields the
error estimate
kf − In (f ) − δIn k∞ ≤ εint
n + εn
per
with
1
εint
n = Cω (In )kf (n+1) k∞ and εper
n = η kIn k kf k∞ .
(n + 1)!
Since η is small (maybe of the size of machine precision), the contribution εintn is
not seen in the beginning. However, with increasing n, the part εintn is assumed to
tend to zero, while εper
n increases to infinity because of the instability of In .
We illustrate this situation in two different scenarios. In both cases we assume
that the analytic function f is such that the exact interpolation error (4.8) decays
−n
like εint
n =e .
(1) Assume a perturbation error εpern = ηen due to an exponential increase of
kIn k. The resulting error is
e−n + ηen .
Regarding n as a real variable, we find a minimum at n = 12 log η1 with the value
√
2 η. Hence, we cannot achieve better accuracy than half the mantissa length.
2
(2) According to (4.7), we assume that εint
n = η(1 + π log(n + 1)), so that the
sum
2
e−n + η(1 + log(n + 1))
π
is the total error. Here, minimising n is the solution to the fixed-point equation
n = log(n + 1) − log(2η/π). For η = 10−16 the minimal value 3.4η of the total
error is taken at the integer value n = 41. The precision corresponds to almost the
full mantissa length. Hence, in this case the instability kIn k → ∞ is completely
harmless.10
9
There are further rounding errors, which we ignore to simplify the analysis.
10
To construct an example, where even for (4.7) the instability becomes obvious, one has to assume
that the interpolation error decreases very slowly like εint
n = 1/ log(n).
4.7 Tensor Product Interpolation 55
4.7 Tensor Product Interpolation
Finally, we give an example where the norm kIn k is required for the analysis of
the interpolation error, even if we ignore input errors and rounding errors. Consider
the function f (x, y) in two variables (x, y) ∈ [0, 1] × [0, 1]. The two-dimensional
polynomial interpolation can easily be constructed from the previous In . The tensor
product11 In2 := In ⊗ In can be applied as follows. First, we apply the interpolation
with respect to x. For any y ∈ [0, 1] we have
n
X
F (x, y) := In (f (·, y))(x) = f (xi,n , y) Φi,n (x).
i=0
In a second step, we apply In with respect to y:

n X
X n
In2 (f )(x, y) = In (F (x, ·)) (y) = f (xi,n , xj,n ) Φi,n (x) Φj,n (y).
i=0 j=0
Inequality (4.8) yields a first error12

n+1
1 ∂
|f (x, y) − F (x, y)| ≤ Cω (In ) f for all x, y ∈ [0, 1].
(n + 1)! ∂xn+1 ∞
The second one is

n
X
F (x, y) − In2 (f )(x, y) = |f (xi,n , y) − In (f (xi,n , ·)(y)| Φi,n (x).
i=0
Again
n+1
1 ∂
|f (xi,n , y) − In (f (xi,n , ·)(y)| ≤ Cω (In ) n+1
f
(n + 1)! ∂y
∞
holds and leads us to the estimate

n+1
F − In2 (f ) ≤ kIn k
1 ∂
∞
Cω (In ) n+1
f
.
(n + 1)! ∂y ∞
The previous estimates and the triangle inequality yield the final estimate
∂ n+1 ∂ n+1

f − In2 (f ) ≤
1
∞
C (I
ω n ) kI n k k f k∞ + k f k ∞ .
(n + 1)! ∂y n+1 ∂xn+1
1
Note that the divergence of kIn k can be compensated by (n+1)! .
11
Concerning the tensor notation see [9].
12
Here, k·k∞ is the maximum norm over [0, 1]2 .
56 4 Interpolation
4.8 Stability of Piecewise Polynomial Interpolation
One possibility to obtain stable interpolations is by constructing a piecewise poly-

nomial interpolation. Here, the degree of the piecewise polynomials is fixed, while
the size of the subintervals approaches zero as n → ∞. Let J = [0, 1] be the
underlying interval. The subdivision is defined by ∆n := {x0 , x1 , . . . , xn } ⊂ J
containing points satisfying
0 = x0 < x1 < . . . < xn−1 < xn = 1.
This defines the subintervals Jk := [xk−1 , xk ] of length hk := xk − xk−1 and

δn := max1≤k≤n hk . In principle, all quantities xk , Jk , hk should carry an
additional index n, since each subdivision of the sequence (∆n )n∈N has different
(n)
xk = xk . For the sake of simplicity we omit this index, except for the grid size
δn , which has to satisfy δn → 0.
Among the class of piecewise polynomial interpolations, we can distinguish two
types depending on the support13 of the Lagrange functions Φj,n . In case of Type I,
Φj,n has a local support, whereas supp(Φj,n ) = J for Type II. The precise definition
of a local support is: there are α, β ∈ N0 independent of n such that
min{n,j+β}
[
supp(Φj,n ) ⊂ Jk . (4.9)
k=max{1,j−α}
4.8.1 Case of Local Support
The simplest example is the linear interpolation where In (f )(xk ) = f (xk ) and f |Jk
(i.e., f restricted to Jk ) is a linear polynomial. The corresponding Lagrange function
Φj,n is called the hat function and has the support14 supp(Φj,n ) = Jj ∪ Jj+1 .
We may fix another polynomial degree d and fix points 0 = ξ0 < ξ1 < . . . <
ξd = 1. In each subinterval Jk = [xk−1 , xk ] we define interpolation nodes ζ` :=
xk−1 + (xk − xk−1 ) ξ` . Interpolating f by a polynomial of degree d at these nodes,
we obtain In (f )|Jk . Altogether, In (f ) is a continuous15 and piecewise polynomial
function on J. Again, supp(Φj,n ) = Jj ∪ Jj+1 holds.
A larger but still local support occurs in the following construction of piecewise
cubic functions. Define In (f )|Jk by cubic interpolation at the nodes14 xk−2 , xk−1 ,
xk , xk+1 . Then the support supp(Φj,n ) = Jj−1 ∪ Jj ∪ Jj+1 ∪ Jj+2 is larger than
before.
13
The support of a function f defined on I is the closed set supp(f ) := {x ∈ I : f (x) 6= 0}.
14
The expression has to be modified for the indices 1 and n at the end points.
15
If In (f ) ∈ C 1 (I) is desired, one may use Hermite interpolation; i.e., also dIn (f )/dx = f 0 at
x = xk−1 and x = xk . This requires a degree d ≥ 3.
4.8 Stability of Piecewise Polynomial Interpolation 57
The error estimates can be performed for each subinterval separately. Transfor-
−(d+1)
mation of inequality (4.8) to Jk yields kIn (f ) − f k∞,Jk ≤ Chk kf (d+1) k∞ ,
where d is the (fixed) degree of the local interpolation polynomial. The overall
estimate is
kIn (f ) − f k∞ ≤ Cδn−(d+1) kf (d+1) k∞ → 0, (4.10)
where we use the condition δn → 0. Pn
Stability is controlled by the maximum norm of Φn := i=1 |Φi,n (·)|. For
the examples from above it is easy to verify that kΦi,n k ≤ K independently
of i and n. Fix an argument x ∈ I. The local support property (4.9) implies
Pn
that Φi,n (x) 6= 0 holds for at most α + β + 1 indices i. Hence i=1 |Φi,n (x)|
≤ C stab := (α + β + 1) K holds and implies supn kIn k ≤ C stab (cf. (4.5)).
4.8.2 Spline Interpolation as an Example for Global Support
The space Vn of the natural cubic splines is defined by
Vn = f ∈ C 2 (I) : f 00 (0) = f 00 (1) = 0, f |Jk cubic polynomial for 1 ≤ k ≤ n .

The interpolating spline function S ∈ Vn has to satisfy S(xk ) = f (xk ) for 0 ≤ k ≤ n.

We remark that S is also the minimiser of
Z
00 2 2
min |g (x)| dx: g ∈ C (J) : S(xk ) = f (xk ) for 0 ≤ k ≤ n .
J
In this case the support of a Lagrange function Φj,n , which now is called a cardinal
spline, has global support:16 supp(Φj,n ) = J. Interestingly, there is another basis
of Vn consisting of so-called B-splines Bj , whose support is local:14 supp(Bj ) =
Jj−1 ∪ Jj ∪ Jj+1 ∪ Jj+2 . Furthermore, they are non-negative and sum up to
n
X
Bj = 1. (4.11)
j=0
17
We choose an equidistantP grid; i.e., Ji = [(i − 1)h, ih] with h := 1/n. The
n
stability estimate kIn k = k i=0 |Φi,n (·)|k∞ ≤ Cstab (cf. (4.5)) is equivalent to
n
X
kSk∞ ≤ Cstab kyk∞ , where S = yi Φi,n ∈ Vn
i=0
is the spline function interpolating yi = S(xi ). In the following, we make use of the
16
Φj,n is non-negative in Jj ∪ Jj+1 and has oscillating signs in neighbouring intervals. One
can prove that the maxima of Φj,n in Jk are exponentially decreasing with |j − k|. This fact can
already be used for a stability proof.
17
For the general case compare [14, §2], [15, §8.7], [19, §2.4].
58 4 Interpolation
B-splines, which easily can be described for the equidistant case.18 The evaluation
at the grid points yields
B0 (0) = Bn (1) = 1, B0 (h) = Bn (1 − h) = 1/6,

B1 (0) = Bn−1 (1) = 0,
(4.12)
B1 (h) = Bn−1 (1 − h) = 2/3, B1 (2h) = Bn−1 (1 − 2h) = 1/6,
Bj (xj ) = 2/3, Bj (xj±1 ) = 1/6 for 2 ≤ j ≤ n − 2.
Pn
One verifies that yj := j=0 Bj (xj ) = 1. Since the constant function S = 1 ∈ Vn
is interpolating, the unique solvability of the splinePninterpolation proves (4.11).
Now we return to a general Pn spline function S = i=0 yi Φi,n . A
Prepresentation
n
by B-splines reads S = j=0 bj Bj . Note that yi = S(xi ) = j=0 bj Bj (xi ).
Inserting the values from (4.12), we obtain
 
6
1 4 1 
1 .. .. .. 

y = Ab with A =  . . . b
6 
 1 4 1
6
n n
for the vectors y = (yi )i=0 and b = (bi )i=0 . A can be written as A = 23 [I + 12 C]
−1kCk
with ∞ = 1; i.e., A is strongly diagonal dominant and the inverse satisfies
A ≤ 3 because of
∞
∞
3 1 3 X −ν ν
A−1 = [I + C]−1 = 2 C .
2 2 2 ν=0
Pn
Using b = A−1 y, we derive from S = j=0 bj Bj that

n n n
X X X
|S(x)| =
bj Bj (x) ≤
|bj | Bj (x) ≤ kbk∞ Bj (x) = kbk∞
j=0 Bj ≥0 j=0 (4.11)
j=0
for all x ∈ J, so that the stability estimate kSk∞ ≤ Cstab kyk∞ is proved with
Cstab := 3.
18
The explicit polynomials are
 

 ξ3 , ξ = x − xj−2 , x ∈ Jj−1 
1 h3 + 3h2 ξ + 3hξ 2 − 3ξ 3 , ξ = x − xj−1 , x ∈ Jj ,
 
Bj = 3 2 2 3 for 2 ≤ j ≤ n − 2,
6h3   h + 3h ξ + 3hξ − 3ξ , ξ = xj+1 − x, x ∈ Jj+1 ,  
 ξ3 , ξ = xj+2 − x, x ∈ Jj+2 , 
 
2
1  6h x − 2x3 , x ∈ J1 , 
B1 = h3 + 3h2 ξ + 3hξ 2 − 3ξ 3 , ξ = 2h − x, x ∈ J2 , , Bn−1 (x) = B1 (1 − x),
6h3  ξ 3 , ξ = 3h − x, x ∈ J3 ,

3
1 h + 3h2 ξ + 3hξ 2 − ξ 3 , ξ = h − x, x ∈ J1 ,
B0 = , Bn (x) = B0 (1 − x).
6h 3 ξ3 , ξ = 2h − x, x ∈ J2 ,
4.9 From point-wise Convergence to Operator-Norm Convergence 59
Remark 4.15. The previous results show that consistency is in conflict with stability.
Polynomial interpolation has an increasing order of consistency, but suffers from in-
stability (cf. Theorem 4.14). On the other hand, piecewise polynomial interpolation
of bounded order is stable.
4.9 From Point-wise Convergence to Operator-Norm

Convergence
As already mentioned in §3.5 in the context of quadrature, only point-wise conver-

gence In (f ) → f (f ∈ X) can be expected, but not operator-norm convergence
kIn − idk → 0. However, there are situations in which point-wise convergence can
be converted into operator-norm convergence.
An operator K : X → Y is called compact if the image B := {Kf : kf kX ≤ 1}
is precompact (cf. page 44). The following theorem is formulated for an arbitrary,
point-wise convergent sequence of operators An : Y → Z.
Theorem 4.16. Let X, Y, Z be Banach spaces, and A, An ∈ L(Y, Z). Suppose
that point-wise convergence An y → Ay holds for all y ∈ Y . Furthermore, let
K : X → Y be compact. Then the products Pn := An K converge with respect to
the operator norm to P := AK; i.e., kPn − P k → 0.
Proof. M := {Kx : kxkX ≤ 1} ⊂ Y is precompact because of the compactness
of K, so that we can apply Lemma 3.49:
kPn − P k = sup{kPn x − P xkZ : x ∈ X, kxk ≤ 1}

= sup{kAn (Kx) − A (Kx)kZ : x ∈ X, kxk ≤ 1}
= sup{kAn y − AykZ : y ∈ M } → 0. t
u
(3.29)
A typical example of a compact operator is the embedding

E : C λ ([0, 1]), k·kC λ ([0,1]) → (C([0, 1]), k·k∞ ) .
For integer λ ∈ N, C λ ([0, 1]) is the space of λ-times continuously differentiable

functions, where the norm k·kC λ ([0,1]) is the maximum of all derivatives up to order
λ. For 0 < λ < 1, C λ ([0, 1]) are the Hölder continuous functions with k·kC λ ([0,1])
explained in Footnote 9 on page 44. The embedding is the identity mapping:
E(f ) = f ; however, the argument f and the image E(f ) are associated with differ-
ent norms. As mentioned in the proof of Theorem 3.50, E ∈ L(C λ ([0, 1]), C([0, 1]))
is compact.
In the case of λ = 4, estimate (4.10) already yields the operator-norm convergence
kIn − idkC([0,1])←C λ ([0,1]) ≤ C/n4 → 0. To show a similar operator-norm con-
vergence for 0 < λ < 1, interpret In − id as (In − id)E : C λ ([0, 1]) → C([0, 1]).
Applying Theorem 4.16 with A = id, An = In , and K = E, we obtain
kIn − idkC([0,1])←C λ ([0,1]) → 0.

60 4 Interpolation
4.10 Approximation
Often, interpolation is used as a simple tool to obtain an approximation; i.e., the

interpolation condition (4.1) is not essential. Instead, we can directly ask for a best
approximation Φn ∈ Vn of f ∈ B, where Vn ⊂ B is an (n + 1)-dimensional
subspace of a Banach space B with norm k·k:
kf − Φn k = inf {kf − gk : g ∈ Vn } . (4.13)
Using compactness arguments one obtains the existence of a minimiser Φ. If the

space B is strictly convex,19 the minimiser is unique (cf. [10]).
A prominent choice of Vn are the polynomials of degree ≤ n, while B =
C([a, b]) is equipped with the maximum norm. Polynomials satisfy the following
Haar condition: any 0 6= f ∈ Vn has at most n zeros (cf. Haar [7]). As a con-
sequence, also in this case, the best approximation problem (4.13) has a unique
solution. For the numerical solution of the best approximation the following equi-
oscillation property is essential (cf. Chebyshev [2]):
Theorem 4.17. Let ε := f − Φn be the error of the best approximation in (4.13).

Then there are n + 2 points xµ with a ≤ x0 < x1 < . . . < xn+1 ≤ b such that
|ε(xµ )| = kf − Φn k and ε(xµ ) = −ε(xµ+1 ) for 0 ≤ µ ≤ n. (4.14)
The second part of (4.14) describes n + 1 = dim(Vn ) equations, which are used
by the Remez algorithm to determine Φn ∈ Vn (cf. Remez [16]).
From (4.14) one concludes that there are n zeros ξ1 < . . . < ξn of ε = f − Φn ;
i.e., Φn can be regarded as an interpolation polynomial with these interpolation
points. However note that the ξµ depend on the function f.
The mapping f 7→ Φn is in general nonlinear. Below, when we consider Hilbert
spaces, it will become a linear projection.
Since the set of polynomials is dense in C([a, b]) (cf. Theorem 3.28), the condi-
tion
[
V0 ⊂ V1 ⊂ . . . ⊂ Vn ⊂ Vn+1 ⊂ . . . and Vn = C([a, b]) (4.15)
n∈N0
is satisfied. Condition (4.15) implies
kf − Φn k & 0 as n → ∞ for Φn from (4.13). (4.16)
Stability issues do not appear in this setting. One may consider the sequence
{kΦn k : n ∈ N0 } , but (4.16) proves convergence kΦn k → kf k; i.e., the sequence
must be uniformly bounded.
The approximation is simpler if B is a Hilbert space with scalar product h·, ·i .
Then the best approximation from (4.13) is obtained by means of the orthogonal
19
B is strictly convex if kf k = kgk = 1 and f 6= g imply kf + gk < 2.
4.10 Approximation 61
projection20 Πn ∈ L(B, B) onto Vn :
Φn = Πn f.
Given any orthonormal basis {φµ : 0 ≤ µ ≤ n} of Vn , the solution has the explicit
representation
Xn
Φn = hf, φµ i φµ . (4.17)
µ=0
The standard example is the Fourier approximation R π of 2π periodic real func-

tions in [−π, π]. The L2 scalar product is hf, gi = −π f gdx. Let n be even. Vn is
spanned by the orthonormal basis functions

1 cos(mx) sin(mx)
√ , √ , √ : 1 ≤ m ≤ n/2 .
2π π π
At first glance there is no stability problem to be discussed, since the operator norm
of orthogonal projections equals one: kΠn kL2 ←L2 = 1. However, if we consider
the operator norm kΠn kB←B for another Banach space, (in)stability comes into
play.
Let Πn be the Fourier projection from above and choose the Banach space
B = C2π := {f ∈ C([−π, π]) : f (−π) = f (π)} equipped with the maximum
norm k·k∞ . We ask for the behaviour of kΠn k∞ , where now k·k∞ = k·kC2π ←C2π
denotes the operator norm. The mapping (4.17) can be reformulated by means of
the Dirichlet kernel,
1 π sin(2n + 1)y
Z
(Φn f ) (x) = [f (x + 2y) + f (x − 2y)]dy.
π 0 sin(y)
From this representation we infer that
1 π sin(2n + 1)y
Z
kΠn k∞ = dy.
π 0 sin(y)
Lower and upper bounds of this integral are

4
log(n + 1) ≤ kΠn k∞ ≤ 1 + log(2n + 1).
π2
This shows that the Fourier projection is unstable with respect to the maximum
norm. The negation of the uniform boundedness theorem 3.38 together with
kΠn k∞ → ∞ implies the well-known fact that uniform convergence Πn f → f
cannot hold for any f ∈ C2π .
The orthogonal Fourier projection Πn is the best choice for the Hilbert space
L2 ([−π, π]). For C2π one may choose another projection Pn from C2π onto Vn .
20
That means (i) Πn Πn = Πn (projection property) and (ii) Πn is selfadjoint: hΠn f, gi =
hf, Πn gi for all f, g ∈ B.
62 4 Interpolation
This, however, can only lead to larger norms kPn k∞ due to the following result of
Cheney et al. [3].
Theorem 4.18. The Fourier projection Πn is the unique minimiser of
min{kPn k∞ : Pn ∈ L(C2π , C2π ) projection onto Vn }.
References
1. Bernstein, S.N.: Sur la limitation des valeurs d’un polynôme p(x) de degré n sur tout un
segment par ses valeurs en (n + 1) points du segment. Izv. Akad. Nauk SSSR 8, 1025–1050
(1931)
2. Chebyshev, P.L.: Sur les questions de minima qui se rattachent à la représentation approxima-
tive des fonctions. In: Oeuvres, Vol. I, pp. 641–644. St. Petersburg (1899)
3. Cheney, E.W., Hobby, C.R., Morris, P.D., Schurer, F., Wulbert, D.E.: On the minimal property
of the Fourier projection. Trans. Am. Math. Soc. 143, 249–258 (1969)
4. Erdös, P.: Problems and results on the theory of interpolation. Acta Math. Acad. Sci. Hungar.
12, 235–244 (1961)
5. Euler, L.: De progressionibus harmonicis observationes. Commentarii academiae scientiarum
imperialis Petropolitanae 7, 150–161 (1740)
6. Faber, G.: Über die interpolatorische Darstellung stetiger Funktionen. Jber. d. Dt. Math.-
Verein. 23, 190–210 (1914)
7. Haar, A.: Die Minkowskische Geometrie und die Annäherung an stetige Funktionen. Math.
Ann. 78, 294–311 (1918)
8. Hackbusch, W.: Hierarchische Matrizen - Algorithmen und Analysis. Springer, Berlin (2009)
10. Meinardus, G.: Approximation of functions: theory and numerical methods. Springer, New
York (1967)
11. Mills, T.M., Smith, S.J.: The Lebesgue constant for Lagrange interpolation on equidistant
nodes. Numer. Math. pp. 111–115 (1992)
12. Natanson, I.P.: Konstruktive Funktionentheorie. Akademie-Verlag, Berlin (1955)
13. Natanson, I.P.: Constructive function theory, Vol. III. Frederick Ungar Publ., New York (1965)
14. Plato, R.: Concise Numerical Mathematics. AMS, Providence (2003)
16. Remez, E.J.: Sur un procédé convergent d’approximations successives pour déterminer les
polynômes d’approximation. Compt. Rend. Acad. Sc. 198, 2063–2065 (1934)
17. Rivlin, T.J.: Chebyshev Polynomials. Wiley, New York (1990)
18. Schönhage, A.: Fehlerfortpflanzung bei Interpolation. Numer. Math. 3, 62–71 (1961)
20. Trefethen, L.N., Weideman, J.A.C.: Two results on polynomial interpolation in equally spaced
points. J. Approx. Theory 65, 247–260 (1991)
21. Turetskii, A.H.: The bounding of polynomials prescribed at equally distributed points (Rus-
sian). Proc. Pedag. Inst. Vitebsk 3, 117–127 (1940)
Chapter 5
Ordinary Differential Equations
The numerical treatment of ordinary differential equations is a field whose scope

has broadened quite a bit over the last fifty years. In particular, a whole spectrum
of different stability conditions has developed. Since this chapter is not the place to
present all details, we concentrate on the most basic concept of stability. As a side-
product, it will lead us to the power bounded matrices, which is a class of matrices
with certain stability properties. More details about ordinary differential equations
can be found, e.g., in [20], [23], [4, 3], [5, §§5-6], and in the two volumes [8], [9].
5.1 Initial-Value Problem
5.1.1 Setting of the Problem
Let f : R × R → R be a continuous function.1 In what follows we are looking

for continuously differentiable functions y(x) satisfying the ordinary differential
equation
y 0 (x) = f (x, y(x)). (5.1a)
The initial-value problem requires finding a solution y of (5.1a) which, in addition,
satisfies
y(x0 ) = y0 (5.1b)
for a given ‘initial value’ y0 .
Usually one is looking for the solution y at x ≥ x0 , either in a finite2 interval
I := [x0 , xE ] or in the unbounded interval I := [x0 , ∞). Correspondingly, f needs
to be defined on I × R.
1
In the case of a system of differential equations, f is defined in R × Rn and the solution y ∈
C 1 (R, Rn ) is vector-valued. For our considerations it is sufficient to study the scalar case n = 1.
2
However, it may happen that the solution exists only on a smaller interval [x0 , xS ) ⊂ [x0 , xE ].

64 5 Ordinary Differential Equations
If f is only continuous, a solution of (5.1a,b) exists due to Peano’s theorem

(at least, in the neighbourhood [x0 , x0 + ε) for some ε > 0); however, there may be
more than one solution. Uniqueness is ensured by the assumption that f is Lipschitz
continuous with respect to the second argument (cf. Lipschitz [16, pp. 500ff]. For
simplicity, we formulate the global Lipschitz continuity:
|f (x, y1 ) − f (x, y2 )| ≤ L |y1 − y2 | for all x ∈ I, y1 − y2 ∈ R. (5.2)
The Lipschitz constant L will appear in later analysis. Unique solvability will be
stated in Corollary 5.10.
5.1.2 One-Step Methods
We choose a fixed step size3 h > 0. The corresponding grid points are
xi := x0 + ih (i ∈ N0 , xi ∈ I).
Next, we define approximations ηi of y(xi ). The notation ηi assumes that an under-

lying step size h is defined. If necessary, we write η(x0 + ih; h) instead of ηi . Note
that η(x; h) is defined only for grid points x = x0 + ih (i ∈ N0 ). The desired
property is η(x; h) ≈ y(x), where the error should tend to zero as h → 0.
Because of the given initial value y0 , we start the computation with
η0 := y0 . (5.3)
The prototype of the one-step methods is the Euler method, which starts with
(5.3) and defines recursively
ηi+1 := ηi + hf (xi , ηi ). (5.4)
Exercise 5.1. Consider the differential equation y 0 = ay (i.e., f (x, y) = ay) with
the initial value y0 = 1 at x0 = 0 (its exact solution is y(x) = eax ). Determine the
solution of (5.4). Does y(x) − η(x; h) → 0 hold for h := x/n > 0, when n → ∞
and h → 0 for fixed nh = x ?
For other one-step methods, one replaces the right-hand side in (5.4) by a more
general expression hφ(xi , ηi , h; f ). Here, the last argument f means that inside of φ
the function f can be used for arbitrary evaluations. The Euler method corresponds
to φ(xi , ηi , h; f ) = f (xi , ηi ).
Definition 5.2. A general explicit one-step method has the form
ηi+1 := ηi + hφ(xi , ηi , h; f ). (5.5)

3
In practical implementations one has to admit varying step widths hi = xi+1 − xi . Usually,
these are chosen adaptively.
5.1 Initial-Value Problem 65
Often, the evaluation of φ is performed in several partial steps. For instance, the
Heun method [10] uses the intermediate step ηi+1/2 :
h h
ηi+1/2 := ηi + f (xi , ηi ), ηi+1 := ηi + hf (xi + , ηi+1/2 ).
2 2
These equations yield φ(xi , ηi , h; f ) := f (xi + h2 , ηi + h2 f (xi , ηi )). The classi-
cal Runge–Kutta method uses four intermediate steps (cf. Runge [19], Kutta [15];
history in [1]; for a modern description see, e.g., [8, §II]).
5.1.3 Multistep Methods
The term ‘one-step method’ refers to the fact that (xi+1 , ηi+1 ) is determined only
by (xi , ηi ). The past values ηj , j < i, do not enter into the algorithm.
On the other hand, since besides ηi the values ηi−r , ηi−r+1 , . . . , ηi−1 are avail-
able (r is a fixed natural number), one may ask whether one can use these data. In-
deed, having more free parameters, one can try to increase the order of the method.
This leads to the r-step method, which is of the form
r−1
X
ηj+r := − αν ηj+ν + hφ(xj , ηj+r−1 , . . . , ηj , h; f ) (5.6)
ν=0
(more precisely, this is the explicit form of a multistep method) with the additional
parameters α0 , . . . , αr−1 ∈ R for j = 0, . . . , r − 1. As we shall see,
r−1
X
αν = −1 (5.7)
ν=0
describes a first consistency condition.
Remark 5.3. Because of (5.7), a multistep method (5.6) with r = 1 coincides with
the one-step method (5.5).
Remark 5.4. In the case of r ≥ 2, the multistep methods (5.6) can only be used
for the computation of ηi for i ≥ r. The computation of η1 , η2 , . . . , ηr−1 must be
defined in another way (e.g., by a one-step method).
An example of a two-step method is the midpoint formula
ηj+2 = ηj + 2hf (xj+1 , ηj+1 ); (5.8)
i.e., r = 2, α0 = −1, α1 = 0, φ(xj , ηj+1 , ηj , h; f ) = 2f (xj + h, ηj+1 ).

A rather dubious proposal is the extrapolation
ηj+2 = 2ηj+1 − ηj ; (5.9)

i.e., r = 2, α0 = −2, α1 = 1, φ(xj , ηj+1 , ηj , h; f ) = 0. Even without any

quantitative error analysis one can see that the latter two-step method cannot be
successful. Since (5.9) does not depend on f , it produces the same linear function
ηj = η0 + j (η1 − η0 ) for all differential equations.
5.2 Fixed-Point Theorem and Recursive Inequalities
We now provide some technical tools required in this chapter.
Theorem 5.5 (Banach’s fixed-point theorem). Let X be a Banach space and

Ψ : X → X a contraction mapping; i.e., there is a constant LΨ ∈ [0, 1) such
that
kΨ (x0 ) − Ψ (x00 )kX ≤ LΨ kx0 − x00 kX for all x0 , x00 ∈ X. (5.10a)
Then the fixed-point equation x = Ψ (x) has exactly one solution, and for all starting
values x0 ∈ X the fixed-point iteration xn+1 = Ψ (xn ) converges to this solution.
Furthermore, the error of the n-th iterant can be estimated by
(LΨ )n
kx − xn kX ≤ kx1 − x0 kX for all n ∈ N0 . (5.10b)
1 − LΨ
Proof. (i) Let x0 , x00 be two solutions of the fixed-point equation; i.e., x0 = Ψ (x0 )
and x00 = Ψ (x00 ). Exploiting the contraction property with L = LΨ , we get
kx0 − x00 kX = kΨ (x0 ) − Ψ (x00 )kX ≤ L kx0 − x00 kX .
From L < 1 we conclude that kx0 − x00 kX = 0; i.e., uniqueness x0 = x00 is proved.
(ii) The iterants xn of the fixed-point iteration satisfy the inequality
kxn+1 − xn kX = kΨ (xn ) − Ψ (xn−1 )kX ≤ L kxn − xn−1 kX
and thereby kxn+1 − xn kX ≤ Ln kx1 − x0 kX . For arbitrary n > m, the multiple

triangle inequality yields the estimate
n
X n
X
kxn − xm kX ≤ kxj − xj−1 kX ≤ Lj−1 kx1 − x0 kX
j=m+1 j=m+1
∞
X
j Lm
≤ L kx1 − x0 kX = kx1 − x0 kX ; (5.10c)
j=m
1−L
i.e., {xn } is a Cauchy sequence. Since X is a Banach space and therefore complete,
a limit x∗ = lim xn exists. Because of the continuity of the function Ψ , the limit in
xn+1 = Ψ (xn ) yields x∗ = Ψ (x∗ ); i.e., x∗ is a fixed-point solution.
Inequality (5.10b) follows from (5.10c) for n → ∞. t u
5.2 Fixed-Point Theorem and Recursive Inequalities 67
In the following analysis, recursive inequalities of the following form will appear:
aν+1 ≤ (1 + hL) aν + hk B for all ν ≥ 0, where L, B, h, k, a0 ≥ 0. (5.11)
The meaning of the parameters is: L Lipschitz constant, h step size, and k local
consistency order.
Lemma 5.6. Any solution of the inequalities (5.11) satisfies the estimate
 
 νh for L = 0 
aν ≤ eνhL a0 + hk−1 B · eνhL − 1 (ν ∈ N0 ).
 for L > 0 
L
Proof. We pose the following induction hypothesis:
ν−1
X µ ν
aν ≤ Aν := (1 + hL) hk B + (1 + hL) a0 . (5.12)
µ=0
The start of the induction is given by a0 ≤ A0 = a0 . For the induction step

ν 7→ ν + 1 insert aν ≤ Aν into (5.11):
Xν
µ ν+1
aν+1 ≤ (1 + hL) Aν + hk B = (1 + hL) hk B + (1 + hL) a0 + hk B
µ=1
Xν
µ ν+1
= (1 + hL) hk B + (1 + hL) a0 = Aν+1 .
µ=0
ν ν
Exercise 3.24a shows that (1 + hL) ≤ ehL = ehLν . For L > 0, the geometric
sum yields the value
ν−1 ν
X µ (1 + hL) − 1 B ν
hk B (1 + hL) = hk B = hk−1 [(1 + hL) − 1]
µ=0
(1 + hL) − 1 L
B hLν
≤ hk−1

e −1 .
L
Therefore, Aν from (5.12) can be estimated by Aν ≤ hk−1 B
hLν
L e − 1 + ehLν a0 .
The particular case L = 0 can be treated separately or obtained from L > 0 by
performing the limit L → 0. t u
Rx
Exercise 5.7. Prove that any solution ϕ of the inequality ϕ(x) ≤ ϕ0 + L x0 ϕ(t)dt
is bounded by Φ; i.e., ϕ(x) ≤ Φ(x), where Φ is the solution of the integral equation
Z x
Φ(x) = ϕ0 + L Φ(t)dt.
x0
Rx
Hint: (a) Define Ψ (Φ) by Ψ (Φ)(x) := ϕ0 + L x0 Φ(t)dt. The integral equation is
the fixed-point equation Φ = Ψ (Φ) for Φ ∈ C(I), I = [x0 , xE ]. Show that Ψ is a
contraction with respect to the norm
kψk := max{|ψ(t)| exp (−2L(t − x0 )) : t ∈ I}
with contraction number LΨ = 12 (independently of the length of the interval I,

including I = [x0 , ∞)).
(b) Apply the fixed-point iteration Φn+1 := Ψ (Φn ) with Φ0 := ϕ and show Φn ≥ ϕ
for all n ≥ 0.
Lemma 5.8. Let X and Y be Banach spaces. The operators S, T ∈ L(X, Y ) are
supposed to satisfy
T −1 ∈ L(Y, X) and kS − T kY ←X kT −1 kX←Y < 1.
Then also the inverse S −1 ∈ L(Y, X) exists and satisfies
kT −1 kX←Y
kS −1 kX←Y ≤ .
1 − kS − T kY ←X kT −1 kX←Y
Proof. Fix y ∈ X. The mapping Φ(x) := T −1 (T − S)x + T −1 y is contracting with

contraction rate q = kT −1 kX←Y kS − T kY ←X < 1. Banach’s fixed-point Theorem
5.5 states the unique solvability of Φ(x) = x, which implies that T x = T Φ(x) =
(T − S)x + y and Sx = y. Hence the inverse S −1 exists. The fixed-point iteration
with starting value x0 produces x1 = T −1 y. Estimate (5.10b) with n = 0 yields
1 1
kS −1 yk = kS −1 y − x0 k = kx − x0 k ≤ kx1 − x0 k = kT −1 yk,
1−q 1−q
which leads to the desired inequality. t
u
5.3 Well-Conditioning of the Initial-Value Problem
Before we start with the numerical solution, we should check whether the initial-
value problem, i.e., the mapping (y0 , f ) 7→ y, is well-conditioned. According to
§2.4.1.2, the amplification of a perturbation of the input data is to be investigated. In
the present case, one can perturb the initial value y0 as well as the function f . The
first case is analysed below, the second one in Theorem 5.12.
Theorem 5.9. Let y1 , y2 ∈ C 1 (I) be two solutions of the differential equation

(5.1a) with the initial values
y1 (x0 ) = y0,1 and y2 (x0 ) = y0,2 respectively.
Assume that f ∈ C(I × R) satisfies (5.2). Then the following estimate4 holds in I:
|y1 (x) − y2 (x)| ≤ |y0,1 − y0,2 | eL(x−x0 ) with L from (5.2). (5.13)
4
By definition of I in §5.1, x ≥ x0 holds. Otherwise, eL(x−x0 ) is to be replaced by eL|x−x0 | .
5.3 Well-Conditioning of the Initial-Value Problem 69
Rx
Proof. yi (x) = y0,i + x0
f (t, yi (t))dt holds for i = 1, 2, so that
Z x

|y1 (x) − y2 (x)| = y0,1 − y0,2 + [f (t, y1 (t)) − f (t, y2 (t))] dt
x
Z 0x
≤ |y0,1 − y0,2 | + |f (t, y1 (t)) − f (t, y2 (t))| dt ≤
x0 (5.2)
Z x
≤ |y0,1 − y0,2 | + L |y1 (t) − y2 (t)| dt.
x0
The function Φ(x) := |y0,1 − y0,2 | eL(x−x0 ) satisfies the equation

Z x
Φ(x) = |y0,1 − y0,2 | + LΦ(t)dt on I.
x0
Hence Exercise 5.7 proves |y1 (x) − y2 (x)| ≤ Φ(x); i.e., (5.13). This ends the proof
of Theorem 5.9. tu
Corollary 5.10. Assumption (5.2) ensures uniqueness of the solution of the initial-
value problem (5.1a,b).
Proof. If y1 , y2 ∈ C 1 (I) are two solutions, Theorem 5.9 yields
|y1 (x) − y2 (x)| ≤ |y0,1 − y0,2 | exp (L(x − x0 )) = 0,

y0,1 =y0,2
hence y1 = y2 on I. t
u
One may denote the solution of the initial-value problem (5.1a,b) by y(x; y0 )
with the initial value y0 as second argument. Then (5.13) states that y(·; ·) as well as
f (·, ·) are Lipschitz continuous with respect to the second argument. This statement
can be generalised.
Exercise 5.11. If f (·, ·) is k-times continuously differentiable with respect to the

second argument y, also the solution y(·; ·) does so with respect to y0 .
A perturbation of the right-hand side f in (5.1a) is studied next.
Theorem 5.12. Let y and ỹ be solutions of y 0 = f (x, y) and ỹ 0 = f˜(x, ỹ), respec-
tively, with coinciding initial values y(x0 ) = ỹ(x0 ) = y0 . Only f (not f˜) has to
fulfil the Lipschitz condition (5.2), while

f (x, y) − f˜(x, y) ≤ ε for all x ∈ I, y ∈ R.

Then
ε

eL(x−x0 ) − 1 if L > 0
|y(x) − ỹ(x)| ≤ L , (L from (5.2)).
ε(x − x0 ) if L = 0
Proof. Set δ(x) := |y(x) − ỹ(x)| and note that

Z x h i Z x h i
˜ f (ξ, y(ξ)) − f˜(ξ, ỹ(ξ)) dξ

δ(x) = f (ξ, y(ξ)) − f (ξ, ỹ(ξ)) dξ ≤

x x0
Z x0
= f (ξ, y(ξ)) − f (ξ, ỹ(ξ)) + f (ξ, ỹ(ξ)) − f˜(ξ, y(ξ)) dξ

x
Z x0 Z x
≤ |f (ξ, y(ξ)) − f (ξ, ỹ(ξ))| dξ + f (ξ, ỹ(ξ)) − f˜(ξ, y(ξ)) dξ

x x0
Z x0
≤ Lδ(ξ)dξ + (x − x0 )ε.
x0
Rx
A majorant of δ is the solution d of d(x) = x0 Ld(ξ)dξ + (x − x0 )ε. In the case of
L > 0, the solution is d(x) = Lε eL(x−x0 ) − 1 . t

u
5.4 Analysis of One-Step Methods
In §5.4.1 we explain why for our analysis in §§5.4.3–5.4.5 it is sufficient to study

explicit one-step methods. In §5.4.2 we discuss the Lipschitz continuity of φ.
5.4.1 Implicit Methods
Definition 5.13. A general implicit one-step method has the form
ηi+1 := ηi + hφ(xi , ηi , ηi+1 , h; f ). (5.14)
An example of (5.14) is the implicit Euler method, where
φ(xi , ηi , ηi+1 , h; f ) = f (xi + h, ηi+1 ). (5.15)
Next we assume that φ is defined for xi ∈ I, ηi , ηi+1 ∈ R, and sufficiently small

h (0 < h ≤ h0 ).
Exercise 5.14. Let φ from (5.14) be Lipschitz continuous with respect to ηi+1 :
|φ(xi , ηi , ηi+1 , h; f ) − φ(xi , ηi , η̂i+1 , h; f )| ≤ L |ηi+1 − η̂i+1 | .
Show that the fixed-point equation (5.14) is uniquely solvable, if h < 1/L.
Assuming unique solvability of (5.14), a function ηi+1 = Ψ (xi , ηi , h; f ) exists

by the implicit function theorem. Inserting ηi+1 = Ψ (xi , ηi , h; f ) into the third
argument of φ(xi , ηi , ηi+1 , h; f ), one obtains formally an explicit one-step method
(5.5) with
5.4 Analysis of One-Step Methods 71
φ̂(xi , ηi , h; f ) := φ(xi , ηi , Ψ (xi , ηi , h; f ), h; f ).

Hence, for theoretical considerations we may restrict ourselves to the explicit
case (5.5). The only additional requirement is that one has to restrict the step sizes
to 0 < h ≤ h0 with sufficiently small h0 . For the theoretical analysis the latter
condition is of no consequence, since we study the limit h → 0.
5.4.2 Lipschitz Continuity of φ
Analogous to the Lipschitz condition (5.2) we shall need Lipschitz continuity of φ:


 xi ∈ I,
|φ(xi , η 0 , h; f ) − φ(xi , η 00 , h; f )| ≤ Lφ |η 0 − η 00 | for all η 0 , η 00 ∈ R, (5.16)
h ≤ h0 .

Since φ is defined implicitly via f , the Lipschitz property of φ is inherited from the
Lipschitz continuity of f . Therefore, we have always to assume that f satisfies (5.2).
Exercise 5.15. (a) Prove for the Euler and Heun methods that (5.2) implies (5.16).
(b) According to §5.4.1, the implicit Euler method (5.15) leads to an explicit method
with φ̂(xi , ηi , h; f ). For sufficiently small h, prove Lipschitz continuity of φ̂.
5.4.3 Consistency
We now motivate the consistency condition. Replacing the discrete solution ηi of

ηi+1 := ηi +hφ(xi , ηi , h; f ) by the exact solution y(xi ) of the differential equation,
we obtain the so-called local discretisation error τ by
y(xi+1 ) = y(xi ) + h [φ(xi , y(xi ), h; f ) + τ (xi , y(xi ); h)] .
For the explicit definition of τ (ξ, η; h) fix ξ ∈ I and η ∈ R, and let Y (· ; ξ, η) be the
solution of (5.1a) with initial value condition
Y (ξ; ξ, η) = η at x = ξ (not at x = x0 ).
Then
Y (ξ + h; ξ, η) − η
τ (ξ, η; h) := − φ(ξ, η, h; f ) (5.17)
h
defines the local discretisation error at (ξ, η; h).
Obviously, we may expect that the one-step method (5.5) is better the smaller τ
is. Note that τ = 0 leads to the ideal result ηi = y(xi ).
Definition 5.16. (a) The one-step method characterised by φ is called consistent if
sup |τ (x, y(x); h)| → 0 (h → 0). (5.18)

x∈I
Here y is the solution of (5.1a,b). The argument f ∈ C(I×R) of φ must satisfy (5.2).
(b) Furthermore, φ is called consistent of order p if τ (x, y(x); h) = O(hp ) holds
uniformly for h → 0 on x ∈ I for all sufficiently smooth f .
Assuming f to be sufficiently smooth,5 one performs the Taylor expansion of
1
h [y(x + h; x, η) − η] and uses
y(x + h; x, η) = y(x; x, η) + hy 0 (x; x, η) + o(h) = η + hf (x, η) + o(h).
Hence (5.18) implies the condition φ(x, η, h; f ) → f (x, η). One easily checks that
this condition is satisfied for the methods of Euler and Heun.
However, the trivial one-step method ηi+1 := ηi (i.e., φ = 0) leads, in general,
to τ (x, η; h) = O(1) and is not consistent.
5.4.4 Convergence
We recall the notation ηi = η(xi , h). The desired property is η(x, h) ≈ y(x).
Concerning the limit h → 0, we restrict ourselves tacitly to (a subsequence of)
hn := (x − x0 ) /n, since then x = nhn belongs to the grid on which η(·, hn ) is
defined.
Definition 5.17 (convergence). A one-step method is called convergent if for all
Lipschitz continuous f and all x ∈ I
lim η(x, h) = y(x) (y solution of (5.1a,b))

h→0
holds. A one-step method has convergence order p if η(x, h) = y(x) + O(hp ) for
sufficiently smooth f.
5.4.5 Stability
Consistency controls the error generated in the i-th step from xi to xi+1 under the
assumption that ηi is the exact starting value. At the start, η0 = y0 is indeed exact,
so that according to condition (5.18) the error ε1 := η1 − y1 is o(h) or O(hp+1 ),
respectively.
During the steps for i ≥ 1 the consistency error, e.g., arising at x1 , is transported
into η2 , η3 , . . . Since the computation proceeds up to x = xn , one has to perform
5
Without p-fold continuous differentiability of f one cannot verify τ (x, η; h) = O(hp ).
5.4 Analysis of One-Step Methods 73
n = O(1/h) steps. If the error would be amplified in each step by a factor c > 1
(c independently of h), ηn had an error O(cn ) = O(c1/h ). Obviously, such an error
would explode exponentially as h → 0. In addition, not only can the consistency
error ε1 be amplified, but also can all consistency errors εi at the later grid points
xi .
Next we state that—thanks to Lipschitz condition (5.16)—the errors are under
control.6
Lemma 5.18 (stability of one-step methods). Assume that the Lipschitz condi-
tion (5.16) holds with constant Lφ and the local discretisation error is bounded by
|τ (xi , y(xi ); h)| ≤ Th (cf. (5.17)). Then the global discretisation error is bounded
by
e(x−x0 )Lφ − 1
|η(x, h) − y(x)| ≤ Th . (5.19)
Lφ
Proof. δi := |ηi − y(xi )| is the global error. The local discretisation error is de-
noted by τi = τ (xi , y(xi ); h). Starting with δ0 = 0, we obtain the recursion formula
δi+1 = |ηi+1 − y(xi+1 )| = |ηi + hφ(xi , ηi , h; f ) − y(xi+1 )|

y(xi+1 ) − y(xi )
= ηi − y(xi ) − h
− φ(xi , ηi , h; f )
h

y(xi+1 ) − y(xi )
= ηi − y(xi ) − h
− φ(xi , y(xi ), h; f )
h

+ h [φ(xi , ηi , h; f ) − φ(xi , y(xi ), h; f )]

y(xi+1 ) − y(xi )
≤ |ηi − y(xi )| + h − φ(xi , y(xi ), h; f )
h
+ h |φ(xi , ηi , h; f ) − φ(xi , y(xi ), h; f )|
≤ δi + h |τi | + hLφ δi = (1 + hLφ ) δi + hTh ,
which coincides with (5.11) for aν = δν , h = h, L = Lφ , k = 1, B = Th . Lemma

5.6 proves δν ≤ Th (eνhLφ − 1)/Lφ because of a0 = δ0 = 0. t u
So far, consistency is not assumed. From consistency we now derive conver-

gence. Furthermore, consistency order and convergence order will coincide.
Theorem 5.19. Let the one-step method (5.5) fulfil the Lipschitz condition (5.16),
and assume consistency. Then (5.5) is also convergent:
lim η(x, h) = y(x).
h→0
If, in addition, the consistency order is p, then also the convergence is of order p.
Proof. Consistency implies Th → 0, so that convergence follows from (5.19).

Consistency order p means Th ≤ Chp and implies |η(x, h) − y(x)| ≤ O(hp ). t
u
6
Note that perturbations in the initial value x0 are already analysed in Theorem 5.9.
We remark that, in general, the Lipschitz condition (5.16) holds only locally.
Then one argues as follows. G := {(x, y) : x ∈ [x0 , xE ], |y − y(x)| ≤ 1} is com-
(x−x0 )Lφ
pact. It suffices7 to require (5.16) on G. For sufficiently small h, Th e Lφ −1
in (5.19) is bounded by 1 and therefore (x, η(x, h)) ∈ G. A view to the proof
of Lemma 5.18 shows that all intermediate arguments belong to G and therefore,
(5.16) is applicable.
According to Theorem 5.19, one may be very optimistic that any consistent one-
step method applied to some ordinary differential equation is working well. How-
ever, the statement concerns only the asymptotic behaviour as h → 0. A problem
arises if, e.g., the asymptotic behaviour is only observed for h ≤ 10−9 , while we
want to apply the methods for a step size h ≥ 0.001. This gives rise to stronger
stability requirements (cf. §5.5.8).
5.5 Analysis of Multistep Methods
The general multistep method is defined in (5.6). Introducing formally αr := 1, we

can rewrite the r-step method as
r
X
αν ηj+ν = hφ(xj , ηj+r , ηj+r−1 , . . . , ηj , h; f ) . (5.20a)
ν=0
A multistep method is called linear if φ has the particular form

r
X
φ(xj , ηj+r , ηj+r−1 , . . . , ηj ; h; f ) = bµ f (xj+µ , ηj+µ ) . (5.20b)
µ=0
| {z }
=:fj+µ
If br = 0, the linear multistep method is explicit, otherwise it is implicit.

The coefficients αν from (5.20a) define the characteristic polynomial
r
X
ψ(ζ) := αν ζ ν . (5.21a)
ν=0
In the case of a linear multistep method, a further polynomial can be introduced:

r−1
X
σ(ζ) := bν ζ ν . (5.21b)
ν=0
Remark 5.20. Consistency condition (5.7) is equivalent to
ψ(1) = 0. (5.22)
7
Locally Lipschitz continuous functions are uniformly Lipschitz continuous on a compact set.
5.5 Analysis of Multistep Methods 75
5.5.1 Local Discretisation Error, Consistency
Using Y (x; ξ, η) from (5.17), we define the local discretisation error by
τ (x, y; h) := (5.23)
" r #
1 X
αν Y (xj+ν ; xj , y) − hφ xj , Y (xj+ν−1 ; xj , y), . . . , Y (xj ; xj , y), h; f .
h ν=0 | {z }
=y
Definition 5.21. A multistep methods is called consistent if
sup |τ (x, y(x); h)| → 0 (h → 0)

x∈I
for all f ∈ C(I × R) with Lipschitz property (5.2). Here y(x) is the solution of
(5.1a,b). Furthermore, the multi-step method (5.20a) is called consistent of order p
if |τ (x, y(x); h)| = O(hp ) holds for sufficiently smooth f .
For f = 0 and the initial valuePy0 = 1, the solution is y(x) = 1 and,
Prin this case,
r
τ (x, y(x); h) → 0 simplifies to ( ν=0 αν − hφ) /h → 0, implying ν=0 αν = 0,
which is condition (5.7).
5.5.2 Convergence
Differently from the case of a one-step method, we cannot assume exact starting
values η1 , . . . , ηr−1 . Therefore, we assume that all starting values are perturbed:
ηj = y(xj ) + εj , ε = (εj )j=0,...,r−1 .
We denote the solution corresponding to these starting values by η(x; ε, h).

Definition 5.22. A multistep methods is called convergent if for all f ∈ C(I × R)
with (5.2) and all starting values y0 , the global error satisfies
sup |η(x; ε, h) − y(x)| → 0 for h → 0 and kεk∞ → 0.

x∈I
A stronger requirement is that we also perturb the equations (5.20a) by εj for

j ≥ r:
r
X
αν ηj+ν = hφ(xj , ηj+r−1 , . . . , ηj , h; f ) + hεj+r for j ≥ 0. (5.24)
ν=0
In this case, ε = (εj )j≥0 is a tuple with as many entries as grid points (note that
the quantities εj for j < r and j ≥ r have a quite different meaning!). Again,
η(x; ε, h) → y(x) can be required for h → 0 and kεk∞ → 0.
5.5.3 Stability
We recall the characteristic polynomial ψ defined in (5.21a).
Definition 5.23. The multistep method (5.20a) is called stable if all roots ζ of the
characteristic polynomial ψ have the following property: either |ζ| < 1 or ζ is a
simple zero with |ζ| = 1.
We check three examples:

(1) The midpoint rule (5.8) yields ψ(ζ) = ζ 2 − 1. Both zeros ζ = ±1 are simple
with |ζ| = 1. Therefore, the midpoint rule is stable.
(2) The two-step method (5.9) describing the extrapolation corresponds to ψ(ζ) =
2
ζ 2 − 2ζ + 1 = (ζ − 1) . Therefore, ζ = 1 is a double root and indicates instability.
(3) In the case of a one-step methods (i.e., r = 1), αr = 1 and (5.7) lead to
ψ(ζ) = ζ − 1. The only zero ζ = 1 satisfies the second condition in Definition 5.23.
This proves the following remark corresponding to the result of Lemma 5.18.
Remark 5.24. One-step methods are always stable in the sense of Definition 5.23.
5.5.4 Difference Equations
The relation between the stability condition from Definition 5.23 and the multistep
methods (5.20a) is not quite obvious. The connection will be given in the study of
difference equations. As preparation we first discuss power bounded matrices.
5.5.4.1 Power Bounded Matrices
Definition 5.25. Let k·k be a matrix norm. A square matrix A is power bounded8 if
sup{kAn k : n ∈ N} < ∞. (5.25)
Because of the norm equivalence in finite-dimensional vector spaces, the choice

of the matrix norm k·k in Definition 5.25 is irrelevant.
To prepare the next theorem, we recall some terms from linear algebra:
Let λ be an eigenvalue of A ∈ Cd×d . λ has algebraic multiplicity k ∈ N0 ,
k
if the characteristic polynomial det(ζI − A) contains the factor (ζ − λ) , but not
k+1
(ζ − λ) . λ has geometric multiplicity k ∈ N0 if dim{e ∈ Cd : Ae = λe} = k.
The inequality ‘geometric multiplicity ≤ algebraic multiplicity’ is always valid.
8
In principle, we would like to use the term ‘stable matrix’; however, this notation is already used
for matrices satisfying sup{kexp(tA)k : t > 0} < ∞.
We remark that there is an extension of power boundedness to a family F of matrices by
sup{kAn k : n ∈ N, A ∈ F } < ∞. Characterisations are given by the Kreiss matrix theorem
(cf. Kreiss [13], Morton [17], and [18, Sect. 4.9]). For a generalisation, see Toh–Trefethen [22].
k·k is an associated matrix norm in Cd×d if there is a vector norm ||| · ||| in Cd
such that kAk = sup{||| Ax ||| / ||| x |||: x 6= 0} for all A ∈ Cd×d .
The norm k·k∞ has two meanings. For vectors x ∈ Cd , it is the maximum norm
kxk∞ = maxi |xi |, while forP matrices it is the associated matrix norm. Because of
the property kM k∞ = maxi j |Mij |, it is also called the row-sum (matrix) norm.
We denote the spectrum of a square matrix M by
σ(M ) := {λ ∈ C : λ eigenvalue of M }.
Exercise 5.26. Suppose that k·k is an associated matrix norm. Prove |λ| ≤ kM k
for all λ ∈ σ(M ).
Theorem 5.27. Equivalent characterisations of the power boundedness of A are
(5.26a) as well as (5.26b):
All eigenvalues of A satisfy either
(a) |λ| < 1 or (5.26a)
(b) |λ| = 1, and λ has coinciding algebraic and geometric multiplicities.
There is an associated matrix norm such that kAk ≤ 1. (5.26b)
Proof. We shall prove (5.25) ⇒ (5.26a) ⇒ (5.26b) ⇒ (5.25).

(i) Assume (5.25) and set Cstab := sup{kAn k : n ∈ N}. Choose an associated
n
matrix norm k·k. Applying Exercise 5.26 to M = An , we obtain |λn | = |λ| ≤
kAn k ≤ Cstab for all n ∈ N, hence |λ| ≤ 1. Assume |λ| = 1. If λ has higher
algebraic than geometric multiplicity, there is an eigenvector e 6= 0 and a generalised
eigenvector h, so that Ae = λe and Ah = e + λh. We conclude that An h =
λn−1 (ne + λh) and kAn hk = kne + λhk ≥ n kek − kλhk → ∞ as n → ∞, in
the contradiction to kAn hk ≤ Cstab khk. Hence, (5.26a) holds.
(ii) Assume (5.26a). Sort the eigenvalues λi of A such that |λ1 | ≤ |λ2 | ≤ . . . <
|λd−m+1 | = . . . = |λd | = 1, where m ≥ 0 is the number the zeros with absolute
value equal to one. Consider the Jordan normal form
  

 λ1 ∗



  λ2 ∗ 
J 0 J = ,
  
J = T −1 AT = 1
with 1  ..
0 D 
 . ∗ 
λ

d−m



D = diag{λd−m+1 , . . . , λd },

where the entries ∗ are either zero or one. Since the algebraic and geometric
multiplicities of λd−m+1 , . . . , λd coincide, D is a diagonal m × m matrix, while
all eigenvalues λi (i = 1, . . . , r − m) have absolute value < 1. Set ∆ε :=
diag{1, ε, ε2 , . . . , εr−1 } with ε ∈ (0, 1 − |λr−m |]. One verifies that ∆−1
ε J∆ε has
the row-sum norm k∆−1 ε J∆ε k∞ ≤ 1. Therefore, a transformation by S := T ∆ε
yields the norm kS −1 ASk∞ ≤ 1. kAk := kS −1 ASk∞ is the associated matrix
norm corresponding to the vector norm kxk := kSxk∞ . This proves (5.26b).
(iii) Assume (5.26b). Associated matrix norms are submultiplicative; i.e.,
n
kAn k ≤ kAk , so that kAk ≤ 1 implies Cstab = 1 < ∞. t u
So far, we have considered a single matrix. In the case of a family A ⊂ Cn×n

of matrices, one likes to have a criterion as to whether there is a uniform bound of
sup{kAn k : n ∈ N, A ∈ A}.
Lemma 5.28. Let A ⊂ Cn×n be bounded, and suppose that there is some γ < 1
such that the eigenvalues satisfy |λ1 (A)| ≤ |λ2 (A)| ≤ . . . ≤ |λn (A)| ≤ 1 and
|λn−1 (A)| ≤ γ < 1 for all A ∈ A. Then sup{kAn k : n ∈ N, A ∈ A} < ∞.
Proof. We may choose the spectral norm k · k = k · k2 . The Schur normal form RA
is defined by A = QRA Q−1 , Q unitary, RA upper triangular, with (RA )ii = λi (A).
By boundedness of A, there is some M > 0 such that RA is bounded entry-wise by
(RA )ij ≤ Rij , where the matrix R is defined by Rij := 0 for i > j, Rii := γ for
 
γ ... M M
 .
 0 ..

1 ≤ i ≤ n − 1, Rnn := 1, Rij := M for i < j; i.e., R = M M. It is easy to
0 ... γ M
0 ... 0 1
verify that kAn k2 = kRA
n
k2 ≤ kRn k2 . Since R is power bounded, we have proved
a uniform bound. t u
5.5.4.2 Solution Space F0
The set F = CN0 consists of sequences x = (xj )j∈N0 of complex numbers. We are
looking for sequences x ∈ F satisfying the following difference equation:
r
X
αν xj+ν = 0 for all j ≥ 0, where αr = 1. (5.27)
ν=0
Lemma 5.29. (a) F forms a linear vector space.

(b) F0 := {x ∈ F satisfies (5.27)} is a linear subspace of F with dim(F0 ) = r.
Proof. (i) The vector space properties of F and F0 are trivial. It remains to prove
dim F0 = r. We define x(i) ∈ F for i = 0, 1, . . . , r − 1 by the initial values
(i)
xj = δij for j ∈ {0, . . . , r − 1}. For j ≥ r we use (5.27) to define
r−1
(i) (i)
X
xj := αν xj−r+ν for all j ≥ r. (5.28)
ν=0
Obviously, x(i) Psatisfies (5.27); i.e., x(i) ∈ F0 for 0 ≤ i ≤ r − 1.

(ii) Assume i βi x(i) = 0. Evaluating the entries at j ∈ {0, . . . , r − 1}, we
P (i) P
obtain 0 = i βi xj = i βi δij = βj , which proves linear independence.
Pr−1
(iii) For each x ∈ F0 we define y := x− i=0 xi x(i) ∈ F0 , which by definition
satisfies y0 = y1 = . . . = yr−1 = 0. Analogous to (5.28), these initial values lead
to yj = 0 for all j ≥ r. The result y = 0 proves that {x(0) , . . . , x(r−1) } already
spans F0 . Together, dim(F0 ) = r follows. t u
5.5.4.3 Representation of the Solutions

Pr ν
Remark 5.30. (a) Let ζ0 ∈ C be a root of the polynomial ψ(ζ) = ν=0 αν ζ .
j
Then x = (ζ0 )j∈N0 is a solution of (5.27); i.e., x ∈ F0 .
(b) Suppose that ψ has r different zeros ζi ∈ C, i = 1, . . . , r. Then the solutions
x(i) = (ζij )j∈N0 form a basis of the space F0 .
Proof. (i) Inserting xj = ζ0j into (5.27) yields

r
X r
X r
X
αν xj+ν = αν ζ0j+ν = ζ0j αν ζ0ν = ζ0j ψ(ζ0 ) = 0;
ν=0 ν=0 ν=0
i.e., x ∈ F0 .
(ii) By Part (i), x(i) ∈ F0 holds. One easily verifies that the solutions x(i) are
linearly independent. Because of dim F0 = r, {x(i) : 1 ≤ i ≤ r} forms a basis. t u
In the case of Remark 5.30, the zeros are simple. It remains to discuss the case
of multiple zeros.
We recall that the polynomial ψ has an (at least) k-fold zero ζ0 if and only if
ψ(ζ0 ) = ψ 0 (ζ0 ) = . . . = ψ (k−1) (ζ0 ) = 0 . The Leibniz rule yields
d ` j
ζ ψ(ζ) = 0 at ζ = ζ0 for 0 ≤ ` ≤ k − 1.
dζ
d `

The explicit representation of ( dζ ) ζ j ψ(ζ) reads
r
(`) X
j
0 = ζ ψ(ζ) = αν ζ0j+ν−` (j + ν) (j + ν − 1) · . . . · (j + ν − ` + 1) .
ζ=ζ0
ν=0
(5.29)
(`)
Define x(`) for ` ∈ {0, 1, . . . , k − 1} via xj = ζ0j j (j − 1) · . . . · (j − ` + 1).
Insertion into the difference equation (5.27) yields
r r
(`)
X X
αν xj+ν = αν ζ0j+ν (j + ν) (j + ν − 1) · . . . · (j + ν − ` + 1) .
ν=0 ν=0
Pr (`)
This is ζ0` times the expression in (5.29); hence ν=0 αν xj+ν = 0; i.e., x(`) ∈ F0 .
Remark 5.31. Let ζ0 6= 0 be a zero of ψ with multiplicity k.

(`) Q`−1
(a) Then x(0) , . . . , x(k−1) with xj = ζ0j ν=0 (j − ν) (` = 0, . . . , k − 1) are k
linearly independent solutions of (5.27).
(`)
(b) Similarly, x̂j = ζ0j j ` represent k linearly independent solutions of (5.27).
Proof. (i) x(`) ∈ F0 is shown above. The linear independence for ` = 0, . . . , k − 1

(`)
follows, e.g., from the different growth of xj /ζ0j for j → ∞.
nQ o
`−1
(ii) ν=0 (x − ν) : ` = 0, . . . , k − 1 as well as x` : ` = 0, . . . , k − 1 are
bases of the polynomials of degree ≤ k − 1. Therefore {x(0) , . . . , x(k−1) } and
{x̂(0) , . . . , x̂(k−1) } span the same space. t
u
The case ζ0 = 0 is excluded in Remark 5.31, since the previous definition leads
(`)
to xj = 0 for j ≥ min{1 − `, 0} and therefore does not yield linearly independent
solutions.
(i)
Remark 5.32. Let ζ0 = 0 be a k-fold zero of ψ. Then x(i) with xj = (δij )j∈N0
(i = 0, . . . , k − 1) are k linearly independent solutions of (5.27).
Proof. Use ψ(0) = ψ 0 (0) = . . . = ψ (k−1) (0) = 0 to obtain α0 = . . . = αk−1 = 0.

Pr (i) Pr (i) (i)
This shows that ν=0 αν xj+ν = ν=k αν xj+ν . By definition, xj+ν = 0 holds
for ν ≥ k; i.e., (5.27) is satisfied. Obviously, the x(i) are linearly independent. t
u
TheoremP 5.33. Let ζi (i = 1, . . . , m) be the different zeros of the polynomial

r ν
ψ(ζ) = ν=0 αν ζ with the corresponding multiplicities ki . Any ζi gives rise
to ki solutions of (5.27), which are defined in Remark 5.31, if ζi 6= 0, and in
Remark 5.32, if ζi = 0. Altogether, r linearly independent solutions spanning F0
are characterised.
Proof. The linear P

independence of the constructed solutions is left as an exercise.
In Theorem 5.33, ki = r solutions are described. They form a basis of F0 , since,
according to Lemma 5.29, dim F0 = r holds. t u
5.5.4.4 Stability
Definition 5.34. The difference equation (5.27) is called stable if any solution of
(5.27) is bounded with respect to the supremum norm:
kxk∞ := sup |xj | < ∞ for all x ∈ F0 .

j∈N0
Remark 5.35. A characterisation equivalent to Definition 5.34 reads as follows:

there is a constant C, so that
kxk∞ ≤ C max |xj | for all x ∈ F0 . (5.30)

j=0,...,r−1
Proof. (i) Eq. (5.30) implies kxk∞ < ∞, since maxj=0,...,r−1 |xj | is always finite.
(i)
Pr−1x ∈(i)F0 as in (ii) of the proof of Lemma 5.29. x has
(ii) We choose the basis
a representation x = i=0 xi x . Assuming stability in the sense of Definition
(i)
Pr−1
5.34, we have Ci := kx k∞ < ∞ and therefore also C := Ci < ∞.
Pr−1 Pr−1 Pi=0
r−1
The estimate kxk∞ = k i=0 xi x(i) k∞ ≤ i=0 |xi | kx(i) k∞ = i=0 Ci |xi | ≤
C maxj=0,...,r−1 |xj | proves (5.30). t
u
Obviously, all solutions of (5.27) are bounded if and only if all basis solutions
given in Theorem 5.33 are bounded. The following complete list of disjoint cases
refers to the zeros ζi of ψ and their multiplicities ki .
1. |ζi | < 1: all sequences (ζij j ` )j∈N0 with 0 ≤ ` < ki are zero sequences and
therefore bounded.
2. |ζi | > 1: for all sequences lead to lim |ζij j ` | = ∞; i.e., they are unbounded.
3. |ζi | = 1 and ki = 1 (simple zero): (ζij )j∈N0 is bounded by 1 in absolute value.
4. |ζi | = 1 and ki > 1 (multiple zero): lim |ζij j ` | = ∞ holds for 1 ≤ ` ≤ ki − 1;
i.e., the sequences are unbounded.
Therefore, the first and third cases characterise the stable situations, while the
second and fourth cases lead to instability. This proves the next theorem.
Theorem 5.36. The difference equation (5.27) is stable if and only if ψ satisfies the
stability condition from Definition 5.23.
5.5.4.5 Companion Matrix

Pr
Definition 5.37. The companion matrix of the polynomial ψ(ζ) = ν=0 αν ζ ν is of
size r × r and is equal to
 
0 1
 .. .. 
A=
 . . .

(5.31)
 0 1 
−α0 . . . −αr−2 −αr−1
Remark 5.38. (a) det(ζI − A) = ψ(ζ).

(b) Any solution (xj )j∈N0 of the difference equation (5.27) satisfies (5.32) and, vice
versa, (5.32) implies (5.27):
   
xj+1 xj
 xj+2   xj+1 
 ..  = A  . (5.32)
   
..
 .   . 
xj+r xj+r−1
By (5.32) we can formally reformulate the r-step method as a one-step method

>
for the r-tuple Xj := (xj , . . . , xj+r−1 ) . Therefore, the stability behaviour must
be expressed by the properties of A. An obvious connection is described in the next
remark.
Remark 5.39. The difference equation (5.27) is stable if and only if A is a power
bounded matrix.
Proof. Let k·k denote a vector norm as well as the associated matrix norm. The
tuple Xj satisfies Xj = AXj−1 (cf. (5.32)). In particular, Xn = An X0 holds.
Uniform boundedness of |xj | and kXj k are equivalent. If (5.27) is stable, kAn X0 k
is uniformly bounded for all X0 with kX0 k ≤ 1 and all n ∈ N; i.e., A is a power
bounded matrix. On the other hand, Cstab := sup{kAn k : n ∈ N} < ∞ yields the
estimate kXn k ≤ Cstab kX0 k and therefore stability. t
u
Using (5.26b), we obtain the next statement.
Lemma 5.40. The difference equation (5.27) is stable if and only if there is an
associated matrix norm, so that kAk ≤ 1 holds for the companion matrix A from
(5.31).
5.5.4.6 Estimates of Inhomogeneous Solutions
Theorem 5.41. Suppose that the difference equation is stable and that the initial
values fulfil |xj | ≤ α for 0 ≤ j ≤Pr − 1. If the sequence (xj )j∈N0 satisfies the
r
inhomogeneous difference equation ν=0 αν xj+ν = βj+r with
|βj+r | ≤ β + γ max{|xµ | : 0 ≤ µ ≤ j + r − 1}
for some γ ≥ 0, then there exist k and k 0 such that

0 jkγ jkβ for γ = 0,
|xj | ≤ kk αe + β jγk
γ e −1 for γ > 0.
Proof. (i) Let A be the companion matrix. k·k denotes the vector norm in Rr as
well as the associated matrix norm k·k from Lemma 5.40. Because of the norm
equivalence,
kXk∞ ≤ k kXk , kXk ≤ k 0 kXk∞ for all X ∈ Rr
holds for suitable k, k 0 .

> >
(ii) Set Xj := (xj , . . . , xj+r−1 ) and e = (0, . . . , 0, 1) ∈ Rr . According to
Lemma 5.40, kAk ≤ 1 holds. Since a scaling of the vector norm does not change
the associated matrixPnorm, we assume without loss of generality that kek = 1. The
r
difference equation ν=0 αν xj+ν = βj+r is equivalent to
Xj+1 = AXj + βj+r e.
(iii) Set ξj := max{|xµ | : 0 ≤ µ ≤ j}. By definition of Xµ , also ξj =

max0≤µ≤j−r+1 kXµ k∞ is valid for j ≥ r − 1. It follows that
kXj+1 k = kAXj + βj+r ek ≤ kAk kXj k + |βj+r | kek ≤ kXj k + |βj+r |

|{z} |{z}
≤1 =1
≤ kXj k + β + γξj+r−1 .
Define ηj by
η0 := kX0 k , ηj+1 := ηj + β + γξj+r−1 . (5.33)

Obviously, kXj k ≤ ηj and ηj+1 ≥ ηj (j ≥ 0) hold. The estimate
ξj+r−1 = max |xµ | = max kXµ k∞

0≤µ≤j+r−1 0≤µ≤j
≤ max k kXµ k ≤ k max ηµ = kηj

0≤µ≤j 0≤µ≤j ηj+1 ≥ηj
together with the definition (5.33) of ηj+1 yields
ηj+1 ≤ (1 + γk) ηj + β.
Apply Lemma 5.6 to this inequality. The corresponding quantities in (5.11) are
ν ≡ j, aν ≡ ηj , h ≡ 1, L ≡ kγ, B ≡ β. Lemma 5.6 yields the inequality

jβ if γ = 0
ηj ≤ η0 ejkγ + β (j ∈ N0 ).
kγ ejkγ − 1 if γ > 0
Furthermore, η0 = kX0 k ≤ k 0 kX0 k∞ ≤ k 0 α and |xj | ≤ kXj k∞ ≤ k kXj k ≤ kηj

holds. Together, the assertion of the theorem follows. t
u
5.5.5 Stability and Convergence Theorems
We shall show that convergence and stability of multistep methods are almost
equivalent. For exact statements one needs a further assumption concerning the
connection of φ(xj , ηj+r−1 , . . . , ηj , h; f ) and f . A very weak assumption is
f =0 =⇒ φ(xj , ηj+r−1 , . . . , ηj , h; f ) = 0. (5.34)
This assumption is satisfied, in particular, for the important class of linear r-step
methods:
r
X
φ(xj , ηj+r , . . . , ηj , h; f ) = bµ fj+µ with fk = f (xk , ηk ). (5.35)
µ=0
Theorem 5.42 (stability theorem). Suppose (5.34). Then the convergence from
Definition 5.22 implies stability.
Proof. (i) We choose f = 0 and the starting value y0 = 0. Therefore, y = 0 is

the exact solution
Pr of the initial-value problem, while the discrete solution satisfies
the equations ν=0 αν ηj+νP= 0 with initial values η0 = ε0 , . . . , ηr−1 = εr−1
r
(cf. Definition 5.22). Since ν=0 αν ηj+ν = 0 is the difference equation (5.27),
(ηj )j∈N0 ∈ F0 holds. However, we have to note that the multistep method visits
only the finite section (ηj )0≤j≤J(h) with J(h) = b(xE − x0 ) /hc, since those j
satisfy xj ∈ I.
(ii) For the indirect proof, assume instability. Then an unbounded solution
x ∈ F0 exists. The divergence
C(h) := max {|xj | : 0 ≤ j ≤ J(h)} → ∞ for h → 0
follows from J(h) → ∞ for h → 0 and kxk∞ = ∞. Choose the initial pertur-
bation ε = (εj )j=0,...,r−1 := (xj /C(h))j=0,...,r−1 . Obviously, kεk∞ → 0 holds
for h → 0. For this initial perturbation the multistep method produces the solution
1
(ηj )0≤j≤J(h) with ηj = C(h) xj . Since
1
sup |η(x; ε, h) − y(x)| = max {|xj | : 0 ≤ j ≤ J(h)} = 1,
x∈I C(h)
this error does not tend to 0 as h → 0, kεk∞ → 0, in contradiction to the assumed

convergence. t u
For the reverse direction we need a Lipschitz condition corresponding to (5.16)
in the case of a one-step method:
for each f ∈ C(I × R) with (5.2) there is Lφ ∈ R such that (5.36)

|φ(xj , ur−1 , . . . , u0 , h; f ) − φ(xj , vr−1 , . . . , v0 , h; f )| ≤ Lφ max |ui − vi | .
i=0,...,r−1
Remark 5.43. Condition (5.36) is satisfied for linear r-step methods (5.35).
Theorem 5.44 (convergence theorem). Let (5.36) be valid. Furthermore, the multi-
step method is supposed to be consistent and stable. Then it is convergent (even in
the stronger sense as discussed below Definition 5.22).
Proof. The initial error is defined by ηj = y(xj ) + εj for j = 0, . . . , r − 1. The
multistep formula with additional errors hεj+r is described in (5.24). The norm of
the error is kεk∞ := max{|εj | : 0 ≤ j ≤ J(h)} with J(h) as in the previous
proof (the usual convergence leads to εj = 0 for r ≤ j ≤ J(h), only in the case
described in brackets can εj 6= 0 appear for r ≤ j ≤ J(h)). We have to show that
η(x; ε, h) → y(x) for h → 0, kεk∞ → 0.
The error is denoted by ej := ηj − y(xj ). The initial values for 0 ≤ j ≤ r − 1
are ej = εj . The equation
r
X
αν y(xj+ν )−hφ(xj , y(xj+r−1 ), . . . , y(xj ), h; f ) = hτ (xj , y(xj ); h) =: hτj+r
ν=0
containing the local discretisation error τj+r is equivalent to (5.23). We form the
difference between the latter equation and (5.24) for j ≥ r and obtain
r
X
αν ej+ν = βj+r := h (εj+r − τj+r )
ν=0
+ h [φ(xj , ηj+r−1 ,. . . , ηj , h; f ) − φ(xj , y(xj+r−1 ),. . . , y(xj ), h; f )] .
From (5.36) we infer that
|βj+r | ≤ hLφ max |eµ | + h (kεk∞ + kτ k∞ ) ,

j≤µ≤,j+r−1
where τ = (τj )r≤j≤J(h) , kτ k∞ = max |τj |.

r≤j≤J(h)
The consistency condition supx∈I τ (x, y(x); h) → 0 implies kτ k∞ → 0 for h → 0.

The assumptions of Theorem 5.41 hold with
xj = ej , α = kεk∞ , β = h (kεk∞ + kτ k∞ ) , γ = hLφ ,
so that the theorem yields
kεk∞ + kτ k∞ jhLφ k
|ej | ≤ kk 0 kεk∞ ejhLφ k +

e −1 (5.37)
Lφ
in the case of hLφ > 0 (the case hLφ = 0 is analogous). The product jh in the
exponent is to be interpreted as xj − x0 and therefore bounded by xE − x0 (or it
is constant and equal to x in the limit process j = n → ∞, h := (x − x0 ) /n). As
part of the definition of convergence, kτ k∞ → 0 (consistency) and kεk∞ → 0
holds for h → 0. According to (5.37), ej converges uniformly to zero; i.e.,
supx∈I |η(x; ε, h) − y(x)| → 0. t u
Corollary 5.45. In addition to the assumptions in Theorem 5.44 assume consistency
of order p. Then also the convergence order is p, provided that the initial errors are
sufficiently small:

p r−1
|η(x; ε, h) − y(x)| ≤ C h + max |εj | .
j=0
Proof. We have kεk∞ = max0≤j≤r−1 |εj | and kτ k∞ ≤ O(hp ). Inequality (5.37)

proves the desired error bound. t
u
5.5.6 Construction of Optimal Multistep Methods
5.5.6.1 Examples
The Adams–Bashforth methods are explicit linear r-step methods of the form
r−1
X
ηj+r = ηj+r−1 + h bµ fj+µ (5.38)
µ=0
(cf. (5.20b)). The associated characteristic polynomial is
ψ(ζ) = ζ r − ζ r−1 = ζ r−1 (ζ − 1) .

Its zeros are ζ1 = . . . = ζr−1 = 0, ζr = 1, so that the Adams–Bashforth methods

are stable. The coefficients bµ (µ = 0, . . . , r − 1) can be used to make the local con-
sistency error as small as possible. The optimal choice yields a multistep method9
of order r.
Exercise 5.46. (a) Euler’s method is the optimal Adams–Bashforth method for r = 1.
(b) What are the optimal coefficients b0 , b1 in (5.38) for r = 2 ?
The general explicit linear two-step method
ηj+2 = −α1 ηj+1 − α0 ηj + h [b1 fj+1 + b0 fj ]
contains four free parameter. Because of the side condition (5.7) (i.e., α0 + α1 = 1),
there remain three degrees of freedom. Thus, the optimal choice α0 = −5, α0 = 2,
α1 = 4, b1 = 4 of the coefficients can reach consistency of order p = 3. The
resulting method is
ηj+2 = −4ηj+1 + 5ηj + h [4fj+1 + 2fj ] . (5.39)
j xj ηj − y(xj ) The associated polynomial is

2 0.02 −0.1610 -8
ψ(ζ) = ζ 2 + 4ζ − 5 = (ζ − 1) (ζ + 5) .
3 0.03 +0.5010 -8
4 0.04 −0.3010 -7 The root −5 proves instability. We demonstrate that this
5 0.05 +0.1410 -6 instability is clearly observed in practice. We apply (5.39)
.. .. ..
. . . to the initial-value problem y 0 = −y, y(0) = 1 =: η0
−x −h
99 0.99 +0.1310 +60 (⇒ y(x) = e ) and choose the exact value η1 := e
100 1.00 −0.6510 +60 to avoid any further error. The step size in I = [0, 1] is
chosen by h = 0.01. The root ζ = −5 is responsible for
alternating signs of the error and for the explosive increase (598 = 3.210 68).
5.5.6.2 Stable Multistep Methods of Optimal Order
The previous example shows that one cannot use all coefficients αν , bµ from (5.6)
and (5.35) in order to maximise the consistency order. Instead stability is a side
condition, when we optimise αν , bµ . The characterisation of optimal stable multi-
step methods is due to Dahlquist [2].
Theorem 5.47. (a) If r ≥ 1 is odd, the highest consistency order of a stable linear
r-step method is p = r + 1.
(b) If r ≥ 2 is even, the highest consistency order is p = r + 2. In this case, all roots
of the characteristic polynomial ψ have absolute value 1.
9
Note the following advantage of multistep methods compared with one-step methods: In spite
of the increased consistency order r, only one function value fj+r−1 = f (xj+r−1 , ηj+r−1 )
needs to be evaluated per grid point xj ; the others are known from the previous steps.
5.5.6.3 Proof
Step 1: A stable linear r-step method of order p = r + 1 is existing.

For this purpose, substitute the variable ζ of the polynomials ψ(ζ) and σ(ζ) from
(5.21a,b) by z:
1+z ζ −1
ζ= , z= . (5.40a)
1−z ζ +1
This defines the functions
r r
1−z 1+z 1−z 1+z
p(z) := ψ , s(z) = σ . (5.40b)
2 1−z 2 1−z
Remark 5.48. (a) p(z) and s(z) are polynomials of degree ≤ r.

(b) If ζ0 6= −1 is a root of ψ(ζ) with multiplicity k, then p(z) has the root z0 = ζζ00 −1
+1
again of multiplicity k.
(c) The transformation (5.40a) maps the complex unit circle |ζ| < 1 onto the left
half-plane <e z < 0; in particular, ζ = 1 is mapped onto z = 0, and ζ = −1 onto
z = ∞.
Proof. (i) Part (c) can be verified directly. For Part (a) note that p(z) is a lin-
1+z ν r r−ν ν
ear combination of the polynomials ( 1−z ) (1 − z) = (1 − z) (1 + z) for
0 ≤ ν ≤ r , which are all of degree r. Similarly for s.
k
(ii) Let k be the multiplicity of ζ0 . Then ψ(ζ) = (ζ − ζ0 ) ψ0 (ζ) holds with
some ψ0 (ζ0 ) 6= 0. Hence,
r k
1−z 1+z 1+z
p(z) = − ζ0 ψ0
2 1−z 1−z

k 1 r−k 1+z
= (1 + z − (1 − z) ζ0 ) (1 − z) ψ0
2r 1−z
" #
r−k
k (1 − z) 1+z
= (z − z0 ) ψ
k 0
.
2r (ζ0 + 1) 1−z
Since ζ0 = ∞ is excluded, z0 = 1 is not a root and the bracket [. . .] does not vanish
at z = z0 . This proves Part (b). t
u
Because of stability and ψ(1) = 0 (cf. (5.22)), ζ = 1 is a simple root. By
Remark 5.48c, p(z) has a simple root at z = 0. Hence, p is of the form
p(z) = α1 z + α2 z 2 + . . . + α` z ` with α1 6= 0, ` = degree(p) ≤ k. (5.41a)
Without loss of generality, we may assume that
α1 > 0. (5.41b)
(otherwise scale the equation of the multistep method by −1 changing ψ(ζ) into
−ψ(ζ)). We shall prove that
αµ ≥ 0 for all 1 ≤ µ ≤ r. (5.41c)
For this purpose, we denote the zeros of p by zν = xν + i yν (xν , yν ∈ R). Then,

Y Y Y 2

p(z) = α` z (z − zν ) = α` z (z − xν ) (z − xν ) + yν2
ν ν with yν =0 ν with yν 6=0
(5.41d)
must hold, where the last product is taken over all pairs of conjugate complex zeros.
Stability implies |ζν | ≤ 1, which by Remark 5.48c implies <e zν ≤ 0; i.e., xν ≤ 0.
Because of z − xν = z + |xν |, all polynomial coefficients of p(z) must carry the
same sign (or vanish). Hence, (5.41b) implies (5.41c).
We define the holomorphic function
ψ(ζ)
ϕ(ζ) := − σ(ζ). (5.42)
log ζ
Note that because of ψ(1) = 0, ϕ has no singularity at ζ = 1. A related function is

r
1−z 1+z p(z)
g(z) :=
2
ϕ
1−z
= 1+z − s(z). (5.43)
log 1−z
Theorem 5.49. Assume ψ(1) = 0 (cf. (5.22)). Then the linear multistep method
(5.20b) has the (local) consistency order p if and only if ζ = 1 is a p-fold root of ϕ.
Proof. Assume consistency order p. We choose the differential equation y 0 = y.

Then we have (up to a factor) z(t) = et and
r r r r
aν ex+νh − h bν ex+νh aν (eh )ν − h bν (eh )ν
P P P P
ν=0 ν=0 x ν=0 ν=0
τ (x, y, h) = =e
h h
ψ(eh )

= ex − σ(eh ) = ex ϕ(eh ).
h
Set δ := eh −1 = h·eθh (0 < θ < h from the mean value theorem). Hence, δ can be
estimated from both sides by const · h. The Taylor expansion of ϕ(eh ) = ϕ(1 + δ)
around δ = 0 exists, since ϕ is holomorphic at ζ = 1:
ϕ(p−1) (1) p−1

ϕ(eh ) = ϕ(1) + ϕ0 (1)δ + . . . + δ + O(δ p ).
(p − 1)!
Since δ k ∼ hk , we conclude from τ (x, y, h) = O(hp ) = O(δ p ) that terms involving

δ k with k < p cannot appear; i.e., ϕ(1) = ϕ0 (1) = . . . = ϕ(p−1) (1) = 0. Hence,
1 is a p-fold zero of ϕ.
On the other hand, if 1 is a p-fold zero, the Taylor expansion shows that
τ (x, y, h) = ex ϕ(eh ) = O(δ p ) = O(hp ) and therefore the method has consistency
order p. t u
Concerning g, the following implications from (5.43) are valid:
g(z) has a p-fold zero at z = 0

⇔ ϕ(ζ) has a p-fold zero at ζ = 1
⇔ p = consistency order.
1+z
We recall the function p(z) from (5.40b). Since p(z)/ log 1−z is holomorphic at
z = 0, there is a power series
z p(z)
1+z = β 0 + β 1 z + β2 z 2 + . . . (5.44a)
log 1−z
z
The function g connects

Pr p and s. To obtain a p-fold zero of g at z = 0, s must be of
the form s(z) = µ=0 βµ0 z µ with βµ0 = βµ for 0 ≤ µ ≤ p − 1. Since we consider
the case p = r + 1 > r = degree(s) of Theorem 5.47a, the polynomial s(z) is
already uniquely determined:
s(z) = β0 + β1 z + . . . + βr z r , (5.44b)
where r = p − 1. Polynomial s from (5.44b) determines a unique σ (cf. (5.40b)).

This ends the construction of a stable method of order p = r + 1.
Step 2: For odd r, there is no stable method of consistency order p > r + 1.

Let the method have an order p > r + 1. The comparison of (5.44a) and (5.44b)
shows that in (5.44a)
βµ = 0 for r + 1 ≤ µ ≤ p − 1 (5.44c)
must hold (the first βµ for 0 ≤ µ ≤ r can be made to zero by the choice (5.44b)).
1+z
The function z/ log 1−z is an even function in z, so that the power series becomes
z
1+z = c0 + c2 z 2 + c4 z 4 + . . .
log 1−z
Let αµ be the coefficients from (5.41a), where we set αµ := 0 for µ > ` =

degree(p). Comparing the coefficients of both sides in (5.44a), we obtain
β0 = c0 α1 , ... β2ν = c0 α2ν+1 + c2 α2ν−1 + . . . + c2ν α1 ,

β1 = c0 α2 , ... β2ν+1 = c0 α2ν+2 + c2 α2ν + . . . + c2ν α2 .
We shall prove that c2ν < 0 for all ν ≥ 1. Hence, for odd r it follows that
βr+1 = c0 αr+2 + c2 αr + c4 αr−2 + . . . + cr+1 α1 < 0 (5.44d)

| {z } |{z}|{z} |{z}| {z } |{z}|{z}
=0 <0 ≥0 <0 ≥0 <0 >0
in contradiction to βr+1 = 0 (cf. (5.44c)).

For the proof of c2ν < 0 we need the following lemma of Kaluza [12].
P∞ P∞
Lemma 5.50. Let f (t) = ν=0 Aν tν and g(t) = ν=0 Bν tν be power series with
the properties
f (t)g(t) ≡ 1, Aν > 0 (ν ≥ 0), Aν+1 Aν−1 > A2ν (ν ≥ 1).
Then Bν < 0 holds for all ν ≥ 1.
Aν+1 Aν
Proof. (i) The assumption Aν+1 Aν−1 > A2ν can be written as Aν > Aν−1 for
An+1 An−ν+1
ν ≥ 1. From this we conclude that An > An−ν for 1 ≤ ν ≤ n. The latter
inequality can be rewritten as
An+1 An−ν − An An−ν+1 > 0. (5.45)
(ii) Without loss of generality assume A0 = 1. This implies B0 = 1. Comparison
of the coefficients in f (t)g(t) ≡ 1 proves that
n
X n
X
0 = An + Bν An−ν (n ≥ 1), −Bn+1 = An+1 + Bν An−ν+1 (n ≥ 0).
ν=1 ν=1
For n = 0, the latter identity shows that B1 <P

0. Multiply the first equation by An+1 ,
n
the second by −An , and add: An Bn+1 = ν=1 Bν (An+1 An−ν − An An−ν+1 ).
Thanks to (5.45), Bn+1 < 0 follows by induction. t u
1 1+z
We apply this lemma to f (z 2 ) = z log 1−z = 2+ 23 z 2 + 25 z 4 +. . . The coefficients
2
Aν = 2ν+1 > 0 satisfy
2 2 4 4 2
Aν+1 Aν−1 = · = 2 > 2 = Aν
2ν + 3 2ν − 1 (2ν + 1) − 4 (2ν + 1)
and therefore the supposition of the lemma. Since Bν = c2ν , the assertion of Step 2
is proved.
Step 3: For even r, stable methods are characterised as in Theorem 5.47b.
For even r, the sum corresponding to (5.44d) becomes
βr+1 = c0 αr+2 + c2 αr + c4 αr−2 + . . . + cr α2 ≤ 0,

| {z } |{z}|{z} |{z}| {z } |{z}|{z}
=0 <0 ≥0 <0 ≥0 <0 ≥0
where βr+1 = 0 holds if and only if α2 = α4 = . . . = αr = 0. The latter property

is equivalent to p(z) being odd:
p(z) = −p(−z) for all z.
Hence, each root zν of p corresponds to a root −zν . As stability requires <e zν ≤ 0,

we obtain <e zν = 0 for all zν , corresponding to the condition |ζν | = 1. On the
other hand, we have the reverse statement: If |ζν | = 1 for all root of ψ, it follows
that <e zν = 0 and
Y 2

p(z) = const · z z 2 − (=m zν )
ν
and therefore p(z) = −p(−z). Hence, the order p = r + 2 can be obtained.

Order p = r + 3 (r even) cannot be reached, since by Step 2 βr+2 < 0.
5.5.7 Further Remarks
5.5.7.1 Systems of Differential Equations
The previous considerations concerning stability remain unchanged for systems

y 0 = f (x, y), where y and f take values in Rn . Nevertheless, further complications
can appear.
First, we consider the case of stiff systems. The previous convergence analysis
provides statements for the limit h → 0. The proven error estimates hold for h
sufficiently small. A more practical question is: how large can we choose h, when
does the asymptotic regime start? For instance, the differential equation y 0 = −y
has the positive and decreasing solution e−x , while the explicit Euler scheme (5.4)
with step size h = 3 yields the values ηj = (−2)j , which are oscillatory in sign
and increasing. A simple analysis shows that for h < 1 also the discrete solution
is positive and decreasing. Now assume the system y 0 = Ay of linear differential
2×2
with A = diag{−1, −1000} ∈ R . The dominant part of the
equations solution
is 0 e related to the eigenvalue one, since the other component y02 e−1000x
y1 −x
decays very strongly. As in the one-dimensional example from above, one would
like to use a step size of the size h < 1. However, the second component enforces
the inequality h < 1/1000. This effect appears for general linear systems, if A
possesses one eigenvalue of moderate size and another one with strongly negative
real part, or for nonlinear systems y 0 = f (x, y), where A(x) := ∂f /∂y has similar
properties. This leads to the definition of A-stability (or absolute stability; cf. [20],
[9], [3]). Good candidates for A-stable methods are implicit ones (cf. 5.4.1). Implicit
methods have to solve a linear system of the form Az = b. Here, another problem
arises. For instance, the spatial discretisation of a parabolic initial value problem (see
Chap. 6) yields a large stiff system of ordinary differential equations. The solution
of the large linear system requires further numerical techniques (cf. [6]).
Next, we consider the formulation By 0 = Cy of a linear system or, in the gen-
eral nonlinear case, F (x, y, y 0 ) = 0. If B is regular (or F solvable with respect to
y 0 ), we regain the previous system with A := B −1 C. If, however, B ∈ Rn×n is
singular with rank k, the system By 0 = Cy consists of a mixture of n − k differ-
ential equations and k algebraic side conditions. In this case, the system is called
a differential-algebraic equations (DAE; [14]). In between there are singularly per-
turbed systems, where B = B0 + εB1 is regular for ε > 0, but ε is small and B0 is
singular.
5.5.8 Other Stability Concepts
The analysis of numerical schemes for ordinary differential equations has created
many further variants of stability definitions. Besides this, there are stability con-
ditions (e.g., Lyapunov stability) which are not connected with discretisations, but
with the (undiscretised) differential equation and its dynamical behaviour (cf. [8,
§I.13], [11, §X], [7], [21]).
References
1. Butcher, J.C.: A history of Runge-Kutta methods. Appl. Numer. Math. 20, 247–260 (1996)
2. Dahlquist, G.: Convergence and stability in the numerical integration of ordinary differential
equations. Math. Scand. 4, 33–53 (1956)
3. Deuflhard, P., Bornemann, F.: Scientific computing with ordinary differential equations.
Springer, New York (2002)
4. Deuflhard, P., Bornemann, F.: Numerische Mathematik II. Gewöhnliche Differentialgleichun-
gen, 3rd ed. Walter de Gruyter, Berlin (2008)
5. Gautschi, W.: Numerical Analysis. An Introduction. Birkhäuser, Boston (1997)
6. Hackbusch, W.: Iterative solution of large sparse systems of equations. Springer, New York
(1994)
7. Hahn, W.: Stability of Motion. Springer, Berlin (1967)
8. Hairer, E., Nørsett, S.P., Wanner, G.: Solving Ordinary Differential Equations I, 2nd ed. North-
Holland, Amsterdam (1993)
9. Hairer, E., Wanner, G.: Solving Ordinary Differential Equations II. Springer, Berlin (1991)
10. Heun, K.: Neue Methoden zur approximativen Integration der Differentialgleichungen einer
unabhängigen Veränderlichen. Z. für Math. und Phys. 45, 23–38 (1900)
11. Heuser, H.: Gewöhnliche Differentialgleichungen. Teubner, Stuttgart (1991)
12. Kaluza, T.: Über Koeffizienten reziproker Potenzreihen. Math. Z. 28, 161–170 (1928)
13. Kreiss, H.O.: Über die Stabilitätsdefinition für Differenzengleichungen die partielle Differen-
tialgleichungen approximieren. BIT 2, 153–181 (1962)
14. Kunkel, P., Mehrmann, V.: Differential-algebraic equations - analysis and numerical solution.
EMS, Zürich (2006)
15. Kutta, W.M.: Beitrag zur näherungsweisen Integration totaler Differentialgleichungen. Z. für
Math. und Phys. 46, 435–453 (1901)
16. Lipschitz, R.O.S.: Lehrbuch der Analysis, Vol. 2. Cohen, Bonn (1880)
17. Morton, K.W.: On a matrix theorem due to H. O. Kreiss. Comm. Pure Appl. Math. 17, 375–
379 (1964)
18. Richtmyer, R.D., Morton, K.W.: Difference Methods for Initial-value Problems, 2nd ed. John
Wiley & Sons, New York (1967). Reprint by Krieger Publ. Comp., Malabar, Florida, 1994
19. Runge, C.D.T.: Ueber die numerische Auflösung von Differentialgleichungen. Math. Ann. 46,
167–178 (1895)
20. Stetter, H.J.: Analysis of Discretization Methods for Ordinary Differential Equations. North-
Holland, Amsterdam (1973)
21. Terrel, W.J.: Stability and Stabilization - An Introduction. Princeton University Press, Prince-
ton and Oxford (2009)
22. Toh, K.C., Trefethen, L.N.: The Kreiss matrix theorem on a general complex domain. SIAM
J. Matrix Anal. Appl. 21, 145–165 (1999)
23. Werner, H., Arndt, H.: Gewöhnliche Differentialgleichungen. Eine Einführung in Theorie und
Praxis. Springer, Berlin (1986)
Chapter 6
Instationary Partial Differential Equations
The analysis presented in this chapter evolved soon after 1950, when discretisations
of hyperbolic and parabolic differential equations had to be developed. Most of the
material can be found in Richtmyer–Morton [21], see also Lax–Richtmyer [16].
All results concern linear differential equations. In the case of hyperbolic equa-
tions there is a crucial difference between linear and nonlinear problems, since in
the nonlinear case many unpleasant features may occur that are unknown in the
linear case. Concerning general hyperbolic conservation laws, we refer, e.g., to
Kröner [14] and LeVeque [19]. However, even in the nonlinear case, the linearised
problems should satisfy the stability conditions described here.
6.1 Introduction and Examples
6.1.1 Notation, Problem Setting, Function Spaces
Replacing the scalar ordinary differential equation by a linear system of ordinary

differential equations, one obtains y 0 = Ay + f , where—in the simplest case—
A is a constant N × N matrix, and y and f have values in RN (or CN ). In this
case, the previous results concerning one-step and multistep methods can easily
be generalised. The situation changes if the (bounded) matrix A is replaced by an
unbounded differential operator, as we do now.
The independent variable x of the ordinary differential equation is renamed by t
(‘time’ variable). The differential operator A contains differentiations with respect
to spatial variables x1 , . . . , xd . Although d = 3 is the realistic case, we restrict our
considerations mainly to d = 1. The case d > 1 will be discussed in §6.5.6.1.
Notation 6.1. The desired solution is denoted by u (instead of y). The independent
variables are t and x. The classical notation for u depending on t and x is u(t, x).
Let B be a space of functions in the variable x. Then u(t) denotes the function
u(t, ·) ∈ B (partially evaluated at t). Therefore, u(t, x) and u(t)(x) are equivalent

94 6 Instationary Partial Differential Equations
notations. Let I = [0, T ] be the time interval in which t varies. If DA ⊂ B is the

domain1 of the differential operators A, the partial differential equation takes the
following form: Find a continuous function u : I → DA ⊂ B such that
∂
u(t) = Au(t) for all t ∈ I. (6.1a)
∂t
The initial-value condition is given by
u(0) = u0 for some u0 ∈ DA ⊂ B. (6.1b)
Concerning the differential operator A, we discuss two model cases:
∂ ∂2
A := a (a 6= 0) and A := a (a > 0). (6.2)
∂x ∂x2
In what follows, the domain of u(·, ·) is the set
Σ =I ×R with I = [0, T ]. (6.3)
Here the time t varies in I = [0, T ], while the spatial variable x varies in R.
I corresponds to the interval I = [x0 , xE ] from §5.1. The spatial domain R is
chosen as the unbounded domain to avoid boundary conditions.2
We restrict our considerations to two Banach spaces, generally denoted by B:
• B = C(R), space of the complex-valued, uniformly3 continuous functions with
finite supremum norm kvkB = kvk∞ = sup{|v(x)| : x ∈ R}.
• B = L2 (R), space of the complex-valued, measurable, and square-integrable
functions. This means that the L2 norm
sZ
2
kvkB = kvk2 = |v(x)| dx
R
is finite. This Banach space is also a Hilbert space with the scalar product
Z
(u, v) := u(x)v(x)dx.
R
For both cases of (6.2) we shall show that the initial-value problem is solvable.
1
The domain of a differential operator A is DA = {v ∈ B : Av ∈ B is defined}. Often, it
suffices to choose a smaller, dense set B0 ⊂ DA and to extend the results continuously onto DA .
2
A similar situation arises if the solutions are assumed to be 2π-periodic in x. This corresponds
to the bounded domain Σ = I × [0, 2π] with periodic boundary condition u(t, 0) = u(t, 2π). In
the 2π-periodic case, the spaces are
Cper (R) := {v ∈ C(R) : v(x) = v(x + 2π) for all x ∈ R},
L2per (R) := {v ∈ L2 (R) : v(x) = v(x + 2π) for almost all x ∈ R}.
3
Note that limits of uniformly continuous functions are again uniformly continuous.
6.1 Introduction and Examples 95
6.1.2 The Hyperbolic Case A = a∂/∂x
∂
First we choose B = C(R). The domain of A = a ∂x is the subspace B0 = C 1 (R),
4
which is dense in B. The partial differential equation
∂ ∂
u=a u
∂t ∂x
is of hyperbolic type.5 The solution of problem (6.1a,b) can be described directly.
Lemma 6.2. For any u0 ∈ B0 = C 1 (R), the unique solution of the initial-value
problem (6.1a,b) is given by u(t, x) := u0 (x + at).
∂
Proof. (i) By ∂t ∂
u = ∂t u0 (x+at) = au00 (x+at) and a ∂x ∂
u0 (x+at) = au00 (x+at),
the differential equation (6.1a) is satisfied. The initial value is u(0, x) := u0 (x).
(ii) Concerning uniqueness, we transform to the (characteristic) direction:
ξ = x + at, τ = t, U (τ, ξ) := u(t(τ, ξ), x(τ, ξ))
with the inverse transformation t(τ, ξ) = τ , x(τ, ξ) = ξ − aτ . The chain rule

∂ ∂ ∂ ∂ ∂ ∂ ∂
yields ∂τ U = ∂t u ∂τ t + ∂x u ∂τ x = ∂t u − a ∂x u = 0. With respect to the new
∂
variables, Eq. (6.1a) becomes the ordinary differential equation ∂τ U (τ, ξ) = 0 (for
each ξ ∈ R). Hence, there is a unique constant solution U (τ, ξ) = U (0, ξ) =
u(t(0, ξ), x(0, ξ)) = u(0, ξ) = u0 (ξ). tu
In the case of the space B = L2 (R), we choose B0 := C0∞ (R) ⊂ DA (or

B0 := C 1 (R) ∩ L2 (R)) as a dense subspace in B. Here C0∞ (R) is the set of all
infinitely often differentiable functions with compact support (cf. Footnote 13 on
page 56). Lemma 6.2 holds also for this B0 .
Remark 6.3. Let t ≥ 0 be arbitrary. If the initial value u0 belongs to C 1 (R) or

C0∞ (R), then also the solution u(t) belongs to C 1 (R) or C0∞ (R), respectively.
Furthermore,
ku(t)kB = ku0 kB
holds for k·kB = k·k∞ as well as for k·kB = k·k2 .
Proof. Since u(t) is a shifted version of u0 and a shift does not change the norm
k·kB , the assertions follow. t
u
Exercise 6.4. Extend the previous statement to any Lp norm

sZ
p p
kukp = |u(x)| dx (1 ≤ p < ∞).
R
4
The case of a = 0 is exceptional, since then ut = 0 can be considered as a family of ordinary
differential equations for each x ∈ R. Hence, the theory of §5 applies again.
5
Concerning the definition of types of partial differential equations, see Hackbusch [8, §1].
6.1.3 The Parabolic Case A = ∂ 2 /∂x2

2
∂
The parabolic differential equation ∂t u = a ∂∂xu2 is called the heat equation, since it
describes the evolution of the temperature u as a function of time t and space x in
the case an infinite wire (one-dimensional case!). The factor a > 0 is the thermal
√ If a is constant, we may assume a = 1 (otherwise, transform by t 7→ at
conductivity.
or x 7→ ax):
∂u ∂2u
= for t > 0. (6.4)
∂t ∂x2
Lemma 6.5. The solution of (6.4) with a continuous initial value (6.1b) is given by
Z ∞
1 2

u(t, x) = √ u0 (ξ) exp −(x−ξ)4t dξ for t > 0 and x ∈ R. (6.5)
4πt −∞
2
1
Proof. (i) One verifies that √4πt exp −(x−ξ)
4t is a solution of (6.4) for each ξ ∈ R
and t > 0. Since the integrand in (6.5) decays exponentially, one can interchange
integration and differentiation, and obtains that u from (6.5) satisfies (6.4).
(ii) It remains to prove that limt&0 u(t, x) = u0 (x). Note that
Z ∞ 2
1 −ζ
√ exp dζ = 1 for t > 0 (6.6)
4πt −∞ 4t
(cf. [27, p. 187]). Substitution ζ = ξ − x yields

∞ 2
− (x − ξ)
Z
1
u(t, x) = √ u0 (ξ) exp dξ
4πt −∞ 4t
∞ 2
− (x − ξ)
Z
1
= u0 (x) + √ [u0 (ξ) − u0 (x)] exp dξ.
4πt −∞ 4t
If we can prove that the last term tends to zero as t & 0, the desired statement
limt&0 u(t, x) = u0 (x) follows.
Let x and ε > 0 be fixed. Because of continuity of u0 , there is a δ > 0 such
that |u0 (ξ) − u0 (x)| ≤ ε/2 for all |ξ − x| ≤ δ. We split the integral into the sum of
three terms:
Z x+δ 2
1 − (x − ξ)
I1 (t, x) := √ [u0 (ξ) − u0 (x)] exp dξ,
4πt x−δ 4t
Z x−δ 2
1 − (x − ξ)
I2 (t, x) := √ [u0 (ξ) − u0 (x)] exp dξ,
4πt −∞ 4t
Z ∞ 2
1 − (x − ξ)
I3 (t, x) := √ [u0 (ξ) − u0 (x)] exp dξ.
4πt x+δ 4t
The first integral is bounded by

6.1 Introduction and Examples 97
x+δ 2
− (x − ξ)
Z
1
|I1 (t, x)| ≤ √ |u0 (ξ) − u0 (x)| exp dξ
4πt x−δ 4t
Z x+δ 2
ε/2 − (x − ξ)
≤√ exp dξ
4πt x−δ 4t
Z ∞ 2
ε/2 − (x − ξ) ε
≤√ exp dξ = .
4πt −∞ 4t (6.6) 2
Set C := supx∈R |u0 (x)| < ∞. Then I2 is bounded by

x−δ 2
− (x − ξ)
Z
1
|I2 (t, x)| ≤ √ |u0 (ξ) − u0 (x)| exp dξ
4πt −∞ 4t
Z x−δ 2
2C − (x − ξ)
≤√ exp dξ
4πt −∞ 4t
Z ∞
2C
exp −τ 2 dτ.

= √ √ √
τ =(x−ξ)/ 4t π δ/ 4t
R∞ R∞
Since the improper integral −∞ exp −τ 2 dτ exists, R exp −τ 2 dτ tends to
zero as R → ∞. Therefore, a sufficiently small t > 0 yields |I2 (t, x)| ≤ 4ε . We
obtain the same bound |I3 (t, x)| ≤ 4ε for I3 . Together,

1 Z ∞ 2
− (x − ξ) ε ε ε
√ [u0 (ξ) − u0 (x)] exp dξ ≤ + + = ε

4πt −∞ 4t 2 4 4
holds for sufficiently small t > 0. As x and ε are arbitrarily chosen,

∞ 2
− (x − ξ)
Z
1
lim √ [u0 (ξ) − u0 (x)] exp dξ = 0
t&0 4πt −∞ 4t
is proved for all x. t

u
The representation (6.5) shows that the solution u(t) at t > 0 is infinitely often
differentiable, although the initial value u0 is only continuous. However, the solution
exists only for t > 0, not for t < 0. Note the different property in the hyperbolic
case, where the representation of the solution from Lemma 6.2 holds for all t ∈ R.
In the hyperbolic case, the norm ku(t)kB is independent of t, whereas in the
parabolic case, only a monotonicity statement holds.
Lemma 6.6. Let u(t) ∈ B = C(R) be a solution of (6.4). Then the inequality
ku(t)k∞ ≤ ku(0)k∞ holds for all t ≥ 0.
Proof. Set C := ku(0)k∞ and let t > 0. From (6.5) we infer that
∞ 2
− (x − ξ)
Z
1
|u(t, x)| ≤ √ |u0 (ξ)| exp dξ
4πt −∞ 4t
∞ 2
− (x − ξ)
Z
C
≤√ exp dξ = C ;
4πt −∞ 4t (6.6)
hence, ku(t)k∞ ≤ C. t
u
Lemma 6.7. Let u(t) ∈ B = L2 (R) be a solution of (6.4). Then ku(t)k2 ≤ ku(0)k2
holds for all t ≥ 0.
Proof. It is sufficient to restrict to u(0) ∈ DA . One concludes either from (6.5) or
2
from general considerations (cf. (6.7e)) that ∂∂xu2 ∈ L2 (R) for t ≥ 0, so that the
00 0
following integrals exist. Let t ≥ t ≥ 0. Because of
Z Z Z Z t00
00 0 ∂
2 2
u(t0 , x)2 dtdx

u(t , x) dx − u(t , x) dx =
R R R t0 ∂t
t00 t00
∂ 2 u(t, x)
Z Z Z Z
∂u(t, x)
=2 u(t, x) dtdx = 2 dtdx u(t, x)
R t0 ∂t R t0 ∂x2
Z t00 Z Z t00 Z 2
∂ 2 u(t, x) ∂u(t, x)
=2 u(t, x) dxdt = −2 dxdt ≤ 0,
t0 R ∂x2 t0 R ∂x
2
ku(t)k2 is weakly decreasing. t
u
6.2 Semigroup of Solution Operators
In the following, B0 = C ∞ (R) ∩ B is chosen as a dense subset of B. We know that

ut = Au has a classical solution for any initial value u0 ∈ B0 . A solution is called
classical or strong if ut (t) and Au(t) are elements of B for all t ≥ 0. We recall that
ku(t)kB ≤ ku0 kB has been proved.
Obviously, the mapping from u0 ∈ B0 into the solution u(t) at some fixed t ≥ 0
is linear. This defines the solution operator
T (t) : u0 7→ u(t) for t ≥ 0. (6.7a)
So far, T (t) is defined on B0 . From ku(t)kB ≤ ku0 kB and the density of B0 we

infer that T (t) is a bounded operator and can be extended uniquely and continuously
onto B.
Remark 6.8. Suppose that T (t) ∈ L(B0 , B) and that B0 is dense in B. Then T (t)
can be extended uniquely and continuously onto B. The extended T (t) ∈ L(B, B)
and the original mapping T (t) ∈ L(B0 , B) have equal operator norms; i.e.,
sup{kT (t)vkB / kvkB : 0 6= v ∈ B0 } = sup{kT (t)vkB / kvkB : 0 6= v ∈ B}.
Proof. Let u0 ∈ B. There is a sequence v0,n ∈ B0 with limn→∞ ku0 − v0,n kB = 0.
One verifies that vn := T (t)v0,n forms a Cauchy sequence, and therefore it defines
a unique u := limn→∞ vn . The definition T (t)u0 =: u defines the continuous
extension T (t) ∈ L(B, B) of the desired form. t u
6.2 Semigroup of Solution Operators 99
Because of
T (t) ∈ L(B, B), (6.7b)
any initial value u0 ∈ B leads to a function u(t) := T (t)u0 ∈ B. If u0 ∈ B\B0 ,
the resulting function is called a ‘generalised’ or ‘weak’ solution in contrast to the
strong solution mentioned above. Note that the descriptions u(t, x) := u0 (x + at)
from Lemma 6.2 and of u(t, x) by (6.5) make sense for any u0 ∈ B.
Next, we show the semigroup property
T (t)T (s) = T (t + s) for all t, s ≥ 0. (6.7c)
For this purpose consider the strong solution u(τ ) = T (τ )u0 for an initial value
u0 ∈ B0 and fix some s ≥ 0. Then û0 := u(s) equals T (s)u0 . Set û(t) := u(t + s).
Since ût (t) = ut (t + s) = Au(t + s) = Aû(t) and û(0) = u(s) = û0 , we conclude
that û(t) = T (t)û0 ; i.e., T (t + s)u0 = u(t + s) = û(t) = T (t)û0 = T (t)T (s)u0
holds for all u0 ∈ B0 . Since B0 is dense in B, the identity (6.7c) follows.
The operators A and T (t) commute:
AT (t) = T (t)A on DA (6.7d)
(see Footnote 1 for the domain DA of A). For a proof, consider the strong solution
u(t) = T (t)u0 for any u0 ∈ B0 . The third line in
AT (t)u0 = Au(t) = ut (t)

u(t + h) − u(t) 1
= lim = lim [T (t + h) − T (t)] u0
h&0 h h&0 h

u(h) − u(0)
= lim T (t) [T (h) − I] u0 = T (t) lim
h&0 h&0 h
= T (t)ut (0) = T (t)Au(0) = T (t)Au0
uses the continuity of T (t) (cf. (6.7b)). Since T (t)Au0 ∈ B is defined for all
u0 ∈ DA , also AT (t) has this property. This proves6
T (t) : DA 7→ DA for t ≥ 0. (6.7e)
The definition of T (t) yields
T (0) = I (identity) (6.7f)
for t = 0; i.e., {T (t) : t ≥ 0} is a semigroup with neutral element. It is called

the semigroup generated by A. The generating operator A can be regained via
Av = limt&0 [(T (t)v − v) /t] for all v ∈ DA . Another notation for T (t) is
etA = exp(tA).
6
The semigroups of hyperbolic and parabolic problems have different properties. In the parabolic
case, T (t) : DA 7→ B holds for positive t (even T (t) : DA → C ∞ (R) ∩ B) as can be seen
from (6.5), whereas in the hyperbolic case the smoothness does not improve with increasing t.
Finally, we discuss the behaviour of generalised solutions u(t) = T (t)u0 with

u0 ∈ B for t → 0. In Lemma 6.5 u0 is assumed to be continuous (not necessarily
uniformly continuous) and point-wise convergence u(t, x) → u0 (x) for all x ∈ R is
shown. Assuming uniform continuity, we obtain uniform convergence by the same
proof; i.e., u(t) = T (t)u0 → u0 in B as t → 0. Under the resulting property
lim k[T (t) − T (0)] u0 kB → 0 for all u0 ∈ B the semigroup is called continuous.
The same situation happens in the hyperbolic case for any u0 ∈ C(R).
Above we used the inequality ku(t)kB ≤ ku0 kB for t ≥ 0. The constant one can
be replaced by any bound and the range t ≥ 0 may be reduced to any interval [0, τ ],
τ > 0.
Exercise 6.9. Suppose that Kτ := sup0≤t≤τ kT (t)kB←B < ∞ for some τ > 0.
dt/τ e
Prove kT (t)kB←B ≤ Kτ for any t ≥ 0, where dxe := min{n ∈ Z : x ≤ n}.
In the next subsections, we shall refer to the following inequality (6.8), which
∂ ∂2
holds with KT = 1 for the model examples A = a ∂x and A = ∂x 2.
Assumption 6.10. Let {T (t) ∈ L(B, B) : t ≥ 0} be a semigroup with neutral

element T (0). Moreover, T (t) is supposed to be uniformly bounded on I = [0, T ]:
kT (t)kB←B ≤ KT for all t ∈ I = [0, T ]. (6.8)
So far, only the homogeneous equation ut (t) = Au(t) has been studied.
∂
Remark 6.11. The solution of the inhomogeneous equation ∂t u(t) = Au(t) + f (t)
can be represented by
Z t
u(t) = T (t)u0 + T (t − s)f (s)ds.
0
6.3 Discretisation of the Partial Differential Equation
6.3.1 Notations
We replace the real axis R of the x variable by an infinite grid of step size ∆x > 0:
G∆x = {x = ν∆x : ν ∈ Z}. (6.9)
Correspondingly, the interval I = [0, T ] is replaced by a finite grid of step size

∆t > 0:
I∆t = {t = µ∆t ≤ T : µ ∈ N0 }.
The Cartesian product of both grids yields the rectangular grid
∆t
Σ∆x := I∆t × G∆x = {(t, x) ∈ Σ : x/∆x ∈ Z, t/∆t ∈ N0 } (6.10)
6.3 Discretisation of the Partial Differential Equation 101
(cf. (6.3)). As we shall see, the step sizes ∆x, ∆t are, in general, not chosen inde-
pendently, but are connected by a parameter λ (the power of ∆x corresponds to the
order of the differential operator A):
∂
∆t/∆x in the hyperbolic case A = a ∂x ,
λ= 2 ∂ 2 (6.11)
∆t/∆x in the parabolic case A = ∂x2 .
∆t
For a grid function U : Σ∆x → C we use the notation
Uνµ := U (µ∆t, ν∆x) ∆t

for (µ∆t, ν∆x) ∈ Σ∆x , µ ∈ N0 , ν ∈ Z.
All grid values at t = µ∆t are collected in
U µ := (Uνµ )ν∈Z .
Let ` = CZ be the linear space of two-sided infinite sequences with component-

wise addition and multiplication by scalars. The Banach spaces B = C(R) (or
L∞ (R)) and B = L2 (R) correspond to the sequence spaces `∞ and `2 :
• `∞ = CZ with the norm kU k`∞ =qsup{|Uν | : ν ∈ Z} is a Banach space,
2
• `2 = CZ with the norm kU k`2 = ∆x ν∈Z |Uν | is a Hilbert space.
P
As a common symbol we use `p (p ∈ {2, ∞}).
6.3.2 Transfer Operators r, p
The continuous Banach space B and the discrete space `p of grid functions are
connected via
r = r∆x : B → `p . (6.12a)
The letter r means ‘restriction’. The index ∆x will be omitted, when the underlying
step size is known.
In the case of B = C(R), an obvious choice of r is the evaluation at the grid
points of G∆x :
u ∈ C(R) ⇒ ru ∈ `∞ with (ru)j = u(j∆x) for j ∈ Z. (6.12b)
In the case of B = L2 (R) this is impossible, since L2 functions have no well-

defined point evaluations. Instead, one can use mean values:
Z (j+1/2)∆x
2 2 1
u ∈ L (R) ⇒ ru ∈ ` with (ru)j = u(x)dx for j ∈ Z.
∆x (j−1/2)∆x
(6.12c)
We suppose that r = r∆x : B → `p is bounded:
kr∆x k`p ←B ≤ Cr for all ∆x > 0. (6.13)

Lemma 6.12. The restrictions (6.12b,c) satisfy condition (6.13) with Cr = 1 with
respect to the respective norms k·k`∞ ←C(R) and k·k`2 ←L2 (R) .
Proof. (i) kruk`∞ = sup{|u(j∆x)| : j ∈ Z} ≤ sup{|u(x)| : x ∈ R} = kukC(R)

yields (6.12b).
P R (j+1/2)∆x 2
2 2 1
(ii) kruk`2 = kruk2 = ∆x j | (ru)j |2 = ∆x
P
j (j−1/2)∆x u(x)dx and

Schwarz’ inequality
Z 2 "Z # "Z #
(j+1/2)∆x (j+1/2)∆x (j+1/2)∆x
2
u(x)dx ≤ 1dx · |u(x)| dx

(j−1/2)∆x (j−1/2)∆x (j−1/2)∆x
Z (j+1/2)∆x
2
= ∆x |u(x)| dx
(j−1/2)∆x
2 P (j+1/2)∆x
R 2 R 2 2
yield (6.12c): kruk`2 ≤ |u(x)| dx = |u(x)| dx = kukL2 (R) . t
u
j (j−1/2)∆x R
The ‘prolongation’ p acts in the reverse direction: p = p∆x : `p → B. We

suppose the existence of a prolongation p with the following properties:
kp∆x kB←`p ≤ Cp for all ∆x > 0 and rp = I. (6.14)
The condition rp = I indicates that p is a right-inverse of r.
Exercise 6.13. Verify that in the cases (6.12b,c) the following choices of p satisfy
condition (6.14) with Cp = 1:
p is piecewise linear interpolation: (6.15a)
∞
v∈` 7→ pv ∈ C(R) with (pv) (x) = ϑvj + (1 − ϑ) vj+1 ,
where x = (j + ϑ) ∆x, j ∈ Z, ϑ ∈ [0, 1),
or
p is piecewise constant interpolation: (6.15b)
2 2
v ∈ ` 7→ pv ∈ L (R) with (pv) (x) = vj ,
where x ∈ (j − 12 )∆x, (j + 12 )∆x , j ∈ Z.

6.3.3 Difference Schemes
Initially, U 0 is prescribed via Uν0 = ru0 (r from (6.12a), u0 from (6.1b)). The
explicit difference scheme
X µ
Uνµ+1 = aj Uν+j (6.16)
j∈Z
6.3 Discretisation of the Partial Differential Equation 103
allows us to compute the next time step U µ+1 from U µ . In practice,

P
j∈Z is a finite
sum; i.e., almost all aj vanish.
∂
Example 6.14. (a) In the hyperbolic case, one may replace ∂t u by the difference
u(t+∆t,x)−u(t,x) ∂u u(t,x+∆x)−u(t,x)
quotient ∆t and A = a ∂x by a ∆x . Solving
u(t + ∆t, x) − u(t, x) u(t, x + ∆x) − u(t, x)

=a
∆t ∆x
with respect to u(t + ∆t, x) and using the relation (6.11), one obtains
a∆t
u(t + ∆t, x) = u(t, x) + u(t, x + ∆x) − u(t, x) ;
∆x
µ
i.e., Uνµ+1 = (1 − aλ) Uνµ + aλUν+1 . (6.17a)
(b) Replacing the right-sided difference u(t,x+∆x)−u(t,x)

∆x by the symmetric differ-
u(t,x+∆x)−u(t,x−∆x)
ence 2∆x , one derives the difference scheme
aλ µ aλ µ
Uνµ+1 = − Uν−1 + Uνµ + U . (6.17b)
2 2 ν+1
u(t+∆t,x)−u(t,x) u(t+∆t,x)−[u(t,x+∆x)+u(t,x−∆x)]/2
(c) A replacement of ∆t by ∆t yields
1 − aλ µ 1 + aλ µ
Uνµ+1 = Uν−1 + Uν+1 . (6.17c)
2 2
2 u(t+∆t,x)−u(t,x)
∂
(d) In the parabolic case of A = ∂x 2 , the difference quotient ∆t for
∂ u(t,x−∆x)−2u(t,x)+u(t,x+∆x) ∂2u
∂t u and the second difference quotient ∆x2 for ∂x2 are
obvious choices and lead together with λ = ∆t/∆x2 from (6.11) to
u(t + ∆t, x) − u(t, x) 2u(t, x) − u(t, x − ∆x) − u(t, x + ∆x)

= ;
∆t ∆x2
i.e.,
µ µ
Uνµ+1 = λUν−1 + (1 − 2λ) Uνµ + λUν+1 . (6.18)
The difference scheme (6.16) describes a linear mapping U µ 7→ U µ+1 and

defines the linear operator
X
C : `p → `p , (CU )ν := aj Uν+j for U ∈ `p . (6.19)
j∈Z
Hence, a shorter form of (6.16) is
U µ+1 = CU µ . (6.20)
Notation 6.15. The difference operator C (and therefore also the coefficients aj )
may depend on the parameter λ and the step size ∆t:
C = C(λ, ∆t).
The dependence on ∆x follows implicitly via (6.11).
An example of a ∆t-depending difference operator is given next.

∂
Example 6.16. Let C(λ) be a suitable difference operator for A = a ∂x . Then the
∂
differential operator A = a ∂x + b can be discretised by
C 0 (λ, ∆t) := C(λ) + ∆t · b;
i.e., a0 from (6.16) is replaced by
a00 := a0 + ∆t · b.
|aj | < ∞, then C(λ, ∆t) ∈ L(`∞ , `∞ ).

P
Exercise 6.17. Prove: (a) If j∈Z
(b) If supj∈Z |aj | < ∞, then C(λ, ∆t) ∈ L(`2 , `2 ).
A µ-fold application of C(λ, ∆t) yields
U µ = C µU 0 U (µ∆t, ν∆x) = C µ U 0

and ν
.
Remark 6.18. So far, the coefficients aj are assumed to be scalars. If the scalar
∂
equation ∂t u = Au for u : I × R → R is replaced by a vector-valued equa-
tion for u : I × R → RN , the coefficients aj in (6.19) become N × N matrices.
The vector-valued case will be discussed in §6.5.5.
The shift operator Ej is defined by
(Ej U )ν := Uj+ν .
Since a shift does not change the `p -norm of U ,
kEj k`p ←`p = 1 (6.21)
holds. The difference operator C from (6.19) can be reformulated as

X
C := aj Ej . (6.22)
j∈Z
We conclude with a numerical computation of the example (6.17b), where we

choose a = 2 and λ = 1; i.e., ∆t = ∆x. The initial value U 0 is the constant value 1
with the exception that at x = 0 we perturb 1 by 1.001. Without this perturbation,
the solution of ut = 2ux is u(t, x) = 1 and also the discrete solution is Uνµ = 1.
We aim at the solution u(1, x) at t = 1. Three computations with the step sizes
6.4 Consistency, Convergence, and Stability 105
1 1
∆t = 1, 10 , 100 are shown. The required number of time steps is 1, 10, or 100,
1/∆t
respectively. The last three rows of the table list the values of Uν = U (1, ν∆x)
for −2 ≤ ν ≤ 2.
u(t, −2∆x) u(t, −∆x) u(t, 0) u(t, ∆x) u(t, 2∆x) ∆t = ∆x

t=0 1.000 1.000 1.001 1.000 1.000
t=1 1.000 1.001 1.001 0.999 1.000 1
t=1 1.045 0.230 0.869 1.770 1.045 1/10
t=1 6.29810 30 −3.754 610 30 6462.5 3.754 610 30 6.29810 30 1/100
Instead of convergence to u(t, x), we observe that completely wrong approxima-

tions are produced. The reason is an obvious instability of scheme (6.17b) (cf. Ex-
ample 6.46b). Note that we may replace the perturbed value 1.001 also by 1+10−16 .
This would change, e.g., 6.29810 30 to 6.29810 17, which is equally wrong.
6.4 Consistency, Convergence, and Stability
6.4.1 Definitions
Again, the restriction r from §6.3.2 is used. Consistency will depend on the under-
lying norm `p , where p = 2, ∞ are the discussed examples corresponding to the
Banach spaces B = L2 (R) and C(R). We say that ‘`p is suited to B’, if the require-
ments (6.13) and (6.14) are satisfied.
Definition 6.19 (consistency). Let B0 ⊂ DA be a dense subset of B. Let `p be

suited to B. The solution of (6.1a,b) is denoted by u(t) = T (t)u0 , u0 ∈ B0 . The
local discretisation error τ is defined by
1
τ (t) = [ru(t + ∆t) − C(λ, ∆t)ru(t)] .
∆t
The difference scheme C(λ, ∆t) is called consistent (with respect to `p ) if for all
u0 ∈ B0 ,
sup {kτ (t)k`p : 0 ≤ t ≤ T − ∆t} → 0 as ∆t → 0.
The latter condition is equivalent to
sup k[rT (∆t) − C(λ, ∆t)r] T (t)u0 k`p := o(∆t) for all u0 ∈ B0 .
0≤t≤T −∆t
Note that the following definition of convergence refers to the whole Banach
space B, not to a dense subspace B0 .
Definition 6.20 (convergence). For all u0 ∈ B, let u(t) = T (t)u0 denote the
generalised solution. The difference scheme C(λ, ∆t) is called convergent (with
respect to `p ), if
kru(t) − C(λ, ∆t)µ ru0 k`p → 0 for ∆t → 0 and µ∆t → t ∈ I = [0, T ].
Definition 6.21 (stability). The difference scheme C(λ, ∆t) is called stable (with
respect to `p ) if
sup{kC(λ, ∆t)µ k`p ←`p : ∆t ≥ 0, µ ∈ N0 , 0 ≤ µ∆t ≤ T } < ∞. (6.23)
If (6.23) holds only for certain values of λ, the scheme is called conditionally stable;
otherwise, the scheme is called unconditionally stable.
In the case of (6.23), the stability constant is defined by
K = K(λ) := sup{kC(λ, ∆t)µ k`p ←`p : ∆t ≥ 0, µ ∈ N0 , 0 ≤ µ∆t ≤ T }.
Instead of ‘stable with respect to `p ’ we say for short ‘`p stable’. Similarly, the
terms ‘`p stability’, ‘`p consistent’, ‘`p consistency’ etc. are used.
6.4.2 Convergence, Stability and Equivalence Theorems
First, we show that consistency and stability—together with some mild technical
assumptions—imply convergence.
Theorem 6.22 (convergence theorem). Suppose (a) to (d):
(a) r bounded with respect to `p (cf. (6.13)),
(b) T (t) satisfies Assumption 6.10,
(c) `p stability of the difference scheme C(λ, ∆t),
(d) `p consistency.
Then the difference scheme is convergent with respect to `p .
Proof. (i) Given an initial value u0 ∈ B0 ⊂ DA (B0 dense subset of B), define
u(t) = T (t)u0 . We split the discretisation error as follows:
ru(t) − C(λ, ∆t)µ ru0 = r [u(t) − u(µ∆t)] + [rT (µ∆t) − C(λ, ∆t)µ r] u0
= r [u(t) − u(µ∆t)] + [rT (∆t)µ − C(λ, ∆t)µ r] u0 .
Use the telescopic sum

µ−1
X
rAµ − B µ r = B ν [rA − Br] Aµ−ν−1
ν=0
with A := T (∆t) and B := C(λ, ∆t). This shows
kru(t) − C(λ, ∆t)µ ru0 k`p ≤ krk`p ←B ku(t) − u(µ∆t)kB +
µ−1
X
kC(λ, ∆t)ν k`p ←`p rT (∆t) − C(λ, ∆t)r u (µ − ν − 1) ∆t `p

+
ν=0
6.4 Consistency, Convergence, and Stability 107
µ−1
X
≤ Kr ku(t) − u(µ∆t)kB + K(λ)∆t τ (µ − ν − 1) ∆t `p
ν=0
with the stability constant K(λ). Since u0 ∈ B0 ⊂ DA , the solution u(t) =

T (t)u0 is strong, therefore differentiable and in particular continuous (cf. (6.7e)):
ku(t) − u(µ∆t)kB → 0 as µ∆t → t. Therefore, the first term is a zero sequence.
Because of the consistency assumption, the local discretisation error τ tends to
zero uniformly. From kτ ((µ − ν − 1) ∆t)k`p ≤ ε we infer that
µ−1
X
K(λ)∆t kτ ((µ − ν − 1) ∆t)k`p ≤ µK(λ)∆tε ≤ K(λ)T ε,
µ∆t≤T
ν=0
so that the whole sum tends to zero. This proves convergence in the case of an initial
value u0 ∈ B0 .
(ii) For a general initial value u0 ∈ B and an arbitrary ε > 0, one finds a u∗0 ∈ B0
with ku0 − u∗0 kB ≤ ε/ [3Kr max {kT (t)kB←B , K(λ)}]. The associated solution
u∗ (t) = T (t)u∗0 satisfies
ε
kr [u(t) − u∗ (t)]k`p ≤ Kr kT (t) [u0 − u∗0 ]kB ≤ Kr kT (t)kB←B ku0 − u∗0 kB ≤
3
and
kC(λ, ∆t)µ ru0 − C(λ, ∆t)µ ru∗0 k`p = kC(λ, ∆t)µ r [u0 − u∗0 ]k`p
≤ K(λ)Kr ku0 − u∗0 kB ≤ ε/3.
Together with kru∗ (t) − C(λ, ∆t)µ ru∗0 k`p ≤ ε/3 from (i) for sufficiently small ∆t
and t − µ∆t, it follows that kru(t) − C(λ, ∆t)µ ru0 k`p ≤ ε, so that also for general
initial values u0 ∈ B convergence is shown. t u
Next, we show that stability is also necessary for convergence.
Theorem 6.23 (stability theorem). Choose B and `p suitably, so that (6.13),
(6.14), and (6.8) hold. Then `p convergence implies `p stability.
Proof. For an indirect proof assume that the difference scheme is unstable. Then
there are sequences ∆tν > 0, µν ∈ N0 with 0 ≤ µν ∆tν ≤ T , so that
kC(λ, ∆tν )µν k`p ←`p → ∞ for ν → ∞.
Since the interval I = [0, T ] is compact, there is a subsequence with µν ∆tν → t ∈ I.

The latter convergence yields
kru(t) − C(λ, ∆tν )µν ru0 k`p → 0 for all u0 ∈ B,
and therefore,
kC(λ, ∆tν )µν ru0 k`p ≤ 1 + kru(t)k`p ≤ 1 + krk`p ←B kT (t)kB←B ku0 kB

u(t)=T (t)u0
≤ 1 + Kr KT ku0 kB =: K1 (u0 ) for ν ≥ ν0

(6.13), (6.8)
with sufficiently large ν0 = ν0 (u0 ). One concludes that Cν := C(λ, ∆tν )µν r
is a point-wise bounded sequence of operators. Corollary 3.39 yields that Cν is
uniformly bounded: there is a constant K with
kC(λ, ∆tν )µν rk`p ←B ≤ K for all ν ∈ N.
Since, according to (6.14), p is a bounded right-inverse of r, it follows that
kC(λ, ∆tν )µν k`p ←`p = kC(λ, ∆tν )µν rpk`p ←`p
≤ kC(λ, ∆tν )µν rk`p ←B kpkB←`p ≤ KKp
(6.14)
in contradiction to the assumption kC(λ, ∆tν )µν k`p ←`p → ∞. t

u
The conclusion from both theorems is the equivalence theorem.
Theorem 6.24 (equivalence theorem). Suppose (6.8), (6.13), (6.14), and `p con-
sistency. Then `p convergence and `p stability are equivalent.
6.4.3 Other Norms
So far, we restrict the analysis to the `p and Lp norms for p = 2 and p = ∞. For
1 ≤ p < ∞, the Lp norm is defined in Exercise 6.4, while `p is defined analogously.
The reason for the restriction to p ∈ {2, ∞} is twofold. If different properties hold
for `2 and `∞ , a more involved analysis is necessary to describe the properties for
`p with 2 < p < ∞. It might be that the separation is between p = 2 and p > 2,
because in the latter case the Hilbert structure is lost. However, it might also happen
that properties change between the cases p < ∞ and p = ∞, since in the first case
`p is reflexive, but not in the latter case.
If stability estimates hold for both cases p = 2 and p = ∞, we are in a very
pleasant situation, since these bounds imply corresponding estimates for the `p and
Lp setting for 2 < p < ∞. This result is based on the interpolation estimate by
Riesz–Thorin. It is proved by Marcel7 Riesz [23] in8 1926/27, while Thorin [25]
(1939) simplified the proof. In the following lemma, let k·kp←p be the operator
norm of L(`p , `p ) or L(Lp , Lp ).
Lemma 6.25 (Riesz–Thorin theorem). Assume 1 ≤ p1 ≤ q ≤ p2 ≤ ∞. Then
α β q − p1 p2 − q
k·kq←q ≤ k·kp1 ←p1 k·kp2 ←p2 with α = , β= .
p2 − p1 p2 − p1
7
Marcel Riesz is the younger brother of Frigyes Riesz, who is the author of, e.g., [22].
8
The article belongs to Volume 49, which is associated to the year 1926. However, it is the last
article of that volume and carries the footnote ‘Imprimé le 11 janvier 1927’. Therefore, one finds
also 1927 as the publication date.
6.5 Sufficient and Necessary Conditions for Stability 109
Note that α + β = 1. This implies that p1 stability (i.e., k. . .kp1 ←p1 ≤ M1 ) and
p2 stability k. . .kp2 ←p2 ≤ M2 imply q stability in the form
k. . .kq←q ≤ M := M1α M2β .
Hence, if a criterion yields both `2 and `∞ stability, then `p stability holds for all
2 ≤ p ≤ ∞.
6.5 Sufficient and Necessary Conditions for Stability
6.5.1 First Criteria
The following results belong to the classical stability theory of Lax–Richtmyer [16].
Criterion 6.26. If
kC(λ, ∆t)k`p ←`p ≤ 1 + Kλ ∆t for all ∆t > 0,
then the difference scheme is `p stable with stability constant K(λ) := eT Kλ .
Proof. Use
µ µ
kC(λ, ∆t)µ k`p ←`p ≤ kC(λ, ∆t)k`p ←`p ≤ (1 + Kλ ∆t)
µ µ
and apply Exercise 3.24a: (1 + Kλ ∆t) ≤ eKλ ∆t = eKλ µ∆t ≤ eKλ T . t
u
µ∆t≤T
The coefficients aj of C(λ, ∆t) can be used to estimate kC(λ, ∆t)k`p ←`p .
P
Remark 6.27. The difference scheme (6.19) satisfies kC(λ, ∆t)k`p ←`p ≤ |aj | .
P P P
Proof. kC(λ, ∆t)k`p ←`p = j aj Ej ≤ j |aj | kEj k`p ←`p = j |aj |

`p ←`p
follows from (6.22) and (6.21). tu
Combining the previous statements, we arrive at the next result.

P
Corollary 6.28. If j |aj | ≤ 1 + Kλ ∆t for all ∆t > 0, the difference scheme is
stable (with respect to the `2 and `∞ norms) with stability constant K(λ) := eT Kλ .
Definition 6.29 (positive difference scheme). A difference scheme (6.19) is called

positive if aj ≥ 0 holds for all coefficients.
Positive difference schemes map non-negative initial values U 0 ∈ `p into non-

negative solutions U µ , which often is an appreciated property. From Corollary 6.28
we infer the next one.
P
Criterion 6.30. Positive difference schemes (6.19) with j aj = 1 + O(∆t) are
stable (with respect to the `2 and `∞ norms).
We can also formulate a criterion for instability.

P
Criterion 6.31. Suppose that aj ≥ 1 + ∆t c(∆t) with lim∆t→0 c(∆t) = ∞.
Then the difference scheme (6.19) is unstable (with respect to `2 and `∞ norms).
Proof. (i) In the case of `∞ , choose the constant grid function U 0 ∈ `∞P , i.e.,
Uj0 = 1 for all j ∈ Z. Then U 1 = C(λ, ∆t)U 0 equals ζU 0 with ζ := aj
and, correspondingly, U µ = C(λ, ∆t)µ U 0 = ζ µ U 0 . Hence, kC(λ, ∆t)µ k`p ←`p ≥
µ
ζ µ ≥ [1 + ∆tc(∆t)] . Exercise 6.32a proves the assertion.
(ii) In the case of `2 , the previous proof cannot be repeated, since U 0 ∈
/ `2 .
Instead, the proof will be given after Theorem 6.44 on page 118. t u
Exercise 6.32. Suppose that lim∆t→0 c(∆t) = ∞. (a) Prove

µ
sup{[1 + ∆t c(∆t)] : µ ∈ N0 , ∆t > 0, µ∆t ≤ T } = ∞.
for any constant K > 0 and √

(b) Show that √ all positive integers µ with µ ≤ T /∆t,
the√µ-th root µ K satisfies the inequality µ K ≤ 1 + Cρ ∆t. Hint: set c(∆t) :=
µ
K − 1 /∆t and prove that Cρ := sup∆t>0 c(∆t) < ∞.
Next, we apply the criteria to the examples in §6.3.3.
Example 6.33. (a) In (6.17a), the non-zero coefficients are a0 = 1 − aλ and

a1 =Paλ. Assume a ≥ 0. For λ with 0 ≤ aλ ≤ 1, (6.17a) is a positive scheme
with aj = 1; hence, it is `2 and `∞ stable according to Corollary 6.28.
(b) In (6.17b), the non-zero coefficients are a−1 = − aλ aλ
2 , a0 = 1, and a1 = 2 .
case a = 0 (cf. Footnote 4 on page 95), we cannot apply
Excluding the trivial P
Corollary 6.28, since |aj | = 1 + aλ > 1. (6.17b) is not a positive scheme.
(c) In (6.17c), the non-zero coefficients are a−1 = 1−aλ
2 and a1 = 1+aλ
2 . Under the
conditionP |aλ| ≤ 1 the scheme is positive, and Criterion 6.30 ensures `2 and `∞
stability: aj = 1.
(d) In (6.18), the non-zero coefficients are a−1 = λ, a0 = 1 − 2λ, a1 = λ. For
λ ∈ (0, 1/2] the scheme is positive, and Criterion 6.30 ensures `2 and `∞ stability.
An interesting question is whether a stable scheme remains stable after a pertur-

bation (cf. §3.5.3).
Lemma 6.34 (perturbation lemma). Let C(λ, ∆t) be `p stable with stability con-
stant K(λ). Suppose that a perturbation D(λ, ∆t) is bounded by
kD(λ, ∆t)k`p ←`p ≤ CD ∆t.
Then
C 0 (λ, ∆t) := C(λ, ∆t) + D(λ, ∆t)
is again `p stable with stability constant 9

K 0 (λ) ≤ K(λ)eK(λ)CD T ,
where I = [0, T ] is the given time interval. This result holds also in the case that
C(λ, ∆t) and D(λ, ∆t) are non-commutative10 operators.
Proof. We have to estimate C 0 (λ, ∆t)µ . The simple binomial formula is only valid
for commutative terms. In the general case, we have
Xµ X
µ
[C + D] = C α1 DC α2 D · · · C αm DC αm+1 ,
m=0 α1 +...+αµ+1 =µ−m
where the second sum runs over all αj ∈ N0 with α1 + . . . + αµ+1 = µ − m. Each
term C α1 DC α2 D · · · C αm DC αm+1 contains m factors
D and µ − m factors C.
µ
For a fixed m ∈ [0, µ] the number of these terms is m . Together with the estimate
kC α1 DC α2 D · · · C αµ DC αµ+1 k
m
≤ kC α1 k kDk kC α2 k kDk · · · kC αm k kDk kC αm+1 k ≤ K(λ)m+1 (CD ∆t)
we arrive at the inequality
kC 0 (λ, ∆t)µ k`p ←`p =

µ
X
µ µ
m
= k[C(λ, ∆t) + D(λ, ∆t)] k`p ←`p ≤ m K(λ)m+1 (CD ∆t)
m=0
µ
X
µ
m µ
= K(λ) m (K(λ)CD ∆t) = K(λ) [1 + K(λ)CD ∆t] .
m=0
µ
with [1 + K(λ)CD ∆t] ≤ eK(λ)CD µ∆t ≤ eK(λ)CD T (cf. Exercise 3.24). t
u
µ∆t≤T
Remark 6.35. (a) A simple application of Lemma 6.34 is the following one. Let
∂
the differential operator A in ∂t u = Au (cf. (6.1a)) be A = A1 + A0 , where
A1 contains derivatives of at least first order, while A0 u = a0 u is the term of order
zero. The discretisation yields correspondingly C(λ, ∆t) = C1 (λ, ∆t)+C0 (λ, ∆t).
A consistent discretisation of C0 satisfies the estimate kC0 (λ, ∆t)k`p ←`p = O(∆t).
By Lemma 6.34, the stability of C1 (λ, ∆t) implies the stability of C(λ, ∆t). There-
fore, it suffices to investigate differential operators A without terms of order zero.
(b) Let A = A1 be the principal part from above. The property A1 = 0 (1 ∈ `∞ :
constant function with value one) shows that u = 1 is a solution. This implies the
special consistency condition
X
aj = 1 (= I in the matrix-valued case ) (6.24)
j∈Z
for the coefficients of C(λ, ∆t) (cf. (6.19)).

9
If commutativity CD = DC holds, the stability constant improves: K 0 (λ) ≤ K(λ)eCD T .
10
Non-commutativity occurs for matrix-valued coefficients aj (cf. §6.5.5).
Next we need the spectral radius11
ρ(A) := sup{|λ| : λ singular value of A}. (6.25)
Criterion 6.36. A necessary condition for stability (with respect to `2 and `∞ ) is
ρ(C(λ, ∆t)) ≤ 1 + O(∆t).
Proof. ρ(Aµ ) = ρ(A)µ holds for µ ∈ N. On the other hand, ρ(A) ≤ kAk is valid
for any associated norm. Therefore, stability yields
ρ(C(λ, ∆t))µ = ρ(C(λ, ∆t)µ ) ≤ kC(λ, ∆t)µ k`p ←`p ≤ K(λ)
for all µ, ∆t with µ∆t ≤ T . Exercise 6.32b proves ρ(C(λ, ∆t)) ≤ 1 + Cρ ∆t. t
u
Remark 6.37. Suppose that the operator C(λ, ∆t) ∈ L(`2 , `2 ) is normal; i.e.,
C(λ, ∆t) commutes with the adjoint operator C(λ, ∆t)∗ . Then ρ(C(λ, ∆t)) =
kC(λ, ∆t)k`2 ←`2 holds and `2 stability is equivalent to ρ(C(λ, ∆t)) ≤ 1 + O(∆t).
Proof. (i) Let C be a normal operator. First we prove kC 2 k`2 ←`2 = kC ∗ Ck`2 ←`2 .
From
hCCu, CCui`2 = hC ∗ CCu, Cui`2 = hCC ∗ Cu, Cui`2 = hC ∗ Cu, C ∗ Cui`2
it follows that
kC 2 k2`2 ←`2 = sup hCCu, CCui`2 = sup hC ∗ Cu, C ∗ Cui`2 = kC ∗ Ck2`2 ←`2 .
kuk`2 =1 kuk`2 =1
Since also
kC ∗ Ck`2 ←`2 = sup hu, C ∗ Cvi`2 = sup hCu, Cvi`2 ≥

kuk`2 =kvk`2 =1 kuk`2 =kvk`2 =1 u=v
≥ sup hCu, Cui`2 = kCk2`2 ←`2 ,

kuk`2 =1
kCk2`2 ←`2 ≥ kC 2 k`2 ←`2 is shown. Because of kC 2 k`2 ←`2 ≤ kCk2`2 ←`2 (submulti-
plicativity of the operator norm), the equality kC 2 k`2 ←`2 = kC ∗ Ck`2 ←`2 is proved.
Analogously, kCkn`2 ←`2 = kC n k`2 ←`2 follows for all n = 2k (k ∈ N).
p
(ii) The characterisation ρ(C) = limn→∞ n kC n k`2 ←`2 together with (i) proves
ρ(C) = kCk`2 ←`2 .
(iii) According to Criterion 6.36, ρ(C(λ, ∆t)) ≤ 1 + O(∆t) is necessary,
while kC(λ, ∆t)k`2 ←`2 ≤ 1 + O(∆t) is sufficient (cf. Criterion 6.26). Since
ρ(C(λ, ∆t)) = kC(λ, ∆t)k`2 ←`2 , both inequalities are identical. t
u
11
λ is regular value of A, if λI − A is bijective and the inverse (λI − A)−1 ∈ L(B, B)
exists. Otherwise, λ is a singular value of A. In the case of finite-dimensional vector spaces (i.e.,
for matrices), the terms ‘singular value’ and ‘eigenvalue’ coincide.
Since ρ(C(λ, ∆t)) ≤ 1 + O(∆t) from Remark 6.37 is relatively easy to check,
one may ask whether a similar criterion holds for more general operators. For this
purpose we introduce the ‘almost normal operators’: C(λ, ∆t) is almost normal if
2 2
kC(λ, ∆t)C(λ, ∆t)∗ − C(λ, ∆t)∗ C(λ, ∆t)k`2 ←`2 ≤ M (∆t) kC(λ, ∆t)k`2 ←`2
for some constant M < ∞.

Criterion 6.38 (Lax [15, p. 41]). If C(λ, ∆t) is almost normal, then `2 stability is
equivalent to the estimate kC(λ, ∆t)k`2 ←`2 ≤ 1 + O(∆t).
Proof. (i) Since kC(λ, ∆t)k`2 ←`2 ≤ 1 + O(∆t) is sufficient according to Criterion
6.26, it only remains to prove the necessity. Below, we abbreviate C(λ, ∆t) by C.
µ µ
(ii) First we prove by induction that (C ∗ ) C µ can be reordered into (C ∗ C) in
2 ∗ ∗
at most µ /2 steps (one step consists of C C 7→ CC ). In the case of µ = 1, no
permutation is required and 0 ≤ 12 /2 proves the start of induction. Suppose the
induction hypothesis for µ. Then
µ+1 µ µ+1
(C ∗ C) ,→ (C ∗ ) C µ C ∗ C ,→ (C ∗ ) C µ+1
µ2 /2 steps µ steps
2 2
and µ2 + µ ≤ (µ+1) 2 prove the hypothesis.
2 2µ
(iii) Each interchange perturbs the operator norm at most by M (∆t) kCk`2 ←`2 .
∗
In the following formula the factors Cj , 1 ≤ j ≤ 2µ, are either C or C :
kC1 · · · Cν CC ∗ Cν+3 · · · C2µ − C1 · · · Cν C ∗ CCν+3 · · · C2µ k`2 ←`2

≤ kC1 · · · Cν k`2 ←`2 kCC ∗ − C ∗ Ck`2 ←`2 kCν+3 · · · C2µ k`2 ←`2
h i
ν 2 2 2µ−ν−2 2 2µ
≤ kCk`2 ←`2 M (∆t) kCk`2 ←`2 kCk`2 ←`2 = M (∆t) kCk`2 ←`2
for all 0 ≤ ν ≤ 2µ − 2 (note that kC ∗ k`2 ←`2 = kCk`2 ←`2 ).

(iv) Forming the supremum over all u ∈ `2 with kuk`2 = 1 in
µ
hC µ u, C µ ui`2 = hu, (C ∗ ) C µ ui`2
µ M 2 2µ 2
≥ hu, (C ∗ C) ui`2 − (µ∆t) kCk`2 ←`2 kuk`2 ,
(ii), (iii) 2
µ 2
the left-hand
h i kC k`2 ←`2 , while the right-hand side yields the lower
side becomes
2 2µ
bound 1 − M 2 (µ∆t) kCk`2 ←`2 because of
µ µ µ 2µ
sup hu, (C ∗ C) ui`2 = k(C ∗ C) k`2 ←`2 = kC ∗ Ck`2 ←`2 = kCk`2 ←`2 .
C ∗C normal
Together with stability, this shows that

M 2 2µ 2
1− (µ∆t) kCk`2 ←`2 ≤ kC µ k`2 ←`2 ≤ K(λ)2 . (6.26)
2
√ 2
Restricting µ and ∆t by µ∆t ≤ min{1/ M , T }, we obtain 1 − M 1
2 (µ∆t) ≥ 2 .
p
Inequality (6.26) implies kCk`2 ←`2 ≤ 2µ 2
2K(λ) . Exercise 6.32b shows that
kCk`2 ←`2 = kC(λ, ∆t)k`2 ←`2 ≤ 1 + O(∆t). t u
The previous statements are connected to estimates of the kind
kC(λ, ∆t)k`2 ←`2 ≤ 1 or kC(λ, ∆t)k`2 ←`2 ≤ 1 + O(∆t).
These inequalities are the most convenient conditions proving stability. However,
even if kC(λ, ∆t)k`2 ←`2 ≥ c > 1, it may happen that the powers stay bounded
(kC(λ, ∆t)µ k`2 ←`2 ≤ const). In that case one may try to find an equivalent norm
2 P 2
which behaves easier. Since the square kU k`2 = i∈Z |Ui | is a quadratic form,
one may introduce another quadratic form Q(U ) such that
1 2 2
kU k`2 ≤ Q(U ) ≤ K1 kU k`2 for all U ∈ `2 . (6.27)
K1
p
This describes the equivalence of the norms k·k`2 and Q(·). The next lemma is a
slight generalisation of the ‘energy method’ stated in [21, pages 139–140].
Lemma 6.39. Let (6.27) be valid. Suppose that the growth of Q(·) for one time step
U µ+1 = C(λ, ∆t)U µ is limited by
2
2
0 ≤ Q(U µ+1 ) − Q(U µ ) ≤ K2 ∆t kU µ k`2 + U µ+1 `2 + K3 ∆t. (6.28a)

Then, for ∆t ≤ 1/(2K1 K2 ), the norms Q(U µ ) and kU µ k`2 stay bounded. More
precisely,
µ
2 2 K3 1 + K∆t K3
kU µ k`2 ≤ 2K12 − 1 U 0 `2 + − for µ ∈ N.
2K2 1 − K∆t 2K2
(6.28b)
Proof. We introduce the following abbreviations:
2 1 + K∆t
qµ := Q(U µ ), sµ := kU µ k`2 , K := K1 K2 , L := K1 K3 , c := .
1 − K∆t
µ−1 µ−1

P P
Since sµ ≤ K1 qµ = K1 q0 + (q`+1 − q` ) ≤ K1 K1 s0 + (q`+1 − q` ) ,
`=0 `=0
we infer from (6.28a) that
µ−1
X
sµ ≤ K12 − K1 K2 ∆t s0 + 2K1 K2 ∆t

s` + K1 K2 ∆t sµ + K1 K3 µ∆t.
`=0
Using the abbreviations K and L, we get

µ−1
X
(1 − K∆t) sµ ≤ K12 − K∆t s0 + 2K∆t

s` + Lµ∆t for all µ ∈ N.
`=0
(6.29a)
Note that 1 − K∆t ≥ 1/2 because of ∆t ≤ 1/(2K1 K2 ). The induction hypothesis
K3
sµ ≤ Sµ := Acµ − B , A := 2K12 − 1 s0 + B

with B := (6.29b)
2K2
is easily verified for µ = 0, since K12 ≥ 1 follows from (6.27):
S0 = Ac0 − B = 2K12 − 1 s0 + B − B ≥ s0 .

Assume that s` ≤ S` holds for all ` ≤ µ − 1. Insertion of s` ≤ S` and (6.29b) into

(6.29a) yields
Xµ−1
(1 − K∆t) sµ ≤ K12 − K∆t s0 + 2K∆t

S` + Lµ∆t
`=0
X µ−1
≤ K12 − K∆t s0 + 2AK∆t c` − 2BKµ∆t + Lµ∆t.

`=0
Pµ−1 µ
c −1 1−K∆t
Using `=0 c` = c−1 = 2K∆t (cµ − 1), we continue:
(1 − K∆t) sµ ≤ K12 − K∆t s0 + (1 − K∆t) A (cµ − 1) + (L − 2KB) µ∆t

K12 −K∆t
n o
= (1 − K∆t) Acµ − B + B + 1−K∆t s0 − A .
L=2KB | {z }
=Sµ
K12 −K∆t
The expression 1−K∆t is monotone in K∆t. From K∆t ≤ 1/2 we infer that
K12 −K∆t
− A ≤ B + 2K12 − 1 s0 − A = 0,

B+ 1−K∆t s0
which proves sµ ≤ Sµ for µ. t

u
To derive further stability criteria in the case of `2 , we need Fourier transforma-

tions, which are provided in the next part.
6.5.2 Fourier Analysis
A 2π-periodic function f ∈ L22π (R) is the 2π-periodic extension of f ∈ L2 (0, 2π)

to R. The associated Fourier series is
Z 2π
1 X 1
√ ϕα e iαξ
with ϕα := √ f (ξ)e−iαξ dξ (α ∈ Z),
2π α∈Z 2π 0
√
where i = −1. In the case of sufficiently smooth 2π-periodic functions f , the
Fourier series converges uniformly (with respect to k·k∞ ) to f . For general L2 func-
tions, the convergence holds only in the sense of the L2 norm. Therefore, the series
1
in f (ξ) = 2π α∈Z ϕα eiαξ has to be understood in the sense of k·kL2 (0,2π) . In
P
√
the sequel, the following isometry (Parseval equality) is essential:
kf kL2 (0,2π) = kϕk`2 , where ϕ = (ϕα )α∈Z . (6.30)
The transfer from the function f ∈ L2 (0, 2π) to its Fourier coefficients ϕ ∈ `2 is
the Fourier analysis, which will be denoted by F:
Ff = ϕ.
7 √12π α∈Z ϕα eiαξ = f is called the Fourier

P
On the other hand, the mapping ϕ →
synthesis and coincides with the inverse F −1 .
For p = 2, the solutions U µ of the difference scheme are `2 sequences. We denote
the associated 2π-periodic functions by Û µ :
1 X µ iαξ
Û µ := F −1 U µ , Û µ (ξ) = √ Uα e .
2π α∈Z
The difference scheme U µ+1 = C(λ, ∆t)U µ is equivalent to
Û µ+1 = F −1 U µ+1 = F −1 C(λ, ∆t)U µ = F −1 C(λ, ∆t)FF −1 U µ = Ĉ(λ, ∆t)Û µ
with
Ĉ(λ, ∆t) := F −1 C(λ, ∆t)F. (6.31)
2 2
While C(λ, ∆t) ∈ L(` , ` ), the transformed operator Ĉ(λ, ∆t) belongs to the
space L(L2 (0, 2π), L2 (0, 2π)).
Exercise 6.40. Using the isometry (6.30), prove
kFk`2 ←L2 (0,2π) = kF −1 kL2 (0,2π)←`2 = 1 (6.32)
(i.e., F and F −1 are unitary) and show that
kAk`2 ←`2 = kF −1 AkL2 (0,2π)←`2 for all A ∈ L(`2 , `2 ),

kBkL2 (0,2π)←`2 = kBFkL2 (0,2π)←L2 (0,2π) for all B ∈ L(`2 , L2 (0, 2π)).
The decisive quantity for stability is the operator norm kC(λ, ∆t)µ k`2 ←`2 . The
previous exercise shows that
kC(λ, ∆t)µ k`2 ←`2 = F −1 C(λ, ∆t)µ L2 (0,2π)←`2

µ
= kF −1 C(λ, ∆t)µ FkL2 (0,2π)←L2 (0,2π) = k FC(λ, ∆t)F −1 kL2 (0,2π)←L2 (0,2π)

= kĈ(λ, ∆t)µ kL2 (0,2π)←L2 (0,2π) . (6.33)

This leads us to the following remark.

Remark 6.41. An equivalent definition of `2 stability with the stability constant
K(λ) is
kĈ(λ, ∆t)µ kL2 (0,2π)←L2 (0,2π) ≤ K(λ) for all µ ∈ N0 , ∆t > 0 with µ∆t ≤ T.
P of Ĉ(λ, ∆t) from (6.31) we consider a term

For the concrete determination
Cj = aj Ej from C(λ, ∆t) = j Cj (cf. (6.22) and (6.19)):
(Cj U )ν := aj Uν+j for U ∈ `2 .
F −1 Cj U holds with Û = F −1 U ; i.e., Û ∈ L2 (0, 2π)

As stated above, Ĉj Û = P
is the Fourier series √12π α∈Z Uα eiαξ with the coefficients U = (Uα )α∈Z ∈ `2 .
Fourier synthesis of Cj U = (aj Uα+j )α∈Z yields
1 X 1 X
F −1 Cj U = √ (aj Uα+j ) eiαξ = aj √ Uα+j eiαξ
2π α∈Z 2π α∈Z
1 X 1 X
= aj √ Uβ ei(β−j)ξ = aj e−ijξ √ Uβ eiβξ = aj e−ijξ Û .
β=α+j 2π β∈Z 2π β∈Z
A comparison with F −1 Cj U = Ĉj Û shows that Ĉj = aj e−ijξ ; i.e., the linear
mapping Ĉj : L2 (0, 2π) → L2 (0, 2π) is the multiplication by the function aj e−ijξ .
Since Ĉ(λ, ∆t) = F −1 C(λ, ∆t)F = F −1 j Cj F = j F −1 Cj F = j Ĉj ,
P P P
we obtain the following remark.
P
Remark 6.42. Consider the difference operator C(λ, ∆t) = j∈Z aj Ej . The
Fourier transformed operator is Ĉ(λ,∆t)= j∈Z aj e−ijξ . Application of Ĉ(λ, ∆t)
P
corresponds to the multiplication by the trigonometric polynomial 12

X
G(ξ) := aj e−ijξ , (6.34)
j∈Z
which is called the ‘characteristic function’ or ‘symbol’ of C(λ, ∆t).

Exercise 6.43. Let φ ∈ C(R) be a bounded continuous function. Define the multi-
plication operator Φ : B → B by (Φ(f )) (ξ) := φ(ξ)f (ξ) for all ξ ∈ R. Prove for
both cases B = C(R) and B = L2 (R) that Φ ∈ L(B, B) possesses the operator
norm kΦkB←B = kφk∞ .
Therefore, the identity (6.33) becomes
kC(λ, ∆t)µ k`2 ←`2 = sup{|G(ξ)µ | : ξ ∈ R}. (6.35)
Here, ξ ∈ R can be replaced by ξ ∈ [0, 2π), since G is 2π-periodic.13

12
The parameters λ, ∆t are omitted in G(ξ) = G(ξ;λ, ∆t). Note that aj depends on λ, ∆t.
13
If the coefficients aj are matrices (cf. Remark 6.18), |G(ξ)µ | is to be replaced by the spectral
norm kG(ξ)µ k2 .
6.5.3 Further Criteria

P µ
−ijξ
Application of the last exercise to Ĉ(λ, ∆t)µ = j∈Z aj e yields
X µ
kĈ(λ, ∆t)µ kL2 (0,2π)←L2 (0,2π) = aj e−ijξ

j∈Z
X µ ∞
−ijξ
= aj e . (6.36)

j∈Z ∞
Note that (6.36) is only valid for scalar coefficients aj . Together with Remark 6.41,
(6.36) leads to the following theorem.
Theorem 6.44. The difference

P scheme (6.19) is `2 stable if and only if the charac-
−ijξ
teristic function G(ξ) = j∈Z aj e satisfies the estimate (6.37) with suitable
Kλ :
|G(ξ)| ≤ 1 + Kλ ∆t for all ξ ∈ R. (6.37)
Proof. (i) (6.37) implies kC(λ, ∆t)k`2 ←`2 = kĈ(λ, ∆t)kL2 (0,2π)←L2 (0,2π) =
kGk∞ ≤ 1 + Kλ ∆t, so that Corollary 6.28 proves stability.
(ii) Set c(∆t) := (kGk∞ − 1) /∆t. If there is no constant Kλ with (6.37),
c(∆t) → ∞ follows for ∆t → 0. Exercise 6.32a and (6.36) imply instability. t u
Proof of Criterion 6.31 in the case of `2 . Since inequality aj ≥ 1P

P
+ ∆t c(∆t)
with lim∆t→∞ c(∆t) = ∞ is supposed and kGk∞ ≥ |G(0)| = | aj | holds,
Exercise 6.32a proves `2 instability. tu
2
The previous analysis refers to ` . However, the characteristic function has also
consequences for `∞ .
Lemma 6.45. Let C(λ, ∆t) be of the form (6.19) with constant coefficients aj . Then
`∞ stability implies `2 stability:
kC(λ, ∆t)µ k`∞ ←`∞ ≥ kGµ k∞ = kC(λ, ∆t)µ k`2 ←`2 for all µ ∈ N0 .
The negation of this statement is: if the difference scheme (6.19) is `2 unstable, it is
also `∞ unstable.
Proof. Choose the special initial value U 0 with Uν0 = eiνξ , where ξ ∈ R is charac-
terised by |G(ξ)| = kGk∞ . Note that U 0 ∈ `∞ and kU 0 k∞ = 1. Application of
C(λ, ∆t) yields U 1 = C(λ, ∆t)U 0 with
X X X
Uν1 = 0
aj Uν+j = aj ei(ν+j)ξ = eiνξ aj eijξ = G(ξ)Uν0 ,
j∈Z j∈Z j∈Z
so that C(λ, ∆t)µ U 0 = G(ξ)µ U 0 and kC(λ, ∆t)µ U 0 k∞ = kG(ξ)µ U 0 k∞ =

µ
|G(ξ)| kU 0 k∞ = kGµ k∞ kU 0 k∞ . This proves the assertion. t
u
Below, the examples from §6.3.3 are again analysed with regard to (in)stability.
Example 6.46. (a) Scheme (6.17a) with a0 = 1 − aλ, a1 = aλ. For λ satisfy-
ing 0 ≤ aλ ≤ 1 , stability is already confirmed in Example 6.33a. The associated
characteristic function
G(ξ) := 1 − aλ + aλe−iξ = 1 − aλ (1 − cos ξ) − iaλ sin ξ
has modulus |G(π)| = |1 − 2aλ| at ξ = π. Since |G(π)| > 1 for all aλ ∈ / [0, 1],
Theorem 6.44 and Lemma 6.45 prove instability with respect to `2 and `∞ for aλ
outside of [0, 1]. Hence, the scheme (6.17a) is conditionally stable (with respect to
`2 and `∞ ) under the restriction aλ ∈ [0, 1], and unstable otherwise.
(b) Scheme (6.17b) with a−1 = − aλ aλ
2 , a0 = 1, a1 = 2 . The associated character-
istic function
aλ iξ aλ −iξ
G(ξ) := − e +1+ e = 1 − iaλ sin ξ
2 2
q
2
has maximum norm kGk∞ = 1 + |aλ| and, therefore, except the trivial case
a = 0, the scheme is always unstable (with respect to `2 and `∞ ).
(c) Scheme (6.17c) with a−1 = 1−aλ 2 , a1 =
1+aλ
2 . If |aλ| ≤ 1, Example 6.33c
shows stability. Because of
1 − aλ iξ 1 + aλ −iξ 2 2
G(ξ) := e + e and |G(ξ)| = cos2 ξ + |aλ| sin2 ξ,
2 2
2
the bound kGk∞ = max{1, |aλ| } follows. Hence, the scheme is conditionally
stable for |aλ| ≤ 1, but unstable for |aλ| > 1 (with respect to `2 and `∞ ).
(d) Scheme (6.18) with a−1 = λ, a0 = 1 − 2λ, a1 = λ. For λ ∈ (0, 1/2], stability
is shown in Example 6.33d. Because of
G(ξ) := λeiξ + 1 − 2λ + λe−iξ = 1 − 2λ(1 − cos ξ)
and kGk∞ = |G(π)| = |1 − 4λ| the scheme is conditionally stable for λ ∈ (0, 1/2],
but unstable for λ > 1/2 (with respect to `2 and `∞ ).
Lemma 6.45 yields a direct relation between G and `2 stability. Because of the
inequality kC(λ, ∆t)µ k`∞ ←`∞ ≥ kC(λ, ∆t)µ k`2 ←`2 , boundedness of powers of G
is necessary for `∞ stability. As we shall see, this condition is not sufficient for `∞
stability. A complete characterisation of `∞ stability is given by Thomée [24].
Theorem 6.47. Assume a difference scheme (6.16) with G(ξ) = j∈Z aj e−ijξ (cf.
P
(6.34)). Then the scheme is `∞ stable if and only if either condition (a) or (b) is
fulfilled:
(a) G(ξ) = ce−ijξ with |c| = 1,
(b) the set {ξ ∈ [−π, π] : |G(ξ)| = 1} = {ξ1 , . . . , ξN } has finite cardinality, and
there are numbers αk ∈ R, βk ∈ N, and γk ∈ C with <e γk > 0, such that
G(ξk + ξ) = G(ξk ) exp iαk ξ−γk ξ 2βk (1 + o(1))

as ξ → 0 (1 ≤ k ≤ N ).
The CFL condition is a necessary condition for convergence. ‘CFL’ abbreviates

the names of the authors Courant, Friedrichs, Lewy of [1].14
Criterion 6.48 (Courant–Friedrichs–Lewy). Let the hyperbolic differential equa-

∂ ∂
tion ∂t u = a ∂x u be discretised by the explicit difference scheme (6.16). Set
J1 = min{j ∈ Z : aj 6= 0},
J2 = max{j ∈ Z : aj 6= 0},
P µ P µ
so that the sum aj Uν+j can be reduced to aj Uν+j . If
j∈Z J1 ≤j≤J2
aλ ∈
/ [J1 , J2 ],
the scheme is not convergent (with respect to any norm). Therefore, aλ ∈ [J1 , J2 ] is
a necessary convergence condition.
Because of the equivalence theorem, stability cannot hold for aλ ∈

/ [J1 , J2 ].
Hence, the CFL criterion is also a stability criterion.
Proof of Criterion 6.48. For an arbitrary space-time point (t, x) with t > 0, the
solution is u(t, x) = u0 (x + at) according to Lemma 6.2. One verifies that a grid
point (t, x) = (µ∆t, ν∆x) of the solution Uνµ depends only on the initial values Uk0
with k ∈ [ν + µJ1 , ν + µJ2 ]. Multiplication by ∆x shows
xk = k∆x ∈ [x + µ∆xJ1 , x + µ∆xJ2 ]

= [x + µ∆xJ1 , x + µ∆xJ2 ]
= [x + tJ1 /λ, x + tJ2 /λ].
If aλ ∈
/ [J1 , J2 ], then x + at ∈
/ [x + tJ1 /λ, x + tJ2 /λ]. Therefore, the computation
of Uνµ does not use those data, on which the solution u(t, x) depends. To be more
precise, choose an initial value u0 ∈ C0∞ (R) with
u0 (a) 6= 0, supp(u0 ) ⊂ R\[J1 /λ, J2 /λ].
Then U (1, x) = 0 holds in a neighbourhood of x = 0, while the true solution

satisfies u(1, 0) = u0 (a) 6= 0 and even u(1, x) 6= 0 in a neighbourhood of x = 0.
As a consequence, Uνµ cannot converge to u. t u
As an example, take the scheme (6.17a). Since a0 = 1 − aλ and a1 = aλ, the

CFL bounds are J1 = 0 and J2 = 1. The necessary CFL condition aλ ∈ [0, 1]
coincides for Example 6.46a with the exact stability property.
We summarise the mentioned necessary and/or sufficient conditions in the fol-

lowing table.
14
The original paper is written in German. An English translation can be found in [3,AppendixC].
statement name or necessary sufficient

comment `2 `∞ `2 `∞
Criterion 6.26 * *
Corollary 6.28 * *
Criterion 6.30 positive scheme * *
Criterion 6.31 * *
Lemma 6.34 perturbation * *
Criterion 6.36 * *
Criterion 6.38 almost normal * *
Theorem 6.44 * *
Lemma 6.45 * *
Theorem 6.47 * *
Criterion 6.48 CFL * *
Criterion 6.52 * *
Criterion 6.54(a) von Neumann * *
Criterion 6.54(b) von Neumann * *
Lemma 6.55 Lax-Wendroff *
Criterion 6.57 Friedrichs *
Theorem 6.58 Friedrichs *
6.5.4 Implicit Schemes
All schemes examined so far are, at best, conditionally stable. In order to obtain
unconditionally stable schemes, one must admit implicit difference schemes.15
∂u ∂2u
In the parabolic case ∂t = ∂x2 , the second x-difference can be formed at time
level t + ∆t, i.e.,
∂2u u(t + ∆t, x − ∆x) − 2u(t + ∆t, x) + u(t + ∆t, x + ∆x)

2
≈ ,
∂x ∆x2
∂ u(t+∆t,x)−u(t,x)
while the time derivative ∂t u becomes ∆t as before. This leads to
µ+1 µ+1
−λUν−1 + (1 + 2λ) Uνµ+1 − λUν+1 = Uνµ . (6.38)
Instead of the explicit form U µ+1 = C(λ, ∆t)U µ , one now obtains an implicit
scheme of the form
C1 (λ, ∆t)U µ+1 = C2 (λ, ∆t)U µ , (6.39)
where in the present case the operators are given by
15
The CFL criterion does not apply to implicit schemes, since formally implicit schemes can be
viewed as explicit ones with an infinite sum (i.e., J1 = −∞, J2 = ∞). Then aλ ∈ [J1 , J2 ] = R
is valid, and the CFL condition is always satisfied.
X
(C1 (λ, ∆t)U )ν = a1,j Uν+j with a1,−1 = a1,1 = −λ, a1,0 = 1 + 2λ and
−1≤j≤1

X all other coefficients = 0;
C2 (λ, ∆t)U = a2,j Uν+j with a2,0 = 1,
i.e., C2 = identity.
j∈Z
One would like to solve (6.39) with respect to U µ+1 :

−1
U µ+1 = C(λ, ∆t)U µ with C(λ, ∆t) := [C1 (λ, ∆t)] C2 (λ, ∆t). (6.40)
−1
Therefore, the existence of the inverse [C1 (λ, ∆t)] is to be investigated.
Lemma 6.49. (a) If the coefficients a1,j of C1 (λ, ∆t) satisfy the inequality
!
X
|a1,j | / |a1,0 | ≤ 1 − ε < 1 for some ε > 0,
j∈Z\{0}
−1 1
then the inverse exists and satisfies k [C1 (λ, ∆t)] k`p ←`p ≤ ε|a1,0 | .
−1
(b) The inverse [C1 (λ, ∆t)] exists in L(`2 , `2 ) if and only if the characteristic
function G1 (ξ) of C1 satisfies an inequality |G1 (ξ)| ≥ η > 0. The norm equals
−1
k [C1 (λ, ∆t)] k`2 ←`2 = 1/ inf |G1 (ξ)| .
ξ∈R
Proof. (i) The equation C1 V = U is equivalent to the fixed-point equation

1
V = (U − C10 V ) =: Φ(V ) with C10 := C1 − a1,0 I.
a1,0
The contraction constant of Φ is kC10 k`p ←`p / |a1,0 |. By Remark 6.27,

P
kC10 k`p ←`p j∈Z\{0} |a1,j |
≤ ≤1−ε
|a1,0 | |a1,0 |
is valid, so that a unique inverse exists. The estimate

1
kV k`p ≤ kU k`p + (1 − ε) kV k`p
|a1,0 |
1 −1
shows that ε kV k`p ≤ |a1,0 | kU k`p , so that k [C1 (λ, ∆t)] k`p ←`p has the bound
1/ (ε |a1,0 |).
(ii) Ĉ1 (λ, ∆t) = F −1 C1 (λ, ∆t)F is the Fourier transformation of C1 . Ĉ1 (λ, ∆t)
is the operator Û 7→ G1 Û . Obviously, the multiplication operator M ∈ L(`2 , `2 )
with M Û := (1/G1 ) Û is the inverse of Ĉ1 . The norm of M is
kM k`2 ←`2 = kĈ1−1 k`2 ←`2 = k1/G1 k∞ ≤ 1/η.

On the other hand, one verifies that kĈ1−1 k`2 ←`2 cannot be finite if inf |G1 | equals
zero. Since the Fourier transform does not change the `2 norm, the equality
kC1−1 k`2 ←`2 = kĈ1−1 k`2 ←`2 = k1/G1 k∞ = 1/ inf ξ∈R |G1 (ξ)| follows. t u
Example 6.50. The scheme (6.38) is `2 stable for all λ = ∆t/∆x2 ; i.e., it is uncon-
ditionally stable.
Proof. G1 (ξ) = −λeiξ +1+2λ−λe−iξ = 1+2λ (1 − cos x) ≥ inf ξ∈R |G1 (ξ)| = 1
is the characteristic function of C1 . Since C2 = I, we obtain
−1 −1
C(λ, ∆t) := [C1 (λ, ∆t)] C2 (λ, ∆t) = [C1 (λ, ∆t)]
and kC1−1 k`2 ←`2 = 1/ inf ξ∈R |G1 (ξ)| = 1. Stability follows by Criterion 6.26. t
u
−1
Remark 6.51. [C1 (λ, ∆t)] from above is equal to the infinite operator
|j|
−1
X 1 2λ
[C1 (λ, ∆t)] = aj (λ)Ej with aj = √ √ ;
j∈Z
1 + 4λ 2λ + 1 + 1 + 4λ
µ
i.e., the implicit scheme (6.38) is identical to Uνµ+1 =
P
aj Uν+j .
j∈Z
P
Proof. Check that C1 (λ, ∆t) j∈Z aj Ej = identity. t
u
The general case of an implicit scheme (6.39) is treated in the next criterion (its
proof is identical to that of Theorem 6.44).
Criterion 6.52. The scheme (6.39) is `2 stable if and only if the characteristic func-
tion
G(ξ) := G2 (ξ)/G1 (ξ)
satisfies condition (6.37).
Example 6.53. A modification of (6.38) is the so-called theta scheme

µ+1 µ+1
− λΘUν−1 + (1 + 2λΘ) Uνµ+1 − λΘUν+1 (6.41)
µ µ
= λ (1 − Θ) Uν−1 + (1 − 2λ (1 − Θ)) Uνµ + λ (1 − Θ) Uν+1
for Θ ∈ [0, 1]. For Θ = 0 and Θ = 1 we regain (6.18) and (6.38), respectively.
For Θ = 1/2, (6.41) is called16 the Crank–Nicolson scheme. The scheme (6.41) is
unconditionally `2 stable for Θ ∈ [1/2, 1], whereas in the case of Θ ∈ [0, 1/2) it is
conditionally `2 stable for
λ ≤ 1/ (2 (1 − 2Θ)) .
In all stable cases, kC(λ, ∆t)k`2 ←`2 = 1 holds.

16
Occasionally, one finds the incorrect spelling ‘Crank–Nicholson’. The second author of [2] is
Mrs. Phyllis Nicolson (1917–1968).
Proof. Let (DU )ν = −Uν−1 + 2Uν − Uν+1 be the negative second difference
operator. The characteristic function of D is GD (ξ) = 2 − 2 cos ξ = 4 sin2 (ξ/2).
The operators C1 , C2 from (6.39) in the case of (6.41) are
C1 (λ, ∆t) = I + λΘD and C2 (λ, ∆t) = I − λ (1 − Θ) D.
The associated functions are
G1 (ξ) = 1 + λΘGD (ξ) and G2 (ξ) = 1 − λ (1 − Θ) GD (ξ),
so that
1 − λ (1 − Θ) GD (ξ)
G(ξ) = .
1 + λΘGD (ξ)
1−λ(1−Θ)X
The function 1+λΘX is monotonically decreasing with respect to X, so that the
maximum of |G(ξ)| is taken at X = 0 = GD (0) or X = 4 = GD (π):

4λ (1 − Θ) − 1
kGk∞ = max{G(0), −G(π)} = max 1, .
1 + 4λΘ
If Θ ∈ [ 12 , 1], then
4λ (1 − Θ) − 1 4λ 4λ 2λ − 1
−G(π) = = −1 ≤ −1= ≤1
1 + 4λΘ 1 + 4λΘ Θ≥ 12 1 + 2λ 2λ + 1
proves kGk∞ = 1.
In the case of Θ ∈ [0, 1/2), the choice of λ must ensure the estimate −G(π) ≤ 1.
Equivalent statements are
4λ (1 − Θ) − 1 ≤ 1 + 4λΘ ⇔ 4λ (1 − 2Θ) ≤ 2 ⇔ λ ≤ 1/ (2 (1 − 2Θ)) . t

u
6.5.5 Vector-Valued Grid Functions
So far, `p has been the set the complex-valued sequences (Uν )ν∈Z , Uν ∈ C. If the
∂
equation ∂t u(t) = Au(t) from (6.1a) is vector-valued (values in CN ), also the grid
functions (Uν )ν∈Z must be vector-valued:
`p = {U = (Uν )ν∈Z : Uν ∈ CN } with the norms

r
X 2
kU k2 = kUν k2 , kU k∞ = sup kUν k∞ ,
ν∈Z ν∈Z
where kUν kp is the Euclidean norm in CN (p = 2) or the maximum norm in CN

(p = ∞, cf. Remark 6.18).
Now, the coefficients aj in the difference scheme (6.19) are N ×N matrices. The
statements concerning consistency, convergence, and stability remain unchanged
(only the norms are to be interpreted differently). However, the criteria (see §6.5
and later) are to be generalised to the case N > 1.
Criterion 6.26 remains valid without any change.
The estimate in Remark 6.27 becomes
X
kC(λ, ∆t)k`p ←`p ≤ kaj kp ,
j
where k.k2 is the spectral norm and k.k∞ is the row-sum norm for N × N matrices.
P P
In Corollary 6.28 one has to replace |aj | by j kaj kp .
When we form G(ξ) = j∈Z aj e−ijξ by the Fourier transform, G(ξ) is now an
P
N × N matrix-valued function. The characterisation (6.35) becomes
kC(λ, ∆t)µ k`2 ←`2 = sup{kG(ξ)µ k2 : ξ ∈ [0, 2π)}. (6.42)
Instead of the relatively abstract operator C(λ, ∆t), one has now to investigate the
boundedness of the N × N matrices G(ξ)µ .
According to Criterion 6.36, ρ(C(λ, ∆t)) ≤ 1 + O(∆t) is a necessary condi-

tion. Since the unitary Fourier transform does not change the spectra, C(λ, ∆t) and
{G(ξ) : ξ ∈ [0, 2π)} have identical eigenvalues. This will lead to the von Neumann
condition. Here the spectral radius ρ(G(ξ)) is defined in (6.25). Part (b) corresponds
to Remark 6.37.
Criterion 6.54 (von Neumann condition). (a) A necessary condition for stability
(with respect to `2 and `∞ ) is
sup{ρ(G(ξ)) : ξ ∈ [0, 2π)} ≤ 1 + O(∆t).
(b) If all matrices G(ξ) are normal, this condition is even sufficient for `2 stability.
The following statement uses the numerical radius of a matrix:

hAv, vi
r(A) := sup 2 .

06=v∈CN kvk2
Lemma 6.55 (Lax–Wendroff condition, [18]). Suppose that there is a constant

KLW such that
r (G(ξ)) ≤ 1 + KLW ∆t for all ξ ∈ [0, 1].
Then the scheme is `2 stable.

Proof. The numerical radius of general square matrices A possesses the properties
kAk2 ≤ 2r(A), r(An ) ≤ r(A)n for n ∈ N0
(cf. §2.9.5 in [6] or [7]). Hence,

n
kG(ξ)n k2 ≤ 2r(G(ξ)n ) ≤ 2r(G(ξ))n ≤ 2 (1 + KLW ∆t) ≤ 2 exp(KLW T )
holds for all n with n∆t ≤ T and all ξ ∈ [0, 2π). Because of (6.42), the assertion is
proved. t u
Definition 6.29 carries over to the matrix case when we replace non-negative
reals by positive semidefinite matrices; i.e., a positive difference scheme is charac-
terised by ‘aj positive semidefinite’.
Exercise 6.56. Suppose that the positive semidefinite coefficients aj are either
(i) all diagonal or
(ii) simultaneously diagonalisable; i.e., there is a transformation S such that the
matrices dj = Saj S −1 are diagonal for all j.
Show that analogous to Criterion 6.30, the difference scheme (6.19) is `2 and
∞
P
` stable if aj = I (cf. (6.24)).
However, also without simultaneous diagonalisability, the criterion can be

generalised. Note that even x-dependent coefficients aj = aj (x) are allowed.
Criterion 6.57 (Friedrichs [5]). Suppose that the difference schemeP(6.19) has
positive semidefinite coefficients aj satisfying the consistency condition j∈Z aj = I
(cf. (6.24)). The coefficients aj must be either constant or the following three
conditions must hold:
(i) the hyperbolic case with λ = ∆t/∆x is given (cf. (6.11)),
(ii) aj (·) P
are globally Lipschitz continuous in R with Lipschitz constant Lj ,
(iii) B := j∈Z jLj < ∞.
Then
kC(λ, ∆t)k`2 ←`2 ≤ 1 + CL ∆t
holds with CL = B/ (2λ) implying `2 stability.
Proof. We prove the general case (constant coefficients correspond to Lj = 0). In

XX
U, V ∈ `2

(C(λ, ∆t)V, U )`2 = ∆x haj (ν∆x)Vν+j , Uν i (6.43)
ν∈Z j∈Z
h·, ·i denotes the scalar product of CN (N is the dimension of Uν , Vν ∈ CN ).

1 1
Any positive semidefinite matrix M fulfils |hM x, yi| ≤ 2 hM x, xi + 2 hM y, yi
for x, y ∈ CN . Application to (6.43) yields

XX

∆x
ha j (ν∆x)Vν+j , U i
ν

ν∈Z j∈Z
∆x X X ∆x X X
≤ haj (ν∆x)Uν , Uν i + haj (ν∆x)Vν+j , Vν+j i .
2 2
ν∈Z j∈Z ν∈Z j∈Z
∆x
PDP E
∆x 1 2
kUν k22
P
The first term is 2 aj (ν∆x)Uν , Uν = = kU k`2 .
ν∈Z j∈Z (6.24) 2 ν∈Z 2
Substitute ν in the second term by µ = ν + j:

∆x X X ∆x X X
haj (ν∆x)Vν+j , Vν+j i = haj ((µ − j) ∆x)Vµ , Vµ i .
2 2
ν∈Z j∈Z µ∈Z j∈Z
Lipschitz continuity, kaj ((µ − j) ∆x) − aj (µ∆x)k2 ≤ Lj j∆x holds.

Because ofP
With B = j∈Z jLj < ∞ (assumption (iii)) we estimate by
XX
∆x
2 haj ((µ − j) ∆x)Vµ , Vµ i
µ∈Z j∈Z
XX 2
∆x ∆x
= h[aj ((µ − j) ∆x) − aj (µ∆x)] Vµ , Vµ i + kV k`2
(6.24) 2 2
µ∈Z j∈Z
1 + B∆x 2
≤ kV k`2 .
2
1 2 1+B∆x 2
From the previous estimates we obtain |(CV, U )`2 | ≤ 2 kU k`2 + 2 kV k`2 .
Characterising the norm by
kC(λ, ∆t)k`2 ←`2 = sup |(CV, U )`2 | ,

kU k`2 =kV k`2 =1
we obtain the inequality kC(λ, ∆t)k`2 ←`2 ≤ 1 + B∆x

2 = 1 + CL ∆t. The last step
uses either ∆x = ∆t/λ from assumption (i) or B = 0 (constant coefficients). t
u
We add some comments concerning the latter criterion.

√
(1) Inequality kC(λ, ∆t)k`2 ←`2 ≤ 1 + O(∆x) = 1 + O( ∆t) is not sufficient for
stability in the parabolic case with λ = ∆t/∆x2 .
(2) If the coefficients aj are constant, the vanishing values Lj = B = CL = 0 show
that kC(λ, ∆t)k`2 ←`2 = 1.
P
(3) If only finitely many coefficients are different from zero, B = j∈Z jLj < ∞
is satisfied.
A positive difference scheme can be obtained as discretisation of the symmetric
hyperbolic system of differential equations
∂ ∂
u + A(x) u = 0 (A(x) symmetric N × N matrix, u ∈ CN ). (6.44)
∂t ∂x
Replacing ∂x∂
u by u(t,x+∆x)−u(t,x−∆x)
2r∆x
∂
and ∂t u by u(t+∆t,x)−ū(t,x)
∆t with the
1
average ū(t, x) = 2 [u(t, x + ∆x) + u(t, x − ∆x)], we are led to the difference
scheme (Friedrichs’ scheme)
1
(C(λ, ∆t)U )ν = {[I − λA(ν∆x)] Uν−1 + [I + λA(ν∆x)] Uν+1 } . (6.45)
2
Choosing 0 < λ ≤ 1/ supx∈R kA(x)k2 , we obtain a positive scheme C(λ, ∆t).
Hence `2 stability follows, provided that A(x) is Lipschitz continuous.
Criterion 6.36 as well as Remark 6.37 and Criterion 6.38 remain valid in the
vector-valued case.
Also in the vector-valued
P case, the Fourier transformed difference operator has
−ijξ
the form Ĉ(λ, ∆t)(ξ) = j∈Z aj e , where now Ĉ is a 2π-periodic function,
whose values are N × N matrices. As in the scalar case, the stability estimate
kC(λ, ∆t)µ k`2 ←`2 ≤ K(λ) holds for all µ∆t ≤ T if and only if
kĈ(λ, ∆t)µ kL2 (0,2π)←L2 (0,2π) ≤ K(λ).

−ijξ µ
P P µ
However, the equality j∈Z aj e ∞
= j∈Z aj e−ijξ ∞ from (6.36) is,
in general, no longer valid for the matrix-valued case, but becomes an inequality:
X µ X µ
aj e−ijξ
∞
≤ aj e−ijξ ∞ .
j∈Z j∈Z
The investigation of `∞ stability of a hyperbolic system is more involved.

Following Mizohata [20], we define regular hyperbolicity. A system
∂ ∂
u + A(t, x) u = 0 (A N×N matrix, u ∈ CN, x ∈ R, 0 ≤ t ≤ T ) (6.46)
∂t ∂x
is called regularly hyperbolic, if the eigenvalues di (t, x) (1 ≤ i ≤ N ) of A(t, x)
are real and distinct; i.e., there is some δ > 0 such that
|di (t, x) − dj (t, x)| ≥ δ for all i 6= j (1 ≤ i, j ≤ N, x ∈ R, 0 ≤ t ≤ T ).
The following stability result is proved by Tomoeda [26].
Theorem 6.58. Suppose that (6.46) is a regularly hyperbolic system satisfying the
smoothness condition
sup{|A(t, x)|, |At (t, x)|, |Ax (t, x)| : x ∈ R, 0 ≤ t ≤ T } < ∞.
Then Friedrichs’ scheme (6.45) with
0 < λ < 1/ sup |di (t, x)|

x∈R,0≤t≤T
is `∞ stable.
6.5.6 Generalisations
In §6.5.5 we have already analysed the generalisation to the case of systems of

differential equations with vector-valued solutions. Next, a series of further gen-
eralisations is discussed.
6.5.6.1 Spatially Multivariate Case
Instead of one spatial variables, we admit d variables x = (x1 , . . . , xd ) (d = 3 is

the standard case). Then the heat equation (6.4) becomes ∂u
∂t = ∆u for t > 0 with
∂2
Pd
the Laplace operator ∆ = k=1 ∂x2 . The representation (6.5) of u can be adapted
k
to the d-dimensional situation.
∂ ∂
The hyperbolic differential equation ∂t u = a ∂x u is generalised to
d
∂ X ∂u
u= Ak . (6.47)
∂t ∂xk
k=1
For the discretisation, the grid G∆x = {x = ν∆x : ν ∈ Z} from (6.9) is to be

replaced by the d-dimensional grid17
G∆x = {x = ν∆x : ν ∈ Zd }
d
with multi-indices ν = (ν1 , . . . , νd ), νj ∈ Z. Now, the Banach space `p is CZ
equipped with the norms
s X
d 2
kU k`∞ = sup{|Uν | : ν ∈ Z } and kU k`2 = ∆xd |Uν | .
ν∈Zd
2
The Fourier transform of U ∈ ` yields
1 X
Û (ξ) = d/2
Uν eiνξ ,
(2π) ν∈Zd
where d
X
νξ = hν, ξi = νj ξj
k=1
denotes the Euclidean scalar product in Rd .
The stability analysis of system (6.47) with N × N matrices Ak leads to the
following complication. In the univariate case d = 1, all coefficient matrices aj
of C(λ, ∆t) are derived from only one matrix A1 , so that in the standard case the
matrices aj are mutually commutable (and thereby simultaneously diagonalisable).
For d > 1 with non-commutative matrices Ak in (6.47), the matrices aj are expected
to be not simultaneously diagonalisable.
17
In principle, different step widths ∆xj make sense. However, after a transformation xj 7→
∆x1
∆xj j
x of the spatial variables we regain a common step size ∆x.
6.5.6.2 Time-Dependent Coefficients
The coefficients in the differential equation (6.2) or (6.47) may depend on t:

a = a(t) or Ak = Ak (t).
In this case the concept of the semigroup of solution operators (cf. §6.2) need to be
modified a bit. Instead of T (t), the operator T (t1 , t0 ) with 0 ≤ t0 ≤ t1 ≤ T appears,
which describes the transition of an initial value at time t0 to the solution at time t1 .
The semigroup property reads T (t2 , t1 )T (t1 , t0 ) = T (t2 , t0 ) for t0 ≤ t1 ≤ t2 and
T (t, t) = I.
The discretisation yields time-dependent difference schemes C(t; λ, ∆t) at
t = µ∆t with U µ = C(µ∆t; λ, ∆t)U µ−1 :
X
(C(t; λ, ∆t)U )ν = d
aj (t)Uν+j for U ∈ `p and ν ∈ Zd .
j∈Z
In the stability definition (6.23), C(λ, ∆t)µ is to be replaced by the product
C(t0 + µ∆t; λ, ∆t) · C(t0 + (µ − 1) ∆t; λ, ∆t) · . . . · C(t0 + ∆t; λ, ∆t)
with 0 ≤ t0 ≤ t0 + µ∆t ≤ T .
Criteria 6.26, 6.36 and Lemma 6.34 hold also in the time-dependent case.
6.5.6.3 Spatially Dependent Coefficients
The differential equation may contain spatially dependent coefficients as, e.g., in
the differential equation (6.44). Correspondingly, all coefficients aj = aj (x) of
C(λ, ∆t) may depend on x. Criterion 6.57 explicitly admits variable coefficients in
the case of positive difference schemes.
There are criteria using again the function
X
G(x, ξ) := aj (x) e−ijξ , (6.48)
j∈Z
which formally corresponds to (6.34). However, G is not the Fourier transform

F −1 C(λ, ∆t)F !
Having an operator C(λ, ∆t) with variable coefficients, we can introduce the
operator
C(x0 ; λ, ∆t)
by replacing all coefficients aj (x) by the constant coefficients aj (x0 ). We say
that C(x0 ; λ, ∆t) is the operator C frozen at x0 . G(x0 , ξ) is the Fourier trans-
form of C(x0 ; λ, ∆t). This leads to the obvious question as to whether stability
of the original scheme C(λ, ∆t) corresponds to the stability of the frozen operator
C(x0 ; λ, ∆t) at all x0 ∈ R. The general answer is negative.
Remark 6.59 (Kreiss [9, §7]). Stability of C(x0 ; λ, ∆t) at all x0 ∈ R is, in general,
neither sufficient nor necessary for the stability of C(λ, ∆t).
Typical sufficient criteria require besides Lipschitz continuity of the coefficients

aj (·) that the scheme be dissipative. Here, C(λ, ∆t) is called dissipative of order
2r (r ∈ N) if there are constants δ, τ > 0 such that
2r
|λν (x, ∆t, ξ)| ≤ 1 − δ |ξ| for all x ∈ R, ∆t ∈ (0, τ ), |ξ| ≤ π, (6.49)
holds for the eigenvalues λν (ν = 1, . . . , N ) of G(x, ξ).
We hint to a connection between condition (6.49) and the Definition 5.23 of the
stability of ordinary differential equations, which requires that zeros λν of ψ with
|λν | = 1 be simple zeros, while otherwise |λν | < 1 must hold. In the case of (6.49),
the powers kG(x, ξ)n k must be uniformly bounded, while in the second case the
companion matrix must satisfy kAn k ≤ const (cf. Remark 5.39). However, the
difference is that in the second case only finitely many eigenvalues exist, so that
max |λν | < 1 holds for all eigenvalues with |λν | 6= 1. In the case of |λν (x, ∆t, ξ)|,
the eigenvalues λν are continuous functions of ξ and their absolute value may tend
to 1. Condition (6.49) describes quantitatively how fast λν (x, ∆t, ξ) approaches 1.
The stability result of Kreiss [10] takes the following form (see also [21, §5.4]).
Theorem 6.60. Suppose the matrices aj of the hyperbolic system

d
X
ut = aj (x)uxj
j=1
to be Hermitian, uniformly bounded, and uniformly Lipschitz continuous with

respect to x. If the difference scheme is dissipative of order 2r and accurate of
order 2r − 1 for some r > 0, then it is `2 stable.
6.5.6.4 Initial-Boundary Value Problem
If we replace the real axis in Σ = [0, T ]×R (cf. (6.3)) by an interval or the half-axis
[0, ∞), the new computational domain Σ = [0, T ] × [0, ∞) has a non-empty spatial
boundary [0, T ] × {0}. Then the parabolic initial-value problem has to be completed
by a boundary condition at x = 0; e.g.,
u(t, 0) = uB (t)
with some given function uB to determine a unique solution.

In the case of hyperbolic problems the situation is more involved. The model
∂ ∂
problem ∂t u(t) = a ∂x u(t) in (6.2) requires the same kind of boundary condi-
tion as above if a > 0. Otherwise, the differential equation is uniquely determined
without further condition at x = 0. In the case of hyperbolic systems, the number

of characteristic directions with positive slope determines the number of necessary
boundary conditions.
After the discretisation, the grid G∆x from (6.9) has to be replaced by the one-
sided infinite grid G∆x = {x = ν∆x : ν ∈ N0 }. The space B becomes `p =
{U = (Uν )ν∈N0 }, where U0 is fixed by the boundary conditions. For an analysis we
refer to Kreiss [11, 12], Richtmyer–Morton [21, §6], and Kreiss–Wu [13].
6.5.6.5 Multistep Schemes
So far, we have considered one-step methods; i.e., U n+1 is computed from U n .

As we already know from ordinary differential equations, instability may arise
from improperly designed multistep methods. As an example of a two-step method
we discuss the popular leap-frog scheme, first applied to the hyperbolic equation
∂ ∂
∂t u = a ∂x u. Both derivatives are discretised by the symmetric difference quotient;
i.e.,
∂ 1
u≈ [u(t + ∆t, x) − u(t − ∆t, x)] and
∂t 2∆t
∂ 1
u≈ [u(t, x + ∆x) − u(t, x − ∆x)].
∂x 2∆x
The resulting two-step method is
µ µ
Uνµ+1 = Uνµ−1 + λa[Uν+1 − Uν−1 ]. (6.50)
As usual, we need starting values for µ = 0 and µ = 1 to proceed with the leap-frog
scheme. The name originates from the fact that the computation of Uνµ+1 involves
only values
Unm with n + m = ν + µ + 1 (modulo 2).
∆t
The grid Σ∆x defined in (6.10) splits into the two colours of the chequer board:
˙ odd , where
Σeven ∪Σ
∆t
Σeven = {(n∆x, m∆t) ∈ Σ∆x : n + m even integer}.
This observation allows us to reduce the computation to Σeven , which halves the
cost of the computation.
The stability analysis follows the idea of Remark 5.38b: we formulate the two-
µ
step method as a one-step method for V µ := UUµ−1 :
µ+1 µ
U λa (E1 − E−1 ) 1 U
= ;
Uµ 1 0 U µ−1
i.e., V µ+1 = CV µ with


λa (E1 − E−1 ) 1 −λa 0 01 λa 1
C := = E−1 + + E1 .
1 0 0 0 10 1 0
Here Ej is the shift operator from (6.22). The corresponding characteristic function
G(ξ) from (6.34) is now matrix-valued:
λa e−iξ − eiξ 1

−2iλa sin ξ 1
G(ξ) = = (0 ≤ ξ ≤ 2π).
1 0 1 0
One easily checks that |λa| > 1 together with ξ = π/2 (i.e., sin ξ = 1) leads to
an eigenvalue λ(ξ) with |λ(ξ)| > 1. Von Neumann’s condition in Criterion 6.54a
implies that |λa| ≤ 1 is necessary for `2 stability.
Proposition 6.61. The leap-frog scheme (6.50) is `2 stable if and only if |λa| < 1.
Proof. It remains to show that |λa| < 1 implies stability. For this purpose we give
an explicit
description of the powers G(ξ)n . Abbreviate x := −λa sin ξ, so that
G(ξ) = 2ix 1

1 0 . We claim that
n
i Un (x) in−1 Un−1 (x)

n
G(ξ) = n−1 for n ≥ 1, (6.51)
i Un−1 (x) in−2 Un−2 (x)
where Un are the Chebyshev polynomials of the second kind. These are defined on
[−1, 1] by
sin((n + 1) arccos(x))
Un (x) := √ (n = −1, 0, 1, . . .)
1 − x2
and satisfy the same three-term recursion
Un+1 (x) = 2xUn (x) − Un−1 (x)
as the Chebyshev polynomials (of the first kind) mentioned in Footnote 7 on
page 52. Statement (6.51) holds for n = 1. The recursion formula together with
G(ξ)n+1 = G(ξ)n G(ξ) proves the induction step.
If |λa| ≤ A < 1, the inequalities |x| ≤ A < 1 and |ϕ| ≤ arccos(A) < 1 follow,
where ϕ := arccos(x). One verifies that
Un (x) = Un (cos ϕ) = sin((n + 1)ϕ)/ sin(ϕ)
is bounded by Un (A) ≈ 1/[2(1 − A)] independently of n. Hence kG(ξ)n k is uni-

formly bounded in n, and `2 stability follows.
If A = 1, the polynomials are bounded by
|Un (x)| ≤ Un (0) = n + 1,
implying (a rather weak form18 of) instability. t

u
18
If the amplification factor is some fixed ζ > 1, the instability effects of kC(λ, ∆t)n k > ζ n
are easily observed and may lead to an overflow in the very end. On the other hand, we conclude
The original leap-frog scheme of Du Fort–Frankel [4] is applied to the parabolic

problem (6.4). Here we start from
∂ 1
u≈ [u(t + ∆t, x) − u(t − ∆t, x)] and
∂t 2∆t
∂2 1
2
u≈ [u(t, x − ∆x) − 2u(t, x) + u(t, x + ∆x)].
∂x ∆x2
In order to obtain the leap-frog pattern, we replace 2u(t, x) by the average
u(t + ∆t, x) + u(t − ∆t, x). Together with λ = ∆t/∆x2 (cf. (6.11)), we obtain the
Du Fort–Frankel scheme
µ µ
(1 + 2λ)Uνµ+1 = (1 − 2λ)Uνµ−1 + 2λ[Uν+1 + Uν−1 ]. (6.52)
The stability analysis can again be based on the representation of G(ξ)n . For
λ ≤ 1/2, a simpler approach makes use of the fact that the coefficients 1−2λ
1+2λ and
2λ µ µ
1+2λ of Uνµ+1 , Uν+1 , Uν−1 are positive. As in Criterion 6.30, one concludes that
the scheme (6.52) is `∞ stable, which implies `2 stability.
For λ > 1/2 we return to the characteristic function
2λ
e−iξ + eiξ 1−2λ 4λ 1−2λ

G(ξ) = 1+2λ 1+2λ = 1+2λ cos ξ 1+2λ (0 ≤ ξ ≤ 2π).
1 0 1 0
All eigenvalues are of absolute value ≤ 1. The eigenvalue 1qappears for ξ = 0

and is a single one (the other is 1−2λ
1+2λ ). A double eigenvalue
2λ−1
2λ+1 < 1 occurs
2
for sin2 ξ = 1/(4λ ). Hence, by Lemma 5.28, the powers of G(ξ) are uniformly
bounded. This proves the following statement.
Proposition 6.62. For any fixed λ =∆t/∆x2 , the Du Fort–Frankel scheme (6.52) is
`2 stable.
As pointed out in [16, §19], it is important to perform the limit ∆t, ∆x → 0

in such a way that λ =∆t/∆x2 stays constant, although any value of λ is allowed.
If one forms the limit ∆t, ∆x → 0 so that the ratio µ = ∆t/∆x is constant, one
obtains the solution of the differential equation
ut = uxx − µutt ;
i.e., consistency is violated.
from kC(λ, ∆t)n k ≈ n that the result at some fixed T = n∆t > 0 contains errors, amplified
by T /∆t. If the initial values are such that the consistency error is of the second order O(∆t2 ),
we have still a discretisation error O(∆t) at t = T . As in §4.6, we have to take into consideration
T
floating point errors, which are also amplified by ∆t . Together, we obtain an error O(∆t2 + eps
∆t
),
eps: machine precision, so that we cannot obtain better results than O(eps2/3 ).
6.6 Consistency Versus Stability 135
6.5.7 Dissipativity for Parabolic Discretisations
Finally, we discuss the dissipativity (6.49) for discretisations of the heat equation
(6.4). Since the solution operator strongly damps high-frequency components, a
smoothing effect occurs: initial values u0 , which are only supposed to be contin-
uous (or to belong to L∞ ), lead to solutions u(t) of (6.5) which are infinitely often
differentiable for all t > 0. A corresponding condition for the discrete schemes is
condition (6.49), which in this case takes the form
2r
|G(ξ)| ≤ 1 − δ |ξ| for all |ξ| ≤ π, (6.53)
since G does not depend on x, and 1 × 1 matrices coincide with the eigenvalue.
Exercise 6.63. (a) The simplest scheme (6.18) satisfies (6.53) with the parameters
r = 1,
δ = min{4λ, 2 (1 − 2λ)}/π 2 for 0 < λ ≤ 1/2.
Dissipativity holds for 0 < λ < 1/2 because of δ > 0, but it fails for λ = 1/2.
(b) The Crank–Nicolson scheme (which is (6.41) with Θ = 1/2) is dissipative for
all λ > 0 with r = 1. For λ → ∞, the `2 stability is uniform (i.e., the stability
constant remains bounded for λ → ∞), but dissipativity vanishes (i.e., δ → 0).
6.6 Consistency Versus Stability
Scheme (6.20) is called consistent of order p if p is maximal with the property
u(t + ∆t) − C(λ, ∆t)u(t) = O(∆tp+1 )
for all sufficiently smooth solutions u of the differential equations.

The Taylor expansion applied to Example (6.17a) yields
u(t + ∆t, x) − (1 − aλ) u(t, x) − aλu(t, x + ∆x) =λ=∆t/∆x

∆t2 ∆x2

∆t ∆t
= u + ∆t ut + utt − 1 − a u−a u + ∆xux + uxx + . . .
2 ∆x ∆x 2
∆t2

∆x
= ∆t [ut − aux ] + utt − auxx + . . .
2 ∆t
Here the functions are to be evaluated at (t, x) and ‘. . .’ are higher-order terms. The
first bracket vanishes, since u is a solution of ut = aux . The second bracket is, in
general, different from zero, so that the discretisation (6.17a) has consistency order
p = 1.
Remark 6.64. There is an exceptional case in the derivation from above. Applying
the differential equation twice, we get
utt = a(ux )t = a(ut )x = a(aux )x = a2 uxx .
Therefore, utt − ∆x ∆t auxx = 0 proves a higher order of consistency if λ = 1/a.

µ
In fact, the difference scheme becomes Uνµ+1 = Uν+1 , and one verifies that the
discrete solution coincides with the exact one restricted to the grid (the consistency
order may be described by p = ∞). This case corresponds to condition (a) in
Theorem 6.47.
Also the other schemes (6.17b,c,d) are of first order consistency. A difference
scheme of the second order is the following Lax–Wendroff scheme (cf. [17, p. 221]):
h 2
i µ h (aλ)2 i
Uνµ+1 = (aλ) aλ µ 2 aλ µ

2 + 2 Uν−1 + 1 − (aλ) Uν + 2 − 2 Uν+1 . (6.54)
Second order consistency is seen from the Taylor expansion
u(t + ∆t, x)
∆t2 (aλ∆x)2
= u + ∆t ut + 2 utt + O(∆t3 ) = u + aλ∆x ux + 2 uxx + O(∆t3 )
= u(t, x) + aλ[u(t, x + ∆x) − u(t, x − ∆x)]
(aλ)2
+ 2 [u(t, x + ∆x) − 2u(t, x) + u(t, x − ∆x)] + O(∆t3 ).
Lemma 6.65. The Lax–Wendroff scheme (6.54) is `2 stable if and only if |aλ| ≤ 1.
Proof. Verify that G(ξ) = 1 − iaλ sin ξ − (aλ)2 (1 − cos ξ) is the characteristic
function. For
2 2
|G(ξ)| = [1 − τ (1 − cos ξ)] + τ sin 2 ξ with τ := (aλ)2
we use sin2 ξ = 1− cos2 ξ and substitute x := cos ξ. Then we have to prove that the
polynomial
2
p(x) = [1 − τ (1 − x)] + τ 1 − x2 = 1 + τ 2 − τ + 2(τ − τ 2 )x + τ 2 − τ x2

remains bounded by 1 for all values x = cos ξ ∈ [−1, 1]. Inserting x = −1 yields
2
(2τ − 1) and proves that τ ≤ 1 is necessary (τ ≥ 0 holds by definition).
2
Since by definition |G(ξ)| = p(x) ≥ 0, the maximum is given by
2
max{p(1), p(−1)} = max{1, (2τ − 1) } = 1,
where for the last step we used |aλ| ≤ 1. Now `2 stability follows from Theorem
6.44. tu
As in §5.5.6, we may ask whether there is a possible conflict between consistency
and stability. While in Theorem 5.47 the consistency order p is limited, it is now the
parity of p which need to be restricted in certain cases.
References 137
∂ ∂
We consider the hyperbolic problem ∂t u = a ∂x u and a general explicit
P differ-
ence scheme (6.16) with coefficients aj . Again, we introduce G(ξ) := j aj e−ijξ
(cf. (6.34)). Furthermore, we choose the Banach space `∞ and ask for `∞ stable
schemes. Using Theorem 6.47, Thomée [24] proves the following implication for
the consistency order.
Theorem 6.66. Under the assumption from above, `∞ stability implies that the con-
sistency order is odd.19 Furthermore, there are `∞ stable schemes for any odd order.
Since the Lax–Wendroff scheme has even consistency order (p = 2), it cannot
be `∞ stable. However, the instability is rather weak. Thomée [24] also proves the
following two-sided inequality for the Lax–Wendroff scheme:
C 0 n1/12 ≤ kC(λ, ∆t)n k`∞ ←`∞ ≤ C 00 n1/6 .
References
1. Courant, R., Friedrichs, K.O., Lewy, H.: Über die partiellen Differentialgleichungen der
mathematischen Physik. Math. Ann. 100, 32–74 (1928)
2. Crank, J., Nicolson, P.: A practical method for numerical evaluation of partial differential
equations of the heat-conduction type. Proc. Cambridge Phil. Soc. 43, 50–67 (1947). (reprint
in Adv. Comput. Math. 6, 207–226, 1996)
3. de Moura, C.A., Kubrusly, C.S. (eds.): The Courant–Friedrichs–Lewy (CFL) condition.
Springer, New York (2013)
4. Du Fort, E.C., Frankel, S.P.: Stability conditions in the numerical treatment of parabolic differ-
ential equations. Math. Tables and Other Aids to Computation 7, 135–152 (1953)
5. Friedrichs, K.O.: Symmetric hyperbolic linear differential equations. Comm. Pure Appl. Math.
7, 345–392 (1954)
6. Hackbusch, W.: Iterative Lösung großer schwachbesetzter Gleichungssysteme, 2nd ed.
Teubner, Stuttgart (1993)
7. Hackbusch, W.: Iterative solution of large sparse systems of equations. Springer, New York
(1994)
8. Hackbusch, W.: Elliptic differential equations. Theory and numerical treatment, Springer
Series in Computational Mathematics, vol. 18, 2nd ed. Springer, Berlin (2003)
9. Kreiss, H.O.: Über die Stabilitätsdefinition für Differenzengleichungen die partielle Differen-
tialgleichungen approximieren. BIT 2, 153–181 (1962)
10. Kreiss, H.O.: On difference approximations of the dissipative type for hyperbolic differential
equations. Comm. Pure Appl. Math. 17, 335–353 (1964)
11. Kreiss, H.O.: Difference approximations for the initial-boundary value problem for hyperbolic
differential equations. In: D. Greenspan (ed.) Numerical solution of nonlinear differential
equations, pp. 141–166. Wiley, New York (1966)
12. Kreiss, H.O.: Initial boundary value problem for hyperbolic systems. Comm. Pure Appl. Math.
23, 277–298 (1970)
13. Kreiss, H.O., Wu, L.: On the stability definition of difference approximations for the initial
boundary value problem. Appl. Numer. Math. 12, 213–227 (1993)
14. Kröner, D.: Numerical Schemes for Conservation Laws. J. Wiley und Teubner, Stuttgart
(1997)
19
This statement does not extend to the parabolic case.
15. Lax, P.D.: Difference approximations of linear differential equations – an operator theoretical
approach. In: N. Aronszajn, C.B. Morrey Jr (eds.) Lecture series of the symposium on partial
differential equations, pp. 33–66. Dept. Math., Univ. of Kansas (1957)
16. Lax, P.D., Richtmyer, R.D.: Survey of the stability of linear finite difference equations. Comm.
Pure Appl. Math. 9, 267–293 (1965)
17. Lax, P.D., Wendroff, B.: Systems of conservation laws. Comm. Pure Appl. Math. 13, 217–237
(1960)
18. Lax, P.D., Wendroff, B.: Difference schemes for hyperbolic equations with high order of ac-
curacy. Comm. Pure Appl. Math. 17, 381–398 (1964)
19. LeVeque, R.J.: Finite Volume Methods for Hyperbolic Problems. Cambridge University Press,
Cambridge (2002)
20. Mizohata, S.: The theory of partial differential equations. University Press, Cambridge (1973)
21. Richtmyer, R.D., Morton, K.W.: Difference Methods for Initial-value Problems, 2nd ed. John
Wiley & Sons, New York (1967). Reprint by Krieger Publ. Comp., Malabar, Florida, 1994
22. Riesz, F., Sz.-Nagy, B.: Functional Analysis. Dover Publ. Inc, New York (1990)
23. Riesz, M.: Sur les maxima des formes bilinéaires et sur les fonctionelles linéaires. Acta Math.
49, 465–497 (1926)
24. Thomée, V.: Stability of difference schemes in the maximum-norm. J. Differential Equations
1, 273–292 (1965)
25. Thorin, O.: An extension of the convexity theorem of M. Riesz. Kungl. Fysiogr. Sällsk. i Lund
Förh. 8(14) (1939)
26. Tomoeda, K.: Stability of Friedrichs’s scheme in the maximum norm for hyperbolic systems
in one space dimension. Appl. Math. Comput. 7, 313–320 (1980)
(2004)
Chapter 7
Stability for Discretisations of Elliptic Problems
In the previous chapter we treated partial differential equations of hyperbolic and

parabolic type. The third type of elliptic differential equations is considered in this
chapter. In the hyperbolic and parabolic cases, the solution operator is a mapping
from a Banach space B into itself. In the elliptic case the solution operator is
given by the inverse differential operator L−1 , which is a mapping between different
spaces. This requires introducing a pair (X, Y ) of spaces.
7.1 Elliptic Differential Equations
The prototype of elliptic differential equations is the Poisson equation
Lu := uxx + uyy = f in Ω , (7.1)
where Ω is a bounded domain in R2 . To obtain a unique solution, we have to add a

boundary condition on Γ := ∂Ω, e.g., the homogeneous Dirichlet data
u=0 on Γ.
Another choice would be the inhomogeneous condition u = g or conditions on

(normal) derivatives of u on Γ . The combination of a differential equation (e.g.,
(7.1)) with a boundary condition is called a boundary value problem.
A more general linear differential equations in d variables x = (x1 , . . . , xd ) is
d
X ∂ ∂
Lu = f with L = aij (x) + L0 , (7.2)
i,j=1
∂x i ∂x j
where L0 may contain further derivatives of order ≤ 1. L is called uniformly elliptic,

if

140 7 Stability for Discretisations of Elliptic Problems
d
X 2
aij (x)ξi ξj ≥ δ kξk2 for all x ∈ Ω and all ξ = (ξ1 , . . . , ξd ) ∈ Rd
i,j=1
with a positive constant δ.

We call u a classical solution of Lu = f and u|Γ = 0, if u ∈ C 2 (Ω) ∩ C0 (Ω),
where
C0 (Ω) = u ∈ C(Ω) : u = 0 on Γ .
As discussed in §7.4.2, in general, we cannot expect the solution to be a classical
one, even for a very smooth boundary.
7.2 Discretisation
As in §6.5.6.1, Ω is covered by a grid. In the case of classical difference schemes,

1
this is a Cartesian grid of step size h = hn = n+1 , while for finite element dis-
cretisations one uses an n-dimensional subspace of functions defined piecewise on
a more general triangulation. It is not necessary to require a discretisation for all
n ∈ N. Instead, we assume that there is an infinite subset
N0 ⊂ N,
so that the discretisation is defined for all n ∈ N0 .

2
The Poisson equation (7.1) in the square Ω = (0, 1) will serve as a model
example. The grid is
Ωn := {(νh, µh) ∈ Ω : ν, µ = 1, . . . , n}.
Instead of u(x, y) from (7.1), we are looking for approximations of u at the nodal
points (νh, µh) ∈ Ωn :
uν,µ ≈ u(νh, µh).
In each nodal point (νh, µh), the second x derivative uxx from (7.1) is replaced by
the second divided difference h12 [uν−1,µ − 2uν,µ + uν+1,µ ]. Correspondingly, uyy
becomes h12 [uν,µ−1 − 2uν,µ + uν,µ+1 ]. Together, we obtain the so-called five-point
scheme:
1
[−4uν,µ + uν−1,µ + uν+1,µ + uν,µ−1 + uν,µ+1 ] = fν,µ for all 1 ≤ ν, µ ≤ n,
h2
(7.3)
where fν,µ := f (νh, µh) is the evaluation1 of the right-hand side f of (7.1). If
ν = 1, equation (7.3) contains also the value uν−1,µ = u0,µ . Note that the point
(0, µh) lies on the boundary Γn := {(x, y) ∈ Γ : x/h, y/h ∈ Z}, and does not
1
A finite element discretisation using piecewise linear functions in a regular triangle grid yields
the same matrix, only the right-hand side fh consists of integral mean-values fν,µ instead of point
evaluations.
7.2 Discretisation 141
belong to Ωn . Because of the boundary condition u = 0, all values uν 0 ,µ0 in (7.3)

with (ν 0 h, µ0 h) ∈ Γh can be replaced by zero. As a result, one obtains a system of
n2 linear equations for the n2 unknowns {uν,µ : (νh, µh) ∈ Ωn }:
Ln un = fn , (7.4)
where un = (uν,µ )(νh,µh)∈Ωn and fn = (fν,µ )(νh,µh)∈Ωn .

Correspondingly, one can discretise more general differential equations as, e.g.,
(7.2) in even more complicated domains by difference or finite element methods
(cf. [8, 10]).
Remark 7.1. The matrix Ln from (7.4) and (7.3) has the following properties:
(a) Ln is sparse, in particular, it possesses at most five non-zero entries per row.
(b) Ln is symmetric.
4
(c) −Ln has positive diagonal elements h2 , while all off-diagonal entries are ≤ 0.
(d) the sum of entries in each row of −Ln is ≥ 0. More precisely: if 2 ≤ ν, µ ≤ n − 1,
the sum is 0; at the corner points (ν, µ) ∈ {(1, 1), (1, n), (n, 1), (n, n)} the sum
equals 2/h2 ; for the other points with ν or µ in {1, n} the sum is 1/h2 .
(e) For a concrete representation of the matrix Ln , one must order the com-
ponents of the vectors un , fn . A possible ordering is the lexicographical one:
(1, 1), (2, 1), . . . , (n, 1), (1, 2), . . . , (n, 2), . . . , (1, n), . . . , (n, n). In this case,
Ln has the block form
   
T I −4 1
I T I   1 −4 1 
1   .. .. ..
 
. . .

Ln = 2  . . .

with T = 
 .. .. .. ,

h 

  
 I T I   1 −4 1 
I T 1 −4
where all blocks T, I are of size n × n. Empty blocks and matrix entries are zero.
In the case of the matrix Ln from (7.3) (but not for any discretisation of elliptic
differential equations) Ln is in particular an M-matrix (cf. [7, §6.4], [8, §4.3]).
Definition 7.2 (M-matrix [17]). A matrix A ∈ Rn×n is called an M-matrix if

(a) Aii > 0 for all 1 ≤ i ≤ n,
(b) Aij ≤ 0 for all 1 ≤ i 6= j ≤ n,
(c) A regular, and A−1 has only non-negative entries: A−1 ij ≥ 0.

Remark 7.1c shows the properties (a) and (b) of A = −Ln . Remark 7.1d, to-
gether with the fact that Ln is irreducible, describes that A = −Ln is irreducibly
diagonal dominant. Irreducibly diagonal dominant matrices with the properties (a)
and (b) already possess property (c); i.e., −Ln is an M-matrix (cf. Hackbusch [7,
Theorem 6.4.10b]).
7.3 General Concept
7.3.1 Consistency
Rn
The consistency condition must ensure that Ln is a discretisa-
X
X −→ Xn tion of the differential operator L. For this purpose we intro-
↓L ↓ Ln duce Banach spaces for the domains and ranges of L and Ln :
X, Y, Xn , Yn , so that
PYn
Y Yn L : X → Y, Ln : Xn → Yn (n ∈ N0 ) (7.5a)
n
RY
are continuous mappings. Furthermore, let

n
RX : X → Xn , RYn : Y → Yn , fn = RYn f (n ∈ N0 ) (7.5b)
be ‘restrictions’ from the function spaces X, Y into the spaces Xn , Yn of ‘grid

functions’. Note that RYn defines the right-hand side fn in (7.4). A mapping in the
opposite direction is the ‘prolongation’
PYn : Yn → Y (n ∈ N0 ). (7.5c)
The consistency condition has to relate L and Ln . Since both mappings act in
n
different spaces, the auxiliary mappings RX , RYn , PYn are needed. They permit us to
formulate consistency in two versions. Condition (7.6a) measures the consistency
error by k · kYn , while (7.6b) uses k · kY :
n
lim k (Ln RX − RYn L) ukYn = 0 for all u ∈ X, (7.6a)
N0 3n→∞
lim k (PYn Ln RX
n
− L) ukY = 0 for all u ∈ X. (7.6b)
N0 3n→∞
The conditions (7.6a,b) are almost equivalent. For this purpose, some of the
following technical assumptions are of interest:
PYn RYn f → f for all f ∈ Y, (7.7a)

kPYn kY ←Yn ≤ CP , (7.7b)
kfn kYn ≤ CP0 kPYn fn kY for all fh ∈ Yh , (7.7c)
RYn PYn = I, (7.7d)
kRYn kYn ←Y ≤ CR . (7.7e)
In the case of (7.7d), PYn is a right-inverse of RYn .
Exercise 7.3. Suppositions (7.7d,e) imply (7.7c) with CP0 = CR .
Under the conditions (7.7a-c), the consistency formulations (7.6a,b) are equiva-
lent; more precisely:
7.3 General Concept 143
Proposition 7.4. (a) If (7.7a,b), then (7.6a) implies (7.6b).

(b) If (7.7a,c), then (7.6b) implies (7.6a).
Proof. (i) Condition (7.7a) implies (PYn RYn − I) Lu → 0. Using the splitting
(PYn Ln RX
n
− L) u = PYn (Ln RX
n
− RYn L) u + (PYn RYn − I) Lu,
we observe equivalence in the following statements:
(PYn Ln RX
n
− L) u → 0 ⇔ PYn (Ln RX
n
− RYn L) u → 0.
(ii) (7.6a) and (7.7b) imply PYn (Ln RX
n
− RYn L) u → 0 and, by (i), (7.6b).
(iii) (7.7a) and (7.6b) imply PY (Ln RX −RYn L) u → 0. Now assumption (7.7c)
n n
proves (7.6a). t u
7.3.2 Convergence
In the sequel, a further norm k · kX̂n on X can be fixed, which may be weaker than
k · kXn (or equal), but stronger than k · kYn (or equal):2
k · kXn & k · kX̂n & k · kYn .
The discrete space X̂n should correspond to a space X̂ with X ⊂ X̂ ⊂ {u ∈ Y :

u|Γ = 0}. The correspondence is expressed by
n
kRX kX̂n ←X̂ ≤ ĈR for all n. (7.8)
Convergence should express that any solution u ∈ X of the differential equation

Lu = f (with boundary condition u = 0) leads to a sequence {un } of discrete
solutions tending to u. Since the discrete grid function un ∈ Xn and the function
n
u ∈ X belong to difference spaces, we compare the restriction RX u ∈ Xn with
n
un and require kRX u − un kX̂n → 0. The corresponding discrete solution takes the
form un = L−1 −1 n −1 n
n fn = Ln RY f = Ln RY Lu. We allow that Ln is not invertible
for finitely many n. Restricting N0 to the remaining set, we suppose in what follows
that L−1 0
n exists for all n ∈ N .
The discussion from above leads to the following definition.
Definition 7.5 (convergence). With X and X̂n as above, the sequence {Ln , RYn }
defining the discretisation is called convergent if
n
k(RX − L−1 n
n RY L)ukX̂n → 0 for all u ∈ X and N0 3 n → ∞. (7.9)
An alternative definition of convergence is the following. We replace u ∈ X by

f = Lu ∈ range(L) ⊂ Y , and require convergence for all f ∈ Y , i.e., for all
solutions u = L−1 f with f ∈ Y :
n −1
k(RX L − L−1 n
n RY )f kX̂n → 0 for all f ∈ Y and N0 3 n → ∞. (7.10)
2
An & Bn means that there is a constant C, independent of n, such that An ≥ CBn .
Lemma 7.6. (a) Statement (7.10) implies (7.9).

(b) If L : X → Y is bijective, both characterisations (7.9) and (7.10) are equivalent.
n −1
Proof. Let u ∈ X be arbitrary. For f := Lu, (RX L − L−1 n
n RY )f from (7.10)
n −1 n
equals (RX − Ln RY L)u, proving (7.9), and therefore, Part (a).
In the bijective case, any f ∈ Y has the representation f = Lu for some u ∈ X.
The insertion proves Part (b). t u
7.3.3 Stability
The following definition of stability depends on the choice of X̂n .
Definition 7.7. Stability is characterised by the uniform boundedness of L−1

n :
sup kL−1
n kX̂n ←Yn ≤ Cstab . (7.11)
N0 3n→∞
Theorem 7.8 (convergence theorem). Consistency (7.6a) of u ∈ X and stability

(7.11) imply convergence (7.9) to u ∈ X.
n
Proof. (RX − L−1 n −1 n n
n RY L)u = Ln (Ln RX − RY L) u yields
n
k(RX − L−1 n −1 n n
n RY L)ukX̂n ≤ kLn kXn ←Yn k (Ln RX − RY L) ukYn
n
≤ Cstab k (Ln RX − RYn L) ukYn → 0,
proving convergence (7.9). t

u
Although the setting is similar to the previous chapters, a stability theorem stating
that ‘convergence implies stability’ is, in general, not valid, as we shall prove in
§7.4.2. However, the stability theorem can be based on the convergence definition
in (7.10). Again, the following technical requirement comes into play:
kfn kYn 0
sup ≥ 1/CR for all 0 6= fn ∈ Yn , n ∈ N0 . (7.12)
f ∈Y n
: fn =RY f kf kY
0
Lemma 7.9. The conditions (7.7b,d) imply (7.12) with CR = CP .
kfn kYn 1 kfn kYn
Proof. f := PYn fn yields sup{. . .} ≥ kPYn fn kY ≥ CP kfn kYn = 1/CP . t
u
Theorem 7.10 (stability theorem). Suppose (7.8), kL−1 kX̂←Y ≤ CL−1 , and
(7.12). Then convergence in the sense of (7.10) implies stability (7.11).
Proof. Inequality (7.10) describes the point-wise convergence of the operator

n −1
RX L − L−1 n
n RY ; hence, by Corollary 3.39, the operator norm is uniformly
bounded: kRX L − L−1
n −1 n
n RY kX̂n ←Y ≤ C. We conclude that
7.4 Application to Difference Schemes 145
kL−1 n n −1
n RY kX̂n ←Y ≤ C + kRX L kX̂n ←Yn ≤ C + ĈR CL−1 .
0
Inequality (7.12) implies kRYn f kYn ≤ CR kf kY . The definition of the operator
norm together with (7.12) yields
kL−1 n
n RY f kX̂n kL−1 n
n RY f kX̂n kRY f kYn
n
kL−1 n
n RY kX̂n ←Y = sup = sup n
06=f ∈Y kf kY 06=f ∈Y kRY f kYn kf kY
kL−1
n fn kX̂n kfn kYn kL−1
n kX̂n ←Yn
= sup sup ≥ 0 ,
nf
06=fn ∈Yn f ∈Y : fn =RY kfn kYn kf kY CR
proving that kL−1

n kX̂n ←Yn is bounded independently of n. t
u
Since the convergence definition (7.10) allows us to derive stability, it seems to be

the more suitable definition. However, without bijectivity of L : X → Y , definition
(7.10) does not lead to a convergence theorem as in Theorem 7.8. In §7.4.2 we
shall study bijectivity of L in more detail and gives examples, where bijectivity
fails because of range(L) $ Y . On the other hand, fixing Y and replacing X by
{L−1 f : f ∈ Y } % X would be rather impractical, since it require a consistency
condition involving generalised solutions (cf. §7.4.2).
Only if bijectivity holds, can the equivalence theorem be formulated.
Theorem 7.11 (equivalence theorem). Suppose that the spaces X̂, X, and Y are
chosen such that L : X → Y is bijective, kL−1 kX̂←Y ≤ CL−1 , and (7.8) and
(7.12) hold. Then convergence (7.9) and stability (7.11) are equivalent.
Proof. The direction ‘stability ⇒ convergence’ is stated in Theorem 7.8. Because

of bijectivity; (7.9) implies (7.10) (cf. Lemma 7.6), and Theorem 7.10 shows
‘(7.10) ⇒ stability’. tu
Remark 7.12. Consistency and convergence can be regarded in two different ways.
The previous setting asked for the respective conditions for all u ∈ X. On the other
hand, we can consider one particular solution u = L−1 f . Because of extra smooth-
ness, the consistency error εn := k (Ln RX n
− RYn L) ukYn may behave as O(n−α )
for some α > 0. Then the proof of Theorem 7.8 shows that also the discretisation
error k(RX n
− L−1 n
n RY L)ukX̂n ≤ Cstab εn is of the same kind.
7.4 Application to Difference Schemes
7.4.1 Classical Choice of Norms
The concrete choice of the Banach spaces depends on the kind of discretisation.
The following spaces correspond to difference methods and classical solutions of
boundary value problems:
X = C 2 (Ω) ∩ C0 (Ω),
Y = C(Ω),
2 (7.13a)
Xn = Rn , Xn 3 un = (uν,µ )(νh,µh)∈Ωn ,
2
Yn = Rn , Yn 3 fn = (fν,µ )(νh,µh)∈Ωn ,
with the norms3
kukX = max{kuk∞ , kux k∞ , kuy k∞ , kuxx k∞ , kuyy k∞ },
kukY = kuk∞ := max{|u(x, y)| : (x, y) ∈ Ω},
(7.13b)
kun kXn = max{kun k∞ , k∂x un k∞ , k∂y un k∞ , k∂xx un k∞ , k∂yy un k∞ },
kfn kYn = kfn k∞ := max{|uν,µ | : (νh, µh) ∈ Ωn },
where ∂x , ∂y , ∂xx , ∂yy are the first and second divided difference quotients with
step size h := 1/n. The associated restrictions are the point-wise restrictions:
n
(RX u)ν,µ := u(νh, µh) , (RYn f )ν,µ := f (νh, µh) .
n
The components of (Ln RX − RYn L)u are
[∂xx + ∂yy ] u(x, y) − [uxx (x, y) + uyy (x, y)] for (x, y) ∈ Ωn .
Since, for u ∈ X0 ⊂ C 2 (Ω) ∩ C0 (Ω), the second differences ∂xx u converge uni-
n
formly to the second derivative uxx , it follows that k(Ln RX − RYn L)ukXn → 0
for n → 0. Therefore, the consistency condition (7.6a) is verified.
Consistency follows immediately for X0 = X = C 2 (Ω) ∩ C0 (Ω), since the
n
difference quotients in Ln RX u tend uniformly to the derivatives in RYn Lu.
The additional space X̂ = C0 (Ω) is equipped with the maximum norm
kun kX̂n = kun kYn = kun k∞ .
Therefore, kL−1 −1
n kX̂n ←Yn becomes the row-sum norm kLn k∞ . For the estimate of
−1
kLn k∞ in the model example (7.3), we use that −Ln is an M-matrix (cf. Definition
7.2). For M-matrices there is a constructive way to determine the row-sum norm of
the inverse. In the following lemma, 1I is the vector with all entries of value one.
Lemma 7.13. Let A be an M-matrix and w a vector such that the inequality
Aw ≥1I holds component-wise. Then kA−1 k∞ ≤ kwk∞ holds.
n
Proof. For u ∈ Rn , the vector (|ui |)i=1 is denoted by |u|. The following inequal-
ities are to be understood component-wise. We have |u| ≤ kuk∞ 1I ≤ kuk∞ Aw.
Because of the M-matrix property (c), A−1 ≥ 0 holds, so that
−1
A u ≤ A−1 |u| ≤ A−1 kuk Aw = kuk w
∞ ∞
and kA−1 uk∞ / kuk∞ ≤ kwk∞ can be obtained. Therefore, the desired estimate
follows: kA−1 k∞ = supu6=0 kA−1 uk∞ / kuk∞ ≤ kwk∞ . t
u
3
Only the norm of Yh appears in (7.6a) and (7.11).
7.4 Application to Difference Schemes 147
In the case of A = −Ln one has to look for a function w(x, y) with Lw(x, y) ≥ 1.
A possible solution is w(x, y) = 12 x (1 − x) with the maximum norm kwk∞ = 1/8.
n
Then the point-wise restriction yields the vector wh = RX w on the grid Ωn . Since
second differences and second derivatives are identical in the case of a quadratic
function, it follows that (−Ln ) wh ≥ 1I and kwh k∞ ≤ 1/8, proving the stability
property
kL−1
n k∞ ≤ 1/8 for all n ∈ N (7.14)
with Cstab = 1/8.
Using the consistency, which is checked above, and the stability result (7.14), we
obtain convergence by Theorem 7.8. As discussed in §7.4.2, the given theory does
not fully correspond to the previous setting, because convergence does not imply
stability.
Another aspect is the order of consistency. So far, only the convergence order
n
o(1) is shown. In general, one likes to show that kRX u − un kXn = O(hκ ) for
some κ > 0. For this purpose, the solution u ∈ X must have additional smoothness
properties: u ∈ Z for some Z ⊂ X = C 2 (Ω) ∩ C0 (Ω). Choosing
Z := X ∩ C 4 (Ω),
n
we conclude that k (Ln RX − RYn L) ukYn = O(h2 ) for all u ∈ Z. Correspondingly,
convergence follows with the same order:
n
kRX u − un kX̂n = O(h2 ).
7.4.2 Bijectivity of L
By choosing the spaces X, Y , the operator L is bounded as a mapping from X into

Y . Assuming uniqueness of solutions, injectivity of L ∈ L(X, Y ) holds. This fact
guarantees the existence of L−1 ∈ L(range(L), X). If, in addition, L is surjective,
the inverse L−1 is a continuous mapping from Y onto X.
Theorem 7.14. The differential operators L from (7.1) or (7.2) belong to

L(C 2 (Ω) ∩ C0 (Ω), C(Ω)), but are not surjective.
The proof is given in [8, 10, Theorem 3.2.7]. The consequences are:
1. The definitions (7.9) and (7.10) of convergence are not equivalent.
2. If L is injective, L−1 exists as bounded mapping from C(Ω), e.g., into
X̂ = C 1 (Ω) ∩ C0 (Ω) or X̂ = C0 (Ω), but it does not belong to the space
L(C(Ω), C 2 (Ω) ∩ C0 (Ω)).
Concerning the first statement, note that the characterisation (7.9) refers to all
solutions in C 2 (Ω) ∩ C0 (Ω), whereas (7.10) refers to all solutions u = L−1 f for
f ∈ C(Ω).
For the proof of the second statement, we have to introduce the generalised
solutions u = L−1 f , which do not belong to C 2 (Ω) ∩ C0 (Ω). For this purpose,
one can use Green’s representation
Z
u(x) = G(x, y)f (y)dy. (7.15)
Ω
The Green function G(·, ·) satisfies LG(·, y) = 0 for all y ∈ Ω and G(x, ·) = 0 for
x ∈ Γ . In the case of Poisson’s equation (7.1) in the circle Ω = {kxk < 1} ⊂ R2 ,
" #
1 1
G(x, y) = − log kx − yk − log kxk y − 2x

2π kxk
∂
holds (cf. §2.2 in [8, 10]). Since G as well as ∂x i
G(x, y) has an integrable singular-
ity, the integral in (7.15) exists and defines a function in X̂ = C 1 (Ω) ∩ C0 (Ω). In
particular, X̂ = C0 (Ω) leads to the (finite) operator norm
Z
−1
L
C(Ω)←C(Ω)
= |G(x, y)| dy.
Ω
If f is Hölder continuous, the function u defined in (7.15) belongs to C 2 and is a

solution of Lu = f . This proves that (7.15) defines generalised solutions in X̂ and
that L−1 ∈ L(C(Ω), X̂).
To apply the equivalence Theorem 7.11, we have to look for spaces X, Y , such
that L : X → Y is bijective. An example of classical function spaces are the
Hölder spaces Y = C λ (Ω), 0 < λ < 1, and X = C 2+λ (Ω). The (unusual)
discrete norm on Yn is defined by the maximum over all |f (x)|, x ∈ Ωn , and
λ
|f (x) − f (y)| / kx − yk , x, y ∈ Ωn .
A canonical choice of X, Y with a bijective L : X → Y , obtained from the
variational setting, are the Sobolev spaces X = H01 (Ω) and Y = H −1 (Ω) (cf.
§7.5). Under suitable conditions (e.g., smooth coefficients aij in (7.2) and convex
Ω), also the choices of X = H 2 (Ω) ∩ H01 (Ω) and Y = L2 (Ω) are valid.
The previous convexity condition excludes domains Ω with reentrant corners.
A class of domains allowing reentrant corners are Lipschitz domains (i.e., there must
be a Lipschitz continuous parametrisation of Γ ). Nečas [15] proves the following
regularity result.
Theorem 7.15. Let Ω be a bounded Lipschitz domain, and suppose that L is uni-
formly elliptic with Hölder continuous coefficients aij ∈ C t (Ω) of order t ∈ (0, 12 ].
Then L : X → Y and L−1 : Y → X are bounded for the Sobolev spaces4
X = H01+s (Ω), Y = H −1+s (Ω) for any s with 0 ≤ s < t ≤ 1/2.
Note that s = 0 yields the trivial statement L : H01 (Ω) H −1 (Ω). By Theorem
3/2−ε
7.15, smooth functions f produce solutions in the space H0 (Ω) for all ε > 0.
4
The Sobolev spaces H t (Ω) and H0t (Ω) are introduced in [8, 10].
7.5 Finite Element Discretisation 149
It is possible to discretise the boundary value problem by a difference scheme

and to analyse Ln with respect to discrete Sobolev spaces Xn and Yn of the re-
spective order 1 + s and −1 + s (cf. Hackbusch [6]). Then, the stability estimate
kL−1 n n
n kYn ←Xn ≤ Cstab holds as well as consistency k (Ln RX − RY L) ukYn → 0
1+s
for all u ∈ X = H0 (Ω). Regularity properties of difference schemes are also
discussed in Jovanovič–Süli [13].
7.5 Finite Element Discretisation
7.5.1 Variational Problem
The variation formulation of the boundary value problem (7.1) and (7.2) is based on
a bilinear form:
Z
a(u, v) = − f (x)v(x)dx for all v ∈ H01 (Ω), where (7.16a)
Ω
Z X
d ∂u(x) ∂v(x)
a(u, v) := dx. (7.16b)
Ω j=1 ∂xj ∂xj
The Sobolev space H01 (Ω) can be understood as the completion of C 1 (Ω) ∩ C0 (Ω)
with respect to the norm
v
uZ d 2
u 2
X ∂
kukH 1 := t |u(x)| +
∂xj u(x) dx

Ω j=1
(cf. §6.2 in [8, 10]). The dual space H −1 (Ω) consists of all functionals with finite
norm
kf kH −1 = sup{|f (v)| : v ∈ H01 (Ω), kvkH 1 = 1}.
The second embedding in H01 (Ω) ⊂ L2 (Ω) ⊂ H −1 (Ω) Ris based on the identifica-
tion of functions f ∈ L2 (Ω) with the functional f (v) := Ω f (x)v(x)dx.
If the bilinear form a(·, ·) is bounded, i.e.,
Ca := sup{a(u, v) : u, v ∈ H01 (Ω), kukH 1 = kvkH 1 = 1} < ∞, (7.17)
and problem (7.16a) is identical to the abstract equation
Au = f with A ∈ L(H01 (Ω), H −1 (Ω)), u ∈ H01 (Ω), f ∈ H −1 (Ω).
If, in general, a is a bilinear form on X × X, the operator A ∈ L(X, X ∗ ) (X ∗ dual

space of X) is defined by
(Au) (v) := a(u, v) for all u, v ∈ X. (7.18)
The norm kAkH 1 ←H −1 coincides with Ca from above. The particular example
(7.16b) satisfies Ca < 1.
If A−1 ∈ L(H −1 (Ω), H01 (Ω)) exists, u := A−1 f is the desired solution (it is
called a ‘weak solution’). The existence of A−1 can be expressed by the so-called
inf-sup conditions for the bilinear form a (more precisely, the inf-sup expression is
an equivalent formulation of 1/kA−1 kH 1 ←H −1 ; cf. §6.5 in [8, 10]).
A very convenient, sufficient condition is the H01 (Ω)-coercivity5
2
a(u, u) ≥ εco kukH 1 with εco > 0 for all u ∈ H01 (Ω). (7.19)
Above, we used the spaces X, Y together with the differential operator L. Now
L will be replaced by A, involving the spaces
X := H01 (Ω), Y := H −1 (Ω).
Note that Y = X ∗ is the dual space with the dual norm
kf kX ∗ := sup{|f (u)| : u ∈ X, kukX ≤ 1}.
As (7.19) is valid for (7.16b), A ∈ L(X, Y ) is bijective; i.e., A−1 ∈ L(Y, X) exists.
7.5.2 Galerkin Discretisation
The Galerkin discretisation is completely characterised by a subspace Un ⊂ X of

dimension n. We consider a family {Un }n∈N0 , where N0 ⊂ N is an infinite subset.
The finite element method is a particular example, where Un consists of piecewise
polynomials. The vector space Un can be equipped with different norms. We set
Xn := Un equipped with the norm k·kH 1 . The discrete solution un ∈ Xn is defined
by
Z
a(un , vn ) = − f (x)vn (x)dx for all vn ∈ Xn .
Ω
The restriction of a bilinear form a(·, ·) on X × X to a subspace Xn × Xn yields

again a bilinear form an (·, ·) on Xn × Xn . If a is bounded by Ca (cf. (7.17)), the
bound Can of an satisfies Can ≤ Ca . In the same way as A ∈ L(X, X ∗ ) is uniquely
determined by a, the bilinear form an corresponds to an operator
An ∈ L(Xn , Xn∗ ). (7.20)
Note that Xn∗ is Un equipped with the norm dual to Xn .

5
The algebraic counterpart is the following statement: if A + AH is positive definite, then A is
regular. More precisely, A + AH ≥ 2εco I implies kA−1 k2 ≤ 1/εco . In the case of A = AH ,
kA−1 k2 = 1/εco holds.
Exercise 7.16. The dual norm of Xn∗ is not the restriction of the dual norm of X ∗ to
Un . Prove that kfn kXn∗ ≤ kfn kX ∗ for fn ∈ Xn∗ .
Remark 7.17. Let Xn ⊂ X. If a(·, ·) is X-coercive with constant ε (cf. (7.19)),

an (·, ·) is Xn -coercive with a constant εn,co satisfying εn,co ≥ εco . As a conse-
quence, X-coercive bilinear forms a(·, ·) lead to stable Galerkin approximations
such that A−1 ∗
n ∈ L(Xn , Xn ) exists. In the symmetric case,
kA−1
n kXn ←Xn
∗ ≤ kA
−1
kX←X ∗
holds.
7.5.3 Consistency
The main consistency condition is described in terms of the subspaces Un : the

distance between Un and any u ∈ X has to satisfy
inf{ku − vn kX : vn ∈ Un } → 0 for n → ∞ and all u ∈ X.
An equivalent statement is the point-wise convergence
ku − Πn ukX → 0 for all u ∈ X as n → ∞, (7.21)
where Πn : X → Xn is the X-orthogonal projection onto Xn . The projection may

be considered as an element of L(X, X) as well as L(X, Xn ). The dual operator
Πn∗ then belongs to L(X ∗ , X ∗ ) as well as L(Xn∗ , X ∗ ).
Remark 7.18. (a) The Riesz isomorphism6 J : X → X ∗ is a unitary operator; i.e.,
J ∗ = J −1 . For any f ∈ X ∗ , the action f (u), u ∈ X, can be described by the scalar
product hJ ∗ f, uiX . X ∗ is a Hilbert space with the scalar product
hf, giX ∗ := hJ ∗ f, J ∗ giX .
(b) Let X = Xn ⊕ Xn⊥ be an X-orthogonal decomposition and set Yn := JXn ,

Yn⊥ := JXn⊥ . Then X ∗ = Yn ⊕ Yn⊥ is an X ∗ -orthogonal decomposition.
(c) The X-orthogonal projection Πn ∈ L(X, X) onto Xn satisfies
JΠn = Πn∗ J;
i.e., the dual projection has the representation Πn∗ = JΠn J ∗ = JΠn J −1 .
(d) Property (7.21) is equivalent to
kv − Πn∗ vkX ∗ → 0 for all v ∈ X ∗ as n → ∞. (7.22)

6
Also called the Fréchet–Riesz isomorphism, since both, Fréchet [4] and Frigyes Riesz [19]
published their results in 1907. See also [20, II.30].
Proof. For Part (b) write yn ∈ Yn and yn⊥ ∈ Yn⊥ as yn = Jxn and yn⊥ = Jx⊥ n
for suitable xn ∈ Xn and x⊥ ⊥
n ∈ Xn . Then
yn , yn⊥ X ∗ = J ∗ yn , J ∗ yn⊥ X = J −1 yn , J −1 yn⊥ X = xn , x⊥

n X = 0.
For Part (c) verify that Πn∗ maps y = yn + yn⊥ into yn (with notations yn ∈ Yn
and yn⊥ ∈ Yn⊥ as in Part (b)).
For Part (d) with v = Ju use kv − Πn∗ vkX ∗ = k (I − Πn∗ ) JukX ∗ =Part (c)
kJ (I − Πn ) ukX ∗ = k (I − Πn ) ukX . t u
We recall the operators A from (7.18) and An from (7.20).
Remark 7.19. The relation between A ∈ L(X, X ∗ ) and An ∈ L(Xn , Xn∗ ) is given
by
An = Πn∗ AΠn .
Proof. This follows from hAn un , vn iX ∗ ×Xn = a(un , vn ) = a(Πn un , Πn vn ) =
n
hAΠn un ,Πn vn iX ∗ ×X = hΠn∗ AΠn un , vn iX ∗ ×X for all un , vn ∈ Xn . t
u
The canonical choice of the space Yn is Xn∗ . The mappings between X, Y, Xn , Yn
are as follows:
Y := X ∗ , Yn := Xn∗ , (7.23a)
n
RX = Πn : X → Xn X-orthogonal projection onto Xn , (7.23b)
RYn : Y → Yn restriction to Xn ⊂ X; i.e., fn = RYn f = f |Xn ∈ Xn∗ , (7.23c)
n ∗
PYn := (RX ) = Πn∗ = JΠn J ∗ : Yn → Y. (7.23d)
n
We consider RX as a mapping onto Xn (not into X). Concerning RYn , note that
the mapping f ∈ Y = X ∗ can be restricted to the subspace Xn ⊂ X. For PYn = Πn∗
compare Remark 7.19a.
Lemma 7.20. Assumption (7.21) implies the consistency statement (7.6a).
n
Proof. Application of the functional (An RX − RYn A) u to vn ∈ Xn yields
n
[(An RX − RYn A) u] (vn ) = an (RX
n
u, vn ) − a(u, vn ) = a(Πn u, vn ) − a(u, vn )
= a(Πn u − u, vn ),
n
proving k (Ln RX − RYn L) ukYn ≤ Ca ku − Πn ukX → 0 with Ca from (7.17)
because of (7.21). t
u
Next, we verify that all conditions (7.7a–e) are valid. The constants are
CP = CP0 = CR = 1. Concerning the proof of (7.7a), we note that
PYn RYn f − f = Πn∗ f − f → 0.

(7.22)
Hence, the consistency statement (7.6a) is equivalent to (7.6b) (cf. Proposition 7.4a),
which reads (PYn An RXn
− A) u = (Πn∗ An Πn − A)u → 0 in Y .
7.5.4 Convergence and Stability
We assume that A is invertible; i.e., the variational problem (7.16a) is uniquely

solvable. First we set X̂n := Xn with norm k · kX̂n = k · kXn . For further conclu-
sions, we have to ensure stability. For simplicity, we assume that the bilinear form is
X-coercive (cf. (7.19)). Then, as stated in Remark 7.17, εn,co ≥ εco proves stability.
From consistency (7.6a), using Theorem 7.8, we infer convergence in both (7.9)
and (7.10), since A : X → Y is bijective (cf. Lemma 7.6b). The convergence
statement (7.10) takes the form
n −1
(RX A − A−1 n
n RY )f = (Πn A
−1
− A−1
n )f
= Πn u − un → 0 in X for N0 3 n → ∞,
where u and un are the exact and discrete solutions, respectively. This is not the
standard Galerkin convergence statement
u − un → 0 in X for N0 3 n → ∞,
but the latter statement follows immediately from condition (7.21): u − Πn u → 0

in X.
7.5.5 Quantitative Discretisation Error and Regularity
Finite element discretisation uses the subspace Un ⊂ X = H01 (Ω) of, e.g., piece-
wise linear functions on a triangulation. Let h = hn be the largest diameter of
the involved triangles (or other elements, e.g., tetrahedra in the three-dimensional
case). In the standard case, hn = O(n−1/d ) is to be expected (d: spatial dimension,
Ω ⊂ Rd ), provided that n is related to the dimension: n = dim(Un ).
The consistency error is bounded by Ca ku − Πn ukX as shown by the proof
of Lemma 7.20. According to Remark 7.12, the discretisation error ku − un kX
can be estimated by7 (1 + Cstab Ca ) ku − Πn ukX . While u ∈ X = H01 (Ω) allows
only point-wise convergence, quantitative error bound can be expected for smoother
functions u. Assume, e.g., u ∈ H 2 (Ω) ∩ H01 (Ω). Then the finite element error can
be proved to be ku − un kH01 (Ω) = O(h). More generally, finite elements of degree
p lead to an error8
ku − un kH01 (Ω) = O(hmin(p,t−1)

n ), if u ∈ H t (Ω) ∩ H01 (Ω), t > 1. (7.24)
The property u ∈ H t (Ω) may hold by accident or for systematic reasons

(‘regularity properties’). Solvability of the boundary value problem is equivalent
7
This is the result of Cea’s lemma, which is usually proved differently (cf. Theorem 8.2.1 in
[8, 10]).
8
See Footnote 4 on page 148.
to A−1 ∈ L(H −1 (Ω), H01 (Ω)) (cf. §7.5.1). It describes that the solution process
increases the degree of smoothness by 2. One may ask whether a similar statement
A−1 ∈ L(H t−2 (Ω), H t (Ω) ∩ H01 (Ω))
holds for certain t > 1. In this case, the problem is called t-regular. Such regularity
statements depend on the smoothness of the coefficients of the differential operator
(7.2) and on the smoothness of the boundary. Theorem 7.15 yields t-regularity for
1 < t < 3/2 under the condition that Ω is a Lipschitz domain. For convex domains
(or domains that are smooth images of convex domains) and sufficiently smooth
coefficients, the problem is 2-regular:
A−1 ∈ L(L2 (Ω), H 2 (Ω) ∩ H01 (Ω)).
Piecewise linear finite elements are of degree p = 1. Assuming 2-regularity,

f ∈ L2 (Ω) ensures u ∈ H 2 (Ω)∩H01 (Ω), so that ku−un kH01 (Ω) = C2 hn kf kL2 (Ω)
follows from (7.24).
7.5.6 L2 Error
The adjoint operator A∗ ∈ L(X, X ∗ ) belongs to the adjoint bilinear form a∗ defined
by a∗ (u, v) := a(v, u). In this section we assume that A∗ is 2-regular.
The representation u − un = A−1 f − A−1
n Πn f = (A
−1
− A−1
n Πn )f shows that
En := A−1 − A−1
n Πn = A
−1
− Πn A−1
n Πn
can be considered as error operator with the property
kEn f kH01 (Ω) ≤ C2 hn kf kL2 (Ω) ,
where C2 depends on details of the finite element triangulation (cf. [8, 10]).
Lemma 7.21. En AEn = En holds for Galerkin discretisation.
Proof. The identity
En AEn = (A−1 − A−1

n Πn )A(A
−1
− Πn A−1
n Πn )
= A−1 − 2A−1 −1 −1
n Πn + An Πn AΠn An Πn = En
follows from Πn AΠn = An . t

u
Theorem 7.22. Assume that A∗ is 2-regular, and kEn∗ gkH01 (Ω) ≤ C2 hn kgkL2 (Ω) .
Then
ku − un kL2 (Ω) = kEn f kL2 (Ω) ≤ Ca C2 hn kEn f kH01 (Ω) .
Proof. For any g ∈ L2 (Ω) with kgkL2 (Ω) = 1 and f ∈ X ∗ , we have
hEn f, giX×X ∗ = hEn AEn f, giX×X ∗ = hAEn f, En∗ giX ∗ ×X

Lemma 7.21
= a(En f, En∗ g),
∗

hEn f, gi
X×X ∗ ≤ Ca kEn f kX kEn gkX ≤ Ca C2 hn kEn f kX .

By identification L2 (Ω) = (L2 (Ω))∗ , we can continue with
hEf, giX×X ∗ = g(Ef ) = (Ef, g)L2 (Ω)
for g ∈ L2 (Ω). Since
kEn f kL2 (Ω) = max{| (Ef, g)L2 (Ω) | : g ∈ L2 (Ω), kgkL2 (Ω) = 1},
the assertion follows. t

u
Theorem 7.22 states that the L2 error is by one factor of hn better than the H 1
error kEn f kH01 (Ω) = ku − un kH01 (Ω) . This result, which traces back to Aubin [1]
and Nitsche [16], is usually proved differently, making indirect use of Lemma 7.21.
7.5.7 Stability of Saddle Point Problems
If coercivity (7.19) holds and if a is symmetric: a(u, v) = a(v, u), the variational
formulation (7.16a) is equivalent to the minimisation of J(u) := 12 a(u, u) − f (u).
Quite another type is the following saddle point problem. We are looking for
functions v ∈ V and w ∈ W (V, W Hilbert spaces) satisfying
a(v, v 0 ) + b(w, v 0 ) = f1 (v 0 ) for all v 0 ∈ V,

(7.25)
b(y, w0 ) = f2 (w0 ) for all w0 ∈ W,
with bilinear forms a : V × V → R and b : W × V → R.

An example of the general form (7.25) is Stokes’ problem, where
Z
V = (H01 (Ω))d , W = {f ∈ L2 (Ω) : f dξ = 0},
Z ZΩ
a(v, v 0 ) = h∇v, ∇xi dξ, b(w, v 0 ) = w div v 0 dξ. (7.26)
Ω Ω
Define
J(v, w) := 12 a(v, v) + b(w, v) − f1 (v) − f2 (w).
The following saddle point properties are proved in [8, 10, Theorem 12.2.4].
Theorem 7.23. Let a(·, ·) be symmetric and coercive. (v, w) ∈ V × W is a solution

of (7.25), if and only if
J(v, w0 ) ≤ J(v, w) ≤ J(v 0 , w) for all v 0 ∈ V, w0 ∈ W.
Furthermore,
J(v, w) = min
0
J(v 0 , w) = max
0
min
0
J(v 0 , w0 ).
v ∈V w ∈W v ∈V
The bilinear forms a and b in (7.25) give rise to operators A ∈ L(V, V ∗ ) and
B ∈ L(W, V ∗ ). The variational setting (7.25) is equivalent to the operator equation

A B v f
= 1 .
B∗ 0 w f2
Solvability of the problem is equivalent to the existence of the inverse of

A B
C := .
B∗ 0
Since C is symmetric, its eigenvalues are real. In the case of a symmetric and coer-
cive a(·, ·), all eigenvalues of the operator A are positive. However, C is indefinite;
i.e., it has positive and negative eigenvalues. One can show that the discrete operator
An has eigenvalues larger or equal to those of A (Lemma 7.17 states this result for
the smallest eigenvalue). In the indefinite case, such a simple result is not available.
To the contrary, a wrong type of discretisation can easily generate zero eigenvalues
of Cn ; i.e., the discrete problem is not solvable.
The following counterexample reveals a paradoxical situation. We choose sub-
spaces Vn ⊂ V and Wn ⊂ W such that
dim Wn > dim Vn .
Seemingly, a high-dimensional subspace Wn should improve the approximation of
the component w, but instead, it spoils the whole system. For a proof consider the
block
Bn
.
0
In the best case, this part has rank dim Vn (number of rows of Bn ). Since the matrix
Cn is of size Nn × Nn with Nn := dim Wn + dim Vn , its rank is bounded by
2 dim Vn < Nn . Therefore, at least dim Wn − dim Vn > 0 vanishing eigenvalues
must exist.
An equivalent statement for the stability
kCn−1 k(Vn ×Wn )←(Vn∗ ×Wn∗ ) ≤ Cstab
is the following inf-sup condition, which is also called the Babuška–Brezzi

condition:
inf sup |a(v0 , x0 )| ≥ α > 0, (7.27a)

v0 ∈V0,n with kv0 kV =1 x0 ∈V0,n with kx0 kV =1
sup |a(x0 , v0 )| > 0 for all 0 6= v0 ∈ V0,n ,

x0 ∈V0,n with kx0 kV =1
(7.27b)
inf sup |b(w, x)| ≥ β > 0, (7.27c)
w∈Wn with kwkW =1 x∈Vn with kvkV =1
where
V0,n := ker(Bn∗ ) = {v ∈ V : b(y, v) = 0 for all y ∈ Wn } ⊂ Vn .
The uniform boundedness of Cn−1 is expressed by the fact that the positive numbers
α, β in (7.27a,c) are independent of n (cf. §12.3.2 in [8, 10]).
The solvability of the undiscretised saddle-point problem is characterised by the
same conditions (7.27a-c), but with V0,n and Vn replaced by V0 = ker(B ∗ ) and V .
In the case of Stokes’ problem (7.26), a(·, ·) is coercive on the whole space V , but
for elasticity problems, one must exploit the fact that coercivity is only needed for
the smaller subspace V0 ⊂ V (cf. Braess [2, §III.4 and §VI]).
7.5.8 Further Remarks
7.5.8.1 Perturbations of An and fh
The determination of An and fh requires the evaluation of integrals. This can be

done exactly for An in cases of constant coefficients as in (7.1). However, for general
f and general coefficients as in (7.2), the integral must be approximated by some
quadrature thereby producing additional perturbations δAn and δfn . Whatever the
origin of these perturbations, the natural requirements are
kδAn kXn∗ ←Xn → 0, kδfn kXn∗ → 0 as N0 3 n → ∞. (7.28)
The following statement has a similar flavour as in Theorems 3.46 and 3.47.
Theorem 7.24. Suppose that the Galerkin approximation, characterised by

{An , fn }n∈N0 , is stable with kA−1
n kXn ←Xn∗ ≤ Cstab . Then, under assumption
(7.28), also the perturbed discretisation {An + δAn , fn + δfn }n∈N0 is stable.
The resulting solution ũn = un + δun (un : unperturbed discrete solution) satis-
fies the asymptotic inequality
kδun kXn (7.29)

h i
≤ kδAn kXn∗ ←Xn Cstab kfn kXn∗ + kδfn kXn∗ Cstab + O(kδAn kXn∗ ←Xn ) .
Proof. Write An + δAn as An (I + A−1

n δAn ) and note that
−1 −1
An δAn
Xn ←Xn
≤ An Xn ←X ∗ kδAn kXn∗ ←Xn ≤ Cstab kδAn kXn∗ ←Xn → 0.
n
Define the subset N00 := {n ∈ N0 : kδAn kX ∗ ←Xn ≤ 2C1stab }. Then, for n ∈ N00 , we
n
have A−1 −1
n δAn Xn ←Xn ≤ 1/2 and k(I + An δAn )
−1
kXn ←Xn ≤ 2 (cf. Lemma
5.8), so that
k(An + δAn )−1 kXn ←Xn∗ = k(I + A−1

n δAn )
−1 −1 0
An kXn ←Xn∗ ≤ 2Cstab =: Cstab .
This proves the stability of An + δAn .

(An + δAn )−1 − A−1 −1
n = [(I + An δAn )
−1
− I]A−1
n can be estimated by
k(An +δAn )−1 −A−1

n kXn ←Xn
∗ ≤ Cstab kδAn k ∗
Xn ←Xn(Cstab +O(kδAn kXn
∗ ←X )),
n
while k(An +δAn )−1 kXn ←Xn∗ ≤ (Cstab +kδAn kXn∗ ←Xn ). Therefore, we conclude
that
δun = ũn − un = (An + δAn )−1 (fn + δfn ) − A−1n fn
= [(An + δAn )−1 − A−1
n ]fn + (An + δAn )
−1
δfn
can be estimated as in (7.29). t

u
As an illustration, we discuss the case of quadrature mentioned above. The non-
vanishing entries aij of the finite element matrix An are of the size O(h−2+d/2 )
when we assume a quasi-uniform discretisation with maximal grid size h in the d-
dimensional domain Ω ⊂ Rd . A general quadrature method of the second order
leads to an absolute error of size O(hd/2 ) and to the spectral norm kδAn k2 =
O(hd/2 ). Since k·kX ∗ ←Xn ≤ k·k2 , the first inequality in (7.28) is satisfied.
n
R Another error treatment follows the idea of backward analysis. Let a(u, v) =
Ω
h∇v, c(x)∇ui dx be the bilinear form with a (piecewise) R smooth coefficient
c. For piecewise
R linear finite elements bi , the integrals ∆
h∇b R j i dx are
i , c(x)∇b
const · ∆ c(x)dx (∆: triangle). Let c∆ be the quadrature result of ∆ c(x)dx
and define a Rnew boundary value problem with piecewise constant coefficients
c̃(x) := c∆ / ∆ dx for x ∈ ∆. Then the finite element matrix An + δAn is the
exact matrix for the new bilinear form ã(·, ·).
Another formulation of the total error including (7.29) is given by the first Strang
lemma (cf. Braess [2, Part III, §1]): for coercive bilinear forms a(·, ·) and an (·, ·) and
right-hand sides f and fn let u and un be the respective solutions of
a(u, v) = f for all v ∈ X, an (un , vn ) = fn for all vn ∈ Xn ⊂ X.
Then
|a(un , wn ) − an (un , wn )|
ku − un kX ≤ C inf ku − vn kX + sup
vn ∈Xn wn ∈Xn kwn kX

|(f − fn )(wn )|
+ sup .
wn ∈Xn kwn kX
7.5.8.2 Nonconforming Methods
The standard conforming finite element methods are characterised by the inclusion
Xn ⊂ X, which allows us to evaluate a(un , vn ) for un , vn ∈ Xn . This is also called
internal approximation. An external approximation uses a sequence of spaces Xn
containing also elements outside of X; i.e.,
Xn 6⊂ X.
The variational formulation of the nonconforming Galerkin method is of the form
find un ∈ Xn with an (un , vn ) = fn (vn ) for all vn ∈ Xn , (7.30)
where the subscript of an indicates that the bilinear form depends on n. The
associated operator An : Xn → Xn0 is defined by
(An un )(vn ) = an (un , vn ) for all un , vn ∈ Xn .
Therefore, the variational formulation (7.30) is equivalent to
An un = fn . (7.31)
Above we defined a vector space Xn , but we have not yet defined a norm. In
general, this norm depends on n:
Xn normed by ||| · |||n .
As mentioned above, the bilinear form an (·, ·) : Xn × Xn → R defines the operator

An : Xn → Xn0 via an (xn , ·) = hAn xn , ·iX 0 ×Xn . Stability is characterised by
n
the existence of the inverse A−1 0
n —at least for an infinite subset N ⊂ N—and the
uniform estimate
Cstab := sup A−1

n
0
X ←Xn
< ∞.
n∈N0 n
Since Xn is finite-dimensional, the latter statement is equivalent to the inf-sup

condition
|an (vn , wn )| 1
inf sup ≥ > 0. (7.32)
06=vn ∈Vn 06=wn ∈Vn |||vn |||n |||wn |||n Cstab
Before we study the connection between an (·, ·) and a(·, ·), we consider the
functionals fn and f. The usual setting a(u, v) = f (v) (v ∈ X) describes the
problem Au = f with a functional f ∈ X 0 . The obvious discrete formulation
an (un , vn ) = f (vn ) for all vn ∈ Xn (7.33)
is not quite sound, since f may be undefined for vn ∈ Xn \X (note that Xn \X 6= ∅

in the nonconforming case). There are two remedies: (a) Extend f continuously to
fn such that f (v) = fn (v) for v ∈ X. (b) Usually, we have a Gelfand triple
X ⊂ U = U 0 ⊂ X0
(e.g., with U = L2 (Ω), cf. §6.3.3 in [8, 10]). If f ∈ U 0 and Xn ⊂ U, the functional
f is well defined on Xn and the variational formulation (7.33) makes sense (i.e., we
may choose fn := f ).
The restriction of f to a smaller subset U 0 $ X 0 implies that also the solution u
belongs to a subspace strictly smaller than X. We denote this space by X∗ . In the
case of f ∈ U 0 , this is
X∗ := {A−1 f : f ∈ U 0 } $ X.
There is always the problem of how to compare the discrete solution un ∈ Xn

and the exact solution u ∈ X if the spaces Xn and X are different. One possibility
is to restrict u ∈ X to Xn and to compare both in Xn (cf. (7.6a)). Another approach
is to extend un ∈ Xn into X (cf. (7.6b)). The latter case is trivial in the conforming
Galerkin case, since we have the natural inclusion Xn ⊂ X. In the nonconforming
case, neither Xn nor X contain both u and un . This problem leads us to the next
construction: we define a common superspace
X∗n ⊃ X∗ and X∗n ⊃ Xn with norm ||| · |||∗n .
Then the difference u − un belongs to X∗n and convergence can be formulated by

|||u − un |||∗n → 0. We require that
||| · |||n ≤ C̄∗ ||| · |||∗n on Xn (7.34)
with a constant C̄∗ . Without loss of generality, we may scale ||| · |||∗n such that
C̄∗ = 1.
The next requirement is that the bilinear form an (·, ·) can be continuously ex-
tended from Xn × Xn onto X∗n × Xn such that
|an (v, w)| ≤ Ca |||v|||∗n |||w|||n for all v ∈ X∗n , w ∈ Xn . (7.35)
Remark 7.25. Assume that (V, k·kV ) and (W, k·kW ) are Banach spaces with an
intersection V ∩ W possibly larger than the zero space {0}. Then a canoni-
cal norm of the smallest common superspace U := span(V, W ) is defined by
kukU := inf{kvkV + kwkW : u = v + w, v ∈ V, w ∈ W }.
Corollary 7.26. Assume that k·kV and k·kW are equivalent norms on V ∩W. Prove
that kvkV ≤ kvkU ≤ C kvkV for all v ∈ V and kwkW ≤ kwkU ≤ C kwkW for
all w ∈ W , where k·kU is the norm from Remark 7.25.
Consistency is expressed by
an (u, vn ) = f (vn ) for all vn ∈ Xn , (7.36)
where u is the solution of a(u, v) = f (v) for v ∈ X (here, we assume that f belongs
to Xn0 ). Then the error estimate (Strang’s second lemma, cf. [18, §1.3])
∗
|||u − un |||n ≤ C̄∗ + Ca Cstab inf |||u − wn |||∗n + Ca Cstab |||fn − f |||n
wn ∈Xn
∗
holds, where ||| · |||n is the norm dual to ||| · |||n .
Proof. Let un be the discrete solution, while wh ∈ Xn is arbitrary. From (7.32) we

infer that
|||un − wn |||n ≤ Cstab sup |an (un − wn , vn )| .

vn ∈Xn with |||vn |||n =1
Split an (un − wn , vn ) into an (un − u, vn ) + an (u − wn , vn ). The first term can be

reformulated in terms of (7.30) and (7.36): an (un − u, vn ) = (fn − f )(vn ); hence
∗
|an (un − u, vn )| ≤ |||fn − f |||n
because |||vn |||n = 1. The second term is estimated by
|an (u − wn , vn )| ≤ Ca |||u − wn |||∗n
(cf. (7.35)). Together, we obtain

∗
|||un − wn |||n ≤ Cstab |||fn − f |||n + Ca |||u − wn |||∗n .
The triangle inequality yields
|||u − wn |||n ≤ |||u − un |||n + |||un − wn |||n

≤ C̄∗ |||u − un |||∗n + |||un − wn |||n
∗
≤ C̄∗ + Ca Cstab |||u − wn |||∗n + Ca Cstab |||fn − f |||n .
Taking the infimum over all wn , we obtain the error estimate. t

u
The consistency condition (7.36) is not always satisfied in practical examples

as, e.g., Wilson’s element (for a definition and analysis compare, e.g., Shi [22]).
As a remedy, one has tried to formulate criteria that are easier to check, the so-
called patch tests. Stummel [24, 25] analysed the patch tests and proved that they
are neither necessary nor sufficient for stability.
7.5.8.3 Discontinuous Galerkin Method
A prominent example of a nonconforming finite element is the discontinuous

Galerkin method (DG), which started in the field of hyperbolic problems and for
elliptic problems with dominant convection. Now it has become a popular method
also for standard elliptic problems (see [18], [3], [14], [21, p. 255]).
The spaces X∗ and X∗n ⊃ Xn ⊃ X∗ are as follows. X∗ = H01 (Ω) is the
standard choice. Given any triangulation Tn of Ω, the broken H 1 space is defined by
sX
2
:= u|T ∈ H 1 (T ) for all τ ∈ Tn with kukXn :=

X∗n kun kH 1 (τ ) .
τ ∈Tn
The remarkable fact is that no conditions are required concerning the connection
of u|T 0 and u|T 00 for neighboured elements T 0 , T 00 ∈ Tn ; i.e., in general, func-
tions from X∗n are discontinuous across the internal boundaries of Tn . Obviously,
X∗n ⊃ X∗ = H01 (Ω) holds.
The DG finite element space is the subspace Xn of X∗n , where all u|T are,
e.g., piecewise linearly affine. The advantage of the discontinuous Galerkin method
becomes obvious when we want to choose different polynomial degrees for different
elements. Hence, the computational overhead for an hp-method is much lower.
Starting from a strong solution of −∆u = f in Ω with u = 0 on Γ = ∂Ω,
multiplication with a test function v ∈ Xn and partial integration in each T ∈ Tn
yield the variational form (7.36) with
Z X Z X Z ∂u
(−∆u) vn dx = h∇un , ∇vn i dx − [vn ] ds
Ω τ E ∂nE
τ ∈Tn E∈En
R
and fn (v) = Ω f vdx. Here, En is the set of all edges of τ ∈ Tn . Each edge E ∈ En
is associated with a normal nE (the sign of the direction does not matter). The curly
bracket {. . .} is the average of the expression in the two elements containing E,
while [. . .] is the difference (its sign depends on the direction of nE ). The right-
hand side in the latter expression is no longer symmetric. Since [u] = 0 for u ∈ X∗ ,
we may add further terms without changing the consistency property:
X Z
an (un , vn ) := h∇un , ∇vn i dx (7.37)
τ ∈Tn τ
X Z ∂un
∂vn

− [vn ] − [un ] ds
∂nE ∂nE
E∈En E
X Z
+η h−1
E [un ] [vn ] ds.
E∈En E
The expression in the second line is antisymmetric. Therefore it vanishes for

un = vn , and
X X Z
2 −1 2
an (un , un ) = k∇un kL2 (τ ) + η hE [un ] ds
τ ∈Tn E∈En E
proves coercivity of an .
Other variants of the discontinuous Galerkin
n methods use a symmetric bilinear
P R n ∂un o ∂vn
o
form with the term E∈En E ∂nE [vn ] + ∂nE [un ] ds in the second line of
(7.37) (cf. [18, pp. 124ff]).
7.5.9 Consistency Versus Stability
7.5.9.1 Convection-Diffusion Equation
For ordinary differential equations, we mentioned stiff differential equations (cf.

§5.5.7.1). In that case, standard stable schemes work in the limit h → 0, but their
results are useless unless h ≤ h0 , where h0 may be a rather small number. A similar
phenomenon can happen for the discretisation of elliptic problems.
The differential equation
−∆u + hc, grad ui = f in Ω ⊂ R2 , u = 0 on Γ = ∂Ω, (7.38)
is called singularly perturbed if kck 1. Because of the dominant convection

term hc, grad ui, the problem can be considered as a perturbation of the hyperbolic
problem hc, grad ui = f. Since the hyperbolic problem cannot be combined with
boundary data u = 0 on all of Γ, the solution of (7.38) develops a boundary layer at
the outflow part of the boundary.
A stable discretisation can be constructed as follows. The part −∆u is discretised
by the five-point formula (7.3). The treatment of hc, grad ui depends on the signs of
the components c1 , c2 in c = (c1 , c2 ). Assume c1 > 0. Then c1 ∂u/∂x1 is replaced
by the backward difference
c1
(uν,µ − uν−1,µ ) .
h
If c1 < 0, the forward difference
c1
(uν+1,µ − uν,µ )
h
is used. The second term c2 ∂u/∂x2 in hc, grad ui is treated analogously. As a
result, the discretisation matrix is an M-matrix (cf. Definition 7.2). This property
helps to prove the stability of the discretisation.However, the consistency order is
one because of the one-sided differences.
Second-order consistency can be obtained by replacing ci ∂u/∂xi with
symmetric differences. The conflict with stability can be checked by the violation
of the M-matrix sign conditions. If c1 > 0, the coefficient of uν+1,µ becomes
1 c1
− +
h2 2h
and fails to be non-positive unless h ≤ 2/c1 . Because of the assumption kck 1,
the requirement h ≤ 2/c1 limits the use of the second-order scheme to rather small
step sizes.
This conflict between consistency and stability also occurs for finite element
discretisation (the corresponding modifications are, e.g., streamline diffusion
methods, cf. John–Maubach–Tobiska [12]).
7.5.9.2 Defect Correction Methods
A possible remedy is the use of the defect correction scheme. It is based on two
different discretisations:
Lh uh = fh (7.39a)
is a stable scheme, possibly of a low order of consistency, whereas the second
scheme
L0h u0h = fh0 (7.39b)
has an order of higher consistency. Scheme (7.39b) is not required to be stable. It
may even be singular (i.e., a solution u0h may not exist).
The method consists of two steps. First, the basic scheme (7.39a) is applied to
(0)
obtain the starting value uh :
(0)
uh := L−1
h fh . (7.40a)
(0)
Next the defect of uh with respect to the second scheme is computed:
(0)
dh := L0h uh − fh0 . (7.40b)
The correction using dh is performed in the second and final step:

(1) (0)
uh := uh − L−1
h dh . (7.40c)
(1) (0)
We shall show that uh has better consistency properties than uh . Note that
(1)
uh is neither a solution of (7.39a) nor of (7.39b). If (7.39a) has consistency order
(1)
one, the defect correction by (7.40c) can increase the consistency order of uh only
by one. Therefore, the steps (7.40b,c) must be repeated a few times, if the consis-
tency order of (7.39b) is larger than two. However, it is essential that the number of
iterations be a small finite number. Otherwise, the possible instability of (7.39b)
would destroy the result.
For the proof, we assume the following properties:
u∗ : exact solution; u∗h := RhX u∗ ,

kLh u∗h − fh kH −1 ≤ Ch, kL0h u∗h − fh0 kH −1 ≤ Ch2 , (7.41)
h h
−1 −1
−1 ≤ C,
L 1 L 2 2 ≤ C,
h H ←H
h h
h H ←L h h
kLh − L0h kH −1 ←H 2 ≤ Ch.

h h
Here L2h , Hh1 , Hh2 are the discrete analogues of L2h (Ω), Hh1 (Ω), Hh2 (Ω) (with
derivatives replaced by divided differences). Hh−1 is the dual space of Hh1 .
The last estimate of Lh − L0h in (7.41) can be obtained as follows. Split Lh − L0h
into
References 165
L0h I − RhX PhX − Rh0Y L − L0h RhX PhX + Rh0Y − RhY LPhX

+ RhY L − Lh RhX PhX − Lh I − RhX PhX

and assume
0Y 0
I − RhX PhX 1 X

2 , Rh L − Lh Rh −1 ≤ Ch,

Hh ←Hh Hh ←H 2
0Y
Rh − RhY −1 2 , RhY L − Lh RhX −1

Hh ←L H ←H 2
≤ Ch,
0
X h
kLh kH −1 ←H 1 , kLh kH −1 ←H 1 , kLkL2 ←H 2 , Ph H 2 ←H 2 ≤ C.
h h h h h
(1)
Lemma 7.27. Under the suppositions (7.41), the result uh of the defect correction
(1)
method satisfies the estimate kuh − u∗h kHh1 ≤ Ch2 .
(1)
Proof. We may rewrite uh − u∗h as
h i
(1) (0) (0)
uh − u∗h = uh − u∗h − L−1h L0h uh − fh0
h (0) i
= I − L−1 0 ∗
+ L−1 0 ∗ 0

h L h u h − u h h [Lh uh − fh ]
h i
(0)
= L−1 0 ∗ −1 0 ∗
h [Lh − Lh ] uh − uh + Lh [Lh uh − fh ] .
0
Consistency kL0h u∗h − fh0 kH −1 ≤ Ch2 and stability L−1

h H ←H −1
1 ≤ C imply
h h h
that the second term satisfies kL−1 0 ∗ 0
h [Lh uh − fh ] kHh
2
1 ≤ Ch .
Similarly,
(0)
kuh − u∗h kHh2 = L−1 ∗
−1
2 2 kLh u∗h − fh k 2 ≤ Ch

h fh − uh H 2 ≤ Lh

H ←L
h h
L h h
holds. The inequalities L−1

0
−1 ≤ C and kLh − Lh k 2 −1 ≤ Ch yield
1
h Hh ←Hh Hh ←Hh
the estimate
(0)
kL−1 0 ∗
h [Lh − Lh ] [uh − uh ]kHh
1 ≤ Ch
2
for the first term. Together, the assertion of the lemma follows. t
u
The defect correction method was introduced by Stetter [23]. A further descrip-
tion can be found in Hackbusch [9, §14.2]. The particular case of diffusion with
dominant convection is analysed by Hemker [11].
The application of the defect correction method is not restricted to elliptic
problems. An application to a hyperbolic initial-value problem can be found in [5].
References
1. Aubin, J.P.: Behaviour of the error of the approximate solution of boundary value problems
for linear operators by Galerkin’s and finite difference methods. Ann. Scuola Norm. Sup. Pisa
21, 599–637 (1967)
2. Braess, D.: Finite Elements: Theory, Fast Solvers, and Applications in Solid Mechanics,
3rd ed. Cambridge University Press, Cambridge (2007)
3. Cockburn, B., Karniadakis, G.E., Shu, C.W. (eds.): Discontinuous Galerkin Methods.
Theory, Computation and Applications, Lect. Notes Comput. Sci. Eng., Vol. 11. Springer,
Berlin (2000).
4. Fréchet, M.: Sur les ensembles de fonctions et les opérations linéaires. C. R. Acad. Sci. Paris
144, 1414–1416 (1907)
5. Hackbusch, W.: Bemerkungen zur iterativen Defektkorrektur und zu ihrer Kombination mit
Mehrgitterverfahren. Rev. Roumaine Math. Pures Appl. 26, 1319–1329 (1981)
6. Hackbusch, W.: On the regularity of difference schemes. Ark. Mat. 19, 71–95 (1981)
7. Hackbusch, W.: Iterative Solution of Large Sparse Systems of Equations. Springer, New York
(1994)
8. Hackbusch, W.: Elliptic differential equations. Theory and numerical treatment, Springer
Series in Computational Mathematics, Vol. 18, 2nd ed. Springer, Berlin (2003)
9. Hackbusch, W.: Multi-grid methods and applications, Springer Series in Computational
Mathematics, Vol. 4. Springer, Berlin (2003)
10. Hackbusch, W.: Theorie und Numerik elliptischer Differentialgleichungen, 3rd ed. (2005).
www.mis.mpg.de/scicomp/Fulltext/EllDgl.ps
11. Hemker, P.W.: Mixed defect correction iteration for the solution of a singular perturbation
problem. Comput. Suppl. 5, 123–145 (1984)
12. John, V., Maubach, J.M., Tobiska, L.: Nonconforming streamline-diffusion-finite-element-
methods for convection diffusion problems. Numer. Math. 78, 165–188 (1997)
13. Jovanovič, B., Süli, E.: Analysis of finite difference schemes for linear partial differential
equations with generalized solutions, Springer Series in Computational Mathematics, Vol. 45.
Springer, London (2013)
14. Kanschat, G.: Discontinuous Galerkin methods for viscous incompressible flows. Advances
in Numerical Mathematics. Deutscher Universitätsverlag, Wiesbaden (2007)
15. Nečas, J.: Sur la coercivité des formes sesquilinéares elliptiques. Rev. Roumaine Math. Pures
Appl. 9, 47–69 (1964)
16. Nitsche, J.: Ein Kriterium für die Quasi-Optimalität des Ritzschen Verfahrens. Numer. Math.
11, 346–348 (1968)
17. Ostrowski, A.M.: Über die Determinanten mit überwiegender Hauptdiagonale. Comment.
Math. Helv. 10, 69–96 (1937)
18. Di Pietro, D.A., Ern, A.: Mathematical aspects of discontinuous Galerkin methods. Springer,
Berlin (2011)
19. Riesz, F.: Sur une espèce de géométrie analytique des systèmes de fonctions sommables.
C. R. Acad. Sci. Paris 144, 1409–1411 (1907)
20. Riesz, F., Sz.-Nagy, B.: Functional Analysis. Dover Publ. Inc, New York (1990)
21. Roos, H.G., Stynes, M., Tobiska, L.: Numerical methods for singularly perturbed differential
equations: Convection-diffusion and flow problems, Springer Series in Computational Math-
ematics, Vol. 24. Springer, Berlin (1996)
22. Shi, Z.C.: A convergence condition for the quadrilateral Wilson element. Numer. Math. 44,
349–361 (1984)
23. Stetter, H.J.: The defect correction principle and discretization methods. Numer. Math. 29,
425–443 (1978)
24. Stummel, F.: The generalized patch test. SIAM J. Numer. Anal. 16, 449–471 (1979)
25. Stummel, F.: The limitations of the patch test. Int. J. Num. Meth. Engng. 15, 177–188 (1980)
Chapter 8
Stability for Discretisations of Integral
Equations
Since the paper of Nyström [7], integral equations are used to solve certain boundary
value problems. Concerning the integral equation method and its discretisation, we
refer to Sauter–Schwab [8] and Hsiao–Wendland [5]. The following considerations
of stability hold as long as the integral kernel is weakly singular.
8.1 Integral Equations and Their Discretisations
8.1.1 Integral Equation, Banach Space
The Fredholm integral equation of the second kind is described by
λf = g + Kf, (8.1)
where λ ∈ C\{0} and the function g are given together with an integral operator K
defined by Z
(Kf ) (x) = k(x, y)f (y)dy for all x ∈ D. (8.2)
D
If λ = 0, Eq. (8.1) is called Fredholm’s integral equation of the first kind (cf. Fred-
holm [3]). The integration domain D is an (often compact) subset of Rn . Other
interesting applications lead to surface integrals; i.e., D = ∂Ω of some Ω ⊂ Rn .
The function g belongs to a Banach space X with norm k·k. The solution f of
(8.1) is also sought in X: g, f ∈ X. The Banach space X must be chosen such that
K ∈ L(X, X)
holds; i.e., K is a bounded mapping from X into itself. An equivalent notation of

problem (8.1) is1
(λI − K) f = g.
1
Equations (λA − K) f = g with a not necessarily bounded operator A are considered in [6].

168 8 Stability for Discretisations of Integral Equations
−1
If λI − K : X → X is bijective, also the inverse (λI − K) belongs to L(X, X)
(this is a consequence of the open mapping theorem; cf. [10, §II.5]) and a formal
description of the solution is
−1
f = (λI − K) g.
The operator norm k·kX←X is also denoted by k·k.

2
The standard choices for X are
qR X = C(D) or X = L (D) with the norms
kf k = supx∈D |f (x)| or kf k = D
|f (x)|2 dx, respectively.
In many cases, the operator K is not only bounded, but also compact (cf. §4.9).
Lemma 8.1. (a) If k ∈ L2 (D × D), then K : X → X is compact for X = L2 (D).

(b) If D is compact and k ∈ C(D ×D), then K : X → X is compact for X = C(D).
Proof. The proof will be postponed to §8.1.2.1. t

u
A more general criterion in the case of X = C(D) is the following.
Theorem 8.2. Suppose that D is compact and that k satisfies

Z Z
|k(x, y)| dy < ∞, lim |k(ξ, y) − k(x, y)| dy = 0 for all x ∈ D. (8.3)
D ξ→x D
Then K is compact for X = C(D).
Proof. (i) Precompactness of M := {Kf : f ∈ C(D), kf k ≤ 1} follows by

Arzelà–Ascoli’s theorem (cf. Theorem 3.51). Part (iii) will show uniform bounded-
ness of M, while Part (iv) will yield equicontinuity.
(ii) We introduce the following auxiliary functions:
Z Z
ϕ(x) := |k(x, y)| dy, Φ(ξ, x) := |k(ξ, y) − k(x, y)| dy.
D D
Claim: ϕ ∈ C(D) is Rbounded, and Φ ∈ C(D × D) is Runiformly continuous. From

|ϕ(ξ) − ϕ(x)| = D (|k(ξ, y)| − |k(x, y)|) dy ≤ D |k(ξ, y) − k(x, y)| dy =
Φ(ξ, x) and (8.3), it follows that ϕ is continuous. A continuous function on a com-
pact set is bounded. An analogous estimate shows |Φ(ξ, x) − Φ(ξ, x0 )| ≤ Φ(x, x0 )
and |Φ(ξ, x) − Φ(ξ 0 , x)| ≤ Φ(ξ, ξ 0 ). Again by (8.3), Φ is continuous on the compact
set D × D; hence, uniformly continuous.
(iii) Each g = Kf ∈ M is bounded by kgk ≤ kϕk, since
Z Z

|g(x)| = k(x, y)f (y)dy ≤ |k(x, y)| |f (y)| dy ≤ ϕ(x) kf k ≤ ϕ(x) ≤ kϕk .
D D
Hence M is uniformly bounded.

(iv) Fix ε > 0. As Φ is uniformly continuous (cf. (ii)), there is some δ > 0 such
that Φ(ξ, x) = |Φ(ξ, x) − Φ(x, x)| ≤ ε for all ξ, x ∈ D with |ξ − x| ≤ δ. For any
g = Kf ∈ M we have
8.1 Integral Equations and Their Discretisations 169
Z Z ≤1
z }| {
|g(ξ) − g(x)| = [k(ξ, y) − k(x, y)] f (y)dy ≤
|k(ξ, y) − k(x, y)| |f (y)|dy
D D
≤ Φ(ξ, x) ≤ ε for |ξ − x| ≤ δ,
proving that all functions g ∈ M are equicontinuous. t

u
We mention that the conditions from (8.3) are also necessary, provided that
k(x, ·) ∈ L1 (D) for all x ∈ D.
8.1.2 Discretisations
There are various discretisation techniques leading to a sequence of ‘discrete’ inte-

gral operators Kn ∈ L(X, X), n ∈ N, with the property that the range of Kn is
n-dimensional. The corresponding ‘discrete’ integral equation is
λfn = g + Kn fn (8.4a)
or λfn = gn + Kn fn , (8.4b)
where in the latter case
gn → g (8.4c)
is assumed.
Because the right-hand sides in (8.4a,b) are contained in an at most (n + 1)-
dimensional subspace, λ 6= 0 implies that fn belongs to this subspace. This leads to
a system of finitely many linear equations.
8.1.2.1 Kernel Approximation
The kernel approximation approximates the kernel function k(·, ·) by a so-called

degenerate kernel
n
X
kn (x, y) = aj (x)bj (y) for x, y ∈ D, (8.5)
j=1
leading to
Z n
X Z
Kn f = kn (·, y)f (y)dy = aj (·) bj (y)f (y)dy.
D j=1 D
Exercise 8.3. Let X = C(D) or X = L2 (D). Suppose that kkn − kkC(D×D) → 0

or kkn − kkL2 (D×D) → 0, respectively. Prove the corresponding operator norm
convergence kK − Kn k → 0.
The suppositions of Exercise 8.3 imply even compactness of K.
Remark 8.4. Suppose that Kn has a finite-dimensional range. Then kK−Kn k→0
implies that K, Kn : X → X are compact operators.
Proof. We recall two properties of compact operators: (i) finite dimension range
implies compactness, (ii) a limit of compact operators is compact. From (i) we
infer that Kn is compact, while (ii) proves that K = lim Kn is compact. t
u
Proof of Lemma 8.1. It is sufficient to verify the supposition of Exercise 8.3.
For Part (b) involving X = C(D), use Weierstrass’ approximation theorem. Since
D × D is compact, there is a sequence of polynomials Pn (x, y) of degree ≤ n − 1
such that kPn − kkC(D×D) → 0. Since kn := Pn has a representation (8.5) with
aj (x) = xj and a polynomial bj (y) of degree ≤ n − 1, Exercise 8.3 can be applied.
Proof. In the Hilbert case of X = L2 (D), K possesses an infinite singular value
decomposition2
X∞
K= σj aj b∗j with orthonormal {aj }, {bj } ⊂ X,
j=1
where the singular values satisfy σ1 ≥ σ2 ≥ . . . ≥ 0 , σ1 = kKk, and

X∞
2
σj2 = kkkL2 (D×D) < ∞
j=1
Pn
(cf. [9], [4, §4.4.3]). The operator Kn = j=1 σj aj b∗j has the kernel
n
X
kn (x, y) = σj aj (x)bj (y)
j=1
2 P∞
and satisfies kK − Kn k = j=n+1 σj2 → 0. t
u
8.1.2.2 Projection Method
The projection method is characterised by a sequence of subspaces Xn ⊂ X and

projections Πn ∈ L(X, X) with
Xn = range(Πn ) with dim(Xn ) = n. (8.6)
Either we construct a sequence of projections {Πn } and define Xn by (8.6),

or we build a sequence of subspaces {Xn } and define Πn as a projection onto Xn .
Setting
Kn := Πn K and gn := Πn g,
we obtain the discrete problem by (8.4b): λfn = gn + Kn fn . Note that the solution
belongs to Xn , provided that λ 6= 0 and the solution exists.
b0j denotes the corresponding functional b0j (ϕ) := D bj (y)ϕ(y)dy; formally, b0j := Jbj is
2
R
obtained by the Riesz isomorphism explained in Remark 7.18.
8.1 Integral Equations and Their Discretisations 171
There are two standard choices of Πn .

(1) Case X = C(D). Define an interpolation at the points {ξk,n : 1 ≤ k ≤ n} ⊂ D
by
n
X
Πn ϕ := ϕ(ξk,n )Lk,n
k=1
with Lagrange functions Lk,n ∈ X (cf. page 48). Since fn ∈ Xn , it suffices to

know the function values fn,k := fn (ξk,n ). The generation of the system of linear
equations requires computing the integrals
Z
(KLk,n ) (ξk,n ) = k(ξk,n , y)Lk,n (y)dy.
D
(2) Case X = L2 (D). The Hilbert space structure of X allows us to define Πn as

the orthogonal projection onto Xn , where Xn = span{ϕk,n : 1 ≤ k ≤ n} is some
Galerkin subspace of dimension n. The generation of the linear system requires
computing
Z Z
ϕk,n (x)k(x, y)ϕk,n (y)dxdy.
D D
8.1.2.3 Nyström Method

R
The integral D
. . . dy suggests applying a quadrature method
Z n
X
ϕ(y)dy ≈ Qn (ϕ) = aj,n ϕ(ξj,n ).
D j=1
This leads to the Nyström discretisation

n
X
λfn (x) = g(x) + ak,n k(x, ξk,n )fn (ξk,n ) for all x ∈ D (8.7)
k=1
(cf. Nyström [7]). Restricting the latter equation to x ∈ {ξk,n : 1 ≤ k ≤ n}, we

obtain a linear system for fn (ξk,n ), 1 ≤ k ≤ n. If this system has a unique solution,
the values fn (ξk,n ) inserted into the right-hand side of (8.7) determine fn (x) for all
x ∈ D, provided that λ 6= 0. The extension of fn (ξk,n ) to fn ∈ C(D) is also called
Nyström interpolation. Because of this interpolation, the Nyström discretisation can
again be written as λfn = g + Kn fn with
n
X
(Kn ϕ) (x) = ak,n k(x, ξk,n )ϕ(ξk,n ). (8.8)
k=1
Since we need point evaluations, the Banach space X = C(D) must be used.
8.2 Stability Theory
8.2.1 Consistency
Let K, Kn ∈ L(X, X) be operators acting in a Banach space X.
Definition 8.5 (consistency). {Kn } is called consistent with respect to K if 3

Kn ϕ → Kϕ for all ϕ ∈ X. (8.9)
{Kn } is called consistent if some K ∈ L(X, X) exists satisfying (8.9).
Exercise 8.6. (a) Prove that {Kn } is consistent with respect to K if and only if
(i) Kn ϕ → Kϕ for all ϕ ∈ M ⊂ X and some M dense in X, and
(ii) supn kKn k < ∞.
(b) Assuming that Kn ϕ is a Cauchy sequence and supn kKn k < ∞, define K by
Kϕ := lim Kn ϕ and prove that K ∈ L(X, X); i.e., {Kn } is consistent with respect
to K.
(c) Operator norm convergence kK − Kn k → 0 is sufficient for consistency.
8.2.2 Stability
Stability refers to the value λ 6= 0 from problem (8.1). This λ is assumed to be fixed.
Otherwise, one has to use the term ‘stable with respect to λ’.
Definition 8.7 (stability). {Kn } is called stable if there exist some n0 ∈ N and
Cstab ∈ R such that 4
−1
k (λI − Kn ) k ≤ Cstab for all n ≥ n0 .
−1
If (λI − K) ∈ L(X, X), the inverse exists also for perturbations of K.
−1
Remark 8.8. Suppose (λI − K) ∈ L(X, X) and
−1
kK − Kn k < 1/k (λI − K) k.
Then
−1
−1 k (λI − K) k
k (λI − Kn ) k≤ −1 .
1 − k (λI − K) k kK − Kn k
Proof. Apply Lemma 5.8 with T := λI − K and S := λI − Kn . t
u
3
In Definition 4.5, consistency involves Kn ϕ → Kϕ only for ϕ from a dense subset X0 ⊂ X.
Then the full statement (8.9) could be obtained from stability: supn kKn k < ∞. Here, stability
will be defined differently. This is the reason to define consistency as in (8.9).
4
k (λI − K)−1 k ≤ Cstab is the short notation for ‘λI − K ∈ L(X, X) is bijective and the
inverse satisfies k (λI − K)−1 k ≤ Cstab ’.
8.2 Stability Theory 173
A simple consequence is the following stability result.

−1
Remark 8.9. If (λI − K) ∈ L(X, X) and kK − Kn k → 0, then {Kn } is stable.
The roles of K and Kn can be interchanged.
−1
Exercise 8.10. If kK − Kn k < 1/k (λI − Kn ) k holds for all n ≥ n0 , then the
−1
inverse (λI − K) ∈ L(X, X) exists. The assumption is, in particular, valid, if
kK − Kn k → 0 and {Kn } is stable.
8.2.3 Convergence
Definition 8.11 (convergence). {Kn } is called convergent if there is some n0 ∈ N

such that
(i) λfn = g + Kn fn is solvable for all g ∈ X and n ≥ n0 , and
(ii) the limit limn fn exists in X.
The next remark shows that f := limn fn satisfies the continuous problem. The
−1
existence of (λI − K) is left open, but will follow later from Theorem 8.16.
Remark 8.12. If {Kn } is consistent and convergent, then f := limn fn satisfies

λf = g + Kf (cf. (8.1)) and the operator λI − K is surjective.
Proof. (i) By Exercise 8.6a, C := supn kKn k < ∞ holds. The solutions fn exist
for n ≥ n0 and define some f := limn fn . Consistency shows that
kKn fn − Kf k = kKn (fn − f ) − (K − Kn ) f k

≤ C kfn − f k + k(K − Kn ) f k → 0;
hence the limit process n → ∞ in λfn = g + Kn fn yields λf = g + Kf .

(ii) Since for all g ∈ X, (i) yields a solution of (λI − K) f = g, surjectivity
follows. t u
A particular result of the Riesz–Schauder theory (cf. Yosida [10, §X.5]) is the
following result.
Lemma 8.13. Suppose that λ 6= 0, K ∈ L(X, X) compact. Then λI−K is injective

if and only if λI − K is surjective.
Exercise 8.14. Let K be compact and {Kn } be consistent and convergent. Prove
−1
(λI − K) ∈ L(X, X).
Lemma 8.15. If {Kn } is stable and consistent, then the operator λI−K is injective.
Proof. Injectivity follows from k(λI − K) ϕk ≥ η kϕk for some η > 0 and all
ϕ ∈ X. For an indirect proof assume that there is a sequence ϕn ∈ X with
kϕn k = 1 and kψn k ≤ 1/n for ψn := (λI − K) ϕn .
By consistency Km ϕn → Kϕn for m → ∞ and fixed n. Hence there is some

mn ∈ N such that kKm ϕn − Kϕn k ≤ 1/n for m ≥ mn . We estimate the norm of
ζn,m := (λI − Km ) ϕn = ψn − (Km ϕn − Kϕn ) by 2/n for m ≥ mn . Stability
−1
and the representation ϕn = (λI − Km ) ζn,m yields 1 = kϕn k ≤ C · 2/n.
The contradiction follows for n > 2C. t
u
8.2.4 Equivalence
Theorem 8.16 (stability theorem). (a) Convergence implies stability.

−1
(b) Convergence and consistency imply stability and (λI − K) ∈ L(X, X).
Proof. Part (a) follows again by Banach–Steinhaus (Corollary 3.39).

In the case of Part (b), Remark 8.12 states that λI − K is surjective. Since, by
Part (a) stability holds, Lemma 8.15 proves that λI − K is also injective. Together,
−1
λI − K is bijective, which proves (λI − K) ∈ L(X, X). t u
Theorem 8.17 (convergence theorem). Suppose consistency, stability, and either

(i) λI − K surjective or (ii) K compact.
(a) Then {Kn } is convergent and
−1 −1
fn = (λI − Kn ) g → f = (λI − K) g for all g ∈ X.
−1 −1
(b) If gn → g, then fn = (λI − Kn ) gn → f = (λI − K) g.
Proof. Lemma 8.15 ensures injectivity of λI − K. Together with assumption (i),

−1
(λI − K) ∈ L(X, X) is proved. In the case of (ii), we conclude that injectivity
of λI − K implies surjectivity (cf. Lemma 8.13), if λ 6= 0 and K compact. This
−1
allows us to define f := (λI − K) g. Set
dn := λf − Kn f − gn ,
where either gn := g (case of (8.4a)) or gn → g. Consistency and gn → g show that

−1
dn → 0. Subtraction of fn := (λI −Kn ) gn implies that (λI −Kn ) (f − fn ) = dn
and
−1
f − fn = (λI − Kn ) dn
for n ≥ n0 . Stability yields the estimate kf − fn k ≤ C kdn k → 0; i.e., fn → f . t
u
Combining the previous theorems, we obtain the next one.

8.3 Projection Methods 175
Theorem 8.18 (equivalence theorem). Suppose consistency and one of the condi-
tions (i) or (ii) from Theorem 8.17. Then stability and convergence are equivalent.
Furthermore, gn → g implies
−1 −1
fn = (λI − Kn ) gn → f = (λI − K) g.
−1
Remark 8.19. The suppositions (λI − K) ∈ L(X, X) and kK − Kn k → 0
imply consistency, stability, and convergence.
Proof. Exercise 8.6c shows consistency. Remark 8.9 yields stability. Finally,
convergence is ensured by Theorem 8.17a. t
u
For the numerical solution of a linear problem Ax = b, the condition of A is

important:
kAkkA−1 k if A, A−1 ∈ L(X, X),

cond(A) = (8.10)
∞ otherwise.
Exercise 8.20. If {Kn } is consistent and convergent, then there is some n0 ∈ N

such that
sup cond(λI − Kn ) < ∞.
n≥n0
8.3 Projection Methods
We recall that a projection method is characterised by the sequences {Πn } and

{Xn = range(Πn )}. The approximation of X by {Xn } is described by the condi-
tion
dist(x, Xn ) := inf{kx − yk : y ∈ Xn } → 0 for all x ∈ X. (8.11)
Definition 8.21. A sequence {Πn } is called convergent if Πn x = x for all x ∈ X.
Exercise 8.22. Prove that a convergent sequence {Πn } generates subspaces Xn

satisfying condition (8.11).
The terms consistency and stability can also be applied to {Πn }.
Exercise 8.23. Define:

(i) {Πn } is consistent if there is a dense subset M ⊂ X such that Πn x = x for all
x ∈ M;
(ii) {Πn } is stable if supn kΠn k < ∞.
Prove that convergence (see Definition 8.21) is equivalent to consistency and
stability.
8.4 Stability Theory for Nyström’s Method
We recall that only the Banach space X = C(D) (or a Banach space with even
stronger topology) makes sense. Therefore, throughout this section X = C(D) is
chosen.
Lemma 8.24. Assume that D is compact and k ∈ C(D×D). Then kK−Kn k ≥ kKk.
R
Proof. The operator norm kKk can be shown to be equal to supx∈D D |k(x, y)| dy.
Since D is compact,
R the supremum is a minimum; i.e., there is some ξ ∈ D
with kKk = D |k(ξ, y)| dy. For any ε > 0, oneR can construct a function
ϕε ∈ X = C(D) such that kϕε k = 1, kKϕε k ≥ D |k(ξ, y)| dy − ε, and in
addition, ϕε (ξk,n ) = 0 for 1 ≤ k ≤ n. The latter property implies Kn ϕε = 0, so
that kK − Kn k ≥ k(K − Kn )ϕε k = kKϕε k ≥ kKk − ε. As ε > 0 is arbitrary,
the assertion follows. t u
The statement of the lemma shows that we cannot use the argument that the
operator norm convergence kK − Kn k → 0 proves the desired properties.
It will turn out that instead of K − Kn , the products (K − Kn )K and
(K −Kn )Kn may still converge to zero. The next theorem of Brakhage [2] proves
the main step. Here we use the operators S, T ∈ L(X, X) which replace K and Kn
(for a fixed n). In this theorem, X may be any Banach space.
Theorem 8.25. Let X be a Banach space and λ 6= 0. Suppose that the operators
−1
S, T , (λI − S) belong to L(X, X), and that T is compact. Under the condition
−1
k(T − S) T k < |λ| /k (λI − S) k, (8.12a)
−1
also (λI − T ) ∈ L(X, X) exists and satisfies
−1
−1 1 + k (λI − S) k kT k
k (λI − T ) k≤ −1 . (8.12b)
|λ| − k (λI − S) k k(T − S) T k
The solutions of (λI − S) fS = g and (λI − T ) fT = g differ by
−1 k(T − S) T k kfS k + k(T − S) gk

kfS − fT k ≤ k (λI − S) k , (8.12c)
−1
|λ| − (λI − S) k(T − S) T k

−1
kfS − fT k ≤ k (λI − T ) k k(T − S) fS k . (8.12d)
−1
Proof. (i) If (λI − T ) ∈ L(X, X), the identity I = λ1 [(λI − T ) + T ] leads to
−1 −1 −1
(λI − T ) = λ1 [I + (λI − T ) T ]. We replace (λI − T ) on the right-hand
−1
side by the inverse (λI − S) , whose existence is assumed, and define
1 −1
A := [I + (λI − S) T ].
λ
B := A (λI − T ) should approximate the identity:
8.4 Stability Theory for Nyström’s Method 177
−1 −1
B = λ1 [I + (λI − S) T ] (λI − T ) = I − λ1 [T − (λI − S) T (λI − T )]
1 −1
=I− λ (λI − S) [(λI − S) T − T (λI − T )]
1 −1
=I− λ (λI − S) (T − S) T.
By assumption,
−1 −1
k λ1 (λI − S) (T − S) T k ≤ k (λI − S) k k(T − S) T k / |λ| < 1
holds and proves the existence of B −1 .

(ii) Since B = A (λI − T ) is injective (even bijective), also λI − T must be
injective. Lemma 8.13 states that λI − T is surjective; hence, λI − T is bijec-
−1
tive and (λI − T ) ∈ L(X, X) exists. In particular, B = A (λI − T ) implies
−1
(λI − T ) = B −1 A:
h i−1
−1 −1 −1
(λI − T ) = λI − (λI − S) (T − S) T [I + (λI − S) T ].
The inverse on the right-hand side can be estimated by

h i−1
λI − (λI − S)−1 (T − S) T

≤ 1
|λ| − k (λI − S)−1 k k(T − S) T k
−1 −1
(cf. Lemma 5.8). Together with kI + (λI − S) T k ≤ 1 + k (λI − S) k kT k,
we obtain inequality (8.12b).
(iii) Subtracting (λI − T ) fT = g from (λI − S) fS = g, we obtain the expres-
sion λ(fS − fT ) = SfS − T fT = T (fS − fT ) + (S − T )fS and
−1
fS − fT = (λI − T ) (S − T )fS .
This proves inequality (8.12c). Interchanging S and T , we arrive at fT − fS =

−1
(λI − S) (T − S)fT . Replace fT on the right-hand side by fT = λ1 (g + T fT ):
1 −1
fT − fS = λ (λI − S) (T − S)(g + T fT )
1 −1 1 −1
= λ (λI − S) (T − S)(g + T fS ) + λ (λI −S) (T −S)T (fT −fS ).
This proves
h i−1
−1 1 −1
fT − fS = λI − (λI − S) (T − S) T λ (λI − S) (T − S)(g + T fS ).
Norm estimates yield the assertion. t

u
The key assumption of Theorem 8.25 is that k(T − S) T k is small enough. Set
S = K and replace the fixed operator T by a sequence {Kn }. Then we have to take
care that (K − Kn )Kn → 0. This will be achieved by the following definition of
collectively compact operators (cf. Anselone [1]).
Definition 8.26 (collective compactness). A sequence {Kn } ⊂ L(X, X) is called

collectively compact if the following set is precompact:
{Kn ϕ : ϕ ∈ X, kϕk ≤ 1, n ∈ N} . (8.13)
Since we have introduced collective compactness to analyse the Nyström method,

we have to demonstrate that the Nyström method produces collectively compact op-
erators Kn . The following assumption that k is continuous is not restrictive, since a
quadrature method requires point evaluations.
Lemma 8.27. Assume that D is compact and k ∈ C(D × D). Then any Nyström
method based on a stable quadrature rule yields a collectively compact set {Kn } of
operators.
Pn (i) Stability of the quadrature rule in (8.7) is expressed by the inequality

Proof.
k=1 |ak,n | ≤ Cstab . Pn
Pn(ii) The set (8.13) is uniformly bounded: | k=1 ak,n k(·, ξk,n )ϕ(ξk,n )| ≤
k=1 |ak,n | |k(·, ξk,n )| |ϕ(ξk,n )| ≤ C := Cstab kkk∞ .
Part i,kϕk≤1
(iii) Since D × D is compact, k(·, ·) is uniformly continuous. In particular,
for all ε > 0 there is some δε > 0 such that x, y ∈ D with kx − yk ≤ δε
satisfy |k(x, ξk,n ) − k(y, ξk,n )| ≤ ε/Cstab . Hence the elements of (8.13) are
equicontinuous:

Xn n
X
ak,n k(x, ξk,n )ϕ(ξk,n ) − ak,n k(y, ξk,n )ϕ(ξk,n )

k=1 k=1
Xn
≤ |ak,n | |k(x, ξk,n ) − k(y, ξk,n )| |ϕ(ξk,n )| ≤ Cstab · ε/Cstab = ε.
k=1
By Theorem 3.51 (Arzelà–Ascoli), the set (8.13) is precompact; i.e., {Kn } is

collectively compact. t
u
We show that the operator norm convergence, which is valid for the previous
discretisations, implies the new property.
Lemma 8.28. Suppose that each operator Kn is compact (e.g., because of a
finite-dimensional range). If kK − Kn k → 0, then {Kn } is collectively compact.
Proof. (i) Remark 8.4 states that K is compact.
(ii) We have to show that M := {Kn ϕ : ϕ ∈ X, kϕk ≤ 1, n ∈ N} is precompact.
A sequence in M has the form {Kn(j) ϕj } with n(j) ∈ N and kϕj k ≤ 1. Let
N := lim sup n(j) ∈ N ∪ {∞}.
Case A: N < ∞. The subsequence can be chosen such that nk := n(jk ) = N
for all k ∈ N. By the compactness of KN , the sequence KN ϕjk = Knk ϕjk has a
convergent subsequence.
Case B: N = ∞. A subsequence can be chosen such that nk+1 > nk . By
the same argument as before, there is a convergent subsequence Kϕjk . Since
8.4 Stability Theory for Nyström’s Method 179
kK − Knk k → 0 implies (K − Knk )ϕjk → 0, the same subsequence {jk } yields

a convergent sequence Knk ϕjk = Kϕjk − (K − Knk )ϕjk . t
u
Next, we list some properties of collectively compact operators.
Lemma 8.29. Let {Kn } be collectively compact.

(a) Each Kn is compact.
(b) If {Kn } is consistent with respect to K (i.e., Kn ϕ → Kϕ for all ϕ ∈ X) ,
then K is compact, {Kn } is uniformly bounded, and
k(K − Kn )Kk → 0 and k(K − Kn )Kn k → 0.
Proof. (i) Let M be the set defined in (8.13). Part (a) follows immediately, since
{Kn ϕ : ϕ ∈ X, kϕk ≤ 1} ⊂ M for any n.
(ii) As Kϕ = lim Kn ϕ, it belongs to the closure M which is compact. Therefore,
K is compact.
(iii) {Kn } is uniformly bounded because of Corollary 3.39.
(iv) The image of E := {ϕ ∈ X, kϕk ≤ 1} under K or Kn is contained in the
compact set M . Hence, Lemma 3.49 and the definition of the operator norm prove
that k(K − Kn )Kk = supϕ∈E k(K − Kn )Kϕk ≤ supψ∈M k(K − Kn )ψk → 0
and similarly for (K − Kn )Kn . t u
Exercise 8.30. Prove: (a) If B ⊂ X is bounded and {Kn } collectively compact,

then {Kn f : f ∈ B, n ∈ N} is precompact.
(b) If {An } is collectively compact and {Bn } uniformly bounded (a weaker
condition is collective compactness; cf. Lemma 8.29), then {An Bn } is collectively
compact.
The statement of Exercise 8.30b corresponds to the fact that AB is compact if

A is compact and B bounded. The conditions on A and B can be interchanged:
AB is compact if A is bounded and B compact. The analogue for collective
compactness is given in the next lemma. Note that consistency of {An } implies
uniform boundedness (cf. Exercise 8.6a).
Lemma 8.31. Suppose that {An } is consistent, while {Bn } is collectively compact.
Then {An Bn } is collectively compact.
Proof. We have to show that the set M := {An Bn f : kf k ≤ 1, n ∈ N} is pre-

compact. A sequence from M has the form
ϕk := Ank Bnk fk with kϕk k ≤ 1 and nk ∈ N.
Set gk := Bnk fk . By collective compactness of {Bn }, there is a subsequence

(denoted again by k) such that
gk = Bnk ϕk → g.
Next, two cases must be distinguished: (A) sup nk < ∞ and (B) sup nk = ∞.
Case A: We can select a subsequence that such nk = n is constant; i.e., gk =

Bn ϕk → g. Continuity of An proves ϕk = An gk → An g; i.e., a convergent
subsequence is found.
Case B: Then ϕk = Ank gk , where, without loss of generality, we may assume
nk+1 > nk . By consistency, there is some A with
Ank g → Ag and sup kAnk k < ∞.

nk ∈N
Therefore, Ank gk − Ag = Ank (gk − g) + (Ank − A)g → 0 shows that Ank gk is

a convergent subsequence. t
u
Combining Theorem 8.25 and Lemma 8.29, we obtain the following main result.
−1
Theorem 8.32. Let λ 6= 0 satisfy5 (λI − K) ∈ L(X, X). Suppose that {Kn } is
consistent with respect to K and collectively compact. Then {Kn } is stable and
convergent. Furthermore, there is some n0 ∈ N such that for all n ≥ n0 the
following statements hold:
−1
−1 1 + k (λI − K) k kKn k
k (λI − Kn ) k≤ −1 (n ≥ n0 );
|λ| − k (λI − K) k k(K − Kn )Kn k
−1 −1
the solutions f = (λI − K) g and fn = (λI − Kn ) g satisfy
−1 k(K − Kn )Kn k kf k + kK − Kn k kgk

kf − fn k ≤ k (λI − K) k −1 ,
|λ| − k (λI − K) k k(K − Kn )Kn k
−1
kf − fn k ≤ k (λI − Kn ) k k(K − Kn )f k (n ≥ n0 ).
Conclusion 8.33. Assume that D is compact and k ∈ C(D × D). Let λ 6= 0 be

a regular value of K. Then, any Nyström method based on a consistent and stable
quadrature is also stable and convergent.
Proof. By Lemma 8.27 {Kn } is collectively compact. Consistency of the quadrature

method implies consistency of {Kn }. Now, Theorem 8.32 yields the statement. t u
8.5 Perturbation Results

−1
Lemma 8.34. Let {Kn } be stable with constant Cstab : k (λI − Kn ) k ≤ Cstab .
Let {Tn } be bounded by kTn k ≤ ζ/Cstab with ζ < 1 for all n ≥ n0 . Then, also
{Kn + Tn } is stable.
Proof. The proof follows the same lines as in Remark 8.8. t

u
5
In Lemma 8.40 we shall show that this assumption is necessary.
8.6 Application to Eigenvalue Problems 181
Theorem 8.35. Suppose that λ 6= 0 is a regular value of K + T ; i.e., the inverse

(λI − K − T )−1 ∈ L(X, X) exists. Assume that {Kn } is convergent and consistent
with respect to K ∈ L(X, X), {Tn } is collectively compact with Tn ϕ → T ϕ for
all ϕ ∈ X. Then, also {Kn + Tn } is stable, consistent, and convergent.
Proof. (i) By Theorem 8.16b, λI − K is bijective and (λI − K)−1 ∈ L(X, X)

exists.
(ii) To prove that {(λI − Kn )−1 Tn } is collectively compact, we apply Lemma
8.31 with An = (λI − Kn )−1 and Bn = Tn . By Theorems 8.16 and 8.17a, An is
consistent with respect to A = (λI − K)−1 ; i.e., the suppositions of Lemma 8.31
are satisfied.
(iii) For any ϕ ∈ X, Theorem 8.17b with gn := Tn ϕ → T ϕ proves that
Cn := (λI − Kn )−1 Tn satisfies Cn ϕ → Aϕ, where C := (λI − K)−1 T . This
shows that {An } is consistent with respect to A.
(iv) Because of λI − K − T = (λI − K)(I − C) and (i), also the inverse
(I − C)−1 ∈ L(X, X) exists. Since {Cn } is collectively compact (cf. (ii)), and
consistent with respect to C (cf. (iii)) with (I − C)−1 ∈ L(X, X), Theorem 8.32
−1
states stability: (I − Cn )
≤ C. From
(λI − Kn − Tn )−1 = (I − Cn )−1 (λI − Kn )−1
we conclude that also {Kn + Tn } is stable.

(v) Consistency of {Kn + Tn } is trivial. Since λI − K − T = (λI − K)(I − C)
is bijective (cf. (iv)), convergence follows from Theorem 8.17a. tu
8.6 Application to Eigenvalue Problems
So far, we have required that (λI − K)−1 and (λI − Kn )−1 , at least for n ≥ n0 ,
exist. Then, λ ∈ C is called a regular value of K and Kn . The singular values of K
[Kn ] are those for which λI − K [λI − Kn ] is not bijective. The Riesz–Schauder
theory states that for compact K, all singular values are either λ = 0 or eigenvalues,
which means that there is an eigenfunction f such that
λf = Kf (0 6= f ∈ X). (8.14a)
Similarly, we have discrete eigenvalue problems
λn fn = Kn fn (0 6= fn ∈ X, n ∈ N). (8.14b)
The sets of all singular values are the spectra σ = σ(K) and σn = σ(Kn ), respec-
tively. The Riesz–Schauder theory states the following.
Remark 8.36. Either the range of a compact operator is finite-dimensional (imply-

ing that σ is a finite set), or λ = 0 ∈ σ is the only accumulation point of the
spectrum.
The obvious question is how λ ∈ σ and λn ∈ σn are related. The following

results require only consistency of {Kn } to K and collective compactness of {Kn },
which holds for all discretisations discussed before.
Theorem 8.37. Suppose that {Kn } is consistent with respect to K and collectively
compact. Let {λn } be a sequence of eigenvalues of (8.14b) with corresponding
eigenfunction fn ∈ X normalised by kfn k = 1. Then, there exists a subsequence
{nk } such that either λnk → 0 or λnk → λ ∈ σ. In the latter case, the subsequence
can be chosen such that fnk → f with f being an eigenfunction of (8.14a).
Proof. (i) First we show uniform boundedness: sup{|λn | : λn ∈ σn , n ∈ N} < ∞.

|λn | ≤ kKn k follows from |λn | = kλn fn k = kKn fn k ≤ kKn kkfn k = kKn k
because of kfn k = 1. Consistency of Kn ϕ → Kϕ for all ϕ ∈ X implies that Kn is
uniformly bounded: C := supn∈N kKn k < ∞. Together, |λn | ≤ C follows.
(ii) Since {z ∈ C : |z| ≤ C} is compact, there is a convergent subsequence
λn → λ (n ∈ N0 ). If λ 6= 0, we have to show that λ ∈ σ. Because of kfn k = 1,
collective compactness of {Kn } implies that there is a second subsequence
n ∈ N00 ⊂ N0 for which Kn fn is convergent. By (8.14b) and λn → λ 6= 0,
this yields convergence of fn to some f :
1
fn = Kn fn → f.
λn
Continuity of the norm yields kf k = 1, implying f 6= 0. The first two terms in
Kn fn = Kn (fn − f ) + (Kn − K) f + Kf vanish (use supn∈N kKn k < ∞ and
consistency), so that Kn fn → Kf as well as Kn fn = λn fn → λf, proving that
λ ∈ σ is an eigenvalue of K. tu
The statement of TheoremS 8.37 about the eigenvalues can be abbreviated by ‘the
set of accumulation points of n∈N σn is contained in σ’. NowSwe prove the reverse
direction: ‘σ is contained in the set of accumulation points of n∈N σn ’.
Theorem 8.38. Suppose that {Kn } is consistent with respect to K and collectively
compact. Then, for any 0 6= λ ∈ σ there is a sequence {λn } of eigenvalues from
(8.14b) with λ = limn∈N λn . Again, a subsequence can be chosen so that the eigen-
functions fnk from (8.14b) converge to an eigenfunction of (8.14a).
We prepare the proof by two lemmata. The following functions on C will be

needed:
1/k(λI − Kn )−1 k if λ ∈

/ σn ,
κn (λ) :=
0 otherwise.
Exercise 8.39. Prove that κn is not only continuous, but also satisfies the Lipschitz
property |κn (λ) − κn (µ)| ≤ |λ − µ| for all λ, µ ∈ C.
The next lemma describes the equivalence of limn∈N κn (λ) = 0 and λ ∈ σ.

8.6 Application to Eigenvalue Problems 183
Lemma 8.40. Suppose that {Kn } is consistent with respect to K.

(a) limn∈N κn (λ) = 0 for each λ ∈ σ.
(b) Suppose, in addition, that the set {Kn } is collectively compact, λ 6= 0, and
limn∈N κn (λ) = 0. Then λ ∈ σ is an eigenvalue of K.
Proof. (i) Let λ ∈ σ with corresponding eigenfunction f 6= 0: λf = Kf . Set
gn := λf − Kn f . Consistency implies that gn → 0. For an indirect proof assume
that limn∈N κn (λ) ≥ 2ε > 0 and choose a subsequence such that κn (λ) ≥ ε; i.e.,
k(λI − Kn )−1 k ≤ 1/ε. Then
kf k = k(λI − Kn )−1 gn k ≤ k(λI − Kn )−1 kkgn k ≤ kgn k/ε → 0
is a contradiction to f 6= 0. This ends the proof of Part (a).

(ii) The assumptions of Part (b) imply k(λI − Kn )−1 k → ∞ (we write formally
k(λI − Kn )−1 k = ∞ for λ ∈ σn ). Hence, there are functions {fn } and {gn } with
(λI − Kn ) fn = gn , kfn k = 1, gn → 0.
By collective compactness, a subsequence {Knk fnk } is convergent. Therefore

fnk = (gnk + Knk fnk )/λ has some limit f with
kf k = 1 and f = lim(gnk + Knk fnk )/λ = Kf /λ;
i.e., λ is an eigenvalue of K (details as in Part (ii) of the proof above). t

u
Lemma 8.41. Let Ω ⊂ C\{0} be compact. Then, either σn ∩ Ω 6= ∅ or κn takes its
minimum on the boundary ∂Ω.
Proof. Assume that σn ∩Ω = ∅ and fix some ϕ ∈ X with kϕk = 1. By assumption,
(zI − Kn )−1 ϕ is well-defined for all z ∈ Ω. Since κn is continuous, and Ω is com-
pact, maxz∈Ω κn (z) = κn (λ) holds for some λ ∈ Ω. Choose Φ ∈ X ∗ according to
∗
Corollary 3.11 such that kΦkX = 1 and Φ((λI − Kn )−1 ϕ) = (λI − Kn )−1 ϕ ,
and define a complex function by
F (z) := Φ((zI − Kn )−1 ϕ) ∈ C.
Its derivative F 0 (z) = −Φ((zI −Kn )−2 ϕ) is defined in Ω, so that F is holomorphic

in Ω. Hence, |F | takes its maximum on ∂Ω:
(λI − Kn )−1 ϕ = F (λ) = |F (λ)| ≤ max{|F (z)| : z ∈ ∂Ω}.

The inequality
∗
|F (z)| = |Φ((zI − Kn )−1 ϕ)| ≤ kΦkX (zI − Kn )−1 ϕ

≤ (zI − Kn )−1 = 1/κn (z)

holds for all z ∈ ∂Ω, so that

(λI − Kn )−1 ϕ ≤ max{1/κn (z) : z ∈ ∂Ω} = 1/ min{κn (z) : z ∈ ∂Ω}.

Since the choice of ϕ ∈ X with kϕk = 1 is arbitrary,
(λI − Kn )−1 ≤
1
min{κn (z) : z ∈ ∂Ω}
follows, which is equivalent to min{κn (z) : z ∈ ∂Ω} ≤ κn (λ); i.e., the minimum
is taken on ∂Ω. t u
Now, we give the proof of Theorem 8.38.
(i) K is compact (cf. Remark 8.29b), and each eigenvalue 0 6= λ ∈ σ is isolated
(Remark 8.36). Choose a complex neighbourhood Ω of λ such that
0∈
/ Ω ⊂ C, Ω ∩ σ = {λ}, λ ∈ Ω\∂Ω, Ω compact. (8.15)
For αn := min{κn (z) : z ∈ ∂Ω} we want to show αn ≥ ε > 0 for n ≥ n0 .

For an indirect proof, assume αnk → 0. Define zn ∈ ∂Ω by αn = κn (zn ). Since
∂Ω is compact, znk → ζ for some ζ ∈ ∂Ω. By |κnk (ζ) − κnk (znk )| ≤ |ζ − znk |
(cf. Exercise 8.39), κnk (znk ) → 0 follows and implies that ζ ∈ Ω ∩ σ (cf. Lemma
8.40b) in contradiction to (8.15).
(ii) Assume αn ≥ ε > 0 for n ≥ n0 . As κn (λ) → 0 (cf. Lemma 8.40a), one can
choose n0 such that κn (λ) ≤ ε/2 for n ≥ n0 . Since the minimum is not attained
on ∂Ω, one concludes from Lemma 8.41 that Ω ∩ σn 6= ∅; i.e., there are eigen-
values λn ∈ σn for all n ≥ n0 . Since Ω can be an arbitrarily small neighbourhood,
there are λn ∈ σn with lim λn = λ. Theorem 8.37 yields the statement about the
eigenfunctions.
References
1. Anselone, P.M.: Collectively Compact Operator Approximation Theory and Applications to

Integral Equations. Prentice-Hall, Englewood Cliffs (1971)
2. Brakhage, H.: Über die numerische Behandlung von Integralgleichungen nach der Quadratur-
formelmethode. Numer. Math. 2, 183–196 (1960)
3. Fredholm, E.I.: Sur une classe d’équations fonctionnelles. Acta Math. 27, 365–390 (1903)
5. Hsiao, G.C., Wendland, W.L.: Boundary Integral Equations. Springer, Berlin (2008)
6. Linz, P.: A general theory for the approximate solution of operator equations of the second
kind. SIAM J. Numer. Anal. 14, 543–554 (1977)
7. Nyström, E.J.: Über die praktische Auflösung von linearen Integralgleichungen mit Anwen-
dungen auf Randwertaufgaben der Potentialtheorie. Soc. Sci. Fenn. Comment. Phys.-Math.
4(15) (1928)
8. Sauter, S.A., Schwab, C.: Boundary element methods, Springer Series in Computational
Mathematics, Vol. 39. Springer, Berlin (2011)
9. Schmidt, E.: Zur Theorie der linearen und nichtlinearen Integralgleichungen. I. Teil: Entwick-
lung willkürlicher Funktionen nach Systemen vorgeschriebener. Math. Ann. 63, 433–476
(1907)
10. Yosida, K.: Functional Analysis. Springer, Berlin (1968)
Index
Adams–Bashforth method, 85 CFL condition, 120

algebra, 32 characteristic function, 117
algorithm, 3 Chebyshev polynomials
finite, 3 of first kind, 52
stable, 10 of second kind, 133
unstable, 10, 11 classical solution, 140
amplification, 8 coercivity, 150
Approximation, 60 collectively compact, 178
approximation compact, 31, 44, 59, see operator, compact
external, 159 collectively, 178
internal, 159 pre-, 44, 45, 59
approximation theorem of Weierstrass, 30 relatively, 44
associated matrix norm, 11 companion matrix, 81
asymptotic expansion, 28 complete, 36
condition, 25
B-spline, 57 of a matrix, 175
Babuška–Brezzi condition, 156 condition (number), 9
Baire’s category theorem, 37 conductivity, 96
Baire–Hausdorff theorem, 37 consistency, 7, 20, 30, 39, 49, 71, 73, 75, 105,
Banach space, 36 142, 172
strictly convex, 60 consistency order, 72, 73, 75, 85, 88
Banach’s fixed-point theorem, 66 consistency versus stability, 42, 59, 86, 135,
basis, dual, 7 163
Bessel function, 4 convection-diffusion equation, 163
bilinear form, 149 convergence, 22, 23, 39, 49, 72, 75, 105, 143,
adjoint, 154 173
bounded, 149 arbitrary slow, 43
coercive, 150 order of, 147
symmetric, 155 convergence order, 73, 85
boundary condition, 139 convergence speed, 43
boundary value problem, 139 convergence theorem, 35, 51, 84, 106, 144,
variation formulation, 149 174
broken Sobolev norm, 161 Courant–Friedrichs–Lewy criterion, 120
Crank–Nicolson scheme, 123, 135
cancellation, 8 criterion
cardinal spline, 57 Courant–Friedrichs–Lewy, 120
Cea’s lemma, 153 Friedrichs, 126

Springer Series in Computational Mathematics 45, DOI 10.1007/978-3-642-39386-0,
186 Index
Lax–Wendroff, 125 explicit, 71

von Neumann, 125 implicit, 70, 71
cylinder function, 4 Euler’s constant, 52
expansion, asymptotic, 28
DAE, 91 extrapolation, Richardson, 28
defect correction, 164
degenerate kernel, 169 finite element method, 141
difference equation, 78, 102 five-point scheme, 140, 163
stable, 80 fixed-point equation, 66
difference scheme, 140 fixed-point iteration, 66
explicit, 102 Fourier analysis, 115
implicit, 121 Fourier series, 115
inhomogeneous, 82 Fourier synthesis, 116
positive, 109, 110, 126 Fredholm integral equation
differential equation first kind, 167
hyperbolic, 95, 165 second kind, 167
inhomogeneous, 100 Friedrichs’ scheme, 128
ordinary, 63 functional, 23
parabolic, 96, 135
partial, 94 Galerkin method, 150
symmetric hyperbolic, 127 discontinuous, 161
uniformly elliptic, 139 Gauss quadrature, 19
differential equations Gelfand triple, 159
stiff, 91 Green’s representation, 148
differential-algebraic equations, 91 grid function, 101
differentiation, numerical, 13
Dirac function(al), 24 hat function, 56
Dirichlet kernel, 61 heat equation, 96, 129
discontinuous Galerkin method, 161 Hermite interpolation, 24, 47
discretisation error Heun method, 65, 71
global, 73 Hölder space, 44
local, 71, 75, 105 hyperbolicity
dissipative, 131 regular, 128
dissipativity, 135 symmetric, 127
domain, 94
dominant convection, 163, 165 inf-sup condition, 150, 156
Du Fort–Frankel scheme, 134 initial-boundary value problem, 131
dual basis, 7 initial-value problem, 63
dual norm, 23 well-conditioning of the, 68
dual space, 23, 149 interpolation, 47
Hermite, 24, 47
eigenvalue problem, 11, 12, 181 Nyström, 171
elementary operations, 3 polynomial, 48
accuracy of, 7 tensor product, 55
energy method, 114 interpolation points, 47
equi-oscillation, 60 Chebyshev, 52
equicontinuous, 45 equidistant, 51
equivalence theorem, 39, 51, 108, 175 interpolatory quadrature, 18
error inverse, right-, 102
absolute, 6
relative, 6 Jordan normal form, 77
error amplification, 8
error analysis, linear, 9 kernel
Euler method, 64 degenerate, 169
Index 187
Peano, 21 normed space, 36

kernel approximation, 169 numerical differentiation, 13
Kreiss matrix theorem, 76 numerical radius, 125
Nyström interpolation, 171
Lagrange function, 18 Nyström method, 171, 176
Lagrange polynomial, 18, 28, 48
Laplace operator, 129 one-step method, 64, 76
Lax–Wendroff scheme, 136 explicit, 64, 70
leap-frog scheme, 132 implicit, 70
Lebesgue constant, 50 operator
Legendre polynomial, 19 almost normal, 113
Lipschitz condition, 84 compact, 59, 168, 170
Lipschitz continuity, 64, 71, 126, 131 multiplication, 117
Lipschitz domain, 148 normal, 112
shift, 104
M-matrix, 141, 146, 163 solution, 98
machine number, 7 transfer, 101
machine precision, 7 operator norm, 36
matrix operator-norm convergence, 59
companion, 81 order
power bounded, 76, 81 consistency, 73
sparse, 141 convergence, 73
stable, 76 overflow, 7
matrix norm, 76
associated, 11, 77, 82 Parseval equality, 116
row-sum, 11, 77 patch test, 161
spectral, 11 Peano kernel, 21, 22
maximum norm, 11, 21 perturbation, 42, 157, 180
midpoint formula, 65 perturbation lemma, 110
multiplication operator, 117 Poisson equation, 139
multiplicity polynomial
algebraic, 76 characteristic, 12, 74, 86
geometric, 76 Chebyshev, 52, 133
multistep method, 74, 132 Lagrange, 18, 28, 48
explicit, 65 Legendre, 19
linear, 74, 83, 85 root, 12, 79, 80
optimal, 85 trigonometric, 117
zeros of a, 11
Newton–Cotes quadrature, 19 polynomial interpolation, 48
nodal points, 47 instability of, 51
norm piecewise, 56
dual, 23 precompact, see compact, pre-, 59
L2 , 94 problem, 3
`2 , 101 boundary value, 139
Lp , 95, 108 eigenvalue, 11, 12, 181
`p , 108 ill conditioned, 9
`∞ , 101 initial-value, 63
matrix, see matrix norm saddle point, 155
maximum, 11, 21, 32, 38 Stokes, 155
operator, 36, 38, 59 t-regular, 154
row-sum, 11, 77 well conditioned, 9, 11, 25, 68
spectral, 11 projection method, 170, 175
supremum, 80 projection, orthogonal, 171
norm equivalence, 76, 82 prolongation, 102
188 Index
quadrature, 17, 171 spline

Gauss, 19, 20, 27 B-, 57
interpolatory, 18, 19, 39 cardinal, 57
Newton–Cotes, 19, 20, 27 cubic, 57
Romberg, 28 stability, 25, 27, 30, 49, 72, 76, 80, 81, 92, 106,
quadrature weights, 17 109, 117, 144, 172
conditional, 106
recurrence relation, 4, 7 unconditional, 106, 123
regular value, 181 stability constant, 106
regularity, 148, 149, 153 stability theorem, 39, 51, 83, 107, 144, 174
Remez algorithm, 60 stiff differential equations, 91
restriction, 101, 142 Stokes problem, 155
Richardson extrapolation, 28 Stone–Weierstrass theorem, 30, 33
Riesz isomorphism, 151 Strang lemma
Riesz–Schauder theory, 173, 181 first, 158
Riesz–Thorin theorem, 108 second, 160
right-inverse, 102 streamline diffusion method, 163
Romberg quadrature, 28 strictly convex, 60
row-sum norm, 11 support of a function, 56
Runge-Kutta method, 65 symbol, 117
system, regularly hyperbolic, 128
saddle point problem, 155
scheme tensor product interpolation, 55
Crank–Nicolson, 123, 135 tensor rank, 14
difference, see difference scheme theorem
Du Fort–Frankel, 134 Arzelà–Ascoli, 45
five-point, 140, 163 Baire’s category, 37
Friedrichs’, 128 Baire–Hausdorff, 37
Lax–Wendroff, 136 Banach’s fixed-point, 66
leap-frog, 132 Brakhage, 176
theta, 123 convergence, see convergence theorem
semigroup Dahlquist, 86
continuous, 100 equivalence, see equivalence theorem
generated by A, 99 Hahn–Banach, 23
with neutral element, 99 Kreiss matrix, 76
separation of points, 30 open mapping, 168
shift operator, 104 Peano, 64
singular value, 181 Riesz–Thorin, 108
singular value decomposition stability, 39, see stability theorem
infinite, 170 Stone–Weierstrass, 30, 33
Sobolev norm, broken, 161 uniform boundedness, 36
Sobolev space, 148, 149 Weierstrass’ approximation, 30, 170
solution theta scheme, 123
classical, 98, 140 transfer operator, 101
generalised, 99 trapezoidal rule, compound, 28, 40
strong, 98, 99
weak, 99, 150 underflow, 7
solution operator, 98 uniform boundedness theorem, 36
space, dual, 23
spectral norm, 11 variation formulation, 149
spectral radius, 112, 125
spectrum, 77, 181 weak solution, 150
speed of convergence, 43 Wilson’s element, 161

The Concept of Stability in Numerical Mathematics

Uploaded by

Copyright:

Available Formats

The Concept of Stability in Numerical Mathematics

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Concept of Stability in Numerical Mathematics

Uploaded by

Copyright:

Available Formats

Springer Series in Computational Mathematics 45

For further volumes:

The Concept of Stability in

Library of Congress Control Number: 2014931977

© Springer-Verlag Berlin Heidelberg 2014

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Usually, the subject of a book on numerical mathematics pertains to a certain field

Leipzig, October 2013 Wolfgang Hackbusch

2 Stability of Finite Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

3.4.7 Uniform Boundedness Theorem . . . . . . . . . . . . . . . . . . . . . . . . 36

5 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.5.6 Construction of Optimal Multistep Methods . . . . . . . . . . . . . . 85

6 Instationary Partial Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . 93

7 Stability for Discretisations of Elliptic Problems . . . . . . . . . . . . . . . . . . . 139

7.5.5 Quantitative Discretisation Error and Regularity . . . . . . . . . . 153

8 Stability for Discretisations of Integral Equations . . . . . . . . . . . . . . . . . . 167

k·k2 spectral norm of a matrix; cf. §2.4.2.2

δa, δA, . . . perturbations of quantities a, A, . . .

ai,n quadrature weights in §3

Ln (·) Legendre polynomial; cf. §3.1.4

In numerical mathematics we have to distinguish between two types of methods.

W. Hackbusch, The Concept of Stability in Numerical Mathematics, 1

2.1 About Algorithms

An algorithm is used to solve a (numerical) problem. For the mathematical formula-

Φ : (x1 , . . . , xn ) ∈ X = domain(Φ) 7→ (y1 , . . . , ym ) ∈ Y = range(Φ). (2.1)

The algorithm uses the intermediate values x(1) = (z1 , z2 ) with z1 := x1 x2 ,

W. Hackbusch, The Concept of Stability in Numerical Mathematics, 3

2.2 A Paradoxical Example

2.2.1 First Algorithm

The chosen task is the computation of J5 (0.6).

J0 (0.6) = 0.9120, J1 (0.6) = 0.2867, (2.4)

2.2.2 Second Algorithm

Before we give an explanation, we show a second ‘paradox’: an algorithm based on

J0 (x) + 2J2 (x) + 2J4 (x) + . . . = 1. (2.10)

The choice m = 10 yields

J5 (0.6) = 1.994819537430010 -5,

where all 14 digits are correct.

For the latter identity use ex e−x = 1.

all jn depend linearly on j0 , j1 . In particular, we obtain

2.3 Accuracy of Elementary Operations

Floating point arithmetic used in computers is characterised by a fixed mantissa

The minimiser x of minx∈M |x − ξ| is the best ‘rounding’ of ξ ∈ R. Denoting

|fl (a op b) − (a op b)| ≤ eps |a op b| for all a, b ∈ M. (2.13)

Remark 2.5. Inequality (2.13) can be regarded as a ‘consistency’ property.

2.4 Error Amplification

The mathematical term ‘amplification’ will be used in a quite general way. If an

η = ξ1 − ξ2 , x1 = rd(ξ1 ), x2 = rd(ξ2 ), ỹ = fl(x1 − x2 ),

where ξ1 and ξ2 are general real numbers. The interesting error is

|η − ỹ| for ỹ = fl(x1 − x2 ).

The absolute error is bounded by

|η − ỹ| ≤ |x1 − x2 − fl(x1 − x2 )| + |∆x1 | + |∆x2 |

2.4.1.1 Linear (Differential) Error Analysis

Which error is of interest (absolute or relative) depends on the particular problem.

2.4.1.2 Condition and Stability

A problem is called well-conditioned if the amplification of the input error is not

x(0) = (x1 , . . . , xn ) 7→ x(1) 7→ . . . 7→ x(p) = {y1 , . . . , ym }.

We regain the mapping Φ of the problem as the composition Φ = ϕp ◦ϕp−1 ◦. . .◦ϕ1 .

Φ(q) := ϕp ◦ ϕp−1 ◦ . . . ◦ ϕq+1 for 0 ≤ q ≤ p

2.4.2 Further Examples