Kress Linear Integral Equations, 1999

Applied Mathematical Sciences
Volume 82
Editors
J.E. Marsden L. Sirovich
Advisors
S. Antman J.K. Hale P. Holmes
T. Kambe J. Keller K. Kirchgassner
BJ. Matkowsky C.S. Peskin
1. John: Partial Differential Equations. 4th ed. 34. Kevorkian/Cole: Perturbation Method.~ in
2. Sirovich: Techniques of Asymptotic Analysis. Applied Mathematics.
3. Hale: Theory of Functional Differential 35. Carr: Applications of Centre Manifold Theory.
Equations. 2nd ed. 36. Bengtsson/GhiVKiillen: Dynamic Meteorology:
4. Percus: Combinatorial Methods. Data Assimilation Methods.
5. von Mises/Friedrichs: Fluid Dynamics. 37. Saperstone: Semidynamical Systems in Infinite
6. FreibergerlGrenander: A Short Course in Dimensional Spaces.
Computational Probability and Statistics. 3S. Lichtenberg/Lieberman: Regular and Chaotic
7. Pipkin: Lectures on Viscoelasticity Theory. Dynamics. 2nd ed.
S. Giacoglia: Perturbation Methods in Non-linear 39. PicciniiStampacchiaiVidossich: Ordinary
Systems. Differential Equations in R".
9. Friedrichs: Spectral Theory of Operators in 40. NaylorlSell: linear Operator Theory in
Hilbert Space. Engineering and Science.
10. Stroud: Numerical Quadrature and Solution of 41. Sparrow: The Lorenz Equations: Bifurcations.
Ordinary Differential Equations. Chaos. and Strange Attractors.
11. Wolovich: linear Multivarlable Systems. 42. GuckenheimerlHolmes: Nonlinear Oscillations.
12. Berkovitz: Optimal Control Theory. Dynamical Systems. and Bifurcations of Vector
13. Bluman/Cole: Similarity Methods for Fields.
Differential Equations. 43. Ockendort/Faylor: Inviscid Fluid Flows.
14. Yoshizawa: Stability Theory and the Existence 44. Pazy: Semigroups of linear Operators and
of Periodic Solution and Almost Periodic Applications to Partial Differential Equations.
Solutions. 45. GlashofflGustafson: Linear Operations and
IS. Braun: Differential Equations and Their Approximation: An Introduction to the
Applications. 3rd ed. Theoretical Analysis and Numerical Treatment
16. Lefschetz: Applications of Algebraic Topology. of Semi-Infinite Programs.
17. Collatz!Wetterling: Optimization Problems. 46. Wilcox: Scattering Theory for Diffraction
IS. Grenander: Pattern Synthesis: Lectures in Gratings.
Pattern Theory. Vol. I. 47. Hale et al: An Introduction to Infinite
19. Marsden/McCracken: Hopf Bifurcation and Its Dimensional Dynamical Systems-Geometric
Applications. Theory.
20. Driver: Ordinary and Delay Differential 4S. Murray: Asymptotic Analysis.
Equations. 49. Ladyzhenskaya: The Boundary-Value Problems
21. CouranrlFriedrichs: Supersonic Flow and Shock of Mathematical Physics.
Waves. 50. Wilcox: Sound Propagation in Stratified Fluids.
22. RoucheIHabets/Laloy: Stability Theory by 51. GolubitskylSchaeffer: Bifurcation and Groups in
liapunov's Direct Method. Bifurcation Theory. Vol. I.
23. Lamperti: Stochastic Processes: A Survey of the 52. Chipot: Variational Inequalities and Flow in
Mathematical Theory. Porous Media.
24. Grenander: Pattern Analysis: Lectures in Pattern 53. Majda: Compressible Flnid Flow and System of
Theory. Vol. II. Conservation Laws in Several Space Variables.
25. Davies: Integral Transform~ and Their 54. Wasow: linear Turning Point Theory.
Applications. 2nd ed. 55. Yosida: Operational Calculus: A Theory of
26. KushnerlClark: Stochastic Approximation Hyperfunctions.
Methods for Constrained and Unconstrained 56. Chang/Howes: Nonlinear Singular Perturbation
Systems. Phenomena: Theory and Applications.
27. de Boor: A Practical Gnide to Splines. 57. Reinhardt: Analysis of Approximation Methods
2S. Keil.wn: Markov Chain Models-Rarity and for Differential and Integral Equations.
Exponentiality. 58. DwoyerlHussainWoigt (eds): Theoretical
29. de Veuheke: A Course in Elasticity. Approaches to Turbulence.
30. Shiatycki: Geometric Quantization and Quantum 59. Sanders/Verhulst: Averaging Method.~ in
Mechanics. Nonlinear Dynamical Systems.
31. Reid: Sturmian Theory for Ordinary Differential 60. GhiVChiidress: Topics in Geophysical
Equations. Dynamics: Atmospheric Dynamics. Dynamo
32. Meis/Markowitz: Numerical Solution of Partial Theory and Climate Dynamics.
Differential Equations.
33. Grenander: Regular Structures: Lectures in
Pattern Theory. Vol. ill. (continued following index)
Rainer Kress
Linear Integral Equations

Second Edition
Springer
Rainer Kress
Institut fUr Numerische und Angewandte
Mathematik
Universităt Gottingen
LotzestraBe 16-18
D-37083 Gottingen
Germany
Editors
J.E. Marsden L. Sirovich
Control and Dynamical Systems, 107-81 Division of Applied Mathematics
California Institute of Technology Brown University
Pasadena, CA 91125 Providence, RI 02912
USA USA
Mathematics Subject Classification (1991): 45A05, 45B05, 45E05, 45L05, 47Gxx, 65110, 65R20
Library of Congress Cataloging-in-Publication Data

Kress, Rainer, 1941-
Linear integral equations / Rainer Kress. - 2nd ed.
p. cm. - (Applied mathematical sciences ; v. 82)
Includes bibliographical references and index.
ISBN 978-1-4612-6817-8 ISBN 978-1-4612-0559-3 (eBook)
DOI 10.1007/978-1-4612-0559-3
1. Integral equations. 1. Title. II. Series: Applied
mathematical sciences (Springer-Verlag New York Inc.) ; v. 82.
QAl.A647 voI. 82 1999
[QA431]
510 s-dc21
[515'.45] 98-51753
Printed on acid-free paper.

© 1989, 1999 Springer Science+Business Media New York
Originally published by Springer-Verlag New York Berlin Heide\berg in 1999
Softcover reprint of the hardcover 2nd edition 1999
Ali rights reserved. This work may not be translated or copied in whole or in part without the
written permission ofthe publisher (Springer Science+Business Media, LLC), except for brief
excerpts in connection with reviews or schoiarly ana1ysis. Use in connection with any form of
information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimiiar methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even if
the former are not especially identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely
byanyone.
Production managed by A. Orrantia; manufacturing supervised by Thomas King.
Photocomposed copy prepared from the author's PostScript files.
987654321
ISBN 978-1-4612-6817-8 SPIN 10707345

To the Memory of My Parents
Preface to the Second Edition
In the ten years since the first edition of this book appeared, integral
equations and integral operators have revealed more of their mathematical
beauty and power to me. Therefore, I am pleased to have the opportunity
to share some of these new insights with the readers of this book. As in the
first edition, the main motivation is to present the fundamental theory of
integral equations, some of their main applications, and the basic concepts
of their numerical solution in a single volume. This is done from my own
perspective of integral equations; I have made no attempt to include all of
the recent developments.
In addition to making corrections and adjustments throughout the text
and updating the references, the following topics have been added: In Sec-
tion 4.3 the presentation of the Fredholm alternative in dual systems has
been slightly simplified and in Section 5.3 the short presentation on the
index of operators has been extended. The treatment of boundary value
problems in potential theory now includes proofs of the jump relations for
single- and double-layer potentials in Section 6.3 and the solution of the
Dirichlet problem for the exterior of an arc in two dimensions (Section 7.6).
The numerical analysis of the boundary integral equations in Sobolev space
settings has been extended for both integral equations of the first kind in
Section 13.4 and integral equations of the second kind in Section 12.4.
Furthermore, a short outline on fast O(n log n) solution methods has been
added in Section 14.4. Because inverse obstacle scattering problems are now
extensively discussed in the monograph [25], in the concluding Chapter 18
the application to inverse obstacle scattering problems has been replaced
by an inverse boundary value problem for Laplace's equation.
viii Preface to the Second Edition
I would like to thank Peter Hahner and Andreas Vogt for carefully read-
ing the manuscript and for a number of suggestions for improving it. Thanks
also go to those readers who helped me by letting me know the errors and
misprints they found in the first edition.
I hope that this book continues to attract mathematicians and scientists
to the field of integral equations and their applications.
G6ttingen, October 1998 Rainer Kress

Preface to the First Edition
I fell in love with integral equations about twenty years ago when I was
working on my thesis, and I am still attracted by their mathematical beauty.
This book will try to stimulate the reader to share this love with me.
Having taught integral equations a number of times I felt a lack of a text
which adequately combines theory, applications and numerical methods.
Therefore, in this book I intend to cover each of these fields with the same
weight. The first part provides the basic Riesz-Fredholm theory for equa-
tions of the second kind with compact operators in dual systems including
all functional analytic concepts necessary for developing this theory. The
second part then illustrates the classical applications of integral equation
methods to boundary value problems for the Laplace and the heat equation
as one of the main historical sources for the development of integral equa-
tions, and also introduces Cauchy type singular integral equations. The
third part is devoted to describing the fundamental ideas for the numerical
solution of integral equations. Finally, in a fourth part, ill-posed integral
equations of the first kind and their regularization are studied in a Hilbert
space setting.
In order to make the book accessible not only to mathematicians but also
to physicists and engineers I have planned it as self-contained as possible
by requiring only a solid foundation in differential and integrals;alculus
and, for parts of the book, in complex function theory. Some background
in functional analysis will be helpful, but the basic concepts of the the-
ory of normed spaces will be briefly reviewed, and all functional analytic
tools which are relevant in the study of integral equations will be devel-
oped in the book. Of course, I expect the reader to be willing to accept
x Preface to the First Edition
the functional analytic language for describing the theory and the numer-
ical solution of integral equations. I hope that I succeeded in finding the
adequate compromise between presenting integral equations in the proper
modern framework and the danger of being too abstract.
An introduction to integral equations cannot present a complete picture
of all classical aspects of the theory and of all recent developments. In
this sense, this book intends to tell the reader what I think appropriate to
teach students in a two-semester course on integral equations. I am willing
to admit that the choice of a few of the topics might be biased by my own
preferences and that some important subjects are omitted.
I am indebted to Dipl.-Math. Peter Hahner for carefully reading the
book, for checking the solutions to the problems and for a number of sug-
gestions for valuable improvements. Thanks also go to Frau Petra Trapp
who spent some time assisting me in the preparation of the Jb.1EX version of
the text. And a particular note of thanks is given to my friend David Colton
for reading over the book and helping me with the English language. Part
of the book was written while I was on sabbatical leave at the Department
of Mathematics at the University of Delaware. I gratefully acknowledge the
hospitality.
Gottingen, September 1988 Rainer Kress

Contents
Preface to the Second Edition vii
Preface to the First Edition ix
1 Normed Spaces 1
1.1 Convergence and Continuity. 2
1.2 Completeness.. 5
1.3 Compactness . . . . 6
1.4 Scalar Products . . . 9
1.5 Best Approximation 11
Problems . . . . . . . . . 13
2 Bounded and Compact Operators 15

2.1 Bounded Operators. 15
2.2 Integral Operators 17
2.3 Neumann Series . . . 18
2.4 Compact Operators 20
Problems . . . . . . . . . 27
3 Riesz Theory 28
3.1 Riesz Theory for Compact Operators. 28
3.2 Spectral Theory for Compact Operators 34
3.3 Volterra Integral Equations 36
Problems . . . . . . . . . . . . . . . . . . . . 38
xii Contents
4 Dual Systems and Fredholm Alternative 39

4.1 Dual Systems via Bilinear Forms .. 39
4.2 Dual Systems via Sesquilinear Forms 41
4.3 The Fredholm Alternative 45
4.4 Boundary Value Problems 50
Problems . . . . . . . . . . . . 53
5 Regularization in Dual Systems 55

5.1 Regularizers . . . . 55
5.2 Normal Solvability 57
5.3 Index 62
Problems . . . . . . 66
6 Potential Theory 67
6.1 Harmonic Functions . . . . . . . . . . . 67
6.2 Boundary Value Problems: Uniqueness. 74
6.3 Surface Potentials . . . . . . . . . . . 78
6.4 Boundary Value Problems: Existence. 82
6.5 Nonsmooth Boundaries 87
Problems . . . . . . . . . . . . . 92
7 Singular Integral Equations 94

7.1 Holder Continuity . . . . 94
7.2 The Cauchy Integral Operator 97
7.3 The Riemann Problem . . . . . 105
7.4 Integral Equations with Cauchy Kernel. 107
7.5 Cauchy Integral and Logarithmic Potential 115
7.6 Logarithmic Single-Layer Potential on an Arc 120
Problems . . . . . . . . . . . . . . . . . . . . . . . 124
8 Sobolev Spaces 125

8.1 The Sobolev Space HP[O, 271"] . . . . . . . . . 125
8.2 The Sobolev Space HP(r) . . . . . . . . . . . 135
8.3 Weak Solutions to Boundary Value Problems 142
Problems . . . . . . . . . . . . . . . . . . . . . . . 151
9 The Heat Equation 152

9.1 Initial Boundary Value Problem: Uniqueness 152
9.2 Heat Potentials . . . . . . . . . . . . . . . . 155
9.3 Initial Boundary Value Problem: Existence 160
Problems . . . . . . . . . . . . . . . . . . . . . . 162
10 Operator Approximations 163

10.1 Approximations via Norm Convergence 164
10.2 Uniform Boundedness Principle . . . . . 165
Contents xiii
10.3 Collectively Compact Operators. . . . . . . 167

10.4 Approximations via Pointwise Convergence 168
10.5 Successive Approximations 170
Problems . . . . . . . . . . . . . . . . . . . . . . 175
11 Degenerate Kernel Approximation 177

11.1 Degenerate Operators and Kernels 178
11.2 Interpolation . . . . . . . . . . . . 179
11.3 Trigonometric Interpolation . . . . 182
11.4 Degenerate Kernels via Interpolation 187
11.5 Degenerate Kernels via Expansions 192
Problems . . . . . . . . . . . . . . . . . . 195
12 Quadrature Methods 197

12.1 Numerical Integration 198
12.2 Nystrom's Method . . 201
12.3 Weakly Singular Kernels . 206
12.4 Nystrom's Method in Sobolev Spaces . 213
Problems . . . . . . . . . . . . . . . . . . . 216
13 Projection Methods 218

13.1 The Projection Method . . . . . . . . . . . . . . . . . 218
13.2 Projection Methods for Equations of the Second Kind 223
13.3 The Collocation Method . . . . . . . . . . . . . . . . 225
13.4 Collocation Methods for Equations of the First Kind 232
13.5 The Galerkin Method 240
Problems . . . . . . . . . . . . . . . 245
14 Iterative Solution and Stability 247

14.1 Stability of Linear Systems 248
14.2 Two-Grid Methods . . . . . . . 251
14.3 Multigrid Methods . . . . . . . 255
14.4 Fast Matrix-Vector Multiplication 260
Problems . . . . . . . . . . . . . 264
15 Equations of the First Kind 265

15.1 Ill-Posed Problems . . . . . 265
15.2 Regularization of Ill-Posed Problems 269
15.3 Compact Self-Adjoint Operators 271
15.4 Singular Value Decomposition 277
15.5 Regularization Schemes 281
Problems . . . . . . . . . . . 289
16 Tikhonov Regularization 290

16.1 The Tikhonov Functional 290
xiv Contents
16.2 Weak Convergence 291

16.3 Quasi-Solutions . . 292
16.4 Minimum Norm Solutions 298
16.5 Classical Tikhonov Regularization 301
Problems . . . . . . . . . . . . . . . . 307
17 Regularization by Discretization 308

17.1 Projection Methods for Ill-Posed Equations 308
17.2 The Moment Method . . . . . . . . . . . 313
17.3 Hilbert Spaces with Reproducing Kernel 315
17.4 Moment Collocation 317
Problems . . . . . . . . . . . . . . . . . 318
18 Inverse Boundary Value Problems 320

18.1 Ill-Posed Equations in Potential Theory 320
18.2 An Inverse Problem in Potential Theory 328
18.3 Approximate Solution via Potentials . . 333
18.4 Differentiability with Respect to the Boundary 341
Problems . . . . . . . . . . . . . . . . . . . . . . . . 345
References 347
Index 361
1
Normed Spaces
The topic of this book is linear integral equations of which
lb K(x, y)cp(y) dy = f(x), x E [a, b],
-l
and
b
cp(x) K(x, y)cp(y) dy = f(x), x E [a, b],
are typical examples. In these equations the function cp is the unknown,

and the so-called kernel K and the right-hand side f are given functions.
The above equations are called Fredholm integral equations of the first and
second kind, respectively. We will regard them as operator equations
Acp = f
and
cp - Acp = f
of the first and second kind in appropriate normed function spaces.
The symbol A : X -t Y will mean a single-valued mapping whose domain
of definition is a set X and whose range is contained in a set Y, i.e., for
every cp E X the mapping A assigns a unique element Acp E Y. The range
A(X) is the set A(X) := {Acp : cp E X} of all image elements. We will use
the terms mapping, function, and operator synonymously.
Existence and uniqueness of a solution to an operator equation can be
equivalently expressed by the existence of the inverse operator. If for each
f E A(X) there is only one element cp E X with Acp = f, then A is said to
R. Kress, Linear Integral Equations

© Springer-Verlag New York, Inc. 1999
2 1. Normed Spaces
be injective and to have an inverse A-I: A(X) --+ X defined by A-I f := 'P.
The inverse mapping has domain A(X) and range X. It satisfies A-I A = I
on X and AA-I = I on A(X), where I denotes the identity operator
mapping each element into itself. If A(X) = Y, then the mapping is said to
be surjective. The mapping is called bijective if it is injective and surjective,
i.e., if the inverse mapping A-I: Y --+ X exists.
In the first part of the book we will present the Riesz-Fredholm theory
for compact operators which, in particular, answers the question of exis-
tence and uniqueness of solutions to integral equations of the second kind
with sufficiently smooth kernels. In order to develop the theory, we will
assume that the reader is familiar with the elementary properties of linear
spaces, normed spaces, and bounded linear operators. For convenience and
to introduce notations, in this chapter, we briefly recall a few basic concepts
from the theory of normed spaces, omitting most of the proofs. For a more
detailed study, see Aubin [10], Heuser [70], Kantorovic and Akilov [80],
Rudin [156], and Taylor [169] among others.
1.1 Convergence and Continuity

Definition 1.1 Let X be a complex (or real) linear space (vector space).
A function II . II : X --+ 1R with the properties
(N1) II'PII > 0, (positivity)
(N2) II'P II =o if and only if 'P = 0, (definiteness)
(N3) II a'P II = lalll'PlI, (homogeneity)
(N4) II'P + 1/111 < II'PII + 111/111, (triangle inequality)

for all 'P, 1/1 EX, and all a E <C (or 1R) is called a norm on X. A linear
space X equipped with a norm is called a normed space.
As a consequence of (N3) and (N4) we note the second triangle inequality
I II'PII - 111/111 I ~ II'P - 1/111· (1.1)

For two elements in a normed space II'P - 1/111 is called the distance between
'P and 1/1.
Definition 1.2 A sequence ('Pn) of elements in a normed space X is called
convergent if there exists an element 'P E X such that limn~oo II'Pn -'PII = 0,
i.e., if for every e: > 0 there exists an integer N(e:) such that II'Pn - 'PII < e:
for all n ~ N(e:). The element'P is called the limit of the sequence ('Pn),
and we write
lim 'Pn = 'P or 'Pn --+ 'P, n --+ 00.
n~oo
1.1 Convergence and Continuity 3
Note that by (N4) the limit of a convergent sequence is uniquely deter-

mined. A sequence that does not converge is called divergent.
Definition 1.3 A function A : U C X --+ Y mapping a subset U of a

normed space X into a normed space Y is called continuous at rp E U if
limn->oo Arpn = Arp for every sequence (rpn) from U with limn->oo rpn = rp.
The function A : U C X --+ Y is called continuous if it is continuous for
all rp E U.
An equivalent definition is the following: A function A : U C X --+ Y

is continuous at rp E U if for every E: > 0 there exists 0 > 0 such that
IIArp-A'l,b1l < E: for all 'l,b E U with IIrp-'l,b11 < o. Here we have used the same
symbol 11·11 for the norms on X and Y. The function A is called uniformly
continuous if for every E: > 0 there exists 0 > 0 such that IIArp - A'l,bll < E:
for all rp, 'l,b E U with Ilrp - 'l,bll < o.
Note that by (1.1) the norm is a continuous function.
In our study of integral equations the basic example of a normed space
will be the linear space C[a, b] of continuous real- or complex-valued func-
tions rp defined on an interval [a, b] c IR furnished either with the maximum
norm
IIrplloo := max Irp(x)1
xE[a,b]
or the mean square norm
Convergence of a sequence of continuous functions in the maximum norm

is equivalent to uniform convergence, and convergence in the mean square
norm is called mean square convergence. Throughout this book, unless
stated otherwise, we always assume that C[a, b] (or C(G), i.e., the space of
continuous real- or complex-valued functions on compact subsets G c IRm)
is equipped with the maximum norm.
Definition 1.4 Two norms on a linear space are called equivalent if they
have the same convergent sequences.
Theorem 1.5 Two norms 11·lla and 11·llb on a linear space X are equivalent
if and only if there exist positive numbers c and C such that
for all rp EX. The limits with respect to the two norms coincide.
Proof. Provided that the conditions are satisfied, from Ilrpn - rplla --+ 0,
n --+ 00, it follows Ilrpn - rpllb --+ 0, n --+ 00, and vice versa.
4 1. Normed Spaces
°
Conversely, let the two norms be equivalent and assume that there is
no C > such that IIcplib ~ CIICPlla for all cP EX. Then there exists a
sequence (CPn) with IICPnlla = 1 and IICPnllb :2: n 2 . Now, the sequence ('l/Jn)
with 'l/Jn := n-1cpn converges to zero with respect to II . lIa, whereas with
respect to II· lib it is divergent because of II'l/Jnllb :2: n. 0
Theorem 1.6 On a finite-dimensional linear space all norms are equiva-

lent.
Proof. In a linear space X with finite dimension m and basis !I, ... , f m
every element can be expressed in the form
m
As is easily verified,
Ilcplloo:= max lakl (1.2)
k=l, ... ,m
defines a norm on X. Let II ·11 denote any other norm on X. Then, by the
triangle inequality we have
for all cP E X, where

m
C := L Ilfkll·
°
k=l
Assume that there is no c > such that cllcplloo :::; Ilcpll for all cP E X.
Then there exists a sequence (CPn) with IICPnl1 = 1 such that IICPniioo :2: n.
Consider the sequence C¢n) with 'l/Jn := IICPnll~lCPn and write
m
Because of II'l/Jnlloo = 1 each of the sequences (akn), k = 1, ... , m, is

bounded in <C. Hence, by the Bolzano-Weierstrass theorem we can select
convergent subsequences ak,n(j) -t ak, j -t 00, for each k = 1, ... , m. This
now implies II'l/Jn(j) - 'l/Jlloo -t 0, j -t 00, where
m
and II'l/Jn(j) - 'l/JII ~ ClI'l/Jn(j) - 'l/Jlloo -t 0, j -t 00. But on the other hand we
have II'l/Jnll = 1/IICPn1100 -t 0, n -t 00. Therefore, 'l/J = 0, and consequently
II'l/Jn(j) 1100 -t 0, j -t 00, which contradicts II'l/Jn 1100 = 1 for all n. 0
1.2 Completeness 5
For an element t.p of a normed space X and a positive number r the set
B( t.pj r) := {'I/J EX: II'I/J - t.pll < r} is called the open ball of radius r and
center t.p, the set B[t.pj r] := {'I/J EX: II'I/J - t.pll ::s r} is called a closed ball.
Definition 1.1 A subset U of a normed space X is called open if for each
element t.p E U there exists r > 0 such that B (t.pj r) cU.
Obviously, open balls are open.

Definition 1.8 A subset U of a normed space X is called closed if it con-
tains all limits of convergent sequences of U.
A subset U of a normed space X is closed if and only if its complement

X \ U is open. Obviously, closed balls are closed. In particular, using the
norm (1.2), it can be seen that finite-dimensional subspaces of a normed
space are closed.
Definition 1.9 The closure U of a subset U of a normed space X is the set
of all limits of convergent sequences of U. A set U is called dense in another
set V if V C U, i.e., if each element in V is the limit of a convergent
sequence from U.
A subset U is closed if and only if it coincides with its closure. By the

Weierstrass approximation theorem (see [31]) the linear subspace P of poly-
nomials is dense in C[a, b] with respect to the maximum norm and the mean
square norm.
Definition 1.10 A subset U of a normed space X is called bounded if
there exists a positive number C such that 1It.p1l ::s C for all t.p E U.
Convergent sequences are bounded.
1.2 Completeness
Definition 1.11 A sequence (t.pn) of elements in a normed space X is
called a Cauchy sequence if for every e > 0 there exists an integer N (e) such
that lI<Pn - <Pm II < e for all n,m;::: N(e), i.e., ifliilln,m~oo IICPn - CPmll = O.
Every convergent sequence is a Cauchy sequence, whereas the converse
in general is not true. This gives rise to the following definition.
Definition 1.12 A subset U of a normed space X is called complete if
every Cauchy sequence of elements in U converges to an element in U. A
normed space X is called a Banach space if it is complete.
Complete sets are closed, and closed subsets of a complete set are com-
plete. Since the Cauchy criterion is sufficient for convergence of sequences of
6 1. Normed Spaces
complex numbers, using the norm (1.2), we observe that finite-dimensional

normed spaces are Banach spaces. The Cauchy criterion is also sufficient for
uniform convergence of a sequence of continuous functions toward a con-
tinuous limit function. Therefore, the space C[a, bj is complete with respect
to the maximum norm. As can be seen from elementary counterexamples,
C[a, bj is not complete with respect to the mean square norm.
It is very important that every normed space can be completed in the
following sense.
Theorem 1.13 For each normed space X there exists a Banach sl!..ace X
such that X is isomorphic and isometric to a dense subspace of X, i. e.,
there is a linear bijective mapping I from X into a dense subspace of X such
that III<pllx = 1I<pllx for all <P E X. The space X is uniquely determined
up to isometric isomorphisms, i. e., to any two such Banach spaces there
exists a linear bijective mapping between the two spaces leaving the norms
invariant.
For a proof of this concept of completion we refer to any introduction to
functional analysis. As a matter of fact it will not playa central role in this
book. Using Lebesgue integration theory, it can be seen that the completion
of C[a, bj with respect to the mean square norm yields the complete space
L2[a, bj of measurable and Lebesgue square-integrable functions, or to be
more precise, of equivalence classes of such functions that coincide almost
everywhere with respect to the Lebesgue measure (see [10, 156, 169)).
1.3 Compactness
Definition 1.14 A subset U of a normed space X is called compact if
every open covering of U contains a finite subcovering, i. e., if for every
family Vj, j E J, of open sets with the property
Uc UVj
jEJ
there exists a finite subfamily Vj(k), j(k) E J, k = 1, ... ,n, such that
n
U C U Vj(k)·
k=1
A subset U is called sequentially compact if every sequence of elements

from U contains a subsequence that converges to an element in U.
A subset U of a normed space is called totally bounded if for each C >0
there exists a finite number of elements <PI, ••• ,<Pn in U such that
n
Uc UB(<pj;c),
j=1
1.3 Compactness 7
i.e., each element cP E U has a distance less than c from at least one of the
elements CPI, ..• , CPn. Note that each sequentially compact set U is totally
bounded. Otherwise there would exist a positive c and a sequence (CPn)
in U with the property IiCPn - CPmll ~ c for all n i= m. This would imply
that the sequence (CPn) does not contain a convergent subsequence, which
contradicts the sequential compactness of U. For each totally bounded set
U, letting c = 11m, m = 1,2, ... , and collecting the corresponding finite
systems of elements CPI, •.. ,CPn depending on m, we obtain a sequence that
is dense in U.
Theorem 1.15 A subset of a normed space is compact if and only if it is

sequentially compact.
Proof. Let U be compact and assume that it is not sequentially compact.

Then there exists a sequence (CPn) in U that does not contain a convergent
subsequence with limit in U. Consequently, for each cP E U there exists
an open ball B( cP; r) with center cP and radius r( cp) containing at most a
finite number of the elements of the sequence (CPn). The set of these balls
clearly is an open covering of U, and since U is compact, it follows that U
contains only a finite number of the elements of the sequence (CPn). This is
a contradiction.
Conversely, let U be sequentially compact and let Vj, j E J, be an open
covering of U. First, we show that there exists a positive number c such
that for every cP E U the ball B (cp; c) is contained in at least one of the sets
Vj: Otherwise there would exist a sequence (CPn) in U such that the ball
B(CPn; lin) is not contained in one of the Vj. Since U is sequentially com-
pact, this sequence (CPn) contains a convergent subsequence (CPn(k») with
limit cP E U. The element cP is contained inJlome Vj, and since Vj is open,
using the triangle inequality, we see that B(CPn(k); 1In(k)) C Vj for suffi-
ciently large k. This is a contradiction. Now, since the sequentially compact
set U is totally bounded, there exists a finite number of elements CPI, .•. , CPn
in U such that the balls B(CPk;c), k = 1,; .. ,n, cover U. But for each of
these balls there exists a set Vj(k), j(k) E J, such that B(cpk;c) C Vj(k)'
Hence, the finite family Vj(k), k = 1, ... , n, covers U. 0
In particular, from Theorem 1.15 we observe that compact sets are

bounded, closed, and complete.
Definition 1.16 A subset of a normed space is called relatively compact

if its closure is compact.
As a consequence of Theorem 1.15, a set U is relatively compact if and

only if each sequence of elements from U contains a convergent subse-
quence. Hence, analogous to compact sets, relatively compact sets are to-
tally bounded.
8 1. Normed Spaces
Theorem 1.17 A bounded and finite-dimensional subset of a normed space

is relatively compact.
Proof. This follows from the Bolzano-Weierstrass theorem using the norm
(1.2). 0
For the compactness in C(G), the space of continuous real- or complex-

valued functions defined on a compact set G c lR,m, furnished with the
maximum norm
1I'Plioo := max
xEG
1'P(x)l,
we have the following criterion.
Theorem 1.18 (Arzela-Ascoli) A set U c C(G) is relatively compact if

and only if it is bounded and equicontinuous, i. e., if there exists a constant
C such that
1'P(x)1 ~ C
for all x E G and all 'P E U, and for every E: >
that
°
there exists 8 > Osuch
1'P(x) - 'P(y) I < E:

for all x, y E G with Ix - yl < 8 and all 'P E U. {Here and henceforth
by Izi := (zr + ... + z;;'y/2 we denote the Euclidean norm of a vector
z = (Zl, ... ,zm) E lR,m.)
Proof. Let U be bounded and equicontinuous and let ('Pn) be a sequence in

U. We choose a sequence (Xi) from the compact and, consequently, totally
bounded set G that is dense in G. Since the sequence ('Pn(Xi)) is bounded
in <D for each Xi, by the standard diagonalization procedure we can choose
a subsequence ('Pn(k)) such that ('Pn(k)(Xi» converges in <D as k -+ 00 for
each Xi. More precisely, since ('Pn(XI)) is bounded, we can choose a subse-
quence ('Pnl(k)) such that ('Pndk)(XI)) converges as k -+ 00. The sequence
('Pnl(k)(X2» again is bounded and we can choose a subsequence ('Pn2(k) of
('Pnl(k) such that ('Pn2(k)(X2» converges as k -+ 00. Repeating this pro-
cess of selecting subsequences, we arrive at a double array ('Pni(k) such
that each row ('Pni(k) is a subsequence of the previous row ('Pni_dk) and
each sequence ('Pni(k)(Xi» converges as k -+ 00. For the diagonal sequence
'Pn(k) := 'Pnk(k) we have that ('Pn(k) (Xi» converges as k -+ 00 for all Xi·
Since the set (Xi), i = 1,2, ... , is dense in G, given E: > 0, the balls
B(Xi; 8), i = 1,2, ... , cover G. Since G is compact we can choose i E IN
such that each point X E G has a distance less than 8 from at least one
element Xj of the set Xl, ... , Xi . Next choose N(E:) E IN such that
1.4 Scalar Products 9
for all k, I 2:: N(e) and all j = 1, ... , i. From the equicontinuity we obtain
+ICPn(I)(Xj) - CPn(I)(x)1 < 3e

for all k, I 2:: N(e) and all x E G. This establishes the uniform convergence
of the subsequence (CPn(k»), i.e., convergence in the maximum norm. Hence
U is relatively compact.
Conversely, let U be relatively compact. Then U is totally bounded, i.e.,
given e > 0 there exist functions CPl, ... ,CPi E U such that
. min .lIcp - cpjlloo < :.

)=1, ... " 3
for all cp E U. Since each of the I.{JI, ••• ,CPi is uniformly continuous on the
compact set G, there exists 6 > 0 such that
e
Icpj(x) - cpj(y)1 < '3
for all X,y E G with Ix - yl < 6 and all j = 1, ... ,i. Then for all cP E U,
choosing jo such that
we obtain
for all x, y E G with Ix - yl < 6. Therefore U is equicontinuous. Finally, U

is bounded, since relatively compact sets are bounded. 0
1.4 Scalar Products

Definition 1.19 Let X be a complex (or real) linear space. Then a func-
tion (.,.) : X X X -t <C (or R) with the properties
(HI) (cp,cp) ~ 0, (positivity)
(H2) (cP, cp) = 0 if and only if cP = 0, (definiteness)
(H3) (cp,1/J) = (1/J, cp), (symmetry)
(H4) (acp+/31/J,x) = a(cp,x)+/3(1/J,X), (linearity)

for all cp,1/J, x E X, and a,/3 E <C (orR) is called a scalar product, or an
inner product, on X. (By the bar we denote the complex conjugate.)
10 1. Normed Spaces
As a consequence of (H3) and (H4) we note the antilinearity
(cp, a'ljJ + J3X) = a( cp, 'ljJ) + i3( cp, X)· (1.3)

Theorem 1.20 A scalar product satisfies the Cauchy-Schwarz inequality
l(cp,'ljJ)1 2 :s (cp,cp)('ljJ,'ljJ)
for all cp, 'ljJ EX, with equality if and only if cp and'ljJ are linearly dependent.
Proof. The inequality is trivial for cp = O. For cp "# 0 it follows from
(acp + J3'ljJ,acp + J3'ljJ) = laI 2 (cp, cp) + 2Re{ai3(cp,'ljJ)} + 1J312('ljJ,'ljJ)

= (cp,cp)('ljJ,'ljJ) -1(cp,'ljJ)1 2 ,
where we have set a = _(cp,cp)-1/2(cp,'ljJ) and J3 = (cp,cp)1/2. Since (.,.) is

positive definite, this expression is nonnegative, and it is equal to zero if
and only if acp + J3'ljJ = O. In the latter case cp and 'ljJ are linearly dependent
because J3 "# o. 0
Theorem 1.21 A scalar product (".) on a linear space X defines a norm

by
for all cp EX. If X is complete with respect to this norm it is called a

Hilbert space; otherwise it is called a pre-Hilbert space.
Proof. We leave it as an exercise to verify the norm axioms. The triangle

inequality follows from the Cauchy-Schwarz inequality. 0
Note that L2[a, b] is a Hilbert space with the scalar product given by
(cp,'ljJ):= lb cp(x)'ljJ(x)dx.
Definition 1.22 Two elements cp and'ljJ of a pre-Hilbert space X are called

orthogonal if
(cp,'ljJ) =0.
Two subsets U and V of X are called orthogonal if each pair of elements
cpE U and 'ljJ E V are orthogonal. For two orthogonal elements or subsets
we write cp -1 'ljJ and U -1 V, respectively. A subset U of X is called an
orthogonal system if (cp, 'ljJ) = 0 for all cp, 'ljJ E U with cp"# 'ljJ. An orthogonal
system U is called an orthonormal system if Ilcpll = 1 for all cp E U. The
set
Ul.. := {'ljJ EX: 'ljJ -1 U}
is called the orthogonal complement of the subset U.
1.5 Best Approximation 11
1.5 Best Approximation

Definition 1.23 Let U c X be a subset of a normed space X and let
<P EX. An element v' E U is called a best approximation to <P with respect
to U if
Ii<p - vii = inf Ii<p - uli,
uEU
i.e., v E U has smallest distance from <po
Theorem 1.24 Let U be a finite-dimensional subspace of a normed space

X. Then for every element of X there exists a best approximation with
respect to U.
Proof. Let <P E X and choose a minimizing sequence (un) for <p, i.e., Un E U
satisfies
1I<p - unll-+ d:= inf 1I<p - ull, n -+ 00.
uEU
Because of II Un II ::; 1I<p - unll + 1I<p1l, the sequence (un) is bounded. Since U
has finite dimension, it is closed and by Theorem 1.17 the sequence (un)
contains a convergent subsequence (Un(k» with limit v E U. Then
1I<p - vII = k-+oo

lim 1I<p - Un(k) II = d
completes the proof. o

Theorem 1.25 Let U be a linear subspace of a pre-Hilbert space X. An
element v is a best approximation to <P E X with respect to U if and only if
(<p-v,U)=O
for all U E U, i. e., if and only if <P - v ..1 U. To each <p E X there exists at
most one best approximation with respect to U.
Proof. This follows from the equality
which is valid for all v, u E U and all Q E JR. o
Theorem 1.26 Let U be a complete linear subspace of a pre-Hilbert space

X. Then to every element of X there exists a unique best approximation
with respect to U.
Proof. Let <p E X and choose a sequence (un) with
(1.4)
12 1. Normed Spaces
where d:= infuEu Ilrp - ull. Then

II(rp - un) + (rp - U m)1I2 + lIun - u m l1 2 = 211rp - un 112 + 211rp - u m l1 2
2 2
<4d
-
2
+-+-
n m
for all n, mE IN, and since ~ (un + um) E U, it follows that
Hence, (un) is a Cauchy sequence, and because U is complete, there exists

an element v E U such that Un -+ v, n -+ 00. Passing to the limit n -+ 00
in (1.4) shows that v is a best approximation of rp with respect to U. 0
We wish to extend Theorem 1.25 to the case of convex subsets of a pre-

Hilbert space. A subset U of a linear space X is called convex if
..\rpl + (1 - ..\)rp2 E U
for all rpl, rp2 E U, and all ..\ E (0,1).
Theorem 1.27 Let U be a convex subset of a pre-Hilbert space X. An
element v E U is a best approximation to rp E X with respect to U if and
only if
Re(rp - v, U - v) ~ 0
for all U E U. To each rp E X there exists at most one best approximation
with respect to U.
Proof. This follows from the equality
Ilrp - [(1 - ..\)v + ..\ulll 2 = Ilrp - vll 2 - 2,,\ Re(rp - v, U - v) + ).211u - v1l 2,
which is valid for all u, v E U, and all ). E (0,1). o
It is left to the reader to carry Theorem 1.26 over from the case of a
linear subspace to the case of a convex subset.
For a subset U of a linear space X we denote the set spanned by all linear
combinations of elements of U by span U.
Theorem 1.28 Let {un : n E IN} be an orthonormal system in a pre-
Hilbert space X. Then the following properties are equivalent:
(a) span{u n : n E IN} is dense in X.
(b) Each rp E X can be expanded in a Fourier series
00
rp= ~(rp,Un)Un'
n=l
Problems 13
(c) For each cp E X we have Parseval's equality

00
IIcpll2 = 2: l(cp,unW·
n=l
The properties (a) -( c) imply that

(d) cp = 0 is the only element in X with (cp, un) = 0 for all n E IN,
and (a), (b), (c), and (d) are equivalent if X is a Hilbert space. An or-
thonormal system with the property (a) is called complete.
Proof. What has to be understood by a series is explained in Problem 1.5.
(a) =} (b): By Theorems 1.25 and 1.26, the partial sum
n
CPn = 2:(cp,Uk)Uk
k=l
is the best approximation to cp with respect to span{ Ul, •.. ,Un}. Since
span {Un : n E IN} is dense in X, the sequence of the best approximations
converges, i.e., CPn -+ cp, n -+ 00.
(b) =} (c): This follows by taking the scalar product of the Fourier series
with cp.
(c) =} (a): This follows from
(c) =} (d): This is trivial.

(d) =} (a): In the case of a Hilbert space X, we set U := span{un : n E IN}
and assume that X i=- U. Then there exists cp E X with cp (j. U. Since X is
complete, U is also complete. Therefore, by Theorems 1.25 and 1.26 the best
approximation v to cp with respect to U exists and satisfies (v - cp, un) = 0
for all n E IN. Hence, we have the contradiction cp = v E U. 0
Problems
1.1 Show that finite-dimensional subspaces of normed spaces are closed and
complete.
1.2 A norm II . lIa on a linear space X is called stronger than a norm II . lib if
every sequence converging with respect to the norm II . lIa also converges with
respect to the norm II . lib. The same fact is also expressed by saying that the
norm II . lib is weaker than the norm II . 1Ja. Show that II . lIa is stronger than
14 1. Normed Spaces
II· lib if and only if there exists a positive number C such that II'Pbli :::; CII'Plia for
all 'P E X. Show that on C[a, b] the maximum norm is stronger than the mean
square norm. Construct a counterexample to demonstrate that these two norms
are not equivalent.
1.3 Show that a continuous real-valued function on a compact subset of a

normed space assumes its supremum and its infimum and that it is uniformly
continuous.
1.4 Construct a counterexample to demonstrate that C[a, b] is not complete

with respect to the mean square norm.
1.5 Let ('Pn) be a sequence of elements of a normed space X. The series

00
is called convergent if the sequence (Sn) of partial sums
Sn:= L'Pk
k=l
converges. The limit S = lim n .... oo Sn is called the sum of the series. Show that
in a Banach space X the convergence of the series
00
is a sufficient condition for the convergence of the series 2:;;0=1 'Pk and that
Hint: Show that (Sn) is a Cauchy sequence.

2
Bounded and Compact Operators
In this chapter we briefly explain the basic properties of bounded linear

operators and then introduce the concept of compact operators that is of
fundamental importance in the study of integral equations.
2.1 Bounded Operators

Definition 2.1 An operator A: X -+ Y mapping a linear space X into a
linear space Y is called linear if
A( aip + (3'lj;) = aAip + (3A'lj;

for all ip, 'lj; E X and all a, (3 E <C (or JR).
Recall Definition 1.3 for the continuity of an operator A : X -+ Y map-
ping a normed space X into a normed space Y.
Theorem 2.2 A linear operator is continuous if it is continuous at one

element.
Proof. Let A : X -+ Y be continuous at ipo EX. Then for every ip E X
and every sequence (ipn) with ipn -+ ip, n -+ 00, we have
since ipn - ip + ipo -+ ipo, n -+ 00. D

16 2. Bounded and Compact Operators
Definition 2.3 A linear operator A: X -+ Y from a normed space X into

a normed space Y is called bounded if there exists a positive number C
such that
for all <P EX. Each number C for which this inequality holds is called a
bound for the operator A. (Again we use the same symbol 11·11 for the norms
on X and Y.)
With the aid of the linearity of A it is easy to see that A is bounded if
and only if
IIAII:= sup IIA<p1I < 00. (2.1)
1I'P1I9
Hence a linear operator is bounded if and only if it maps bounded sets in X
into bounded sets in Y. The number IIAII is the smallest bound for A and is
called the norm of A. Every linear combination of bounded linear operators
again is a bounded linear operator, i.e., the set L(X, Y) of bounded linear
operators from X into Y forms a linear space.
Theorem 2.4 The linear space L(X, Y) of bounded linear operators from
a normed space X into a normed space Y is a normed space with the norm
{2.1}. IfY is a Banach space then L(X, Y) also is a Banach space.
Proof. The proof consists in carrying over the norm axioms and the com-
pleteness from Y onto L(X,Y). For the second part, let (An) be a Cauchy
sequence in L(X, Y), i.e., IIAm - Anll -+ 0, m, n -+ 00. Then for each
<P E X the sequence (An<P) is a Cauchy sequence in Y and converges, since
Y is complete. Then A<p := limn -+ oo An<P defines a bounded linear operator
A : X -+ Y, which is the limit of the sequence (An), i.e., IIAn - All -+ 0,
n -+ 00. 0
Note that for sequences of operators we have to distinguish between

convergence with respect to the operator norm (2.1) and pointwise conver-
gence. A sequence (An) of operators An: X -+ Y from a normed space X
into a normed space Y is called pointwise convergent if for every <P E X
the sequence (An<P) converges in Y. Norm convergence implies pointwise
convergence, but the converse is not true.
Theorem 2.5 A linear operator is continuous if and only if it is bounded.
Proof. Let A : X -+ Y be bounded and let (<Pn) be a sequence in X with
<Pn -+ 0, n -+ 00. Then from IIA<Pnll :::; CII<Pnll it follows that A<Pn -+ 0,
n -+ 00. Thus, A is continuous at such that
IIA<p1I :::; CII<p1l for all <P E X. Then there exists a sequence (<Pn) in X with
lI<Pnll = 1 and IIA<Pnll ~ n. Consider the sequence 'l/Jn := IIA<Pnll- 1 <Pn. Then
2.2 Integral Operators 17
1/Jn -+ 0, n -+00, and since A is continuous, A1/Jn -+ A(O) = 0, n -+ 00.

This is a contradiction to IIA1/Jnll = 1 for all n. Hence, A is bounded. 0
Theorem 2.6 Each linear operator A : X -+ Y from a finite-dimensional

normed space X into a normed space Y is bounded.
Proof. Let il, ... , fm be a basis of X. Then, writing
and using the norm (1.2), we have

m
IIA'PII ~ 1: IIAikllll'Plioo
k=l
for all 'P E X. Now the statement is a consequence of Theorem 1.6. 0
Remark 2.7 Let X, Y, and Z be normed spaces and let A : X -+ Y and

B : Y -+ Z be bounded linear operators. Then the product BA : X -+ Z,
defined by (BA)'P := B(A'P) for all 'P E X, is a bounded linear operator
with IIBAII ~ IIAIIIIBIi.
Proof. This follows from II(BA)'PII = IIB(A'P)II ~ IIBII II All II'PII·
2.2 Integral Operators

Now we want to introduce integral operators. We assume that the reader is
familiar with the Riemann integral for real- and complex-valued functions
in Rm. A set GeRm is called Jordan measurable if the characteristic
function Xa, defined by xa(x) = 1 for x E G and Xa(x) = 0 for x fI. G,
is Riemann integrable. The Jordan measure IGI is the integral of Xa. For
each Jordan measurable set G the closure 0 and the boundary aG also
are Jordan measurable with 101 = IGI and laGI = O. If G is compact and
Jordan measurable, then each function f E C(G) is Riemann integrable. In
addition, we assume that G is the closure of an open set or, equivalently,
that G coincides with the closure of its interior. This ensures that each
nonnegative function f E C(G) satisfying fa f(x) dx = 0 must vanish
identically, i.e., f(x) = 0 for all x E G.
Theorem 2.8 Let GeRm be a nonempty compact and Jordan measur-
able set that coincides with the closure of its interior. Let K : G x G -+ ~ be
a continuous function. Then the linear operator A: C(G) -+ C(G) defined
by
(A'P)(x):= fa K(x, Y)'P(Y) dy, x E G, (2.2)
is called an integral operator with continuous kernel K. It is a bounded

operator with
IIAlloo = max
xEC lcr IK(x, y)1 dy.
Proof. Clearly (2.2) defines a linear operator A : C(G) -+ C(G). For each
rp E C(G) with Ilrplloo ::; 1 we have
I(Arp)(x)1 ::; kIK(x, y)1 dy, x E G,
and thus
IIAlloo =
sup IIArplloo::; max
11'1'11009 xEC lc
IK(x, y)1 dy. r
Since K is continuous, there exists Xo E G such that
lcr IK(xo,y)ldy = max r IK(x,y)ldy.

xEC lc
For c > 0 choose 'l/J E C(G) by setting

K(xo,y)
'l/J(y):= IK(xo,y)1 +c' y E G.
Then 11'l/Jlloo ::; 1 and
IIA'l/Jlloo:::: I(A'l/J) (xo)1 = r IK(xo,y)1 dy:::: lcr IK(xo,y)1
lc IK(xo,y)l+c
2
_c
IK(xo,y)l+c
2 2
dy
= fa IK(xo, y)1 dy - clGI·

Hence
IIAlloo = sup IIArplloo::::

11'1'11009
IIA'l/Jlloo:::: r IK(xo,y)1 dy - clGI,
lc
and because this holds for all c > 0, we have
IIAlloo::::
lcr IK(xo,y)1 dy = max r IK(x,y)1 dy.
xEC lc
This concludes the proof. o
2.3 Neumann Series

For operator equations of the second kind
rp-Arp=J
existence and uniqueness of a solution can be established by the Neumann
series provided that A is a contraction, i.e., IIAII < 1.
2.3 Neumann Series 19
Theorem 2.9 Let A : X -+ X be a bounded linear operator on a Banach

space X with IIAII < 1 and let I : X -+ X denote the identity operator.
Then I -A has a bounded inverse on X that is given by the Neumann series
00
(I - A)-l = L:Ak
k=O
and satisfies
II(I - A) II
-1 1
~ 1_ IIAII
(The iterated operators Ak are defined recursively by AO .- I and
Ak :=AA k - 1 for k E IN.)
Proof. Since IIAII < 1, in view of Remark 2.7, we have absolute convergence
in the Banach space L(X, X), and therefore, by Problem 1.5, the Neumann
series converges in the operator norm and defines a bounded linear operator
00
with IISII ~ (1 - IIAID- 1 . S is the inverse of I - A, as can be seen from

n
(I-A)S=(I-A) lim L:Ak= lim (I_An+1)=I
n-+oo n-+oo
k=O
and
n
S(I - A) = lim L:Ak(I - A) = lim (I - An+!)
n-too n-too
= I,
k=O
o
Obviously, the partial sums
of the Neumann series satisfy ipn+1 = Aipn + f for n 2': O. Hence, the
Neumann series is related to successive approximations by the following
theorem.
Theorem 2.10 Under the assumptions of Theorem 2.9 for each f E X

the successive approximations
<Pn+l := A<Pn + f, n = 0,1,2, ... , (2.3)
with arbitrary <Po E X converge to the unique solution <P of <P - A<p = f·
Proof. By induction it can be seen that

n-l
<Pn=An<po+ LAkf, n=1,2, ... ,
k=O
whence
follows. o
We explicitly want to state this result for integral equations.

Corollary 2.11 Let K be a continuous kernel satisfying
maxlIK(x,Y)ldY < l.
xEG G
Then for each f E C(G) the integral equation of the second kind
<P(X)-!aK(X,Y)<P(Y)dY=f(x), XEG,
has a unique solution <P E C(G). The successive approximations
<Pn+l(X):= fa K(x,Y)<Pn(y)dY+f(x), n=0,1,2, ... ,
with arbitrary <Po E C( G) converge uniformly to this solution.

The method of successive approximations has two drawbacks. First, the
Neumann series ensures existence of solutions to integral equations of the
second kind only for sufficiently small kernels, and second, in general, it
cannot be summed in closed form. Later in the book we will have more to
say about using successive approximations to obtain approximate solutions
(see Section 10.5).
2.4 Compact Operators

To provide the tools for establishing the existence of solutions to a wider
class of integral equations we now turn to the introduction and investigation
of compact operators.
Definition 2.12 A linear operator A : X ---+ Y from a normed space X

into a normed space Y is called compact if it maps each bounded set in X
into a relatively compact set in Y.
Since by Definition 1.16 and Theorem 1.15 a subset U of a normed space

Y is relatively compact if each sequence in U contains a subsequence that
converges in Y, we have the following equivalent condition for an operator
to be compact.
Theorem 2.13 A linear operator A: X ---+ Y is compact if and only if for

each bounded sequence (tpn) in X the sequence (Atpn) contains a convergent
subsequence in Y.
We proceed by establishing the basic properties of compact operators.
Theorem 2.14 Compact linear operators are bounded.
Proof. This is obvious, since relatively compact sets are bounded (see The-
orem 1.15). 0
Theorem 2.15 Linear combinations of compact linear operators are

compact.
Proof. Let A, B : X ---+ Y be compact linear operators and let a, f3 E <C.

Then for each bounded sequence (tpn) in X, since A and B are compact,
we can select a subsequence (tpn(k)) such that both sequences (Atpn(k))
and (Btpn(k)) converge. Hence (aA + f3B)tpn(k) converges, and therefore
aA + f3B is compact. 0
Theorem 2.16 Let X, Y, and Z be normed spaces and let A; X ---+ Y and
B ; Y ---+ Z be bounded linear operators. Then the product BA ; X ---+ Z is
compact if one of the two operators A or B is compact.
Proof. Let (tpn) be a bounded sequence in X. If A is compact, then there

exists a subsequence (tpn(k)) such that Atpn(k) ---+ 'IjJ E Y, k ---+ 00. Since
B is bounded and therefore continuous, we have B(Atpn(k)) ---+ B'IjJ E Z,
k ---+ 00. Hence BA is compact. If A is bounded and B is compact, the
sequence (Atpn) is bounded in Y, since bounded operators map bounded
sets into bounded sets. Therefore, there exists a subsequence (tpn(k)) such
that (BA)tpn(k) = B(Atpn(k)) ---+ X E Z, k ---+ 00. Hence, again BA is
compact. 0
Theorem 2.17 Let X be a normed space and Y be a Banach space. Let

the sequence An ; X ---+ Y of compact linear operators be norm convergent
to a linear operator A : X ---+ Y, i.e., IIAn - All ---+ 0, n ---+ 00. Then A is
compact.
Proof. Let ('Pm) be a bounded sequence in X, i.e., II 'Pm II ~ C for all m E IN

and some C > O. Because the An are compact, by the standard diagonal-
ization procedure (see the proof of Theorem 1.18), we can choose a subse-
quence ('Pm(k» such that (An'Pm(k» converges for every fixed n as k ~ 00.
Given e > 0, since IIAn - All ~ 0, n ~ 00, there exists no E IN such that
llAno - All < e/3C. Because (Ano'Pm(k» converges, there exists N(e) E IN
such that
e
IIAno'Pm(k) - Ano'Pm(l) II < "3
for all k, I ~ N(e). But then we have
+IIAno'Pm(l) - A'Pm(l) II < e.

Thus (A'Pm(k» is a Cauchy sequence, and therefore it is convergent in the
Banach space Y. 0
Theorem 2.18 Let A : X ~ Y be a bounded linear operator with finite-

dimensional range A(X). Then A is compact.
Proof, Let U c X be bounded. Then the bounded operator A maps U into
the bounded set A(U) contained in the finite-dimensional space A(X). By
the Bolzano-Weierstrass Theorem 1.17 the set A(U) is relatively compact.
Therefore A is compact. 0
Lemma 2.19 (Riesz) Let X be a normed space, U C X a closed subspace

with U =f. X, and a E (0,1). Then there exists an element't/J E X with
1I't/J11 = 1 such that 11't/J - 'PII ~ a for all 'P E U,
Proof. Because U =f. X, there exists an element f E X with f 1. U, and
because U is closed, we have
f3 := inf
cpEU
Ilf - 'PII > O.
We can choose g E U such that
f3 ~ Ilf - gil ~ !ia .

Now we define
f-g
't/J:= IIf-gll'
Then 11't/J1l = 1, and for all 'P E U we have
1 f3
I/'t/J-cpl/= IIf_glll/f-{g+l/f-gl/cp}lI~ Ilf-gil ~a,
since g + IIf - gil 'P E U. o
Theorem 2.20 The identity operator I : X -+ X is compact if and only

if X has finite dimension.
Proof. Assume that I is compact and X is not finite-dimensional. Choose

an arbitrary 'PI E X with II'PIII = 1. Then UI := span{ 'PI} is a finite-
dimensional and therefore closed subspace of X. By Lemma 2.19 there
exists 'P2 E X with 11'P211 = 1 and 11'P2 - 'PIli 2: 1/2. Now consider U2 :=
span{'PI' 'P2}. Again by Lemma 2.19 there exists 'P3 E X with 11'P311 = 1
and 11'P3 - 'PIli 2: 1/2, 11'P3 - 'P211 2: 1/2. Repeating this procedure, we ob-
tain a sequence ('Pn) with the properties lIr.pn II = 1 and II'Pn - r.pm II 2: 1/2,
n =1= m. This implies that the bounded sequence (r.pn) does not contain a
convergent subsequence. Hence we have a contradiction to the compactness
of I. Therefore, if the identity operator is compact, X has finite dimension.
The converse statement is an immediate consequence of Theorem 2.18. 0
This theorem, in particular, implies that the converse of Theorem 2.14 is

false. It also justifies the distinction between operator equations of the first
and second kind, because obviously for a compact operator A the operators
A and I-A have different properties. Note that by Theorems 2.16 and 2.20
the compact operator A cannot have a bounded inverse unless its range has
finite dimension.
Theorem 2.21 The integral operator with continuous kernel is a compact

operator on C (G).
Proof. Let U c C(G) be bounded, i.e., 11r.plloo S C for all 'P E U and some
C > O. Then
I(Ar.p)(x)1 S CIGI max IK(x, y)1
x,yEG
for all x E G and all r.p E U, i.e., A(U) is bounded. Since K is uniformly
continuous on the compact set G X G, for every e > 0 there exists 8 > 0
such that
e
IK(x,z) - K(y,z)1 < ClGI
for all x, y, z E G with Ix - yl < 8. Then
I(Ar.p)(x) - (A'P)(y)1 < e
for all x, y E G with Ix - yl < 8 and all r.p E U, i.e., A(U) is equicontinuous.
Hence A is compact by the Arzela-Ascoli Theorem 1.18. 0
We wish to mention that the compactness of the integral operator with

continuous kernel also can be established by finite-dimensional approxi-
mations using Theorems 2.17 and 2.18 in the Banach space C(G). In this
context note that the proofs of Theorems 2.17 and 1.18 are similar in struc-
ture. The finite-dimensional operators can be obtained by approximating
either the continuous kernel by polynomials through the Weierstrass ap-

proximation theorem or the integral through a finite sum (see [24]).
Now we extend our investigation to integral operators with a weakly
singular kernel, i.e., the kernel K is defined and continuous for all
x, y E G c lRm , x =I- y, and there exist positive constants M and a E (0, m]
such that
IK(x, y)1 ::; Mix - ylo-m, x, y E G, x =I- y. (2.4)
Theorem 2.22 The integral operator with a weakly singular kernel is a

compact operator on C(G).
Proof. The integral in (2.2) defining the operator A exists as an improper

integral, since
and
[ Ix - ylo-m dy ::; Wm t d
po-m pm-l dp = Wm dO,
la io a
where we have introduced polar coordinates with origin at x, d is the di-
ameter of G, and Wm denotes the surface area of the unit sphere in IRm.
Now we choose a piecewise linear continuous function h : [0,00) -t IR by
setting
°: ;
1
0, t ::; 1/2,
h(t) := 2t - 1, 1/2 ::; t ::; 1,
1, 1 ::; t < 00,

and for n E IN we define continuous kernels Kn : G x G -t CO by
h(nlx - yI)K(x, y), x =I- y,

Kn(x,y) := {
0, x=y.
The corresponding integral operators An : C(G) -t C(G) are compact by

Theorem 2.21. We have the estimate
I(Acp)(x) - (Ancp)(x)1 = IlanB[x;l/n)

[ {l- h(nlx - y)}K(x, y)cp(y) dyl
[lin
::; Mllcplloowm10 po-m pm-l dp
= Mllcplioo : ; , x E G.
From this we observe that AnCP -+ Acp, n -+ 00, uniformly, and therefore
Acp E e(O). Furthermore it follows that
IIA - Anlloo :S M Wm -+ 0, n -+ 00,

ana
and thus A is compact by Theorem 2.17. 0
Finally, we want to expand the analysis to integral operators defined on

surfaces in lRm. Having in mind applications to boundary value problems,
we will confine our attention to surfaces that are boundaries of smooth
domains in lRm. A bounded open domain D C lRm with boundary aD is
en,
said to be of class n E IN, if the closure jj admits a finite open covering
p
DcUVq
q=l
such that for each ofthose Vq that intersect with the boundary aD we have
the properties: The intersection Vq n D can be mapped bijectively onto the
half-ball H := {x E lRm : Ixl < 1, Xm 2: O} in lRm, this mapping and its
inverse are n times continuously differentiable, and the intersection VqnaD
is mapped onto the disk H n {x E lR1n : Xm = O}.
In particular, this implies that the boundary aD can be represented
locally by a parametric representation
mapping an open parameter domain U C lR1n - 1 bijectively onto a surface

patch S of aD with the property that the vectors
ax i = 1, ... ,m -1,
are linearly independent at each point x of S. Such a parameterization

we call a regular parametric representation. The whole boundary aD is
obtained by matching a finite number of such surface patches.
On occasion, we will express the property of a domain D to be of class
en also by saying that its boundary aD is of class en.
The vectors aX/aUi, i = 1, ... , m - 1, span the tangent plane to the
surface at the point x. The unit normal v is the unit vector orthogonal to
the tangent plane. It is uniquely determined up to two opposite directions.
The surface element at the point x is given by
where g is the determinant of the positive definite matrix with entries
gij
ax ax
:= ~ . ~, i, j = 1, ... ,m - l.
UUi uUj
In this book, for two vectors a = (ab ... , am) and b = (bb"" bm ) in JR,m
(or <em) we denote by a· b = alb l + ... + ambm the dot product.
Assume that aD is the boundary of a bounded open domain of class C l .
In the Banach space C(aD) of real- or complex-valued continuous functions
defined on the surface aD and equipped with the maximum norm
IIcplloo := xEaD
max Icp(x)l,
we consider the integral operator A : C(aD) -+ C(aD) defined by
(Acp)(x):= r
laD
K(x,y)cp(y)ds(y), xEaD, (2.5)
where K is a continuous or weakly singular kernel. According to the di-

mension of the surface aD, a kernel K is said to be weakly singular if it is
defined and continuous for all x, y E aD, x =f:. y, and there exist positive
constants M and a E (0, m - 1] such that
IK(x, y)1 :::; Mix - ylo:-m+1, x, y E aD, x =f:. y. (2.6)
Analogously to Theorems 2.21 and 2.22 we can prove the following theorem.
Theorem 2.23 The integral operator with continuous or weakly singular
kernel is a compact operator on C(aD) if aD is of class c 1 •
Proof. For continuous kernels the proof of Theorem 2.21 essentially remains
unaltered. For weakly singular kernels the only major difference in the proof
compared with the proof of Theorem 2.22 arises in the verification of the
existence of the integral in (2.5). Since the surface aD is of class C 1 , the
normal vector v is continuous on aD. Therefore, we can choose R E (0,1]
such that
1
v(x) . v(y) 2:: '2 (2.7)
for all x, y E aD with Ix - yl :::; R. Furthermore, we can assume that R

is small enough such that the set S[x; R] := {y E aD : Iy - xl :::; R} is
connected for each x E aD. Then the condition (2.7) implies that S[x; R]
can be projected bijectively onto the tangent plane to aD at the point x.
By using polar coordinates in the tangent plane with origin in x, we now
can estimate
I1r
S[x;R]
K(x,y)cp(y) ds(y) I : :; Mllcplloo1r
S[x;R]
Ix - ylo:-m+l ds(y)
Problems 27
Here we have used the facts that Ix - yl 2: p, that the surface element
pm- 2 dpdw
ds(y) = v(x). v(y) ,
expressed in polar coordinates on the tangent plane, can be estimated with

the aid of (2.7) by ds(y) ::; 2pm- 2 dpdw, and that the projection of S[x; R]
onto the tangent plane is contained in the interior of the disk of radius R
and center x. Furthermore, we have
I{
JaD\S[x;Rj
K(x,y)<p(y) ds(y) I::; MIi<plioo (
JaD\S[x;Rj
R o - m+1ds(y)
Hence, for all x E aD the integral (2.5) exists as an improper integral. For
the compactness of A, we now can adopt the proof of Theorem 2.22.
Problems
2.1 Let A: X --t Y be a £ounde2 linear operator from a normed space X into
a normed space Y and let X and Y be the completions of X and Y, 1Zsp'Ztivel~
Then there exists a uniquely determined bounded linear operator A : X --t Y
such that Acp = Acp for all cp E X. Furthermore, IIAII = IIAII. The operator A is
called the continuous extension of A. (In the sense of Theorem 1.13 the space X
is interpreted as a dense subspace of its completion X.)
Hint: For cp E X define Acp = lim n -+oo Acpn, where (CPn) is a sequence from X
with cpn --t cp, n --t 00.
2.2 Show that Theorem 2.10 remains valid for operators satisfying IIAkll < 1
for some k E IN.
2.3 Write the proofs for the compactness of the integral operator with contin-
uous kernel in G(G) using finite-dimensional approximations as mentioned after
the proof of Theorem 2.21.
2.4 Show that the result of Theorem 2.8 for the norm of the integral operator
remains valid for weakly singular kernels.
Hint: Use the approximations from the proof of Theorem 2.22.
2.5 For the integral operator A with continuous kernel use the Cauchy-Schwarz
inequality to establish that each set U C G(G) that is bounded with respect to
the mean square norm is mapped into a set A(U) C C(G) that is bounded with
respect to the maximum norm and equicontinuous. From this, deduce that the
integral operator with continuous kernel is compact with respect to the mean
square norm. Use the same method and the approximations from the proof of
Theorem 2.22 to extend this result to'weakly singular kernels with a > m/2.
3
Riesz Theory
We now present the basic theory for an operator equation
c.p - Ac.p = f
of the second kind with a compact linear operator A : X -+ X on a normed
space X. This theory was developed by Riesz [153] and initiated through
Fredholm's [42] work on integral equations of the second kind.
3.1 Riesz Theory for Compact Operators

We define
L:=I-A,
where I denotes the identity operator.
Theorem 3.1 (First Riesz Theorem) The nullspace of the operator L,
i.e.,
N(L) := {c.p EX: Lc.p = O},
is a finite-dimensional subspace.
Proof. The nullspace of the bounded linear operator L is a closed subspace

of X, since for each sequence (c.pn) with c.pn -+ c.p, n -+ 00, and Lc.pn = 0
we have that Lc.p = O. Each c.p E N(L) satisfies Ac.p = c.p, and therefore the
restriction of A to N(L) coincides with the identity operator on N(L). The
operator A is compact on X and therefore also compact from N(L) onto

3.1 Riesz Theory for Compact Operators 29
N(L), since N(L) is closed. Hence N(L) is finite-dimensional by Theorem

2.20. 0
Theorem 3.2 (Second Riesz Theorem) The range of the operator L,

i.e.,
L(X) := {L¥? : ¥? EX},
is a closed linear subspace.
Proof. The range of the linear operator L is a linear subspace. Let f be an
element of the closure L( X). Then there exists a sequence (¥?n) in X such
that L¥?n -t f, n -+ 00. By Theorem 1.24 to each ¥?n we choose a best
approximation Xn with respect to N(L), i.e.,
The sequence defined by

ipn := ¥?n - Xn, n E lN,
is bounded. We prove this indirectly, i.e., we assume that the sequence
(ipn) is not bounded. Then there exists a subsequence (ipn(k») such that
lIipn(k) II ~ k for all k E IN. Now we define
¥?n(k)
k E IN.
'l/Jk := I/ipn(k) 1/ '
Since I/'l/Jkl/ = 1 and A is compact, there exists a subsequence ('l/Jk(j») such
that
A'l/Jk(j) -+ 'I/J E X, j -+ 00.
Furthermore,
I/L'l/Jkll = IIL!n(k)1/ ~ II Lipn(k) II -t 0, k -+ 00,

lI¥?n(k) II k
since the sequence (L¥?n) is convergent and therefore bounded. Hence
L'l/Jk(j) -+ 0, j -+ 00.
Now we obtain
'l/Jk(j) = L'l/Jk(j) + A'l/Jk(j) -+ 'I/J, j -+ 00,
and since L is bounded, from the two previous equations we conclude that
L'I/J = O. But then, because Xn(k) + lIipn(k) II 'I/J E N(L) for all k in lN, we
find
lI'l/Jk - 'l/JI/ = 11_ 1 II I/¥?n(k) - {Xn(k) + lIipn(k) II 'I/J }II

¥?n(k)
30 3. Riesz Theory
This contradicts the fact that 'ljJk(j) -7 'IjJ, j -700.

Therefore the sequence (iPn) is bounded, and since A is compact, we can
select a subsequence (iPn(k») such that (A0'n(k») converges as k -7 00. In
view of L0'n(k) -7 f, k -7 00, from 0'n(k) = L0'n(k) +A0'n(k) we observe that
0'n(k) -7 'P E X, k -7 00. But then L0'n(k) -7 £rp E X, k -7 00, and there-
fore f = L'P E L(X). Hence L(X) = L(X), and the proof is complete. 0
For n ;::: 1 the iterated operators Ln can be written in the form
where
An = f)-l)k-l (~) Ak
k=l
is compact by Theorems 2.15 and 2.16. Therefore by Theorem 3.1 the
nullspaces N(Ln) are finite-dimensional subspaces, and by Theorem 3.2
the ranges Ln(x) are closed subspaces.
Theorem 3.3 (Third Riesz Theorem) There exists a uniquely deter-

mined nonnegative integer r, called the Riesz number of the operator A,
such that
and
x = LD(X) ~ Ll(X) ~ ... ~ LT(X) = U+1(X) = .... (3.2)
Furthermore, we have the direct sum
X = N(U) ffi U(X),
i.e., for each 'P E X there exist uniquely determined elements 'IjJ E N(LT)
and X E LT(X) such that 'P = 1/J + X·
Proof. Our proof consists of four steps:
1. Because each 'P with Ln'P = 0 satisfies Ln+l'P = 0, we have
Now assume that
Since by Theorem 3.1 the nullspace N(Ln) is finite-dimensional, the Riesz

Lemma 2.19 implies that for each n E IN there exists 'Pn E N(Ln+l) such
that II'Pnll = 1 and
1
!!'Pn - 'P!! ;::: "2
for all 'P E N(Ln). For n >m we consider
A'Pn - A'Pm = 'Pn - ('Pm + L'Pn - L'Pm)'

Then 'Pm + L'Pn - L'Pm E N(Ln), because
Ln( 'Pm + L'Pn - L'Pm) = L n- m- 1L m+1'Pm + L n+1'Pn - L n - m Lm+1'Pm = O.
Hence
1
IIA'Pn - A'Pmll ~ "2
for n > m, and thus the sequence (A'Pn) does not contain a convergent
subsequence. This is a contradiction to the compactness of A.
Therefore in the sequence N(Ln) there exist two consecutive nullspaces
that are equal. Define
Now we prove by induction that
Assume that we have proven N(Lk) = N(Lk+l) for some k ~ r. Then

for each 'P E N(Lk+2) we have Lk+1 L'P = Lk+2'P = O. This implies that
L'P E N(Lk+l) = N(Lk). Hence Lk+l'P = Lk L'P = 0, and consequently
'P E N(Lk+l). Therefore, N(Lk+ 2) c N(Lk+l), and we have established
that
{O} = N(Lo) ~ N(Ll) ~ ... ~ N(F) = N(F+l) = ....
2. Because for each 'ljJ = Ln+l'P E Ln+1(x) we can write 'ljJ = Ln L'P, we
have
Now assume that
Since by Theorem 3.2 the range Ln(x) is a closed subspace, the Riesz
Lemma 2.19 implies that for each n E 1N there exist 'ljJn E Ln(x) such that
II'ljJnll = 1 and
1
II'ljJn -'ljJ11 ~"2
for all'ljJ E Ln+1(x). We write 'ljJn = Ln'Pn and for m > n we consider
Then 'ljJm + L'ljJn - L'ljJm E Ln+l(X), because
'ljJm + L'ljJn - L'ljJm = L n+1(Lm- n- 1'Pm + 'Pn - Lm-n'Pm)'

32 3. Riesz Theory
Hence
for m > n, and we can derive the same contradiction as before.

Therefore in the sequence Ln(x) there exist two consecutive ranges that
are equal. Define
Now we prove by induction that
Assume that we have proven Lk(X) = Lk+1(X) for some k 2: q. Then for
each'ljJ = Lk+1<p E Lk+l(X) we can write Lk q and let <p E N(LT). Then,
because LT-l<p E LT-l(X) = LT(X), we can write LT-l<p = LTrp for some
rp E X. Since LT+lrp = LT<p = 0, we have rp E N(U+ 1 ) = N(U), i.e.,
U-1<p = LTrp = 0. Thus <p E N(U- 1 ), and hence N(LT-l) = N(U). This
contradicts the definition of r.
On the other hand, assume that r < q and let 'ljJ = Lq-l<p E Lq-l(X).
Because L'ljJ = Lq<p E Lq(X) = Lq+l(X), we can write L'ljJ = Lq+lrp for
some rp E X. Therefore Lq(<p - Lrp) = L'ljJ - Lq+lrp = 0, and from this
we conclude that Lq-l(<p - Lrp) = 0, because N(Lq-l) = N(Lq). Hence
'ljJ = Lqrp E Lq(X), and consequently Lq-l(X) = Lq(X). This contradicts
the definition of q.
4. Let 'ljJ E N(U) n U(X). Then 'ljJ = LT<p for some <p E X and U'ljJ = 0.
Therefore L2T<p = 0, whence <p E N(L2T) = N(U) follows. This implies
'ljJ=U<p=O.
Let <p E X be arbitrary. Then U<p E U(X) = L2T(X) and we can
write LT<p = L2Trp for some rp E X. Now define 'ljJ := Urp E U(X) and
X := <p - 'ljJ. Then LTX = LT<p - L2Trp = 0, i.e., X E N(LT). Therefore the
decomposition <p = X + 'ljJ proves the direct sum X = N(LT) EB U(X). D
We are now ready to derive the following fundamental result of the Riesz
theory.
Theorem 3.4 Let A : X --+ X be a compact linear operator on a normed

space X. Then I - A is injective if and only if it is surjective. If J - A
is injective (and therefore also bijective), then the inverse operator
(J - A)-I: X --+ X is bounded.
Proof. By (3.1) injectivity of I - A is equivalent to r = 0, and by (3.2)

surjectivity of I - A is also equivalent to r = 0. Therefore injectivity of
I - A and surjectivity of I - A are equivalent.
It remains to show that L -1 is bounded when L = I - A is injective.
Assume that L- 1 is not bounded. Then there exists a sequence Un) in X
with lifnll = 1 such that IlL -1 fnll ~ n for all n E IN. Define
fn
gn := IIL-1 fnll '
n E IN.
Then gn -+ 0, n -+ 00, and Iicp,:,,11 = 1 for all n. Since A is compact, we can

select a subsequence (CPn{k)) such that A<Pn{k) -+ cP E X, k -+ 00. Then,
since
CPn - ACPn = gn,
we observe that CPn{k) -+ cP, k -+ 00, and cP E N(L). Hence cP = 0, and this
contradicts II<Pnll = 1 for all n E IN. 0
We can rewrite Theorems 3.1 and 3.4 in terms of the solvability of an

operator equation of the second kind as follows.
Corollary 3.5 Let A : X -+ X be a compact linear operator on a normed
space X. If the homogeneous equation
cP - Acp =0 (3.3)
only has the trivial solution cP = 0, then for each f E X the inhomogeneous
equation
cP - A<p = f (3.4)
has a unique solution cP E X and this solution depends continuously on f.
If the homogeneous equation (3.3) has a nontrivial solution, then it has
only a finite number m E lN of linearly independent solutions CP1, ... ,<Pm
and the inhomogeneous equation (3.4) is either unsolvable or its general
solution is of the form
m
<P = cp + 2: ak<Pk,
k=1
where al, ... ,am are arbitrary complex numbers and cp denotes a particular
solution of the inhomogeneous equation.
The main importance of the Riesz theory for compact operators lies in the
fact that it reduces the problem of establishing the existence of a solution
to (3.4) to the generally much simpler problem of showing that (3.3) has
only the trivial solution cP = o.
It is left to the reader to formulate Theorem 3.4 and its Corollary 3.5 for
integral equations of the second kind with continuous or weakly singular
kernels.
34 3. Riesz Theory
Corollary 3.6 Theorem 3.4 and its Corollary 3.5 remain valid when I - A
is replaced by S - A, where S : X -+ Y is a bounded linear operator that
has a bounded inverse S-l : Y -+ X and A : X -+ Y is a compact linear
operator from a normed space X into a normed space Y.
Proof. This follows immediately from the fact that we can transform the
equation
S'P - A'P = f
into the equivalent form
'P - S-l A'P = S-l f,
where S-lA : X -+ X is compact by Theorem 2.16. o
The decomposition X = N(Lr) EB Lr(x) of Theorem 3.3 generates an

operator P : X -+ N(U) that maps 'P E X onto Pcp := 'tf.; defined by the
unique decomposition 'P = 'tf.; + X with 'tf.; E N(Lr) and X E U(X). This
operator is called a projection operator, because it satisfies p2 = P (see
Chapter 13). We conclude this section with the following lemma on this
projection operator.
Lemma 3.7 The projection operator P : X -+ N(Lr) defined by the de-

composition X = N(U) EB Lr(x) is compact.
Proof. Assume that P is not bounded. Then there exists a sequence ('Pn)
in X with II'Pnll = 1 such that IfP'Pnll :::: n for all n E IN. Define
.1. 'Pn
'f/n := IIP'Pnll' n E IN.
Then 'tf.;n -+ 0, n -+ 00, and IfP'tf.;nll = 1 for all n E IN. Since N(U) is
finite-dimensional and (P'tf.;n) is bounded, by Theorem 1.17 there exists a
subsequence ('tf.;n(k)) such that P'tf.;n(k) -+ X E N(Lr), k -+ 00. Because
'tf.;n -+ 0, n -+ 00, we also have P'tf.;n(k) - 'tf.;n(k) -+ X, k -+ 00. This implies
that X E U(X), since P'tf.;n(k) - 'tf.;n(k) E U(X) for all k and U(X) is
closed. Hence X E N(Lr) n U(X), and therefore X = 0, i.e., P'tf.;n(k) -+ 0,
k -+ 00. This contradicts IfP'tf.;nll = 1 for all n E IN. Hence P is bounded,
and because P has finite-dimensional range P(X) = N(U), by Theorem
2.18 it is compact. 0
3.2 Spectral Theory for Compact Operators

We continue by formulating the results of the Riesz theory in terms of
spectral analysis.
3.2 Spectral Theory for Compact Operators 35
Definition 3.8 Let A : X -t X be a bounded linear operator on a normed

space X. A complex number A is called an eigenvalue of A if there exists
an element <p EX, <p -=I 0, such that A<p = A<p. The element <p is called an
eigenelement of A. A complex number A is called a regular value of A if
(AI - A)-1 : X -t X exists and is bounded. The set of all regular values of
A is called the resolvent set p(A) and R(A; A) := (AI - A)-1 is called the
resolvent of A. The complement of p(A) in <C is called the spectrum o-(A)
and
r(A):= sup IAI
AElT(A)
is called the spectral radius of A.
For the spectrum of a compact operator we have the following properties.
Theorem 3.9 Let A: X -t X be a compact linear operator on an infinite-

dimensional normed space X. Then A = 0 belongs to the spectrum o-(A) and
o-(A) \ {OJ consists of at most a countable set of eigenvalues with no point
of accumulation except, possibly, A = o.
Proof. Suppose that A = 0 is a regular value of A, i.e., A-1 exists and is
bounded. Then I = A -1 A is compact by Theorem 2.16, and by Theorem
2.20 we obtain the contradiction that X is finite-dimensional. Therefore
A = 0 belongs to the spectrum o-(A).
For A -=I 0 we can apply the Riesz theory to the operator AI - A. Either
N(AI - A) = {OJ and (AI - A)-1 exists and is bounded by Corollary 3.6
or N(AI - A) -=I {OJ, i.e., A is an eigenvalue. Thus each A -=I 0 is either a
regular value or an eigenvalue of A.
It remains to show that for each R > 0 there exist only a finite num-
ber of eigenvalues A with IAI 2 R. Assume, on the contrary, that we
have a sequence (An) of distinct eigenvalues satisfying IAnl 2 R. Choose
eigenelements <Pn such that A<Pn = An<Pn for n = 0,1, ... , and define finite-
dimensional subspaces
It is readily verified that eigenelements corresponding to distinct eigenval-

ues are linearly independent. Hence, we have Un-1 C Un and Un-1 -=I Un
for n = 1,2, .... Therefore, by the Riesz Lemma 2.19 we can choose a
sequence ('¢n) of elements '¢n E Un such that II'¢nll = 1 and
1
II'¢n -'¢112 2
for all '¢ E Un - 1 and n = 1,2, .... Writing
36 3. Riesz Theory
we obtain
n-i
>'n'ljJn - A'ljJn = 2)>'n - >'k)ankCPk E Un-i·
k=O
Therefore, for m < n we have
for m < n, and the sequence (A'ljJn) does not contain a convergent subse-
quence. This contradicts the compactness of A. 0
3.3 Volterra Integral Equations

Integral equations of the form
lX K(x,y)cp(y)dy = f(x), x E [a, b),

and
cp(x) - lX K(x, y)cp(y) dy = f(x), x E [a, b),
with variable limits of integration are called Volterra integral equations of

the first and second kind, respectively. Equations of this type were first
investigated by Volterra [180). One can view Volterra equations as special
cases of Fredholm equations with K(x, y) = 0 for y > x, but they have
some special properties. In particular, Volterra integral equations of the
second kind are always uniquely solvable.
Theorem 3.10 For each right-hand side f E Ora, b) the Volterra integral
equation of the second kind
cp(x) -lX K(x,y)cp(y) dy = f(x), x E [a, b),
with continuous kernel K has a unique solution cp E Ora, b).

Proof. We extend the kernel onto [a, b) x [a, b] by setting K(x, y) := 0 for
y > x. Then K is continuous for x =f=. y and
IK(x,y)1 ~ M:= max IK(x,y)1
a~y~x~b
for all x =f=. y. Hence, K is weakly singular with a = 1.

3.3 Volterra Integral Equations 37
Now let <p E eta, b] be a solution to the homogeneous equation
<p(x) -lx K(x, y)<p(y) dy = 0, x E [a, b].
By induction we show that
(3.5)
for n = 0,1,2, .... This certainly is true for n = O. Assume that the
inequality (3.5) is proven for some n ;::: O. Then
1<p(x)1 = Ilar K(x, y)<p(y) dy I:::; 11<plloo Mn+1(x a)n+l

(n + I)!
Passing to the limit n -+ 00 in (3.5) yields <p(x) = 0 for all x E [a,b]. The
statement of the theorem now follows from Theorems 2.22 and 3.4. 0
In terms of spectral theory we can formulate the last result as follows:

A Volterra integral operator with continuous kernel has no spectral values
different from zero.
Despite the fact that, in general, integral equations of the first kind are
more delicate than integral equations of the second kind, in some cases
Volterra integral equations of the first kind can be treated by reducing
them to equations of the second kind. Consider
l x
K(x,y)<p(y)dy = f(x), x E [a,b], (3.6)
and assume that the derivatives Kx = oK/ax and f' exist and are con-
tinuous and that K(x,x) i= 0 for all x E [a,bj. Then differentiating with
respect to x reduces (3.6) to
<p(x) + la
r Kx(x,y) f'(x)
K(x,x) <p(y)dy= K(x, x) , x E [a,bj. (3.7)
Equations (3.6) and (3.7) are equivalent if f(a) = O. If Ky = oK/oy exists

and is continuous and again K(x,x) i= 0 for all x E [a,b], then there is a
second method to reduce the equation of the first kind to one of the second
kind. In this case, setting
'ljJ(x):= l x
<p(y) dy, xE[a,b],
and performing an integration by parts in (3.6) yields
'ljJ(x) - r
la
Ky(x, y) 'ljJ(y) dy = f(x) ,
K(x,x) K(x,x)
x E [a, b]. (3.8)
We leave it as an exercise to extend this short discussion of Volterra

integral equations to the case of Volterra integral equations for functions
of more than one independent variable.
38 3. Riesz Theory
Problems
3.1 Let A : X ---+ Y be a compact linear operator from a normed space X into
a normed space Y. The continuous extension A : X ---+ Y of A is compact with
A(X) C Y (see Problem 2.1).
3.2 Let X be a linear space, let A, B : X ---+ X be linear operators satisfying
AB = BA, and let AB have an inverse (AB)-1 : X ---+ X. Then A and B have
inverse operators A-I = B(AB)-l and B- 1 = A(AB)-I.
3.3 Prove Theorem 3.4 under the assumption that An is compact for some
n?:.l.
Hint: Use Problem 3.2 to prove that the set (I7(A))n := {An : ,\ E I7(A)} is
contained in the spectrum I7(An). Then use Theorem 3.9 to show that there
exists an integer m ?:. n such that each of the operators
27rik
Lk := exp - - I - A, k = 1, ... , m - 1,
m
has a bounded inverse. Then the equations R(I - A)cp = RJ and (I - A)cp = J,
where R:= I1:~1 L k , are equivalent.
3.4 Let Xi, i = 1, . .. , n, be normed spaces. Show that the Cartesian product
X:= Xl X ... X Xn ofn-tuples II' = (cpI, ... ,CPn) is a normed space with the
maximum norm
1111'1100:= i=l,
max IIcpill·
... ,n
Let Aik : Xk ---+ Xi, i, k = 1, ... , n, be linear operators. Show that the matrix
operator A : X ---+ X defined by
n
(Acp)i := L AikCPk
k=1
is bounded or compact if and only if each of its components Aik : Xk ---+ Xi is

bounded or compact, respectively. Formulate Theorem 3.4 for systems of operator
and integral equations of the second kind.
3.5 Show that the integral operator with continuous kernel
L
00
K(x,y):= (k~l)2 {cos(k+1)xsinky-sin(k+1)xcosky}

k=O
on the interval [O,27r] has no eigenvalues.

4
Dual Systems
and Fredholm Alternative
In the case when the homogeneous equation has nontrivial solutions, the
Riesz theory, i.e., Theorem 3.4 gives no answer to the question of whether
the inhomogeneous equation for a given inhomogeneity is solvable. This
question is settled by the Fredholm alternative, which we shall develop in
this chapter. Rather than presenting it in the context ofthe Riesz-Schauder
theory for the adjoint operator in the dual space we will consider the Fred-
holm theory for compact adjoint operators in dual systems generated by
nondegenerate bilinear or sesquilinear forms. This symmetric version is
more elementary and better suited for applications to integral equations.
4.1 Dual Systems via Bilinear Forms

Throughout this chapter we tacitly assume that all linear spaces under
consideration are complex linear spaces; the case of real linear spaces can
be treated analogously.
Definition 4.1 Let X, Y be linear spaces. A mapping (',.) : X X Y -+ <C
is called a bilinear form if
for all 'PI, 'P2, 'P EX, 'IPI, '¢2, '¢ E Y, and ab a2, /3b /32 E <C. The bilinear
form is called nondegenerate if for every 'P E X with 'P =f. 0 there exists

40 4. Dual Systems and Fredholm Alternative
'Ij; E Y such that (cp, 'Ij;) #- 0; and for every 'Ij; E Y with 'Ij; #- 0 there exists
cp E X such that (cp, 'Ij;) #- o.
Definition 4.2 Two normed spaces X and Y equipped with a nondegen-

erate bilinear form (.,.) : X X Y -+ <C are called a dual system and denoted
by (X, Y).
Theorem 4.3 Let G c JRm be as in Theorem 2.8. Then (C(G), C(G) is

a dual system with the bilinear form
(cp,'Ij;):= fa cp(x)'Ij;(x)dx, cp,'Ij; E C(G). (4.1)
Proof. This is obvious from Definition 4.2. o
Definition 4.4 Let (Xl, Yl ) and (X2' Y2) be two dual systems. Then two
operators A : Xl -+ X 2 , B : Y 2 -+ Yl are called adjoint (with respect to
these dual systems) if
(Acp, 'Ij;) = (cp, B'Ij;)
for all cp E Xl, 'Ij; E Y 2. (We use the same symbol (".) for the bilinear
forms on (Xl, Yl ) and (X2' Y 2).)
Theorem 4.5 Let (Xl, Yl ) and (X2' Y2) be two dual systems. If an op-
erator A : Xl -+ X 2 has an adjoint B : Y2 -+ Yl , then B is uniquely
determined, and A and B are linear.
Proof. Suppose that there exist two adjoints to A and denote these by Bl
and B 2. Let B:= Bl - B 2. Then
for all cp E Xl and'lj; E Y 2 • Hence, because (.,.) is nondegenerate, we have

B'Ij; = 0 for all 'Ij; E Y 2, i.e., Bl = B 2.
To show that B is linear we observe that
(cp,/3lB'Ij;l + /32 B 'Ij;2) = /31 (cp, B'Ij;l) + /32(cp, B'Ij;2)

= /31 (Acp, 'lj;l) + /32 (Acp, 'lj;2)
= (Acp, /31 'lj;l + /32'1j;2) = (cp, B(/3l 'lj;l + /32'1j;2»
In a similar manner, it is seen that A is linear. o

Theorem 4.6 Let K be a continuous or a weakly singular kernel. Then in

the dual system (C (G), C (G» the (compact) integral operators defined by
(Acp)(x):= fa K(x, y)cp(y) dy, x E G,
(B'lj;)(x):= fa K(y,x)'lj;(y)dy, x E G,
are adjoint.
Proof The theorem follows from
(Acp, 'lj;) = fa (Acp) (x)'lj;(x) dx = fa (fa K(x, y)cp(y) dY) 'lj;(x) dx

= fa cp(y) (fa K(x, y)'lj;(x) dX) dy = fa cp(y) (B'lj;) (y) dy = (cp, B'lj;).
In the case of a weakly singular kernel, interchanging the order of integra-
tion is justified by the fact that AnCP ~ Acp, n ~ 00, uniformly on G,
where An is the integral operator with continuous kernel Kn introduced in
the proof of Theorem 2.22. 0
4.2 Dual Systems via Sesquilinear Forms

Definition 4.7 Let X, Y be linear spaces. A mapping (. , .) : X X Y ~ <C
is called a sesquilinear form if
(0!1CP1 + 0!2CP2,'lj;) = O!l(cpI,'lj;) + 0!2(CP2,'lj;),
(cp, /31 'lj;1 + /32'lj;2) = i31 (cp, 'lj;1) + i32( cp, 'lj;2)
for all CP1,CP2,CP E X, 'lj;1,'1P2,'lj; E Y, and 0!1,0!2,/31,/32 E <C. (Here, the bar
indicates the complex conjugate.)
We leave it as an exercise to formulate Definition 4.4 and Theorem 4.5
in dual systems generated by nondegenerate sesquilinear forms.
Of course, there is a close relation between bilinear and sesquilinear
forms. Assume that there exists a mapping * : Y ~ Y with the properties
(/31'lj;1 +/32'lj;2)* = i31'lj;i +i32'lj;z and ('lj;*)* = 'lj; for all 'lj;1,'lj;2,'lj; E Y and
/31, /32 E <C. Such a mapping is called an involution and provides a one-to-one
correspon'llence between bilinear and sesquilinear forms by (cp, 'lj;) = (cp, 'lj;*).
In the space C(G) the natural involution is given by 'lj;*(x) := 'lj;(x) for all
x E G and all 'lj; E C(G). Again we leave it as an exercise to formulate
Theorems 4.3 and 4.6 for the corresponding sesquilinear form
(cp,'lj;):= fa cp(x)'lj;(x)dx, cp,'lj; E C(G). (4.2)

The operator A: C[O, 1] --+ C[O, 1] defined by (Arp)(x) := rp(l) provides

an example of a compact operator that does not have an adjoint operator
with respect to the dual systems (C[O,l],C[O, 1]) or (C[O, l],C[O, 1]). As-
sume that B : C[O,l] --+ C[O,l] is an adjoint of A and choose a function
'lj; E C[O, 1] with f01 'lj;(x) dx = 1. Then, by the Cauchy-Schwarz inequality
we have
Irp(l) I = I(Arp, 'lj;) I = I(rp, B'lj;) I :::; Ilrp11211B'lj;112
for all rp E C[O, 1]. Considering this inequality for the sequence (rpn) with
rpn(x) := xn we arrive at a contradiction, since the right-hand side tends
to zero as n --+ 00.
In the sequel, we will demonstrate that in Hilbert spaces for bounded
linear operators the adjoint operators always exist. From Definition 1.19
we observe that each scalar product on a linear space X may be considered
as a nondegenerate sesquilinear form that is symmetric, i.e., (rp, 'lj;) = ('lj;, rp)
°
for all rp, 'lj; EX, and positive definite, i.e., (rp, rp) > for all rp E X with
rp -I- 0. Thus each pre-Hilbert space canonically is a dual system.
Theorem 4.8 (Riesz) Let X be a Hilbert space. Then for each bounded
linear function F : X --+ <c there exists a unique element f E X such that
F(rp) = (rp,f) (4.3)
for all rp EX. The norms of the element f and the linear function F
coincide; i. e.,
Ilfll = IIFII· (4.4)
°°
Proof. Uniqueness follows from the fact that because of the positive definite-
ness of the scalar product, f = is the only representer of the zero function
°
F = in the sense of (4.3). For F -I- choose wE X with F(w) -I- 0. Since
F is continuous, the nullspace N(F) = {rp EX: F(rp) = O} can be seen
to be a closed, and consequently complete, subspace of the Hilbert space
X. By Theorem 1.26 there exists the best approximation v to w with re-
spect to N(F), and by Theorem 1.25 it satisfies w - v -.l N(F). Then for
9 := w - v we have that
(F(g)rp - F(rp)g,g) = 0, rp E X,
because F(g)rp - F(rp)g E N(F) for all rp E X. Hence,
F(9)9)
F(rp) = ( rp'lf9IT2
for all rp E X, which completes the proof of (4.3).

From (4.3) and the Cauchy-Schwarz inequality we have that
IF(rp) I :::; Ilfllllrpll, rp E X,

whence IIFII :S Ilfll follows. On the other hand, inserting f into (4.3) yields
IIfl12 = F(f) :S 11F1i11f11,

and therefore Ilfll :S IIFII· This concludes the proof of the norm equality
~~. 0
Theorem 4.9 Let X and Y be Hilbert spaces, and let A : X -+ Y be
a bounded linear operator. Then there exists a uniquely determined linear
operator A * : Y -+ X with the property
for all rp E X and 'ljJ E Y, i. e., A and A * are adjoint with respect to the
dual systems (X, X) and (Y, Y) generated by the scalar products on X and
Y. The operator.A* is bounded and IIA*II = IIAII. (Again we use the same
symbol (. , .) for the scalar products on X and Y.)
Proof. For each 'ljJ E Y the mapping rp f-t (Arp,'ljJ) clearly defines a bounded
linear function on X, since I(Arp,'ljJ)1 :S IIAllllrpllll'ljJll. By Theorem 4.8 we
can write (Arp, 'ljJ) = (rp, f) for some f EX. Therefore, setting A *'ljJ := f we
define an operator A * : Y -+ X that is an adjoint of A. By Theorem 4.5,
the adjoint is uniquely determined and linear. Using the Cauchy-Schwarz
inequality, we derive
IIA*'ljJ112 = (A*'ljJ, A*'ljJ) = (AA*'ljJ, 'ljJ) ::; IIAIIIIA*'ljJIIII'ljJ11

for all'ljJ E Y. Hence, A* is bounded with IIA*II :S IIAII. Conversely, since
A is the adjoint of A*, we also have IIAII:S IIA*II. Hence IIA*II = IIAII. 0
Theorem 4.10 Let X and Y be Hilbert spaces and let A : X -+ Y be a
compact linear operator. Then the adjoint operator A * : Y -+ X is also
compact.
Proof. Let ('ljJn) be a bounded sequence in Y, i.e, II'ljJnll :S C for all n E IN
and some C > O. By Theorem 4.9 the adjoint operator A* : Y -+ X is
bounded and, consequently, the operator AA * : Y -+ Y is compact by The-
orem 2.16. Hence there exists a subsequence ('ljJn(k») such that (AA*'ljJn(k»)
converges in Y. But then from
IIA*('ljJn(k) - 'ljJn(j») 112 = (AA*('ljJn(k) - 'ljJn(j»), 'ljJn(k) - 'ljJn(j»)
:S 2C1IAA*('ljJn(k) - 'ljJn(j») II
we observe that (A*'ljJn(k») is a Cauchy sequence, and therefore it converges
in the Hilbert space X. 0
The following theorem is due to Lax [114] and provides a useful tool
to extend results on the boundedness and compactness of linear operators
from a given norm to a weaker scalar product norm.
Theorem 4.11 (Lax) Let X and Y be normed spaces, both of which have
a scalar product (. , .), and assume that there exists a positive constant c
such that
(4.5)
for all cp, 'Ij; EX. Let A : X -+ Y and B : Y -+ X be bounded linear

operators satisfying
(Acp,'Ij;) = (cp,B'Ij;) (4.6)
for all cp E X and 'Ij; E Y. Then A : X -+ Y is bounded with respect to the

norms II . lis induced by the scalar products and
(4.7)
Proof. Consider the bounded operator M : X -+ X given by M := BA with

IIMII S IIBIIIIAII· Then, as a consequence of (4.6), M is self-adjoint, i.e.,
(Mcp, 'Ij;) = (cp, M'Ij;) for all cp, 'Ij; E X. Therefore, using the Cauchy-Schwarz
inequality, we obtain
for all cp E X with Ilcplls S 1 and all n E IN. From this, by induction, it
follows that
IIMcplls S IIM2n cpll;-n.
By (4.5) we have Ilcplls S vIC Ilcpll for all cp E X. Hence,
Passing to the limit n -+ 00 now yields
IIMcplis S IIMII
for all cp E X with I/cpl/s S 1. Finally, for all cp E X with I/cpl/s S 1, we have
from the Cauchy-Schwarz inequality that
From this the statement follows. o
For an example of an application of Lax's theorem let X = Y = C( G) be

equipped with the maximum norm and the L2 scalar product (4.2). Using
the approximations from the proof of Theorem 2.22, without any further
analysis, from Theorems 2.17 and 4.11 it can be seen that integral operators
with weakly singular kernels are compact in the completion of C(G) with
respect to the scalar product (4.2), i.e., in L2(G) (see Problem 4.5).
4.3 The Fredholm Alternative

Now we proceed to develop the Fredholm theory for compact operators,
which we will write for a dual system generated by a bilinear form. We
begin with the following technical lemma.
Lemma 4.12 Let (X, Y) be a dual system. Then to every set of linearly
independent elements !P1, ... ,!Pn E X there exists a set 1/11, ... ,1/1n E Y
such that
(!pi,1/1k) = 8ik , i, k = 1, ... , n,
where 8i k = 1 for i = k and 8ik = 0 for if. k. The same statement holds
with the roles of X and Y interchanged.
Proof. For one linearly independent element the lemma is true, since (.,.)
is nondegenerate. Assume that the assertion of the lemma has been proven
for n ~ 1 linearly independent elements. Let !PI, ... , <pn+1 be n + 1 lin-
early independent elements. Then, by our induction assumption, for every
m = 1, ... , n + 1, to the set !PI, ... , !Pm-I. !Pm+1, ... , !Pn+1 of n elements
in X there exists a set of n elements 1/1im), ... ,1/1;;;~1' 1/1;;:;'1' ... ,1/1~~i in Y
such that
(!pi,1/1im)} = 8ik , i, k = 1, ... ,n + 1, i, k f. m. (4.8)
Since (" -) is nondegenerate, there exists Xm E Y such that
because otherwise
n+1
!Pm - L (!Pm, 1/1im )}!Pk = 0,
k=l
ki=m
which is a contradiction to the linear independence of the !PI,· .. , !Pn+1'
Define
1/1m := -f-
m
{xm - I: 1/1im)
k=l
(!pk, Xm)} .
ki=m
Then (!Pm,1/1m) = 1, and for i f. m we have
because of (4.8). Hence we obtain 'I/J!, ... , 'l/Jn+l such that

(CPi,'l/Jk)=bik, i,k=1, ... ,n+1,
and the lemma is proven. 0
Theorem 4.13 (First Fredholm Theorem) Let (X, Y) be a dual system

and A : X --t X, B : Y --t Y be compact adjoint operators. Then the
nullspaces of the operators J - A and J - B have the same finite dimension.
Proof. By the first Riesz Theorem 3.1 we have
m:= dimN(I - A) < 00, n:= dimN(J - B) < 00.
We assume that m < n. Then we choose a basis CPI, ... , CPm for N(I - A)
(if m > 0) and a basis 'l/Jl, ... , 'l/Jn for N(I - B). By Lemma 4.12 there exist
elements al, ... , am E Y (if m > 0) and bl , ... , bn E X such that
i,k = 1, . .. ,m,
i,k = 1, ... ,no

Define a linear operator T : X --t X by
m
Tcp:= L(cp,ai)bi , cP E X, (4.9)
i=l
if m > 0 and T := 0 if m = 0. Recall the compact projection operator

P : X --t N[(J - At] from Lemma 3.7. Since T : N[(I - At] --t X is
bounded by Theorem 2.6, from Theorem 2.16 we have that T P : X --t X
is compact. Therefore, in view of Theorem 2.15, we can apply the Riesz
theory to the operator A - T P.
For m > 0, from
we find that
k= 1, ... ,m,
(4.10)
k = m+ 1, ... ,n.
Now let cP E N(I - A + TP). Then from (4.10) we see that
(Pcp, ak) = 0, k = 1, ... , m, (4.11)
and therefore TPcp = O. Hence cP E N(J - A) and, consequently, we can

write
m
cP = L L1!iCPi,
i=l
where D:i = (I.{?, ai) for i = 1, ... , m. Now from PI.{? = I.{? for I.{? E N(I - A)
and (4.11) we conclude that I.{? = O. Thus we have proven that I - A + TP
is injective. This, of course, is also true in the case when m = O.
Since I - A +T P is injective1 by Theorem 3.4 the inhomogeneous equation
I.{? - AI.{? + TPI.{? = bn
has a unique solution I.{? Now, with the aid of the second line of (4.10)
(which is also true for m = 0) we arrive at the contradiction
Therefore m ~ n. Interchanging the roles of A and B shows that n ~ m.

Hence m = n. 0
Theorem 4.14 (Second Fredholm Theorem) Let (X, Y) be a dual sys-

tem and A : X --t X, B : Y --t Y be compact adjoint operators. Then
(I - A)(X) = {f EX: (I,1/;) = 0, 1/; E N(I - Bn

and
(I - B)(Y) = {g E Y : (I.{?,g) = 0, I.{? E N(I - An.
Proof. It suffices to carry out the proof for the range of I - A, and by
Theorems 3.4 and 4.13 we only need to consider the case where m > O. Let
f E (I - A)(X), i.e, f = I.{? - AI.{? for some I.{? E X. Then
(I,1/;) = (I.{? - AI.{?, 1/;) = (I.{?,1/; - B1/;) =0
for all1/; E N(I - B).
Conversely, assume that f satisfies (I,1/;) = 0 for all1/; E N(I - B). From
the proof of the previous theorem we know that the equation
I.{? - AI.{? + T PI.{? = f

has a unique solution I.{? Then, in view of (4.10) and the condition on f,
we have
Hence T PI.{? = 0, and thus I.{? also satisfies I.{? - AI.{? = f. 0
We now summarize our results in the so-called Fredholm alternative.

Theorem 4.15 Let A: X --t X, B : Y --t Y be compact adjoint operators
in a dual system (X, Y). Then either I - A and 1- B are bijective or I - A
and I - B have nontrivial nullspaces with finite dimension
dimN(I - A) = dimN(I - B) E 1N
and the ranges are given by
(J - A)(X) = {f EX: (f,'Ij;) = 0, 'Ij; E N(J - B)}
and
(J - B)(Y) = {g E Y : (cp,g) = 0, cp E N(J - An.
Choosing the dual system introduced in Theorem 4.3 and the integral
operators with continuous or weakly singular kernels considered in Theo-
rem 4.6, our results include the classical Fredholm alternative for integral
equations of the second kind that was first obtained by Fredholm [42] and
which we now state as a corollary.
Corollary 4.16 Let G c lRm be as in Theorem 4.3 and let K be a con-

tinuous or weakly singular kernel. Then either the homogeneous integral
equations
cp(x) - i K(x, y)cp(y) dy = 0, x E G,
and
'Ij;(x)- iK(y,X)'Ij;(y)dY=O, XEG,
only have the trivial solutions cp =0 and 'Ij; =0 and the inhomogeneous
integral equations
cp(X) - i K(x, y)cp(y) dy = f(x), x E G,
and
'Ij;(X) - i K(y,x)'Ij;(y) dy = g(x), x E G,
have a unique solution cp E C(G) and'lj; E C(G) for each right-hand side
f E C(G) and g E C(G), respectively, or the homogeneous integral equa-
tions have the same finite number m E :IN of linearly independent solutions
and the inhomogeneous integral equations are solvable if and only if the
right-hand sides satisfy
i f(x)'Ij;(x) dx = 0
for all solutions 'Ij; of the homogeneous adjoint equation and
i cp(x)g(x) dx = 0
for all solutions cp of the homogeneous equation, respectively.

The original proof by Fredholm for this result for continuous kernels
is based on interpreting the integral equations as a limiting case of linear
systems by considering the integrals as a limit of Riemann sums and passing
to the limit in Cramer's rule for the solution of linear systems (see [79]). We
wish to mention that this original proof is shorter than our more general
approach, however, it is restricted to the case of integral equations with
continuous kernels.
Our results also include the so-called Riesz-Schauder theory as devel-
oped by Schauder [160) (see also [70, 79, 80, 169]) by taking Y = X*, the
dual space of X, which is defined as the normed space X* = L(X, CC) of
bounded linear functionals, i.e., bounded linear operators F : X --+ CC. A
normed space X together with its dual space X* forms a canonical dual
system (X,X*) with the bilinear form defined by (cp,F) := F(cp) for all
elements cp E X and all functionals F E X*. It is a consequence of the
Hahn-Banach theorem (see [70, 79, 80,169)) that this bilinear form is non-
degenerate. Our more general form of the Fredholm alternative seems to
be more appropriate for the discussion of integral equations because of its
symmetric structure. In particular, in our setting the adjoint of an integral
equation in a function space again is an integral equation in a function
space, whereas in the Schauder theory the adjoint equation is an equation
in the dual space of bounded linear functionals. Hence, the Schauder theory
does not immediately include the classical results of Fredholm on integral
equations with continuous kernels.
The Fredholm alternative in dual systems was first proven by Wend-
land [181, 183] under the assumption that the bilinear or sesquilinear form
is bounded. An elementary proof, which does not use the Hahn-Banach
theorem, i.e., Zorn's lemma, and which also does not require the bilinear or
sesquilinear form to be bounded, was first given in [24) (see also [99)). The
current version of the proof of Theorem 4.13, which differs slightly from the
proof in [24] and in the first edition of this book, is due to Martensen [119).
For a history ofthe Fredholm alternative in dual systems we refer to [101).
Example 4.17 C~msider the integral equation
cp(x) -lab eX-Ycp(y) dy = f(x), x E [a, b). (4.12)
Obviously a solution of (4.12) must be of the form
cp(x) = f(x) + ce x , (4.13)
where c is a constant. Inserting (4.13) into (4.12), we observe that cp solves

the integral equation provided that c satisfies
c{l - (b - an = 1b e-Yf(y) dy. (4.14)

Now either b - a =f:. 1 or b - a = 1. In the first case (4.14) has a unique

solution leading to the unique solution
1 (b
cp(x) = f(x) + 1- (b _ a) Ja e- Yf(y) dy eX
of the integral equation. In the second case, (4.14) has a solution if and
only if
(4.15)
and then any c satisfies (4.14). Note that, for b - a = 1, the function
,¢(x) = e- x is the solution of the homogeneous adjoint equation
,¢(x) -l b
eY-X,¢(y) dy = 0, x E [a,b],
and therefore (4.15) coincides with the solvability condition of the Fredholm
alternative. 0
4.4 Boundary Value Problems

We conclude this chapter by giving some flavor of the use of the Riesz-
Fredholm theory to solve boundary value problems by considering the or-
dinary differential equation of the second order
(4.16)
on the interval [0,1] with coefficients at, a2, v E C[O, 1] subject to the
boundary conditions
u(O) = uo, u(1) = U1. (4.17)
The general idea in the application of integral equations in the treatment of
boundary value problems is to equivalently transform the boundary value
problem into an integral equation and then solve the integral equation.
Let u be twice continuously differentiable and set cp := -u". Then, by
partial integration we find the relations
u(x) = u(O) + u'(O)x -loX (x - y)cp(y) dy,
u(x) = u(1) - u'(1)(1 - x) + 11 (x - y)cp(y) dy,
0= 10 1
cp(y) dy + u'(1) - u'(O).
4.4 Boundary Value Problems 51
Multiplying the first equation by (1 - x), the second by x, the third by

(1 - x)x, and then adding, we obtain
U(x) = u(O)(l-x)+u(l)x+ lox (l-x)y<p(y) dy+ i 1

x(l-y)<p(y) dy. (4.18)
Differentiating this equation yields
U'(x) = u(l) - u(O) -loX y<p(y) dy + i 1

(1 - y)<p(y) dy. (4.19)
Now let u be a solution to the boundary value problem. Then, using the
differential equation (4.16) and the boundary condition (4.17), from (4.18)
and (4.19) we deduce that <p = -u" satisfies the integral equation
<p(x) -10 K(x,y)<p(y) dy

1
= f(x), x E [0,1]' (4.20)
with the kernel

o ~ y < x ~ 1,
o~ x <y ~ 1,
and the right-hand side
f(x):= (Ul - uo)al(x) + {uo(l- x) + ulx}a2(x) - v(x), x E [0,1).
Conversely, let <p E C[O, 1] be a solution to the integral equation and

define a twice continuously differentiable function u by
u(x) := uo(I-X)+UIX+ lox (l-x)y<p(y) dy+ i 1

x(l-y)<p(y) dy, x E [0,1).
Then u(O) = uo, u(l) = UI, and by construction of the integral equation
we have -u" = <p = al u' + a2U - v. Therefore, the boundary value problem
and the integral equation are equivalent. In particular, via
w(x) := fox (1 - x)y<p(y) dy + 11 (1 - y)x<p(y) dy, x E [0,1),
the homogeneous boundary value problem
w" + alw' + a2W = 0, w(O) = w(l) = 0 (4.21)
and the homogeneous integral equation
<p(x)- fol K(x,y)<p(y)dy=O, xE [0,1], (4.22)

are equivalent. Note that W -=f. 0 implies that tp -=f. 0 and vice versa.
Because the kernel K is continuous and bounded on 0 ::; y < x ::; 1
and on 0 ::; x < y ::; 1, it is weakly singular with a = 1. Hence, the
Fredholm alternative is valid: Either the inhomogeneous integral equa-
tion (4.20) is uniquely solvable for each right-hand side f E C[O, 1] and
therefore the boundary value problem itself also is uniquely solvable for
all inhomogeneities v E C[O, 1] and all boundary values Uo and Ul, or the
homogeneous integral equation (4.22), and consequently the homogeneous
boundary value problem, have nontrivial solutions.
The homogeneous boundary value problem (4.21) has at most one lin-
early independent solution. To show this, let WI, W2 be two solutions of
(4.21). Then there exist constants Al,A2 such that AIW~(O) + A2W~(0) = 0
and IA1/ + IA21 > O. From the homogeneous boundary conditions we also
have that AIWl(O) + A2W2(0) = o. Now the Picard-Lindelof uniqueness
theorem for initial value problems implies that Al WI + A2W2 = 0 on [a, b],
i.e., two solutions to the homogeneous boundary value problem (4.21) are
linearly dependent. Therefore, if (4.21) has nontrivial solutions, it has one
linearly independent solution and the homogeneous integral equation and
its adjoint equation both also have one linearly independent solution.
Let 'lj; be a solution to the homogeneous adjoint equation
'lj;(x) - 10 1
K(y, x)'lj;(y) dy = 0, x E [0,1]'
i.e.,
Then 'lj;(0) = 'lj;(1) = 0 and for the derivative we find
whence
('lj;' - al'lj;)' + a2'lj; = 0
follows. By the Picard-Lindelof theorem this homogeneous adjoint bound-
ary value problem again admits at most one linearly independent solution.
Therefore, in the case when the homogeneous boundary value problem has
a nontrivial solution, the homogeneous adjoint integral equation and the
homogeneous adjoint boundary value problem are equivalent. By the Fred-
holm alternative, the inhomogeneous integral equation (4.20) and therefore
the inhomogeneous boundary value problem (4.16) and (4.17) are solvable
if and only if
10 1
f(x)'lj;(x) dx = 0 (4.23)
Problems 53
is satisfied for the solutions 'lj; of the homogeneous adjoint boundary value
problem. Using the differential equation and the boundary condition for 'lj;,
we find
Hence, the condition (4.23) and
10 1
v(x)'lj;(x) dx = uo'lj;'(O) - ul'lj;'(l)
are equivalent, and we can summarize our results in the following form.
Theorem 4.18 Either the inhomogeneous boundary value problem
is uniquely solvable for all right-hand sides v and boundary values uo, ul or
the homogeneous boundary value problem
wI! + al W ' + a2 W = 0, w(O) = w(l) = 0,

and the homogeneous adjoint boundary value problem
each have one linearly independent solution wand 'lj;, respectively. In the
latter case, the inhomogeneous boundary value problem is solvable if and
only if
10 1
v(x)'lj;(x) dx = uo'lj;'(O) - ul'lj;'(l).
Problems
4.1 Let (.,.) : C[a, b] x C[a, b] -+ 1R be a degenerate bilinear form. Then there
exists a function f E C[a, b] such that (1, 'Ij;) = 0 for all 'Ij; E C[a, b]. Since f i= 0,
without loss of generality we may assume that f(a) = 1. The compact operators
A,B : C[a,b] -+ C[a,b] defined by A'P := 'P(a)f and B'Ij; := 0 are adjoint with
respect to (', -). By showing that N(I - A) = span{f} and N(J - B) = {O}
demonstrate that for the validity of the Fredholm alternative the bilinear form
necessarily must be nondegenerate.
4.2 Let X be the linear space of all functions 'I' E C(O, 1] for which positive
numbers M and Q: (depending on '1') exist such that 1'P(x)'1 ::; MX",-1/2 for all
x E (0,1]. Then X is a normed space with the norm
11'1'11:= sup v'X1'P(x)l,

xE(O,lj
and (X, X) is a dual system with the bilinear form
(rp,'I/J):= 11 rp(x)'I/J(x)dx.
Show that the integral operators A, B : X --t X with continuous kernel K defined
as in Theorem 4.6 are compact and adjoint. By using the sequence (rpn) given by
rpn(X) := X 1 / n - 1 / 2 show that the bilinear form is not bounded.
4.3 Formulate and prove the Fredholm alternative for a pair of operators S - A
and T-B, where Sand T each have a bounded inverse and A and B are compact.
4.4 Show that under the assumptions of Theorem 4.15 the operators A and
B both have Riesz number one if and only if for each pair of basis rp1,· .. , rpm
and 'l/J1, ... , 'l/Jm of the nullspaces N(J - A) and N(J - B) the matrix (rpi, 'l/Jk),
i, k = 1, ... , m, is nonsingular.
4.5 Use Lax's Theorem 4.11 to show that the integral operator with weakly
singular kernel of Theorem 2.22 is a compact operator from L2(G) into L2(G)
(see also Problem 2.5).
5
Regularization in Dual Systems
In this chapter we will consider equations that are singular in the sense
that they are not of the second kind with a compact operator. We will
demonstrate that it is still possible to obtain results on the solvability of
singular equations provided that they can be regularized, i.e., they can be
transformed into equations of the second kind with a compact operator.
5.1 Regularizers
The following definition will say more precisely what we mean by regular-
izing a bounded linear operator.
Definition 5.1 Let Xl> X 2 be normed spaces and let K : Xl -+ X 2 be a
bounded linear operator. A bounded linear operator RI. : X2 -+ Xl is called
a left regularizer of K if
ReK = I -AI., (5.1)
where AI. : Xl -+ Xl is compact; a bounded linear operator Rr : X 2 -+ Xl
is called a right regularizer of K if
(5.2)
where Ar : X 2 -+ X 2 is compact. A bounded linear operator R : X 2 -+ Xl
is called a regularizer of K if
RK=I-A£ and KR=I-Ar' (5.3)
where Ae : Xl -+ Xl and Ar : X 2 -+ X 2 are compact.

56 5. Regularization in Dual Systems
The difference between a left and a right regularizer is compact, since by

multiplying (5.1) from the right by Rr and (5.2) from the left by R£ and
then subtracting we obtain Rr - R£ = A£Rr - R£Ar' which is compact by
Theorems 2.15 and 2.16. Again by Theorem 2.16, we observe that adding
a compact operator to a regularizer preserves the regularizing property.
Therefore, provided that there exist a left and a right regularizer we may
always assume that R£ = Rr = R.
Let us first consider regularizing from the left. Any solution to the orig-
inal equation
K<p= f (5.4)
also solves the regularized equation
<p - AN = Rd· (5.5)
Therefore, by regularizing from the left we do not lose any solutions. Con-
versely, let <p be a solution of the regularized equation (5.5). Then
R£(K<p - f) = 0,
and <p solves the original equation (5.4) provided that N(Rt} = {o}. We
call a left regularizer an equivalent left regularizer if the original and the
regularized equation have the same solutions. Then, we have the following
theorem.
Theorem 5.2 A left regularizer is an equivalent left regularizer if and only
if it is injective.
Now let us treat regularizing from the right. Here we have to compare
the original equation (5.4) and the regularized equation
(5.6)
Provided that 'if; solves the regularized equation, <p := Rr'if; is a solution
to the original equation (5.4). Therefore, by regularizing from the right we
do not create additional solutions. Conversely, to each solution <p of the
original equation (5.4) there corresponds a solution 'if; of the regularized
equation (5.6) with Rr'if; = <p provided that Rr(X2 ) = Xl. We call a right
regularizer an equivalent right regularizer if it maps the solutions of the
regularized equation onto the solutions of the original equation. Then, we
have the following theorem.
Theorem 5.3 A right regularizer is an equivalent right regularizer if and

only if it is surjective.
From Theorems 5.2 and 5.3 we conclude that in a situation where we can
establish the existence of an injective left regularizer or a surjective right
regularizer the results of the Riesz theory partially carryover to the singular
equation K cp = f. In particular, if K is injective and has an equivalent left

regularizer, then K is surjective by Theorem 3.4. If K is surjective and has
an equivalent right regularizer, then K is injective. The transformation of
a Volterra equation of the first kind into Volterra equations of the second
kind as described in Section 3.3 may serve as a first example (see Problem
5.1).
5.2 Normal Solvability

In this section we want to demonstrate that it is also possible to obtain
solvability results by regularization that is not equivalent. To this end, we
first prove the following lemma.
Lemma 5.4 Under the assumptions of Theorem 4.13 let 1 - A have a

nontrivial nullspace. Then the Riesz number r(A) of A and the Riesz num-
ber r(B) of B coincide, i.e., r(A) = r(B) = r E IN. The corresponding
projection operators
P: X-+ N[(1 - An and p': Y -+ N[(1 - Bn

defined by the direct sums
X = N[(1 - An EB (1 - At(X) and Y = N[(1 - Bn EB (1 - Bt(Y),

respectively, are adjoint. The elements al,' .. ,am and bl , ... ,bm entering
in the definition (4.9) of the operator T can be chosen such that the operator
S:= (1 -A+TP) : X -+ X and its inverse S-1 : X -+ X both have adjoint
operators S' : Y -+ Y and [S-I]' : Y -+ Y, respectively. The adjoint S' has
an inverse and [S-I]' = [S']-I.
Proof. By Theorem 4.13, for q E lN, the two nullspaces N[(1 - A)q] and
N[(1 - B)q] have the same finite dimension. Therefore, in view of Theorem
3.3, the Riesz numbers of A and B must coincide.
As a consequence of Theorem 4.14, applied to (1 - At and (1 - Bt, we
have (Pcp, 'lj; - P''lj;) = 0 for cp E X and'lj; E Y, since Pcp E N[(1 - At] and
'lj; - P''lj; E (1 - Bt(Y). Therefore
(Pcp, 'lj;) = (Pcp, p' 'lj; + 'lj; - p' 'lj;) = (Pcp, P''lj;)
for all cp E X and 'lj; E Y. Analogously, we have (cp, P''lj;) = (Pcp, P''lj;), and
consequently (Pcp, 'lj;) = (cp, P''lj;) for all cp E X and 'lj; E Y.
Let cp E N[(1 - At] such that (cp, 'lj;) = 0 for all 'lj; E N[(I - Bn. Then,
by Theorem 4.14 we also have cp E (1 - At(X), and therefore cp = o.
Analogously, 'lj; E N[(1 - Bn and (cp, 'lj;) = 0 for all cp E N[(1 - At]
implies that 'lj; = O. Therefore, the bilinear form generating the dual system
(X, Y) is nondegenerate on N[(I - An

x N[(I - B)'l Using Lemma 4.12,
this implies that we can choose the elements entering in the definition of
T such that a1,"" am E N[(I - Bn
and b1 , •.. , bm E N[(I - At], and
consequently PT = T.
The operator T' : Y ~ Y defined by
m
T''Ij;:= I:(bi,'Ij;)ai' 'Ij; E Y,

i=l
clearly is the adjoint of T and P'T' = T'. Then we have
(TPtp,'Ij;) = (PTPtp,'Ij;) = (tp,P'T'P''Ij;) = (tp,T'P''Ij;)

for all tp E X and 'Ij; E Y, i.e., the operators TP and T' P' are adjoint.
Analogous to T P, the operator T' P' is also compact. Therefore, by the
Fredholm alternative, the bijectivity of S := I - A + T P implies bijectivity
of S' := I - B + T' P'. Now the last statement of the lemma follows from
(S-l f, g) = (S-l f, S'[S']-l g) = (SS-l f, [S']-lg) = (I, [S'rlg)

for all f E X and g E Y, i.e., [S- l ], = [S']-l. o
Theorem 5.5 Let Xl, X 2 be normed spaces, let K : Xl ~ X 2 be a bounded
linear operator and let R: X 2 ~ Xl be a regularizer of K. Then K and R
have finite-dimensional nullspaces.
Proof. By assumption we have that RK = I - Ae and KR = I - A r ,
where Ae and Ar are compact. Let K tp = 0. Then RK tp = 0, and therefore
N(K) C N(I - Ae). By Theorem 3.1, N(I - Ae) has finite dimension.
Therefore dimN(K) ::; dimN(I - Ac) < 00. Analogously we have that
N(R) C N(I - A r ), which implies dimN(R) ::; dimN(I - Ar) < 00. 0
Theorem 5.6 Let (Xl, YI ) and (X2' Y 2) be dual systems, let K : Xl ~ X 2,

K' : Y2 ~ YI and R : X 2 ~ Xl, R' : Yi ~ Y2 be two pairs of bounded and
adjoint operators such that R is a regularizer of K and R' is a regularizer
of K'. Then the nullspaces of K and K' have finite dimension, and the
ranges are given by
K(X 1 ) = {f E X 2 : (I, 'Ij;) = 0, 'Ij; E N(K')}
and
K'(Y2 ) = {g E YI : (tp,g) = 0, rp E N(K)}.
Proof. The finite dimension of the nullspaces follows from Theorem 5.5. By
symmetry it suffices to prove the statement on the ranges for the operator
K. By assumption we have
RK= I -A, K'R' = I -B,

where A: Xl -+ Xl and B : YI -+ YI are compact and adjoint.

Let f E K(X I ), i.e., f = K<p for some <p E Xl' Then
(j,'lj;) = (K<p,'lj;) = (<p,K''lj;) = 0
for all 'lj; E N(K'), and therefore
K(X I ) C {f E X 2 : (j, 'lj;) = 0, 'lj; E N(K')}.
Conversely, assume that f E X 2 satisfies (j, 'lj;) = 0 for all 'lj; E N(K').
Then
(Rf, X) = (j,R'X) = 0
for all X E N(J - B), since K'R'X = X - BX = O. Therefore, by Theorem
4.15, the regularized equation
<p - A<p = Rf (5.7)
is solvable. If the regularizer R is injective, then by Theorem 5.2 each

solution of (5.7) also satisfies the original equation K <p = f, and therefore
in this case the proof is complete.
If R is not injective, by Theorem 5.5 its nullspace is finite-dimensional.
Let n := dimN(R) and choose a basis hI"'" h n of N(R). By Lemma 4.12
there exist elements Cll ... ,Cn E Y 2 such that
We first consider the case where J - A is injective. Then, for the solution
<p = (I - A)-lRf of (5.7) we can write
n
K<p - f = LJ3kh k,
k=l
where
fA = (K(J - A) -1 Rf - f, Ck) = (j, gk)
and
gk := R'(I - B)-lK'Ck - Ck
for k = 1, ... ,n. For all 'lj; E Xl we have
whence K' gk = 0 follows for k = 1, ... , n. This implies (j, gk) = 0 for
k = 1, ... ,n, and therefore K <p = f, i.e., the proof is also complete in this
case.
It remains to consider the case where both Rand J - A have a nontrivial
nullspace. Let m := dimN(I - A), choose a basis <PI,"" <Pm of N(J - A),
and recall the analysis of Theorems 4.13 and 4.14 and Lemma 5.4. The
general solution of the regularized equation (5.7) has the form
m
<P = LOWi + (I - A+TP)-lRj (5.8)

i=l
with complex coefficients a1,' .. , am. Since for each solution <P of (5.7) we
have that K<p - j E N(R), as above we can write
n
K<p - j = L(3khk,
k=l
where
m
(3k = Lai(K<pi,Ck) + (K(I - A +TP)-lRj - j,Ck), k = 1, ... ,no
i=l
Therefore the solution (5.8) of the regularized equation solves the original
equation if and only if the coefficients aI, ... ,am solve the linear system
m
L ai(K<pi, Ck) = (I - K(I - A + TP)-l Rj, Ck), k = 1, ... , n. (5.9)
i=l
If the matrix (K <Pi, Ck) of this linear system has rank n, then (5.9) is solv-
able. Otherwise, i.e., for rank p < n, from elementary linear algebra we
know that the conditions that are necessary and sufficient for the solvabil-
ity of (5.9) can be expressed in the form
n
LPik(l - K(I - A+TP)-lRj,Ck) = 0, i = 1, ... ,n - p,
k=l
for some matrix Pik, or in short form
(I, gi) = 0, i = 1, ... , n - p, (5.10)
where
n
gi := LPik {Ck - R'(I - B +T'p,)-lK'ck} , i = 1, ... ,n - p.
k=l
Now the proof will be completed by showing that gi E N(K')

For'ljJ E Xl, observing that 'ljJ - A'ljJ = RK'ljJ, analogously to (5.8) we can
write
m
'ljJ= Lai<Pi + (I-A+TP)-lRK'ljJ

i=l
for some D:l, ••• , D: m E <C, and consequently

m
L D:iKCPi + K(I - A + TP)-lRK'IjJ - K'IjJ = o.
i=l
This implies that the linear system (5.9) with f replaced by K'IjJ has a
solution. Therefore, since the conditions (5.10) are necessary for the solv-
ability,
('IjJ,K'gi) = (K'IjJ,gi) = 0, i = 1, ... ,n - p,
for all 'IjJ E Xl. This now implies that K' gi = 0, i = 1, ... ,n - p. 0
Comparing Theorem 5.6 with the Fredholm alternative for compact op-
erators we note that the only difference lies in the fact that the dimensions
of the nullspaces, in general, are no longer the same. In particular, this im-
plies that from injectivity of K, in general, we cannot conclude surjectivity
of K as in the Riesz theory. This also will give rise to the introduction of
the index of an operator in the following section.
As in Chapter 4 our analysis includes the special case of the canonical
dual systems (XI, Xi) and (X2' X z) with the dual spaces Xi and X of z
Xl and X 2 • In this setting the solvability conditions of Theorem 5.6 usu-
ally are referred to as normal solvability of the operator K, and in this
case the results of Theorem 5.6 were first obtained by Atkinson [7]. For the
reasons already mentioned in Chapter 4, our more general setting again
seems to be more appropriate for the discussion of integral equations. This
will become obvious from the various examples discussed in the following
chapters, which will include the classical results by Noether [138] on singu-
lar integral equations with Hilbert kernels. For convenience, we reformulate
Theorem 5.6 in terms of solvability conditions.
Corollary 5.7 Under the assumptions of Theorem 5.6 each of the homo-
geneous equations
K cP = 0 and K' 'IjJ = 0
has at most a finite number of linearly independent solutions. The inhomo-
geneous equations
K cP = f and K' 'IjJ = 9
are solvable if and only if the conditions
(j,'IjJ) =0 and (cp,g) =0

are satisfied for all solutions 'IjJ and r.p of the homogeneous equations
K' 'IjJ =0 and K cP = 0,

respectively.
5.3 Index
We conclude this chapter by introducing the concept of the index of an
operator. Let U and V be subspaces of a linear space X such that
X=UEElV.
Let n := dim V < 00 and assume that W is another subspace W with the
property X = U EEl W. Choose a basis Vl, ... ,Vn of V and let Wl, .•. , Wn+1
be n + 1 elements from W. Then, since X = U EEl V, there exist elements
UI, ... , U n +1 from U and a matrix Pik such that
n
Wi = Ui + E PikVk, i = 1, ... , n + 1.
k=1
The homogeneous linear system

n+1
E PikAi = 0, k = 1, ... , n,
i=1
of n equations for n + 1 unknowns has a nontrivial solution Al, .•. , An+l.

Then
n+1 n+1
L AiWi = L AiUi,
i=1 i=1
and consequently
n+1
LAiWi = 0,
i=1
since UnW = {O}. Hence the elements WI. ..• , W n +1 are linearly dependent,
and therefore dim W ::; dim V. Interchanging the roles of V and W we also
have dim V ::; dim W, whence dim V = dim W follows.
Therefore, the codimension of a subspace U of X is well defined by setting
codim U := dim V
if there exists a finite:dimensional subspace V such that X = U EEl V,

and codimU = 00 otherwise. In particular, codimX = O. Obviously, the
co dimension is a measure for the deviation of the subspace U from the
whole space X.
Assume that U I and U2 are subspaces of X such that codim U I < 00
and U I C U2. Theil X = U I EEl VI, where VI is a finite-dimensional sub-
space. For the finite-dimensional subspace VI n U2 of VI there exists an-
other finite-dimensional subspace V2 such that VI = (VI n U2) EEl V2. Now let
<p E U2 n V2. Then, since V2 C VI, we have <p E VI n U2. Therefore <p = 0,
since (VI n U2) n V2 = {O}. Now let <p E X be arbitrary. Then, since
5.3 Index 63
x = Ul EI1 VI, we have 'P = Ul + VI for some Ul E Ul and VI E VI' Further-

more, since VI = (VI nU2 ) EB V2 , we have VI = W2 +V2 for some W2 E VI nU2
and V2 E V2 . Then 'P = U2 + V2, where U2 := Ul + W2 E U2 . Hence we have
proven that X = U2 EI1 V2 , and this implies that codimU2 ~ codimUl if
Ul c U2·
Definition 5.8 A linear operator K : X -+ Y from a linear space X
into a linear space Y is said to have finite defect if its nullspace has finite
dimension and its range has finite codimension. The number
indK := dimN(K) - codimK(X)
is called the index of the operator K.
Note that in the proof of the following theorem we do not use an adjoint
operator, i.e., the analysis is based only on the Riesz theory.
Theorem 5.9 Let A be a compact operator on a normed space X. Then

1 - A has index zero.
Proof. Since by Theorem 3.4 the statement is obvious when 1 - A is in-

jective, we only need to consider the case where m := dimN(1 - A) E IN.
From X = N(1 - At EB (1 - At(X) and (1 - At(X) c' (I - A)(X) we
conclude that there exists a finite-dimensional subspace U of X such that
X = (I - A)(X) EI1 U. We need to show that dimU = m.
Let n := dimU and note that n :f:. 0 by Theorem 3.4. We choose bases
'PI, ... , 'Pm of N(I - A) and bl , ... , bn of U. In the case where r > 1
we choose additional elements 'Pm+!,"" 'Pm r with mr > m such that
'PI, ... ,'Pm, 'Pm+l,···, 'Pm r is a basis of N(1 - At. Then we define a linear
operator T: N(I - At -+ U by prescribing
k ~ min(m,n),
(5.11)
k > min(m, n).
By Theorem 2.6 the operator T is bounded. Consequently, since the pro-
jection operator P : X -+ N(1 - At is compact, the operator T P : X -+ U
is compact.
Now assume that m < n and let
'P - A'P + T P'P = O.
Then, since (1 - A)(X) n U = {O}, we have 'P - A'P =0 and TP'P = O.

Therefore we can write
and the definition (5.11) of T implies that

m
Lakbk = TPep = O.
k=l
From this we conclude that ak = 0, k = 1, ... , m, since the bk are linearly

independent. Hence ep = 0, and therefore the operator 1 - A + T P is injec-
tive, and consequently surjective by Theorem 3.4. Therefore the equation
is uniquely solvable. For its solution ep we conclude that TPep = bn , since

(I - A)(X) n U = {OJ. In view of the definition (5.11) of T this is a
contradiction to the linear independence of bI, ... , bn .
Now assume that m > n. Since X = (1 - A)(X) ffi U and T P(X) C U,
we can represent each f E X in the form
n
f = (1 - A+TP)ep+ Lakbk
k=l
for some ep E X and complex coefficients al,"" an' From (5.11) we con-
clude that (1 - A + T P)epk = Tepk = bk, k = 1, ... ,n, and therefore
f = (I - A + T P) (ep + ~ akepk) ,
i.e., the operator 1 - A + T P is surjective, and consequently injective by
Theorem 3.4. Since m > n, from (5.11) we have (1 -A+TP)epm = Tepm = 0,
and the injectivity of I - A + T P leads to the contradiction epm = O. 0
Theorem 5.10 Under the assumptions of Theorem 5.6 the operators K

and K' have finite defect with index
indK = dimN(K) - dimN(K') = -indK'.
Proof. Let n:= dimN(K') and choose a basis 'lj;I, ... ,'lj;n of N(K') ifn > O.
By Lemma 4.12 there exist elements bl , ••. , bn E X 2 such that
(5.12)
Define
U := span{bI, ... ,bn }.
Then K(XI)nU = {OJ by Theorem 5.6 and (5.12). Furthermore, for f E X 2
we have
n
u:= L(J,'lj;k)bk E U
k=l
and f -u E K(X I ) by Theorem 5.6. Hence X 2 = K(Xd ffiU, and therefore
codimK(XI ) = dimU = dimN(K'). 0
5.3 Index 65
Theorem 5.11 For two operators Kl and K2 satisfying the assumptions

of Theorem 5.6 we have
indK1 K 2 = indK I + indK2·
Proof. For the sake of notational brevity we confine ourselves to the case
of a dual system (X, Y) and two operators K I , K2 : X -+ X, which satisfy
the assumptions of Theorems 5.6 and 5.10, i.e., together with their adjoint
operators K~, K~ : Y -+ Y they possess regularizers. Then KIK2 : X -+ X
and its adjoint K~K~ : Y -+ Y satisfy the assumptions of Theorems 5.6
and 5.10.
Denote mj := dimN(Kj ), mj := dimN(Kj), and choose bases of the
nullspaces
N(Kj) = span{ 'Pj,l, ... , 'Pj,mj}' N(Kj) = span{'¢'j,b ... , 'l,Uj,mj}

for j = 1,2. Let 'P E N(KIK2). Then K 2'P E N(KI)' i.e.,
ml
K 2'P = L CXi'PI,i·
i=l
By Theorem 5.6 this equation is solvable if and only if

ml
L CXi('PI,i, 'l,U2,k) = 0, k = 1, ... , m;. (5.13)
i=1
By p we denote the rank of the mi x m; matrix ('PI,i, 'l,U2,k). Then the

solution space of (5.13) has dimension mi - p. Therefore
dimN(KIK2) = dimN(K2) + mi - P = m2 + mi - p.
Similarly, let 'l,U E N(K~KD. Then K~'l,U E N(K~), i.e.,
,
m2
K~'l,U = Lf3k'l,U2,k.
k=1
This equation is solvable if and only if

m2
,
L f3k('PI,i, 'l,U2,k) = 0, i = 1, ... , mI· (5.14)
k=1
The solution space of (5.14) has dimension m; - p. Therefore

dimN(K~KD = dimN(KD + m; - p = m~ + m; - p.
In view of Theorem 5.10, combining the two results yields
indKI K 2 = (mi - m~) + (m2 - m;) = indKI + indK2,
and the proof is complete. o
Corollary 5.12 Under the assumptions of Theorem 5.6 the index is stable
with respect to compact perturbations, i. e., for compact operators C with a
compact adjoint C' we have
ind(K + C) = indK.
Proof. Let K and K' be adjoint operators with adjoint regularizers R and
R'. Then, since RK = ] - A, where A is compact, Theorems 5.9 and 5.11
imply that
indR + indK = indRK = ind(] - A) = 0,
i.e., ind K = - ind R. For a compact operator C the operator R also regu-
larizes K + C and the operator R' regularizes K' + C'. Therefore
ind(K + C) = - indR = indK,

For the history of the development of the notion of the index of an
operator we refer to [35].
Of course, this chapter can provide only a first glance into the theory
of singular operators. For a detailed study, in the canonical dual system
(X, X*), we refer to the monograph by Mikhlin and Prossdorf [124].
Problems
5.1 Show that the transformations of the Volterra integral equation of the first
kind (3.6) into the Volterra equations of the second kind (3.7) and (3.8) can be
interpreted as regularizations from the left and from the right, respectively.
Hint: Use the space Gl[a, b] of continuously differentiable functions furnished with
the norm IIcplh := IIcplI"" + IIcp'II",,·
5.2 Convince yourself where in the proof of Theorem 5.6 use is made of the
fact that the operators K and K' possess regularizers from the left and from the
right.
5.3 Use Theorem 5.9 for an alternative proof of Theorem 4.15.
5.4 Let Xl,X2 be Banach spaces, let K: Xl -+ X 2 be a bounded operator, and
let R : X2 -+ Xl be a left (right) regularizer of K. Show that for all operators
G : Xl -+ X2 with IIGII < IIRII the operator K + G has a left (right) regularizer.
5.5 Use Problem 5.4 to show that in Banach spaces under the assumptions of
Theorem 5.6 the index is stable with respect to small perturbations, i.e., there
exists a positive number 'Y (depending on K and K') such that
ind(K + G) = ind K
for all operators G with adjoint G' satisfying max(IIGII, IIG'II) < 'Y (see [7, 34]).
6
Potential Theory
The solution of boundary value problems for partial differential equations

is one of the most important fields of applications for integral equations.
About a century ago the systematic development of the theory of inte-
gral equations was initiated by the treatment of boundary value problems
and there has been an ongoing fruitful interaction between these two ar-
eas of applied mathematics. It is the aim of this chapter to introduce the
main ideas of this field by studying the basic boundary value problems of
potential theory. For the sake of simplicity we shall confine our presenta-
tion to the case of two and three space dimensions. The extension to more
than three dimensions is straightforward. As we shall see, the treatment
of the boundary integral equations for the potential theoretic boundary
value problems delivers an instructive example for the application of the
Fredholm alternative, since both its cases occur in a natural way.
6.1 Harmonic Functions

We begin with a brief outline of the basic properties of harmonic functions
going back to the early development of potential theory at the beginning of
the 19th century with contributions by Dirichlet, Gauss, Green, Riemann
and Weierstrass. For a more comprehensive study of potential theory we
refer to Courant and Hilbert [27), Folland [41), Helms [66], Kellogg [82],
Martensen [118], and Mikhlin [123).

68 6. Potential Theory
Definition 6.1 A twice continuously differentiable real-valued function u,

defined on a domain D C IRm, m = 2,3, is called harmonic if it satisfies
Laplace's equation
~u=O in D,
where
m {)2u
~u:=L {) 2 •
j=l
Xj
Harmonic functions describe time-independent temperature distributions,

potentials of electrostatic and magnetostatic fields, and velocity potentials
of incompressible irrotational fluid flows.
There is a close connection between harmonic functions in IR2 and holo-
morphic functions in <C. From the Cauchy-Riemann equations we readily
observe that both the real and imaginary parts of a holomorphic function
J(z) = U(XI,X2) + iv(xl,x2), z = Xl + iX2, are harmonic functions.
Most of the basic properties of harmonic functions can be deduced from
the fundamental solution that is introduced in the following theorem. Recall
that by Ixi we denote the Euclidean norm of a vector X E IRm.
Theorem 6.2 The function
1 1
- In--- m=2,
271" Ix - yl '
1 1
m=3,
471" Ix - yl '
is called the fundamental solution of Laplace's equation. For fixed y E IRm
it is harmonic in IRm \ {y}.
Proof. This follows by straightforward differentiation. o

For n E lN, by en(D) we denote the linear space of real- or complex-
valued functions defined on the domain D, which are n times continuously
differentiable. By en(D) we denote the subspace of all functions in en(D),
which with all their derivatives up to order n can be extended continuously
from D into the closure D. In this chapter, we mostly deal with real-
valued functions but with proper interpretation our results remain valid for
complex-valued functions. From p. 25 we recall what is meant by saying a
bounded domain D or its boundary aD belong to class en for n E IN.
One of the basic tools in studying harmonic functions is provided by
Green's integral theorems. Recall that for two vectors a = (at. ... , am) and
b = (b l , ... , bm ) in IRm we denote by a . b = alb l + ... + ambm the dot
product.
6.1 Harmonic Functions 69
Theorem 6.3 (Green's Theorem) Let D be a bounded domain of class

C 1 and let 1/ denote the unit normal vector to the boundary aD directed into
the exterior of D. Then, for u E C 1 (f)) and v E C 2 (f)) we have Green's
first theorem
! D
{uL.v+gradu.gradv}dx = r
laD
u ~v
ul/
ds (6.1)
and for u, v E C 2 (f)) we have Green's second theorem
L (u L. v - v L. u) dx = faD (u ~~ - v ~~) ds. (6.2)
Proof. We apply Gauss' divergence theorem
rdiv A
lD
dx = r
laD
1/' A ds
to the vector field A E Cl(f)) defined by A:= ugradv and use
dive u grad v) = grad u . grad v + u div grad v

to establish (6.1). To obtain (6.2) we interchange u and v and then sub-
tract. D
Note that our regularity assumptions on D are sufficient conditions for

the validity of Gauss' and Green's theorems and can be weakened. In par-
ticular, the boundary can be allowed to have edges and corners. For a
detailed study, see, for example, [123, 130].
Corollary 6.4 Let v E C 2 (fJ) be harmonic in D. Then
1 aD ul/
av
-;:;- ds = O. (6.3)
Proof. This follows by choosing u = 1 in (6.1). D
Theorem 6.5 (Green's Formula) Let D be as in Theorem 6.3 and let

u E C 2 (fJ) be harmonic in D. Then
u(x) = laD {~~ (y) <t>(x, y) - u(y) a:~(;~)} ds(y), xED. (6.4)
Proof. For xED we choose a sphere D(x; r) := {y E JR'm : Iy - xl = r}

of radius r such that D(x;r) c D and direct the unit normal 1/ to D(x;r)
into the interior of D(x;r). Now we apply Green's second theorem (6.2) to
the harmonic functions u and <t>(x,') in the domain {y ED: Iy - xl > r}
to obtain
1 aDUr!(x;r)
{
u(y)
a<t>(x, y) au }
a () - -;:;- (y) <t>(x, y) ds(y) = O.
1/ Y ul/
Since on n(x; r) we have
v(y)
grady ~(x, y) = wmrm- 1 ' (6.5)
where W2 = 27l', W3 = 47l', a straightforward calculation, using the mean

value theorem and (6.3), shows that
lim f
r-tO In(x;r)
{u(y) a:(~,))
v y
- aau
v
(y)~(x,Y)}dS(Y) =u(x),
whence (6.4) follows. D
From Green's formula we can conclude that harmonic functions are an-
alytic functions of their independent variables.
Theorem 6.6 Harmonic functions are analytic.

Proof. We make use of the fact that each holomorphic function of several
complex variables, i.e., a function satisfying the Cauchy-Riemann equa-
tions with respect to each of the complex variables is also analytic, i.e., it
has a local power series expansion and vice versa (see [50]). The theorem
now follows from the observation that the fundamental solution ~(x, y) is
an analytic function of the Cartesian coordinates Xj, j = 1, ... , m, of x
and the fact that the integrands in (6.4) and their derivatives with respect
to x are continuous with respect to y if x is contained in a compact subset
of D. Therefore the Cauchy-Riemann equations for u can be verified by
differentiating with respect to x under the integral. D
From Theorem 6.6 it follows that a harmonic function that vanishes in

an open subset of its domain of definition must vanish identically.
Theorem 6.7 (Mean Value Theorem) Let u be harmonic in an open

ball B(x;r) = {y E lRm : Iy-xl < r} with boundary n(x;r) and continuous
in the closure B[x; rJ. Then
u(x) = mm f u(y) dy = 1m _ 1 f u(y) ds(y), (6.6)

wmr JB[x;r] wmr In(x;r)
i.e., the value of u at the center of the ball is equal to the integral mean
values over both the ball and its boundary surface (W2 = 27l', W3 = 47l').
Proof. For each 0 < p < r we have u E C 2 (B[x;p]) and can apply (6.3) and
(6.4) with the result
u(x) = 1m _ 1 f u(y) ds{y), (6.7)

WmP J1y-xl=p
whence the second mean value formula follows by passing to the limit
p ~ r. Multiplying (6.7) by pm-1 and integrating with respect to p from 0
to r we obtain the first mean value formula. 0
Theorem 6.S (Maximum-Minimum Principle) A harmonic function

on a domain cannot attain its maximum or its minimum unless it is
constant.
Proof. It suffices to carry out the proof for the maximum. Let u be a
harmonic function in the domain D and assume that it attains its max-
imum value in D, i.e., the set DM := {x ED: u(x) = M} where
M := sUPxED u(x) is not empty. Since u is continuous, DM is closed rela-
tive to D. Let x be any point in DM and apply the mean value Theorem
6.7 to the harmonic function M - u in a ball B(xjr) with B[xjr) c D.
Then
O=M-u(x)=
wmr
mm r
lB[x;r)
{M-u(y)}dy,
so that u = M in B(xj r). Therefore DM is open relative to D. Hence

D = D M , i.e., u is constant in D. 0
Corollary 6.9 Let D be a bounded domain and let u be harmonic in D

and continuous in jj. Then u attains both its maximum and its minimum
on the boundary.
For the study of exterior boundary value problems we also need to in-
vestigate the asymptotic behavior of harmonic functions as Ixi ~'OO. To
this end we extend Green's formula to unbounded domains.
Theorem 6.10 Assume that D is a bounded domain of class C 1 with a

connected boundary aD and outward unit normal v and let u E C2(lRm\D)
be a bounded harmonic function. Then
u(X) = U oo
r {u(y)
+ laD aq,(x, y) au }
av(y) - av (y) q,(x, y) ds(y) (6.8)
for x E lRm \ D and some constant U oo • For m = 2, in addition,
r
laD ov
auds=O (6.9)
and the mean value property at infinity
U oo = -2
1 r
1rr llyl=r
u(y) ds(y) (6.10)
for sufficiently large r is satisfied.

Proof. Without loss of generality we may assume that the origin x = 0 is

contained in D. Since u is bounded, there exists a constant M > 0 such
that lu(x)1 ~ M for all x E IRm \ D. Choose Ro large enough to ensure
that y E IRm \ D for all Iyl ~ Ro/2. Then for a fixed x with Ixl ~ Ro we
can apply the mean value Theorem 6.7 to the components of grad u. From
this and Gauss' integral theorem we obtain
gradu(x)= mm
wmr
r
1B[x;rJ
gradu(y)dy=-
wmr
mm r
In(x;r)
v(y)u(y)ds(y),
where v is the unit normal to 0 (Xi r) directed into the interior of O( Xi r)

and where we choose the radius to be r = Ixl/2. Then we can estimate
mM 2mM
Igradu(x)1 ~ -r- = TxI (6.11)
for all Ixl ~ R o·

For m = 2, we choose r large enough such that Or := 0(0; r) is contained
in lR? \ iJ and apply Green's second theorem (6.2) to u and cI>(0,·) in the
annulus r < Iyl < R and use (6.5) to obtain
~
r
r uds -In ~r r auav ds =..!..R r uds -In..!..R r auav ds.
lnr lnr lnR lnR
(Note that v is the interior normal to Or and OR.) From this, with the aid
of Corollary 6.4 applied in the annulus between aD and Or, we find
11
-
r nr
uds+ln- 11
r
au
l'lds=R
aD uV
11 nR
uds+ln R 11 aD uV
au
l'lds. (6.12)
Since the first term on the right-hand side is bounded by 21r M, letting
R -+ 00 in (6.12) implies that the integral in (6.9) must be zero. Note that
(6.9) only holds in the two-dimensional case and is a consequence of the
fact that in IR2 the fundamental solution is not bounded at infinity.
For x E IRm \ iJ, m = 2,3, we now choose r large enough such that
jj C B(x; r). Then by Green's formula (6.4), applied in the domain between
aD and O(x; r), we have that
u(x) = r
laDun(x;r)
{U(y) a:«(,)) _
v y
~u (Y)cI>(x,Y)}dS(Y).
uV
(6.13)
With the aid of Corollary 6.4 we find
1 n(x;r)
8u (y) cI>(x, y) ds(y) = -4
-8
v
l1
1rr aD uV
au
l'l (y) ds(y) -+ 0, r -+ 00,
ifm = 3, and
1 n(x;r) uV
au
l'l
1 11
(y) cI>(x, y) ds(y) = -2 In -
1r r aD uV
au
l'l (y) ds(y) = °
if m = 2, where we have made use of (6.9). With the aid of (6.5) we can
write
f
Jf!(x;r)
u(y) o:(~'
v y
f) ds(y) =
wmr
Im _ 1 f
J1y-xl=r
u(y) ds(y).
From the mean value theorem we have
u(x + y) - u(y) = grad u(y + Ox) . x

for some 0 E [0,1], and using (6.11) we can estimate
2mMlxl
lu(x + y) - u(y)1 ~ Iyl- Ixl
provided that Iyl is sufficiently large. Therefore
wmr
Im _ 1 IJ1y-xl=r
f u(y) ds(y) - f
J1yl=r
u(y) dS(y)1 ~ Cr
for some constant C > 0 depending on x and all sufficiently large r. Now
choose a sequence (rn) of radii with rn --+ 00. Since the integral mean values
J.Ln := Im _ 1 f u(y) ds(y)

wmrn J1yl=r n
are bounded through lJ.Lnl ~ M, n E lN, by the Bolzano-Weierstrass the-
orem we may assume that the sequence (J.Ln) converges, i.e., J.Ln --+ u oo ,
n --+ 00, for some U oo E JR. From this, in view of the above estimates, we
now have that
[ { u(y) oq>(x,y) ou }
0 () - -;;) (y) q>(x, y) ds(y) --+ u oo , n --+ 00.
f!(x;rn) V Y uV
Hence (6.8) follows by setting r = rn in (6.13) and passing to the limit

n --+ 00. Finally, (6.10) follows by setting R = rn in (6.12), passing to the
limit n --+ 00, and using (6.9). 0
From (6.8), using the asymptotic behavior of the fundamental solution
(6.14)
and
oq>(x,y) _
OXj -
0(_1_)
Ixl m- 1 '
(6.15)
for Ixl -t 00, which holds uniformly for all directions x/lxl and all y E aD,
and the property (6.9) if m = 2, we can deduce that bounded harmonic
functions in an exterior domain satisfy
u(x) = u oo + 0 C~I)' grad u(x) =0 Cxl~-l ), Ixl-t 00, (6.16)
uniformly for all directions.
6.2 Boundary Value Problems: Uniqueness

Green's formula (6.4) represents any harmonic function in terms of its
boundary values and its normal derivative on the boundary, the so-called
Cauchy data. In the subsequent analysis we shall see that a harmonic func-
tion is already completely determined by either its boundary values or, up
to a constant, its normal derivative alone.
In the sequel, let D C lRm be a bounded domain of class C2. For the
sake of simplicity for the rest of this chapter we assume that the boundary
aD is connected. Again by v we denote the unit normal of aD directed
into the exterior domain lRm \ D.
Interior Dirichlet Problem. Find a function u that is harmonic in D,

is continuous in D, and satisfies the boundary condition
u= f on aD,
where f is a given continuous function.
Interior Neumann Problem. Find a function u that is harmonic in D,

is continuous in D, and satisfies the boundary condition
au =g
- onaD
av
in the sense
lim v(x)· grad u(x - hV(x)) = g(x), x E aD,

h-++O
of uniform convergence on aD, where 9 is a given continuous function.
Exterior Dirichlet Problem. Find a function u that is harmonic in

lRm\ D, is continuous in lRm \ D, and satisfies the boundary condition
u= f on aD,
6.2 Boundary Value Problems: Uniqueness 75
where f is a given continuous function. For Ixl -+ 00 it is required that

u(x) = 0 (1), m = 2, and u(x) = 0 (1), m = 3,
uniformly for all directions.
Exterior Neumann Problem. Find a function U that is harmonic in

lRm\ 15, is continuous in lRm \ D, and satisfies the boundary condition
au
-= 9 on aD
av
in the sense of uniform convergence on aD, where 9 is a given continuous
function. For Ixl -+ 00 it is required that u(x) = 0 (1) uniformly for all
directions.
Note that for the exterior problems we impose that U oo = 0, with the
exception of the Dirichlet problem in lR2 , where u is only required to be
bounded.
These boundary value problems carry the names of Dirichlet, who made
important contributions to potential theory, and Neumann, who gave the
first rigorous existence proof (see Problem 6.5). From the numerous appli-
cations we mention:
(1) Determine the stationary temperature distribution in a heat-conduct-
ing body from the temperature on the boundary or from the heat flux
through the boundary.
(2) Find the potential of the electrostatic field in the exterior of a perfect
conductor.
(3) Find the velocity potential of an incompressible irrotational flow
around an obstacle.
Our aim is to establish that each of the above potential theoretic bound-
ary value problems has a unique solution depending continuously On the
given boundary data, i.e., they are well-posed in the sense of Hadamard
(see Section 15.1).
In our uniqueness proofs we need to apply Green's Theorem 6.3. Since for
solutions to the boundary value problems we do not assume differentiability
up to the boundary, we introduce the concept of parallel surfaces. These
are described by
aDh := {z = x + hv(x) : x E aD},
with a real parameter h. Because aD is assumed to be of class C 2 , we
observe that aDh is of class Cl. For m = 3, let x( u) = (Xl(U), X2( u), X3(U)),
u = (Ub U2), be a regular parametric representation of a surface patch of
aD . Then straightforward differential geometric calculations show that the
determinants
aX ax]
g( u) := det [ aUi' aUj and g(u;h):= det _ . -
aUi aUj
[ az az ]
are related by
g(u; h) = g(u){1- 2hH(u) + h 2K(U)}2,
where Hand K denote the mean and Gaussian curvature of aD, respec-
tively (see [118, 130]). This verifies that the parallel surfaces are well defined
provided the parameter h is sufficiently small to ensure that 1- 2hH + h 2K
remains positive. This also ensures that in a sufficiently small neighborhood
of aD each point z can be uniquely represented in the form z = x + hv( x),
where x E aD and h E JR.
In particular, the surface elements ds on aD and dSh on aD h are related
by
dSh(Z) = {I - 2hH + h 2K}ds(x). (6.17)
Since vex) . vex) = 1, we have
av(x) . vex) = 0, i = 1,2,

aUi
for all x E aD, and therefore the tangential vectors
for all (sufficiently small) h lie in the tangent plane to aD at the point
x, i.e., the normal vector Vh(Z) of the parallel surface aDh coincides with
the normal vector vex) of aD for all x E aD. Hence, in view of (6.17),
Theorems 6.5 and 6.10 remain valid for harmonic functions U E C(D) and
u E C(JRm \ D), respectively, provided they have a normal derivative in
the sense of uniform convergence.
Note that in two dimensions the equation (6.17) has to be replaced by
dsh(z) = (1 - /'l,h)ds(x), where /'l, denotes the curvature of aD, i.e., for the
representation aD = {xes) : So ~ s ~ SI} in terms of the arc length we
have /'l, = v . x" .
Theorem 6.11 Both the interior and the exterior Dirichlet problems have
at most one solution.
Proof. The difference U := UI - U2 of two solutions to the Dirichlet problem

is a harmonic function that is continuous up to the boundary and satis-
fies the homogeneous boundary condition U = 0 on aD. Then, from the
maximum-minimum principle of Corollary 6.9 we obtain U = 0 in D for
the interior problem, and observing that u(x) = 0(1), Ixl ~ 00, we also
obtain u = 0 in JR3 \ D for the exterior problem in three dimensions.
For the exterior problem in two dimensions, by the maximum-minimum
principle Theorem 6.8 the supremum and the infimum of the bounded
harmonic function u are either attained on the boundary or equal to u oo .
When the maximum and minimum are both attained on the boundary then
6.2 Boundary Value Problems: Uniqueness 77
from the homogeneous boundary condition we immediately have u = 0 in

IRm \ D. If the supremum is equal to U oo , then from u{x) S U oo for all
x E IR2 \ D and the mean value property (6.1O) we observe that u = U oo
in the exterior of some circle. Now we can apply the maximum principle to
see that u = U oo in all of IR2 \ D and the homogeneous boundary condition
finally implies u = 0 in IR2 \ D. The case where the infimum is equal to U oo
is settled by the same argument. 0
Theorem 6.12 Two solutions of the interior Neumann problem can dif-
fer only by a constant. The exterior Neumann problem has at most one
solution.
Proof. The difference u := Ul - U2 of two solutions for the Neumann prob-
lem is a harmonic function continuous up to the boundary satisfying the
homogeneous boundary condition aujav = 0 on aD in the sense of uni-
form convergence. For the interior problem, suppose that u is not con-
stant in D. Then there exists some closed ball B contained in D such that
fB Igrad ul 2 dx > O. From Green's first theorem (6.1), applied to the interior
Dh of some parallel surface aDh := {x - hv{x) : x E aD} with sufficiently
small h > 0, we derive
f Igrad ul 2 dx S f Igrad ul 2 dx = f u ~u ds.

lB lDh laDh v
Passing to the limit h -+ 0, we obtain the contradiction fB Igrad ul 2 dx SO.
Hence, u must be constant.
For the exterior problem, assume that grad u i- 0 in IRm \ D. Then
again, there exists some closed ball B contained in IRm \ fJ such that
fB Igrad ul 2 dx > O. From Green's first theorem, applied to the domain
Dh,r between some parallel surface aDh := {x + hv{x) : x E aD} with
sufficiently small h > 0 and some sufficiently large sphere Or of radius r
centered at the origin (with interior normal v), we obtain
f Igrad ul 2 dx S f Igrad ul 2 dx = - f u ~~ ds - f u ~~ ds.

lB lDh,r lOr laDh
Letting r -+ 00 and h -+ 0, with the aid of the asymptotics (6.16), we
arrive at the contradiction fBI grad ul 2 dx S O. Therefore, u is constant in
IRm \ fJ and the constant must be zero, since U oo = O. 0
From the proofs it is obvious that our uniqueness results remain valid
under weaker regularity conditions on the boundary. Uniqueness for the
Dirichlet problem via the maximum-minimum principle needs no regularity
of the boundary, and uniqueness for the Neumann problem holds for those
boundaries for which Green's integral theorem is valid. We have formulated
the boundary value problems for C2 boundaries, since we shall establish
the existence of solutions under these conditions.
6.3 Surface Potentials

Definition 6.13 Given a function <p E C(aD), the functions
u(x):=
laD
r <p(y)<I?(x, y) ds(y), x E lRm \ aD, (6.18)
and
( a<I?(x, y)
v(x):= laD <p(y) av(y) ds(y), x E lRm\aD, (6.19)
are called, respectively, single-layer and double-layer potential with density

<po In two dimensions, occasionally, for obvious reasons we will call them
logarithmic single-layer and logarithmic double-layer potential.
For fixed y E lRm the fundamental solution u = <I?(.,y) represents the
potential of a unit point source located at the point y, i.e., grad x <I?(x, y)
gives the force-field of this point source acting at the point X. The single-
layer potential is obtained by distributing point sources on the boundary
aD. For h > 0, by the mean value theorem we have
<I?(x, y + hv(y)) - <I? (x, y - hv(y)) = 2h /I(y) . grad <I? (x, y + ()hv(y))
for some () = ()(y) E [-1,1]. Therefore, the double-layer potential can be
interpreted as the limit h -t 0 of the superposition of the single-layer po-
tentials Uh and U-h with densities <pj2h on aDh and -<pj2h on aD_h, re-
spectively, i.e., the double-layer potential is obtained by distributing dipoles
on the boundary aD.
Since for points x ft aD we can interchange differentiation and integra-
tion, the single- and double-layer potentials represent harmonic functions in
D and lRm \ D. For the solution of the boundary value problems we need
to investigate the behavior of the potentials at the boundary aD where
the integrals become singular. The boundary behavior is expressed by the
following so-called jump relations.
Theorem 6.14 Let aD be of class C 2 and <p E C(aD). Then the single-
layer potential u with density <p is continuous throughout lRm. On the
boundary we have
u(x) =
laD
r <p(y)<I?(x,y)ds(y), x E aD, (6.20)
where the integral exists as an improper integral.
Proof. Analogous to the proofs of Theorems 2.22 and 2.23, by using the
cut-off function h, it can be shown that the single-layer potential u is the
uniform limit of a sequence of functions Un that are continuous in lRm. 0
6.3 Surface Potentials 79
For the further analysis of the jump relations we need the following
lemma. The inequality (6.21) expresses the fact that the vector x - y for x
close to y is almost orthogonal to the normal vector v(x).
Lemma 6.15 Let aD be of class C2. Then there exists a positive constant
L such that
Iv(x). {x - y}1 :::; Llx _ Yl2 (6.21)
and
Iv(x) - v(y)1 :::; Llx - yl (6.22)
for all x, y E aD.
Proof. We confine ourselves to the two-dimensional case. For the three-
dimensional case we refer to [24]. Let r = {x(s) : s E [0, so]} be a regular
°
parameterization of a patch reaD, i.e., x : [0,1] -+ reaD is injective
and twice continuously differentiable with x'(s) =1= for all s E [0, so]. Then,
by Taylor's formula we have
1
Iv(x(t))· {x(t) - x(r)}1 :::; - max IX"(S)llt - r12,
2 OSsSso
Iv(x(t)) - v(x(r))1 :::; o~~~o I! v(x(s))llt - rl,
Ix(t)) - x(r)1 ~ min Ix'(s)llt - rl.

OSsSso
The statement of the lemma is evident from this. o
Example 6.16 For the double-layer potential with constant density we
have
1=::
XED,
2 ( 8<J>(x, y) ds(y) = xE8D, (6.23)

JOD av(y)
0,
This follows for x E IRm \ [) from (6.3) applied to <J>(x,·) and for xED
from (6.4) applied to u = 1 in D. The result for x E aD is derived by
excluding x from the integration by circumscribing it with a sphere D(x; r)
of radius r and center x with the unit normal directed toward the center.
Let H(x; r) := D(x; r) n D. Then, by (6.3) applied to <J>(x, .), we have
( a<J>(x, y) ds(y) + ( 8<J>(x, y) ds(y) = 0,
J{YEOD:ly-xl?r} 8v(y) JH(x;r) av(y)
. 1 1
and from
hm2 a(x, y) dsy=hm

(). 2 1 dsy=1
( )
r--+O H(x;r) av(y) r--+O wmr m - H(x;r)
the result follows. 0

Theorem 6.17 For aD of class C2, the double-layer potential v with con-
tinuous density r.p can be continuously extended from D to fJ and from
:rn.m \ jj to :rn.m \ D with limiting values
( 8(x, y) 1
v±(x) = laD r.p(y) 8v(y) ds(y) ±"2 r.p(x), xE8D, (6.24)
where
v±(x):= lim vex ± hv(x))
h-++O
and where the integral exists as an improper integral.
Proof. Because of Lemma 6.15 we have the estimate
8 (X'Y)I=lv(Y).{X- Y}I< L =I- (6.25)
1
8v(y) wmlx - ylm - wmlx _ ylm-2' x y,
i.e., the integral in (6.24) has a weakly singular kernel. Therefore, by The-
orem 2.23 the integral exists for x E 8D as an improper integral and
represents a continuous function on 8D.
As pointed out on p. 76, in a sufficiently small neighborhood U of 8D
we can represent each x E U uniquely in the form x = z + hv(z), where
z E 8D and h E [-ho, hol for some ho > O. Then we write the double-layer
potential v with density r.p in the form
vex) = r.p{z)w{x) + u{x), x = z + hv{z) E U \ 8D,
where
( 8 {x, y)
w{x) := laD 8v(y) ds(y)
and
( 8(x,y)
u{x):= laD {r.p(y) - r.p(z)} 8v{y) ds{y). (6.26)
For x E 8D, i.e., for h = 0, the integral in (6.26) exists as an improper

integral and represents a continuous function on 8D. Therefore, in view of
Example 6.16, to establish the theorem it suffices to show that
lim u{z + hv{z))
h-+O
= u(z), z E 8D,
uniformly on 8D.
From (6.21) we can conclude that
1
Ix - Yl2 ~ "2 {Iz - Yl2 + Ix - Z12}
for x = z + hv{z) and h E [-ho,hol provided that ho is sufficiently small.

Therefore, writing
8{x,y)
- v{y)·{z-y}
--:-~...::....:...
v(y)·{x-z}
+ --"'-'---;---"----;--..e...
av{y) - wmlx - ylm wmlx _ ylm '
6.3 Surface Potentials 81
and again using (6.21), we can estimate
1
8 (x,y)
8v(y) S;
I C{I Ix-zl
Ix _ ylm-2 + [lz _ Yl2 + Ix - zI2]m/2
I
}
for some constant C I > o. Recalling the proof of Theorem 2.23 and denoting
8D(z; r) := 8DnB[z; r], for sufficiently small r we project onto the tangent
plane and deduce that
( I8(x, y) I {r r Ix - zlpm-2 dp }
JaD(z;r) 8v(y) ds(y) S; C I Jo dp + Jo (p2 + Ix _ zI2)m/2
(6.27)
S;CI
{r+Joroo (),2+1)m/2·
),m-2 d)' }
From the mean value theorem we obtain that

8 (x,y) _ 8(z,y) I < C2 Ix-zl
1
8v(y) 8v(y) - Iz - ylm
for some constant C 2 > 0 and 21x - zl S; Iz - YI. Hence we can estimate
( 1 I
8 (x,y) _ 8(z,y) ds(y) < C3 Ix-zl (6.28)
JaD\aD(z:r) 8v(y) 8v(y) - rm
for some constant C3 > 0 and Ix - zl S; r/2. Now we can combine (6.27)
and (6.28) to find that
lu(x) - u(z)1 S; C { max l<p(y) _ <p(z)1 + Ix - zl}

ly-zl:Sr rm
for some constant C> 0, all sufficiently small r, and Ix - zl S; r/2. Given
E > 0 we can choose r > 0 such that
E
max l<p(y) - <p(z)1 S; 2C
Iy-zl::;r
for all z E 8D, since <p is uniformly continuous on 8D. Then, taking
8 < Er m /2C, we see that lu(x) - u(z)1 < E for all Ix - zl < 8, and the
proof is complete. 0
Theorem 6.18 Let 8D be of class C 2. Then for the single-layer potential

u with continuous density (x, y) 1
8v (x) = J aD <p(y) 8v(x) ds(y) =F "2 <p(x), x E8D, (6.29)
where
8:±
uV
(x):= lim v(x)· grad u(x ± hV(x))
h--++O
is to be understood in the sense of uniform convergence on 8D and where
the integral exists as an improper integral.
Proof. Let v denote the double-layer potential with density r.p and let U be
as in the proof of Theorem 6.17. Then for x = Z + hll(z) E U \ aD we can
write
II(Z) . grad u(x) + vex) = r {1I(Y) - lI(z)} . grady cI>(x, y) r.p(y) ds{y),
l&D
where we have made use of grad x tl>(x, y) = - grady tl>(x, y). Using (6.22),
analogous to the single-layer potential in Theorem 6.14, the right-hand side
can be seen to be continuous in U. The proof is now completed by applying
Theorem 6.17. 0
Theorem 6.19 Let aD be of class C 2 . Then the double-layer potential v

with continuous density r.p satisfies
lim lI(x), {grad vex + hll(x)) - grad vex - hll(x))}

h~+O
=0 (6.30)
uniformly for all x E aD.
Proof. We omit the rather lengthy proof, which is similar in structure to

the proof of Theorem 6.17. For a detailed proof we refer to [24]. 0
6.4 Boundary Value Problems: Existence

Green's formula shows that each harmonic function can be represented ru;
a combination of single- and double-layer potentials. For boundary value
problems we try to find a solution in the form of one of these two potentials.
To this end we introduce two integral operators K, K' : C(aD) --t C(aD)
r r.p(y) atl>(x,
by
y)
(Kr.p)(x) := 2 l&D all(Y) ds(y) , x E aD, (6.31)
and
(6.32)
Because of (6.25) the integral operators K and K' have weakly singular
kernels and therefore are compact by Theorem 2.23. Note that in two di-
mensions for C 2 boundaries the kernels of K and K' actually turn out to be
continuous (see Problem 6.1). As seen by interchanging the order of integra-
tion, K and K' are adjoint with respect to the dual system (C(aD), C(aD))
defined by
(r.p, 't/J):= r r.p't/J ds,
l&D r.p, 'l/J E C(aD).
6.4 Boundary Value Problems: Existence 83
Theorem 6.20 The operators J - K and J - K' have trivial nullspaces

N{J - K) = N{J - K') = {a}.
The nullspaces of the operators J + K and J + K' have dimension one and
N(J + K) = span{l}, N(I + K') = span{7Po}
with
{ '¢o ds i= 0,
laD
i. e., the Riesz number is one.
Proof Let cp be a solution to the homogeneous equation cp - K cp = 0
and define a double-layer potential v by (6.19). Then by (6.24) we have
2v_ = K cp - cp = 0 and from the uniqueness for the interior Dirichlet
problem (Theorem 6.11) it follows that v = 0 in D. From (6.30) we see
that av+lall = 0 on aD, and since vex) = 0(1), Ixl -t 00, from the
uniqueness for the exterior Neumann problem (Theorem 6.12) we find that
v = 0 in IRm \ D. Hence, from (6.24) we deduce cp = v+ - v_ = 0 on aD.
Thus N(J - K) = {O} and, by the Fredholm alternative, N(J - K') = {a}.
Now let cp be a solution to cp + Kcp = 0 and again define v by (6.19).
Then by (6.24) we have 2v+ = K cp + cp = 0 on aD. Since vex) = 0 (1),
Ixl -t 00, from the uniqueness for the exterior Dirichlet problem it follows
that v = 0 in IRm \ D. From (6.30) we see that av_Iall = 0 on aD and
from the uniqueness for the interior Neumann problem we find that v is
constant in D. Hence, from (6.24) we deduce that cp is constant on aD.
Therefore, N(J + K) c span{l}, and since by (6.23) we have 1 + K1 = 0,
it follows that N(J + K) = span{l}.
By the Fredholm alternative, N(J+K') also has dimension one. Therefore
N(J+K') = span{7Po} with some function '¢O E C(aD) that does not vanish
identically. Assume that (1, '¢o) = 0 and define a single-layer potential u
with density '¢O. Then by (6.20) and (6.29) we have
au_ -0 au+
all - , and all = -'¢o on aD (6.33)
in the sense of uniform convergence. From au_lOll = 0 on aD, by the

uniqueness for the interior Neumann problem (Theorem 6.12), we conclude
that u is constant in D. Assume that u is not constant in IRm\D. Then there
exists a closed ball B contained in IRm \ D such that IB Igrad ul 2 dx > O.
By Green's theorem (6.1), using the jump relations (6.33), the assumption
(1, '¢o) = 0 and the fact that u+ is constant on aD, we find
1I
B
grad ul 2 dx :::; - [ auU
Or
£l ds -
vII
1aD
au+ ds
u+ ~
vII
= - [
Or
aU ds +
u £l
vII
1aD
u+ '¢o ds = - [
Or
au ds
u £l
vII
where Or denotes a sphere with sufficiently large radius r centered at the

origin (and interior normal v). With the help of faD 1/Jods = 0, using (6.14)
and (6.15), it can be seen that u has the asymptotic behavior (6.16) with
u oo = O. Therefore, passing to the limit r -+ 00, we arrive at the contradic-
tion fB Igrad ul 2dx :S O. Hence, u is constant in lRm\D and from the jump
relation (6.33) we derive the contradiction 1/Jo = O. Therefore, (l,1/Jo) =I- O.
The statement on the Riesz number is a consequence of Problem 4.4. 0
Theorem 6.21 The double-layer potential
u(x)
r
= laD cp(y)
a(x, y)
av(y) ds(y), xED, (6.34)
with continuous density cp is a solution of the interior Dirichlet problem

provided that cp is a solution of the integral equation
r a(x,y)
cp(x) - 2 laD cp(y) av(y) ds(y) = -2f(x), x E aD. (6.35)
Proof. This follows from Theorem 6.17. o

Theorem 6.22 The interior Dirichlet problem has a unique solution.
Proof. The integral equation cp - Kcp = -2f of the interior Dirichlet prob-
lem is uniquely solvable by Theorem 3.4, since N(J - K) = {O}. 0
From Theorem 6.14 we see that in order to obtain an integral equation of

the second kind for the Dirichlet problem it is crucial to seek the solution
in the form of a double-layer potential rather than a single-layer potential,
which would lead to an integral equation of the first kind. Historically, this
important observation goes back to Beer [14].
The double-layer potential approach (6.34) for the exterior Dirichlet
problem leads to the integral equation cp + K cp = 2f for the density cp.
Since N(J +K') = span{1/Jo}, by the Fredholm alternative, this equation is
solvable if and only if (I,1/Jo) = O. Of course, for arbitrary boundary data
f we cannot expect this condition to be satisfied. Therefore we modify our
approach as follows.
Theorem 6.23 The modified double-layer potential
r
u(x) = laD cp(y)
{a(x, y)
av(y)
I} ds(y),
+ Ixl m - 2 (6.36)
with continuous density cp is a solution to the exterior Dirichlet problem

provided that cp is a solution of the integral equation
r
cp(x) + 2 laD cp(y)
{a(x, y)
av(y) +
I} ds(y)
Ixl m - 2 = 2f(x), x E aD. (6.37)
Here, we assume that the origin is contained in D.

6.4 Boundary Value Problems: Existence 85
Proof. This again follows from Theorem 6.17. Observe that u has the
required behavior for Ixl ---+ 00, namely, u(x) = 0 (1) if m = 2 and
u(x)=o(1)ifm=3. D
Theorem 6.24 The exterior Dirichlet problem has a unique solution.

Proof. The integral operator K : C(aD) ---+ C(aD) defined by
-
(K<p)(x)
r <p(y) {a(x,y)
:= 2 laD av(y) + Ixl
I} ds(y), m- 2 x E aD,
is compact, since the difference K - K has a continuous kernel. Let <p be a

solution to t~ homogeneous equation <p + K<p = 0 and define u by (6.36).
Then 2u = K<p + <p = 0 on aD, and by the uniqueness for the exterior
Dirichlet problem it follows that u = 0 in lRm \ D. Using (6.15), we deduce
the asymptotic behavior
uniformly for all directions. From this, since u = 0 in lRm \ fJ, we obtain
faD <p ds = O. Therefore <p + K <p = 0, and from Theorem 6.20 we conclude
that <p is constant on aD. Now faD (x, y) ds(y), XED, (6.38)
with continuous density 'lj; is a solution of the interior Neumann problem

provided that 'lj; is a solution of the integral equation
'lj;(x)
r
+ 2 laD 'lj;(y)
a(x,y)
av(x) ds(y) = 2g(x), x E aD. (6.39)
Proof. This follows from Theorem 6.18. D
Theorem 6.26 The interior Neumann problem is solvable if and only if
r
laD
gds =0 (6.40)
is satisfied.
Proof. The necessity of condition (6.40) is a consequence of Green's theorem
(6.3) applied to a solution u. The sufficiency of condition (6.40) follows from
the fact that by Theorem 6.20 it coincides with the solvability condition of
the Fredholm alternative for the inhomogeneous integral equation (6.39),
i.e., for 'lj; + K''lj; = 2g. D
Theorem 6.27 The single-layer potential
u(x) = ( 1/J(y)if?{x, y) ds(y), x E JRm \ D, (6.41)

laD
with continuous density 1/J is a solution of the exterior Neumann problem
provided that 1/J is a solution of the integral equation
( aif?(x,y)
1/J(x) - 2 laD 1/J{y) av(x) ds(y) = -2g(x), x E aD, (6.42)
and, if m = 2, also satisfies
( 1/J ds = o. (6.43)
laD
Proof. Again this follows from Theorem 6.18. Observe that for m = 2
the additional condition (6.43) ensures that u has the required behavior
u(x) = 0(1), Ixl-t 00, as can be seen from (6.14). 0
Theorem 6.28 In JR3 the exterior Neumann problem has a unique solu-
tion. In JR2 the exterior Neumann problem is uniquely solvable if and only
if
( gds =0 (6.44)
laD
is satisfied.
Proof. By Theorems 3.4 and 6.20 the equation 1/J - K'1/J = -2g is uniquely
solvable for each right-hand side g. If (6.44) is satisfied, using the fact that
1 + Kl = 0, we find
2(1,1/J) = (1 - Kl, 1/J) = (1,1/J - K'1/J) = -2(1, g) = O.
Hence, the additional property (6.43) is satisfied in JR2. That condition

(6.44) is necessary for the solvability in JR2 follows from (6.9). 0
We finally show that the solutions depend continuously on the given

boundary data.
Theorem 6.29 The solutions to the Dirichlet and Neumann problems de-
pend continuously in the maximum norm on the given data.
Proof. For the Dirichlet problem the assertion follows from the maximum-
minimum principle (Theorem 6.8). In two dimensions, for the exterior prob-
lem, from the form (6.36) of the solution u we observe that we have to
incorporate the value U oo at infinity through faD
<p ds. But this integral de-
pends continuously on the given boundary data, since the inverse (1 +K)-l
of I + K is bounded by Theorem 3.4.
For the Neumann problem we first observe that for single-layer potentials
U with continuous density 1/J for any closed ball B in IRm we have an
estimate of the form
\\U\\oo,B ~ \\W\\oo,B \\1/J\\oo,aD,

where the function
W(x):= ( \(x,y)\ds(y), xEIRm,

JaD
is continuous in IRm by Theorem 6.14. Then for the exterior problem choose
a sufficiently large ball B and the continuous dependence of the solution
on the boundary data in B follows from the boundedness of the inverse
(1 - K,)-l of 1 - K'. In the remaining exterior of B, continuity then
follows from the maximum-minimum principle.
For the interior problem we can expect continuity only after making the
solution U unique by an additional condition, for example, by requiring
that faD Uds = O. From (1, K'1/J) = (K1,1/J) = -(1, 1/J) we observe that K'
maps the closed subspace Co(8D):= {1/J E C(8D): faD1/Jds = O} into it-
self. By Theorem 6.20 the operator 1 +K' has a trivial nullspace in Co(8D).
Hence, the inverse (I + K')-l is bounded from Co(8D) onto Co(8D), Le.,
the unique solution 1/Jo of 1/Jo + K'1/Jo = g satisfying faD 1/Jo ds = 0 de-
pends continuously on g. Therefore, as above, the corresponding single-
layer potential Uo depends continuously on g in the maximum norm. Fi-
nally, u := Uo - faD Uo ds/\8D\ yields a solution vanishing in the integral
mean on the boundary, and it depends continuously on g. 0
6.5 Nonsmooth Boundaries

Despite the fact that the integral equation method provides an elegant ap-
proach to constructively prove the existence of solutions for the boundary
value problems of potential theory we do not want to disguise its major
drawback: the relatively strong regularity assumption on the boundary to
c
be of class 2 . It is possible to slightly weaken the regularity and allow
Lyapunov boundaries instead of C 2 boundaries and still remain within the
framework of compact operators. The boundary is said to satisfy a Lya-
punov condition if at each point x E 8D the normal vector l/ exists and
there are positive constants L and a such that for the angle -o(x, y) between
the normal vectors at x and y the estimate -o(x,y) ~ L\x - y\'" holds for
all x, y E 8D. For the treatment of the Dirichlet and Neumann problem for
Lyapunov boundaries, which does not differ essentially from that for C 2
boundaries, we refer to [123].
However, the situation changes considerably if the boundary is allowed
to have edges and corners. This effects the form of the integral equations
and the compactness of the integral operators as we will demonstrate by

considering the interior Dirichlet problem in a two-dimensional domain D
with corners. We assume that the boundary aD is piecewise twice differ-
entiable, i.e., aD consists of a finite number of closed arcs r I, ... , r p that
are all of class G 2 and that intersect only at the corners Xl, ... , xp. At the
corners the normal vector is discontinuous (see Fig. 6.1 for a domain with
three corners).
For simplicity, we restrict our analysis to boundaries that are straight
lines in a neighborhood of each of the corners. In particular, this includes
the case where aD is a polygon. The interior angle at the corner Xi we de-
note by 'Yi and assume that 0 < 'Yi < 27r, i = 1, ... ,p, i.e., we exclude cusps.
For a boundary with corners, the continuity of the double-layer potential
with continuous density as stated in Theorem 6.17 remains valid, but at
the corners the jump relation (6.24) has to be modified into the form
i = 1, ... ,p, (6.45)
where 8: = 'Yi/7r and 8; = 2 - 'Yi/7r. It is a matter of straightforward

application of Green's theorem as in Example 6.16 to verify (6.45) for
constant densities. For arbitrary continuous densities, the result can be
obtained from the G 2 case of Theorem 6.17 by a superposition of two
double-layer potentials on two G 2 curves intersecting at the corner with
the density c.p equal to zero on the parts of the two curves lying outside
aD.
Trying to find the solution to the interior Dirichlet problem in the form
of a double-layer potential with continuous density c.p as in Theorem 6.21 re-
duces the boundary value problem to solving the integral equation
c.p - Kc.p = -2/, where the operator K : G(aD) -+ G(aD) is given by
(Kc.p) (x), X =f. Xi, i = 1, ... ,p,

(Kc.p)(x) := { .
(Kc.p)(x) + (~ -1 )c.p(Xi), X = Xi, i = 1, ... ,p.
Note that for c.p E G(aD), in general, Kc.p is not continuous at the corners.
However K c.p is continuous, since it is the sum K c.p = v+ + v_ of the
continuous boundary values of the double-layer potential v.
By Problem 6.1 the kernel
k(x,y) := v(y)· {x - y}
7rlx - Yl2
of the integral operator K is continuous on r i x r i for i = 1, ... ,p. Sin-
gularities of the kernel occur when X and y approach a corner on the two
different arcs intersecting at the corner.
For n E 1N we use the continuous cutoff function h introduced in the

proof of Theorem 2.22 to define the operators Kn : C(8D) -+ C(8D) by
(KnCP)(x):=
l8D
r h(nlx - yl)k(x, y)cp(y) ds(y), x E 8D.
For each n E 1N the operator Kn is compact, since its kernel is continuous

on 8D x r i fori = 1, ... ,p, i.e., we can interpret Kn as the sum of p integral
operators with continuous kernels on 8D x r i by subdividing the integral
over 8D into a sum of integrals over the arcs r i for i = 1, ... ,p.
FIGURE 6.1. Domain with a corner
Now consider Kn := K - Kn and assume that n is large enough that

for each x E aD the disk B[x; lin] = {y E m? : Ix - yl :S lin} intersects
only either one or, in the vicinity of the corners, two of the arcs rio By our
assumption on the nature of the corners we can assume n is large enough
that in the second case the intersection consists of two straight lines A and
B (see Fig. 6.1). Let
M := . max max Ik(x, Y)I .

• =l, ... ,p x,yEr i
Then, by projection onto the tangent line, for the first case we can estimate
-
I(Kncp)(x)1 :S Mllcplloo 1 8DnB[x;1/n]
ds(y) :S Mllcplloo -4 .
n
In the second case, we first note that for x E B \ {Xi}, by Green's theorem
(6.3) applied in the triangle with the corners at x and at the endpoints Xi
and z of A we have
r188v(y)
JA
cI>(x, y) I ds(y) I r 8cI>(x, y) ds(y) I
=
JA 8v(y)
= a(x) ,
211"
where a(x) denotes the angle of this triangle at the corner x (see Fig. 6.1).
Elementary triangle geometry shows that a(x)+'i :S 7f, where, without loss
of generality, we have assumed that Ii < 7f. Therefore, since for x E B\ {Xi}
we have k(x,y) = 0 for all y E B \ {xd, we obtain
Finally, for the corner Xi at the intersection of A and B we have
since k(Xi' y) = 0 for all y E (A U B) \ {xd. Combining these results we

observe that we can choose n large enough that IIKnlloo :S q where
q := .EJ-ax
t-l, ... ,p
11 - Ii
7r
I < 1.
Hence, we have a decomposition 1- K = 1- Kn - K n , where 1- Kn has
a bounded inverse by the Neumann series Theorem 2.9 and where Kn is
compact. It is left to the reader to carryover the proof for injectivity of the
operator I - K from Theorem 6.20 to the case of a boundary wigt corners.
Then existence of a solution to the inhomogeneous equation 'P - K'P = - 2f
follows by Corollary 3.6.
This idea of decomposing the integral operator into a compact operator
and a bounded operator with norm less than one reflecting the behavior
at the corners goes back to Radon [151] and can be extended to the gen-
eral two-dimensional case and to three dimensions. For details we refer
to Cryer [28], Knil [94], and Wendland [182]. For a more comprehensive
study of boundary value problems in domains with corners we refer to
Grisvard [57]. For the integral equation method in Lipschitz domains we
refer to Verchota [179J.
Finally, we wish to mention that the integral equations for the Dirichlet
and Neumann problems can also be treated in the space L2(aD) allowing
boundary data in L 2 (aD). This requires the boundary conditions to be un-
derstood in a weak sense, which we want to illustrate by again considering
the interior Dirichlet problem. We say that a harmonic function u in D
assumes the boundary values f E L2(aD) in the L2 sense if
lim
h-HO
r [u(x - hv(x)) - f(x)Fds(x)
J{}D
= O.
To establish uniqueness under this weaker boundary condition, we choose

parallel surfaces aD h := {x-hv(x) : x E aD} to aD with h > 0 sufficiently
small. Then, following Miranda [125J, for
J(h):= r
J&Dh
u 2 ds, h > 0,
we can write
J(h) = r
loD
{I + 2hH(x) + h2K(x)} [u(x - hv(x)))2ds(x)
and differentiate to obtain

1 dJ r fJu r
2" dh = - loDh U fJv ds + laD {H(x) + hK(x)} [u(x - hv(x)Wds(x).
Hence, using Green's theorem (6.1), we have
(6.46)
where Dh denotes the interior of the parallel surface fJD h. Now let u van-
ish on the boundary fJD in the £2 sense and assume that grad u #- 0
in D. Then there exists some closed ball B contained in D such that
1 := IB Igrad ul 2dx > 0, and from (6.46) we deduce that dJ/dh ::; -I
for all 0 < h ::; ho and some sufficiently small ho > O. Since J is continuous
on [0, hoL is continuously differentiable on (0, hoL and satisfies J(O) = 0, we
see that J(h) ::; -Ih for all 0 < h ::; ho. This is a contradiction to J(h) ~ 0
for all h > O. Therefore u must be constant in D, and from J(O) = 0 we
obtain u = 0 in D.
Using the fact that, due to Theorem 6.6, on the parallel surfaces fJD h
for h > 0 there is more regularity of u, a different approach to establishing
uniqueness under weaker assumptions was suggested by Calderon [19). It is
based on representing u in terms of the double-layer operator Kh on fJDh
and establishing IIKh - KII£2(aD) -+ 0 as h -+ O.
To prove existence of a solution for boundary conditions in the £2 sense
via the surface potential approach, it is necessary to extend the jump rela-
tions of Theorem 6.14,6.17,6.18, and 6.19 from C(fJD) onto £2(fJD). This
can be achieved quite elegantly through the use of Lax's Theorem 4.11 as
worked out by Kersten [83). In particular, for the double-layer potential v
with density cp E £2(fJD), the jump relation (6.24) has to be replaced by
lim r
h-l-+O laD
[2v(x ± hV(x)) - (Kcp) (x) =F cp(xWds(x) = o. (6.47)
From this, we see that the double-layer potential with density cp E £2(fJD)
solves the Dirichlet problem with boundary values f E £2 (fJD) provided the
density solves the integral equation cp - Kcp = -2f in the space £2(fJD).
Noting that integral operators with weakly singular kernels are compact
from £2(fJD) into £2(fJD) (see Problem 4.5), for existence and uniqueness
of a solution to this integral equation we need to establish that the homo-
geneous equation admits only the trivial solution. From Theorem 6.20 we
know that the operator 1 - K has a trivial nullspace in C(fJD). Therefore,
by the Fredholm alternative applied in the dual system (C(8D),L2(8D))

with the L2 bilinear form, the adjoint operator 1- K' has a trivial nullspace
in L 2 (8D). Again by the Fredholm alternative, but now applied in the dual
system (L2(8D), L2(8D)) with the L2 bilinear form, the operator 1- K
has a trivial nullspace in L2(8D). This idea to use the Fredholm alterna-
tive in two different dual systems for showing that the nullspaces for weakly
singular integral operators of the second kind in the space of continuous
functions and in the L2 space coincide is due to Hahner [64].
We will come back to the potential theoretic boundary value problems
in the next two chapters. In Section 7.5 we will solve the interior Dirichlet
and Neumann problems in two dimensions by integral equations of the first
kind in a Holder space setting. And in Section 8.3, again in two dimensions,
we will solve the integral equations of the second kind in Sobolev spaces
leading to weak solutions of the boundary value problems.
For integral equation methods for boundary value problems for the Helm-
holtz equation Dou + k 2 u = 0, i.e., for acoustic and electromagnetic scat-
tering problems, we refer to [24, 25].
Problems
6.1 Use a regular 271'-periodic parameterization aD = {x(t) : 0 :::; t:::; 271'} with
counterclockwise orientation for the boundary curve to transform the integral
equation (6.35) of the interior two-dimensional Dirichlet problem into the form
cp(t) - 10r" k(t, r)cp(r) dr = -2f<t),

where cp(t) := <p(x(t)), f<t) := f(x(t)) and the kernel is given by
.!. [x'(r)]..L . {x(t) - x(r)} t i=
71' Ix(t) _ x(r)12 ' r,
k(t,r) =
{
1 [x' (t)]..L . x" (t)
271' Ix'(t)12
t = r,
where [x']..L := (x~, -xD. Show that this kernel is continuous provided aD is of
class 0 2 •
6.2 Show that for an ellipse with parametric representation XI(t) = a cos t,
X2(t) = bsin t, the kernel k of Problem 6.1 is given by
ab 1
k(t, r) = -;- a2 + b2 _ (a2 _ b2) cos(t + r) .
6.3 Extend Theorem 6.20 to domains D with nonconnected boundaries and, in
particular, show that dim N (/ - K) = p, where p denotes the number of bounded
components of IRm \ D. For the interior Dirichlet problem establish existence of
a solution through a modification of the integral equation (6.35) analogous to
(6.36) by adding a point source in each of the bounded components of IRffl \ D.
Problems 93
6.4 Let D C IR? be of class C 2 and strictly convex in the sense that the cur-
vature of the boundary aD is strictly positive. Show that there exists a constant
o < 8 < 1 such that
( Ia~(Xi, y) - a~(X2, y) Ids(y) < 1 - 8
laD av(y) av(y) -
for all Xi,X2 E aD.
Hint: Use Example 6.16, Problem 6.1, and the property that
a~(x,y) v(y) . {x - y}
av(y) 27rlx - Yl2
is negative on aD x aD to verify that
Ilr{ {a~(Xi'
av(y)
y) - a~(X2, y)} dS(y)1
av(y)
:s .! - alaDI,
2
for each Jordan measurable subset reaD, where
.
a:= x,yEaD
min la~(x'Y)1
aV ()y > o.
6.5 In 1870 Neumann [137] gave the first rigorous proof for the existence of a
solution to the two-dimensional interior Dirichlet problem in a strictly convex
domain of class C 2 • By completely elementary means he established that the
successive approximations
1 1
<pn+i := 2" <pn + 2" K<Pn - J, n = 0, 1,2, ... ,
with arbitrary rpo ~ C(aD) converge uniformly to the unique solution <p of the
integral equation <p - K <p = - 2J. In functional analytic terms his proof amounted
to showing that the operator L given by L := ~ (1 + K) is a contraction with
respect to the norm
IIrpll := I sup rp(z) - inf <p(z) I + a sup Irp(z)l,
zEaD zEaD zEaD
where a > 0 is appropriately chosen. This norm is equivalent to the maximum

norm. Derive the above results for yourself.
Hint: Use Problem 6.4 to show that IILII :s
(2 - 8 + a)/2 by writing
(Lrp)(x) = laD [rp(y) - rp(x)] a:~~~f) ds(y)
and
1 1
(L<p)(Xi) - (Lrp)(X2) = 2" [<P(Xi) - rp(x)]- 2" [<P(X2) - rp(x)]
{ [ ( )_ ()] {a~(Xi'Y) _ a~(X2'Y)} d ( )

+ laD rp y rp x av(y) av(y) sy ,
where x E aD is chosen such that rp(x) = {suPzEaD rp(z) + infzEaD rp(z)}/2 (see
[81]).
7
Singular Integral Equations
In this chapter we will consider one-dimensional singular integral equations

involving Cauchy principal values that arise from boundary value problems
for holomorphic functions. The investigations of these integral equations
with Cauchy kernels by Gakhov, Muskhelishvili, Vekua, and others have
had a great impact on the further development of the general theory of
singular integral equations. For our introduction to integral equations they
will provide an application of the general idea of regularizing singular oper-
ators as described in Chapter 5. We assume the reader is acquainted with
basic complex analysis.
For a comprehensive study of singular integral equations with Cauchy
kernels we refer to Gakhov [46], Meister [121], Mikhlin and Prossdorf [124],
Muskhelishvili [133], Prossdorf [147], Vekua [178], and Wendland [184].
7.1 Holder Continuity

For our investigation of singular integrals we first need to introduce the
concept of Holder continuity.
Definition 7.1 A real- or complex-valued function <p defined on a set

G c lRm is called uniformly Holder continuous with Holder exponent
o < a ::; 1 if there exists a constant C such that
l<p(x) - <p(y)1 ::; Clx _ yin
7.1 Holder Continuity 95
for all x, y E G. By cO,a (G) we denote the linear space of all functions de-
fined on G that are bounded and uniformly Holder continuous with exponent
Q. The space CO,a(G) is called a Holder space.
Note that each uniformly Holder continuous function is uniformly con-

tinuous, but the converse is not true. We illustrate by two examples that
uniformly Holder continuous functions live between continuous and differ-
entiable functions. The function cP : [0,1/2] ---+ JR, given by cp(x) = 1/lnx
for x E (0,1/2] and cp(o) = 0, is uniformly continuous but not uniformly
Holder continuous. The function 't/J : [0,1] ---+ JR, given by 't/J(x) = Vx,
is uniformly Holder continuous with exponent 1/2 but not continuously
differentiable on [0,1]. In general, by the mean value theorem, a continu-
ously differentiable function on a convex set with bounded derivatives is
uniformly HOlder continuous with exponent 1.
Theorem 7.2 The Holder space CO,a(G) is a Banach space with the norm
Ilcplla := sup Icp(x)1 + sup Icp(x) - cp(Y)1 .

xEG x,yEG Ix - yla
xf.y
Proof. It is clear that
ICPla:= sup Icp(x) - cp(y) I

x,yEG Ix - yla
xf.y
defines a semi-norm on CO,a(G), i.e., it satisfies all norm axioms with the
exception of the definiteness (N2). Then II ·lIa = II ·1100 + 1·la is a norm,
since IIcplioo := sUPxEG Icp(x)1 defines a norm. Convergence in the supremum
norm II ·1100 is equivalent to uniform convergence on G. If G is compact,
the supremum norm and the maximum norm on C(G) coincide.
It remains to show that CO,a(G) is complete. Let (CPn) be a Cauchy se-
quence in CO,a(G). Then obviously (CPn) also is a Cauchy sequence with
respect to the supremum norm and, by the sufficiency of the Cauchy crite-
rion for uniform convergence, there exists a function cP E C(G) such that
IICPn - cplloo ---+ 0, n ---+ 00. Because (CPn) is a Cauchy sequence in CO,a(G),
given c: > 0, there exists N(c:) E 1N such that ICPn - CPkla < e for all
n, k 2: N(c:), i.e.,
for all n, k 2: N(c:) and all x, y E G. Since CPn ---+ cP, n ---+ 00, uniformly on
G, by letting k ---+ 00 we have
I{CPn(X) - cp(x)} - {CPn(Y) - cp(y)}1 :::; c:lx - yla
for all n 2: N(c:) and all X,y E G. From this we conclude that cP E CO,a(G)
and ICPn -cpla :::; dor all n 2: N(c:), which implies IICPn -CPlia ---+ 0, n ---+ 00. 0
96 7. Singular Integral Equations
Note that the product of two uniformly HOlder continuous functions tp

and 'Ij; is again uniformly Holder continuous with
By the following technical lemma we illustrate that Holder continuity is

a local property.
Lemma 1.3 Assume that the function tp satisfies Itp(x)1 :::; M for all x E G
and
Itp(x) - tp(y)1 :::; Clx - yla
for all x, y E G with Ix-yl :::; a and some constants a, C, M, and 0 < O! :::; 1.
Then tp E cO,a (G) with
Iitplia :::; M+ max (C, 2a~) .

Proof. If Ix - yl > a then
Itp(x) - tp(y)1 :::; 2M :::; 2M (Ix: YI) a,
whence Itpla :::; max(C,2Mjaa) follows. o
In particular, from Lemma 7.3, we see that for O! < {3 each function
tp E Co,fJ(G) is also contained in CO,a(G). For this imbedding we have the
following compactness property.
Theorem 1.4 Let 0 < O! < {3 :::; 1 and let G be compact. Then the imbed-
ding operators
IfJ : CO,fJ(G) -+ C(G)
and
are compact.
Proof. Let U be a bounded set in Co,fJ(G), i.e., IitplifJ :::; C for all tp E U
and some positive constant C. Then we have
for all x E G and

Itp(x) - tp(y)1 :::; Glx - YlfJ (7.1)
for all x, y E G and tp E U, i.e., U is bounded and equicontinuous. There-
fore, by the Arzela-Ascoli Theorem 1.18, the set U is relatively compact in
C(G), which implies that the imbedding operator IfJ : Co,fJ(G) -+ C(G) is
compact.
It remains to verify that U is relatively compact in CO'O!(G) . From (7.1)

we deduce that
I{<p(x) - ,¢(x)} - {<p(y) - '¢(y)}1
= I{ <p(x) -'¢(x)}- {<p(y) -'¢(y) }i0!/.B1{<p(x) -,¢(x)} - {<p(y) -'¢(y) }1 1-0!/.B
~ (2C)0!/.Blx - yI0!(211<p - '¢lIoo)1-O!/.B
for all <p, '¢ E U and all x, y E G. Therefore
for all <p, '¢ E U, and from this we can conclude that each sequence taken
from U and converging in C(G) also converges in CO'O!(G). This completes
the proof. 0
7.2 The Cauchy Integral Operator

Let D be a bounded and simply connected domain in the complex plane. We
denote its boundary by r := aD and assume it to be of class C 2 • The normal
vector v is directed into the exterior of D. For complex integration along the
contour r we assume that the direction of integration is counterclockwise.
We confine our analysis to this basic configuration and note that it can
be extended, for example, to multiply connected domains with less regular
boundaries.
The Cauchy integral
fez) := ~ f <p() d(, z E {) \ r,

27l'~ lr (- z
with density <p E C(r) defines a function that is holomorphic in {)\D and in
D. Obviously, for z varying in an open domain not intersecting with r, the
integrand is continuous with respect to ( and continuously differentiable
with respect to z. Therefore, we can differentiate under the integral to verify
the Cauchy-Riemann equations for f. Occasionally, we will call a function
that is holomorphic in {) \ jj and in D sectionally holomorphic.
As in the case of the single- and double-layer potentials of Chapter 6
we are interested in the behavior of the Cauchy integral for points on the
boundary, where the integral becomes singular. By integrating the equation
(1n In x)' = (x In x) -1 between 0 ~ x ~ 1 we observe that the Cauchy
integral for points on the boundary, in general, will not exist if the density is
merely continuous. Therefore we assume the density to be uniformly Holder
continuous. Note that the introduction of Holder continuity by Definition
7.1 also covers functions on subsets of the complex plane by identifying
IR? and ~. For the remainder of this chapter we will always assume that
a E (0,1) for the HOlder exponent, despite the fact that part of our results
remain valid for a = 1.
Theorem 7.5 For densities <p E CO''''(r), 0< a < 1, the Cauchy integral
exists as a Cauchy principal value
_1 r <p«) d( = _1 lim r <p«) d(

27ri lr ( - Z 27ri p-tO lr\r(z;p) ( - Z
for all Z E r. Here, r(ZiP) := {( E r: I( - zi ~ pl.

Proof. Let Z E r and set H(ZiP) := {( ED: = p} with counter- I( - zi
clockwise orientation. Then, by Cauchy's integral theorem,
Writing ( = Z + peiiJ , d( = ipe iiJ dfJ, we find
1·1m
p-tO
l -d(- = 1·1m
H(z;p) ( - Z p-tO
l H(z;p)
·dfJ·
~ = ~7r,
since r is of class C 2 • Hence
_1_ [~_~ - , Z Er, (7.2)

27ri r (- Z 2
in the sense of a Cauchy principal value.
The normal vector v is continuous on r. Therefore, as in the proof of
Theorem 2.23, we can choose R E (0,1] such that the scalar product of the
normal vectors satisfies
1
lI(z) . v«) 2: "2 (7.3)
for all z, ( E r with Iz - (I ::; R. Furthermore, we can assume that R is

small enough that r(Zi R) is connected for each Z E r. Then the condition
(7.3) implies that r(z; R) can be bijectively projected onto the tangent line
to r at the point z. The line elements ds«) on r and do- on the tangent
are related by
do-
ds«) = vZ·lI(
() ()::; 2 do-.
Hence,
r
lr(z;R)
I<p«) -
( - z
I
<p(z) ds ::; 21<p1", jR-R
10-1",-1 do- = 4R'"
a
l<pl",
and
( lep(() - ep(z) IdS::S lepl", ( I( - ZI",-l ds::S R",-llrllepl",·

Jr\r(z;R) ( - z Jr\r(z;R)
Therefore
(ep(()-ep(z) d(
Jr (- z
exists as an improper integral, and by the decomposition
( ep(() d( = (ep(() - ep(z) d( + ep(z) { ~

Jr ( - z Jr (- z Jr ( - z
the proof is completed. o
After establishing the existence of the Cauchy integral for points on the
boundary, we want to show that the function f defined by the Cauchy
integral is uniformly Holder continuous in jj and in <C \ D. For this we
choose h > 0 small enough that in the parallel strip
Dh := {w + 1Jhv(w) : W E r, 1J E [-1, 1]}

each point z is uniquely representable through projection onto r in the
form z = w+1Jhv(w) with wE r and 1J E [-1,1] (see p. 75). Then we show
that the function 9 : Dh -t <C defined by
g(z) := ~ ( ep(() - ep(w) d(, z E Dh,

27rZ Jr (- z
is uniformly Holder continuous with
(7.4)
where C is some constant independent of ep.
We begin by noting that we can choose h sufficiently small that
(7.5)
for each pair Zl = Zl(1J1) = W1 + 1J1hv(W1)' Z2 = Z2(1J2) = W2 + 1J2hv(W2).

Since r is of class C2, the normal v is continuously differentiable. Therefore,
by the mean value theorem, there exists a constant M > 0 such that
IV(Wl) - v(w2)1 ::s Mlw1 - w21 for all W1,W2 E r. Then, from
IW1 - w21 ::s IZ1(1J) - Z2 (1J) I + hlv(W1) - v(w2)1, 1J E [-1,1],
we observe that IW1 - w21 ::s 2Iz1(1J) - Z2(1J) I for all 1J E [-1,1]' if we
restrict h ::s 112M. Now (7.5) follows from the observation that for all
1J1, 1J2 E [-1, 1] by elementary geometry we have
We can use (7.5) to estimate l<p( () -<p( w)1 ::; 1<Pla I( -wla ::; 2a 1<Pla I( -zia
and then, as in the proof of Theorem 7.5, by splitting the integral into the
two parts over r(W; R) and r \ r( W; R), we obtain
(7.6)
for all Z E Dh and some constant Co.
Now let Zl,Z2 E Dho with 0 < IZI - z21 ::; R/4 and set p := 41z1 - z21.
Then we can estimate
(7.7)
where C 1 is some constant depending on a. Here we have made use of the

fact that r(Wl;p) C r(w2;2p). For I( - wli ~ p = 41z1 - z21 we have
1 1
I(-Zll::; 1(-Z21+lz2- Z11::; 1(-Z21+41(-Wll::; 1(-Z21+21(-zd.
Hence, I( - zd ::; 21( - z21 for I( - wli ~ p, and this inequality can now be
used to obtain
::; IZI - z211<Pla r

Jr\r(Wl;P)
I( - w21 a
I( - zlll( - z21
ds (7.8)
::; C~IZI - z211<Pla {;:R (}"~~a + ~;~a }::; C21z1 - z2I a l<Pla,
where C2 and C~ are some constants depending on a and r. Note that for
this estimate we need the restriction a < 1, and we have split the integral
into the two parts over r( WI; R) \ r( WI; p) and r \ r( WI; R). Finally, since
from (7.5) and the proof of Theorem 7.5 it is obvious that
r d(
Jr\r(Wl;p) ( - Zl
is bounded by 211", we have the estimate
for some constant 03. Combining (7.7)-(7.9), we obtain
211" Ig(ZI) - g(z2)1 ~ (01 + OJ. + a3)lzl - z21 alcpla (7.10)

for all ZI,Z2 E Dh with IZI-Z21 ~ R/4. Now the desired result (7.4) follows
from (7.6) and (7.10) with the aid of Lemma 7.3.
By Cauchy's integral theorem and (7.2) we have
1, zED,
1
2~ilr(~Z= 2'
ZE r,
0, ZE cc \ D,
Therefore, from the decomposition
I(z) = g(z)
cp(w)
+ -2· 11"~
Irr .. - z
d(
-r- ,
we observe that I E aO,a(D h n D) and I E aO,a(Dh n (CC \ D)) with

boundary values
1
I±(z) = -2.
11"~
Ir cp«() d( =t= -21 cp(z),
-r-
r .. - z
z E r.
Holder continuity of I in jj and CC \ D then follows by again employing

Lemma 7.3. We summarize our results in the following theorem.
Theorem 7.6 (Sokhotski-Plemelj) For densities cp E aO,a(r) the holo-
morphic function I defined by the Cauchy integral
1
I(z) := -2.
11"~
Irr .. -
cp«()
-r-
z
de, ZECC\r, (7.11)
can be unilormly Holder continuously extended from D into jj and from

CC \ jj into CC \ D with limiting values
J±(z) = -2·
1
11"~
1 cp«()
-r-
r .. - z
1
d( =t= -2 cp(z), zE r, (7.12)
where J±(z) = h-++O

lim I(z ± hV(z)). Furthermore, we have the inequalities
lor some constant a depending on 0: and r.

The formula (7.12) was first derived by Sokhotski [167] in his doctoral
thesis in 1873. A sufficiently rigorous proof was given by Plemelj [143] in
1908.
Corollary 7.7 The Cauchy integral operator A Co,<> (r) ---t Co,<> (r),
defined by
(Aip)(z) := ~
7ft
r:(() de,
Jr " - z
zE r,
is bounded.
Proof. We write the Sokhotski-Plemelj formula (7.12) in the short form
1 1
J± = -2 Aip =f -
2
ip (7.13)
and obtain the boundedness of A from Theorem 7.6. D
As a first application of the Sokhotski-Plemelj Theorem 7.6 we solve

the problem of finding a sectionally holomorphic function with a given
discontinuity along the contour r.
Theorem 7.8 Let ip E CO,"'(r). Then there exists a unique function f that
is holomorphic in D and <C \ jj, which can be extended continuously from
D into jj and from <C \ 15 into <C \ D satisfying the boundary condition
f - - 1+ = ip on r,
and for which fez) ---t 0, z ---t 00, uniformly for all directions. This function
is given by {7.11}.
Proof. By Theorem 7.6 the function f given by (7.11) has the required
properties. To establish uniqueness, from f _ - f + = on r, by Morera's°
theorem, we conclude that f is holomorphic everywhere in <C, i.e., f is an
entire function. Then from fez) ---t 0, z ---t 00, by Liouville's theorem it
follows that f = 0 in <C. From this proof we see that the general sectionally
holomorphic function f satisfying f- - 1+ = ip on r is obtained from (7.11)
by adding an arbitrary entire function. D
As a second application of Theorem 7.6 we state necessary and sufficient

conditions for the existence of a holomorphic function in D or in <C \ 15 with
given boundary values.
Theorem 7.9 For a given function ip E co,<>(r), there exists a function f

that is holomorphic in D and continuous in jj with boundary values f = ip
on r if and only if ip is a solution of the homogeneous integral equation of
the second kind
ip - Aip = O.
The solution is given by {7.11}.
Proof. Let 1 be holomorphic with 1= r.p on f. Then by Cauchy's integral

formula we have
I(z) =~ r I(()
2nk(-z
d( =~
2nk(-z
r r.p(() de, ZED,
and from Theorem 7.6 it follows that 2r.p = 2/- = Ar.p + r.p on f, and
therefore r.p - Ar.p = o.
Conversely, if r.p is a solution of r.p - Ar.p = 0, then again by Theorem 7.6
the function 1 defined by (7.11) has boundary values 2/- = Ar.p + r.p = 2r.p
onf. 0
Obviously, for the corresponding exterior problem in <C\D with 1(00) = 0

we have the homogeneous integral equation r.p + Ar.p = O. From these results
we observe that the operator A provides an example for nullspaces N(I -A)
and N(I + A) with infinite dimension, since there exist infinitely many
linearly independent holomorphic functions. This, in particular, implies
that A is not compact, and it also means that I - A and I + A cannot be
regularized.
We now can prove a property of the Cauchy integral operator A, which
is of central importance in our study of singular integral equations with
Cauchy kernels, since it will allow the construction of regularizers.
Theorem 7.10 The Cauchy integral operator A satisfies
Proof. For r.p E CO'O«f) define f by (7.11). Then, by Theorem 7.9, we have
f+ + Af+ = 0 and f- - Af- = O. Hence, using (7.13), we derive
and this is the desired result. o
Note that Theorem 7.10 is a further indication that the Cauchy integral
operator A cannot be compact.
Theorem 7.11 The operators A and - A are adjoint with respect to the
dual system (Co,o< (f), CO,O< (f)) with the nondegenerate bilinear form
(r.p,,¢):= l r.p(z)'¢(z)dz, r.p,,¢ E CO'O«f).
Proof. It is left as an exercise to show that (.,.) is nondegenerate on

Co,O< (f). For r.p, '¢ E CO,O< (f) define
f(z) := ~
27r~
r r.p(()
lr (- z
de, 1 ['¢(()
g(z):= -2.
7r~
-;-- de,
r.,-z
z E <C\f.
Then, by Theorem 7.6 and Cauchy's integral theorem we find that
= -2 ( J(z)g(z) dz -+ 0, R -+ 00,
J1z1=R
since J(z), g(z) = O(l/izi), z -+ 00. D
Note that Theorem 7.11 justifies interchanging the order of integration

in
Example 7.12 Let r be the unit circle. Then, substituting z = eit , ( = eir ,
we find
~ = ! (cot T - t + dT,
(- z 2 2
i)
i.e., the operator A can be expressed in the form
(A<p) .
(e't) 1 127r { cot -T - t
= -. +i} <p .
(e'r) dT, t E [0,27rJ,
27rZ 0 2
with the integral to be understood as a Cauchy principal value. Consider
the integral equation of the first kind
1 (27r T - t -
27r Jo cot -2- 0(T) dT = 'lj;(t), t E [0,27rJ, (7.14)
in the space CO,,,, [0, 27r]. Since
{27r t
Jo cot"2 dt =O,
by integrating (7.14) with respect to t and interchanging the order of inte-

gration (use Theorem 7.11 for 'lj;(z) = liz and transform the integrals), we
find that
{27r
Jo :¢(t) dt = ° (7.15)
is a necessary condition for the solvability of (7.14). It is also sufficient,

because by making use of A2 = I we see that when (7.14) is satisfied the
general solution of (7.14) is given by
1 (27r T - t -
ip(t) = - 27r Jo cot -2- 'lj;(T) dT + C, t E [0,27rJ, (7.16)
where C is an arbitrary constant. D

7.3 The Riemann Problem 105
The formulas (7.14) and (7.16) represent the inversion formulas of Hilbert.
The kernel in (7.14) and (7.16) is a called Hilbert kernel, since singular equa-
tions with this kernel were first encountered by Hilbert [71]. Splitting the
integral equation of Theorem 7.9 into its real and imaginary part, we read-
ily find that the Hilbert inversion formulas relate the real and imaginary
part of holomorphic functions in the unit disk.
7.3 The Riemann Problem

The following boundary value problem was first formulated by Riemann in
his inaugural dissertation. Because a first attempt toward a solution was
made by Hilbert [71] in 1904 through the use of integral equations, what
we will denote as the Riemann problem in the literature is sometimes also
called the Hilbert problem.
Riemann Problem. Find a junction f that is holomorphic in D and in

<C \ D, which can be extended uniformly Holder continuously from D into
D and from <C \ D into <C \ D satisfying the boundary condition
f-=gf++h onr (7.17)
and
f(z) -+ 0, z -+ 00, (7.18)
uniformly for all directions. Here, g and h are given uniformly Holder con-
tinuous junctions on r.
Before we give a solution to this Riemann problem for the case where
g vanishes nowhere on r, we want to indicate its connection to integral
equations.
Theorem 7.13 Let a,b,h E CO,a(r). Then the Cauchy integral (7.11)
maps solutions cp E cO,a (r) of the integral equation
acp + bAcp = h (7.19)
linearly and bijectively onto solutions f of the Riemann problem with the
boundary condition
(a + b)f- = (a - b)f+ + h on r. (7.20)
Proof. By Theorems 7.6 and 7.8, the Cauchy integral maps cO,a (r) linearly
and bijectively onto the linear space of sectionally holomorphic functions
that vanish at infinity and can be extended uniformly Holder continuously
with Holder exponent a from D into jj and from <C \ jj into <C \ D. From
(7.13) and Theorem 7.9 we have the relation
a<p + bA<p = a(f- - f+) + b(f- + 1+) = (a + b)f- - (a - b)f+
between the density <p and the function f defined by the Cauchy integral.
This ends the proof. 0
Definition 7.14 Let 9 be a complex valued and nowhere vanishing func-

tion defined on the contour r. The index of the function 9 is the integer
indg given by the increment of its argument along the contour in a coun-
terclockwise direction divided by 27r, i. e.,
indg:= ~ { dargg.
27r lr
The index can also be expressed by the logarithm of 9 through
1 . ( dIng.
indg = -2
mlr
Now, with the aid of the Sokhotski-Plemelj jump relations, we explicitly
solve the homogeneous Riemann problem.
Theorem 7.15 Let 9 E CO,a(r) be a nowhere vanishing function with

ind 9 = K. Then there exists a unique sectionally holomorphic function fo
satisfying the homogeneous boundary condition
fo- = gfo+ on r
and
lim ZK fo(z) = 1
Z-400
uniformly for all directions. It is called the canonical solution to the homo-
geneous Riemann problem, has the property that fo(z) "# 0 for all z E <C,
is
and uniformly Holder continuous up to the boundary.
Proof. Choose a point a E D and define G E cO,a (r) by
G(z) := (z - a)-Kg(z), z E r.
Because G has index ind G = 0, we can reduce the homogeneous boundary
condition F _ = G F+ by taking logarithms to obtain In F _ - In F+ = In G,
where we may take any branch of the logarithm to arrive at a single-valued
function InG E CO,a(r). According to Theorem 7.8 we set
'lj;(z):= ~ ( InG(() d(, zE <c\r,

27rZ lr (- z
7.4 Integral Equations with Cauchy Kernel 107
and F(z) := e"pCz), z E <C \ r. Then, from 'l/J- - 'l/J+ = InC on r we derive
F_ = CF+ on r, and since 'l/J(z) --+ 0, z --+ 00, we have F(z) --+ 1, z --+ 00.
Now the function fo, defined by,
F(z), ZED,
fo(z) := {
(z - a)-I< F(z), z E <C \ b,
has the required properties.
Let f be any other sectionally holomorphic function satisfying the homo-
geneous boundary condition f - = gf+ on r. Then the quotient q := f / fo
is sectionally holomorphic (observe fo(z) -/:- 0 for all z E <C) satisfying
q_ = q+ on r. Hence, by Morera's theorem, q is an entire function. From
zl< fo(z) --+ 1 and zl<f(z) --+ 1 for z --+ 00, it follows that q(z) --+ 1 for
z --+ 00, and Liouville's theorem implies q = 1 on <C. D
From the proof we note that the general sectionally holomorphic func-
tion f satisfying the homogeneous boundary condition f- = gf+ on r is
obtained from the canonical solution fo by multiplying it by an arbitrary
entire function.
Theorem 7.16 Under the assumptions of Theorem 7.15, the homogeneous

Riemann problem admits max(K, 0) linearly independent solutions.
Proof. For p an entire function, the product f := pfo vanishes at infinity if

and only if
Hence, p must be a polynomial of degree less than or equal to K- 1 if K > 0

and it must be zero if K :::; O. D
By the procedure used in Theorem 7.15 it is also possible to treat the

inhomogeneous Riemann problem. But here we will proceed differently and
use Theorem 5.6 to obtain results on the existence of solutions to the inho-
mogeneous integral equation (7.19), which then by the equivalence stated
in Theorem 7.13 imply results on the existence of solutions to the inhomo-
geneous Riemann problem (see Problem 7.1).
7.4 Integral Equations with Cauchy Kernel

For 0 < a, f3 :::; 1 by 0 0 ,(3,0 (r x r) we denote the set of all functions k
defined on r x r satisfying
(7.21 )
for all points Zl! Z2, (1, (2 E f and some constant M depending on k. Let
o < a < (3 ~ 1 and assume that a E CO,OI. (r) and k E CO,{3,OI. (f x f). Then
the operator K : CO,OI.(f) -t CO,OI.(f), defined by
1
(K<p)(z) := a(z)<p(z) + ---:
7rZ
Ir -;--
r
k(z, () <p«() d(,
.,-z
z E f,
is called a singular integral operator with a Cauchy kernel. The operator

KO : CO,OI.(f) -t CO,OI.(r), given by
(K°<p)(z) := a(z)<p(z) + b(~)

7rZ
r <p«() d(,
lr (- z
z E f,
where b(z) := k(z, z), z E f, is called the dominant part of the operator
K. For the coefficients a and b of the dominant part we will assume that
a 2 - b2 vanishes nowhere on f. The dominant part can be written in the
short form KO = aI + bA and is bounded by Corollary 7.7. The splitting
of the operator K into its dominant part KO and the remainder K - KO is
justified by the following theorem.
Theorem 7.17 The singular integral operator K is bounded and the dif-
ference between K and its dominant part KO is compact from CO,OI. (f) into
CO,OI.(f).
Proo/.We choose 0 < 'Y < a and show that the difference H := K - KO
is bounded from CO,'Y(r) into CO,OI.(r). Then the assertion follows from
Theorems 2.16 and 7.4.
We write
(H<p)(z) =~
7r~ lr(h(z,()<p«()d(, z E f,
with kernel
h( ;-)._ k(z, () - k(z, z) (7.22)
z,.,.- (_ z '
Since k E Co,{3,OI.(r x f), there exists a constant M such that
Ih(z, ()I ~ Mlz - (101.-1, Z =F (. (7.23)
Hence, the kernel h is weakly singular and therefore we have
(7.24)
for some constant Co.

To establish HOlder continuity, we proceed analogously to the proof of
Theorem 7.6. Let ZI, Z2 E f with 0 < IZI -z21 ~ R/4 and set p:= 41z1 -z21.
Then, using (7.23), we can estimate
::; C~lIrplioo 1
2P
uo-1du ::; C11lrplloolzl - z21°
for some constants C 1 and C~ depending on M and a. Using (7.21), we can
estimate
(7.26)
for some constants C2 , C~, and C~ depending on M, a, /3, and r. Because

for ( E r with I( - zll 2:: p we have that I( - zll ::; 21( - z21, we derive the
further estimate
(7.27)
::; C~llrplloolzl - z21 {1: u o - 2 du + ~;~o }

::; C3 11rplloolzl - z21°
for some constants C3 and C~ depending on M,a, and r. Finally, we write
and note from the proof of Theorem 7.5 that the second integral on the
right-hand side is bounded by 271". For the first integral we have
for some constant C~ depending on r and 'Y. Hence, using again (7.21), we
obtain
(7.28)
for some constant C4 depending on M, a and r. Combining (7.25)-(7.28),

we find
for all Zb Z2 E r with IZI - z21 :::; R/4. Now the desired result
follows from (7.24) and (7.29) with the aid of Lemma 7.3. o
Theorem 7.18 The operators KO = aI + bA and KOI := aI - Ab are
adjoint with respect to the dual system (Co,a. (r), Co,a. (r). They both have
finite-dimensional nullspaces and the index of the operator KO is given by
. dKo =In
. d --b.
- a b
In
a+
Proof. That KO and KOI are adjoint is easily derived as a consequence of
Theorem 7.11. From Theorems 7.13 and 7.16 we obtain
(7.30)
where
a-b
K:= ind a+b.
Analogous to Theorem 7.13, the homogeneous adjoint equation
a'if;-Ab'if; =0 (7.31)
is equivalent to the Riemann problem with homogeneous boundary condi-

tion
(a - b)f- = (a + b)f+ on r. (7.32)
Let 'if; solve the integral equation (7.31). Then <p := b'if; solves a<p - bA<p = 0
and therefore, by Theorem 7.13, the sectionally holomorphic function f
defined by (7.11) satisfies the boundary condition (7.32). Conversely, let

f be a solution to the Riemann problem with boundary condition (7.32).
Then for the function 'lj; := 2(a + b)-l f- we have f- + 1+ = a'lj; and
f- - 1+ = b'lj;. Hence, by Theorem 7.9 and its analog for <C \ b, we find
a'lj; - Ab'lj; = f- + f+ - A(f- - 1+) = O.
With the aid of ind[(a + b)-l(a - b)] = - ind[(a - b)-l (a + b)], Theorem
7.16 now yields
(7.33)
and indKo = dimN(KO) - dimN(KO,) = /'\, follows by combining (7.30)
and (7.33). 0
From now on, for symmetry, we will assume that k E CO,(3,(3 (r x r) with
o < 0: < (3 :S 1. The singular integral operators with Cauchy kernel
(Krp)(z) := a(z)rp(z) 11k(z,()
+ ----: - ( - rp(() d(, z E r,
7ft r '> - Z
and
(K''lj;)(z) := a(z)'lj;(z) - ~
7ft
r~((, z) 'lj;(() d(,
lr '> - Z
z E r,
are adjoint. This follows by writing K = KO + Hand K' = KO, + H', where
Hand H' have weakly singular kernels h(z, () and h((, z), respectively, with
h given by (7.22). For weakly singular kernels, as ill Theorem 4.6, the order
of integration may be interchanged.
Theorem 7.19 The operator (a 2 - b2 )-1 K' is a regularizer of K.
Proof. First, we observe that for each c E cO,a (r) the difference Ac - cA
is compact from CO,a(r) into CO,a(r). This follows from Theorem 7.17,
applied to the kernel c(() contained in CO,(3,a(r x r). Actually, the need for
the compactness of the commutator Ac - cA was our reason for allowing
different exponents 0: and {3 in Theorem 7.17. Now, using A2 = I, we derive
KO, KO = (aI - Ab)(aI + bA) = (a 2 - b2 )I + M
where M := abA - Aba + (b 2 A - Ab2 )A is compact, since A is bounded by
Corollary 7.7. Hence, (a 2 _b2 )-1 KO' is a left regularizer of KO. Analogously,
KO KO, = (aI + bA)(aI - Ab) = (a 2 - b2 )I + M

where M:= bAa-aAb is compact, Le., KO'(a 2 _b2 )-1 is a right regularizer
of KO. Again, the difference (a 2 - b2 )-lKO, - KO'(a 2 - b2 )-1 is compact.
Therefore, (a 2 - b2 ) -1 KO, is a regularizer of KO. Now the assertion of the
theorem follows from the fact that K = KO + Hand K = KO, + H' with
compact Hand H'. 0
We now are in a position to derive the classical Noether theorems from

our general theory on regularization in dual systems of Chapter 5.
Theorem 7.20 (First Noether Theorem) The singular integral opera-

tor with Cauchy kernel has a finite-dimensional nullspace.
Proof. This follows from Theorems 5.6 and 7.19. o

Theorem 7.21 (Second Noether Theorem) The index of the singular
integral operator K with Cauchy kernel is given by
a-b
indK = ind --b'
a+
Proof. This follows from Theorems 7.18 and 7.19 and Corollary 5.12. 0
Theorem 7.22 (Third Noether Theorem) The inhomogeneous singu-

lar integral equation
Kcp = h
is solvable if and only if
is satisfied for all 'lj; E N(K').

°
Theorem 7.23 Let < a < (3 S 1, a E CO,a(r), and k E CO,i3,i3(r x r)
and assume that a 2 -b2 vanishes nowhere on r, where b(z) := k(z, z), z E r.
Then, in the Holder space CO,a(r), the number of linearly independent
solutions of the homogeneous singular integral equation
a(z)cp(z) +~ f ~z, () cp(() d( = 0, z E r,

7rZ lr . , - z
and of its adjoint equation
a(z)'lj;(z) - ~ f :((, z) 'lj;(() d( = 0, z E r,

7rZ lr . , -z
are both finite, and their difference is given by the index
_1_fdlna-b
27ri lr a +b .
The inhomogeneous singular integral equation
a(z)cp(z) +~ r,
7rZ lrf k;Z'
. , - ()z cp(() d( = h(z), z E
is solvable if and only if
Ir h(z)'lj;(z) dz = °
is satisfied for all solutions 'lj; of the homogeneous adjoint equation.
Proof. This is a reformulation of the three preceding theorems. 0
For equations of the first kind, i.e., equations with a = 0, we note the
following corollary.
Corollary 7.24 Singular integral operators of the first kind with Cauchy
kernel have index zero. In particular, injective singular integral operators
of the first kind are bijective with bounded inverse.
Proof. The index zero follows from Theorem 7.21. The bounded inverse for
an injective operator of the first kind is a consequence of the Riesz Theo-
rem 3.4 together with the fact that A provides an equivalent regularizer. 0
With the aid of Example 7.12, choosing for r the unit circle, we obtain the
following corollary, which contains the equations for which Noether [138]
proved the theorems named after him.
Corollary 7.25 Let a and k be real-valued and 27r-periodic and assume

that a 2 + b2 is strictly positive, where b(t) := k(t, t) for t E [0, 27r]. Then
for the singular integral equation with Hilbert kernel
1
a(t)cp(t) - 27r 10
rrr k(t, r) cot -2-
r-t
cp( r) dr = h(t), t E [0,27rJ,
and its adjoint equation
a(t),¢(t)
1 r
+ 27r 10
21r r - t
k(r, t) cot -2- ,¢(r) dr = g(t), t E [0,27rJ,
the three Noether theorems are valid, i.e., the numbers of linearly indepen-
dent solutions of the homogeneous equation and the homogeneous adjoint
equation are both finite, and their difference is given by the index
1
-
12'11' darctan -b .
7r 0 a
The inhomogeneous equation is solvable if and only if
r
10
21r
h(t),¢(t) dt = °
is satisfied for all solutions '¢ of the homogeneous adjoint equation.
Proof. After transforming them onto the unit circle r by setting z = eit ,
( = eir , the two equations read
a(z)<P'(z) -.!.
7r
rk(z, () <P'(() d(
lr (- z
= h(z), zE r,
and
a(z)~(z) + ~1[' lrrk;C, z) ~(e) de = g(z),
,,- z
z E f,
for ip(e it ) := cp(t) and ~(eit) := 'lj;(t)je it . Here, we have set a(e it ) := a(t),
h(e it ) := h(t), g(e it ) := g(t)je it , and k(eit,e iT ) := (e it + eiT )k(t,r)j2eiT .
Hence, the corollary is established through transforming the Noether the-
orems for the pair of equations on the unit circle back into the equations
on [0,21[']. For the index we compute
+ 'b 1211" = 21n( a + ib)

In a _ ~b
1211"
= 2i arctan -b 1211" ,
a 2 ° ° a °
and the proof is finished. o
Concluding this section, we wish to mention that the above integral equa-
tions can also be treated in various other normed function spaces. In this
context we make use of Lax's Theorem 4.11 to establish the boundedness
of the Cauchy integral operator A in L 2 (f).
Theorem 7.26 The Cauchy integral operator A is bounded from L2(r)
into L 2 (f).
Proof. We want to apply Lax's Theorem 4.11 in the positive dual system
(Go,D: (f), GO,D: (f)) generated by the scalar product
(cp, 'lj;) := Ir cp,¢ ds, cp, 'lj; E GO,D: (f).
Therefore, we need the adjoint A * of the Cauchy integral operator A with

respect to this scalar product. Complex integration and contour integration
along f are related by
dz = ei,),(z) ds(z) (7.34)
for all z E f, where ,( z) denotes the angle between the tangent to f at the
point z and the real axis. Using Theorem 7.11 and (7.34), we see that
Ir Acp¢ds = -Ir CPXA('lj;X)ds,
where X(z) := ei/(Z) , z E f. Hence, the adjoint A* is given by
A*'lj; = -XA('lj;X),
and A* : GO,e>(f) -7 Go,D:(r) is bounded, since A : Go,D:(r) -7 Go,e>(r) is
bounded. Now the statement follows from Theorem 4.11 and the denseness
of Go,D:(r) in L 2 (f). 0
In particular, as a consequence of Theorem 7.26, the property A2 = I of

Theorem 7.10 remains valid in L2(f) and can be used to construct regu-
larizers as in Theorem 7.19.
7.5 Cauchy Integral and Logarithmic Potential

We now want to use the close connection between holomorphic and two-
dimensional harmonic functions to derive regularity results on the logarith-
mic single- and double-layer potential from the Sokhotski-Plemelj theorem.
In the Cauchy integral for fixed z E <C we substitute ( - z = reiiJ , where
r = I( - zl and f) = arg«( - z). Then
lrr (-
d(
z = lrr dln«( - z) = lrr · r In r . Of)) (0
d(ln r + zf)) = lr ---a;- + z os ds.
By the Cauchy-Riemann equations, applied to In( ( - z) = In r + if), we

have
of) = 8Inr = -~ln~
os OV OV r·
Consequently,
1
211"i
r cp(() d(
lr (- z = - 211"
1 r
lr cp«()
0
ov«() In
1
I( _ zl ds«()
(7.35)
1
- 211"i
r 0 1
lr cp«() os«() In I( _ zl ds«().
This formula indicates that for real-valued densities the real part of the
Cauchy integral coincides with the logarithmic double-layer potential
1
v(z) := 211"
r 0
lr cp«() ov«() In I( _ zl
1
ds«(),
Note that we use the same symbol z for the complex number z = Xl + iX2
and the vector z = (Xl,X2) E lR? Hence, from Theorem 7.6 and the
Sokhotski-Plemelj formula (7.12), we derive the following theorem.
Theorem 7.27 The logarithmic double-layer potential v with Holder con-

tinuous density cp can be extended uniformly Holder continuously from D
into iJ and from lR? \ iJ into lR? \ D with limiting values
1
v±(z) = 211"
r
lr cp«()
0
ov«() In
1 1
I( _ zl ds«() ± "2 cp(z), z E r. (7.36)
Recall the relation (7.34) between complex integration and contour in-
tegration along r. Since r is of class C2, the angle 'Y is continuously dif-
ferentiable with respect to z. Therefore, from Theorem 7.6 we immediately
deduce that
1
f(z) := -2 -;--
11" r.,,-z
1
cp«() ds«(), (7.37)
is holomorphic in D and in (: \ jj and that it can be extended uniformly

HOlder continuously from D into jj and from <C \ jj into <C \ D with limiting
values
J±(z) := ~ ( ep«() ds«() T!:. e-h(z)ip(z), z E f. (7.38)

271" lr (- z 2
The gradient of the logarithmic single-layer potential
1
u(z) := 271"
r 1
lr ep«() In I( _ zl ds«(), z E IR? \ f,
is given by
1 { (- z
gradu(z) = 271" lr ep«() I( _ Zl2 ds«(),
We observe that the gradient of the single-layer potential u corresponds to

the complex conjugate of the holomorphic function f given by the Cauchy
integral (7.37). Therefore, observing the fact that the complex conjugate of
ie-h(z) corresponds to the normal v(z) at z E f, we can again use Theorem
7.6 to prove the following theorem.
Theorem 7.28 The first derivatives of the logarithmic single-layer poten-
tial u with Holder continuous density can be extended uniformly Holder
continuously from D into jj and from rn? \ jj into rn? \ D with limiting
values
1 r
gradu±(z) = 271" lr ep«() grad In
1
I( _ zl ds«() T
1
2" v(z)ep(z), z E f. (7.39)
In particular, we see that the logarithmic single-layer integral operator

S defined by
1
(Sep)(z) := 271" lrrep«() In I( _1 zl ds«(), z Ef, (7.40)
is a bounded operator from Co,O:(f) into Cl,O:(r). By Cl,O:(f) we denote

the normed space of all functions defined on f that have a uniformly HOlder
continuous first derivative furnished with the norm
where the prime indicates differentiation with respect to the arc length.
Consider again the logarithmic double-layer potential v and assume that
ep E Cl,O:(f). For a = (al, a2) E rn? we denote a.L = (a2' -al). Then, for
r = I( - zl, we write
a l l
r r
av«() In = v«()· grad, In = -v«()· grad z In r1= - (1r
div z In [r«()).L
)
where we have made use of the relation v = r.l. between the tangent vector
r and the normal vector v. From this, using the vector identity
grad div w = bow + [grad div w.l.]\
the identity [w.l.]l. = -w, and 61n l/r = 0, we see that
grad z av~() IIi ~ = [grad (r( () .grad In ~ ) ] .l.

z z
Then, by partial integration, for the gradient of the double-layer potential

we obtain that
1[
= 271" grad ir(d<p 1
ds (() In I( _ zl ds(()
].l. , z E ]R? \ r.
Therefore, from Theorem 7.28 we derive the following result.

Theorem 7.29 The first derivatives of the logarithmic double-layer poten-
tial v with density <p E CI,a(r) can be extended uniformly Holder continu-
ously from D into jj and from m? \ jj into m? \ D. The normal derivative
is given by
av± 1 d ( dip 1
r,
av (z) = 271" ds(z) ir ds (() In I( _ zl ds((), zE (7.41)
and the tangential derivative by
This implies that the operator T, defined by the normal derivative of the
logarithmic double-layer potential
1 a ( a 1
r,
(T<p)(z) := 271" av(z) ir <pee) av(() In I( _ zl ds((), z E (7.43)
can be expressed through the single-layer integral operator S in the form
(7.44)
Hence, T is a bounded operator from CI,a(r) into CO,a(r).

We wish to mention that these results on the regularity of single- and
double-layer potentials with uniformly Holder continuous densities can be
extended to the three-dimensional case by techniques similar to those used

in the proofs of Theorems 7.6 and 7.17 (see [24]).
Now we want to demonstrate how the two-dimensional Dirichlet and
Neumann problem for harmonic functions can be solved through the use
of integral equations of the first kind. For notational brevity we introduce
the operator M : cO,a (r) ---7 cO,a (r) by
M~:= ~ - I~I h~ds.

The single-layer potential
u(z) = 27f
1 r 1
lr (M~)(() In I( _ zl ds(()
1 r
+ 1fT lr ~ ds, z E ]R? \ f, (7.45)
solves the interior and the exterior Dirichlet problem with boundary con-
dition u = f on f provided the density ~ E cO,a (r) solves the integral
equation
h
2~ (M~)(() In I( ~ zl ds(() + I~I h~ ds = fez), z E f. (7.46)
We abbreviate (7.46) into the form So~ = f where we have set
So~:= SM~ + I~I h~ds. (7.47)
Note that our modification ofthe logarithmic single-layer potential in (7.45)

ensures the correct behavior of u at infinity for the exterior problem. As we
will see it is also crucial for the interior problem since it ensures uniqueness.
Because S maps CO,a(r) into C 1 ,a(f), the equation (7.46) can be solved
only for inhomogeneities f E C 1 ,a(r).
Theorem 7.30 The modified single-layer operator So: CO,a(f) ---7 C 1 ,a(r)
is bijective with bounded inverse SOl: C 1 ,a(r) ---7 CO,a(f).
Proof. Let f E c 1,a (f) and assume that ~ E cO,a (r) solves
So~ = f. (7.48)
Then ~ also solves
d
ds So~
1 r
+ Iff lr So~ ds =
df
ds
1
+ 1fT lr f
r ds. (7.49)
The bounded operator R : c 1,a (f) ---7 cO,a (f) given by
df
Rf := ds
1
+ 1fT lr f
r ds
is injective, since Rf = 0, by integration, implies fr f ds == 0 and df / ds = O.

Hence, the equations (7.48) and (7.49) are equivalent. By the jump relation
(7.39), we may differentiate under the integral in (7.46) to see that equation
(7.49) has the form of a singular integral equation with Cauchy kernel
1 r k(z, () df 1
21r Jr (- z 'l/J(() ds(()+(H'l/J)(z) = ds (z)+1R Jr f ds,
r z E r. (7.50)
Here, H : Co,o: (r) --7 Co,o: (r) given by
is a compact operator, since it is bounded and has finite-dimensional range.

The kernel k is given by
8 1 Re{h(z) (( - z)}
k(z, () = (( - z) 8s(z) In I( _ zl = I( _ zl2 (( - z), z =1= (,
where h(z) := Tl(Z) + iT2(Z) E <C. With the aid of a + ii = 2Rea, this can
be rewritten into the form
Since r is of class C 2 , the tangent vector T is continuously differentiable

on r, and using a parametric representation, it can be seen that
(( - z)2
z =1= (,
I( - z12'
can be extended as a continuously differentiable function onto r X r. This
implies k E CO,l,l(r x r). Hence, we can apply Noether's theorems in the
form of Problem 7.3. Because the integral equation is of the first kind, its
index is zero.
Let 'l/J be a solution to the homogeneous equation (7.49). Then 'l/J also sat-
isfies the homogeneous form of (7.48) and therefore the single-layer poten-
tial u given by (7.45) solves the homogeneous interior and exterior Dirichlet
problem. Hence, by the uniqueness Theorem 6.11, we have u = 0 in D and
in lR? \ D. From the behavior at infinity we deduce fr 'l/J ds = 0, and then
the jump relations imply 'l/J = 0 on r. Now the assertion of the theorem
follows from Corollary 7.24 and the boundedness of the differentiation op-
erator R. 0
We now turn to the Neumann problem. The double-layer potential
1 r 8 1
u(z) = 21r Jr rp(() 8v(() In I( _ zl ds((), ZElR?\r, (7.51 )
solves the interior and exterior Neumann problem with boundary condition
au/av = 9 on r provided the density c.p E C1,"'(r) solves the integral
equation
1 a
2n av(z)
r
lr <p(()
a
av(() In
1
I( _ zl ds(() = g(z), z E r. (7.52)
Theorem 7.31 For each 9 E CO''''(r) satisfying the solvability condition
h9dS = 0 (7.53)
of the interior and exterior two-dimensional Neumann problem there exists

a solution c.p E C1,"'(r) to the integral equation of the first kind {7.52}. Two
solutions can differ only by a constant.
Proof. The condition (7.53) ensures the existence of a single-valued func-

tion G E Cl,"'(r) with dG/ds = g. Now consider the equation So'l/J = G.
By Theorem 7.30 it is uniquely solvable. We set X := 'l/J - 'l/J ds/lrl· Ir
The property Ir X ds = 0 ensures the existence of a single-valued function
<p E C1''''(r) such that d<p/ds = X. Then
d dc.p d d dG
T<p = ds S ds = ds Sx = ds So'l/J = ds = g.
It is left as an exercise for the reader to show that the nullspace of T is
given by N(T) = span{I}. 0
Our analysis follows Hsiao and MacCamy [75], who modified an earlier
approach by Fichera [40] to make the integral equations uniquely solvable.
7.6 Logarithmic Single-Layer Potential on an Arc

In the final section of this chapter we wish to indicate how the modifications
of the integral equations of the first kind of Section 7.5 arising from the
Dirichlet or Neumann problem for the exterior of an arc can be reduced
to the case of a closed boundary contour by using a cosine substitution as
suggested by Multhopp [131]. For brevity we confine ourselves to the case
of the single-layer potential for the Dirichlet problem. For the Neumann
problem we refer to Manch [126].
Assume that r c lR? is an arc of class C 3 , i.e.,
r = {x(s) : s E [-1, I]},
where x : [-1,1] --* lR? is an injective and three times continuously differ-
entiable function with x'(s) i- 0 for all s E [-1,1]. We consider the Dirichlet
problem for the Laplace equation in the exterior of the arc f: Given a func-
tion f E C 1,a(f), find a bounded solution U E C2(lR? \ r) n C(R2) to the
Laplace equation
aU=o in R2\f (7.54)
satisfying the Dirichlet boundary condition
U = f on f. (7.55)
Note that besides continuity we do not explicitly assume any condition for
the behavior of the solution at the two endpoints Z1 := x(l) and Z-1 :=
x( -1) of f.
Theorem 7.32 The Dirichlet problem for the exterior of an arc has at
most one solution.
Proof. The statement follows analogously to the proof of Theorem 6.11
from the maximum-minimum principle and the boundedness condition at
infinity. 0
As in the case of a closed boundary curve (compare (7.45)) we will es-

tablish existence by seeking the solution in the form
1
u(z) = 271"
r 1
lr (M'lj;)«() In I( _ zl ds«()
1
+ 1fT lr'lj;ds,
r z E R2 \ f, (7.56)
of a single-layer potential. For the case of the arc f the density 'lj; is assumed
to be of the form
.I.(Z)
0/ = ;j(z) , Z E f \ {Z1, Z-1 }, (7.57)
viz - z111z - z-11
where ;j E CO,a(r). Note that the integrals in (7.56) exist as improper
integrals with respect to the two endpoints of f.
Recall the cut-off function h from the proof of Theorem 2.22 and, for
n E IN, denote by Un the single-layer potential (7.56) with the density M'lj;
replaced by Xn, where Xn«() := h(nl( - z1I)h(nl( - z-11)(M'lj;)«(). Then,
r
by interpreting f as part of a closed contour and setting Xn = 0 on f, r\
from Theorem 6.14 we conclude that Un is continuous in all of R2. For
sufficiently large n we have
r
lr(zl;1/n)
1
I( - z11 3/ 4
ds«() < C 1
- n 1/ 4
for some constant C 1 • Therefore, using Holder's inequality for integrals and
(7.57), we can estimate
for sufficiently large n and some constant C. Analogous to the logarithmic

single-layer potential the integral on the right-hand side is continuous and
therefore is bounded on compact subsets oflR? Therefore we can conclude
that
frq z 1 ;1/n)
I(M'IjJ)(() In Ir ~ II ds(() --+ 0,
., z
n --+ 00,
uniformly on compact subsets of IR? From this and the corresponding

property for the integral over a neighborhood of the other endpoint Z-l,
it follows that u is the uniform limit of a sequence of continuous functions
and therefore is continuous in all of IR? (see also [68], p. 276). Therefore,
as in the case of a closed boundary curve (compare (7.46», the single-
layer potential (7.56) solves the Dirichlet problem (7.54)-(7.55) provided
the density satisfies the integral equation
2~ h (M'IjJ)(() In I( ~ zi ds(() + I~I h 'IjJ ds = f(z), zE f. (7.58)
From the uniqueness Theorem 7.32 and the jump relations of Theorem 6.18
we can conclude that the integral equation (7.58) has at most one solution.
Following Multhopp [131] we substitute s = cos t into the parameteriza-
tion f = {x(s) : s E [-1, 1]} to transform the integral equation (7.58) into
the parameterized form
1 r
41f fo {-In( 4[cos t - COST]2) + p(t, T)} cp(T) dT = g(t), t E [0,1f]. (7.59)
Here we have set
cp(t) := Isin tx'(cost)1 'IjJ(x(cost», g(t):= f(x(cost», t E [0,1f],
and the kernel p is given by
4[cos t - cos TF 41f

p(t T) '= In
,. Ix(cost) - X(COST)12
+-
If I
1
--IfI lITo In Ix (cost) - x (cosO")12 Ix'(cosO")lsinO"dO"
1
for t =1= T. (Compare Problem 7.2 and use (7.60).) Solving the integral
equation (7.58) for the function 'IjJ on f is equivalent to solving the integral
equation (7.59) for the even 21f-periodic function cpo
Writing
x(s) - x(O") =
s- 0" fo
r x'[s + '\(0" - s)] d,\,
1
it can be seen that the kernel p can be extended as a twice continuously

differentiable function on IR? that is even and 21f-periodic with respect to
both variables. Hence, in view of Theorem 7.4 the integral operator
1 ('
(Lcp)(t) := 47r io {p(t, r) - 4}cp(r) dr, t E [0,7r],
is a compact operator from the Holder space cg~ae c CO,a(JR) of even

27r-periodic Holder continuous functions into the 'Holder space of even
27r-periodic Holder continuously differentiable functions ci~ae c C1,a(JR).
From the identity ,
In(4[cost _ cosr]2) = In (4sin 2 t ~ r) + In (4sin2 t ~ r) (7.60)
it follows that for even and 27r-periodic functions <p and g the integral
equation
1('
47r io {-In( 4[cos t - cos rj2) + 4} <p( r) dr = g(t), t E [0,7r],
is equivalent to
47r r
1 io 2
", { t-
-In ( 4sin2 -2-r) + }2 <p(r) dr = g(t), t E [0,27r].
From Problem 7.2, by choosing x(t) = (cost,sint) as the parameterization

of the unit circle, it can be seen that the latter equation corresponds to the
parameterized form of the single-layer integral equation (7.46) for the case
of the unit disk D. Hence, as a consequence of Theorem 7.30, the integral
operator
(Locp)(t):= - 1
47r
1'"
0
{-In(4[cost-cosrj2)+4}cp(r)dr, t E [0,7rJ,
is a bounded and bijective operator from cg~ae into ci~ae with a bounded
inverse. Therefore, writing (7.59) in the form Lo<p+L<p =:: g, from Corollary
3.6 and the injectivity for (7.58) we have the following theorem.
Theorem 7.33 For each f E C1,a(r) the single-layer integral equation
(7.58) has a unique solution 'lj; of the form (7.57) with ;j; E cO,a(r).
From this we finally conclude the following result.
Theorem 7.34 The Dirichlet problem (7. 54}-(7. 55} for the exterior of an
arc is uniquely solvable.
From the form (7.57) of the density of the single-layer potential we expect
that the gradient of the solution develops singularities at the endpoints of
the arc. For a closer study of the nature of these singularities we refer to
Grisvard [57].
Problems
7.1 Let 9 E CO''''(r) be a nowhere vanishing function. Show that the inhomoge-
neous Riemann problem with boundary conditions J- = gf+ +h on r is solvable
if and only if the condition fr h1-
dz = 0 is satisfied for all solutions to the 1
homogeneous adjoint Riemann problem with boundary condition 1+ = g1-.
7.2 Use a regular 27r-periodic parameterization r = {x(t) : 0 ~ t ~ 27r} with
counterclockwise orientation for the boundary curve to transform the integral
equation of the first kind (7.46) for the Dirichlet problem into the form
4~ 127r {-In (4sin 2 t ; r) + pet, r)} :;fer) dr = !<t),
where :;f(t):= Ix'(t)I,¢(x(t», !<t):= J(x(t», and where the kernel is given by
• 2 t - r
4sm -2-
p(t,r) := In Ix(t) _ x(r)12 -
1 1
1fT °
2" 1 , 47r
In Ix(t) _ x(o-}12 Ix (0-)1 do- + 1fT' t =1= r.
Show that the kernel p is continuously differentiable provided r is of class C 2 •

'fransform the integral equation of the first kind (7.52) for the Neumann problem
into the form
4~ 1 2
" {cot r;t + q(t,r)} cP'(r)dr = g(t) , 0~t~27r,
where cP(t) := cp(x(t», get) := Ix'(t)lg(x(t», and where the kernel is given by
._ x'(t). [x(r) - x(t)] r- t
q(t,r) .- 2 Ix(t) _ x(r)12 - cot -2-' t =1= r.
Show that the kernel q is continuous provided r is of class C 2 •

Hint: Use the integrals of Lemma 8.21 from the next chapter for m = o.
7.3 Formulate and prove Noether's theorems for the singular integral equa-
tion with Cauchy kernel where the complex integration is replaced by contour
integration.
Hint: Use (7.34).
7.4 Evaluate the dominant parts of the singular integral operator occurring in
(7.50) and its adjoint. Then use (7.35) together with Theorem 7.17 to deduce
that the integral operators K and K', given by (6.31) and (6.32) are compact
from CO''''(r) into CO''''(r).
7.5 Show that for each real-valued function cp E CO''''(r) there exists a function
J that is holomorphic in D, which can be extended continuously from D into jj
and that has real part Re J = cp on the boundary r. Show that two functions
with this property can differ only by a purely imaginary constant.
Hint: Solve a Dirichlet problem for the real part of J and use (7.35) and Problem
7.4.
8
Sobolev Spaces
In this chapter we study the concept of weak solutions to boundary value

problems for harmonic functions. We shall extend the classical theory of
boundary integral equations as described in the two previous chapters from
the spaces of continuous or Holder continuous functions to appropriate
Sobolev spaces. For the sake of brevity we will confine ourselves to interior
boundary value problems inJwo dimensions.
We wish to mention that our introduction of Sobolev spaces of periodic
functions and Sobolev spaces on a closed contour differs from the usual
approach in two regards. Firstly, instead of using the Fourier transform we
only need Fourier series, and secondly, the Sobolev spaces in the contour
case are defined via a global rather than a local procedure. The motivation
for our approach is the hope that this analysis might make the basic ideas
on these Sobolev spaces more easily accessible. We also want to emphasize
that we do not rely heavily on Lebesgue integration as a necessary pre-
requisite. For our purpose, it will be sufficient to understand the Hilbert
space L2[0, 211"] as the completion of the space e[o, 211"] of continuous func-
tions with respect to the mean square norm.
8.1 The Sobolev Space HP[O, 27r]

As the basis of our presentation of Sobolev spaces we begin with a brief
review of the classical Fourier series expansion. For a function <p E L2[0, 211"]

126 8. Sobolev Spaces
the series
(8.1)
where
am := 2~ fo27r rp(t)e-imtdt,
is called the Fourier series of rp, its coefficients am are called the Fourier
coefficients of rp. On L2[0, 27f], as usual, the mean square norm is introduced
by the scalar product
r
(rp,1j;):= 1o rp(t)1j;(t) dt.
27r
We denote by fm the trigonometric monomials
fm(t) := eimt
for t E lR and mE 7L. Then the set (fm : mE 7L} is an orthogonal system.
By the Weierstrass approximation theorem (see [31]), the trigonometric
polynomials are dense with respect to the maximum norm in the space of
27r-periodic continuous functions, and e[0,27r] is dense in L2[0, 27r] in the
mean square norm. Therefore, by Theorem 1.28, the orthogonal system is
complete and the Fourier series (8.1) converges in the mean square norm.
Because of the orthonormality factor Ilfmll~ = 27r, Parseval's equality as-
sumes the form
(8.2)
We will now define subs paces HP[O, 27f] of L2[0, 27f] by requiring for their
elements a certain decay of the Fourier coefficients am as Iml ---+ 00.
Definition 8.1 Let °: :;p < 00. By HP[0,27r] we denote the space of all
functions rp E L2[0, 27r] with the property
L
00
(1 + m2 )Plam l2 < 00
m=-oo
for the Fourier coefficients am of rp. The space HP[O, 27f] is called a Sobolev
space. Frequently we will abbreviate HP = HP[O, 27f]. Note that HO[O, 27f]
coincides with L2[0, 27r].
Theorem 8.2 The Sobolev space HP[0,27f] is a Hilbert space with the
scalar product defined by
L
00
(rp,1j;)p:= (1+m 2 )pambm

m=-oo
8.1 The Sobolev Space HP[O, 271"] 127
for 'P, 'ljJ E HP[O, 271"] with Fourier coefficients am and bm , respectively. Note
that the norm on HP[O, 271"] is given by
The trigonometric polynomials are dense in HP[O, 271"].
Proof. We leave it as an exercise to verify that HP is a linear space and

that (. , .)p is a scalar product. That (. , .)p is well defined can be concluded
from the Cauchy-Schwarz inequality
To prove that HP is complete, let ('Pn) be a Cauchy sequence, i.e., given

c > 0, there exists N(c) E 1N such that II'Pn - 'Pkllp < c for all n, k ~ N(c),
or in terms of the Fourier coefficients am,n of 'Pn,
L
00
(1 + m 2 )Plam,n - am,kJ 2 < c 2

m=-oo
for all n, k ~ N(c). From this we observe that
L
M2
(1 + m 2 )plam,n - am,kl 2 < c2 (8.3)
m=-Ml
for all M I , M2 E 1N and all n, k ~ N(c). Therefore, since ~ is complete,
there exists a sequence (am) in ~ such that am,n -7 am, n -7 00, for each
m E '/L;. Passing to the limit k -7 00 in (8.3) now yields
M2
L (1 + m 2 )Plam,n - amJ2 :::; c2
m=-Ml
for all MlJ M2 E 1N and all n ~ N(c). Hence
L
00
(1 + m 2 )PJam,n - aml2 :::; c2

m=-oo
for all n ~ N(c). From this we conclude that

00
m=-oo
defines a function cP E HP with Ilcp - CPnllp -t 0, n -t 00. Hence, HP is

complete.
Let cP E HP with Fourier coefficients am' Then for the partial sums
CPn := E~=-n amfm of the Fourier series we have that
00
IIcp - CPnll; = L (1 + m 2 )Plaml2 -t 0, n -t 00,
Iml=n+l
since the series E:=_00(1 + m 2)Plam l2 converges. Therefore, the trigono-
metric polynomials are dense in HP. 0
Theorem 8.3 If q > p then Hq[O, 271'] is dense in HP[O, 271'] with compact
imbedding from Hq[O, 271'] into HP[O, 271'].
Proof. From (1 + m 2 )P ::; (1 + m 2 )q for mE'll it follows that Hq c HP
with bounded imbedding
(8.4)
for all cP E Hq. Then the denseness of Hq in HP is a consequence of the
denseness of the trigonometric polynomials in HP.
We denote the imbedding operator by I : Hq -t HP. For n E 1N we
define finite dimensional bounded operators In : Hq -t HP by setting
InCP := E~=-n amfm for cP E Hq with Fourier coefficients am. Then
00
II(In - I)cpll~ = L (1 + m2 )Plam l2
Iml=n+l
Hence, compactness of the imbedding follows from Theorem 2.17. 0
Theorem 8.4 Let p > 1/2. Then the Fourier series for cP E HP[O,271']
converges absolutely and uniformly. Its limit is continuous and 271'-periodic
and coincides with cP almost everywhere. This embedding of the Sobolev
space HP[0,271'] in the space C27r of 271'-periodic continuous functions is
compact.
Proof. For the Fourier series of cP E HP[0,211'], by the Cauchy-Schwarz
inequality, we conclude that
m=-oo
for all t E JR. Since the series Z:::=-oo (1 +m 2)-P converges for p > 1/2, this
estimate establishes the absolute and uniform convergence of the Fourier
series. Its limit is continuous and 211"-periodic because it is the uniform limit
of a sequence of continuous and 211"-periodic functions. It must coincide with
cp almost everywhere, because we already have convergence of the Fourier
series to cp in the mean square norm, which is weaker than the maximum
norm.
The above estimate also shows that
for all cp E HP[O, 211"] and some constant C depending on p. Therefore, the
imbedding from HP[O, 211"] into C27r is bounded, and consequently, by The-
orems 2.16 and 8.3, is compact for p > 1/2. 0
For the introduction of the Sobolev spaces HP(r) for a closed contour r
in JR2 we shall need additional norms that are equivalent to II· lip- We will
denote them by II· IIp,[p] , where [P] is the largest integer less than or equal to
p. For their definition we have to distinguish three cases. For k E 1N, by C~7r
we will denote the space of k times continuously differentiable 211"-periodic
functions from JR into (IJ.
Theorem 8.5 For k E 1N we have C~7r C Hk[O, 211"] and on C~7r the norm
II ·llk is equivalent to
(8.5)
Proof. Using
fo27r cp(k) (t)e-imtdt = (im)k fo27r cp(t)e-imtdt (8.6)
and Parseval's equality (8.2) for the Fourier coefficients of cp and cp(k) we
obtain 00
m=-oo
Now the inequalities
(1 + m 2k ) :::; (1 + m 2)k :::; (2m2)k :::; 2k(1 + m 2k ), m E 7l,
yield the equivalence of the norms 1I·lIk and 1I·lIk,k. 0
Theorem 8.5 shows that for k E 1N the Sobolev spaces H k [O,211"] are
homeomorphic to the classical Sobolev spaces given by the completion of
the space of k times continuously differentiable 211"-periodic functions with
respect to the norm II ·lIk,k. Two normed spaces are called homeomorphic
if there exists a bijective bounded linear mapping from one space onto the
other that has a bounded inverse.
Theorem 8.6 For 0 < p < 1 on Gi". the norm II . lip is equivalent to
Proof. We first observe that the norm II . IIp,o corresponds to the scalar
product
r2". <p(t)'lj;(t) dt + r2". r2". {<p(t) - <p(r)}{'IjJ~t)+~ 'lj;(r)} drdt

I.
(<p,'lj;):=
io io io t - riP
sm--
2
on C}".. Note that the second term on the right-hand side is well defined,
since by the mean value theorem continuously differentiable functions are
uniformly Holder continuous with Holder exponent 1. Straightforward in-
tegration shows that the trigonometric monomials f m are orthogonal with
respect to this scalar product, Le.,
1".
where
sin2 m t
"1m := 1611" 2 t dt,
o sin2p+l -
2
8mk = 1 for k = m, and 8mk = 0 for k =f. m. Using 2t/7r < sin t < t for
o < t < 7r /2, we can estimate
1". 1'"
.2 m ·2 m
sm - t sm - t
22p+5 7r 2 dt < "Im < 167r 2p+2 2 dt.
o t 2p+1 0 t 2p+1
For m > 0, we substitute mt = r and use
o< 1
". .2 r
SIn -
2 +i dr :::;
1 m ".
.2 r
SIn -
2 +i dr <
1 00
·2 r
SIn -
2 +i dr < 00
o r P 0 r P 0 r P
to obtain
(8.8)
for all m =f. 0 with some constants Co and CI depending on p. Using the
estimate leimt_eimrl :::; Im(t-r)l, from Parseval equality (8.2) for <p E Gi".
8.1 The Sobolev Space HP[O, 27r] 131
and the Cauchy-Schwarz inequality, proceeding as in the proof of Theorem

8.4 we see that the series
I:
00
'P(t)-'P(7)= am{eimt_eimT}
m=-oo
is absolutely and uniformly convergent for all t,7 E [0, 271} Therefore we
can integrate termwise to derive
r r
27r
io io
27r 1'P(t) - '1'(7)12
-'-'--t---7-'----'<12;--'p--i-+71 d7dt =
-'-I'--.
00
Slll-- m=-oo
2
Now the statement of the theorem follows from the inequalities (8.8) and
together with Parseval's equality (8.2). 0
Corollary 8.7 When p = k + q, where k E 1N and 0 < q < 1, then on

C;:1 the norm II . lip is equivalent to
(8.9)
Proof. Combine the proofs of Theorems 8.5 and 8.6 (see Problem 8.1). 0
Corollary 8.8 For a nonnegative integer k let f E C~7r and assume that
o :S p :S k.
Then for all 'I' E HP[O, 27r] the product f'P belongs to HP[O, 27r]
and
for some constant C depending on p.
Proof. Use the equivalent norms of Theorems 8.5 and 8.6 and Corollary 8.7
(see Problem 8.1). 0
Definition 8.9 For 0 :S p < 00 by H-P[O, 27r] we denote the dual space of
HP[0,27r], i.e., the space of bounded linear functionals on HP[0,27r].
The space H-P is characterized by the following theorem.
Theorem 8.10 For FE H-P[O, 27r] the norm is given by

where Cm = FUm). Conversely, to each sequence (c m ) in <c satisfying

00
m=-(X)
there exists a bounded linear functional F E H-P[O, 27r] with FUm) = cm .

Proof. Assume the sequence (c m ) satisfies the required inequality and define
a functional F : HP[O, 27r]-+ <C by
00
m=-oo
for rp E HP[O,27r] with Fourier coefficients am. Then from the Cauchy-
Schwarz inequality we have
00 00
m=-oo m=-oo
Hence, F is well defined and bounded with
On the other hand, for n E IN the function rpn := E:=_n(l +m2)-Pcm fm

has norm
Therefore
Hence,
L~w (1+ m')-PIc",I'}'" 00 IlFllp,

and this completes the proof. o
Theorem 8.11 For each function 9 E L2[O, 27r] the duality pairing
ior
1 2rr
G(rp):= 27r rp(t)g(t)dt, rp E HP[O, 27r], (8.10)
defines a linear functional G E H-P[O, 27r]. In this sense, L2[O,27r] is a

subspace of each dual space H-P[O, 27rJ, and the trigonometric polynomials
are dense in H-P[O, 27r].
Proof. Denote by bm the Fourier coefficients of 9. Then, since GUm) = bm,

by the second part of Theorem 8.10 we have G E H-p. Let F E H-P with
FUm) = em and define a sequence (Fn) in H-P by
where 9n is the trigonometric polynomial 9n := 2:~=-n cmlm. Then we

have
L
00
JIP - Fnll; = (1 + m2)-pJemJ 2 -70, n -7 00,

Iml=n+l
analogous to Theorem 8.2. o
Obviously, H-P becomes a Hilbert space by appropriately extending the

definition of the scalar product from Theorem 8.2 to negative p. For p = 0
the duality map described in Theorem 8.11 is bijective with IIGlio = 119110.
Therefore, we can identify H-o and HO and obtain a Hilbert scale of Hilbert
spaces HP for all real p with compact imbedding from Hq into HP for q > p.
The Sobolev spaces HP frequently are called interpolation spaces because
of the interpolating properties indicated by the following two theorems.
Theorem 8.12 Let p < q and r = >.p + (1 - >.)q with 0 < >. < 1. Then
Proof. With the aid of Holder's inequality we estimate
L
00
1I<p1I~ = {(I + m2)PJam J2}A{(1 + m2)qJam J2}1-A

m=-oo
'5 L~~ (1 + m')'I....I' } {l=~ (1 + m')'laml' }

A H
= 1I<pII~A 1I<p1I~-2A,

For real s consider the operator D S transferring the trigonometric mono-
mials Im(t) = eimt to D S1m := (1 + m 2 )S 1m. For each p E 1R it clearly
defines a bounded linear operator from HP+2s onto HP with
(8.11)
for all cp E HP+2s. Obviously, Ds+ t = DS Dt for all s, t E JR. In particular,

D-s is the inverse of DS. For s = 1 the operator D = D1 corresponds
to the second derivative Dcp = cp - cp" if cp is twice continuously differen-
tiable. Therefore, we may interpret D as a generalization of this derivative
for elements of the Sobolev spaces and the operators DS as intermediate
derivatives. With the aid of D S we can transform the scalar products by
(8.12)
for all cp, 't/J E HP+s and
(8.13)
for all cp E HP+s and 't/J E HP+2 s .

Noting that Dfm = (1 + m 2)fm, i.e., the fm are eigenfunctions of the
operator D with eigenvalues 1 + m 2 , we indicate obvious possibilities for
extending this procedure to introduce interpolation spaces generated by
other operators.
Theorem 8.13 Let p, q E JR, r > 0, and let A be a linear operator mapping
HP boundedly into Hq and HP+r boundedly into Hq+r, i. e.,
for all cp E HP and

IIAcpllq+r S 0211cpllp+r
for all cp E HP+r and some constants 0 1 and O2. Then
for all 0 < ). < 1 and all cp E HP+>.r, i. e., A maps HP+>.r boundedly into
Hq+>.r for 0 < ). < 1.
Proof. By A* : Hq -+ HP we denote the adjoint of A: HP -+ Hq, i.e.,
(Acp,'t/J)q = (cp,A*'t/J)p (8.14)
for all cp E HP and 't/J E Hq. By Theorem 4.9 we have
for all 't/J E Hq. From this, making use of (8.11), we observe that the operator
.Ii:= D- r / 2A* Dr/2 is bounded from Hq+r into HP+r with
for all 't/J E Hq+r.

8.2 The Sobolev Space HP(r) 135
For all c.p E HP+r and '¢ E Hq+r, with the aid of (8.13) and (8.14), we
deduce
(Ac.p, ,¢)q+r/2 = (Ac.p, Dr/2'¢)q = (c.p, A* Dr/2'¢)p = (c.p, D- r / 2A* D r / 2,¢)p+r/2.
Hence, the bounded operators A : HP+r -t Hq+r and A : Hq+r -t HP+r

are adjoint with respect to the positive dual systems (HP+r, HP+r) and
(Hq+r, Hq+r) generated by the (p+r/2)- and (q+r/2)-scalar products.
Therefore, applying Lax's Theorem 4.11 and using the denseness of HP+r
in HP+r/2, we conclude that A is bounded from HP+r/2 into Hq+r/2 with
for all c.p E HP+r/2.

Repeating this argument, we see that
IIAc.pllq+>.r ::; Ci->'C~IIc.pllp+>.r

°
for all c.p E HP+>.r and all rational binary numbers < A < 1. For arbitrary
0< A < 1 we can choose a monotonic decreasing sequence (An) of binary
rationals such that An -t A, n -t 00. Then, using (8.4), we have
IIAc.pllq+>.r ::; IIAc.pllq+>'nr ::; ci->'nc~n 1Ic.pIlP+>'nr

for all c.p E HP+r and n E IN. Passing to the limit n -t 00 we see that
IIAc.pllq+>.r ::; Ci->'C~IIc.pllp+>.r

for all c.p E HP+r, and because HP+r is dense in HP+>.r, for all c.p E HP+>.r.
The idea to use Lax's theorem to prove this interpolation theorem is due
to Kirsch [85]. 0
8.2 The Sobolev Space HP(r)

Let r be the boundary of a simply connected bounded domain D c lR?
of class Ck, k E IN. With the aid of a regular and k-times continuously
for°: ;
differentiable 27r-periodic parametric representation r = {x(t) : t E [0,27r)}
p ::; k we can define the Sobolev space HP(r) as the space of all
functions c.p E L2(r) with the property that c.p 0 x E HP[O, 27r]. By c.p 0 x, as
usual, we denote the 27r-periodic function given by (c.p 0 x)(t) := c.p(x(t)),
t E JR. The scalar product and norm on HP(r) are defined through the
scalar product on HP[O, 27r] by
(c.p,'¢)HP(r):= (c.p o X,,¢OX)HP[O,21rj.

Without loss of generality we have restricted the parameter domain to
be the interval [0, 27r). However, we must allow the possibility of different
regular parametric representations for the boundary curve r. Therefore, we

have to convince ourselves that our definition is invariant with respect to
the parameterization.
Theorem 8.14 Assume that x and x are two different regular 21r-periodic
parametric representations for the boundary r of a simply connected bounded
domain D C IR? of class Ck, k E lN, i.e., r = {x(t) : t E [0,21r)} and
°
r = {x(t) : t E [0, 21r)}. Then, for ~ p ~ k the Sobolev spaces
HP(r):= {rp E L2(r): rpox E HP[0,21r]}
with scalar product
(rp, '¢)HP(r) := (rp 0 x, '¢ 0 X)HP[O,211'j
and
jjP(r) := {rp E L2(r) : rp 0 x E HP[O, 21r]}
with scalar product
are homeomorphic.
Proof. Assume that both parameterizations have the same orientation.
(The case of two opposite parameterizations is treated analogously.) Be-
cause of the 21r-periodicity, without loss of generality, we may assume that
x(O) = x(O). Denote by x-I: r --+ [0,21r) the inverse of x : [0,21r) --+ r.
Then the regularity of the parameterization implies that the mapping
f : [0,21r) --+ [0,21r) given by f := X-loX is bijective and k-times con-
tinuously differentiable with f(O) = 0. By setting f(t + 21r) := f(t) + 21r it
can be extended as a k-times continuously differentiable function on JR..
To establish the statement of the theorem it suffices to show that for all
rp E HP[O, 21r] we have rp 0 f E HP[O, 21r] with
Ilrp 0 flip::; Cllrpllp, (8.15)
where C is some constant depending on k and p. For this we make use of
the equivalent norms of Theorems 8.5 and 8.6 and Corollary 8.7. If p E lN,
we use the norm given by (8.5). Because of Theorem 8.5 there exist positive
constants CI and C2 such that
cIllrpllL ::; IIrpll; ::; c211rpllL
for all rp E C~1I' and all j = 1, ... ,po Then, using the chain rule and the
Cauchy-Schwarz inequality, we can estimate
IIrp 0 fll~ ~ C2 10211' {lrp(f(t))1 + I:; rp(f(t))r} dt

2
where M is some constant depending on p that contains bounds on the

derivatives of f up to order p. Substituting s = f(t) on the right-hand side
and using (8.4), we get the further estimates
Here, we have set 'Y := 1/1/f'l/co. This completes the proof of (8.15) for
p E lN, since C~7r is dense in HP[O, 71'].
If 0 < p < 1, we use the norm given by (8.7). From l'Hopital's rule and
the property f(t + 271') := f(t) + 271' we can conclude that the function
. f(t) - f(r)
sm 2
t-r
b(t,r) := sin--
2
t-r
If'(t)l, 2;- E'll,
is continuous and 271'-periodic with respect to both variables. Hence, b is
bounded and we can estimate
Icp(f(t)) - cp(f(r))12 < B 2p+1 Icp(f(t)) - cp(f(r)) 12
I.
sm-
2-
t_rI2P+1 - I'
f(t)-f(r)1 2P+1 '
sm 2
where B is a bound on b. From this, substituting s = f(t) and a = f(r),

we obtain
r27r r27r lcp(f(t)) - cp(f(r))12 drdt < 'Y2B2p+1 r27r r27r Icp(s) - cp(a)12 dads
10 10 I'sln-
t - -r 1
2
2P 1
+ - 10 10 I'sm-
s - -a 1
2
2P 1
+ •
With this inequality, (8.15) follows from the equivalence of the norms II· lip
and II . IIp,o on C}7r and the denseness of Ci7r in HP[O, 271'].
Finally, the case of an arbitrary noninteger positive p is settled by com-
bining the two previous cases with the aid of the norm given by (8.9). 0
The classical Sobolev space H1(D) for a bounded domain D C lR? with
boundary aD of class C 1 is defined as the completion of the space C1(D)
of continuously differentiable functions with respect to the norm
For a detailed description of this space, in particular, its equivalent def-

inition through the concept of weak derivatives, and its extension to the
Sobolev spaces Hk(D) of arbitrary order k E 1N we refer to Adams [1],
Gilbarg and Trudinger [49], and Treves [176]. Since each Cauchy sequence
with respect to 11'IIH1(D) is also a Cauchy sequence with respect to 11'1\£2(D),
we may interpret HI(D) as a subspace of L2(D). Obviously the gradient
can be extended from CI(f)) as a bounded linear operator from HI(D)
into L2(D) (see Problem 2.1).
We want to illustrate the connection between Sobolev spaces on the
domain D and Sobolev spaces on its boundary r through the simplest case
of the trace theorem. For functions defined on the closure f) their values
on the boundary are clearly defined and the restriction of the function to
the boundary r is called the trace. The operator mapping a function onto
its trace is called the trace operator.
As a first step, we consider continuously differentiable functions u in
the strip JR x [0, 1] that are 21f-periodic with respect to the first variable,
i.e., u(t + 21f,') = u(t,·) for all t E JR. By Q we denote the rectangle
Q := [0,21f) x [0,1]. For °: :; ry :::; 1 we consider the Fourier coefficients
am(ry) := ~ r
21f 10
27r
u(t, ry)e-imtdt.
By Parseval's equality we have
0:::; ry :::; 1.
Because the am and u are continuous, by Dini's theorem (see Problem 8.2),
the series is uniformly convergent. Hence, we can integrate term by term
to obtain
(8.16)
Similarly, from
10r
27r
I () 1 au ( ) -imtd
am ry = 21f ary t, ry e t
and
we see that
(8.17)
and
(8.18)
In the second step we shall show that
IIu(.,0)IIH1/2[O,211"] ~
2 1 2
-; IluIIH1(Q) (8.19)
for all u with u(·, 1) = 0. The latter property implies that am (1) = for
all mE'll. In particular, using the Cauchy-Schwarz inequality, we can
°
estimate
L
00
Ilu(.,0)lltl/2[O,211"] = (1 + m2)1/2Iam (0)12

m=-oo
We are now ready to establish the following theorem.

Theorem 8.15 Let D be a simply connected bounded domain in JR? with
boundary r of class 0 2 • Then there exists a positive constant 0 such that
for all u E 01(fJ).
°
Proof. As in the proof of Theorem 7.6, we choose a parallel strip Dh :=
{x + TJhv(x) : x E r, TJ E [0, I]} with some h> such that each y E Dh is
uniquely representable through projection onto r in the form y = x+TJhv(x)
with x E rand TJ E [0,1]. Here, for convenience, deviating from our general
rule, we assume the normal v is directed into D. By rh we denote the
interior boundary rh := {y = x + hv(x) : x E r} of D h . Then, through a
regular 211"-periodic parametric representation r = {x(t) : ~ t < 211"} of
the contour r we have a parameterization
°
x(t,TJ) = x(t) + TJhv(x(t)) , °t< ~ 211", ° TJ
~ ~ 1,
of the strip D h . In particular, x provides a bijective and continuously dif-

ferentiable mapping from Q onto Dh with a continuously differentiable
inverse. Thus, applying the chain rule, for all u E C1(Dh) with u = 0 on
rh from (8.19) we conclude that
1
lI u IlH I / 2 (r) = lIu 0 xIlH I / 2[0,211") :::; V1i lIu 0 xIIHl(Q) :::; cliuIlHl(Dh)'
where c is some constant containing a bound on the first derivatives of the

mapping x and its inverse.
Finally, to extend the estimate to arbitrary u E Cl(fJ), we choose a
function 9 E Cl(fJ) given by g(y) = 0 for y f/. Dh and g(y) = f(TJ) for
y E Dh with Y = x + TJhv(x). Here, f is defined by
f(TJ) := (1 - TJ)2(1 + 3TJ)
and satisfies f(O) = 1'(0) = 1 and f(l) = f'(l) = O. The property 1'(0) = 1
will not be used in this proof but will be needed later, in the proof of
Theorem 8.17. Then, applying the product rule, we obtain
lIuIlHI/2(r) = IIguIlHI/2(r) :::; cliguIlHl(D) :::; Cli u IlHl(D)

for all u E C1(fJ) and some constant C containing a bound on 9 and its
first derivatives. 0
Corollary 8.16 The trace operator can be uniquely extended as a contin-

uous operator from Hl(D) into Hl/2(r).
Proof. This follows from Theorem 8.15 and Problem 2.1. o
In the following theorem, by 1·IL1(r) we will denote the semi-norm
lul£1(r) := Ii udsl·
Theorem 8.17 Let D be as in Theorem B.15. Then there exists a positive

constant C such that
lI u lli2(D) :::; C {luli1(r) + II grad ulli2(D)}
and
lI u lli2(D) :::; C {liulli2(r) + II grad ulli2(D)}
for all u E Hl(D).
Proof. Since C 1 (fJ) is dense in Hl(D), it suffices to establish the inequalities
for all u E C 1 (fJ). In addition, obviously, we only need to verify the first
inequality.
We use the notations introduced in the proof of Theorem 8.15 and con-
sider a function u in the strip lR x [0, 1]. From
[211" [211" [211" t" au
10 u(t, TJ) dt - 10 u(t, 0) dt = 10 10 ae (t, e) dedt,
using the Cauchy-Schwarz inequality, we derive
It u(t,~) dtl' ~ {It 2 u(t, 0) dtl' + I ~ [(Q)} .

2x
Integrating this inequality with respect to 'f] we obtain
fo lfo27r U(t,'f])d{ d'f] S C {lfo27r U(t,O)d{ + IlgradUlll2(Q)}

I
I
for some constant C I . Combining this with (8.16) and (8.18), we see that
lI u lll2(Q) S 27r foIlao('f])12d'f] + II ~~ [2(Q)
27r
S C2{lfo u(t, 0) d{ + II grad ulll2(Q) }
for some constant C2 . Substituting back into the domain D as in the proof
of Theorem 8.15, we see that there exists a constant C3 such that
(8.20)
For this transformation, without loss of generality we have assumed that

the parameter t in the representation of the boundary r is given through
the arc length multiplied by 27r/lfi.
Now let u E CI (f» satisfy u = 0 and grad u = 0 on the boundary
r. Then we can extend u to a continuously differentiable function on lR?
by setting u = 0 in the exterior of D. Choose R large enough that D is
contained in the disk with radius R and center at the origin. Then from
u(x) = j Xl
-R UXI
OU
-;;- dXI,
i:
with the aid of the Cauchy-Schwarz inequality, we see that
lu(x)12 S 2R 1:~ 12 dXI'

Integrating this inequality and using u = 0 outside of D yields
lIulll2(D) S 4R211 grad ulll2(D) (8.21)
for all u E CI(fJ) with u = 0 and gradu = 0 on r.

Finally, for arbitrary u E CI(fJ) we use the function 9 introduced in the
proof of the previous theorem to decompose
u=gu+(1-g)u
and then apply (8.20) to the first term and (8.21) to the second term on
the right-hand side. Observing that
for some constant C 4 depending on g ends the proof. o

Corollary 8.18 On HI(D) the norm II . IIH1(D) is equivalent to each of
the two norms
1/2
lIull:=
( lulil(r) + II grad ulli 2 (D) )
and
2 2 ) 1/2
lIull:= ( lI u ll£2(r) + II grad ull£2(D) .
Proof. This follows from Theorems 8.3, 8.15, and 8.17.
8.3 Weak Solutions to Boundary Value Problems

We will demonstrate how the Sobolev spaces H I / 2 (r) and H- I / 2 (r) occur
in a natural way through the solution of boundary value problems in a
weak formulation. For a function u E HI(D), the integral
Ivl grad ul dx 2
is called the Dirichlet integral. For a harmonic function u the Dirichlet

integral represents the energy of the potential u. Therefore, it is natural to
attempt to develop an approach for the solution of boundary value problems
for harmonic functions in which it is required that the solutions have finite
energy, i.e., they belong to the Sobolev space HI(D). Since the functions
contained in HI (D), in general, are not twice continuously differentiable
in D and since they do not attain boundary values or normal derivatives
on the boundary in the classical sense, we have to extend the classical
formulation of the boundary value problems by generalized versions. Our
procedure will be quite typical in the sense that we use specific properties
of classical solutions to formulate what we mean by a weak solution.
For each solution u E C 2 (D) n CI(fJ) to the Dirichlet problem 6u = 0
in D with u = f on r, by the first Green's Theorem 6.3 we have
Iv gradu· gradvdx = 0
for all v E CI(fJ) with v = 0 on r. Since the integral in this equation
is well defined for all u, v E HI(D), we can introduce the following weak
formulation of the interior Dirichlet problem.
Weak Interior Dirichlet Problem. Find a function u E Hl(D) that
L
satisfies
gradu· gradvdx = 0 (8.22)
for all v E Hl(D) with v = 0 on rand
u =f on r, (8.23)
where f E Hl/2(r) is a given function.
Note that the boundary values for functions in Hl(D) have to be under-
stood in the sense of the trace operator according to Corollary 8.16.
For each solution u E C 2 (D) n C1(D) to the Neumann problem 6.u = 0
in D with au/all = 9 on r by the first Green's theorem we have
L grad u . grad v dx = l gv ds
for all v E C 1 (b). Through the duality pairing (8.10) and the trace the-
orem, the integral on the right is well defined for all 9 E H-l/ 2 (r) and
l
v E Hl(D) by
gv ds := g(v).
Therefore we can introduce the following weak formulation of the interior

Neumann problem.
Weak Interior Neumann Problem. Find a function u E Hl(D) that
L l
satisfies
grad u . grad v dx = gv ds (8.24)
for all v E Hl(D), where g E H- 1 / 2 (r) is given.
Theorem 8.19 The weak interior Dirichlet problem has at most one solu-
tion. Two solutions to the weak interior Neumann problem can differ only
by a constant.
Proof The difference u := Ul - U2 between two solutions to the Dirichlet
problem is a weak solution with homogeneous boundary condition u = 0
on r. Then we may insert v = u in (8.22) and obtain II grad ull£2(D) = O.
Since u = 0 on r, from Corollary 8.18 we see that u = 0 in D. As a conse-
quence of (8.24) for two solutions of the Neumann problem the difference
u := Ul - U2 - c again satisfies II grad ull£2(D) = 0 for all constants c. We
choose c such that fr u ds = O. Then from Corollary 8.18 we obtain u = 0
in D. 0
To establish the existence of weak solutions we will proceed analogously

to the classical case and try to find the solution in the form of a logarithmic
single- or double-layer potential with densities in H- 1/ 2(r) or Hl/2(r).
Therefore, we first need to investigate the properties of these potentials
in our Sobolev space setting. This will be achieved through making use
of corresponding properties in the spaces of uniformly HOlder continuous
functions. For the remainder of this section, by a we will always denote a
Holder exponent with 0 < a < 1.
Theorem 8.20 The operators K : H 1/ 2(r) -+ Hl/2(r) defined by the
logarithmic double-layer potential
1[
(Kip)(x) := -
a 1
( ) In -I- I ds(y),
ip(Y) - a xEr,
7rr vy x-y
and K' : H- 1 / 2 (r) -+ H-l/2(r) defined by the normal derivative of the

logarithmic single-layer potential
(K',¢)(x) := r
~7r lr '¢(y) a va(x) In -Ix_1_1
- y ds(y), x E r,
are compact and adjoint in the dual system (Hl/2(r), H-l/2(r))p(r) , i.e.,
(8.25)
for, all ip E Hl/2(r) and'¢ E H-l/2(r). Here, again, we use the notation
(f,g)P(r) := g(f) for f E Hl/2(r) and g E H- 1/ 2(r).
Proof. We shall show that K is bounded from Hl/2(r) into Hl(r). Then
the assertion of the compactness of K follows from Theorems 2.16 and 8.3.
For functions ip E C1,a«r) from Theorem 7.29, we know that the gradient
of the double-layer potential v with density ip can be expressed in terms
of the derivatives of the single-layer potential w with density dip/ds. In
particular, by (7.42) the tangential derivative is given by
av_
as (x) = -2
1 dip 1
ds (x) - 27r
rdipds (y) av(x)
lr
a 1
In Ix _ yl ds(y), x E r.
From this, with the aid of the jump relations of Theorem 7.27 for double-
layer potentials with uniformly Holder continuous densities, we conclude
that
dK ip 1 [ dip a 1
-d- (x) = - - ( ) In -I- I ds(y),
-d (y) - a x E r.
s 7r r s vx x-y
Performing a partial integration yields
dKip
d; (x)
1
= ;;:
r
lr {ip(Y) -
a a
ip(X)} as(y) av(x) In
1
Ix _ yl ds(y), xEr,
where the integral has to be understood in the sense of a Cauchy principal

value. Now we use a 21r-periodic parameterization r = {x(t) : t E [0,21r)}
of r to transform the integral into
dKcp
Ts(x(t)) =
1211" k(t T)
. t'-T {cp(X(T))-cp(x(t))}dT, tE [0,21r],
o sm--
2
where k is a bounded weakly singular kernel. Thus, by the Cauchy-Schwarz
inequality, we can estimate
r2 IdKcp (X(t))1 2 dt ::; M r r2

11"
10 ds 10 10
21r
11" Icp(X(T)) -
. 2-
t -
T
~(x(t)W dTdt
sm
2
for some constant M. Hence, by Theorem 8.6,
II d~S cp II £2(r) ::; cllcpIIHl/2(r)

for some constant c. Since K has continuous kernel (see Problem 6.1), it is
bounded from L2(r) into L2(r). Therefore we conclude
11KcpllHl (r) ::; CllcpIlHl/2(r)

for all cp E C1,"'(r) and some constant C. This completes the proof of
the boundedness of K from Hl/2(r) into H1(r), since C1,"'(r) is dense in
Hl/2(r) by Theorem 8.2.
For cp, 'l/J E C1,"'(r), clearly (8.25) holds and we can estimate
I(cp, K''l/J)L2(r) I = I(Kcp, 'l/J)£2 (r)I
::; II K CPIlHl(r) 11'l/JIIH-l(r)
::; CllcpIIHl/2(r) 11'l/JIIH-l(r).

From this it follows that
11K''l/JIiH-l/2(r) ::; ClI'l/JIIH-l (r)

for all 'l/J E C1,"'(r) and, because Cl,"'(r) is dense in H-1(r), also for all
'l/J E H-1(r). Therefore K' is bounded from H-1(r) into H- 1/ 2(r) and
the statement on the compactness of K' again follows from Theorems 2.16
and 8.3.
Finally, the validity of (8.25) for all cp E H 1/ 2(r) and 'l/J E H- 1/ 2(r)
follows from the boundedness of K and K' by the denseness of C1''''(r) in
Hl/2(r) and in H-l/2(r). 0
Lemma 8.21 For the trigonometric monomials we have the integrals

a, m=O,
-21f 10r
1 2rr t.
In (4sin2 _) e,mtdt = {
2 __1_ m= ±1,±2, ....
Iml'
Proof. Integrating the geometric sum
m-l
1+2 L eikt + eimt = i (1 - eimt ) cot ~, 0< t < 21f, (8.26)
k=l
1
we obtain
2rr . t
e'mt cot - dt = 21fi m = 1,2, ... , (8.27)
o 2 '
in the sense of Cauchy principal values. Integrating the identity
:t { [e imt - 1] In (4 sin2 ~) } = im eimt In ( 4 sin2 ~) + [e imt - 1] cot ~

yields
r
10
2rr
eimt In (4 sin 2 !) dt
2
= _~ r e cot! dt = _ 21f
2m 1 0 2 m
2rr imt
for m = 1,2, ... , and the statement is proven for m =f. O. To evaluate the
integral for m = a we set
1:= 1 2rr
In ( 4 sin 2 ~) dt.
Then
= 1 2rr
In(4sin 2 t)dt = ~ 14rr
In (4sin2~) dt = I,
and therefore 1=0. o
By Problem 7.2 the operator So in the following theorem corresponds to

the modified single-layer operator of Theorem 7.30 for the case of the unit
circle.
Theorem 8.22 The operator So : HP[O, 21fj -7 HP+1 [0, 21fJ, defined by
(So<p )(t) := 2~ 1 2rr

{In (4 sin 2 t ~ T) - 2} <p( T) dT, t E [0,21fj, (8.28)
is bounded and has a bounded inverse for all p E JR. (For negative p the
integral in {8.28} has to be understood as the duality pairing.)
Proof. From Lemma 8.21, for f m (t) = e imt , we have
Sofm = f3mfm' mE 7J"
°
where 13m = -l/lml for m =/:. and 130 = -2. This implies the bounded-
ness of So : HP[0,27fj-+ HP+l[0,27fj and its invertibility with the inverse
operator SOl: HP+1 [0, 27fj -+ HP[O, 27fj given by
-1 1
So fm = (3m fm' m E 7J"
which again is bounded. 0
Theorem 8.23 The operators S : H- 1/ 2(r) -+ H1/2(r) defined by the

single-layer potential
(Sip)(x) := -
7f r
11 x-y
1
ip(y) In -I- I ds(y), x E r,
and T : H 1 / 2 (r) -+ H- 1 / 2 (r) defined by the normal derivative of the

double-layer potential
1 & ( & 1
(T'ljJ)(x) := :;;: &v(x) lr'ljJ(y) &v(y) In Ix _ yl ds(y), x E r,
are bounded. The operator S is self-adjoint with respect to the dual systems
(H- 1/ 2(r), H1/2(r)h2(r) and (H1/2(r), H- 1/2(r))p(r)' i.e.,
(8.29)
for all ip,'ljJ E H- 1 / 2 (r). The operator T is self-adjoint with respect to the
dual systems (H1/2(r), H- 1/ 2(r))p(r) and (H- 1/ 2(r), H1/2(r)h2(r), i.e.,
(8.30)
for all ip, 'ljJ E H1/2 (r).
Proof. We use a 27f-periodic parameterization r = {x(t) : t E [0,27f)} of r.

By Theorem 8.14, we may assume that the parameter t is given by the arc
length on r multiplied by 27f/lrl. Then, for ip E CO,"'(r), we can write (see
Problem 7.2)
(Sip) (x(t)) = l~~ 121< {_In(4sin2t~T) +P(t,T)}ip(X(T))dT

for t E [0, 27fj, where pis 27f-periodic and continuously differentiable. By the
previous Theorem 8.22, the leading term with the logarithmic singularity
represents a bounded operator from H- 1 / 2 (r) into H 1 / 2 (r). Because the
kernel function p is continuously differentiable, using the Cauchy-Schwarz
inequality and an integration by parts, the second term on the right is seen
to be bounded from HO(r) into Hl(r) and from H-1(r) into HO(r). Hence,
by the interpolation Theorem 8.13, it is also bounded from H- 1 / 2 (r) into
H 1/ 2 (r).
For <p, 't/J E C1,"(r) we use the relation (7.44) between Sand T and a
partial integration to write
(T<p,'t/J)p(r) = ( -d d<P)
d S -d ,'t/J = - (S d<p '-d
-d d't/J) . (8.31)
s s L2(r) S S per)
I
I(T<p,'t/J)lp(r)::; S -d<p
ds
I Hl/2(r)
I -d't/Jds I H-l/2(r)
for some constant C. Since C1'''(r) is dense in H 1 / 2 (r), this implies that
T is bounded from Hl/2(r) into H- 1 /2(r).
The self-adjointness property (8.29) follows as in the proof of Theorem
8.20 by extending it from the case of smooth functions. Finally, combining
(8.29) and (8.31) yields (8.30). 0
Theorem 8.24 The logarithmic single-layer potential defines a bounded

linear operator from H- 1/2(r) into H1(D). The logarithmic double-layer
potential defines a bounded linear operator from Hl/2(r) into Hl(D).
Proof. Let u be the single-layer potential with density <p E CO,"(r). Then,
by Green's theorem and the jump relations of Theorem 7.28, we have
r Igradul dx= lrr

1D
2 ii aau ds=
l/
~
41r
rS<p (<p + K'<p) ds.
Therefore, by Theorems 8.20 and 8.23, with a constant c we can estimate
Since S is bounded from L2(r) into L2(r), using Theorem 8.17, we see that
for some constant C. Now the statement on the single-layer potential follows
from the denseness of CO,"(r) in H- 1 / 2 (r).
The case of the double-layer potential v with density '¢ is treated analo-
gously through the relation
which again follows from Green's theorem and the jump relations. 0
Theorem 8.25 The weak Dirichlet problem has a unique solution. The
mapping taking the given boundary data f E Hl/2(r) into the solution
U E Hl(D) is bounded.
Proof. Analogous to the classical treatment we try to find the solution in

the form of a double-layer potential
1
u(x) = 27r
r
lr 'P(y)
8 1
8v(y) In Ix _ yl ds(y), xED.
If 'P E c1,o:(r), by Theorem 7.29 and Green's theorem, we have U E C1(D)

and
jD
grad U . grad v dx = ~
21r
r T'P v ds
for all v E C1(fJ) and

2u = K'P - 'P on r.
In the first equation, by Theorems 8.23 and 8.24 and the trace theorem,
both sides are continuous bilinear mappings from Hl/2(r) x Hl(D) ---7 <C.
Hence, the double-layer potential with density 'P E Hl/2(r) satisfies (8.22).
In the second equation both sides are again bounded from Hl/2 (r) into
Hl/2(r) by Theorems 8.20 and 8.24 and Corollary 8.16. Therefore u is a
weak solution to the Dirichlet problem provided the density 'P solves the
integral equation
'P - K'P = -2J.
Let 'P E Hl/2(r) be a solution to the homogeneous equation 'P - K'P = o.
Then, since K has continuous kernel (see Problem 6.1), we have 'P E C(r)
and from the classical Theorem 6.20 we deduce that N(I - K) = {O} in
H 1 / 2(r). Now existence of a solution in Hl/2(r) to the inhomogeneous
equation follows from the Riesz theory by Theorems 3.4 and 8.20.
The statement on the continuous dependence of the solution on the given
boundary data is a consequence of the boundedness of the inverse of I - K
(Theorem 3.4) and Theorem 8.24. 0
Theorem 8.26 The weak interior Neumann problem is solvable if and

only if
The opemtor mapping the given boundary data 9 E H-l/2(r) into the
unique solution u E HI (D) satisfying the additional condition Ir u ds = 0
is bounded.
Proof. By Theorem 7.28, the single-layer potential
1
u(x) = -2
7r
1
r
'l/J(y) In -I-
1 I ds(y),
x-y
XED,
with density 'l/J E GO,<>(r) belongs to GI(D). It satisfies
l grad u . grad v dx = ~ £+
('l/J K' 'l/J) v ds
for all v E GI(D). As in the previous proof, both sides of this equation
are continuous from H- I / 2 (r) x HI(D) -+ <C by Theorems 8.20 and 8.24
and the trace theorem. Hence, the single-layer potential u with density
'l/J E H- 1/ 2 (r) is a weak solution to the Neumann problem provided 'l/J
solves the integral equation
'l/J+K''l/J = 2g.
As in the proof of Theorem 8.25 from Theorem 6.20 we can conclude
that N(I +K) = span{1} in H I / 2 (r). Therefore, by Theorem 8.20 and the
Fredholm alternative, the inhomogeneous equation 'l/J+K''l/J = 2g is solvable
Ir
in H- 1 / 2 (r) provided 9 satisfies the solvability condition 9 ds = o. The
necessity of the latter condition for the existence of a weak solution to the
Neumann problem follows from (8.24) for v = 1.
The continuous dependence of the solution on the given boundary data is
a consequence of Theorem 8.24 analogous to the classical case of Theorem
6.29. 0
It is also possible to obtain the weak solvability to the Dirichlet and

Neumann problem through integral equations of the first kind as in Section
7.5 (see Problem 8.5). For the solution of the single-layer potential integral
equation of Section 7.6 for an arc in a Sobolev space setting, we refer to
[107].
Our presentation clearly demonstrates that the results on weak solutions
heavily rely on the classical analysis on single- and double-layer potentials
with uniformly Holder continuous densities. The mapping properties of the
integral operators in the Sobolev spaces essentially follow from the corre-
sponding properties in the Holder spaces through applications of functional
analytic tools, in particular, denseness and continuation arguments. For a
proof of the mapping property of the single-layer operator from Theorem
8.23 with the aid of Lax's Theorem 4.11, we refer to [25).
The major advantage of the weak approach to the boundary integral
equations, besides the fact that it allows less regular boundary data, stems
Problems 151
from the fact that Sobolev spaces are Hilbert spaces rather than just
normed spaces. This, as we have seen, adds some elegance to the analysis.
For a more detailed study of boundary integral equations in Sobolev
space settings in the framework of pseudodifferential operators (see Tay-
lor [170, 171] and Treves [177]) including the extensions of the analysis to
exterior problems and to three-dimensional problems we refer to Wend-
land [185, 186, 187] and the literature therein. For extensions of the ap-
proach presented in this section we refer to Kirsch [84, 86].
Problems
8.1 Work out the proofs of Corollaries 8.7 and 8.8.
8.2 Prove Dini's theorem: Let D C lRffl be compact and let (<pn) be a non-
decreasing sequence of continuous real-valued functions on D converging point-
wise to a continuous function <po Then (<pn) converges uniformly on D to <po
8.3 Let K : [0,211"] x [0,211"] -+ lR be n-times continuously differentiable. Show

that the integral operator A with kernel K is compact from HP[O, 211"] into Hq[O, 211"]
for all Ipl, Iql ~ n.
8.4 Formulate and prove the analog of the Noether theorems in the form of
Corollary 7.24 in the Sobolev spaces HP[O,211"].
Hint: Use the integrals (8.27).
8.5 Show that the operators K, K' , S, and T are related by
ST= -I+K2
on H 1 / 2 (r) and
TS=-I+K'2
on H- 1 / 2 (r). Use this result and Theorems 5.6, 8.20, 8.23, and 8.24 to solve the
weak Dirichlet problem through a single-layer potential with H- 1 / 2 (r) density
and the weak Neumann problem through a double-layer potential with Hl/2(r)
density.
Hint: Use Green's theorem and the jump relations to establish
1 T<pSt/Jds = 1 (K - I)<p(K' + I)t/Jds
for all <p E C1''''(r) and t/J E CO''''(r). In the case of the Dirichlet problem the
single-layer approach has to be amended analogously to Theorem 7.30.
9
The Heat Equation
The temperature distribution u in a homogeneous and isotropic heat con-

ducting medium with conductivity k, heat capacity c, and mass density p
satisfies the partial differential equation
au =
- /'i,D.u
at
where /'i, = k/cp. This is called the equation of heat conduction or, shortly,
the heat equation; it was first derived by Fourier. Simultaneously, the heat
equation also occurs in the description of diffusion processes. The heat
equation is the standard example for a parabolic differential equation. In
this chapter we want to indicate the application of Volterra-type integral
equations of the second kind for the solution of initial boundary value
problems for the heat equation. Without loss of generality we assume the
constant /'i, = 1. For a more comprehensive study of integral equations of the
second kind for the heat equation we refer to Cannon [20], Friedman [45],
and Pogorzelski [145].
9.1 Initial Boundary Value Problem: Uniqueness

The mathematical modeling of the heat conduction in a medium with a
given temperature distribution at some initial time and a given temperature
distribution on the boundary of the medium for all times leads to the
following initial boundary value problem for the heat equation.

9.1 Initial Boundary Value Problem: Uniqueness 153
Initial Boundary Value Problem. Let D c R m , m = 1,2,3, be a

bounded domain with boundary r := aD and let T denote a positive number.
Find a function u E C(tJ x [0, T]) that is twice continuously differentiable
with respect to the space variable x and continuously differentiable with
respect to the time variable t in D x (0, Tj and that satisfies the heat equation
at = D.u
au in D x (O,Tj, (9.1)
the initial condition

u(·,O)=w inD, (9.2)
and the Dirichlet boundary condition
u =f on r x [0, Tj. (9.3)

Here, w E C(lJ) and f E C(r x [0, T]) are given continuous functions
subject to the compatibility condition
w=f(·,O) onr. (9.4)
Our aim is to establish that this initial boundary value problem has a
unique solution that depends continuously on the given initial and bound-
ary data. As in the case of the Dirichlet problem for the Laplace equation,
uniqueness and continuous dependence follow from a maximum-minimum
principle.
Theorem 9.1 (Weak Maximum-Minimum Principle) Assume that
u E C(tJ x [0, T]) is twice continuously differentiable with respect to the
space variable and continuously differentiable with respect to the time vari-
able and solves the heat equation in D x (0, Tj. Then u attains both its
maximum and its minimum on the parabolic boundary
B := {(x, 0) : XED} U {(x, t) : x E r, t E [0, T]}.

Proof. We confine ourselves to the proof for the maximum. Define a con-
tinuous function v on tJ x [0, Tj by
v(x, t) := u(x, t) + e(T - t)

with e > 0. Assume that v attains its maximum value in the parabolic
interior D x (0, Tj, i.e., there exists a point (x, t) E D x (0, Tj such that
v(x, t) = mp.x v(y, r).
(y,r)EDx[O,T)
Then the necessary conditions for a maximum

a2 v
a 2 (x, t) ~ 0,
Xi
i = 1, ... ,m,
154 9. The Heat Equation
and
must be fulfilled. Hence, we have
t6.v(x, t) -
av
at (x, t) :::; 0,
which is a contradiction to
av = c > 0
t6.v - -
at
throughout D X (0, T].
Therefore, v attains its maximum on the parabolic boundary B. Then
we can conclude that
u(x, t) :::; v(x, t):::; max v(y, T):::; max u(y, T)

(y,r)EB (y,r)EB
+ cT
for all (x, t) E jj x [0, T]. Because c can be chosen arbitrarily, it follows
that
u(x,t):::; max U(y,T),
(y,r)EB
which ends the proof. o
We wish to mention that analogous to the maximum-minimum princi-

ple Theorem 6.8 for harmonic functions, there is also a strong maximum-
minimum principle, stating that, under the conditions of Theorem 9.1, a
solution to the heat equation cannot assume its maximum or minimum in
the parabolic interior unless it is constant. For a proof in the case m = 1
we refer to Cannon [20].
The maximum-minimum principle implies uniqueness and continuous de-
pendence for the initial boundary value problem as formulated in the fol-
lowing two theorems.
Theorem 9.2 The initial boundary value problem for the heat equation
has at most one solution.
Theorem 9.3 The solution to the initial boundary value problem for the
heat equation depends continuously in the maximum norm on the given
initial and boundary data.
In the statement of Theorem 9.3 we have tacitly assumed that we will

be able to establish existence of the solution.
9.2 Heat Potentials 155
9.2 Heat Potentials

The function
G(x,t;y,r):=
v'47f(t-r)
1 m exp{-~~-YI~},
t-r
t > r, (9.5)
is called fundamental solution of the heat equation. Straightforward dif-

ferentiation shows that G satisfies the heat equation with respect to the
variables x and t. With the aid of this fundamental solution, as in the
analysis of Laplace's equation in Chapter 6, we will define so-called heat
potentials. Then we shall use these potentials to solve the initial boundary
value problem. For the remainder of this chapter, in the case m 2 2, we will
always assume that the bounded domain D c lRm is of class C 2 • We will
take care of the initial condition through the Poisson integral introduced
in the following theorem.
Theorem 9.4 Let w be a continuous function on lRm with compact sup-
port, i. e., w vanishes outside some compact set. Then the Poisson integral
u(x, t) := ~m
47ft irlR", exp {_IX ~tY12} w(y) dy (9.6)
defines an infinitely differentiable solution of the heat equation in

lRm x (0,00). It can be continuously extended into lRm x [0,00) with initial
values
u(· ,0) = w. (9.7)
Proof. Obviously, for all x E lRm and t E (0,00) differentiation and inte-
gration can be interchanged to show that the Poisson integral is infinitely
differentiable and satisfies the heat equation. Substituting y = x + 20 z,
we obtain
~m
47ft
r
ilR'"
exp{_IX~YI2}w(Y)dY= ~m
t v 7r
r
ilR",
w(x+20z)e- 1zI2 dz.
°
Since w has compact support, it is bounded and uniformly continuous. Let
M be a bound for w. Then for € > choose R such that
~
...fi
1 Izl::::R
e- 1z12 dz < ~.
4M
Because w is uniformly continuous there exists 8 > ° such that
€
Iw(x) - w(y)1 < "2
for all x and y with Ix - yl < 8. Setting TJ = 82 /4R 2 and making use of the
integral (see Problem 9.1)
(9.8)
and the fact that for Izl ::; Rand t < TJ we have 2Vtlzl < 2,fiiR = 8, we
deduce that
I~m Lm exp{-IX~tYI2}W(Y)dY-W(X)1
= I~m Lm {W(X+2v'iZ)-w(x)}e-lzI2dZI
< ~ r e-1z12 dz + 2Mm r e- 1z12 dz < e

2Vii Jlzl~R Vii J1z1?R
for all x E lRm
~D.
and all t < TJ. This implies continuity of u at t = ° and
0
In order to deal with the boundary condition we will need surface heat
potentials. Analogous to the terminology used for harmonic functions we
define single- and double-layer potentials. For a function <p E C(r x [0, T])
the single-layer heat potential is given by
u(x,t):= lot i G(x,t;y,r)<p(y,r)ds(y)dr (9.9)

and the double-layer heat potential by
( ) rt r8G(x,t;y,r)
u x,t := J o J r
8v(y)
( ) ()
<p y,r ds y dr. (9.10)
These potentials are well defined for xED and t E (0, T] with the time
integral to be understood as an improper integral with respect to' the upper
limit. For the double-layer potential we assume the unit normal vector v
to the boundary surface r to be directed into the exterior of D.
In the case of one space d:mension for an interval D := (a, b) the two
potentials have the form
u(x, t) := i t 1
o y'4n(t - r)
{ (x - a)2 }
exp - 4(
t- r
) <p(a, r) dr
(9.11)
t 1(x - b)2} {
+ Jo y'4n(t-r) exp -4(t-r) <p(b,r)dr
and
u(x, t) := i t a-x
o 4 Vii y'(t - r)
3
{ (x-a)2}
exp - 4(
t- r
) <p(a, r) dr
i
(9.12)
t X - b {(x - b)2 }
+ 3 exp - 4( ) <p(b,r) dr.
o 4Viiy'(t-r) t-r
Standard analysis again shows that both the single- and double-layer
heat potentials are infinitely differentiable solutions to the heat equation
with initial values u(· ,0) = °

in D X (0, Tj. In addition they can be continuously extended into D X [0, Tj
in D. For the discussion of the boundary
values we will confine our analysis to the double-layer potential.
Theorem 9.5 The double-layer heat potential with continuous density <p
can be continuously extended from D X (0, Tj into j) X (0, Tj with limiting
values
( )= 10t 1r(8G(x,8v(y)tj y, T) <P (y, T) ds ()Y dT -

u x, t
1 (
2" <P x, t
) (9.13)
for x E rand t E (0, Tj. Here the time integral exists as improper integral.
Proof. We carry out the proof for the cases m = 2, 3, and leave it to the
reader to work out the details for the simpler case m = 1 (see Problem
9.3). For xED we may interchange the integrations over rand [0, tJ, since
the integrand is continuous on r X [0, tj with value zero at the upper limit
T = t. Then we substitute for the time integral
u=
Ix-yl
--'2v't
--:======
- T
to obtain
u(x, t) = i t
oJ47r(t-T)
1
m
[v(y) . (x - y)
r
2(
t-T
{Ix - Yl2) } <p(y, T) ds(y)dT
) exp - 4(
t-T
= ~ (v(y) . (x - y) roo u m- 1e- 002 <p (y, t _ Ix - Y12) duds(y).

v'i 1r Ix - ylm 1Ix-yl/2Vt 4u 2
Therefore, we can view the double-layer heat potential as a harmonic
double-layer potential with the density
./,( ) 1
'f' x, y, t := '- m
V 7r
1 00
Ix-yl/2Vt
U
m-l 2
e _00 <p ( y, t - Ix4- 2Y12) dU
U
= Ix -
t
ylm 10 J47r(t _ T)
1
m
1
2(t _ T) exp - 4(t _ T)
{Ix - Yl2 }
<p(y, T) dT
depending on the time t as a parameter. First, we will show that 'IjJ is

continuous on JRm x r x (0, Tj with
lim 'IjJ(x, y, t) = 'Ym<P(y, t) (9.14)
x-+y
for all y E r and t E (0, Tj, where
and the limit holds uniformly on r and compact subintervals of (0, T].
Clearly the function 'ljJ is continuous for all x -=f. y and all t E (0, T]. We
establish the limit (9.14) by splitting
'ljJ(x, y, t) = 1m
r;;; lVix=YT O'm-1 e-a 2 <p (
y, t - Ix 4- 2Y12) dO'
y 7f Ix-YI/20 0'
+
Y
1m
r;;;
7f
1Vix=YT
00
0' m-1 e _a
2 {(
<p y, t - Ix4-
0'
Y12)
2 - <p (y,)t } d0'
+ <p (y, t ) 1m
r;;;
y7f
1Vix=YT
00
0' m-1 e -
a2 d0' =: I 1 + I 2 + I 3·
Obviously we have limx-ty 11 (x, y, t) = °
uniformly on r and on com-
pact subintervals of (0, T] and limx-ty h(x, y, t) = ,m<P(y, t) uniformly on
r x [0, T]. Since there exists 8 >
such that
° °
l<p(y, t1) - <p(y, t2)1 < c
for all y E r and all tb t2 E [0, T] with IiI -t21 < 8. Then for alllx-yl < 48
and all 0' 2: ~ we have
Ix - Yl2 < Ix - yl <8

40'2 - 4
and therefore
I<p (y, t - Ix 40'2

- Y12) - <p(y, t) I< c.
Hence,
Ih(x, y, t)1 < ~ m /00 O'm-1 e -a 2 dO' :::; Imc
y7f Vix=YT
and thus limx-ty 12 (x, y, t) = °
uniformly on r x [0, T]. Now, using
12 = 1/27f and 13 = 1/47f, from Theorem 6.17 we can deduce that
lim u(x - hV(x), t)

h-tO
=
10t lrr 8G~,v tr'
y
1 <p(x, t) (9.15)
T) <p(y, T) ds(y)dT- -2
with uniform convergence on r and on compact subintervals of (0, T]. Note

that we need a slightly generalized version of Theorem 6.17 where the
density is allowed to also depend on x and a parameter t. It is left to the
reader to go over the proof of Theorem 6.17 to see that it can be extended
to this more general case.
Finally, the statement on the continuity of the double-layer heat poten-
tial on jj x (0, T] follows from the fact that the right-hand side of (9.15) is
continuous on r x [0, T] by the following Theorem 9.6. 0
Consider the integral operator H : C(r x [0, T]) -+ C(r x [0, T]) defined
by
(H cp ) x,t
( ) := 2 r (8G(x,t;y,r)
10 lr 8v(y)
( ) ()
cp y,r ds y dr (9.16)
for x E rand t E (0, T] with improper integral over [0, t]. For its kernel
2 8G(x,t;y,r) = 1 m v(y)·(x-y) exp{_lx-YI2}

8v(y) J47r(t - r) t- r 4(t - r) ,
using Lemma 6.15, we can estimate
21 8G (x,t;y,r) 1< L Ix - Yl2 {IX - Y12}

8v(y) -...;t-::7-m t _ r exp - 4(t _ r) , t > r,
with some constant L. Applying the inequality (see Problem 9.2)
(9.17)
°
which is valid for all < s, (3 < 00, for the special case s = Ix - Yl2 / 4( t - r)
and (3 = 1 + m/2 - a, we derive the further estimate
2 1 8G (x,t;y,r) I M t > r, x f:. y, (9.18)

8v(y) :::; (t - r)o<lx - ylm-20<'
°
for all < a < 1+m/2 and some constant M depending on L and a. Hence,
choosing 1/2 < a < 1, the kernel of H is seen to be weakly singular with
respect to both the integrals over r and over time. Therefore, proceeding
as in the proofs of Theorems 2.22 and 2.23, we can establish the following
result.
Theorem 9.6 The double-layer heat potential operator H is compact from

C(r x [0, T]) into C(r x [0, T]).
Summarizing, the double-layer heat potential with continuous density is

continuous in jj x [0, T] with the exception of possible discontinuities on
the boundary r for t = 0. From (9.18) we observe that (Hcp)(., 0) = 0.
Therefore we can expect continuity in [) x [0, T] only if cp(. ,0) = on r. °
°
And indeed, in this case we can extend the density cp continuously onto
r x (-00, T] by setting cp(. ,t) = for t < 0. Then, as a consequence of
Theorem 9.5, we can state the following corollary.
Corollary 9.7 The double-layer heat potential is continuous on [) x [0, T]

provided the continuous density satisfies cp(. ,0) = 0.
9.3 Initial Boundary Value Problem: Existence

We now return to the initial boundary value problem. First we reduce it to
the special case of a homogeneous initial condition. To this end, consider
u(x, t) := vex) + ~m
47ft JlRm
r exp {_Ix -4tY12} {w(y) - v(y)} dy. (9.19)
Here, v denotes the unique solution to the Dirichlet problem for Laplace's
equation in D with boundary condition v = f(', 0) on r (see Theorem
6.22). Because of the compatibility condition (9.4), the function w - v can
be continuously extended into lRm by setting it to zero outside D. Then,
from Theorems 6.6 and 9.4, we deduce that (9.19) defines an infinitely
differentiable solution to the heat equation in D x (0, T] that is continuous
in D x [0, T] and satisfies the initial condition u(·, 0) = w in D. Hence, by
superposition, it suffices to treat the special case of the initial boundary
value problem with initial condition
u(',O) = 0 in D (9.20)
and boundary condition
u =f on r x [0, TJ, (9.21)
where f satisfies the compatibility condition
f(· ,0) =0 on r. (9.22)
In order to deal with the boundary condition (9.21) we seek the solution
in the form of a double-layer heat potential. Again we will leave the case
where m = 1 as an exercise for the reader (see Problem 9.4).
Theorem 9.8 The double-layer heat potential (9.10) solves the initial bound-
ary value problem (9.20)-(9.22) provided that the continuous density 'P for
all x E rand t E (0, T] solves the integral equation
() t Jrr 8C(x,t;y,r)
'P x, t - 2 Jo 8v(y)
( ) () f()
'P y, r ds y dr = -2 x, t . (9.23)
Proof. This follows from Theorem 9.5 and Corollary 9.7. The compatibility
condition (9.22) for f ensures that 'P(' ,0) = 0 for solutions to (9.23). 0
By Theorem 9.6 and the Riesz theory Theorem 3.4, the inhomogeneous
integral equation 'P - H'P = - 2f is solvable if the corresponding homoge-
neous integral equation 'P - H'P = 0 has only the trivial solution. From the
weak singularity (9.18) we see that for the integral operator H the integra-
tion over r corresponds to a compact operator. Therefore we can estimate
for the maximum norm with respect to the space variable by
t1
II(H'P)(', t)lloo,r ~ C Jo (t _ r)a II'P(' ,r)lloo,r dr (9.24)
9.3 Initial Boundary Value Problem: Existence 161
for all 0 < t ~ T and some constant 0 depending on rand a. Note that
(see Problem 9.1)
t dr 1 (ds
1u (t - r)a(r - a)(3 - (t - a)a+(3-1 10
s(3(1- s)a
(9.25)
1 r 1 ds
~ (t - a)a+,8-1 10 [s(1- s)]a
for f3 ~ a < 1. Then, from (9.24), by induction and by interchanging the

integrals in
t r (t - r)a(r
10 10
da t t dr
- a)(3 dr = 10 1u (t _ r)a(r _ a)(3 da,
we find that
II (Hk'PK ,t)lloo,r ~ Ok I k- 1 lot (t _ r);(a-1)+1 11'1'(" r)lIoo,r dr
for all k E IN and all 0 <t ~ T, where
r1 ds
1:= 10 [s(1 - s)]a .
Hence there exists an integer k such that for H := Hk we have an estimate

of the form
II(H'P)(.,t)lIoo,r ~ M lot 11'P(',r)lIoo,r dr (9.26)
for all 0 ~ t ~ T and some constant M.

Now, let 'I' be a solution to the homogeneous equation 'I' - H'P = O.
Then, by iteration, 'I' also solves 'I' - H'P = O. Proceeding as in the proof
of Theorem 3.10, i.e., as for Volterra equations with a continuous kernel,
from (9.26) we can derive that
for all n E IN. Hence 11'1'(" t)lIoo,r = 0 for all t E [0, Tj, i.e., 'I' = O. Thus we
have established the following existence theorem.
Theorem 9.9 The initial boundary value problem for the heat equation
has a unique solution.
We wish to mention that, by using a single-layer heat potential, the
analysis of this chapter can be carried over to an initial boundary value
problem where the Dirichlet boundary condition is replaced by a Neumann
boundary condition. For the one-dimensional case, see Problem 9.5.
The idea to proceed along the same line in the case of the heat equation
as for the Laplace equation is due to Holmgren [73, 74] and Gevrey [48].
The first rigorous existence proof was given by Muntz [132] using succes-
sive approximations for the integral equation in two dimensions. Integral
equations of the first kind for the heat equation have been investigated by
Arnold and Noon [5], Costabel [26], and Hsiao and Saranen [76] in Sobolev
spaces, and by Baderko [11] in Holder spaces.
Problems
9.1 Prove the integrals (9.8) and (9.25).
9.2 Prove the inequality (9.17).
9.3 Carry out the proof of Theorem 9.5 in the one-dimensional case.
9.4 Show that the double-layer heat potential (9.12) solves the initial boundary
value problem for the heat equation in (a, b) x (0, T] with homogeneous initial
condition u(·, 0) = 0 and Dirichlet boundary conditions u(a,·) = f(a,·) and
u(b,·) = f(b,·) (with compatibility condition f(a,O) = f(b,O) = 0) provided the
densities satisfy the system of Volterra integral equations
<pea, t) -it h(t, T)<p(b, T) dT = -2f(a, t),
<pCb, t) -it h(t, T)<p(a, T) dT = -2f(b, t)

for 0 :::; t :::; T with kernel
._ a - b 1 { (a -
h(t,T) .- 2ft ~)3 exp - 4(t _ T)
b?} , 0:::; T < t.
Establish existence and uniqueness for this system.
9.5 Show that the single-layer heat potential (9.11) solves the initial boundary
value problem for the heat equation in (a, b) x (0, T] with homogeneous initial
condition u(- ,0) = 0 and Neumann boundary conditions -8u(a, ·)/8x = g(a,·)
and 8u(b, ·)/8x = g(b,·) (with compatibility condition g(a,O) = g(b,O) = 0)
provided the densities satisfy the system of Volterra integral equations
<pea, t) + it h(t, T)<p(b, T) dT = 2g(a, t),
<p(b,t) + it h(t,T)<p(a,T)dT = 2g(b,t)

for 0 :::; t :::; T with the kernel h given as in Problem 9.4. Establish existence and
uniqueness for this system.
10
Operator Approximations
In subsequent chapters we will study the numerical solution of integral

equations. It is our intention to provide the basic tools for the investigation
of approximate solution methods and their error analysis. We do not aim
at a complete review of all the various numerical methods that have been
developed in the literature. However, we will develop some of the principal
ideas and illustrate them with a few instructive examples.
A fundamental concept for approximately solving an operator equation
Acp = I
with a bounded linear operator A : X -+ Y mapping a Banach space X

into a Banach space Y is to replace it by an equation
with an approximating sequence of bounded linear operators An : X -+ Y

and an approximating sequence In -+ I, n -+ 00. For practical purposes,
the approximating equations will be chosen so that they can be reduced to
solving a finite-dimensional linear system. In this chapter we will provide
an error analysis for such general approximation schemes. In particular,
we will derive convergence results and error estimates for the cases where
we have either norm or pointwise convergence of the sequence An -+ A,
n -+ 00. For the latter case, we will present the concept of collectively
compact operators.

164 10. Operator Approximations
10.1 Approximations via Norm Convergence

Theorem 10.1 Let X and Y be Banach spaces and let A : X ~ Y be a
bounded linear operator with a bounded inverse A-l : Y ~ X. Assume the
sequence An : X ~ Y of bounded linear operators to be norm convergent
IIAn - All ~ 0, n ~ 00. Then for sufficiently large n, more precisely for
all n with
IIA-1(A n - A)II < 1,
the inverse operators A;;l : Y ~ X exist and are bounded by
(10.1)
For the solutions of the equations
Arp =f and Anrpn = fn

we have the error estimate
IIA-11I
IIrpn - rpll :::; 1 _ IIA-l(An _ A)II {II (An - A)rpll + Ilfn - fll}·
Proof. If IIA-l(A n - A)II < 1, then by the Neumann series Theorem 2.9,
the inverse [1 - A-l(A - An)]-l of 1 - A-l(A - An) = A-l An exists and
is bounded by
But then [1 - A-l(A - An)]-l A-l is the inverse of An and bounded by

(10.1). The error estimate follows from

In Theorem 10.1 from the unique solvability of the original equation we
conclude unique solvability of the approximating equation provided the
approximation is sufficiently close in the operator norm. The converse sit-
uation is described in the following theorem.
Theorem 10.2 Assume there exists some N E IN such that for all n 2:: N
the inverse operators A;;l : Y ~ X exist and are uniformly bounded. Then
the inverse operator A-l : Y ~ X exists and is bounded by
(10.2)
10.2 Uniform Boundedness Principle 165
for all n with IIA~l(An - A)II < 1. For the solutions of the equations
A<p =f and An<Pn = fn

Proof. This follows from Theorem 10.1 by interchanging the roles of A and
An. 0
Note that Theorem 10.2 provides an error bound that, in principle, can
be evaluated, because it involves A~l and <Pn but neither A-l nor <po The
error bound of Theorem 10.1 shows that the accuracy of the approximate
solution depends on how well An<P approximates A<p for the exact solution
as expressed through the following corollary.
Corollary 10.3 Under the assumptions of Theorem 10.1 we have the error
estimate
(10.3)
for all sufficiently large n and some constant C.
Proof. This is an immediate consequence of Theorem 10.1. o

In Chapter 11 we will apply the error analysis of Theorems 10.1 and 10.2
to the approximation of integral equations of the second kind by replacing
the kernels through so-called degenerate kernels.
10.2 Uniform Boundedness Principle

To develop a similar analysis for the case where the sequence (An) is merely
pointwise convergent, i.e., An<p --+ A<p, n --+ 00, for all <p, we will have to
bridge the gap between norm and pointwise convergence. This goal will be
achieved through compactness properties and one of the necessary tools
will be provided by the following principle of uniform boundedness.
Theorem 10.4 Let the sequence An : X --+ Y of bounded linear operators

mapping a Banach space X into a normed space Y be pointwise bounded,
i.e, for each <p E X there exists a positive number Ctp depending on <p such
that IIAn<P11 ~ Ctp for all n E IN. Then the sequence (An) is uniformly
bounded, i.e., there exists some constant C such that IIAnl1 ~ C for all
n E IN.
Proof. In the first step, by an indirect proof we establish that positive

constants M and P and an element 'l/J E X can be chosen such that
IIAn'Pll ~ M (10.4)
for all 'P E X with II'P - 'l/JII ~ P and all n E IN. Assume that this is not
possible. Then, by induction, we construct sequences (nk) in IN, (Pk) in JR,
and ('Pk) in X such that
IIAnk'Pll ~ k
for k = 0,1,2, ... and all 'P with II'P - 'Pkll ~ Pk and
1 1
0< Pk ~ '2 Pk-l, lI'Pk - 'Pk-ll1 ~ '2 Pk-l
for k = 1,2, .... We initiate the induction by setting no = 1, Po = 1,
and 'Po = o. Assume that nk E IN, Pk > 0, and 'Pk E X are given. Then
there exist nk+l E IN and 'Pk+l E X satisfying lI'Pk+1 - 'Pkll ~ Pk/2 and
IIAnk +1 'Pk+ll1 ~ k + 2. Otherwise, we would have IIAn'Pll ~ k + 2 for all
'P E X with II'P - 'Pkll ~ Pk/2 and all n E IN, and this contradicts our
assumption. Set
Pk+1 ;= min (p; , IIAn~+ll1) ~ P; .

Then for all cp E X with II'P - 'Pk+111 ~ Pk+!. by the triangle inequality we
have
IIAnk+1 'PII ~ IIAnk+1 'Pk+1l1- IIAnk+1 (cp - 'Pk+dll ~k + 1,

since IIAnk+l ('P - 'Pk+l)1I ~ IIAnk+1 I1pk+1 ~ 1.
For j > k, using the geometric series we have
lI'Pk - 'Pj II ~ lI'Pk - 'Pk+ll1 + ... + lI'Pj-l - 'Pj II
1 1
~ '2 Pk + ... + '2 pj-l ~ Pk·
Therefore, ('Pk) is a Cauchy sequence and converges to some element 'P in

the Banach space X. From II'Pk - cpjll :5 Pk for all j ~ k by passing to the
limit j -+ 00 we see that II'Pk - 'PII :5 Pk for all k E IN. Therefore, we have
IIAn,,'P11 ~ k for all k E IN, which is a contradiction to the boundedness of
the sequence (An'P).
Now, in the second step, from the validity of (10.4) we deduce for each
'P E X with II'PII ~ 1 and for all n E IN the estimate
IIAn'Pll = !. IIAn(P'P + 'l/J) - An'l/JII ~ 2M .

P P
This completes the proof. o
10.3 Collectively Compact Operators 167
As a simple consequence of the uniform boundedness principle we note

that the limit operator of a pointwise convergent sequence of bounded linear
operators again is bounded.
Corollary 10.5 Let X be a Banach space, Y be a normed space, and the

bounded linear operators An : X -+ Y be pointwise convergent with limit
operator A : X -+ Y. Then the convergence is uniform on compact subsets
U of X, i.e.,
sup IIAn consider the open balls B(<p;r) = {'I/I EX: 11'1/1- <p1I < r}
with center <P E X and radius r = c/3C, where C is a bound on the
operators An. Clearly
Uc U
B(<p;r)
<pEU
forms an open covering of U. Since U is compact, there exists a finite

sub covering
m
U c U B(<pj;r).
j=l
Pointwise convergence of (An) guarantees the existence of an integer N(c)

such that
c
IIAn<Pj - A<pj II < 3
for all n Z N (c) and all j = 1, ... ,m. Now let <p E U be arbitrary. Then <p
is contained in some ball B(<pj;r) with center <Pj. Hence for all n Z N(c)
we have
Therefore the convergence is uniform on U. o
10.3 Collectively Compact Operators

Motivated through Corollary 10.5 and following Anselone [3], we introduce
the concept of collectively compact operators.
Definition 10.6 A set A = {A : X -+ Y} of linear operators mapping a

normed space X into a normed space Y is called collectively compact if for
each bounded set U c X the image set A(U) = {A<p : <p E U, A E A} is
relatively compact.
Clearly, every operator in a collectively compact set is compact. Each

finite set of compact operators is collectively compact. A sequence (An) is
called collectively compact when the corresponding set is. Pointwise con-
vergence An -+ A, n -+ 00, of a collectively compact sequence implies
compactness of the limit operator, since
A(U) c {AnCP : cP E U, n E IN}.
Theorem 10.7 Let X, Z be normed spaces and let Y be a Banach space.

Let A be a collectively compact set of operators mapping X into Y and let
Ln : Y -+ Z be a pointwise convergent sequence of bounded linear operators
with limit operator L : Y -+ Z. Then
uniformly for all A E A, i. e.,
sup II(Ln - L)AII-+ 0, n -+ 00.

AEA
Proof. The set U := {Acp : Ilcpli :::; 1, A E A} is relatively compact. By

Corollary 10.5 the convergence Ln'I/J -+ L'I/J, n -+ 00, is uniform for all
'I/J E U. Hence, for every e > 0 there exists an integer N(e) such that
II(Ln - L)Acpli <e

for all n 2: N(e), all cp E X with IIcpll :::; 1, and all A E A. Therefore
for all n 2: N(e) and all A E A. D
Corollary 10.8 Let X be a Banach space and let An : X -+ X be a

collectively compact and pointwise convergent sequence with limit operator
A : X -+ X. Then
II (An - A)AII-+ 0 and II (An - A)Anll-+ 0, n -+ 00.
lOA Approximations via Pointwise Convergence

The following error analysis for equations of the second kind takes ad-
vantage of Corollary 10.8 and is due to Brakhage [17] and Anselone and
Moore [4].
Theorem 10.9 Let A : X -+ X be a compact linear operator in a Banach

space X and let I - A be injective. Assume the sequence An : X -+ X to
lOA Approximations via Pointwise Convergence 169
be collectively compact and pointwise convergent An<P --+ A<p, n --+ 00, for
all <p EX. Then for sufficiently large n, more precisely for all n with
the inverse operators (1 - An)-l : X --+ X exist and are bounded by
II( I - An
)-111 <
-
1 + 11(1 - A)-l An II
1 -II(I _ A)-l(An _ A)Anll . (10.5)
<P - A<p = f and <Pn - An<Pn = fn

1 + 11(1 - A)-l An II
II<Pn - <p11 :::; 1_ II(I _ A)-l(An _ A)Anll {II (An - A)<p11 + IIfn - fll}·
Proof. By the Riesz Theorem 3.4 the inverse operator (I - A)-l : X --+ X
exists and is bounded. The identity
(I - A)-l = I + (I - A)-l A
suggests
B n := I + (1 - A)- lA n
as an approximate inverse for I - An. Elementary calculations yield
(10.6)
where
Sn := (1 - A)-l(An - A)An-
From Corollary 10.8 we conclude that IISnl1 --+ 0, n --+ 00. For IISnl1 < 1 the
Neumann series Theorem 2.9 implies that (1 - Sn)-l exists and is bounded
by
-1 1
II(I - Sn) II:::; 1- IISnll .
Now (10.6) implies first that I - An is injective, and therefore, since An is
compact, by Theorem 3.4 the inverse (1 - An)-l exists. Then (10.6) also
yields (I - An)-l = (I - Sn)-l B n , whence the estimate (10.5) follows. The
error estimate follows from
(I - An)(<Pn - <p) = fn - f + (An - A)<p,
Analogous to Theorem 10.1, in Theorem 10.9 from the unique solvability

of the original equation we conclude unique solvability of the approximat-
ing equation provided the approximation is sufficiently close. The converse
situation is described through the following theorem.
Theorem 10.10 Assume there exists some N E 1N such that for all n 2:: N
the inverse operators (1 - An)-l exist and are uniformly bounded. Then the
inverse (1 - A)-l exists and is bounded by
(10.7)
for all n with
<p - A<p = f and <Pn - An<Pn = fn
Proof. This follows from Theorem 10.9 by interchanging the roles of A and
An. 0
We note that Theorem 10.10 provides an error estimate that, in prin-

ciple, can be evaluated, since it involves (J - An)-l and <Pn but neither
(J - A)-l nor <po From Theorem 10.9 we can conclude that the accuracy
of the approximate solution essentially depends on how well An<P approxi-
mates A<p for the exact solution as expressed in the following corollary.
Corollary 10.11 Under the assumptions of Theorem 10.9 we have the

error estimate
(10.8)
Proof. This is an immediate consequence of Theorem 10.9. o
In Chapter 12 we will apply the error analysis of Theorems 10.9 and

10.10 to the approximate solution of integral equations of the second kind
by using numerical quadratures.
10.5 Successive Approximations

We now briefly revisit the investigation of the convergence of successive
approximations
to solve the equation of the second kind cp - Acp = I, where A : X -+ X is a

bounded linear operator in a Banach space X. So far we have used IIAII < 1
as a sufficient condition for convergence (see Theorem 2.10). It is our aim
now to show that, similar to the case of finite-dimensional linear equations,
convergence of successive approximations can be characterized through the
spectral radius of the operator A. For the notion of the resolvent set p(A),
the resolvent R(A; A), the spectrum u(A), and the spectral radius r(A) of
a bounded linear operator A recall Definition 3.8.
We first note that the basic concepts of classical holomorphic function
theory, i.e., derivative, integral, analyticity, etc., and its fundamental the-
orems, i.e., Cauchy's integral theorem and integral formula, Taylor and
Laurent series, etc., can be canonically extended to functions I with val-
ues in a Banach space X depending on a complex variable varying on an
open domain D C <C. In principle, all definitions and theorems, including
their proofs, are obtained by replacing absolute values II(z)1 in the classical
theory through the norm II I (z) II in X.
Theorem 10.12 Let A: X -+ X be a bounded linear operator mapping a
Banach space X into itself. Then the Neumann series
L
00
(AI - A)-l = A- k- 1Ak (10.9)

k=O
converges in the operator norm for all IAI > r(A) and diverges for all
IAI < r(A).
Proof. Provided the series (10.9) converges it defines a bounded linear op-
erator S{A). As in the proof of Theorem 2.9, this operator S{A) can be seen
to be the inverse of AI - A, since IAI-n-1llAnll -+ 0, n -+ 00, is necessary
for the convergence of the series (10.9).
Let AO belong to the resolvent set p{A) of A. Then, by Theorem 2.9, for
all A with
IA - Aol IIR{Ao; A) II < 1
the series 00
k=O
converges in the operator norm and defines a bounded linear operator with
T{A) = R{AO; A) [I -{AO-A)R{AO; A)]-l = [I -(AO-A)R(AO; A)]-l R{AO; A).

Hence,
(AI - A)T{A) = [AoI - A - (AO - A)I] T(A)
= [I - (AO - A)R(AO; A)] (AoI - A)T(A) =I

and similarly T(A)(>.J - A) = I. Therefore A E p(A) and T(A) = R(A; A).

In particular, this means that the resolvent set p(A) is open and that the
resolvent R(A; A) is an analytic function from the resolvent set into the
Banach space L(X, X) of bounded linear operators on X.
As in classical analytic function theory, we can expand R(A; A) into a
uniquely determined Laurent expansion
00
R(A;A) = LA-k-1Ak
k=O
with bounded linear operators Ak : X -+ X such that the series converges
with respect to the operator norm for alllAI > r(A). For IAI > IIAII, again
by Theorem 2.9, we already know that R(A; A) is given by the Neumann
series (10.9). Therefore from the uniqueness of the Laurent expansion we
conclude that the Neumann series is the Laurent expansion and hence
converges for all IAI > r(A).
On the other hand, assume that the Neumann series converges for some
IAol < r(A). Then there exists a constant M such that IAol-k-lIlAkll ~ M
for all k E IN. Hence, for all A with IAI > IAol the Neumann series has
a convergent geometric series as a majorant, and thus it converges in the
Banach space X (see Problem 1.5). Its limit represents the inverse of AI -A.
Hence all A with IAI > IAol belong to the resolvent set p(A), which is a
contradiction to IAol < r(A). 0
Theorem 10.13 Let A: X -+ X be a bounded linear operator in a Banach

space X with spectral radius r(A) < 1. Then the successive approximations
CPn+l := ACPn + f, n = 0,1,2, ... ,
converge for each f E X and each CPo E X to the unique solution of
cP - Acp = f·
Proof. This follows, as in the proof of Theorem 2.10, from the convergence
of the Neumann series for A = 1 > r(A). 0
That the condition r(A) < 1 for convergence cannot be weakened is

demonstrated by the following remark.
Remark 10.14 Let A have spectral radius r(A) > 1. Then the successive
approximations with CPo = 0 cannot converge for all f EX.
Proof. Assume the statement is false. Convergence of the successive ap-
proximations for all f E X is equivalent to pointwise convergence of the
Neumann series for A = 1. Hence, by the uniform boundedness principle
Theorem lOA, there exists a positive constant C such that
for all n E IN. Then, by the triangle inequality, IIAnll ~ 20 for all n E IN.
Since r(A) > 1, there exists Ao E a(A) with IAol > 1. For the partial sums
o
8n := I:~=I A k - I Ak of the Neumann series we have
11 8 - 8 II < 20IAol- n- 1
m n - IAol-1
for all m > n. Therefore (8n ) is a Cauchy sequence in the Banach space
L(X, X) (see Theorem 2.4). This implies convergence of the Neumann se-
ries for Ao, which contradicts Ao E a(A). 0
If A is compact then the successive approximations, in general, cannot

converge if r(A) = 1. By Theorem 3.9 all spectral values of a compact
°
operator different from zero are eigenvalues. Hence, for r(A) = 1, there
°
exists f E X with f ;f; and Ao with IAol = 1 such that Af = Aof. But
then the successive approximations with CPo = satisfy
n-I
CPn = 2: A~f, n = 1,2, ... ,
k=O
and therefore they diverge.

Usually, in practical calculations, some approximation of the operator A
will be used. Therefore we have to establish that our results on convergence
remain stable under small perturbations. This is done through the following
two theorems.
Theorem 10.15 Let A: X -+ X be a bounded linear operator in a Banach
space X with spectral radius r(A) < 1 and let the sequence Am : X -+ X
of bounded linear operators be norm convergent IIAm - All -+ 0, m -+ 00.
Then for all sufficiently large m the equation cp - AmCP = f can be solved
by successive approximations.
Proof. Choose Ao E (r(A), 1). Then R(A; A) = (AI - A)-I exists and
is bounded for all A E <C with IAI ~ Ao. Since R(A; A) is analytic and
IIR(A;A)II-+ 0, IAI-+ 00 (this follows from the Neumann series expansion
of the resolvent), we have
C:= sup II (AI - A)-III < 00.

1>'1~>'o
Therefore, because IIAm - All -+ 0, m -+ 00, there exists an integer N

such that II (AI - A)-I(Am - A)II < 1 for all m ~ Nand alllAI ~ Ao. But
then from Theorem 10.1, applied to AI - A and AI - Am, we deduce that
(AI - Am)-I exists and is bounded for all m ~ N and all IAI ~ Ao. Hence
r(Am) < 1 for all m ~ N and the statement follows from Theorem 10.13. 0
Similarly, based on Theorem 10.9, we can prove the following theorem.

Theorem 10.16 Let A : X -t X be a compact linear operator in a Banach

space X with spectral radius rCA) < 1 and let the sequence Am: X -t X of
collectively compact linear operators be pointwise convergent Am'P -t A'P,
m -t 00, for all 'P EX. Then for all sufficiently large m the equation
'P - Am'P = f can be solved by successive approximations.
We conclude this section with two examples for the application of The-
orem 10.13.
Theorem 10.17 Volterra integral equations of the second kind with

continuous or weakly singular kernels can be solved by successive
approximations.
Proof. Volterra integral operators have spectral radius zero. For continu-
ous kernels this is a consequence of Theorems 3.9 and 3.10. The case of
a weakly singular kernel can be dealt with by reducing it to a continuous
kernel through iteration as in the proof of Theorem 9.9. 0
Historically, the first existence proofs for Volterra integral equations of

the second kind actually were obtained by directly establishing the conver-
gence of successive approximations.
The following classical potential theoretic result goes back to Plemelj [144].
Theorem 10.18 The integral operators K, K' : C(aD) -t C(aD), defined

by (6.31) and (6.32), have spectrum
-1 E a(K) = a(K') c [-1,1).

Proof. By Theorem 3.9, the spectrum a(K') \ {O} of the compact operator
K' consists only of eigenvalues. Let >. =J -1 be an eigenvalue of K' with
eigenfunction'P and define the single-layer potential u with density 'P. Then,
from the jump relations of Theorems 6.14 and 6.18, we obtain u+ = u_
and
au± 1, 1 1
av = 2K'P =f 2 'P = 2(>' =f 1)'P
on aD. From this we see that
In addition, by the Gauss theorem Corollary 6.4, we deduce faD 'P ds = O.

Therefore, in view of (6.14) the single-layer potential u has the asymptotic
behavior
u(x) = 0 CXI!-l)' gradu(x) =0 Cx~m)' Ixl-t 00,

Problems 175
uniformly for all directions. Now we can apply Green's Theorem 6.3 to
obtain
(10.10)
Assume that fnm\D Igrad ul 2dx = fD Igrad ul 2dx = O. Then grad u = 0 in

R m and from the jump relations we have cP = 0, which is a contradiction to
the fact that cp is an eigenfunction. Hence, from (10.10) we conclude that A
is contained in [-1,1]. From Theorem 6.20 we already know that -1 is an
eigenvalue of K' whereas 1 is not an eigenvalue. Finally, using the fact that
a(K') is real, the equality a(K) = a(K') is a consequence of the Fredholm
alternative for the adjoint operators K and K'. 0
Corollary 10.19 The successive approximations

1 1
CPn+1 := 2" CPn + 2" KCPn - f, n = 0, 1,2, ... ,
converge uniformly for all CPo and all f to the unique solution cP of the inte-
gral equation cP - Kcp = -2f from Theorem 6.21 for the interior Dirichlet
problem.
Proof. We apply Theorem 10.13 to the equation ..p - Acp = - f with the
operator A := 2- 1 (1 + K). Then we have a(A) = {1/2 + A/2 : A E a(K)}
and therefore rCA) < 1 by Theorem 10.18. 0
We mentioned in Problem 6.5 that Neumann used the successive approx-

imation scheme of Corollary 10.19 to give the first rigorous existence proof
for the Dirichlet problem in two-dimensional convex domains.
Problems
10.1 Prove the Banach-Steinhaus theorem: Let A : X -t Y be a bounded linear
operator and let An : X -t Y be a sequence of bounded linear operators from a
Banach space X into a normed space Y. For pointwise convergence An<,O -t A<,O,
n -t 00, for all <,0 E X it is necessary and sufficient that IIAnll ::; C for all n E 1N
and some constant C and that An<,O -t A<,O, n -t 00, for all <,0 E U, where U is
some dense subset of X.
10.2 Show that a sequence An : X -t Y of compact linear operators mapping

a normed space X into a normed space Y is collectively compact if and only if
for each bounded sequence (<,On) in X the sequence (An<,On) is relatively compact
in Y.
10.3 Let X and Y be Banach spaces, 8 : X -t Y be a bounded linear operator
with bounded inverse 8- 1 : Y -t X, and A : X -t Y be compact. Formulate
and prove extensions of Theorems 10.9 and 10.10 for the approximation of the
equation
Sr.p-Ar.p= i
by equations
Snr.pn - Anr.pn = in,
where the An : X -t Yare collectively compact with pointwise convergence
An -t A, n -t 00, the Sn : X -t Y are bounded with pointwise convergence
Sn -t S, n -t 00, and with uniformly bounded inverses S:;;l : Y -t X, and
in -t i, n -t 00.
10.4 Solve the Volterra integral equation
by successive approximations.
10.5 How can the integral equation from Theorem 6.25 for the interior Neumann
problem be solved by successive approximations?
11
Degenerate Kernel Approximation
In this chapter we will consider the approximate solution of integral equa-

tions of the second kind by replacing the kernels by degenerate kernels,
i.e., by approximating a given kernel K(x, y) through a sum of a finite
number of products of functions of x alone by functions of y alone. In par-
ticular, we will describe the construction of appropriate degenerate kernels
by interpolation of the given kernel and by orthonormal expansions. The
corresponding error analysis will be settled by our results in Section 10.1.
We also include a discussion of some basic facts on piecewise linear inter-
polation and trigonometric interpolation, which will be used in this and
subsequent chapters.
The degenerate kernel approximation is the simplest method for numer-
ically solving integral equations of the second kind as far as its under-
standing and error analysis are concerned. However, because the actual
implementation requires the numerical evaluation of either a single or dou-
ble integral for each matrix coefficient and for each right-hand side of the
approximating linear system, in general, it cannot compete in efficiency
with other methods, for example, with the Nystrom method of Chapter 12
or the projection methods of Chapter 13.
For further studies on degenerate kernel methods, we refer to Atkin-
son [9], Baker [12], Fenyo and Stolle [39], Hackbusch [60], and Kantorovic
and Krylov [81].

178 11. Degenerate Kernel Approximation
11.1 Degenerate Operators and Kernels

Let (X, X) be a dual system, generated by a bounded nondegenerate bi-
linear form, and let An : X -+ X be a bounded linear operator of the
form
n
An<P = L(<p,bj}aj, (11.1)
j=l
where aI, ... , an and b1 , ••. , bn are elements of X such that the aI, ... , an
are linearly independent. For the sake of notational clarity we restrict our-
selves to the case where the index n for the operator An coincides with the
number of elements aI, ... , an and b1 , •.. , bn . The solution of the equation
of the second kind <Pn - An<Pn = f with such a degenerate operator reduces
to solving a finite-dimensional linear system.
Theorem 11.1 Each solution of the equation
n
<Pn - L(<pn,bj}aj = f (11.2)
j=l
has the form
n
<Pn =f +L "Ikak, (11.3)
k=l
where the coefficients "11, .•. , "In satisfy the linear system
n
"Ij - L(ak,bjhk = (f,bj ), j = 1, ... ,n. (11.4)
k=l
Conversely, to each solution "11, . .• , "In of the linear system {11.4} there
corresponds a solution <Pn of {11.2} defined by {11.3}.
Proof. Let <Pn be a solution of (11.2). Writing "Ik := (<Pn, bk) for k = 1, ... , n,
and solving for <Pn we obtain the form (11.3). Taking the bilinear form of
(11.3) with bj we find that the coefficients "11, ... ,"In must satisfy the linear
system (11.4). Conversely, let "11, •. . , "In be a solution to the linear system
(11.4) and define <Pn by (11.3). Then

11.2 Interpolation 179
For a given operator A we will use finite-dimensional approximations

An of the form (11.1) to obtain approximate solutions of equations of the
second kind by using Theorems 10.1 and 10.2 for the operators I - A
and I - An. We wish to mention that this approach via finite-dimensional
approximations can also be used to establish the Fredholm alternative for
the case of linear operators in Banach spaces, which can be split into the
sum of an operator of the form (11.1) and an operator with norm less than
one (see Problems 11.1 and 11.2). For integral operators with continuous or
weakly singular kernels such an approximation can be achieved by Problem
2.3 and by using an approximation of weakly singular kernels by continuous
kernels, as in the proof of Theorem 2.22.
In the case of the dual system (C(G), C(G)) introduced in Theorem 4.3
the finite-dimensional operator An has the form
(An<P)(X) = fc Kn(x,y)<p(y)dy, x E G,
of an integral operator with a degenerate kernel

n
Kn(x,y) = L:aj(x)bj(y).
j=l
The solution of the integral equation of the second kind
(11.5)
with such a degenerate kernel is described through Theorem 11.1. We will

use integral operators An with degenerate kernels Kn as approximations
for integral operators A with continuous kernels K in the sense of Section
10.1. In view of Theorems 2.8, 10.1, and 10.2 we need approximations with
the property that
IIAn - Alloo = max

xEG lGr IKn(x, y) - K(x, y)1 dy
becomes small.
Here, we do not attempt to give a complete account of the various pos-
sibilities for degenerate kernel approximations. Instead, we will introduce
the reader to the basic ideas by considering a few simple examples.
11.2 Interpolation
One important method to construct degenerate kernels approximating a
given continuous kernel is by interpolation. For its investigation, and for
use in subsequent chapters, we collect some basic facts on interpolation.
Theorem 11.2 Let Un C C(G) be an n-dimensional subspace and let

Xl, ... ,Xn be n points in G such that Un is unisolvent with respect to
X1, ... ,Xn, i.e., each lunction from Un that vanishes in Xl, ... ,Xn van-
ishes identically. Then, given n values gl, ... , gn, there exists a uniquely
determined function u E Un with the interpolation property
u(Xj) = gj, j = 1, ... ,no
With the interpolation data given by the values 9j = 9 (x j ), j = 1, ... , n, 01

a function 9 E C(G) the mapping 9 I-t U defines a bounded linear operator
P n : C(G) -t Un which is called interpolation operator.
Proof. Let Un = span{U1, ... , un}. Then the solution to the interpolation
problem is given by
z=
n
U= 'YkUk,
k=1
where the coefficients 'Y1, . .. ,'Yn are determined by the uniquely solvable
linear system
n
E 'YkUk(Xj) = gj, j = 1, ... ,no
k=1
Let L 1, ... , Ln denote the Lagrange basis for Un, i.e., we have the interpo-
lation property
Lk(Xj) = bjk' j, k = 1, ... , n,
where bjk = 1 for k = j, and bjk = 0 for k f=. j. Then, from
n
Png = E g(xk)Lk (11.6)
k=1
we conclude the linearity and boundedness of Pn . o
From (11.6) we have that
Now choose Z E G such that
and a function 1 E C(G) with 11/1100 = 1 and

n n
E l(xk)Lk(z) = E ILk(z)l·
k=1 k=l
11.2 Interpolation 181
Then
n
\\Pnll oo ~ \\Pnf\\oo ~ \(Pnf)(z) \ = macx

xE
L \Lk(x)\,
k=l
and therefore, in terms of the Lagrange basis, we have the norm
(11. 7)
For examples we confine ourselves to interpolation on a finite interval

[a, b] c 1R with a < b. First we consider linear splines, i.e., continuous
piecewise linear functions. Let Xj = a + j h, j = 0, ... , n, denote an equi-
distant subdivision with step size h = (b - a)/n and let Un be the space of
continuous functions on [a, b] whose restrictions on each of the subintervals
[Xj-l, Xj], j = 1, ... , n, are linear. Existence and uniqueness for the corre-
sponding linear spline interpolation are evident. Here, the Lagrange basis
is given by the so-called hat functions
1
h (x - Xj-l), x E [Xj-l,Xj], j ~ 1,
Lj(x) = 1 (11.8)
h (XHl - x), x E [Xj, XH1], j:::; n -1,
0, otherwise.
Because the L j are nonnegative, we have I:.'j=o \Lj(x)\ = I:.'j=o Lj(x) = 1

for all x E [a,b]. Therefore, (11.7) implies that
(11.9)
for piecewise linear interpolation. For the interpolation of twice continu-

ously differentiable functions 9 we have the following error estimate.
Theorem 11.3 Let 9 E C 2[a, b]. Then, for the error in piecewise linear
interpolation we have the estimate
(11.10)
Proof. Obviously, the maximum of jPng - g\ in [Xj, Xj+1] is attained at an

interior point ~ where g'(~) = (Png)/(~) = {g(xHd - g(xj)}/h. Without
loss of generality we may assume that ~ - Xj :::; h/2. Then, using Taylor's
formula, we derive
(Png)(~) - g(~) = g(Xj) + (Png)'(O(~ - Xj) - g(~)
= g(Xj) - g(~) - (Xj - ~)g'(~) = ~2 (Xj - ~)2g"(7J),

with some 'f/ E (X j, e). Hence,
and the estimate (11.10) follows. o
11.3 Trigonometric Interpolation

As we have seen in Chapters 6-8 the parameterization of boundary integral
equations for two-dimensional boundary value problems leads to integral
equations for periodic functions. For these, instead of a local approxima-
tion by piecewise linear interpolation (or higher-order spline interpolation),
quite often it will be preferable to use a global approximation in the form of
trigonometric interpolation, since this leads to a more rapid convergence.
Let tj = j7r In, j = 0, ... ,2n - 1, be an equidistant subdivision of the in-
terval [0,27r] with an even number of grid points. Then, given the values
90, ... ,92n-b there exists a unique trigonometric polynomial of the form
n-l
u(t) = ~o + L[akcoskt + !Asinkt]+ ~n cosnt (11.11)
k=l
with the interpolation property u(tj) = 9j, j = 0, ... , 2n -1. Its coefficients
are given by
1 2n-l
ak = - L 9j cos ktj, k = 0, ... , n,
n j=O
1 2n-l
(3k = - L 9j sin ktj, k = 1, ... , n - 1.
n j=O
From this we deduce that the Lagrange basis for the trigonometric inter-
polation has the form
1 {
1+2L cosk(t - tj) + cosn(t - tj)
n-l }
Lj(t) = 2n (11.12)
k=l
for t E [0,27r] and j = 0, ... , 2n - 1. Using the real part of the geometric
sum (8.26), we can transform (11.12) into
Lj(t)
1
= 2n sin n(t - tj) cot T'
t - t·
t =I tj. (11.13)
11.3 Trigonometric Interpolation 183
Theorem 11.4 For n ~ 2, the trigonometric interpolation operator with

2n equidistant interpolation points has norm
2
IlFnl/co < 3 + -7r In(2n). (11.14)
Proof. The function A : [0, 7r] ~ ill. defined by

2n-l
A(t) := L ILj{t)1
j=O
is even and has period 7r/n. Therefore, in view of (11.7), we have
IIPni/ co = max A{t).

09::51r/2n
For 0 :::; t : :; 7r /2n, using (11.12) and (11.13), we find

1 2n-2
A{t) :::; 3 + 2n ~ 1
ttl
cot ~ j .
3=2
Since the cotangent function s 1--+ cot(s/2) is positive and monotonically

decreasing on the interval (0, 7r], for 0 :::; t :::; 7r /2n in view of h - t ~ 7r /2n
we can estimate
-2n1 Lnit
j=2
- t ·1 < -1
cot-_
2
3
27r
itn
t1
S - t
cot--
2
1
ds < --lnsin-.
7r
7r
4n
Analogously we have
L 2 cot _
-2n1 .2n- 1
t - t .,
1
2
_ <-
1 [t2n-1
27r
3
+1
t- s 1 7r
cot - - ds < - - In sin - .
2 t 7r 4n
3=n+ n
Now, using the inequality

. 7r 1
sm->-
4n 2n
the statement of the lemma follows by piecing the above inequalities
together. 0
The estimate (11.14) gives the correct behavior for large n, because pro-
ceeding as in the proof of Theorem 11.4 we can show that
IlFnl/co ~ A(.!!.-.) > ~ cot'!!'-' - ~ In sin .!!.-. (11.15)

2n 2n 4n 7r 4n
for all n ~ 2. This estimate implies that IlFnllco ~ 00 as n ~ 00. Therefore,
from the uniform boundedness principle Theorem 10.4 it follows that there
exists a continuous 2IT-periodic function 9 for which the sequence (Png)

does not converge to 9 in 0[0,2IT). However, by the Banach-Steinhaus
theorem (see Problem 10.1) and the denseness of the trigonometric poly-
nomials in 0[0, 2IT), the following lemma implies mean square convergence
of the trigonometric interpolation polynomials for continuous functions.
Lemma 11.5 The trigonometric interpolation operator satisfies
(11.16)
for all n E IN and all continuous 2IT-periodic functions g.
Proof. Using (11.12), elementary integrations yield
(11.17)
for j, k = 0, ... , 2n - 1. Then, by the Cauchy-Schwarz inequality we can

estimate
and (11.16) is proven. o
Provided the interpolated function g has additional regularity, we also

have uniform convergence of the trigonometric interpolation polynomi-
als as expressed by the following three theorems. For 0 < a ~ 1, by
og;,<> c OO'<>(lR) we denote the space of 2IT-periodic Holder continuous
functions. Furthermore, for m E IN by 0;;;;<> C om'<>(lR) we denote the
space of 2IT-periodic and m-times Holder continuously differentiable func-
tions 9 equipped with the norm
Ilgllm,<> = Ilglloo + Ilg(m) 110,<>.

The following theorem provides a uniform error estimate and implies uni-
form convergence for the trigonometric interpolation of Holder continuous
functions.
Theorem 11.6 Let m E IN U {O} and 0 < a ~ 1. Then for the trigono-
metric interpolation we have
Inn
IIPng - glloo ~ 0
n
m+<> Ilgllm,<> (11.18)
for all 9 E o;;;t and some constant 0 depending on m and a.

11.3 Trigonometric Interpolation 185
Proof. For 9 E C;;;O< let Pn be a best approximation of 9 with respect to

trigonometric polynomials of the form (11.11) and the maximum norm (see
Theorem 1.24). Then, by Jackson's theorem (see [31, 120)) there exists a
constant c that does not depend on g, such that
c
IIPn - glloo :::; nm+o< IIgllm,o<' (11.19)
Now, writing
Png - 9 = Pn(g - Pn) - (g - Pn),
the estimate (11.18) follows from (11.14) and (11.19). o
Theorem 11.6 can be extended to
IIPng - gll£,fj :::; C nm~~:O<-fj IIgllm,o< (11.20)
for 9 E C;,:r'o<, m, £ E JNU{O} with £ :::; m, 0 < f3 :::; a :::; 1 and some constant
C depending on m, £, a, and f3 (see Prossdorf and Silbermann [149, 150)).
For the trigonometric interpolation of 211"-periodic analytic functions we
have the following error estimate (see [96)).
Theorem 11.1 Let 9 : IR -+ IR be analytic and 211" -periodic. Then there
exists a strip D = IR x (-s, s) c <C with s > 0 such that 9 can be extended
to a holomorphic and 211"-periodic bounded function 9 : D -+ <C. The error
for the trigonometric interpolation can be estimated by
s
coth"2
IlPng - glloo:::; M -.-h-' (11.21)
sm ns
where M denotes a bound for the holomorphic function 9 on D.
Proof. Because 9 : IR -+ IR is analytic, at each point t E IR the Taylor
expansion provides a holomorphic extension of 9 into some open disk in the
complex plane with radius r(t) > 0 and center t. The extended function
again has period 211", since the coefficients of the Taylor series at t and
at t + 211" coincide for the 211"-periodic function 9 : IR -+ IR. The disks
corresponding to all points of the interval [0,211"] provide an open covering
of [0,211"]. Because [0,211"] is compact, a finite number of these disks suffices
to cover [0,211"]. Then we have an extension into a strip D with finite width
28 contained in the union of the finite number of disks. Without loss of
generality we may assume that 9 is bounded on D.
Let 0 < (J" < s be arbitrary. By r we denote the boundary of the rect-
angle [-11" /2n, 211" - 11" /2n] x [-(J", (J"] with counterclockwise orientation. A
straightforward application of the residue theorem yields
7 -t
1 [cot - -
-. . 2 g(7) d7 =
2n r smn7
2 (t)
smnt
L (-1)j g(tj) cot _
-.f!-- - -1 2n-l
n j=O
t - t·
_1
2
for -7r /2n ~ t < 27r - 7r /2n, t #- tk, k = 0, ... ,2n - 1. Hence, in view of
(11.13) we obtain
sin nt
g(t) - (Png)(t) = -4-'
7r2
1 r - t
cot - -
. 2 g(r) dr,
r sm nr
where we can obviously drop the restriction that t does not coincide with an
interpolation point. From the periodicity of the integrand and since, by the
Schwarz reflection principle, 9 enjoys the symmetry property g(;;;=) = g( r ),
we find the representation
1 ia+27r i cot -
r-- t }
g(t)-(Png)(t)=-sinntRe { [ . 2 g(r)dr (11.22)
27r Jia sm nr
°
with < a < s. Because of \ sin nr\ 2: sinh nO' and \ cot r /2\ ~ coth 0'/2 for
1m r = a, the bound (11.21) now follows by estimating in (11.22) and then
passing to the limit a -+ s. 0
We can summarize Theorem 11.7 by the estimate
(11.23)
for the trigonometric interpolation of periodic analytic functions, where C

and s are some positive constants depending on g, i.e., the interpolation
error decays at least exponentially.
We conclude this section on trigonometric interpolation with an error
estimate in the Sobolev spaces HP[O, 27r]. Note that by Theorem 8.4 each
function 9 E HP[O, 27r] is continuous if p > 1/2.
Theorem 11.8 For the trigonometric interpolation we have
° ~ q ~p,
1
"2 < p, (11.24)
for all g E HP[O, 27r] and some constant C depending on p and q.

Proof. Consider the monomials fm(t) = eimt and write m = 2kn+J.L, where
k is an integer and -n < J.L ~ n. Since fm(tj) = fJ-!(tj) for j = 0, ... , 2n-1,
we have
J.L =1= n,
J.L=n.
Therefore the trigonometric interpolation polynomial is given by
in terms of the Fourier coefficients ak of 9 E HP[O, 2rr]. Hence,
Ilg - Pngll~ = S1 + S2 + S3,

where
2
n-1
S1:= L (1+m 2)qlam I2, S2:= L (1+tt 2)q L a2kn+JL ,
Iml~n+1 JL=-n+1 k#O
Since p 2: q, we can estimate
S1 :S (1 + n 2 )q-p L (1 + m2)Plam l2:S c1n2Q-2PllglI;

Iml~n+1
with some constant Cl depending on p and q. For the second term, with
the aid of the Cauchy-Schwarz inequality we find
(1 + tt 2 )Q "
S2 :S C2
n-1
L n 2p
2
L.,.,12kn + ttl Pl a2kn+JLI
2
JL=-n+l k#O
with some constant C2 depending on p and q. Similarly we can estimate
with some constant C3 depending on p and q. Since q 2: 0, we have
(1 :~2)Q :S 22Qn2Q-2p, Ittl::; n,
and thus we obtain

n
L L
00
S2 + S3 :S c4n2Q-2p 12kn + ttl2Pla2kn+JLl2 ::; c5n2Q-2PllglI;

JL=-n+1 k=-oo
with some constants C4 and C5 depending on p and q. Summarizing the two
estimates on S1 and S2 + S3 concludes the proof. 0
11.4 Degenerate Kernels via Interpolation

Now we return to integral equations of the second kind of the form
<p(x) -lb K(x, y)<p(y) dy = I(x), a:S x :S b.

Recalling the interpolation Theorem 11.2, we approximate a given contin-

uous kernel K by the kernel Kn interpolating K with respect to x, i.e.,
K n (·, Y) E Un and
for each Y in [a, b]. Then we can write

n
Kn(x,y) = LLj(x)K(xj,Y).
j=1
Hence, Kn is a degenerate kernel with aj = Lj and bj = K(xj, -). In this

case the linear system (11.4) reads
(11.25)
for j = 1, ... ,n, and the solution of the integral equation (11.5) is given by
n
<{In = f +L 'YkLk·
k=1
We illustrate the method by two examples, where we use piecewise lin-

ear interpolation and trigonometric interpolation as discussed in the two
previous sections.
From Theorems 2.8 and 11.3, for the approximation of an integral oper-
ator A with a twice continuously differentiable kernel K by a degenerate
kernel operator An via linear spline interpolation we conclude the estimate
(11.26)
Therefore, by Theorem 10.1, the corresponding approximation of the so-

lution to the integral equation is of order O(h2). In principle, using The-
orem 10.2, it is possible to derive computable error bounds based on the
estimate (11.26). In general, these bounds will be difficult to evaluate in
applications. However, in most practical problems it will be sufficient to
judge the accuracy of the computed solution by refining the subdivision
and then comparing the results for the fine and the coarse grid with the
aid of the convergence order, i.e., in the case of linear splines with the aid
of the approximation order O(h2).
Only in special cases the integrals for the coefficients and the right-hand
side of the linear system (11.25) can be evaluated in closed form. Therefore
the degenerate kernel method, in principle, is only a semi-discrete method.
For a fully discrete method we have to incorporate a numerical evaluation
of the integrals occurring in (11.25). As a general rule, these numerical

quadratures should be performed so that the approximation order for the
solution to the integral equation is maintained.
To be consistent with our approximations, we replace K(xj,.) by its
linear spline interpolation, i.e., we approximate
lb a
K(xj, y)Lk(Y) dy ~ L
m=O
n
K(xj, xm)
lb Lm(y)Lk(Y) dy,
a
(11.27)
for j, k = 0, ... , n. Straightforward calculations yield the tridiagonal matrix

2 1
1 4 1
1 4 1
W=~
6
1 4 1
1 2
for the weights Wmk = J:

Lm(y)Lk(Y) dy. For the right-hand side of (11.25)
we simultaneously replace f and K(xj, .) by its spline interpolations. This
leads to the approximations
1 a
b
K(Xj, y)f(y) dy ~ L
n
k,m=O
K(xj, xm)f(Xk) 1
a
b
Lm(y)Lk(Y) dy (11.28)
for j = 0, ... ,n. We now investigate the influence of these approximations

on the error analysis. For this we interpret the solution ofthe system (11.25)
with the approximate values for the coefficients and the right-hand sides
as the solution 'Pn of a further approximate equation 'Pn - An<pn = In,
namely, of the degenerate kernel equation
with
n n
Kn(x, y) := L K(xj, xm)Lj(x)Lm(Y) and In(x):= L f(xm)Lm(x).
j,m=O m=O
Provided that the kernel K is twice continuously differentiable, writing

K(x,y) - Kn(x,y) = K(x,y) - [PnK(· ,y)](x)
(11.29)
+[Pn[K(· ,y) - [PnK(., ·)](y)](x),
by Theorem 11.3 and (11.9), we can estimate
for all a ~ x, y ~ b. Hence, for the integral operator An with kernel Kn we

have IIAn - Alloo = O(h2). When f is twice continuously differentiable, we
also have IIin - flloo = O(h2). Now, Theorem 10.1 yields the error estimate
IIrpn - ipll = O(h2). Therefore we have the following theorem.
Theorem 11.9 The degenerate kernel approximation via equidistant piece-
wise linear interpolation with the matrix entries and the right-hand sides of
the approximating linear system (11.25) replaced by the quadratures (11. 27)
and (11.28) approximates the solution to the integral equation with order
O(h2) if the kernel K and the right-hand side f are twice continuously
differentiable.
Example 11.10 Consider the integral equation
ip(x) -
r
1
2"1 Jo (x+ l)e- xy ip(Y) dy = e- x - 2"1 + 2"e-(x+l)
1
, o~ x ~ 1. (11.30)
Obviously it has the solution ip(x) = e- X • For its kernel we have
max
O:S;x:S;1
1
0
1
IK(x, y)ldy =
x+ 1
sup -2- (1- e- X )
O<x:S;1 X
< 1.
Therefore, by Corollary 2.11, the solution is unique.

Table 11.1 shows the error between the approximate and the exact so-
lution at the points x = 0, 0.25, 0.5, 0.75, and 1 for various n. It clearly
exhibits the behavior O(h2) as predicted by Theorem 11.9. 0
TABLE 11.1. Numerical results for Example 11.10
n x=o x = 0.25 x = 0.5 x = 0.75 x=l
4 0.004808 0.005430 0.006178 0.007128 0.008331

8 0.001199 0.001354 0.001541 0.001778 0.002078
16 0.000300 0.000338 0.000385 0.000444 0.000519
32 0.000075 0.000085 0.000096 0.000111 0.000130
From Theorem 11.6, for the approximation of an integral operator A

with a 21f-periodic and m-times continuously differentiable kernel K by
degenerate kernels via trigonometric interpolation we conclude the estimate
(11.31)
for all c > O. If the kernel is analytic, then from Theorem 11.7 we have
(11.32)
for some 8 > O. For the derivation of (11.32) we have to use the fact that
for the kernel function, which is assumed to be analytic and periodic with
respect to both variables, there exists a strip D with width 28 such that
K (. , r) can be continued holomorphicly and uniformly bounded into D for
all r E [0,211"]. We leave it to the reader to go over the argument given in
the proof of Theorem 11.7 to verify this property. Now, by Theorem 10.1,
the corresponding approximation of the solution to the integral equation is
also of order 0 (n e - m ) or O{e- ns ), respectively.
Again we have to describe the numerical evaluation of the coefficients and
the right-hand side of the linear system (11.25). We proceed analogously
to the preceding example and approximate the matrix entries by replacing
K(tj,·) by its trigonometric interpolation polynomial, i.e.,
(11.33)
for j, k = 0, ... , 2n -
1 with the integrals on the right-hand side given by
(11.17). For the right-hand side of (11.25) we simultaneously replaceJ and
K(tj,·) by its trigonometric interpolation. This leads to
for j = 0, ... , 2n -1. Note that despite the global nature of the trigonomet-
ric interpolation and its Lagrange basis, due to the simple structure of the
weights (11.17) in the quadrature rule, the approximate computation of the
matrix elements and right-hand side is not too costly. In view of (11.18)
and splitting analogous to (11.29), we remain with a total error 0 (n e - m )
or O(e- ns ), respectively, for the approximate solution of the integral equa-
tion. Here we use that In n e- ns = O(e- nu ) for all 0 < 0' < 8. Hence we
have the following theorem.
Theorem 11.11 The degenerate kernel approximation via trigonometric
interpolation with the matrix entries and the right-hand sides of the ap-
proximating linear system (11.25) replaced by the quadratures (11.33) and
(11.34) approximates the solution to the integral equation oj order 0 (n e - m )
or O(e- ns ), respectively, iJ the 211"-periodic kernel K and the 211"-periodic
right-hand side J are m-times continuously differentiable or analytic,
respectively.
ab r 21r <per) dr
o ~ t ~ 211",
<pet) + -; 10 a2 + b2 _ (a 2 _ b2 ) cos(t + r) = J(t), (11.35)
corresponding to the Dirichlet problem for the Laplace equation in the

interior of an ellipse with semi-axis a ~ b > 0 (see Problem 6.2). We
numerically want to solve the case where the unique solution is given by
L
00
<p(t) = ecostcos(sint) = Re ~ eimt , 0::::; t::::; 27r.

m.
m=O
Then, the right-hand side becomes

f(t) = <p(t) + eccost cos(csin t), 0::::; t ::::; 27r,
where c = (a - b)/(a + b) (compare also Problem 11.4).
Table 11.2 shows the error between the approximate and the exact so-
lution at the three points t = 0, 7r /2, and 7r for various n and depending
on the ratio b/a. It clearly shows the exponential decay of the error: dou-
bling the number of grid points doubles the number of correct digits in the
approximate solution. The rate of the exponential decay depends on the
parameters a and b, which describe the location of the singularities of the
integrands in the complex plane, i.e., they determine the value for the strip
parameter s. 0
2n t=O t = 7r /2 t = 7r
4 -0.10752855 -0.03243176 0.03961310
a=1 8 -0.00231537 0.00059809 0.00045961
b=0.5 16 -0.00000044 0.00000002 -0.00000000
32 0.00000000 0.00000000 0.00000000
4 -0.56984945 -0.18357135 0.06022598

a=1 8 -0.14414257 -0.00368787 -0.00571394
b=0.2 16 -0.00602543 -0.00035953 -0.00045408
32 -0.00000919 -0.00000055 -0.00000069
11.5 Degenerate Kernels via Expansions

A second possibility for constructing approximate degenerate kernels is by
expansions, in particular, by orthonormal expansions. Let (".) denote a
scalar product on C( G) and let {Ul' U2, ... } be a complete orthonormal sys-
tem (recall Theorem 1.28). Then a given continuous kernel K is expanded
with respect to x for each fixed y E G, i.e., K(x, y) is approximated by the
partial sum
n
Kn(x,y):= LUj(x)(K(.,y),Uj)
j=l
11.5 Degenerate Kernels via Expansions 193
of the Fourier series. For this degenerate kernel the linear system (11.4)
reads
'"fj - t Jar
k=l
'Yk uk(y)(K(·, y), Uj) dy = r f(y)(K(·, y),
Ja
Uj) dy (11.36)
for j = 1, ... , n. Usually, the scalar product will be given in terms of an

integral. Therefore the system (11.36) requires a double integration for
each coefficient and for each right-hand side. Since it will turn out that the
degenerate kernel method via orthonormal expansion is closely related to
the Galerkin method (see Section 13.5), we omit further discussion and a
description of the numerical implementation.
For the orthonormal expansion degenerate kernel approximations, in
principle, we can derive error estimates from Theorems 10.1 and 10.2 only
with respect to the scalar product norm. In the subsequent analysis, for
later theoretical use, we will describe how estimates in the maximum norm
can be obtained under additional regularity assumptions on the kernel.
The Chebyshev polynomials of the first kind are defined by
Tn(z) := cos(narccosz), -1:::; z :::; 1, n = 0,1, .... (11.37)
From To (z) = 1 and T1 (z) = z and from the recursion formula
Tn+1(z) + Tn- 1(z) = 2zTn(z),
which follows from the cosine addition theorem, we observe that the Tn
indeed are polynomials of degree n and therefore, in particular, well defined
on the whole complex plane. A substitution x = cos t readily shows that
r1 Tn(x)Tm(x)
J-1 vI _
d
x2
-
x -
~ [8
2 nm +
8]
nO,
i.e., the Tn form an orthogonal system with respect to the scalar product
(<p, 'ljJ):= r <p(x)~ dx

1
1-1 Vl- x 2
(11.38)
on C[-I, 1]. By the same substitution, the denseness of the trigonomet-

ric polynomials in C[0,2rr] implies the denseness of span{To, T}, ... } in
C[-1, 1] with respect to the scalar product (11.38), i.e., the Tn form a
complete orthogonal system.
Theorem 11.13 Let 9 : [-1,1] -t lR be analytic. Then there exists an
ellipse E with foci at -1 and 1 such that 9 can be extended to a holomorphic
and bounded function 9 : D -t <C, where D denotes the open interior of E.
The orthonormal expansion with respect to the Chebyshev polynomials
(11.39)
/1
with coefficients
an = ~ g(x)Tn(x) dx
71' -1 -/1- x 2
is uniformly convergent with the estimate
~ 2M
II 9 - 2' To -
aD -n
L..t amTm ~ R _ 1R .
m=1 00
Here, R is given through the semi-axis a and b of E by R = a + band M

is a bound on 9 in D.
Proof The existence of the holomorphic extension of the analytic function

9 : [-1,1] --+ 1R into the open interior D of some ellipse with foci at -1
and 1 is shown analogously to the proof of Theorem 11.7.
The function
w 1
z=-+-2 2w
maps the annulus 1/R < Iwl < R of the w-plane onto the interior D of
the ellipse E of the z-plane. For each holomorphic function g on D the
function g, defined by
g(w):= 2g (~ + 2~)'
is holomorpic in the annulus. Therefore it can be expanded into a Laurent
series
n=-oo
with coefficients
an = - 11
71'i JwJ=r
9 (w-+-
2
1) --,
dw
2w w + n 1
where 1/R < r < R. By substituting ill = l/w and using Cauchy's integral
theorem we find that an = a_ n for all n E :IN. Hence,
In particular, this Laurent expansion converges uniformly on the circle

Iwl = 1. Writing w = eit , we derive
Tn ( -+-
w
2 2w
1) 1(n +-1)
= -
2
w
wn
Problems 195
first on the unit circle, and then, since both sides of the equation represent
rational functions, for all w i- O. Inserting this into the previous series, we
obtain the expansion (11.39) and its uniform convergence on [-1, 1].
Estimating the coefficients of the Laurent expansion and then passing to
the limit r --+ R yields
2M
lanl ::; Rn' n = 0,1,2 ... ,
with a bound M for g on D. Hence, using ITn(x)1 ::; 1 for all -1 ::; x ::; 1
and all n E IN, the remainder can be estimated by
Im~+l amTm(X) I::; m~+1laml::; R-l

00 00 2M
R- n ,
which finishes the proof. o
From Theorem 11.13, analogous to the derivation ofthe estimate (11.32),

for the approximation of an integral operator A with an analytic kernel by
orthonormal expansion with respect to Chebyshev polynomials we find the
estimate
(11.40)
with some constant R > 1 depending on the kernel of A.
In closing this chapter, we wish to mention the use of Taylor expansions
as another method to construct degenerate kernels that may be useful in
special cases (see Problem 11.3).
Problems
11.1 Let (X, Y) be a dual system with two normed spaces X and Y and let
An : X --+ X and Bn : Y --+ Y be adjoint finite-dimensional operators of the form
n n
An<P = L(<p,bj)aj, Bn'IjJ = L(aj,'IjJ)bj
j=l j=l
with linearly independent elements al, ... ,an E X and bl , ... ,bn E Y. By reduc-
tion to linear systems as in Theorem 11.1, establish the Fredholm alternative for
the operators 1 - An and 1 - Bn.
11.2 Let (X, Y) be a dual system with two Banach spaces X and Y, let
An : X --+ X and Bn : Y --+ Y be adjoint finite-dimensional operators as in
Problem 11.1, and let 8 : X --+ X and T : Y --+ Y be adjoint operators with
norm less than one. With the aid of Theorem 2.9 and Problem 11.1 establish the
Fredholm alternative for the operators 1 - A and 1 - B where A = An + 8 and
B=Bn+ T .
Hint: Transform the equations with the operators 1 - A and 1 - B equivalently
into equations with 1 - (1 - 8)-1 An and 1 - Bn(I - T)-l.
11.3 Solve the integral equation (11.30) approximately through degenerate ker-
nels by using the Taylor series for e XY •
11.4 Show that the eigenvalues of the integral equation (11.35) are described
by
ab 1211" ein-r dr (a - b) n -int

n = 0,1,2, ....
--;- 0 (a 2 +b2 )-(a2 -b2 )cos(t+r) = a+b e ,
11.5 Show that, analogous to (11.40), we have IIAn - Alloo = 0 (lifo) for the
approximation of continuously differentiable kernels through degenerate kernels
via Chebyshev polynomials.
Hint: Use Dini's theorem from Problem 8.2.
12
Quadrature Methods
In this chapter we shall describe the quadrature or Nystrom method for the
approximate solution of integral equations of the second kind with contin-
uous or weakly singular kernels. As we have pointed out in Chapter 11, the
implementation of the degenerate kernel method, in general, requires some
use of numerical quadrature. Therefore it is natural to try the application
of numerical integration in a more direct approach to approximate integral
operators by numerical integration operators. This will lead to a straight-
forward but widely applicable method for approximately solving equations
of the second kind. The reason we placed the description of the quadra-
ture method after the degenerate kernel method is only because its error
analysis is more involved.
Through numerical examples we will illustrate that the Nystrom method
provides a highly efficient method for the approximate solution of the
boundary integral equations of the second kind for two-dimensional bound-
ary value problems as considered in Chapters 6-8. We also wish to point
out that, despite the fact that our numerical examples for the application
of the Nystrom method are all for one-dimensional integral equations, this
method can be applied to multi-dimensional integral equations provided ap-
propriate numerical quadratures are available. The latter requirement puts
some limits on the application to the weakly singular integral equations
arising for boundary value problems in more than two space dimensions.

198 12. Quadrature Methods
12.1 Numerical Integration

We start with a brief account of the basics of numerical quadrature. In
general, a quadrature formula is a numerical method for approximating an
integral of the form
Q(g) := fa w(X)g(X) dx,
where w is some weight function. Throughout this chapter G c IRm will

be compact and Jordan measurable. We consider only quadrature rules of
the form
n
Qn(g) := L o:)n) g(X)n))
j=l
with quadrature points xin ), ... ,x~n) contained in G and real quadrature
weights o:in ), ... , o:~n). For the sake of notational clarity we restrict our-
selves to the case where the index n of the quadrature rule coincides with
the number of quadrature points. Occasionally, we also will write Xl, ... , Xn
.
lnstea d 0 f Xl(n) , ... , Xn(n) and 0:1,"" O:n lnstea
. d 0 f 0: 1(n) , ..• , O:n.
(n) The b aSlC
.
numerical integrations are interpolatory quadratures. They are constructed
by replacing the integrand g by an interpolation with respect to the quadra-
ture points Xl, ..• , X n , usually a polynomial, a trigonometric polynomial,
or a spline, and then integrating the interpolating function analytically.
The classical Newton-Cotes rules are a special case of these interpolatory
quadratures with polynomial interpolation on an equidistantly spaced sub-
division of the interval [a, b]. Since the Newton-Cotes rules have unsatisfac-
tory convergence behavior as the degree of the interpolation increases, it is
more practical to use so-called composite rules. These are obtained by sub-
dividing the interval of integration and then applying a fixed rule with low
interpolation order to each subinterval. The most frequently used quadra-
ture rules of this type are the composite trapezoidal rule and the composite
Simpson's rule. For convenience we will discuss their error analysis, which
we base on the representation of the remainder terms by means of Peano
kernels. Let Xj = a + jh, j = 0, ... , n, be an equidistant subdivision with
step size h = (b - a)/n.
Theorem 12.1 Let g E C 2 [a, b]. Then the remainder
for the composite trapezoidal rule can be estimated by

12.1 Numerical Integration 199
Proof. Define the Peano kernel for the trapezoidal rule by

1
KT(x) := 2 (x - xj-d(xj - x), Xj-l ~ X ~ Xj,
for j = 1, ... , n. Then, straightforward partial integrations yield
Now the estimate follows from
and the observation that KT is nonnegative on [a, b]. o

Theorem 12.2 Let 9 E C 4 [a, b] and let n be even. Then the remainder
Rs(g) := lb g(x) dx - ~ [g(xo) + 4g(Xl) + 2g(X2) + 4g(X3) + 2g(X4)

+ ... + 2g(Xn-2) + 4g(Xn-l) + g(xn)]
for the composite Simpson's rule can be estimated by
Proof. With the Peano kernel of Simpson's rule given by
~ (x - X2j_2)3 - 2~ (x - X2j_2)4,
Ks(x) := {
h (X2j - x )3 - 24
18 1 ( 4
X2j - x) ,
for j = 1, ... ,n/2, via
the proof is analogous to the proof of Theorem 12.1. o

Definition 12.3 A sequence (Qn) of quadrature rules is called convergent
if Qn(g) --+ Q(g), n --+ 00, for all 9 E C(G), i.e., if the sequence of linear
functionals (Qn) converges pointwise to the integral Q.
With the aid of the uniform boundedness principle we can state the
following necessary and sufficient conditions for convergence.
Theorem 12.4 (Szego) The quadrature formulas (Qn) converge if and

only if Qn(g) -t Q(g), n -t 00, for all 9 in some dense subset U c C(G)
and
n
sup L laJnJ I <
nEIN j=l
00.
Proof. It is left as an easy exercise to show that

n
IIQnlloo = L lajnJI.
j=l
Then the statement follows from the Banach-Steinhaus theorem (see Prob-
lem 10.1). 0
Corollary 12.5 (Steklov) Assume that Qn(1) -t Q(l), n -t 00, and that
the quadrature weights are all nonnegative. Then the quadrature formulas
(Qn) converge if and only if Qn(g) -t Q(g), n -t 00, for all 9 in some
dense subset U c C(G).
Proof. This follows from
L la)nJ I = L a)nJ = Qn(l) -t Q(l),

n n
n -t 00,
j=l j=l
and the preceding theorem. 0
In particular, from Theorems 12.1 and 12.2 and Corollary 12.5, we ob-
serve that the composite trapezoidal and the composite Simpson's rule are
convergent.
We conclude our remarks on numerical quadrature by describing an error
estimate for the integration of periodic analytic functions due to Davis [30)
(see also Problem 12.2). Note that for periodic functions the composite
trapezoidal rule coincides with the composite rectangular rule.
Theorem 12.6 Let 9 be as in Theorem 11. 7. Then the error
1 {27r 1 2n-l ( .
RT(9) := 21f Jo g(t) dt - 2n ~ 9 J: )
for the composite trapezoidal rule can be estimated by

IRT(g)1 ::; M (coth ns - 1).
Proof. Using the integrals (8.27), we integrate the remainder term (11.22)
for the trigonometric interpolation to obtain
1 { {ilT+27r }
R T (g)=21f Re JilT (l-icotnr)g(r)dr
12.2 Nystrom's Method 201
for all 0 < (1 < s. This can also be shown directly through the residue
theorem. Now the estimate follows from 11 - icotnTI ::; cothna - 1 for
1m T = (1 and passing to the limit a ---+ s. 0
We can summarize Theorem 12.6 by the estimate
IRT(g)1 ::; C e- 2ns (12.1)
for the composite trapezoidal rule for periodic analytic functions, where
C and s are some positive constants depending on g. The improvement
by the factor 2 in the exponent as compared with (11.23) reflects the fact
that the trapezoidal rule integrates trigonometric polynomials not only of
degree less than or equal to n but also of degree less than or equal to 2n - 1
exactly.
For m-times continuously differentiable 2n-periodic functions 9 the Euler-
MacLaurin expansion yields the estimate
(12.2)
for some constant C depending on m (see [106]).

For a comprehensive study of numerical integration, we refer the reader
to Davis and Rabinowitz [32) and Engels [36).
12.2 Nystrom's Method

We choose a convergent sequence (Qn) of quadrature rules for the integral
Q(g) = fG g(x) dx and approximate the integral operator
(Arp)(x) := fa K(x, y)rp(y) dy, x E C, (12.3)
with continuous kernel K, as introduced in Theorem 2.8, by a sequence of

numerical integration operators
n
(Anrp)(x):= I>:t~n) K(x,x~n))<p(x~n)), x E C. (12.4)
k=l
Then the solution to the integral equation of the second kind
is approximated by the solution of
IOn - An<Pn = j,
which reduces to solving a finite-dimensional linear system.

Theorem 12.7 Let 'Pn be a solution of

n
'Pn(x) - L CtkK(X, Xk)'Pn(Xk) = f(x), x E G. (12.5)
k=l
Then the values 'Pjn) 'Pn(Xj), j 1, ... ,n, at the quadrature points
satisfy the linear system
n
'Pjn) - L CtkK(Xj, Xk)'P~n) = f(xj), j = 1, ... , n. (12.6)
k=l
Conversely, let 'Pjn) , j = 1, ... ,n, be a solution of the system {12.6}. Then
the function 'Pn defined by
n
'Pn(x):= f(x) + LCtkK(X,Xk)'P};), x E G, (12.7)
k=l
solves equation {12.5}.
Proof. The first statement is trivial. For a solution 'Pjn), j = 1, ... , n, of

the system (12.6), the function 'Pn defined by (12.7) has values
= f(xj) + L CtkK(xj, Xk)'P~n) = 'P)n) ,

n
'Pn (Xj) j = 1, ... ,n.
k=l
Inserting this into (12.7), we see that 'Pn satisfies (12.5). o

The formula (12.6) may be viewed as a natural interpolation ofthe values
'P)n), j = 1, ... , n, at the quadrature points to obtain the approximating
function 'Pn and goes back to Nystrom [140].
The error analysis will be based on the following result.
Theorem 12.8 Assume the quadrature formulas (Qn) are convergent. Then
the sequence (An) is collectively compact and pointwise convergent (i. e.,
An'P -+ A'P, n -+ 00, for all 'P E C(G)}, but not norm convergent.
Proof. Because the quadrature formulas (Qn) are assumed to be convergent,

by Theorem 12.4, there exists a constant C such that the weights satisfy
n
L ICt~n) I s:: C
k=l
for all n E 1N. Then we can estimate
IIAn'Plloo s:: C x,yEO

max IK(x, Y)III'Plloo (12.8)
and
for all Xl, X2 E G. Now let U C C (G) be bounded. Then from (12.8) and
(12.9) we see that {An<P : <P E U, n E IN} is bounded and equicontinuous
because K is uniformly continuous on G x G. Therefore, by the Arzela-
Ascoli Theorem 1.18 the sequence (An) is collectively compact.
Since the quadrature is convergent, for fixed <P E C(G) the sequence
(An<P) is pointwise convergent, i.e., (An<P)(x) -+ (A<p)(x), n -+ 00, for all
X E G. As a consequence of (12.9), the sequence (An<P) is equicontinuous.
Hence it is uniformly convergent IIAn °choose a function '¢C E C(G) with °:s
'¢c(x) :s
1 for all
x E G such that '¢c(x) = 1 for all x E G with minj=1, ... ,n Ix - xjl ~ e and
'¢c(Xj) = 0, j = 1, ... , n. Then
IIA<p'¢c - A<plloo :s x,yEa

max IK(x, y)1 { {1- '¢c(Y)} dy -+ 0,
la
e -+ 0,
for all 0
sup sup IIA<p'¢clloo ~ sup IIA<plloo = IIAlloo,

1I'P1l00=1 1£>0 1I'P1l00=1
whence we see that the sequence (An) cannot be norm convergent. 0
Corollary 12.9 For a uniquely solvable integral equation of the second

kind with a continuous kernel and a continuous right-hand side, the Nystrom
method with a convergent sequence of quadrature formulas is uniformly
convergent.
In principle, using Theorem 10.10, it is possible to derive cOlnputable

error bounds (see Problem 12.3). Because these, in general, will be too
complicated to evaluate, as already mentioned in Chapter 11, in applica-
tions it will usually be sufficient to estimate the error by extrapolation from
the convergence order. For the discussion of the error based on the estimate
(10.8) of Corollary 10.11 we need the norm II(A - An)<Plloo. It can be ex-
pressed in terms of the error for the corresponding numerical quadrature
by
and requires a uniform estimate for the error of the quadrature applied
to the integration of K(x, -)ip. Therefore, from the error estimate (lO.8),
it follows that under suitable regularity assumptions on the kernel K and
the exact solution cp, the convergence order of the underlying quadrature
formulas carries over to the convergence order of the approximate solutions
to the integral equation. We illustrate this through the case of the trape-
zoidal rule. Under the assumption cp E C 2 [a, b] and K E C 2 ([a, b] x [a, b]),
by Theorem 12.1, we can estimate
II(A-An)CPlloo::; -2
1 x,yEG uy
I
1 h 2 (b-a) max !'l[)22 K(x,y)cp(y) . I
TABLE 12.1. Numerical solution of (11.30) by the trapezoidal rule
n x=o x = 0.25 x = 0.5 x = 0.75 x=1
4 0.007146 0.008878 0.010816 0.013007 0.015479

8 0.001788 0.002224 0.002711 0.003261 0.003882
16 0.000447 0.000556 0.000678 0.000816 0.000971
32 0.000112 0.000139 0.000170 0.000204 0.000243
Example 12.10 We use the composite trapezoidal rule for approximately

solving the integral equation (11.30) of Example l1.lO. By the numerical
results for the difference between the approximate and exact solutions given
in Table 12.1 the expected convergence rate O(h2) is clearly demonstrated.
We now use the composite Simpson's rule for the integral equation (11.30).
The numerical results in Table 12.2 show the convergence order O(h4),
which we expect from the error estimate (lO.8) and Theorem 12.2. 0
TABLE 12.2. Numerical solution of (11.30) by Simpson's rule
n x=o x = 0.25 x =0.5 x = 0.75 x=1
4 0.00006652 0.00008311 0.00010905 0.00015046 0.00021416

8 0.00000422 0.00000527 0.00000692 0.00000956 0.00001366
16 0.00000026 0.00000033 0.00000043 0.00000060 0.00000086
After comparing the last two examples, we wish to emphasize the ma-
jor advantage of Nystrom's method compared with the degenerate kernel
method of Chapter 11. The matrix and the right-hand side of the linear
system (12.6) are obtained by just evaluating the kernel K and the given
function f at the quadrature points. Therefore, without any further com-

putational effort we can improve considerably on the approximations by
choosing a more accurate numerical quadrature. For example, in the de-
generate kernel method, it is much more involved to design a scheme with
convergence order O(h4) rather than O(h2) from the linear spline interpo-
lation.
Example 12.11 For the integral equation (11.35) of Example 11.12 we use
the trapezoidal rule. Since we are dealing with periodic analytic functions,
from Theorem 12.6 and the error estimate (12.1), we expect an exponen-
tially decreasing error behavior that is exhibited by the numerical results in
Table 12.3. Note that for periodic analytic functions the trapezoidal rule,
in general, yields better approximations than Simpson's rule. 0
TABLE 12.3. Numerical solution of (11.35) by the trapezoidal rule
2n t=O t = 71"/2 t = 71"

4 -0.15350443 0.01354412 -0.00636277
a=l 8 -0.00281745 0.00009601 -0.00004247
b=0.5 16 -0.00000044 0.00000001 -0.00000001
32 0.00000000 0.00000000 0.00000000
4 -0.69224130 -0.06117951 -0.06216587

a=l 8 -0.15017166 -0.00971695 -0.01174302
b=0.2 16 -0.00602633 -0.00036043 -0.00045498
32 -0.00000919 -0.00000055 -0.00000069
Example 12.11 is quite typical for the efficiency of the Nystrom method
with the composite trapezoidal rule for solving two-dimensional bound-
ary value problems for the Laplace equation by using parameterized ver-
sions of the boundary integral equations analogous to the integral equation
of Problem 6.1 for the interior Dirichlet problem. For analytic boundary
curves and boundary data or for boundary curves of class C m +2 , in gen-
eral, the trapezoidal rule yields an approximation order of the form (12.1)
or (12.2), respectively. Hence, this method is extremely efficient, with the
main advantage being its simplicity and its high approximation order. As
a consequence of the latter, in general, the linear systems required for a
sufficiently accurate approximation will be rather small, i.e., the number of
quadrature points will be small. Therefore, despite the fact that they have
full matrices, the linear systems may be solved conveniently by elimination
methods and do not require additional sophisticated efforts for its efficient
solution.
For variants of the Nystrom method for the two-dimensional double-

layer potential integral equation for the Dirichlet problem in domains with
corners, including an error analysis that mimics the existence approach
as described in Section 6.5, we refer to Kress [100] and Kress, Sloan, and
Stenger [110] (see also Atkinson [9]).
We confine ourselves to the above examples for Nystrom's method for
equations of the second kind with a continuous kernel. For a greater variety,
the reader is referred to Anderssen et al. [2], Atkinson [9], Baker [12], Delves
and Mohamed [33], Golberg and Chen [51], Hackbusch [60], and Prossdorf
and Silbermann [150].
12.3 Weakly Singular Kernels

We will now describe the application of Nystrom's method for the approxi-
mate solution of integral equations of the second kind with weakly singular
kernels of the form
(Acp)(x):= faw(lx-yI)K(X,Y)CP(Y)dY, xEGClRm • (12.10)
Here, we assume the weight function w : (0,00) --+ lR to represent the weak
singularity, i.e., w is continuous and satisfies Iw(t)1 ~ Mt Ot - m for all t > 0
and some positive constants M and Q. The remaining part K of the kernel
is required to be continuous. We choose a sequence (Qn) of quadrature
rules n
(Qng)(x):= LQjn)(x)g(xjn)), X E G,
j=l
for the weighted integral
(Qg)(x):= fa w(lx - yl)g(y) dy, x E G,
with quadrature weights depending continuously on x. Then we approx-

imate the weakly singular integral operator by a sequence of numerical
integration operators
n
(AnCP)(x):= L Qkn)(x)K(x, xkn))cp(xkn)), x E G. (12.11)
k=l
Appropriately modified, Theorem 12.7 for the solution of the approximating

equation of the second kind, of course, remains valid. The linear system
corresponding to (12.6) now assumes the form
n
cpjn) _ LQkn)(xj)K(xj,Xk)CPkn) = f(xj), j = 1, ... ,no (12.12)
k=l
12.3 Weakly Singular Kernels 207
Due to the appearance of the weight function w in the quadrature Q, this

form of Nystrom's method for weakly singular kernels is also known as
product integration method.
For the error analysis we will assume that the sequence (Qn) of quadra-
ture formulas converges, i.e., Qng -+ Qg, n -+ 00, uniformly on G for all
g E G(G). Then applying the Banach-Steinhaus theorem to the sequence
of linear operators Qn : G(G) -+ G(G), we observe that for convergence of
the sequence (Qn) it is necessary and sufficient that Qng -+ Qg, n -+ 00,
uniformly on G for all 9 in some dense subset of G (G), and that there exists
a constant G such that the weights satisfy
n
L lakn)(x)1 ::; G (12.13)
k=l
for all n E IN and all x E G. Since the weights are not constant, this bound
alone will not ensure collective compactness of the operator sequence (An).
Therefore, we will also assume that the weights satisfy
(12.14)
uniformly for all x E G. Then, writing

n
(AnCP)(Xl) - (Ancp)(x2) = L akn)(xl){K(xl, xkn») - K(X2,x1n»)}cp(x1n»)
k=l
n
+L {a1n )(Xl) - akn)(x2)}K(X2, X1n»)cp(X1n»)
k=l
we can estimate
n
+ max
x,yEO
IK(x, y)1 sup
nEIN k=l
Liar) (Xl) - a1n ) (x2)lIlcplioo.
Hence, analogous to Theorem 12.8, we can prove the following result.

Theorem 12.12 Assume that the quadrature formulas (Qn) converge and
satisfy the condition (12.14). Then the sequence (An), given by (12.11), is
collectively compact and pointwise convergent AnCP -+ Acp, n -+ 00, for all
cP E G(G), but not norm convergent.
For a systematic study of approximations satisfying condition (12.14) we

refer to Sloan [165]. We confine ourselves to a special case by considering
a weakly singular operator with a logarithmic singularity
10r °
27r t - T) K(t,T)CP(T)dT,
1
(Acp)(t):= 27r In ( 4sin 2 -2- ~ t ~ 27r, (12.15)
in the space C 2tr C C(ffi) of 27r-periodic continuous functions. The ker-

nel function K is assumed to be continuous and 27r-periodic with respect
to both variables. The fundamental solutions to elliptic partial differential
equations, for example the Laplace or the Helmholtz equation, in two space
dimensions contain a logarithmic singularity. Therefore, the boundary inte-
gral equation approach to two-dimensional boundary value problems via a
periodic parameterization of the boundary curve, in general, leads to such
logarithmic singularities (see also Problem 7.2).
According to the general ideas outlined in Section 12.1 and following
Kussmaul [112J and Martensen [117], we construct numerical quadratures
for the improper integral
(Qg)(t) := - 1 127r In (4sin t --

2- T) geT) dT
27r 0 2
by replacing the continuous periodic function g by its trigonometric inter-
polation polynomial described in Section 11.3. Using the Lagrange basis
we obtain
2n-l
(Qng)(t) = L R)n)(t)g(tj) (12.16)
j=O
with the equidistant quadrature points tj = j7r/ n and the quadrature

weights
R)n)(t) = ~ r27r In (4sin2 t -2 T) Lj(T) dT,

27r 10
j = 0, ... , 2n - 1. (12.17)
Using Lemma 8.21, from the form (11.12) of the Lagrange basis we derive
(n) 1 n-l 1 1 }
R j (t) = -;; { ~ m cosm(t - tj) + 2n cosn(t - tj) (12.18)
for j = 0, ... ,2n - 1.

This quadrature is uniformly convergent for all trigonometric polyno-
mials, since by construction Qn integrates trigonometric polynomials of
degree less than or equal to n exactly. We will now establish convergence
for all 27r-periodic continuous functions and the validity of (12.14).
Using the integrals (8.21) and Parseval's equality it can be seen that for
each t E [0,27rJ the function
ft(T) 1
:= -In ( 4sin 2 -
t -- T) , T E ffi,
t-T d71
27r 'F ,
27r 2
belongs to L2[0, 27r] with
f ~m = ~6 .
IIftll~ = ~7r m=l
For fixed t E [0,27r] we choose a 27r-periodic continuous function g with
IIglioo = 1 satisfying g(tj) = 1 if Rjn)(t) ~ 0 and g(tj) = -1 if Rjn)(t) < O.
Then, in view of (12.17), we can write
From this, using Lemma 11.5 and the Cauchy-Schwarz inequality, we can
estimate
2n-l
~ IRjn) (t)1 $llftl121lPngll2 $ ~1I!t1l2 = ~,
and this establishes the uniform boundedness (12.13) for the weights (12.18).
Similarly,
2n-l
L IRjn) (t 1 ) - Rjn)(t2)1 $ ~ IIfh - ft2112,
j=O
and again from (8.21) and Parseval's equality we see that
1 1
leimt1 - eimt2 12 ,
00
IIftl - f t2 1122 = 27r "L....J m2

m=-oo
m;60
whence the validity of condition (12.14) for the weights (12.18) follows.
Hence, for the sequence
2n-l
(An<p)(t) := L Rjn) (t)K(t, tj)<p(tj) (12.19)
j=O
generated by the quadratures (12.16) we can state the following theorem.

Theorem 12.13 The sequence (An) given by {12.19} is collectively com-
pact and pointwise convergent to the integral operator A with logarithmic
singularity given by {12.15}.
Therefore our general convergence and error analysis based on Theorems
10.9 and 10.10 is applicable. By Theorem 11.7, in the case of analytic
functions the quadrature error is exponentially decreasing. By the estimate
(10.8), this behavior is inherited by the approximate solution to the integral
equation provided that the kernel K and the exact solution 'P are analytic
(see Problem 12.4). The matrix entering into (12.12) is a circulant matrix,
i.e.,
R (n)()
k tj =
R(n)
ij-kl' j, k = 0, ... ,2n,
with the weights
R (n) .. = _2. {n2:-1 ~ cos mj7r + (-1)i}, j = 0, ... , 2n - l. (12.20)

J n m n 2n
m=l
Example 12.14 We consider the Neumann boundary value problem for
the reduced wave equation or Helmholtz equation
6.u + K 2 U = 0
with a positive wave number K in the open unit disk D with given normal
derivative
au
a// = 9 on aD.
We assume the unit normal // to be directed into the exterior of D. The
fundamental solution to the two-dimensional Helmholtz equation is given
by
(x, y) ;= ~ H6 1 ) (Klx - yl), x -I- y,
where H6
1 ) denotes the Hankel function of the first kind and of order zero.
We decompose
H6
1 ) = J o + iNo,
where Jo and No are the Bessel and Neumann functions of order zero, and
note the power series
(_1)k (Z)2k
2:
00
Jo(z) = (k!)2 "2 (12.21 )

k=O
and
No(z) = -
2 (InZ )
- +C
2
Jo(z) - - "" ak
00 (_1)k (Z)
-- -
2k (12.22)
7r 2 7r L.; (k!)2 2
k=l
with ak = L:;;"=l 11m and Euler's constant C = 0.57721 .... From these
expansions we deduce the asymptotics
(x, y) = - 1 In -1- + -i - -1(K)

In - + C + 0 ( Ix - yl 2
In1 --)
27r Ix - yl 4 27r 2 Ix - yl
for Ix - yl -+ O. Therefore, the fundamental solution to the Helmholtz
equation has the same singular behavior as the fundamental solution of
Laplace's equation. As a consequence, Green's representation Theorem 6.5

and the potential theoretic jump relations of Theorems 6.14 and 6.17 can
be carried over to the Helmholtz equation. For details we refer to [24]. In
particular, it can be shown that the unknown boundary values c.p = u on
aD of the solution u satisfy the integral equation
c.p(x) +2 1 aD
c.p(y)
aiP(x, y)
a () ds(y)=2
v Y
1 aD
g(y)iP(x,y)ds(y) (12.23)
for x E aD. With the exception of a countable set of wave numbers", ac-
cumulating only at infinity, for which the homogeneous Neumann problem
admits nontrivial solutions, this integral equation is uniquely solvable.
We use the representation x(t) = (cost,sint), 0 :::; t :::; 211", for the unit
circle and, by straightforward calculations, transform the integral equation
(12.23) into the parametric form
cp(t) - -
1 127r K(t, r)cp(r) dr = -21 127r L(t,r)g(r)dr (12.24)
211" 0 11" 0
for 0:::; t :::; 211". Here we have set cp(t) := c.p(x(t)) and g(t) := g(x(t)), and
the kernels are given by
and
L(t, r) := i1l" H~l) (2'" !sin t ~ r J)
for t f=. r. For computing the normal derivative of the fundamental solution
we used the relation
H o(l)1 -- _H(l) -
1 -
-J1 - Z·N1,
where H~1) denotes the Hankel function of the first kind and of order one
represented in terms of the Bessel and Neumann functions J 1 and N1 of
order one.
Observing the expansions (12.21) and (12.22) and their term-by-term
derivatives, we can split the kernels into
t
K(t,r) = K1(t,r)ln ( 4sin2 -2- -r) + K (t,r) 2 (12.25)
and
t-
L(t,r)=L1(t,r)ln ( 4sin2 -2- +L 2 (t,r), r) (12.26)
where
t-T J I ( 2Ksin -2-
KI(t, T) := -Ksin -2- t-T) ,
t-
LI(t,T) := -Jo ( 2Ksin-
2-
T) ,
The kernels K I , K 2 , L}, and L2 turn out to be analytic. In particular, from
lim ZH(I)(Z) = ~
z--+o 1 7rZ
and
2i ) 2i
lim ( Ho(1) (z)--lnzJo(z) =-(0-ln2)+1,
z--+o 7r 7r
we deduce the diagonal terms
K2(t, t) = K(t, t) = 1
and
L 2 (t, t) = -2ln ~ - 20 + i7r
for °: ;
t ::; 27r. Note that despite the continuity of the kernel K, for
numerical accuracy it is advantageous to incorporate the quadrature (12.16)
because of the logarithmic singularities of the derivatives of the kernel K.
Now, in the spirit of Theorem 12.7, we approximate the integral equation
(12.24) by the linear system
j = 0, ... ,2n - 1,
for approximating values 'Pj for 'P( tj).

For a numerical example we consider the case in which the exact solution
to the Neumann problem is given by u(x) = NO(Klx-xo\), where xo = (q,O)
with q > 1. Then the Neumann boundary data are given by
_ 1-qcost
get) = -K vet) NI(KV(t))
12.4 Nystrom's Method in Sobolev Spaces 213
and the exact solution of the integral equation by ip(t) = No(/w(t)), where
v2(t) = 1 + q2 - 2qcost. The numerical results contained in Table 12.4
confirm the exponentially decreasing behavior of the error we expect from
our general error analysis. 0
2n t=O t = 7r/2 t = 7r
4 -0.07907728 0.12761991 0.24102137

11:=1 8 -0.01306333 0.00867450 0.01155067
q=2 16 -0.00023494 0.00003924 0.00004517
32 -0.00000019 0.00000001 0.00000000
4 -0.13910590 0.39499045 0.68563472

11:=1 8 -0.07111636 0.07511294 0.10540051
q = 1.5 16 -0.00659502 0.00277753 0.00386074
32 -0.00005924 0.00000558 0.00000616
8 0.35729406 0.21301358 -0.16596385

11:=5 16 0.01138634 0.00181974 0.00377600
q=2 32 -0.00000558 -0.00000006 -0.00000010
64 0.00000003 0.00000002 0.00000000
8 -0.54146680 -0.35298932 0.09706015

11:=5 16 -0.05669554 -0.02916764 -0.02480085
q = 1.5 32 -0.00021246 -0.00000877 -0.00001332
64 -0.00000006 0.00000000 0.00000003
As at the end of Section 12.2, we wish to point out that Example 12.14
is typical for the efficiency of the Nystrom method based on trigonometric
interpolatory quadratures for logarithmic singularities applied to boundary
value problems for the Helmholtz equation. For details we refer to [25] (see
also Problem 12.5).
12.4 Nystrom's Method in Sobolev Spaces

In this final section we wish to illustrate that the Nystrom method can
also be analyzed in the Sobolev spaces HP[O,27r] and in Holder spaces.
We confine our presentation to the weakly singular integral operator with
logarithmic singularity defined by (12.15) and begin by investigating its
mapping properties in Sobolev spaces.
Theorem 12.15 Assume that K is infinitely differentiable and 27r-periodic

with respect to both variables. Then the operator A given by (12.15) is
bounded from HP[O, 27r] into HP+l [0, 27r] for all p ~ o.
Proof. We write
(12.27)
m=-oo
where
am(t) := 2~ 127r K(t, r)e- imr dr
denote the Fourier coefficients of K(t, .). The infinite differentiability of K
implies that
sup ImIPlla}:;:) 1100 < 00 (12.28)
mE71
for all k E IN U {O} and all p ~ 0 (see (8.6)). Therefore, the series (12.27)
and its term-by-term derivatives of arbitrary order converge uniformly. Re-
call the trigonometric monomials f m (t) = eimt . Then, for trigonometric
polynomials rp we have
m=-oo
where
(Aorp)(t) := 27r
1 [27r
io
( t-
In 4sin 2 -2-
r) rp(r) dr, o :::; t :::; 27r.
By Theorem 8.22 the operator Ao is bounded from HP[O, 27r] into HP+l [0, 27r]
for all p E IR. Using Corollary 8.8 we can estimate
00
IIArpllp+l :::; C1 :L
m=-(X)
[II a}:;:) 1100 + Ilamiloo] II A o(rpfm)llp+l,
where k is the smallest integer larger than or equal to p+ 1 and C 1 is some

constant depending on p. The boundedness of Ao and again Corollary 8.8
yield that
IIAo(rpfm)llp+l :::; C 2 11rpfmllp :::; C 3 1mlkllrpllp
for all mE 7f, and some constants C2 and C3 depending on p. The last two
inequalities, together with (12.28), imply that
for all trigonometric polynomials rp and some constant C depending on p.

Now, in view of Theorem 8.2, the proof is complete. 0
12.4 Nystrom's Method in Sobolev Spaces 215
Corollary 12.16 Under the assumptions of Theorem 12.15, the operator

A : HP[O, 271"] -+ HP[O, 271"] is compact for all p 2: 0.
Proof. This follows from the imbedding Theorem 8.3. o
Corollary 12.16 allows the application of the Riesz Theorem 3.4. In view
of Theorem 8.4 and the smoothing property of A from Theorem 12.15, the
inverse operator (J - A)-I: HP[O, 271"] -+ HP[O, 271"] exists and is bounded
°
if and only if the homogeneous equation r.p - Ar.p = has only the trivial
continuous solution r.p = 0.
The operator Ao corresponds to the single-layer potential operator for
the unit disk. Therefore, by Theorem 7.30 the operator Ao is bounded from
the Holder space C~;..a into c~;..a, and proceeding as in the proof of Theorem
12.15 we can establish the following result.
°
Theorem 12.17 Under the assumptions of Theorem 12.15 the operator A
is bounded from C~;..c< into C~;"C< for all < a < 1.
°
Theorem 12.18 Assume that 5 q 5 p and p > 1/2. Then under the
assumptions of Theorem 12.15 for the operator sequence (An) given by
(12.19) we have the estimate
IIAnr.p - Ar.pllq+1 5 Cnq-Pllr.pllp (12.29)
for all r.p E HP[O, 271"] and some constant C depending on p and q.
Proof. As in the proof of Theorem 12.15 we can write
00
m=-oo
and estimate analogously. Using the boundedness of A o, Theorem 11.8, and
Corollary 8.8, we obtain
IIAo[Pn(r.pfm) - r.pfmlllq+1 5 cIllPn(r.pfm) - r.pfmllq
5 c2nq-PIIr.pfmllp 5 c3Iml p+ l nq- P Ilr.pllp

for all mE 7l and some constants CI, C2, and C3 depending on p and q. Now
the proof can be completed as in Theorem 12.15. 0
Corollary 12.19 Under the assumptions of Theorem 12.15, for the opera-
tors An : HP[O, 271"] -+ HP[O, 271"], we have norm convergence II An - Allp -+ 0,
n -+ 00, for all p 2: 1.
Proof. Setting q = p - 1 in (12.29) yields
C
IIAnr.p - Ar.plip 5 - 1Ir.pllp
n
for all r.p E HP[O, 271"] and all n E IN. This establishes the assertion. 0
Now we are in the position to apply Theorem 10.1 and the error estimate
(10.3) for the approximate solution of <p - A 1/2 also implies uniform convergence. As
opposed to the case of the convergence and error analysis in the space of
continuous functions, due to the regularity properties in the Sobolev spaces,
here we do not rely on collective compactness. We leave it as an exercise
to the reader to extend this analysis to the integral operator
= 10r
27r
(A<p)(t) K(t,r)<p(r)dr
with infinitely differentiable 27r-periodic kernel K and its approximations

2n-l
(Ancp)(t) = ~ L K(t, tk)<p(tk)
n k=O
by the composite trapezoidal rule. We also leave it as an exercise to carry

Theorem 12.18 and its Corollary 12.19 over to Holder spaces as in Theorem
12.17 by using the error estimate (11.20).
For an error analysis of a Nystrom method for Cauchy-type singular
integral equations in Sobolev spaces, which imitates the regularization of
singular equations considered in Chapter 7, we refer to Kirsch and Rit-
ter [93].
Problems
12.1 Let X and Y be normed spaces, let U C X be compact, and let (CPn) be
an equicontinuous sequence of functions cpn : U -t Y converging pointwise on U
to some function cP : U -t Y. Then the sequence (CPn) converges uniformly on U.
12.2 Let D = ill. x (-8, s) denote a strip in the complex plane. Show that the
space H(D) of 21T-periodic holomorphic functions defined on D with the property
lim
0'-+8
1 0
211"
{If(t + iuW + I!(t - iu)12} dt < 00
is a Hilbert space with the scalar product
(f, g) := lim
0'-+8
10
211"
{f(t + iu)g(t + iu) + f(t - iu)g(t - iu)} dt.
Show that the functions
f () 1 int n E 7l,
n t = (41Tcosh2ns)1/2 e ,
Problems 217
form a complete orthonormal system in H(D). Use Cauchy's integral formula to

verify that the remainder for the trapezoidal rule is a bounded linear functional
on H(D). Apply the Riesz representation Theorem 4.8 and Parseval's equality
from Theorem 1.28 to derive the error bound
2M
IRn(g)1 ::; (e4ns _ 1)1/2
which is slightly better than the estimate given in Theorem 12.6 (see [36J and
[97]).
12.3 Derive bounds on II(An - A)Alloo and II(An - A)Anlloo for the numerical
integration operators using the trapezoidal rule.
12.4 Let D =lR x (-s, s) denote a strip in the complex plane. Consider the
Banach space B(D) of 27r-periodic continuous functions defined on [) that are
holomorphic in D, furnished with the maximum norm
Il/lIoo.D := rna]{
tED
I/(t)l·
Prove that under suitable assumptions on K the integral operator given by (12.15)
is compact in B(D). Use this result to deduce that solutions to integral equations
of the second kind with the logarithmic kernel of (12.15) are analytic provided
the kernel K and the right-hand side are analytic.
Hint: By the Arzela-Ascoli theorem, show that integral operators with analytic
kernels are compact in B(D). Use
(Acp)(t) := 2~ 127r In (4sin2 i) K(t, t + s)cp(t + s) ds

for the definition of A on B(D). Approximate Insin2 (s/2) by its Fourier series
and use Theorem 2.17 and the Cauchy-Schwarz inequality.
12.5 Use a regular 21r-periodic parameterization aD = {x(t) : 0 ::; t ::; 27r}

with counterclockwise orientation for an arbitrary boundary curve aD to derive
a parametric form of the integral equation (12.23) for the Neumann problem
for the reduced wave equation corresponding to (12.24). Describe an appropriate
treatment ofthe logarithmic singularity as in Example 12.14 (see [25, 98, 104, 112J
for related integral equations).
13
Projection Methods
The application of the quadrature method, in principle, is confined to equa-

tions of the second kind. To develop numerical methods that can also be
used for equations of the first kind we will describe projection methods as a
general tool for approximately solving operator equations. After introduc-
ing into the principal ideas of projection methods and their convergence
and error analysis we shall consider collocation and Galerkin methods as
special cases. We do not intend to give a complete account of the numerous
implementations of collocation and Galerkin methods for integral equa-
tions that have been developed in the literature. Our presentation is meant
as an introduction to these methods by studying their basic concepts and
describing their numerical performance through a few typical examples.
For a more exhaustive study of projection methods in general and of
collocation and Galerkin methods, we refer to Atkinson [9], Baker [12],
Fenyo and Stolle [39], Hackbusch [60], Krasnoselski et al. [95], Mikhlin and
Prossdorf [124], and Pr6ssdorf and Silbermann [149, 150]. For an intro-
duction to projection methods for boundary integral equations with an
extensive bibliography, we refer to the review paper by Sloan [166].
13.1 The Projection Method

We describe the approximate solution of linear operator equations by pro-
jecting them onto subspaces, which for practical calculations we assume to
be finite-dimensional.

13.1 The Projection Method 219
Definition 13.1 Let X be a normed space and U c X a nontrivial sub-

space. A bounded linear operator P : X -7 U with the property Pcp = cp for
all cp E U is called a projection operator from X onto U.
Theorem 13.2 A nontrivial bounded linear operator P mapping a normed

space X into itself is a projection operator if and only if p 2 = P. Projection
operators satisfy IIPII 2 1.
Proof. Let P : X -7 U be a projection operator. Then, from Pcp E U it

follows that p2cp = P(Pcp) = Pcp for all cp E X. Conversely, let p2 = P and
set U := P(X). Then, for all cp E U we may write cp = P'tf; for some 'tf; E X
and obtain Pcp = cp. Finally P = p2, by Remark 2.7, implies IIPII :S IIPII 2,
whence IIPII 2 1. 0
An important example for projection operators is given by the so-called

orthogonal projection, i.e., by the best approximation in pre-Hilbert spaces
in the sense of Theorem 1.26.
Theorem 13.3 Let U be a nontrivial complete subspace of a pre-Hilbert

space X. Then the operator P mapping each element cp E X into its unique
best approximation with respect to U is a projection operator. It is called
orthogonal projection onto U and satisfies IIPII = 1.
Proof. Trivially, we have Pcp = cp for all cp E U. From the orthogonality

condition of Theorem 1.25 for the best approximation in pre-Hilbert spaces
we readily verify that P is linear and
for all cp E X. Hence IIPII :S 1, and Theorem 13.2 yields IIPII = 1. 0
A second important example for projection operators is given by inter-

polation operators.
Theorem 13.4 The interpolation operator introduced in Theorem 11.2 is

a projection operator.
Note that for polynomial or trigonometric polynomial interpolation, be-

cause of the general nonconvergence results due to Faber (see [135, 162])
combined with the uniform boundedness principle Theorem 10.4, the inter-
polation operators are not uniformly bounded with respect to the maximum
norm (see also the estimate (11.15)).
Definition 13.5 Let X and Y be Banach spaces and let A : X -7 Y be

an injective bounded linear operator. Let Xn C X and Y n C Y be two
sequences of subspaces with dim Xn = dim Y n = n and let Pn : Y -7 Yn
be projection operators. Given fEY, the projection method, generated by
220 13. Projection Methods
Xn and Pn , approximates the equation
A<p=f (13.1)
for <P E X by the projected equation
(13.2)
for <Pn E X n . This projection method is called convergent for the opemtor
A if there exists no E IN such that for each f E A(X) the approximating
equation (13.2) has a unique solution <Pn E Xn for all n 2': no and these
solutions converge <Pn -+ <p, n -+ 00, to the unique solution <P of A<p = f.
In terms of operators, convergence of the projection method means that

for all n 2': no the finite-dimensional operators PnA : Xn -+ Yn are invert-
ible and we have pointwise convergence (PnA)-l PnAEXn
for all <p EX. Therefore, in the subsequent analysis we will always assume
that this condition is fulfilled.
Since PnA : Xn -+ Y n is a linear operator between two finite-dimensional
spaces, carrying out the projection method reduces to solving a finite-
dimensional linear system. In the following sections we shall describe the
collocation and the Galerkin method as projection methods in the sense of
this definition obtained via interpolation and orthogonal projection oper-
ators, respectively. Here, we first proceed with a general convergence and
error analysis.
Theorem 13.6 A projection method converges for an injective linear op-

emtor A : X -+ Y from a Banach space X into a Banach space Y if and
only if there exist no E IN and a positive constant M such that for all
n 2': no the finite-dimensional opemtors
are invertible and the opemtors (PnA)-l PnA : X -+ Xn are uniformly

bounded, i.e.,
(13.4)
for some constant M. In case of convergence we have the error estimate
li<Pn - <p1I :s; (1 + M) 'l/>EX

inf 11'1/1 - <p1I· (13.5)
n
13.1 The Projection Method 221
Proof. Provided the projection method converges for the operator A, the
uniform boundedness (13.4) is a consequence of Theorem 10.4. Conversely,
if the assumptions of the theorem are satisfied we can write
Since for all 'IjJ E X n , trivially, we have (Pn A)-1 PnA'IjJ = 'IjJ, it follows that
Hence, we have the error estimate (13.5) and, with the aid of the denseness
(13.3), the convergence follows. 0
The uniform boundedness condition (13.4) is also known as the stability

condition for the projection method. The error estimate (13.5) of Theorem
13.6 is usually referred to as Cea's lemma (see [21]). It indicates that the
error in the projection method is determined by how well the exact solution
can be approximated by elements of the subspace X n .
We now state the main stability property of the projection method with
respect to perturbations of the operator.
Theorem 13.7 Assume that A : X -+ Y is a bounded linear operator
with a bounded inverse A-I: Y -+ X and that the projection method is
convergent for A. Let B : X -+ Y be a bounded linear operator such that
either
(1) IIPnBllxn-+Yn = sup IIPnB<P1I is sufficiently small for all suf-
cpEX n , 11'1'11=1
ficiently large n, or
(2) B is compact and A + B is injective.
Then the projection method also converges for A + B.
Proof. The operator A satisfies the conditions of Theorem 13.6, i.e., for
all sufficiently large n, the operators PnA : Xn -+ Yn are invertible and
satisfy II(PnA)-IPnAII:::; M for some constant M. Since A has a bounded
inverse, the pointwise convergence (Pn A)-1 PnA -+ I, n -+ 00, on X implies
pointwise convergence (Pn A)-IPn -+ A-I, n -+ 00, on Y. We will show
that for sufficiently large n the inverse operators of I + (Pn A)-1 PnB :
Xn -+ Xn exist and are uniformly bounded if (1) or (2) is satisfied.
If (1) is satisfied, we apply the uniform boundedness principle Theorem
10.4 to the pointwise convergent sequence (Pn A)-1 Pn . Then, by Theorem
2.9, the inverse operators [1 + (PnA)-1 PnBj-l : Xn -+ Xn exist and are
uniformly bounded for all n E 1N for which II (Pn A)-1 PnlllIPnBllxn-+Yn :::; q
for some q < 1. Here, we have used the property P~ = Pn for projection
operators.
If (2) is satisfied, by the Riesz theory 1 + A-I B : X -+ X has a bounded
inverse, since A-I B is compact. From the pointwise convergence of the se-
quence (Pn A)-1 Pn and the compactness of B, by Theorem 10.7, we derive
norm convergence 11[1 + A-I B]- [I + (PnA)-1 Pn B ]II ~ 0, n ~ 00. There-

fore, by Theorem 10.1, the inverse [I + (PnA)-1 PnB]-1 : Xn -+ Xn exists
and is uniformly bounded for all sufficiently large n.
We set S := A + B. Then from PnS = PnA[I + (PnA)-1 PnB] it follows
that PnS : Xn -+ Y n is invertible for sufficiently large n with the inverse
given by
(PnS)-1 = [I + (PnA)-1 PnBtl(PnA)-I.
From (PnS)-1 PnS = [I + (PnA)-1 PnB]-I(PnA)-IPnA(I +A-I B) we can
estimate
and observe that the condition (13.4) is satisfied for S. This completes the
~~ 0
Because, in general, projection methods are only semi-discrete methods,

in actual numerical calculations instead of (13.2) an approximate version
of the form
PnAn~n = Pnfn (13.6)
will usually be solved, where An is some approximation to A and fn ap-
proximates f. For this we can state the following result.
Theorem 13.8 Assume that the projection method converges for the in-
jective bounded linear operator A : X -+ Y and that the approximating
bounded linear operators An : X -+ Y in addition to pointwise convergence
(PnAn - PnA)'¢ -+ 0, n ~ 00, for all '¢ E X satisfy
IlPnAn - PnAllxn-+Yn -+ 0, n -+ 00.
Then for sufficiently large n the approximate equation (13.6) is uniquely

solvable and we have an error estimate
for the solution i.p of {13.1} and some constant C.

Proof. By setting B = An - A in Theorem 13.7 it follows that for sufficiently
large n the inverse operators of PnAn : Xn -+ Yn exist and are uniformly
bounded. With the aid of
~n - i.pn = [(PnAn)-l - (PnA)-I]PnArp + (PnAn)-1 PnUn - f)
= (PnAn)-l[PnA - PnAn](rpn - rp)
+(PnAn)-I[PnA - PnAn]rp + (PnAn)-IPnUn - f)

the error estimate follows from Theorem 13.6 and the uniform boundedness
principle Theorem 10.4. 0
13.2 Projection Methods for Equations of the Second Kind 223
13.2 Projection Methods for Equations

of the Second Kind
For an equation of the second kind
cP - Acp =f (13.8)
with a bounded linear operator A : X --+ X we need only a sequence of

subspaces Xn C X and projection operators Pn : X --+ X n . Then the
projection method assumes the form
(13.9)
Note that each solution CPn E X of (13.9) automatically belongs to X n .

When A is compact, from Theorem 13.7 we have the following convergence
property.
Corollary 13.9 Let A : X --+ X be compact, 1 - A be injective, and the
projection opemtors Pn : X --+ Xn converge pointwise Pncp --+ cP, n --+ 00,
for all cP EX. Then the projection method for 1 - A converges.
Proof. We apply the second case of Theorem 13.7. o
We wish to give an alternate proof of the latter convergence result based

directly on Theorem 10.1. This will also give us the opportunity to point out
that the projection method for equations of the second kind may converge
without pointwise convergence of the projection operators on all of X.
Theorem 13.10 Let A : X --+ X be compact and 1 - A be injective.
Assume that the projections Pn : X --+ Xn satisfy IlPnA - All --+ 0, n --+ 00.
Then, for sufficiently large n, the approximate equation (13.9) is uniquely
solvable for all f E X and we have an error estimate
(13.10)
for some positive constant M depending on A.

Proof. Theorems 3.4 and 10.1, applied to 1 - A and 1 - PnA, imply that
for all sufficiently large n the inverse operators (1 - PnA)-l exist and are
uniformly bounded. To verify the error bound, we apply the projection
operator Pn to (13.8) and obtain
cP - PnAcp = Pnf + cP - Pncp.
Subtracting this from (13.9) we find
(1 - PnA)(CPn - cp) = Pn'P - cP,
whence the estimate (13.10) follows. o
Note that for a compact operator A pointwise convergence Pn<P --+ <p,
n --+ 00, for all <P EX, by Theorem 10.7 implies that IlPnA - All --+ 0,
n --+ 00. But IlPnA - All --+ 0, n --+ 00, may be satisfied without pointwise
convergence of the projection operator sequence (Pn ) as we will see in our
discussion ofthe collocation method (see Theorems 13.15 and 13.16). Then,
of course, we can assure convergence only for those cases where we have
convergence Pn<P --+ <p, n --+ 00, for the exact solution <po
Since projection methods are only semi-discrete, in actual numerical cal-
culations instead of (13.9) an approximate version of the form
(13.11)
needs to be solved, where An is an approximation of A and f n approximates

f. Then we can state the following result.
Corollary 13.11 Under the assumptions of Theorem 13.10 on the opem-
tor A and the projection opemtors Pn assume that (PnAn - PnA)'I/J --+ 0,
n --+ 00 for all 'I/J E X and IlPnAn - PnAllxn-+xn --+ 0, n --+ 00. Then, for
sufficiently large n, the approximate equation (13.11) is uniquely solvable
and we have an error estimate
lI<Pn - <p1I :::; M{llPn<P - <p1I + IlPn(An - A)<p1I + IlPn(Jn - J)II} (13.12)
for some positive constant M.

Proof. We apply Theorem 10.1 to the operators 1 - PnA : Xn --+ Xn and
1 - PnAn : Xn --+ Xn to establish the existence and uniform boundedness
of the inverse operators (1 - PnAn)-l : Xn --+ Xn for sufficiently large n.
From Theorem 10.1 we also obtain the error estimate
for some constant C. Now (13.12) follows from (13.10) and the uniform
boundedness principle Theorem 10.4. 0
We conclude with a theorem that will enable us to derive convergence

results for collocation and Galerkin methods for equations of the first kind.
Consider the equation
8<p - A<p = f, (13.13)
where 8 : X --+ Y is assumed to be linear and bounded with a bounded
inverse 8- 1 : Y --+ X and where A : X --+ Y is compact and 8 - A is
injective. Then, for the projection method
(13.14)
with subspaces Xn and Yn and projection operators Pn : Y --+ Yn chosen

as in Definition 13.5 we have the following theorem.
13.3 The Collocation Method 225
Theorem 13.12 Assume that Yn = S(Xn) and IlPnA - All -+ 0, n -+ 00.

Then, for sufficiently large n, the approximate equation (13.14) is uniquely
solvable and we have an error estimate
(13.15)
for some positive constant M depending on Sand A.
Proof. The equation (13.14) is equivalent to
S-1 PnS(I - S-1 A)CPn = S-1 Pn SS- 1f,
i.e.,
Qn(I - S-1 A)CPn = Q n S- 1f,
where Qn := S-1 PnS : X -+ Xn obviously is a projection operator. Using
IIQnS-1A - S-1AII = IIS- 1(Pn A - A)II and QnCP - cP = S-1(Pn Scp - Scp),
the assertion follows from Theorem 13.10. 0
The essence of Theorem 13.12 lies in the fact that the projection method
is applied directly to the given equation whereas the convergence result is
based on the regularized equation of the second kind.
In actual numerical calculations, instead of (13.14) an approximate ver-
sion of the form
(13.16)
will usually be solved, where An and fn are approximations of A and f,
respectively. Then, analogous to Corollary 13.11, we have the following
result.
Corollary 13.13 Under the assumptions of Theorem 13.12 on the oper-
ators S and A assume that (PnAn - PnA)'if; -+ 0, n -+ 00 for all 'if; E X
and IlPnAn - PnAllxn-+Yn -+ 0, n -t 00. Then, for sufficiently large n,
the approximate equation (13.16) is uniquely solvable and we have an error
estimate
lI~n - cpli ~ M{IIPnScp - Scpll + IIPn(An - A)cpli + IIPn(fn - f)1I} (13.17)

for some positive constant M.
13.3 The Collocation Method

The collocation method for approximately solving the equation
Acp = f, (13.18)
roughly speaking, consists of seeking an approximate solution from a finite-

dimensional subspace by requiring that (13.18) is satisfied only at a finite
number of so-called collocation points. To be more precise, let Y = C( G)

and A : X ~ Y be a bounded linear operator. Let Xn C X and Yn C Y
denote a sequence of subspaces with dim Xn = dim Yn = n. Choose n points
(n) (n).In G (we aso
Xl , ... ,Xn
1 WI·11 WrIte . t ead 0 f Xl(n) , ... ,xn(n»)
. XI, ... ,Xn Ins
such that the subspace Yn is unisolvent (see Theorem 11.2) with respect
to these points. Then the collocation method approximates the solution of
(13.18) by an element CPn E Xn satisfying
(13.19)
Let Xn = span{UI, ... , un}. Then we express CPn as a linear combination
n
CPn = L "IkUk
k=l
and immediately see that (13.19) is equivalent to the linear system
n
TI"Ik(Auk)(Xj) = f(xj)' j = 1, ... ,n, (13.20)
k=l
for the coefficients "11, .•. , "In. The collocation method can be interpreted as
a projection method with the interpolation operator Pn : Y ~ Yn described
in Theorem 13.4. Indeed, because the interpolating function is uniquely
determined by its values at the interpolation points, equation (13.19) is
equivalent to
PnACPn = Pnf.
Hence, our general error and convergence results for projection methods
apply to the collocation method.
Note that the collocation method can also be applied in function spaces
other than the space C(G), for example in Holder spaces or Sobolev spaces.
For an equation of the second kind cP - Acp = f in X = C( G) with
a bounded linear operator A : X -7 X we need only one sequence of
subspaces Xn C X and assume Xn to be unisolvent with respect to the
collocation points. Here, the equations (13.20) assume the form
n
L"Iduk(Xj) - (AUk)(Xj)} = f(xj), j = 1, ... ,n. (13.21)
k=l
We shall first consider the collocation method for integral equations of
the second kind
cp(X) - fa K(x, y)cp(y) dy = f(x), X E G,
with continuous or weakly singular kernel K. If we use the Lagrange basis

for Xn and write
then "Ij = <fIn(Xj), j = 1, ... , n, and the system (13.21) becomes
(13.22)
From the last equation we observe that the collocation method for equations
of the second kind is closely related to the degenerate kernel method via
interpolation. Obviously, the matrix of the system (13.22) coincides with
the matrix (11.25) from the degenerate kernel approach. But the right-
hand sides of both systems are different and the solutions "11, ... ,"In have
different meanings. However, if the right-hand side f happens to belong to
X n , then f = Pnf and we may express the collocation solution <fin also in
the form
n
<fin = f + I>YkLk.
k=1
Now, for the coefficients 71.'" ,7n we obtain the same system as (11.25),
and in this case the collocation method and the degenerate kernel method
via interpolation will yield the same numerical results.
Analogous to the degenerate kernel method the collocation method is
only a semi-discrete method. In principle, the simplest approach to make
the collocation method fully discrete is to use a quadrature operator An of
the form (12.4) or (12.11) from the previous chapter with the collocation
points as quadrature points to approximate A in the sense of Corollary
13.11. In this case, however, the solution (j5n of
will coincide at the collocation points with the solution 1/Jn of the corre-
sponding Nystrom equation
since both equations lead to the same linear system if we use the Lagrange
basis for the collocation method. In particular, we have the relation
This follows from the fact that Xn := f + An(j5n satisfies
and the observation that AnPnX = AnX for all X, since the collocation and
quadrature points coincide. Hence, this type of approximation can be ana-
lyzed and interpreted more conveniently as a Nystrom method. Therefore,
in the sequel for integral equations of the second kind we will consider only
fully discrete collocation methods which are based on numerical integration

of the matrix elements.
A broad variety of collocation methods exists corresponding to various
choices for the subspaces X n , for the basis functions UI, •.. , Un, and for the
collocation points Xl, •.. , X n . We briefly discuss two possibilities based on
linear splines and trigonometric interpolation. Whe:Q. we use linear splines
(or more general spline interpolations) we have pointwise convergence of
the corresponding interpolation operator as can be seen from the Banach-
Steinhaus theorem by using Theorem 11.3 and the norm (11.9) of the linear
spline interpolation operator. Therefore, in this case, Corollary 13.9 applies,
and we can state the following theorem.
Theorem 13.14 The collocation method with linear splines converges for
integral equations of the second kind with continuous or weakly singular
kernels.
Provided the exact solution of the integral equation is twice continuously
differentiable, from Theorems 11.3 and 13.6 we derive an error estimate of
the form
for the linear spline collocation approximate solution 'Pn. Here, M denotes
some constant depending on the kernel K.
TABLE 13.1. Collocation method for Example 11.10
n x=o X = 0.25 x=0.5 X = 0.75 x=l
4 0.003984 0.004428 0.004750 0.004978 0.005135

8 0.000100 0.001112 0.001193 0.001250 0.001291
16 0.000250 0.000278 0.000298 0.000313 0.000323
32 0.000063 0.000070 0.000075 0.000078 0.000081
In most practical problems the evaluation of the matrix entries in (13.22)

will require numerical integration. For twice continuously differentiable ker-
nels, as in the case of the degenerate kernel method, we suggest the quadra-
ture rule (11.27). Then, according to our remarks concerning the connection
between the equations (11.25) and (13.22), the collocation method will pro-
duce the same numerical results as the degenerate kernel method. So we
do not have to repeat the numerical calculations for equation (11.30) of
Example 11.10 using the quadrature (11.27). However, to illustrate that it
does not payoff to evaluate the matrix elements more accurately, we give
the numerical results for the spline collocation for equation (11.30) using
the exact matrix entries. Comparing Tables 11.1 and 13.1, we observe that
the improvement in the accuracy for the solution of the integral equation
is only marginal. In general, instead of investing computing efforts in a

highly accurate evaluation of the matrix elements it is more efficient to
increase the number of collocation points, i.e., to increase the size of the
linear system.
We proceed discussing the collocation method based on trigonometric
interpolation with equidistant knots tj = j7r In, j = 0, .. , 2n - 1.
Theorem 13.15 The collocation method with trigonometric polynomials
converges for integral equations of the second kind with 27r-periodic contin-
uously differentiable kernels and right-hand sides.
Proof. The integral operator A with 27r-periodic continuously differentiable
kernel K maps C27r into CI7r , and we have
IIA~I/oo + II (A~ )'1/00 S; 27r {I/Klloo + II OJ{ lloo} I/~I/oo

for all ~ E C27r • From Theorems 8.4 and 11.8 we can conclude that
for all ~ E C27r , all 1/2 < q < 1, and some constants Cl and C2 depending
on q. Combining these two inequalities and using Theorem 8.5 yields
for all 112 < q < 1 and some constant C depending on q. Now the assertion
follows from Theorem 13.10. 0
One possibility for the implementation of the collocation method is to

use the monomials fm(t) = eimt as basis functions. Then the integrals
J: 7r
K(tj,r)eikTdr have to be integrated numerically. Using fast Fourier
transform techniques (see Nussbaumer [139]) these quadratures can be car-
ried out very rapidly. A second even more efficient possibility is to use the
Lagrange basis as already described in connection with the degenerate ker-
nel method. This leads to the quadrature rule (11.33) for the evaluation of
the matrix elements in (13.22). Because of the simple structure of (11.17),
the only computational effort besides the kernel evaluation is the computa-
tion of the row sums E;:~~ (-I)m K(tj, t m ) for j = 0, ... 2n -1. Of course,
with this quadrature the collocation method again will coincide with the
degenerate kernel method via trigonometric interpolation.
We now proceed to illustrate the application of the collocation method
for equations of the second kind with weakly singular kernels. As in Sections
12.3 and 12.4 we consider the operator
1 127r
(A~)(t) := 27r 0
t -
In ( 4sin2 -2- r)
K(t, r)~(r) dr, os; t S; 27r, (13.23)
where the kernel function K is assumed to be 21r-periodic and infinitely

differentiable.
Theorem 13.16 The collocation method with trigonometric polynomials
converges for integral equations of the second kind with a logarithmic sin-
gularity of the form (13.23) provided the kernel function K is 21r-periodic
and infinitely differentiable and the right-hand side is 21r-periodic and con-
tinuously differentiable.
Proof. From Theorems 11.8 and 12.15 it follows that
IIPnA<p - A<pIlHq ::; clnq-1IlA<pIlHl ::; C2nq-11l<pIlHO ::; c3 nq - 1 1!<plloo
for all <p E e[0,21r], all 1/2 < q < 1, and some constants Cll C2, and C3
depending on q. Now the proof is completed as in Theorem 13.15. 0
To evaluate the matrix elements, as usual, we replace K(tj,·) by its

trigonometric interpolation. This yields the approximations
where
Skm := ;1r 127r In (4sin2~) Lk(r)Lm(r) dr, k,m = 0, .. . ,2n -1,
(and Skm = S-k,m = Sk,-m). A practical numerical evaluation of the

weights in (13.24) rests on the observation that the quadrature rule (12.16)
with the number of grid points 2n doubled to 4n integrates trigonomet-
ric polynomials of degree less than or equal to 2n exactly, provided the
coefficient of the term sin 2nt vanishes. Therefore we have
4n-l
Skm = L R;2n) Lk (t;2n))L m (t;2n)), k, m = 0, ... , 2n - 1.
j=O
With the quadrature rule (13.24) for the evaluation of the matrix el-
ements, the system (13.22) actually corresponds to solving the equation
(13.11) with
1
(An<p)(t) := 21r Jo
r
27r ( t -
In 4sin2 -2-
r) E K(t, tm)Lm(r)<p(r) dr
2n-l
for 0 ::; t ::; 21r. Since the kernel K is assumed infinitely differentiable, from
Theorems 8.4 and 11.8, for the kernel function
2n-l
Kn(t) := L K(t, tm)Lm(r)
m=O
of An, we have IIKn - Klloo = 0 (n- m) for each m E lN, and conse-
quently IIAn - Alloo = 0 (n- m). Therefore Theorem 11.4 implies conver-
gence IIPn(An - A)lloo -+ 0, n -+ 00. Hence Corollary 13.11 and its error
estimate (13.12) applies. In the case when the exact solution 'P is infinitely
differentiable, then from Theorems 8.4 and 11.8 and (13.12) we conclude
that Ilrpn - 'PII = O(n-m) for each m E IN. If both the kernel function K
and the exact solution 'P are analytic, then from Theorem 11.7 and (13.12)
we conclude that the convergence is exponential.
2n t=O t = 7f /2 t = 7f
4 -0.10828581 0.15071347 0.22404278
K=l 8 -0.01368742 0.00827850 0.01129843
q=2 16 -0.00023817 0.00003702 0.00004366
32 -0.00000019 0.00000001 0.00000000
4 -0.19642139 0.43455120 0.66382871

K=l 8 -0.07271273 0.07400151 0.10462955
q = 1.5 16 -0.00662395 0.00275639 0.00384562
32 -0.00005937 0.00000548 0.00000609
8 0.18032618 0.02704260 -0.36058686

K=5 16 0.00617287 -0.00852966 -0.00357971
q=2 32 -0.00000686 -0.00000015 -0.00000009
64 0.00000003 0.00000002 -0.00000000
8 -0.17339263 0.01532368 0.46622097

K=5 16 -0.04130869 -0.01351387 -0.01226315
q = 1.5 32 -0.00022670 -0.00000972 -0.00001335
64 -0.00000006 0.00000000 0.00000003
Example 13.17 We apply the collocation method based on trigonometric

polynomials for the integral equation (12.24) of Example 12.14. We use the
splitting (12.25) of the kernel in a logarithmic and a smooth part, and
then evaluate the corresponding matrix elements with the quadrature rules
(13.24) and (11.33), respectively. The right-hand side of the system (13.22)
is evaluated by numerical integration, as in Example 12.14. The numerical
results in Table 13.2 show the exponential convergence, which we expect
from our error analysis. 0
In general, the implementation of the collocation method as described by

our two examples can be used in all situations where the required numerical
quadratures for the matrix elements can be carried out in closed form
for the chosen approximating subspace and collocation points. In all these
cases, of course, the quadrature formulas that are required for the related
Nystrom method will also be available. Because the approximation order for
both methods will usually be the same, Nystrom's method is preferable; it
requires the least computational effort for evaluating the matrix elements.
However, the situation changes in cases where no straightforward quadra-
ture rules for the application of Nystrom's method are available. This, in
particular, occurs for the boundary integral equations described in Chapter
6 in the case of three space dimensions. Here, the collocation method is the
most important numerical approximation method. Usually, the boundary
surface is subdivided into a finite number of segments, like curved trian-
gles and squares. Then the approximation space is chosen to consist of
some kind of low-order polynomial splines with respect to these surface el-
ements, which possess appropriate smoothness properties across the bound-
ary curves of the segments. Within each segment, depending on the degree
of freedom in the chosen splines, a number of collocation points are selected.
Then the integrals for the matrix elements in the collocation system (13.22)
are evaluated using numerical integration. Since the integral equations usu-
ally have a weakly singular kernel, calculation of the improper integrals
for the diagonal elements of the matrix, where the collocation points and
the surface elements coincide, needs special attention. With these few re-
marks we have given a somewhat vague outline of a very effective method
known as boundary element method. For a detailed description and a re-
view of the corresponding literature we refer the reader to Atkinson [9],
Brebbia, Telles, and Wrobel [18], Chen and Zhou [22], Hackbusch [60], and
Wendland [185, 186, 187]. For global numerical approximation methods for
three-dimensional boundary integral equations using spherical harmonics,
we refer to Atkinson [9] and Colton and Kress [25].
13.4 Collocation Methods for Equations

of the First Kind
We complete our study of collocation methods by indicating its applications
to integral equations of the first kind, basing the convergence and error
analysis on Theorem 13.12 and its Corollary 13.13.
Consider a singular integral equation of the first kind with Hilbert kernel
ior T
21r
1
27r {
cot r t + K(t, r) } <p(r) dr = f(t), o ~ t ~ 27r, (13.25)
as discussed in Corollary 7.25. Because of the Cauchy-type singularity, this

equation, and correspondingly its error analysis, has to be treated in the
Holder space C~~c< or in a Sobolev space HP[O, 271"J. For convenience, the 271"-
periodic kernel function K is assumed to be infinitely differentiable with
respect to both variables. We express the leading singular part of (13.25)
in the form
(R<p)(t) 1
:= 271" 10rrr{ cot --T-
r t
+ 2} <p(r) dr,
For the constant kernel function K = 2, equation (13.25) corresponds to
the Cauchy singular integral equation (7.50) for the unit circle (see also
Problem 7.2). Therefore, from our analysis in the proof of Theorem 7.30
we conclude that the singular integral operator R : C~~c< --t C~~c< is bounded
and has a bounded inverse. Analogously, from the integrals (8.27) it can be
seen that R : HP[O, 271"J --t HP[O, 21l"J is bounded and has a bounded inverse
for all p ~ 0.
Choose Xn to be the subspace of trigonometric polynomials of the form
n n-1
<p(t) = L O!j cosjt + L,8j sinjt, (13.26)
j=O j=1
and note that Xn is unisolvent with respect to the equidistantly spaced grid
tj = j7l"/n, j = 0, ... , 2n -1 (see Section 11.3). From the integrals (8.27) it
follows that Yn := R(Xn) is given by the trigonometric polynomials of the
form
n-1 n
<p(t) = L iij cosjt + L Pj sinjt. (13.27)
j=O j=1
Observe that Yn is not unisolvent with respect to the tj, j = 0, ... , 2n -1,
that, however, it is unisolvent with respect to the intersparsed set of points
t; = tj + 71" /2n, j = 0, ... , 2n - 1. Hence, we use the latter points as
collocation points.
Denote by Pn : C~;.c< --t Yn the trigonometric interpolation operator with
respect to the points t;, j = 0, ... , 2n - 1. Then, from the error estimate
(11.20), for the integral operator A given by
1
(A<p)(t):= 271" 10
r27f
K(t, r)<p(r) dr, (13.28)
we can conclude that

IIPnA - Allo,c< :::; C1 0~~7f IIPnK(.,r) - K(·,r)lIo,c<
for all m E 1N and some constants C1 and C2 depending on m. This implies

that IIPnA-Allo,c< --t 0, n --t 00. Analogously, for the numerical quadrature
operator An given by
1
:L K(t, tk)cp(tk),
2n-l
(Ancp)(t) := 2n 0:::; t :::; 211", (13.29)
k=O
it can be seen that
Hence we can apply Theorem 13.12 and its Corollary 13.13 to obtain the
following result.
Theorem 13.18 Both the semi-discrete collocation method (with trigono-
metric polynomials of the form (13.26) and collocation points j1l" /n + 11" /2n,
j = 0, ... , 2n - 1) and the corresponding fully discrete method using the
numerical quadrature operator An both converge in C~;..a for the integral
equation of the first kind (13.25) with Hilbert kernel provided K and fare
infinitely differentiable.
From the error estimates (13.15) and (13.17) and Theorems 8.4 and 11.8
we can deduce convergence of order O(n-m) for all m E 1N if the exact
solution cp is infinitely differentiable. Analogously, from Theorem 11.7 we
conclude exponential convergence when the kernel function K and the exact
solution cp are both analytic.
Note that here, as opposed to in equations of the second kind, it makes
sense to consider a fully discrete version via quadrature operators, since for
equations of the first kind there is no Nystrom method available. Of course,
a fully discrete version of the collocation method for (13.25) can also be
based on numerical evaluation of the matrix elements as described in the
previous section.
We-leave it as an exercise for the reader to prove a variant of Theorem
13.18 in the Sobolev spaces HP[O, 211"] for p > 1/2.
For the numerical implementation we need the integrals
1 r T t. Lk(r) dr,
21r
Ijk := 211" Jo cot
r-
j,k = 0, ... ,2n -1,
for the Lagrange basis (11.12). With the aid of the integrals (8.27) it can
be seen that
1 - tk - t.
Ijk =-2 {1- cosn(tk - tj)} cot _ _1 , j, k = 0, ... ,2n - 1.
n 2
Then the fully discrete collocation method for the integral equation of the
first kind (13.25) using the numerical quadrature operator An leads to the
linear system
j = 0, ... ,2n - 1,
for the coefficients 'Yo, ... , 'Y2n-l in the representation CPn = L~-:Ol 'YkLk of
the approximate solution.
Example 13.19 The integral equation (13.25) with kernel
K( ) 2 (a 2 - b2 ) sin(t + r)
t,r = - a2 +b2 -(a2 -b2 )cos(t+r)
corresponds to the Cauchy singular integral equation (7.50) for an ellipse
with semi-axis a ~ b > O. Therefore, in particular, we have uniqueness.
Table 13.3 gives some numerical results (for the difference between the
approximate and the exact solution) in the case where the exact solution
is given by
<p(t) = 1- ecost cos(sint), 0:::; t:::; 271'.
Then, setting c = (a - b)/(a + b), the right-hand side becomes
f(t) = eCcos t sin(csin t) + ecost sin (sin t), 0:::; t :::; 271',
as can be derived from (8.27) and Problem 11.4 0
2n t=O t = 71'/2 t=7I'
4 0.16740241 -0.00528236 0.00992820

a=1 8 0.00258801 -0.00022139 -0.00003859
b=0.5 16 0.00000040 -0.00000004 -0.00000001
32 0.00000000 0.00000000 0.00000000
4 0.26639845 -0.04603478 -0.00756302

a=1 8 0.06215996 -0.01726862 -0.00546249
b= 0.2 16 0.00267315 -0.00071327 -0.00022010
32 0.00000409 -0.00000109 -0.00000034
Finally, we consider an integral equation of the first kind

So<P - A<p = f (13.30)
where So and A have logarithmic singularities and are given by
(So<p)(t) 1
:= 271' Jo r
21r
{In(4sin2 t-
-2- r) }<p(r) dr
- 2
and
1
(A<p)(t):=271'Jo r 21r
{ K(t,r)ln 4sin2 ( t -r) +L(t,r) }<p(r)dr
- 2-
°
for ~ t ~ 211". Here, we assume K and L to be infinitely differentiable
and 211"-periodic with respect to both variables such that K(t, t) = for °
°
all ~ t ~ 211". For applications of integral equations with such a logarith-
mic singularity we recall the remarks in Section 12.3 on two-dimensional
boundary value problems.
By Theorem 8.22 the operator So : HP[O, 211"] -t HP+I[O, 211"] is bounded
and has a bounded inverse for all p 2: 0. From the integrals in Lemma
8.21 we conclude that So maps the space of trigonometric polynomials of
the form (13.26) into itself. Hence, here we can use the set tj = j1l"/n,
j = 0, ... , 2n -1, as collocation points.
Theorem 13.20 The semi-discrete collocation method (with trigonomet-
ric polynomials (13.26) and collocation points j1l" In, j = 0, ... , 2n - 1)
converges for the integml equation of the first kind (13.30) with logarithmic
kernel in the Sobolev space HP[O, 211"] for each p > 1/2 provided K, L, and
f are infinitely differentiable.
Proof. For the derivative of Acp we can write
! (Acp)(t) = 2~ 127r {K(t, r) In (4sin2t; r) + L(t, r)} cp(r) dr,

where
t , r-.:..)
K- (t,r ) -_ _aK-,(c...:....
at
and
- aL(t,r) t- r
L(t,r)= at +K(t,r)cot- 2-
are infinitely differentiable as a consequence of the assumption K(t, t) = 0.
Therefore, for integers p 2: 0, from Theorem 12.15 we have that
II! ACPllp+l ~ cllicplip

for all cp E HP[O, 211"] and some constant Cl depending onp. This implies that
A is bounded from HP[O, 211"] into HP+2[O, 211"] for all integers p 2: 0. Then, by
the interpolation Theorem 8.13, it follows that A : HP[O, 211"] -t HP+2[O, 211"]
is bounded for arbitrary p 2: 0.
Now, for p > 1/2, from Theorem 11.8 we conclude that
ilPnAcp - Acp//p+1 ~ C2 IiAcpilp+2 ~ C3 ilcplip

n n
for all cp E HP[O,211"] and some constants C2 and C3 depending on p. This
implies norm convergence ilPnA - All -t 0, n -t 00, for the operator
PnA - A: HP[O, 211"] -t HP+l[O, 211"], and the assertion follows from Theo-
rem 13.12. (Note that instead of Theorem 13.12, we could have used the
stability Theorem 13.7.) 0
Finally, we need to describe a fully discrete variant. For this, as in Chap-

ter 12 (see 12.19), we use the quadrature operators defined by
(13.31 )
for 0 ~ t ~ 27r. In order to apply Corollary 13.13 we prove the following

lemma, which improves on the estimates of Theorem 12.18.
Lemma 13.21 Assume that 0 ~ q ~ p and p > 1/2. Then, for the quadra-
ture operators An, we have the estimate
(13.32)
for all trigonometric polynomials cP of the form (13.26) and some constant
C depending on p.
Proof. For simplicity we only treat the case where L = O. Recalling the
proof of Theorem 12.15 we write
00
m=-oo
From this, using the property Pn{fg) = Pn{fPng) of interpolation opera-

tors, we deduce that
00
m=-oo
By Theorem 11.8, the operators Pn : Hq+l [0, 27r] --+ Hq+l [0, 27r] are uni-
formly bounded. Therefore, using Corollary 8.8 and the property (12.28)
of the Fourier coefficients am and denoting by k the smallest integer larger
than or equal to q, we can estimate
1
L
00
II(PnAn-PnA)CPllq+l ~ Cl (1 + Iml)3+k IlPnAo[Pn{fmCP)- fmcp]lIq+l

m=-oo
for all cP E Xn and some constant Cl.

In the case where Iml ~ n/2 we can estimate
for all cP E Xn and some constants C2 and C3. From this we conclude that
" 1
L...J (1 + Iml)3+k C4
IlPnAo[Pn{fmCP) - fmcplllq+l ~ -; IIcpllq (13.33)
Iml~n/2
for all r.p E Xn and some constant C4. For 0 ~ m ~ n/2, writing
2n
r.p = Lan-jfn-j
j=O
in view of Pnfm+n-j = Pnfm-j-n, j = 0, ... , m, and Pnfm+n-j = fm+n-j,

j = m + 1, ... , 2n, we deduce that
m
Pn(fmr.p) - fmr.p = Lan-j[Pnfm-j-n - fm-Hn]'
j=O
From this, with the aid of the integrals in Lemma 8.21, we find that
and, using the inequalities
1 1 8m
-----, < - (n + j - m)2 ~ n 2 ~ 4(n _ j)2
n-m+j n+m-j - 3n 2 '
for 0 ~ j ~ m ~ n/2, we can estimate
for all r.p E X n , all 0 ~ m ~ n/2, and some constants C5 and C6. From this
and the corresponding inequality for -n/2 ~ m ~ 0 we deduce that
for all r.p E Xn and some constant C7. The last inequality, together with
(13.33), yields
for all r.p E X n . (The constants Cl, ••• , Cs all depend only on q.) From this
the statement of the lemma follows from 11r.pl/q ~ cnq-PI/r.pl/p for all r.p E Xn
and some constant c depending on p and q. 0
Theorem 13.22 The fully discrete collocation method using the quadra-
ture operators (13.31) converges for the integral equation of the first kind
(13.30) with logarithmic kernel in the Sobolev space HP[0,27r] for each
p> 1/2 provided K, L, and f are infinitely differentiable.
Proof. From Theorems 11.8 and 12.18, we conclude uniform boundedness

of the operators PnAn - PnA : HP[0,27f] -+ HP+l[0,27f], and pointwise
convergence (PnAn -PnA)cp -+ 0, n -+ 00, for all trigonometric polynomials
cpo By the Banach-Steinhaus theorem this implies pointwise convergence for
all '¢ E HP[O, 27f]. From Lemma 13.21 we have that
C
II(PnAn - PnA)CPllp+l ~ - Ilcplip,
n
for all cp E Xn = Sol(Yn). Now the statement follows from Corollary 13.13
in connection with Theorem 13.20. (We note that instead of Corollary 13.13
we may use Theorem 13.8.) 0
From the error estimate (13.17) and Theorems 8.4 and 11.8 we have con-
vergence of order O( n -m) for all m E 1N if the exact solution cp is infinitely
differentiable. From Theorem 11. 7 we obtain exponential convergence when
the kernel functions K and L and the exact solution cp are analytic.
The above convergence and error analysis for the fully discrete trigono-
metric polynomial collocation for integral equations of the second kind
with a logarithmic singularity follows Saranen and Vainikko [159]. It dif-
fers from an earlier approach by Kress and Sloan [109] and Prossdorf and
Saranen [148], because it does not require a further decomposition of the
logarithmic singularity, which complicates the numerical implementation
of the fully discrete scheme.
Since the quadrature operators (13.31) are exact for constant kernel func-
tions K and L and trigonometric polynomials 'P E X n , the fully discrete
collocation method for (13.30) leads to the linear system
~1{ R j _k[I-K(tj, tk)]- 2~ [L(tj, t k )+2]} 'Yk = l(tj), j = 0, ... ,2n-1,
J:
lor t h e coeffi Clents
. . -
III CPn L..Jk=O 'Yk L k·
= ~2n-l
Example 13.23 The integral equation (13.30) with kernel K(t,T) 0
and
L(t, T) = -In{a + b2 2 - (a 2 - b2 ) cos(t + Tn - 3
essentially corresponds to the integral equation (7.46) for an ellipse with
semi-axis 0 < b :S a :S 1. Therefore, in particular, it is uniquely solvable.
Table 13.4 gives some numerical results for the case in which the exact
solution is given by
cp(t) = ecos t {costcos(sint) - sintsin(sintn, O:S t ~ 27f.
Then the right-hand side becomes
f(t) = 2 - ecos t cos(sint) - eC cos t cos(csint), 0 ~ t ~ 27f,
with c = (a - b)j(a + b). 0
2n t=O t = 7r/2 t = 7r
4 -0.62927694 -0.19858702 0.20711238
a=1 8 -0.02649085 0.00602690 0.01355414
b = 0.5 16 -0.00000724 0.00000110 0.00000457
32 0.00000000 0.00000000 0.00000000
4 -0.77913195 -0.22694775 0.20990539

a=1 8 -0.08896065 0.02823001 0.03008962
b = 0.2 16 -0.00167596 0.00010607 0.00000226
32 -0.00000114 0.00000009 -0.00000001
13.5 The Galerkin Method

For operator equations in Hilbert spaces the projection method via or-
thogonal projection into finite-dimensional subspaces leads to the Galerkin
method, named after the Russian engineer Galerkin. Let X and Y be Hilbert
spaces and let A : X -+ Y be an injective bounded linear operator. Let
Xn C X and Yn C Y be subspaces with dimXn = dim Yn = n, and let
P n : Y -+ Yn be the orthogonal projection operator described in Theorem
13.3. Then <Pn E Xn is a solution of the projection method, generated by
Xn and Pn , for the equation A<p = f if and only if
(A<pn,g) = (f,g), g E Yn. (13.34)
This follows immediately from the fact that (13.34), by Theorem 1.25, is
equivalent to Pn(A<Pn - f) = O. The equation (13.34) is called the Galerkin
equation.
In the literature, the Galerkin method is also known as the Petrov-
Galerkin method and the special case where X = Y and Xn = Y n is also
called the Bubnov-Galerkin method. When the operator A is self-adjoint
and positive definite, as we will briefly explain, the Bubnov-Galerkin
method coincides with the Rayleigh-Ritz method. A bounded linear op-
erator A : X -+ X is called self-adjoint if A = A *; it is called positive
definite if it is self-adjoint and satisfies
(A<p, <p) >0 (13.35)
for all <p E X with <p ::f. O. A positive definite operator clearly is injective.
We can define an additional scalar product on X by
(13.36)
with the corresponding norm II· liE called the energy norm. Consider the
so-called energy functional, defined on X by
E(cp):= (Acp,cp) - 2Re(j,cp) (13.37)
for all cp E X with f E X a given element. For f E A(X) we can transform
E(cp) - E(A- I f) = IIcp - A-I fll~.
Hence, solving the equation Acp = f is equivalent to minimizing the energy
functional E over X. Based on the classical work of Rayleigh [152] from
1896 and Ritz [154] from 1908, the Rayleigh-Ritz method consists of an
approximation by minimizing E only over a finite-dimensional subspace
X n . This, obviously, is equivalent to finding a best approximation to the
exact solution with respect to Xn in the energy norm. By Theorems 1.25
and 1.26, the best approximation exists and is uniquely determined by the
orthogonality condition
for all 9 E X n . But this coincides with the Galerkin equation (13.34), i.e.,
the Rayleigh-Ritz method is a special case of the Galerkin method. Note
that we have convergence of the Rayleigh-Ritz method_ if the denseness
condition (13.3) is satisfied with respect to the energy norm.
First Bubnov in 1913 and then, in more details, Galerkin [47] in 1915
approached and extended this approximation method without relying on a
minimization formulation. Later, Petrov [141] first considered the general
situation of the form (13.34) of the Galerkin method.
Assume that Xn = span{uI, ... ,Un } and Yn = span{vI, ... ,Vn }. Then
we express CPn as a linear combination CPn = L:~=I 7kUk and find that
solving the equation (13.34) is equivalent to the linear system
n
L7k(Auk,Vj) = (j,Vj), j = 1, ... ,n, (13.38)
k=1
for the coefficients 71> ... , 7n. Usually, the scalar product will have the
form of an integral. Therefore, in the case of integral equations, the system
(13.38) requires a double integration for each of the matrix elements. This
is a disadvantage of the Galerkin method, making it considerably inferior
in numerical efficiency compared with the collocation method. Its major
advantage is based on the simplicity of the orthogonal projection and the
exploitation of the Hilbert space structure. In particular, for integral equa-
tions of the second kind, Corollary 13.9 can always be applied, because the
denseness assumption (13.3) implies pointwise convergence of the orthogo-
nal projection operators. Hence, we can state the following theorem.
Theorem 13.24 Let A : X -7 X be a compact linear operator and assume
that I - A is injective. Then the Bubnov-Galerkin method converges.
In contrast to our previous methods, the Galerkin method applies more

easily to equations of the first kind. We proceed with a few general results.
Definition 13.25 A bounded linear operator A : X -+ X in a Hilbert space

X is called strictly coercive if there exists a constant c > 0 such that
(13.39)
for all 'P E X.
Theorem 13.26 (Lax-Milgram) In a Hilbert space X a strictly coercive

bounded linear operator A : X -+ X has a bounded inverse A-I : X -+ X.
Proof. Using the Cauchy-Schwarz inequality, we can estimate
Hence
(13.40)
for all 'P EX. From (13.40) we observe that A'P = 0 implies 'P = 0, i.e., A
is injective.
Next we show that the range A(X) is closed. Let '¢ E A(X) and let
('¢n) be a sequence from A(X) with '¢n -+ ,¢, n -+ 00. Then we can write
'¢n = A'Pn with some 'Pn EX, and from (13.40) we find that
for all n, mE IN. Therefore, ('Pn) is a Cauchy sequence in X and converges:

'Pn -+ 'P, n -+ 00, with some 'P E X. Then,¢ = A'P, since A is continuous
and A(X) = A(X) is proven.
Knowing now that A(X) is complete, denote by P : X -+ A(X) the
orthogonal projection operator. Let f E X be arbitrary. Then by Theorem
1.25, we have Pf - f 1.. A(X). In particular, (Pf - f, A(Pf - f)) = O.
Hence, from (13.39) we see that f = P f E A(X). Therefore, A is surjective.
Finally, the boundedness of the inverse IIA- 1 11 :::; lie is a consequence of
(13.40). 0
Theorem 13.27 Let A : X -+ X be a strictly coercive operator. Then the

Bubnov-Galerkin method converges.
Proof. As in the previous proof we can estimate
for all 'P E X n , since orthogonal projections are self-adjoint (see Problem
13.1). Hence,
(13.41 )
for all cP E X n . This implies that An = PnA : Xn -+ Xn is injective and,

consequently, surjective. For all cP E X, using (13.41) and Theorem 13.3,
we can estimate
Therefore, IIA;:;-l PnAl1 ~ IIAli/c for all n E lN, and the statement follows
from Theorem 13.6. 0
Theorems 13.26 and 13.27 can be extended to strictly coercive operators

mapping a Hilbert space X into its dual space X* (see Problem 13.4).
We conclude our presentation of the Galerkin method by describing one
more special case.
Theorem 13.28 Let X and Y be Hilbert spaces and let A : X -+ Y be

an injective bounded linear operator. Let Xn C X be a finite-dimensional
subspace. Then lor each lEY there exists a unique element CPn E Xn such
that
IIACPn - III = inf IIA'¢ - III·
1/JEXn
It is called the least squares solution 01 Acp = I with respect to Xn and

it coincides with the Petrov-Galerkin solution lor the subspaces Xn and
Y n := A(Xn).
Proof. Obviously, CPn is a least squares solution of Acp = I with respect to

Xn if and only if ACPn is a best approximation of I with respect to Y n . By
Theorem 1.26, the best approximation exists and is unique. The injectivity
of A then implies uniqueness for the least squares solution. By Theorem
1.25, the best approximation ACPn is characterized by the orthogonality
condition (ACPn - I,g) = 0 for all g E Yn . But this in turn is equivalent to
the Galerkin equation (13.34) for the subspaces Xn and Y n . 0
Note that the numerical implementation of the least squares solution in

the case of integral equations requires a triple integration for each matrix
element and a double integration for each right-hand side of the linear sys-
tem (13.38), since gj = AUj requires an additional integration. Therefore,
in general, it cannot compete in efficiency with other methods.
We now outline the use of the Galerkin method for integral equations.
We start with the Bubnov-Galerkin method for equations of the second
kind
cp(x) - fa K(x,y)cp(y) dy = I(x), x E G,
with continuous or weakly singular kernel K in the Hilbert space L2(G).

Let Xn = span{ul, ... , Un} be an n-dimensional subspace of L2(G). Then
we write cP = ~~=l 'YkUk and the Bubnov-Galerkin equations assume the
form
(13.42)
= kf(x)Uj(X)dX, j=l, ... ,n,
for the coefficients 'Y1, ... , 'Yn. For this system, the first term in the ex-
pression for the matrix elements can usually be evaluated in closed form,
whereas the second term needs a double numerical integration. Before we
proceed, we wish to point out that the matrix of the system (13.42) coin-
cides with the system (11.36) obtained by the degenerate kernel method
via orthogonal expansion.
Without entering too deeply into the numerical implementation, we con-
sider as an example the Galerkin method using linear splines. If we pursue
the same idea as in the collocation method and use interpolatory quadra-
tures with respect to both variables, it turns out that we solve the same
approximate finite-dimensional equation on the space of spline functions as
in the degenerate kernel and in the collocation method via spline interpola-
tion. Therefore the numerical results will be the same. Only the method of
solution for this equation is different: In the collocation method we equate
spline functions by requiring them to coincide at the interpolation grid; in
the Galerkin method we require equality of scalar products with respect to
a basis. In general, we may say that most of the fully discrete implemen-
tations of Galerkin methods may be interpreted as implementations of a
related collocation method.
For integral equations for periodic functions using trigonometric polyno-
mials we note that instead of setting up the Bubnov-Galerkin method in
L2[0,27r] we may use the Sobolev spaces HP[0,27r] with p ;::: O. Since the
trigonometric monomials are orthogonal in each of these Sobolev spaces,
the orthogonal projection onto trigonometric polynomials is always given
through truncation of the Fourier series. Therefore, the Bubnov-Galerkin
equations are the same for all p ;::: O. In particular, this means that we
automatically have convergence with respect to all Sobolev norms in the
case of infinitely differentiable kernels and right-hand sides.
In principle, for the Galerkin method for equations of the second kind
the same remarks as for the collocation method apply. As long as numerical
quadratures are available, in general, the Galerkin method cannot compete
in efficiency with the Nystrom method. Compared with the collocation
method, it is less efficient, since its matrix elements require double integra-
tions. Therefore, in practical problems, the collocation method is the most
widely used projection method for solving integral equations of the second
kind, despite the fact that its error analysis is often less satisfactory than
the error analysis for the Galerkin method.
Problems 245
We conclude with an indication of the Petrov-Galerkin method for equa-

tions of the form (13.25) and (13.30). From the previous section we recall
that for both equations the leading parts R : HP[O,21f) -7 HP[O,21f) and
So : HP[O, 21f] -7 HP+1 [0, 21f] are bijective with a bounded inverse for each
p~ O.
For the first equation, as subspaces Xn and Yn of HP[O, 21f], we choose
trigonometric polynomials of the form (13.26) and (13.27), respectively.
Because R maps Xn bijectively onto Y n , the Petrov-Galerkin equation
PnR'Pn = Pnf
has a unique solution 'Pn E Xn given by 'Pn = R- 1 Pnf. Therefore, the
Petrov-Galerkin method for R converges, and by the stability Theorem
13.7 it also converges for equation (13.25). The equation (13.30) is treated
analogously, using as subspaces Xn of HP[O,21f) and Yn of HP+l [0, 21f]
trigonometric polynomials of the form (13.26).
We summarize these observations in the following theorem.
Theorem 13.29 The Petrov-Galerkin method using trigonometric poly-
nomials converges in HP[O,21f) for equations of the first kind with Hilbert
or logarithmic kernels.
Problems
13.1 Show that orthogonal projection operators are self-adjoint.
13.2 Let A : X -+ X be a bounded positive self-adjoint operator in a Hilbert
space X. Choose Uo E X and define Uj = AUj-1 for j = 1, ... , n - 1. Show
that the Bubnov-Galerkin equations for Acp = f with respect to the so-called
Krylov subspaces Xn = span{uo, ... ,Un-I} are uniquely solvable for each n E IN.
Moreover, if f is in the closure of span{Ajuo : j = 0,1, ... }, then the Bubnov-
Galerkin approximation cpn converges to the solution of Acp = f.
Show that in the special case Uo = f the approximations CPn can be computed
iteratively by the formulas cpo = 0, po = f and
cpn+1 = cpn - anPn,
Pn = Tn + !3n-IPn-l,
Tn = Tn-l - an-IApn-l,
an-I = (Tn-l,Pn-I)/(Apn-I,Pn-I),
!3n-1 = -(Tn' Apn-d/(Apn-I,Pn-t).
Here rn is the residual Tn = ACPn - f. This is the conjugate gradient method of
Hestenes and Stiefel [69].
13.3 Under the assumptions of Theorem 13.24 let cpn be the Bubnov-Galerkin
solution. Consider the iterated Bubnov-Galerkin solution 'Pn, defined by
'Pn := Acpn +f
and show that

lIiPn - cpll ~ cnllCPn - cpll
where Cn -t 0, n -t 00. This more rapid convergence of the iterated Bubnov-
Galerkin solution is called superconvergence (see Sloan (164)).
Hint: Show first that iPn - APniPn = J, then that IIAPn - All -t 0, n -t 00, and
finally that iPn - cP = (I - APn)-l(APn - A)(cp - Pncp).
13.4 A bounded linear operator A : X -t X* mapping a Hilbert space X into

its dual space X* is called strictly coercive if there exists a constant c > 0 such
that
Re(Acp)(cp) :::: cllcpll2
for all cp E X. Formulate and prove extensions of Theorems 13.26 and 13.27.
13.5 Apply the results of Problem 13.4 to the integral equation (13.30) in the
Sobolev space H-l/2[O, 27r].
14
Iterative Solution and Stability
The approximation methods for integral equations described in Chapters

11-13 lead to full linear systems. Only if the number of unknowns is reason-
ably small may these equations be solved by direct methods like Gaussian
elimination. But, in general, a satisfying accuracy of the approximate so-
lution to the integral equation will require a comparatively large number
of unknowns, in particular for integral equations in more than one dimen-
sion. Therefore iterative methods for the resulting linear systems will be
preferable. For this, in principle, in the case of positive definite symmet-
ric matrices the classical conjugate gradient method (see Problem 13.2)
can be used. In the general case, when the matrix is not symmetric more
general Krylov subspace iterations may be used among which a method
called generalized minimum residual method (GMRES) due to Saad and
Schultz [158] is widely used. Since there is a large literature on these and
other general iteration methods for large linear systems (see Freud, Golub,
and Nachtigal [43], Golub and van Loan [52], Greenbaum [53], Saad [157],
and Trefethen and Bau [175], among others), we do not intend to present
them in this book. At the end of this chapter we will only briefly de-
scribe the main idea of the panel clustering methods and the fast multipole
methods based on iterative methods and on a speed-up of matrix-vector
multiplications for the matrices arising from the discretization of integral
equations.
Instead of describing general iterative methods for linear systems, how-
ever, we will discuss two-grid and multigrid iterations that are especially
suitable for solving the linear systems arising from the numerical solution
of integral equations of the second kind.

248 14. Iterative Solution and Stability
14.1 Stability of Linear Systems

We will start this chapter by briefly considering the question of stability of
the linear systems arising in the discretization of integral equations. Usu-
ally, in the course of any numerical procedure for solving linear systems,
errors will be introduced. For example, in the collocation and Galerkin
method, errors occur through the numerical approximations of the matrix
elements and the right-hand side of the linear systems. A linear system is
called stable if small changes in the data of the system cause only small
changes in the solution. We will make this notion more precise by intro-
ducing the concept of a condition number.
Definition 14.1 Let X and Y be normed spaces and let A: X -t Y be a

bounded linear operator with a bounded inverse A-I: Y -t X. Then
cond(A) := IIAIIIIA- 111
is called the condition number of A.
Since 1 = 11111 = IIAA- 111 :::; IIAIIIIA- 111, we always have cond(A) ~ l.
The following theorem shows that the condition number may serve as a
measure for stability.
Theorem 14.2 Let X and Y be Banach spaces, let A : X -t Y be a
bounded linear operator with a bounded inverse A-I : Y -t X, and let
AO : X -t Y be a bounded linear operator such that IIA- 1111IAo - All < l.
Assume that 'P and 'Po are solutions of the equations
A'P =f (14.1)
and
(14.2)
respectively. Then
< cond(A) {lifO - fll IIAo - All}

(14.3)
1_ d(A) IIAo - All Ilfll + IIAII .
con IIAII
Proof. Writing AO = A[I + A- 1(AO - A)], by Theorem 10.1 we observe

that the inverse operator (AO)-1 = [I + A-I (AO - A)]-1 A -1 exists and is
bounded by
IIA- 111
II(AO)-111 :::; 1-IIA-1111IAo _ All. (14.4)
From (14.1) and (14.2) we find that
AO('P0 - 'P) = fO - f - (AO - A)'P,

14.1 Stability of Linear Systems 249
whence
vi - <P = (A5)-l{/5 - I - (A5 - A)<p}
follows. Now we can estimate
and insert (14.4) to obtain
Ilcl- <p1I cond(A) 11/5- III IIA5 - All }

{
11<p1I ~ l-IiA- l IIIIA5 - All IIAIIII<p1I + II All .
From this the assertion follows with the aid of IIAIIII<p1I ~ 11/11. 0
For finite-dimensional approximations of a given operator equation we

have to distinguish three condition numbers: namely, the condition numbers
of the original operator, of the approximating operator as mappings from
the Banach space X into the Banach space Y, and of the finite-dimensional
system that has to be set up for the actual numerical solution. This lat-
ter system we can influence, for example, in the collocation and Galerkin
methods by the choice of the basis for the approximating subspaces.
Consider an equation of the second kind <P - A<p = I in a Banach space
X and approximating equations <Pn - An<Pn = In under the assumptions
of Theorem 10.1, i.e., norm convergence, or Theorem 10.9, i.e., collective
compactness and pointwise convergence. Then from Theorems 10.1, 10.4,
and 10.9 it follows that the condition numbers cond(! - An) are uniformly
bounded. Hence, for the condition of the approximating scheme, we mainly
have to be concerned with the condition of the linear system for the actual
computation of the solution of <Pn - An<Pn = In.
For the discussion of the condition number for the Nystrom method we
recall the notations of Section 12.2. We have to relate the condition number
for the numerical quadrature operator
n
(An<P)(x) := :L>l(kK(X,Xk)<P(Xk), x E G,
k=l
in the Banach space C (G) to the condition number for the matrix operator
An : ~n -+ ~n given by
n
(An<P)j:= LCYkK(Xj,Xk)<Pk, j = 1, . .. ,n,
k=l
for <P = (<pI, ... , <pn) E ~n. We choose functions L j E C(G) with the
properties IILjlloo = 1 and Lj(xj) = 1 for j = 1, ... ,n, such that LjLk = 0
for j =I k. Then we introduce linear operators Rn : C( G) -+ ~n and
Mn : ~n -+ C(G) by
Rnl := (I(Xl), ... , I(x n », IE C(G),

and
n
Mnip := L ipjL j , ip = (ipr, ... , ipn) E (Cn,
j=1
and have IIRniioo = IIMniioo = 1 (see Problem 14.1). From Theorem 12.7
we conclude that
and
From these relations we immediately obtain the following theorem.
Theorem 14.3 For the Nystrom method the condition numbers for the
linear system are uniformly bounded.
This theorem states that the Nystrom method essentially preserves the
stability of the original integral equation.
For the collocation method we recall Section 13.3 and introduce operators
En, An : (Cn --+ (Cn by
n
(Enip)j:= LUk(Xj)ipk, j = 1, ... ,n,
k=1
and
n
(Anip)j := L(Auk)(Xj)ipk, j = 1, ... ,n,
k=1
for ip = (ipr, ... ,ipn) E (Cn. Since Xn = span{ul, ... ,Un } is assumed to
be unisolvent, the operator En is invertible. In addition, let the operator
Wn : (Cn --+ C(G) be defined by
n
W nl := LlkUk
k=1
for I = (,1, ... , 'Yn) and recall the operators Rn and Mn from above. Then
we have Wn = PnMnEn. From (13.19) and (13.21) we can conclude that
and
(En - An)-1 = E;;1 Rn(I - P n A)-1 PnMn.
From these three relations and the fact that by Theorems 10.4 and 13.10
the sequence of operators (1 - Pn A)-1 Pn is uniformly bounded, we obtain
the following theorem.
14.2 Two-Grid Methods 251
Theorem 14.4 Under the assumptions of Theorem 13.10, for the colloca-
tion method the condition number of the linear system satisfies

This theorem suggests that the basis functions must be chosen with cau-
tion. For a poor choice, like monomials, the condition number of En can
grow quite rapidly. However, for the Lagrange basis, En is the identity
matrix with condition number one. In addition, IlFnll enters in the esti-
mate on the condition number of the linear system, and, for example, for
polynomial or trigonometric polynomial interpolation we have IlFnll ~ 00,
n ~ 00 (see the estimate (11.15)). A discussion of the condition number
for the Galerkin method can be carried out quite similarly (see Problem
14.2).
Summarizing, we can state that the Nystrom method is generically stable
whereas the collocation and Galerkin methods may suffer from instabilities
due to a poor choice of the basis for the approximating subspaces.
14.2 Two-Grid Methods

Let X be a Banach space and let A : X ~ X be a bounded linear operator
such that I - A has a bounded inverse. As described in previous chapters
we replace the operator equation of the second kind
'P - A'P = f (14.5)
by a sequence of approximating equations
'Pn - An'Pn = fn (14.6)
leading to finite-dimensional linear systems. We assume the sequence (An)

of bounded linear operators An : X ~ X to be either norm convergent or
collectively compact and pointwise convergent such that either Theorem
10.1 or Theorem 10.9 may be applied to yield existence and uniqueness of
a solution to the approximating equation (14.6) for all sufficiently large n.
Our analysis includes projection methods in the setting of Theorem 13.10.
We note that the index n indicates different levels of discretization, i.e.,
different numbers of quadrature points in the Nystrom method, i)r different
dimensions of the approximating subspaces in the degenerate kernel or
projection methods. Referring to the quadrature and collocation methods,
we will use the term grid instead of level. Deviating from the notation in
preceding chapters, here the number of unknowns in the linear system on
the level n will be denoted by Zn and will be different from n. Frequently, in
particular for one-dimensional integral equations, the number of unknowns

Zn on the level n will double the number of unknowns Zn-l on the preceding
level n - 1 as in all our numerical tables in Chapters 11-13.
Assume that we already have an approximate solution 'Pn,O of (14.6) with
a residual
rn := in - (I - An)'Pn,o.
Then we try to improve on the accuracy by writing
'Pn,l = 'Pn,O + on- (14.7)
For 'Pn,l to solve equation (14.6) the correction term on has to satisfy the
residual correction equation
We observe that the correction term on, in general, will be small compared
with 'Pn,O, and therefore it is unnecessary to solve the residual correction
equation exactly. Hence we write
where the bounded linear operator Bn is some approximation for the inverse
(1 - An)-l of 1 - An. Plugging this into (14.7) we obtain
'Pn,l = [1 - Bn(I - An»)'Pn,O + Bnin

as our new approximate solution to (14.6). Repeating this procedure yields
the defect correction iteration defined by
(14.8)
for the solution of (14.6). By Theorem 2.10, the iteration (14.8) converges
to the unique solution 'ljJn of Bn(1 - An)'ljJn = Bnin provided
or by Theorem 10.13, if the spectral radius of 1 - Bn(I - An) is less than

one. Since the unique solution 'Pn of 'Pn - An'Pn = fn trivially satisfies
Bn(I -An)'Pn = Bnfn' if the iteration scheme (14.8) converges it converges
to the unique solution of (14.6). For a rapid convergence it is desirable
that the norm or spectral radius be close to zero. For a more complete
introduction into this residual correction principle, we refer to Stetter (168).
Keeping in mind that the operators An approximate the operator A,
an obvious choice for an approximate inverse Bn is given by the correct
inverse Bn = (1 - Am)-l of 1 - Am for some level m < n corresponding to
a coarser discretization. For this so-called two-grid method, we can simplify
(14.8) using
1 - (1 - Am)-1(1 - An) = (1 - Am)-l(An - Am).

14.2 Two-Grid Methods 253
Following Brakhage [17] and Atkinson [8], we will consider the two special
cases where we use either the preceding level m = n - 1 or the coarsest
level m = 0.
Theorem 14.5 Assume that the sequence of operators An : X -+ X is
either norm convergent IIAn - All -+ 0, n -+ 00, or collectively compact
and pointwise convergent An<P -+ A<p, n -+ 00, for all <p E X. Then the
two-grid iteration
using two consecutive grids converges, provided n is sufficiently large. The

two-grid iteration
<Pn,i+l := (I - AO)-l{ (An - Ao)<Pn,i + fn}, i = 0,1,2, ... , (14.10)
using a fine and a coarse grid converges, provided the approximation Ao is

already sufficiently close to A.
Proof. Let (An) be norm convergent. Then from the estimate (10.1) we
observe that II (I - An) -111 ::; C for all sufficiently large n with some constant
C. Then the statement on the scheme (14.9) follows from
11(1 - An_l)-l(An - An-l)1I ::; CllAn - An-III -+ 0, n -+ 00. (14.11)
For the scheme (14.10) we assume that the coarsest grid is chosen such that
IIAn - Aoll < 1/2C for all n ~ 0. Then 11(1 - Ao)-l(An - Ao)11 < 1 for all
n E IN. Note that
11(1 - AO)-l(An- Ao)1I -+ II(I - Ao)-l(A - Ao)1I > 0, n -+ 00. (14.12)
Now assume that the sequence (An) is collectively compact and pointwise
convergent. Then the limit operator A is also compact and, by Theorem
10.9, the inverse operators (I - An)-l exist and are uniformly bounded for
sufficiently large n. Hence the sequence (An) defined by
is collectively compact. From the pointwise convergence An<P- An-l<P -+ 0,

n -+ 00, for all <P E X, by Theorem 10.7, we conclude that
(14.13)
Again by Theorem 10.7, we may choose the coarsest grid such that
1
II(I - AO)-l(Am - A)(I - AO)-l(An - Ao)1I < "2
for all m, n ~ 0. This implies
(14.14)
for all n E IN. Now the statement follows from (14.13) and (14.14) and the
variant of Theorem 2.10 discussed in Problem 2.2. 0
Note that, in view of Theorem 12.8, for the Nystrom method we cannot
expect that II(I - An_I)-I(An - An-I)1I -70, n -7 00, instead of (14.13)
or that II (I - Ao)-I(An - Ao)11 < 1 instead of (14.14). Based on this obser-
vation, Brakhage [17] suggested a further variant of the defect correction
iteration (see Problem 14.3).
Comparing the two variants described in Theorem 14.5, we observe from
(14.11) to (14.14) that, in general, we have a more rapid convergence in
the first case. Actually the convergence rate for (14.9) tends to zero when
n -7 00. But for a value judgement we also have to take into account the
computational effort for each iteration step. For the actual computation we
write the defect correction equation in the form
(14.15)
indicating that, given t.pn,i, we have to solve a linear system for the unknown
CPn,i+!' From this we observe that for the evaluation of the right-hand side
in (14.15) we need to multiply a Zn x Zn matrix with a Zn vector requiring,
in general, O(z;) operations. Then solving (14.15) for CPn,HI means directly
solving a linear system on the mth level with Zm unknowns and, in general,
requires O(z~) operations. Therefore, for the second variant, working with
the coarsest grid we may neglect the computational effort for the solution of
the linear system and remain with O(z;) operations for each iteration step.
However, for the first variant, the effort for the solution of the linear system
dominates: Each iteration step needs O(Z!_I) operations. In particular,
when the discretization is set up such that the number of unknowns is
doubled when we proceed from the level n - 1 to the next level n, then
the computational effort for each iteration step in (14.9) is roughly 1/8th
of the effort for solving the linear system on the level n directly. Hence, we
have rapid convergence but each iteration step is still very costly.
We indicate th~ numerical implementation of the defect correction iter-
ation for the Nystrom method using the notations introduced in Chapter
12. Here the approximate operators are given by
L Qin)K(x, xin))cp(xin)),
Zn
(Ant.p)(x) = x E G.
k=l
Each iteration step first requires the evaluation of the right-hand side
of (14.15) at the Zm quadrature points xJm ) , j = 1, ... ,Zm, on the level m

and at the Zn quadrature points xJn), j = 1, ... , Zn, on the level n. The
14.3 Multigrid Methods 255
corresponding computations
Zn
9 n,t_(x~m))
J
= f n (x~m))
J
+ "'"'
L..J o:(n)
k
K(x~m) X(n))lfl _(x(n))
J' k yn,t k
k=l
Zm
_ ",",o:(m)K(x~m) X(m))lfl _(x(m))
L..J k J' k yn,t k , J. = 1, ... , Zm,
k=l
and
9n,t_(x~n))
J
= f n (x~n))
J
+"'"'
Zn
L..J o:(n)
k
K(x~n) X(n))lfl _(x(n))
J' k yn,t k
k=l
Zm
- LO:km) K(xjn) , Xkm))'Pn,i(Xkm )), j = 1, ... , Zn,

k=l
require O(z;') operations. Then, according to Theorem 12.7, we have to
solve the linear system
Zm
(m)) "'"' (m)K( (m) (m)) ( (m))
'Pn,Hl ( Xj - L..J O:k xj ,xk 'Pn,Hl x k = gn,i ( Xj(m))
k=l
for the values 'Pn,Hl (xjm)) at the Zm quadrature points xjm) , j = 1, ... , Zm.
The direct solution of this system needs O(z~) operations. Finally, the
values at the Zn quadrature points xjn), j = 1, ... , Zn, are obtained from
the Nystrom interpolation
Zm
'Pn,i+1(xjn)) = L O:km ) K(xjn) , Xkm ))'Pn,Hl (xim )) + gn,i(xjn))

k=l
for j = 1, ... , Zn, requiring O(znzm) operations. All together, in agreement
with our previous operation count, we need O(z;')+O(z~) operations. It is
left as an exercise to set up the corresponding equations for the degenerate
kernel and the collocation method (see Problem 14.4).
14.3 Multigrid Methods

Multigrid methods have been developed as a very efficient iteration scheme
for solving the sparse linear systems that arise from finite difference or
finite element discretization of elliptic boundary value problems. Following
Hackbusch [59) and Schippers (161) we now briefly sketch how multigrid
methods can also be applied for integral equations of the second kind.
The two-grid methods described in Theorem 14.5 use only two levels; the
multigrid methods use n + 1 levels.
Definition 14.6 The multigrid iteration is a defect correction iteration of

the form (14.8) with the approximate inverses defined recursively by
Bo := (I - AO)-I,
p-I (14.16)
Bn := L [I - B n- I (I - An_I»)m Bn-I, n = 1,2, ... ,
m=O
with some p E IN.
Apparently, Bo stands for the exact solution on the coarsest grid. The
approximate inverse Bn for n + 1 levels represents the application of p
steps of the multigrid iteration B n - I on n levels for approximately solving
the residual correction equation starting with initial element zero, since it
is given by the pth partial sum of the Neumann series. This observation,
simultaneously, is the motivation for the definition of the multigrid method
and the basis for its actual recursive numerical implementation. In practical
calculations p = 1,2, or 3 are appropriate values for the iteration number.
It is our aim to illustrate that the multigrid iteration combines the advan-
tages of the two-grid methods of Theorem 14.5. It has the fast convergence
rate of the scheme (14.9) and the computational effort for one step is es-
sentially of the same order as for one step of the scheme (14.10). First we
note that the approximate inverses satisfy
(14.17)
for all n E IN. For the iteration operator Mn corresponding to one step of
the multigrid iteration using n + 1 levels, from Definition 14.6 and from
(14.17), we deduce that
Mn = 1- Bn(1 - An) = 1- Bn{I - An-I)(1 - An_I)-I(1 - An}
= I - (I -An_I)-I(1 -An)+{1 -Bn{I -An-I)}{I -An_I)-I(1 -An)
After introducing the two-grid iteration operator
(14.18)
we can write the recursion for the iteration operator Mn in the form
(14.19)
Theorem 14.7 Assume that
(14.20)
for all n E IN and some constants q E (0,1] and C > 0 satisfying
C5, 21q (J1+Q2-1). (14.21 )
Then, if p ~ 2, we have
Proof. Note that J1 + q2 - 1 < q for all q > O. Therefore the statement is
correct for n = 1. Assume that it is proven for some n E IN. Then, using
the recurrence relation (14.19), we can conclude that
5, qnC + (2qn-lC)2(1 + qnc)
5, qnC { 1 + t C(1 + qC) } 5, 2qnc < 1,
since 4t(1 + qt) 5, q for all 0 5, t 5, ( J1+Q2 - 1) /2q. o

In particular, setting q = 1, we obtain that
sup IITnll 5, -21 (V2 - 1) = 0.207 ...

nEIN
implies that liMn II 5, V2 - 1 < 1 for all n E IN and all p ~ 2. Roughly

speaking, this ensures convergence of the multigrid method, provided the
approximation on the coarsest level is sufficiently accurate.
More generally, Theorem 14.7 implies that the convergence rate of the
multigrid method has the same behavior as the two-grid method based on
the two finest grids. Consider as a typical case an approximation of order
s in the number Zn of grid points on the nth level
IIAn - All = 0 (Z~)

and assume a relation
(14.22)
between the number of grid points on two consecutive levels with a constant
« 1. Then
Hence, in this case, (14.20) is satisfied with q = (s. In one dimension,

typically the number of grid points on each level will be doubled, Le.;
( = 1/2. In this case for an approximation of second order s = 2 we have
q = 1/4 and (14.21) requires C :::; (..jf7 - 4) /2 = 0.061 ....
For a discussion of the computational complexity, let an denote the num-
ber of operations necessary for the matrix-vector multiplication involved in
the application of An. In most cases we will have an = az;, where a is
some constant. Denote by bn the number of operations for one step of the
multigrid iteration with n + 1 levels. Then, from the recursive definition,
we have
bn = an + pbn-b (14.23)
since each step first requires the evaluation of the residual on the level nand
then performs p steps of the multigrid iteration with n levels. Neglecting
the effort for the direct solutions on the coarsest grid, from (14.22) and
(14.23) we obtain that
(14.24)
provided p(2 < 1. In particular, for the canonical case where p = 2 and
(= 1/2 we have bn :::; 2az;. In general, provided the number of grid points
grows fast enough in relation to the iteration number p, the computational
complexity for the multigrid method is of the same order as for the two-grid
method using the finest and coarsest grids.
Using iterative methods for the approximate solution of 'fJn -An'fJn = fn,
in general, it is useless to try and make the iteration error lI'fJn,i - 'fJnll
smaller than the discretization error lI'fJn - 'fJ11 to the exact solution of
'fJ - A'fJ = f· Therefore, the number of iterations should be chosen such
that both errors are of the same magnitude. Unfortunately, the quanti-
tative size of the discretization error, in most practical situations, is not
known beforehand, only its order of convergence is known. However, the
nested iteration, or full multigrid scheme, which we will describe briefly,
constructs approximations with iteration errors roughly of the same size as
the discretization error. The basic idea is to provide a good initial element
'fJn,o for the multigrid iteration on n + 1 levels by multigrid iterations on
coarser levels as described in the following definition.
Definition 14.8 Starting with 'Po := (1 - AO)-l fo, the full multigrid
scheme constructs a sequence ('Pn) of approximations by performing k steps
of the multigrid iteration on n + 1 levels using the preceding 'Pn-l as initial
element.
First, we note that the computational complexity of the full multigrid
method is still of order O(z;). Using (14.23) and (14.24), the total number
en of operations up to n + 1 levels can be estimated by
~ a 1 2
en = k L..J bm :::; k 1- (2 1- 1"2 zn·
m=l p.,
Theorem 14.9 Assume that the discretization error satisfies
(14.25)
for some constants 0 < q < 1 and C > 0, and that t := SUPnElN liMn II
satisfies
(14.26)
Then for the approximation 'Pn obtained by the full multigrid method we
have that
- _ "<
'Pn 'Pn _ C (q + l)tk n
" q-
tk q, n E IN. (14.27)
Proof. We set
(q + l)tk
a'- "':"=--7.-
.- q - tk
and note that

(a + 1 + q)tk = qa.
Trivially, (14.27) is true for n = O. Assume that it has been proven for
some n 2: O. Since 'Pn+l is obtained through k steps of the defect correction
iteration (14.8) on the level n + 1 with 'Pn as initial element, we can write
k-l
'PnH = M~H'Pn + L M:'HBn+Iin+l'

m=O
From 'Pn+l - An+1'PnH = inH we deduce that
and can insert this into the previous equation to arrive at
II'PnH - 'PnHIl = IIM~H('Pn - 'Pn+l) II :S tkll'Pn - 'PnHIl
:S t k (II'Pn - 'Pnll + lI'Pn - 'PII + II'P - 'Pn+lll)
Now, indeed, Theorem 14.9 implies that the iteration error, even for
k = 1, is of the same size as the discretization error, provided the ap-
proximation on the coarsest level is sufficiently accurate to ensure (14.26)
by Theorem 14.7. In the typical situation where 8 = 2 and ( = 1/2, an

estimate of the form (14.25) is satisfied with q = 1/4.
In connection with the two-grid iterations of Theorem 14.5 we observed
already that for the Nystrom method we cannot expect IITnll-+ 0, n -+ 00.
Therefore, in view of Theorem 14.7, modifications of the multigrid scheme
are necessary for the quadrature method. The multigrid iteration proposed
by Hemker and Schippers [67] replaces (14.16) by
Eo := (I - AO)-l
p-l
En := L [I - E n- l (I - An_d]m E n- l (I - A n - l + An), n = 1,2, ....
m=O
This means that at each level a smoothing operation 1- A n- l + An is

included before the application of p steps of the multigrid iteration at the
preceding level. We write
En = Qn(I - An- l + An)

and, as above, we have the property
Proceeding as in the derivation of (14.19), ~ can !?e shown that the iteration
operators satisfy the recurrence relation Ml = Tl and
Mn+! = Tn+! + M~(I - Tn+ l ), n = 1,2, ... ,

where the two-grid iteration operator Tn is defined by
Tn := (I - An_l)-l(An - An-l)An , n = 1,2, ....
Now Theorem 14.7 holds with Tn and Mn replaced by Tn and Mn. From
Corollary 10.8 we have that IITnll -+ 0, n -+ 00.
For further variants and a more detailed study of multigrid methods for
equations of the second kind including numerical examples, we refer the
reader to Hackbusch [59].
14.4 Fast Matrix-Vector Multiplication

If A is a compact operator then the eigenvalues of I - A are clustered at one
(see Theorem 3.9). Furthermore, from Theorems 14.3 and 14.4, we observe
that the condition number of the linear systems arising from the Nystrom
method (or the collocation method with piecewise linear interpolation) is
independent of the size n of the matrix, i.e., of the degree of discretization.
Therefore, using conjugate gradient-type methods like GMRES (or other

Krylov subspace methods) for the iterative solution of these systems, we
can expect that the number of iterations required for a certain accuracy
is independent of n (see [54]). Hence, the computational complexity of the
iterative solution of integral equations of the second kind by these iteration
methods is determined by the amount of work required for the computation
of matrix-vector products. Note that these and subsequent considerations
also apply to the matrix-vector multiplications occurring in the two-grid
and multigrid methods of the two previous sections.
Since the discretization of integral equations leads to dense matrices, each
matrix-vector multiplication, if it is performed in the naive way, requires
O(n 2 ) operations. Therefore, even solving the linear systems by iteration
methods such as the conjugate gradient method or GMRES might become
too expensive for realistic applied problems. A remedy is to try and perform
the matrix-vector multiplication only approximately within some tolerance.
Indeed, in many cases this concept allows a reduction of the computational
complexity to almost linear growth of order O( n). Examples of this type are
the panel clustering methods suggested by Hackbusch and Nowak [61) and
the closely related fast multipole methods of Rokhlin [155) and Greengard
and Rokhlin [55). Another approach has been initiated through Beylkin,
Coifman, and Rokhlin [16) by using a wavelet basis for spline spaces in the
Galerkin method for integral equations of the second kind. In the latter
method, the main idea is to replace the exact Galerkin matrix through a
matrix that is nearly sparse by setting all entries below a certain threshold
equal to zero. For a survey on these wavelet methods we refer to Dah-
men [29). Since developing the main ideas of wavelets is beyond the aim of
this introduction, we will confine ourselves to explaining the basic principle
of the panel clustering and fast multipole methods by considering a simple
model problem of a one-dimensional integral equation.
For one-dimensional integral equations, in general, the discrete linear
system can be solved efficiently by either elimination methods or the two-
grid and multigrid methods with traditional matrix-vector multiplication.
Nevertheless, the following example is suitable for explaining the basic ideas
of panel clustering and fast multipole methods. This is due to the fact that
in higher dimensions there are more technical details to consider, which
distract from the basic principles. However, these basic principles do not
depend on the dimension of the integral equation.
Denote by r the unit circle in m? and consider an integral equation of
the second kind
cp(x) - [K(X,Y)CP(Y)dS(Y) = f(x), x E r, (14.28)
with sufficiently smooth kernel K and right-hand side f. Using the compos-
ite trapezoidal rule in the Nystrom method the integral equation (14.28) is
approximated by the linear system
(14.29)
with the equidistant discretization points Xj = (cos(2j7r/n),sin(2j7r/n)),

j = 1, ... ,n. Assume that for each z E r we have degenerate kernels
=L
M
KM(X,y;Z) am(x;z)bm(y;z) (14.30)
m=1
available such that

c
IK(x,y) -KM(x,y;z)1 ~!Vi (14.31)
a
for Iy - zl ~ 21x - zi and some constants c > 0 and a > 1. Such degenerate
kernels can be obtained, for example, by Taylor series expansions. In the
fast multipole method for the kernel of the logarithmic double-layer po-
tential operator the approximating kernels are obtained by taking the real
part of Taylor expansions of the complex logarithm in <C. Now the main
idea of both the panel clustering and the fast multi pole methods consists in
approximating the exact kernel in (14.29) by the degenerate kernels (14.30)
for Xk away from Xj. As a result, large submatrices of the n x n matrix
K(Xj,Xk) are well approximated by matrices of low rank. Multiplying an
n x n matrix of rank M with a vector requires only O(Mn) operations
as opposed to O(n 2 ) for a matrix of full rank. This observation leads to a
scheme for the approximate evaluation of the matrix-vector multiplication
in (14.29) with the computational cost reduced to order O(Mn log n).
To explain this scheme we assume that n = 2P with an integer p ~ 3.
For q = 1, ... ,p-1 and f = 1,2, ... , 2q+l we introduce subsets A~, Bj C r
by setting
A~:= {(Coscp,sincp): Icp- ~:I <~} and Bj :=r\A~.

Then we have that A~l C A~-1 and A~l_1 C A~-1 and that Bgl ::> Bj-l
and Bgl _ 1 ::> Br 1 for q = 2, ... ,p - 1 and f = 1,2, ... , 2q. Each set A~
contains a total of 2P - q - 1 of the discretization points centered around
Correspondingly, the complement set Bj contains 2P - 2P - q + 1 discretiza-

tion points. For notational convenience we also define A~ := {xl!} for
f = 1, ... , n. Finally, we set C} := Bl for f = 1, ... ,4, and recalling that
B~l ::> Br 1 and B~l_1 ::> Brl, we further define
q .- Bq2£ \ Bq-l
C2£.- q .- Bq
C 2£-1·- \ Bq-l
l' 2£-1 l
e
for q = 2, ... ,p -1 and = 1,2, ... , 2q • Then, for q 2: 2, the set C'f contains
2P - q discretization points.
Now, we set SO(Xj) := 0, j = 1, ... , n, and compute recursively
Sq(Xj) := Sq-l(Xj) + 2::= am(xj; z'f) 2::= bm(Xk; Z'f)'Pk' Xj E Ai, (14.32)
m=l
e
for q = 1, ... ,p -1 and = 1, 2, ... , 2q +1. Since each of the sets C'f contains
2P - q discretization points and since we have 2q +1 sets C'f, the computa-
tions of the Sq(Xj) from the preceding Sq-l(Xj) requires a total of O(Mn)
operations. Therefore, the total cost to compute Sp-l (Xj) for j = 1, ... ,n
is of order O(Mpn), i.e., of order O(Mnlogn).
Since 21xj - z'fl :sIXk - z'fl for Xj E A~tl and Xk E Ci,from (14.31) we
can conclude that
for Xj E A~tl and e= 1, ... , 2q+1 . From this, by induction, we obtain that
for Xj E A~tl and £ = 1, ... , 2q +1. In particular, for q = p - 1, we have

that
j = 1, ... ,n, (14.33)
e
since A~-l = {xR} for = 1, ... , n. Therefore, the above scheme indeed
computes the required matrix-vector multiplication with a total computa-
tional cost of order 0 (M n log n).
From (14.33) we also observe that the computations correspond to re-
placing the matrix with the entries K(Xj,xk)/n by an approximating ma-
trix such that the difference has maximum norm less than or equal to
ca- M . This, by Theorem 14.2, implies that the resulting solution of the
perturbed linear systems differs from the solution of (14.29) by an error
of order O(a- M ). We want to ensure that the additional error induced by
the approximation of the matrix elements is of the same magnitude as the
error for the approximation of the integral equation (14.28) by the Nystrom
method (14.29). If for the latter we assume a convergence order 0 (n -m) for
some mE IN, then we need to make sure that a- M = O(n- m ), i.e., we have
to choose M = O(mlogn). Therefore, we have achieved a computational
scheme preserving the accuracy of the numerical solution of the integral
equation with the computational cost of order O(mn(logn)2).
For analytic kernels and right-hand sides, for the Nystrom method, we
have exponential convergence O(e- sn ) for some 8 > o. Therefore, in this
case, the above argument leads to the requirement that M = O(n), i.e., we
do not obtain a reduction of the computational cost compared to the naive
matrix-vector multiplications. Hence, as a rule of thumb, panel clustering
and fast multi pole methods are only efficient for low-order approximations.
For further studies of panel clustering and fast multi pole methods,
we refer to Hackbusch [60], Hackbusch and Sauter [62], Greengard and
Rokhlin [56], and the references therein.
Problems
14.1 For the operator M n , from the proof of Theorem 14.3, show that
IIMnlioo = sup IIMnlloo = 1

11<1>1100=1
and verify the existence of the functions L j used in the definition of Mn.
14.2 For the condition number of the Bubnov-Galerkin method for an equation
of the second kind, define matrix operators En, An : ([:n -t ([:n by
n
(En'Y)j := ~)Uj, ukhk

k=l
and n
(An'Y)j := L(Uj, AUkhk

k=l
for j = 1, ... , n. Proceding as in Theorem 14.4, show that in the Euclidean norm
the condition number of the linear system satisfies
cond(En - An) :$ CcondEn
14.3 Brakhage [17] suggested using En = 1 + (1 - Am)-l An as an approximate
inverse for the Nystrom method in the defect correction method (compare also
the proof of Theorem 10.9). Prove a variant of Theorem 14.5 for this case.
14.4 Set up the equations for one step of the two-grid iteration for the colloca-
tion method.
14.5 Formulate and prove an analog of Theorem 14.7 for the case of exponential
convergence IITnll = O(exp(-szn-d) with some s > o.
15
Equations of the First Kind
Compact operators cannot have a bounded inverse. Therefore, equations

of the first kind with a compact operator provide a typical example for so-
called ill-posed problems. In this chapter we introduce regularization meth-
ods for the stable solution of equations of the first kind in a Hilbert space
setting. For a more comprehensive study of ill-posed problems, we refer to
Baumeister [13], Engl, Hanke, and Neubauer [38], Groetsch [58], Kirsch [87],
Louis [116], Morozov [129], and Tikhonov and Arsenin [174].
15.1 Ill-Posed Problems

For problems in mathematical physics, in particular for initial and bound-
ary value problems for partial differential equations, Hadamard [63] postu-
lated three properties:
(1) Existence of a solution.
(2) Uniqueness of the solution.
(3) Continuous dependence of the solution on the data.
The third postulate is motivated by the fact that in all applications
the data will be measured quantities. Therefore, one wants to make sure
that small errors in the data will cause only small errors in the solution. A
problem satisfying all three requirements is called well-posed. The potential
theoretic boundary value problems of Chapter 6 and the initial boundary
value problem for the heat equation of Chapter 9 are examples of well-

266 15. Equations of the First Kind
posed problems. We will make Hadamard's concept of well-posedness more

precise through the following definition.
Definition 15.1 Let A : U -t V be an operator from a subset U of a
normed space X into a subset V of a normed space Y. The equation
Acp= f (15.1)
is called well-posed or properly posed if A is bijective and the inverse oper- .
ator A-I: V -t U is continuous. Otherwise, the equation is called ill-posed
or improperly posed.
According to this definition we may distinguish three types of ill-posed-
ness. If A is not surjective, then equation (15.1) is not solvable for all f E V
(nonexistence). If A is not injective, then equation (15.1) may have more
than one solution (nonuniqueness). Finally, if A-I: V -t U exists but is
not continuous, then the solution cp of (15.1) does not depend continuously
on the data f (instability). The latter case of instability is the one of pri-
mary interest in the study of ill-posed problems. We note that the three
properties, in general, are not independent. For example, if A : X -t Y
is a bounded linear operator mapping a Banach space X bijectively onto
a Banach space Y, then by the inverse mapping theorem (see [156, 169])
the inverse operator A-I : Y -t X is bounded and therefore continuous.
For a long time the research on improperly posed problems was neglected,
since they were not considered relevant to the proper treatment of applied
problems.
Note that the well-posedness of a problem is a property of the operator
A together with the solution space X and the data space Y including the
norms on X and Y. Therefore, if an equation is ill-posed one could try to
restore stability by changing the spaces X and Y and their norms. But, in
general, this approach is inadequate, since the spaces X and Y, including
their norms are determined by practical needs. In particular, the space Y
and its norm must be suitable to describe the measured data and, especially
the measurement errors.
Example 15.2 A classic example of an ill-posed problem, given by
Hadamard, is the following initial value problem for the Laplace equation:
Find a harmonic function u in R x [0, 00) satisfying the initial conditions
8
u(· ,0) = 0, 8y u(· ,0) = f,
where f is a given continuous function. If we choose as data
fn(x) = .!.n sin nx, x E R,
for n E lN, then we obtain the solution
un(x,y) = -\ sinnxsinhny, (x,y) E R x [0,00).

n
15.1 Ill-Posed Problems 267
Obviously, the data sequence (In) converges uniformly to zero, whereas the
solution sequence (un) does not converge in any reasonable norm. There-
fore, the solution to the initial value problem for harmonic functions is ill-
posed. But a number of applications lead to such an initial value problem,
for example, the extrapolation of measured data for the gravity potential
from parts of the surface of the earth to the space above or below this
surface. 0
Example 15.3 A second typical example of an ill-posed problem is the

backward heat conduction. Consider the heat equation
in a rectangle [0,7r] X [0, T] subject to the boundary conditions

u(O,t) = u(7r,t) = 0, 0:::; t:::; T,
and the initial condition
u(x, 0) = c,o(x) , O:::;x:::;7r,

where c,o is a given function. For existence and uniqueness, we refer back to
Chapter 9. Here, the solution can be obtained by separation of variables in
the form
L
00
u(x, t) = an e- n2t sin nx (15.2)

n=l
with the Fourier coefficients
an = - 2111" c,o(y) sin ny dy (15.3)

7r 0
of the given initial values. This initial value problem is well-posed: The final
temperature f := u(· ,T) depends continuously on the initial temperature.
For example, in an £2 setting, using Parseval's equality, we have
i n=l
L a~e-2n2T :::; i L
00 00
IIfll~ = e- 2T a~ = e-2Tllc,oll~·
n=l
However, the inverse problem, i.e., the determination of the initial tem-
perature c,o from the knowledge of the final temperature f, is ill-posed. Here,
we can write
L
00
u(x, t) = bn en2 (T-t) sin nx

n=l
with the Fourier coefficients
bn = -2111" f(y)sinnydy
7r 0
of the given final temperature. Then we have
iL
00
1I<p1I~ = b;e 2n2T

n=l
and there exists no positive constant C such that
i L b;.
00
11<p1I~ ~ Cllfll~ = C
n=l
We may interpret this inverse problem as an initial value problem for the
backward heat equation, where t is replaced by -to The ill-posedness re-
flects the fact that heat conduction is an irreversible physical process. 0
Inserting (15.3) into (15.2), we see that the last example can be put into
the form of an integral equation of the first kind
101< K(x,y)<p(y)dy = f(x), 0 ~ x ~ 71",

where the kernel is given by
~71" L
00
K(x,y) = e- n2T sinnxsinny, 0 ~ x,y ~ 71".

n=l
In general, linear integral equations of the first kind with continuous or
weakly singular kernels provide typical examples for ill-posed problems with
respect to both the maximum and the mean square norm. They are special
cases of the following theorem.
Theorem 15.4 Let X and Y be normed spaces and let A : X -+ Y be
a compact linear opemtor. Then the equation of the first kind A<p = f is
improperly posed if X is not of finite dimension.
Proof. In view of Theorems 2.16 and 2.20 the compact operator A cannot
have a continuous inverse. 0
The ill-posed nature of an equation, of course, has consequences for its

numerical treatment. We may view a numerical solution of a given equation
as the solution to perturbed data. Therefore, straightforward application
of the methods developed in previous chapters to ill-posed equations of the
first kind will usually generate numerical nonsense. In terms of the con-
cepts introduced in Section 14.1, the fact that an operator does not have a
bounded inverse means that the condition numbers of its finite-dimensional
approximations grow with the quality of the approximation. Hence, a care-
less discretization of ill-posed problems leads to a numerical behavior that
only at first glance seems to be paradoxical. Namely, increasing the degree
of discretization, i.e., increasing the accuracy of the approximation for the
operator A, will cause the approximate solution to the equation A<p = f to
become less reliable.
15.2 Regularization of Ill-Posed Problems 269
15.2 Regularization of Ill-Posed Problems

Methods for a stable approximate solution of ill-posed problems are called
regularization methods. Note that in the context of improperly posed prob-
lems the term regularization has a different meaning than in the theory of
singular equations described in Chapter 5. It is our aim to introduce a few
of the classical regularization concepts for linear equations of the first kind.
In the sequel, we will mostly assume that the linear operator A is injec-
tive. This is not a principal loss of generality, since uniqueness for a linear
equation can always be achieved by a suitable modification of the solution
space X. We wish to approximate the solution cp to the equation Acp = f
from the knowledge of a perturbed right-hand side fO with a known error
level
lifO - fll ~ 8. (15.4)
When f belongs to the range A(X), there exists a unique solution cp of
Acp = f. For the perturbed right-hand side, in general, we cannot expect
fO E A(X). Using the erroneous data fO we want to construct a reason-
able approximation cpo to the exact solution cp of the unperturbed equa-
tion Acp = f. Of course, we want this approximation to be stable, i.e.,
we want cpo to depend continuously on the actual data fO. Therefore our
task requires finding an approximation of the unbounded inverse operator
A-I: A(X) -t X by a bounded linear operator R: Y -t X.
Definition 15.5 Let X and Y be normed spaces and let A : X -t Y be an

injective bounded linear operator. Then a family of bounded linear operators
Ra : Y; --t X, O! > 0, with the property of pointwise convergence
lim RaAcp = cp, cp E X, (15.5)

a~O
is called a regularization scheme for the operator A. The parameter O! is

called the regularization parameter.
Of course, (15.5) is equivalent to Raf -t A-I f, O! -t 0, for all f E A(X).

Occasionally, we will use regularization parameter sets other than the pos-
itive real numbers.
The following theorem illustrates properties that cannot be fulfilled by
regularization schemes for compact operators.
Theorem 15.6 Let X and Y be normed spaces, A : X -t Y be a com-

pact linear operator, and dim X = 00. Then for a regularization scheme
the operators Ra cannot be uniformly bounded with respect to O!, and the
operators RaA cannot be norm convergent as O! -t O.
Proof. For the first statement, assume that IIRall ~ C for all O! > 0 with
some constant C. Then from Raf --t A-I f, O! --t 0, for all f E A(X) we
deduce IIA- 1 III ~ GII/II, i.e., A-I: A(X) --t X is bounded. By Theorem
15.4 this is a contradiction to dimX = 00.
For the second statement, assume that we have norm convergence. Then
there exists a > 0 such that IIRaA - III < 1/2. Now for all I E A(X) we
can estimate
IIA -1 III ~ IIA -1 1- RaAA -1 III + IIRa/1l ~ ~ IIA -1 III + liRa 1111/11,
whence IIA-l III ~ 211Rallll/il follows. Therefore A-I: A(X) --t X is
bounded, and we have the same contradiction as above. D
The regularization scheme approximates the solution <p of A<p =I by

the regularized solution
(f/5 . -
Ta'-
R a 16. (15.6)
Then, for the total approximation error, writing
<p~ - <p = Ra/ 6 - Ral + RaA<p - <p,
by the triangle inequality we have the estimate
(15.7)
This decomposition shows that the error consists of two parts: the first
term reflects the influence of the incorrect data and the second term is due
to the approximation error between Ra and A-I. Under the assumptions
of Theorem 15.6, the first term cannot be estimated uniformly with respect
to a and the second term cannot be estimated uniformly with respect to
<po Typically, the first term will be increasing as a --t 0 due to the ill-posed
nature of the problem whereas the second term will be decreasing as a --t 0
according to (15.5). Every regularization scheme requires a strategy for
choosing the parameter a in dependence on the error level 8 in order to
achieve an acceptable total error for the regularized solution. On one hand,
the accuracy of the approximation asks for a small error IIRaA<p - <p1I, i.e.,
for a small parameter a. On the other hand, the stability requires a small
IIRall, i.e., a large parameter a. An optimal choice would try and make the
right-hand side of (15.7) minimal. The corresponding parameter effects a
compromise between accuracy and stability. For a reasonable regularization
strategy we expect the regularized solution to converge to the exact solution
when the error level tends to zero. We express this requirement through
the following definition.
Definition 15.7 A strategy lor a regularization scheme R a , a > 0, i.e.,
the choice 01 the regularization parameter a = a(8) depending on the error
level 8 and, possibly, on 16 , is called regular il lor all I E A(X) and all
16 E Y with 11/6 - III ~ 8 we have
Ra (6)!6 --t A-I I, 8 --t o.
In the discussion of regularization schemes, One usually has to distin-

guish between an a priori and an a posteriori choice of the regularization
parameter a. An a priori choice would be based on some information on
smoothness properties of the exact solution that, in practical problems, will
not generally be available. Therefore, a posteriori strategies based on some
considerations of the data error level 5 are more practical.
A natural a posteriori strategy is given by the discrepancy or residue
principle introduced by Morozov [127, 128]. Its motivation is based on the
consideration that, in general, for erroneous data the residual IIA<p~ - p5 11
should not be smaller than the accuracy of the measurements of f, i.e., the
regularization parameter a should be chosen such that
with some fixed parameter, 2: 1 multiplying the error level 5. In the case of
a regularization scheme Rm with a regularization parameter m = 1, 2, 3, ...
taking only discrete values, m should be chosen as the smallest integer
satisfying
IIARmf8 - f 8 11 ::; ,5.
Finally, we also need to note that quite often the only choice for selecting
the regularization parameter will be trial and error, i.e., one uses a few
different parameters a and then picks the most reasonable result based on
appropriate information on the expected solution.
In the sequel we will describe some regularization schemes including reg-
ular strategies in a Hilbert space setting. Our approach will be based on
the singular value decomposition.
15.3 Compact Self-Adjoint Operators

Throughout the remainder of this chapter, X and Y will always denote
Hilbert spaces. From Theorem 4.9 recall that every bounded linear operator
A: X -+ Y possesses an adjoint operator A* : Y -+ X, i.e.,
for all <p E X and 'ljJ E Y. For the norms we have IIAII = IIA*II.
Theorem 15.8 For a bounded linear operator we have
A(X)i. = N(A*) and N(A*)i. = A(X).

Proof. 9 E A(X)i. means (A<p,g) = 0 for all <p E X. This is equiva-
lent to (<p, A *g) = 0 for all <p EX, which is equivalent to A *9 = 0, i.e.,
9 E N(A*). Hence, A(X)i. = N(A*). We abbreviate U = A(X) and, triv-
ially, have fj C (Ui.)i.. Denote by P : Y -+ fj the orthogonal projection
operator. For arbitrary cp E (U 1.)1., by Theorem 1.25, we have orthogo-

nality Pcp - cp 1- U. But we also have Pcp - cp 1- U1., because we already
know that tJ c (U1.)1.. Therefore it follows that cp = Pcp E tJ, whence
tJ = (U1.)1., i.e., A(X) = N(A *)1.. 0
An operator A : X -+ X mapping a Hilbert space X into itself is called

self-adjoint if A = A*, i.e., if
(Acp,1/J) = (cp,A1/J)
for all cp,1/J EX. Note that for a self-adjoint operator the scalar product
(Acp,cp) is real-valued for all cp E X, since (Acp,cp) = (cp,Acp) = (Acp,cp).
Theorem 15.9 For a bounded self-adjoint operator we have
IIAII = sup I(Acp, cp)l· (15.8)

IIcpll=l
Proof. We abbreviate the right-hand side of (15.8) by q. By the Cauchy-

Schwarz inequality we find
I(Acp,cp)l::; IIAcpllllcpll::; II All

for allllcpli = 1, whence q ::; II All follows. On the other hand, for all cp, 1/J E X
we have
(A(cp+1/J) , cp+1/J)-(A(cp-1/J), cp-1/J) = 2{(Acp,1/J)+(A1/J,cp)} = 4Re(Acp,1/J).

Therefore we can estimate
4 Re(Acp,1/J) ::; q{llcp + 1/J1I2 + IIcp -1/J1I2} = 2q{lIcpll2 + 1I1/J112}.
Now let IIcplI = 1 and Acp =f:. O. Then choose 1/J = IIAcpll-l Acp to obtain
IIAcplI = Re(Acp, 1/J) ::; q.
This implies IIAII ::; q and completes the proof. o
Recall the spectral theoretic notions of Definition 3.8.
Theorem 15.10 All eigenvalues of a self-adjoint operator are real and

eigenelements to different eigenvalues are orthogonal.
Proof. Acp = ACP and cp =f:. 0 imply A(cp, cp) = (Acp, cp) E JR, whence A E JR.
Let Acp = ACP and A1/J = j.t1/J with A =f:. j.t. Then, from
(A - j.t) (cp, 1/J) = (Acp,1/J) - (cp, A1/J) =0

it follows that (cp, 1/J) = O. o
Theorem 15.11 The spectral radius of a bounded self-adjoint operator A

satisfies
r(A) = IIAII. (15.9)
If A is compact then there exists at least one eigenvalue with 1>,1 = IIAII.
Proof. For each bounded linear operator we have r(A) ::; IIAII, since all
A E <C with IAI > IIAII are regular by the Neumann series Theorem 2.9. By
Theorem 15.9 there exists a sequence (~n) in X with lI~nll = 1 such that
We may assume that (A~n' ~n) ---t A, n ---t 00, where A is real and IAI = IIAII.
Then
°: ; IIA~n - A~n112 = IIA~n112 - 2A(A~n' ~n) + A211~n1l2

::; IIAII2 - 2A(A~n' ~n) + A2 = 2A{A - (A~n' ~n)} ---t 0, n ---t 00.
Therefore
A~n - A~n ---t 0, n ---t 00. (15.10)
This implies that A is not a regular value because if it was we would have
the contradiction
1 = lI~nll = II (AI - A)-l(A~n - A~n)lI---t 0, n ---t 00.
Hence, r(A) 2:: IAI = IIAII·

If A is compact, then the bounded sequence (~n) contains a subsequence
°
(~n(k») such that A<Pn(k) ---t 1/J, k ---t 00, for'some 1/J E X. We may assume
that A =1= 0, since for A = the statement of the theorem is trivial. Then,
from (15.10) it follows that <Pn(k) ---t ~, k ---t 00, for some <P E X, and
II <Pn(k) II = 1 for all k implies that 1I<p1I = 1. Finally, again from (15.10), by
the continuity of A, we obtain A<p = A~, and the proof is finished. 0
Now we are ready to summarize our results into the following spectral
theorem for self-adjoint compact operators.
Theorem 15.12 Let X be a Hilbert space and let A : X ---t X be a self-

adjoint compact operator (with A =1= 0). Then all eigenvalues of A are real.
A has at least one eigenvalue different from zero and at most a countable
set of eigenvalues accumulating only at zero. All eigenspaces N(AI - A)
for nonzero eigenvalues A have finite dimension and eigenspaces to differ-
ent eigenvalues are orthogonal. Assume the sequence (An) of the nonzero
eigenvalues to be ordered such that
and denote by Pn : X -+ N(>\nI - A) the orthogonal projection operator

onto the eigenspace for the eigenvalue An. Then
00
(15.11)
in the sense of norm convergence. Let Q : X -+ N(A) denote the orthogonal

projection operator onto the nullspace N{A). Then
00
(15.12)
'P
for all EX. {When there are only finitely many eigenvalues, the series
(15.11) and (15.12) degenerate into finite sums.}
Proof. The first part ofthe statement is a consequence of Theorems 3.1,3.9,
15.10, and 15.11. The orthogonal projection operators Pn are self-adjoint
by Problem 13.1 and bounded with IlPnll = 1 by Theorem 13.3. Therefore,
by Theorem 2.18, they are compact, since the eigenspaces N{AnI - A)
have finite dimension. Hence the operators Am := A - 2:::=1 AnPn are
°
self-adjoint and compact.
Let A =1= be an eigenvalue of Am with eigenelement
Then for 1 ::; n ::; m we have
'P,
i.e., Am'P = A'P.
APn'P = PnAm'P = Pn(A'P - An'P),

since PnPk = °for n =1= k. From this it follows that
since Pn{A'P - An'P) E N{AnI - A). Therefore Pn'P = 0, whence A'P = A'P.
'P
Conversely, let E N{AnI - A). Then Am'P = An'P if n > m and Am'P =
if n ::; m. Therefore, the eigenvalues of Am different from zero are given
°
by Am+l, Am+2, .... Now Theorem 15.11 yields IIAmli = IAm+11, whence
(15.11) follows.
From
11'1' - t, Pn{ ~ 11'1'11' - t, IlPn'PII'

we observe that 2:::=1 IlPn'P1I2 converges in JR. As in the proof of Theorem
8.2 this implies convergence of the series 2:::=1 Pn'P in the Hilbert space
X. Then, by the continuity of A and (15.11), we obtain
A ('P - ~ Pn'P) = A'P - ~ AnPn'P = 0.

Observing that QPn = °for all n completes the proof of (15.12). 0
Let U be a finite-dimensional subspace of X and let ~I. ... , ~m be an

orthonormal basis, i.e., (~m ~k) = 0, n, k = 1, ... , m, n "# k and II~nll = 1,
n = 1, ... ,m. Then the orthogonal projection operator P : X --+ U has the
representation
m
P~ = 2)~, ~n)~n'
n=l
Hence, we can rewrite the series (15.11) and (15.12) into a more explicit
form. For this, deviating from the numbering in Theorem 15.12, we repeat
each eigenvalue in the sequence (An) according to its multiplicity, i.e., ac-
cording to the dimension of the eigenspace N(AnI - A). Assume (~n) to
be a sequence of corresponding orthonormal eigenelements. Then for each
~ E X we can expand
00
A~ = L An(~' ~n)~n (15.13)

n=l
and
00
~ = L(~' ~n)~n + Q~. (15.14)

n=l
By Theorem 15.12, the orthonormal eigenelements of a compact self-adjoint
operator, including those for the possible eigenvalue zero, are complete in
the sense of Theorem 1.28.
Example 15.13 Consider the integral operator A : L2[0, 11"] --+ L2[0, 11"]
with continuous kernel
~ (1I"-x)y,
K(x,y) := {
1
- (1I"-y)x, O:S; x :s; y :s; 11".
7r
This kernel is the so-called Green's function to the boundary value problem
for the simple ordinary differential equation
~"(x) = - f(x), O:S; x :s; 7r,

with homogeneous boundary condition ~(O) = ~(7r) = O. To each function
f E C[O,7r] there exists a unique solution ~ E C 2 [0,1I"] of the boundary
value problem, which is given by
~(x) = fa" K(x, y)f(y) dy, O:S; x :s; 7r. (15.15)
Uniqueness of the solution is obvious, and straightforward differentiation

shows that (15.15) indeed represents the solution. The compact integral
operator A with this so-called triangular kernel is self-adjoint, since its

real-valued kernel is symmetric K(x, y) = K(y, x). The eigenvalue equation
A<p = A<p is equivalent to the differential equation
A<p"(X) + <p(x) = 0, O:S x :S 7r,
with homogeneous boundary condition <p(0) = <p(7r) = O. In particular,

taking A = 0 yields a trivial nullspace N (A) = {O}. The only nontrivial
solutions to the boundary value problem are given through
An = ~2 and <Pn(x) = ~ sin nx, O:S x :S 7r,
for n E IN. Therefore, in this example, (15.13) becomes
71" K(x, y)<p(y) dy = -2 L 1 171" sin ny <p(y) dy

1
00
2" sin nx
o 7r n=1 n 0
corresponding to the classical Fourier expansion of the kernel K. 0
We now wish to characterize the eigenvalues of compact self-adjoint op-

erators by a minimum-maximum principle. We confine our presentation to
the case of a nonnegative operator. A self-adjoint operator A : X --t X is
called nonnegative if
(A<p, <p) ~ 0, <p E X.
Note that nonnegative operators only have nonnegative eigenvalues.
Theorem 15.14 (Courant) Let X be a Hilbert space, A : X --t X be

a nonnegative self-adjoint compact operator, and (An) denote the nonin-
creasing sequence of the nonzero eigenvalues repeated according to their
multiplicity. Then
Al = IIAII = sup (A<p, <p) (15.16)
11'1'11=1
and
An+l = inf sup (A<p, <p), n = 1,2, .... (15.17)

'Pl, ... ,'if;nEX cpJ..'if;l, ... ,'if;n
11'1'11=1
Proof. We need only to prove (15.17). To this end, we abbreviate the right-
hand side of (15.17) by !-In+l. If we choose ¢1 = <PI, ... ,¢n = <Pn, then for
all <p ..L ¢l,"" ¢n from (15.13) and (15.14) we obtain that
00 00
k=n+1 k=n+l
with equality holding for <p = <Pn+l' Therefore /-In+l :S An+!'
On the other hand, to each choice 'l/Jl, ... ,'l/Jn E X there exists an element
'P E span{'Pl, ... , 'Pn+!} with II'PII = 1 and 'P 1.. 'l/JI, ... , 'l/Jn. To see this we
write 'P = E~~~ Ik'Pk and have to solve the linear system of n equations
n+l
L Ik('Pk, 'l/Jj) = 0, j = 1, ... , n,
k=l
for the n + 1 unknowns 11, ... "n+!' This system always allows a solution
that can be normalized such that 1I'P1I2 = E~~~ l'kl 2 = 1. Then, again
from (15.13) and (15.14), it follows that
n+l n+l
(A'P, 'P) = L Akl('P, 'PkW 2: An+! L I('P, 'PkW = An+ll1'P1I2 = An+!'
k=1 k=1
Hence J.Ln+1 2: An+!' and the proof is complete. o
15.4 Singular Value Decomposition

We now will describe modified forms of the expansions (15.11) and (15.12)
for arbitrary compact operators in Hilbert spaces. From Theorem 4.10 we
recall that the adjoint operator of a compact linear operator is compact.
Definition 15.15 Let X and Y be Hilbert spaces, A: X ~ Y be a compact

linear operator, and A * : Y ~ X be its adjoint. The nonnegative square
roots of the eigenvalues of the nonnegative self-adjoint compact operator
A * A : X ~ X are called singular values of A.
Theorem 15.16 Let (J.Ln) denote the sequence of the nonzero singular
values of the compact linear operator A (with A -# 0) repeated accord-
ing to their multiplicity, i.e., according to the dimension of the nullspaces
N(J.L~I -A* A). Then there exist orthonormal sequences ('Pn) in X and (gn)
in Y such that
(15.18)
for all n E IN. For each <p E X we have the singular value decomposition
00
'P = L ('P, 'Pn)'Pn + Q'P (15.19)

n=l
with the orthogonal projection operator Q : X ~ N(A) and

00
A'P = L J.Ln('P, 'Pn)gn' (15.20)

n=1
Each system (/Ln, tpn, gn), n E lN, with these properties is called a singular
system of A. When there are only finitely many singular values, the series
(15.19) and (15.20) degenerate into finite sums. (Note that for an injective
operator A the orthonormal system {tpn : n E IN} provided by the singular
system is complete in X.)
Proof. Let (tpn) denote an orthonormal sequence of the eigenelements of

A* A, i.e., A* Atpn = /L~tpn and define a second orthonormal sequence by
1
gn:= - Atpn.
/Ln
Straightforward computations show that the system (/Ln, tpn, gn), n E lN,
satisfies (15.18). Application of the expansion (15.14) to the self-adjoint
compact operator A * A yields
00
tp = L(tp, tpn)tpn + Qtp, tp E X,

n=l
where Q denotes the orthogonal projection operator from X onto N(A* A).
Let'ljJ E N(A*A). Then (A'ljJ,A'ljJ) = ('ljJ,A*A'ljJ) = 0, and this implies that
N(A* A) = N(A). Therefore, (15.19) is proven and (15.20) follows by ap-
plying A to (15.19). 0
Note that the singular value decomposition implies that for all tp E X
we have
IItpl12 = L I(tp, tpnW + IIQtpl12 (15.21)
n=l
and
00
IIAtpl12 = L/L;I(tp,tpnW. (15.22)

n=l
Theorem 15.17 Let A, B : X -t Y be compact linear operators. Then for

the nonincreasing sequence of singular values we have
/Ll (A) = IIAII = sup IIAtpll (15.23)

IIcpll=l
and
/Ln+l (A) = inf sup II Atp II , n = 1,2, .... (15.24)

,pl, ... ,,pnEX cp.L,pl, ... ,,pn
IIcplI=l
Furthermore
/Ln+m+l (A + B) ::; /Ln+l (A) + /Lm+l (B), n, m = 0,1,2, . . . . (15.25)

Proof. (15.23) and (15.24) follow immediately from Theorem 15.14, since
the squares J.t; of the singular values of A are given by the eigenvalues of
the nonnegative self-adjoint operator A* A, and since (A* A<p, <p) = IIA<p1l2
for all <p EX. The inequality (15.25) is a consequence of
::; inf sup IIA<p1I + inf sup IIB<p1I

.pl, ... ,.pn+mEX ",1..pl, ... ,.pn .pl, ... ,.pn+mEX ",1.1{1n+l,. .. ,1{1n+m
"","=1 "","=1
and (15.24). o
In the following theorem, we express the solution to an equation of the
first kind with a compact operator in terms of a singular system.
Theorem 15.18 (Picard) Let A: X ~ Y be a compact linear operator
with singular system (J.tn, <Pn, gn). The equation of the first kind
A<p = f (15.26)
is solvable if and only if f belongs to the orthogonal complement N(A*)1.

and satisfies
(15.27)
In this case a solution is given by

1
=L - (I,9n)<Pn'
00
<p (15.28)
n=1 J.tn
Proof. The necessity of f E N(A*)1..follows from Theorem 15.8. If <P is a

solution of (15.26), then
and (15.21) implies
whence the necessity of (15.27) follows.

Conversely, assume that f ..L N(A*) and (15.27) is fulfilled. Then, as in

the proof of Theorems 8.2 and 15.12, the convergence of the series (15.27)
in lR implies convergence of the series (15.28) in the Hilbert space X. We
apply A to (15.28), use (15.19) with the singular system (f.1.n, 9n, CPn) ofthe
operator A*, and observe f E N(A*).l to obtain Acp = 'L.':=1(f,9n)9n = f.
This ends the proof. 0
Picard's theorem clearly demonstrates the ill-posed nature of the equa-

tion Acp = f. If we perturb the right-hand side f by fO = f + 89n we obtain
the solution cpo = cp + 8f.1.;;1CPn. Hence, the ratio Ilcpo - cpll/llfo - fll = 1/ f.1.n
can be made arbitrarily large due to the fact that the singular values tend
to zero. The influence of errors in the data f is obviously controlled by
the rate of this convergence. In this sense we may say that the equation
is mildly ill-posed if the singular values decay slowly to zero and that it is
severely ill-posed if they decay very rapidly.
Example 15.19 Consider the integral operator A : L2[O,1) -+ L2[O, 1)
defined by
(Acp)(x):= loX cp(y) dy, 0 ~ x ~ 1.
Then the inverse operator A -1 corresponds to differentiation. The adjoint
operator is given by
(A*~)(x) = 11 ~(y) dy, 0 ~ x ~ 1.

Hence
(A* Acp)(x) = 11 loY cp(z) dzdy, 0 ~ x ~ 1,
and the eigenvalue equation A* Acp = f.1.2cp is equivalent to the boundary
value problem for the ordinary differential equation
f.1.2cp" + cp = 0
with homogeneous boundary conditions cp(1) = cp'(O) = O. The nontrivial
solutions are given by
2 tn2 (2n - l}7rX
f.1.n = (2n - 1)'71" ' CPn ()
X = V L;COS 2 , n E IN.
The singular system is completed by 9n = f.1.;;1 ACPn through
9n ()
x =
V2'sm (2n -2 1)'7I"x .
In this example we have a decay f.1.n = O(1/n) for an integral operator

with ~ discontinuity of the kernel along the diagonal x = y. The kernel
of the integral operator in Example 15.13 has a discontinuity in the first
derivatives at the diagonal and the singular values decay f.1.n = O(1/n2). 0
In general, for compact integral operators the smoothness of the kernel

controls the degree of ill-posedness. Roughly speaking, the smoothness of
the kernel of the operator A determines the smoothness of the range of
A, and this effects the regularity condition on f, which is required for the
solvability of Acp = f. We illustrate this statement with the following result
due to Little and Reade [115].
Theorem 15.20 Let A : L2[-I,I] -+ L2[-1,1] be an integml opemtor

with analytic kernel on [-1,1] x [-1,1]. Then the singular values of A
decay at least exponentially ILn = O(R- n ) for some constant R > 1.
Proof. Let K denote the kernel of A and
1
2" To(x)ao(y) + 1: Tm(x)am(y)
n
Kn(x, y) :=
m=l
its approximation by the orthonormal Chebyshev expansion as considered

in Section 11.3. For the integral operator An with the degenerate kernel
K n , by (11.40) we have IIAn - Alloo = O(R- n ) with some R > 1. Since
dimAn(X) ~ n + 1, from the singular value decomposition (15.20) we
observe that An has at most n + 1 nonzero singular values. Therefore we
can apply Theorem 15.17 to obtain
15.5 Regularization Schemes

As already pointed out, Picard's Theorem 15.18 illustrates the fact that the
ill-posedness of an equation of the first kind with a compact operator stems
from the behavior of the singular values ILn -+ 0, n -+ 00. This suggests
trying to regularize the equation by damping or filtering out the influence
of the factor 1/ILn in the solution formula (15.28).
Theorem 15.21 Let A : X -+ Y be an injective compact linear opemtor

with singular system (ILn, CPn, 9n), n E IN, and let q : (0,00) x (0, IIAII] -+ IR
be a bounded function such that for each a > 0 there exists a positive
constant c(a) with
(15.29)
and
lim q(a, IL)
0<-+0
= 1, 0 < IL ~ IIAII· (15.30)
Then the bounded linear operators Ra : Y ~ X, a > 0, defined by

00 1
Raf:= L - q(a,f-Ln)(f,gn)epn, fEY, (15.31)
n=l f-Ln
describe a regularization scheme with
IIRall ~ c(a). (15.32)
Proof. From (15.21) and (15.29) we have
IIRafll 2 = I: ~ [q(a'f-LnW 1(f,gnW ~ [c(aWI:I(f,gn)1 2 ~ [c(aWllfIl2

~1~ ~1
for all fEY. Therefore, the series (15.31) converges, i.e., the operator Ra
is well defined, and Ra is bounded by (15.32). With the aid of
1
(RaAep, epn) =-
q(a, f-Ln) (Aep, gn) = q(a, f-Ln) (ep, epn)
f-Ln
and the singular value decomposition for RaAep - ep we obtain
00 00
IIRaAep - epll2 = L 1(RaAep - ep, epn)1 2 = L[q(a,f-Ln) - If I(ep, epn)12.

n=l n=l
Here we have used the fact that A is injective. Let ep E X with ep =I- 0 and
c > 0 be given and let M denote a bound for q. Then there exists N (e) E 1N
such that 00
L I(ep, epn)1 2 < 2(M ~ 1)2 .

n=N+l
By the convergence condition (15.30), there exists ao(e) > 0 such that
[q(a, f-Ln) - 1]2 < 211~1I2

for all n = 1, ... , N and all 0 < a ~ ao. Splitting the series in two parts
and using (15.21), it follows that
N
IIRaAep - epll2 < 211~1I2 ~ I(ep, epnW + ~ ~ e

for all 0 < a ~ ao. Thus we have established that RaAep ~ ep, a ~ 0, for
all ep E X, and the proof is complete. 0
Remark 15.22 If we replace condition (15.29) by the stronger condition

q(a'f-L) ~ c(a)f-L2, 0 < f-L ~ IIAII, (15.33)
and denote by M a bound on q, then instead of (15.32) we have
(15.34)
Proof. From
00
ARaf = Lq(o:,f.1,n)(f,gn)gn
n=l
we find
00 00
IIARafll 2 = L[q(o:,f.1,n)F 1(f,gnW:::; M2 L 1(f,gnW :::; M 2 11f11 2 ,

n=l n=l
whence IIARal1 :::; M follows. For the operators Ra : Y -+ Y, 0: > 0, defined
by
as in the proof of Theorem 15.21, the condition (15.33) implies the bound
liRa II :::; 2'(0:). Now, using A* Ra = R a , by the Cauchy-Schwarz inequality
we obtain
for all fEY, whence (15.34) follows. o
We now describe some classical regularization schemes by choosing the

damping or filter function q appropriately.
Theorem 15.23 Let A : X -+ Y be a compact linear operator. Then for

each 0: > 0 the operator 0:1 + A * A : X -+ X has a bounded inverse.
Furthermore, if A is injective then
Ra := (aI + A* A)-l A*
describes a regularization scheme with IIRall :::; 1/2via.

Proof. From
allcpll2:::; (o:cp+ A*Acp,cp), cp E X,
we conclude that for a > 0 the operator 0:1 + A *A is injective. Hence, by
the Riesz Theorem 3.4 we have a bounded inverse (0:1 + A* A)-I, because
A* A is compact.
Now assume that A is injective and let (f.1,n, CPn, gn), n E lN, be a singular
system for A. Then the unique solution CPa of
can be written in the form

Indeed, using A* A'Pn = f..l;''Pn and the singular value decomposition (15.20)
applied to A * f, we find
= l: f..ln(J,gn)'Pn = A* f.
00
(0:1 + A* A)'Pa
n=l
Hence, Ra can be brought into the form (15.31) with
f..l2
q(o:, f..l) = --2 .
O:+f..l
This function q is bounded by 0 < q( 0:, f..l) < 1 and satisfies the conditions
(15.29) and (15.30) with
1
c(o:) = 2vfa
because of the arithmetic-geometric mean inequality vfaf..l :S (0: + f..l2)/2.
The proof of the theorem follows from Theorem 15.21. 0
The regularization described in Theorem 15.23 is called Tikhonov regu-

larization, since it was introduced by Tikhonov [172]. We will analyze it in
more detail in Chapter 16.
Theorem 15.24 Let A : X -+ Y be an injective compact linear operator

with singular system (f..ln, 'Pn, gn), n E IN. Then the spectral cut-off
(15.35)
describes a regularization scheme with IIRml1 = l/f..lm.

Proof. The function q with q(m, f..l) = 1 for f..l 2: f..lm and q(m, f..l) = 0 oth-
erwise satisfies the conditions (15.29) and (15.30) with q(m,f..l) :S f..l/f..lm.
Hence, the proof is completed by Theorem 15.21 and the observation that
Rm(gm) = 'Pm/f..lm. 0
Here, the regularization parameter m determines the number of terms

in the sum (15.35). Accuracy of the approximation requires this number to
be large, and stability requires it to be small. In particular, the following
discrepancy principle turns out to be a regular a posteriori strategy for
determining the stopping point for the spectral cut-off.

with dense range, fEY, and let 8 > O. Then, for the spectral cut-off, there
exists a smallest integer m = m( 8) such that
IIARm(o)f - fll :S 8.
Proof. By Theorem 15.8, the dense range A(X) = Y implies that A* is

injective. Hence, the singular value decomposition (15.19) with the singular
system (ftn, gn, CPn) for the adjoint operator A * yields
00
f = I:-(f,gn)gn, fEY, (15.36)

n=l
and consequently
II(ARm - 1)f11 2 = I:- 1(f,gnW -+ 0, m -+ 00. (15.37)

Pn<p",
From this we conclude that there exists a smallest integer m = m(8) such
that IIARm(eI)! - fll :S 8. 0
From (15.36) and (15.37), we see that
IIARmf - fl12 = IIfll2 - I:- 1(f,gn)1 2 •

Pn?:.P",
This allows a stable determination of the stopping parameter m( 8) by ter-

minating the sum when the right-hand side becomes smaller than or equal
to 82 for the first time.
The regularity of the discrepancy principle for the spectral cut-off de-
scribed through Theorem 15.25 is established in the following theorem.

with dense range. Let f E A(X) and fel E Y satisfy Ilfel - fll :S 8 with
°
8 > and let 'Y > 1. Then, for the spectral cut-off, there exists a smallest
integer m = m( 8), depending on 8 and fel, such that
(15.38)
is satisfied and
(15.39)
Proof. In view of Theorem 15.25, we only need to establish the convergence

(15.39). We first note that (15.37) implies III - ARml1 = 1 for all mE IN.
Therefore, writing
we have the triangle inequalities
(15.40)
and
(15.41 )
From (15.38) and (15.40) we obtain
IIARm(5)i - fll ~ 8 + IIARm(5)f 5 - f 511 ~ (1 + ,)8 -+ 0, 8 -+ 0. (15.42)

Therefore, from the expansion (15.37), we conclude that either the cut-off
number m(8) -+ 00,8 -+ 0, or the expansion for f degenerates into a finite
sum
f = (f,gn)gnL
Jl.nc.Jl.mo
and m(8) ~ mo. In the first case, from IIARm(5)_d 5 - f 511 > ,8 and
(15.41), we conclude that ,8 < 8 + IIA(Rm(5)-d - A-II)II, whence
(15.43)
follows. In order to establish the convergence (15.39), in this case, in view

of (15.7) and (15.43), it suffices to show that
IIRmIlIlA(Rm-IA<p - <p)II-+ 0, m -+ 00,
for all <p EX. But the latter property is obvious from
IIRmIl 2 1IA(Rm_IA<p-<p)1I 2 = ~ L p;I(<P,<PnW ~ L 1(<P,<PnW·

Pm Jl.n<Jl.m-l Jl.n$Jl.m
In the case where f has a finite expansion, clearly

A -If = 'L...J
" -1 ( f
Pn
) <Pn
,gn = Rm f
Jl.nc.Jl.mo
for all m ~ mo. Hence
for all m ~ mo, and therefore m(8) ~ mo. This implies m(8) = mo, because
m(8) ~ mo, as noted above. Observing
IIRm(5)f5 - A-I fll = IIRmo(f5 - 1)11 ~ ~ -+ 0, 8 -+ 0,

Pmo
the proof is finished. 0
°
and let < a < 1/IIAII2. Then the bounded linear operators
m
Rm:= a L(I - aA* A)k A*
k=O
describe a regularization scheme with regularization parameter m -+ 00 and

IIRml1 ~ Ja(m+ 1).
Proof. Using a singular system for A and the singular value decomposition
(15.20) applied to A* f, with the aid of
m
aJL2 2:(1 - aJL2)k =1- (1 - aJL 2)m+1,

k=O
we can write
00 1
Rmf = 2: - {I - (1 - aJL!)m+1 } (J,9n)CPn. (15.44)
n=l JLn
The corresponding function q is given by
It is bounded by 0 S q(m,JL) S 1 for 0 < JL S IIAII, and it satisfies the con-

ditions (15.33) and (15.30) with q(m, JL) S a(m + I)JL2. Hence, the proof is
completed by Remark 15.22. 0
The evaluation of the approximation 'I/1m = Rmf corresponds to m steps

of the iteration scheme
'I/1k := (J - a A* A)'I/1k-l + a A* f, k = 1,2, ... , m, (15.45)
starting with '1/10 := a A *f. This scheme goes back to Landweber [113] and
Fridman [44] and consequently is known as the Landweber-Fridman itera-
tion. The regularization parameter is given by the number m of iteration
steps. Accuracy of the approximation requires m to be large, and stability
requires m to be small. Again, we consider the discrepancy principle as a
strategy for terminating the iterations.
Theorem 15.28 Let A : X --+ Y be an injective compact linear operator
with dense range, fEY, and 0 > O. Then, for the Landweber-Fridman
iterations, there exists a smallest integer m = m(0) such that
Proof. (Compare the proof of Theorem 15.25.) From (15.44) it follows that
00
IIARmf - fll2 = 2:(1- aJL!)2m+21(J,9nW· (15.46)

n=l
Now, because of aJL; S allAII2 < 1 for n E lN, we can choose N E lN such
that
f:
n=N+l
(1- aJL;)2m+21(J,9n)1 2 S f:
n=N+l
1(J,9nW S 0; .
Since (1 - ajt;)2m+2 -t 0, m -t 00, for n = 1, ... , N, there exists mo E IN

such that
N 82
2)1- ajt;ym+2!U,gn)!2 ::; 2"
n=l
for all m 2: mo. In view of (15.46), the last two inequalities imply that
IIARmf - fll ::; 8 for all m 2: mo, and the theorem is proven. 0
In the next theorem we establish the regularity of this discrepancy prin-

ciple for the Landweber-Fridman iteration.
Theorem 15.29 Let A : X -t Y be an injective compact linear operator

with dense range. Let f E A(X) and fO E Y satisfy lifo - fll ::; 8 with
8 > 0 and let 'Y > 1. Then, for the Landweber-Fridman iterations, there
exists a smallest integer m = m( 8}, depending on 8 and f O, such that
(15.47)
is satisfied and
(15.48)
Proof. We only need to establish the convergence (15.48). From (15.46) we
conclude that IIARm - III ::; 1 for all m E IN. Therefore, proceeding as in
the proof of Theorem 15.26 (see 15.42) we obtain that IIARm(oJi - fll -t 0
as 8 -t O. Hence, from (15.46) we conclude that either m(8) -t 00, 8 -t 0,
or f = O. Since in the latter case the statement of the theorem is obvious,
we only need to be concerned with the case where m( 8) -t 00, 8 -+ O.
Analogously to (15.43) we obtain b - 1)8 < IIARm(o)-d - fll, and hence
b -1) IIRrn(o) 118 ::; .ja(m(8) + 1) IIARm(o)-d - fll·

Therefore, in view of (15.7) the proof is completed by showing that
vm + 1I1ARm- 1Acp - Acpll -+ 0, m-+ 00, (15.49)
for all cp E X. From (15.46) we conclude that

00
(m + 1) IIARm-1Acp - Acpll2 = "l)m + 1)(1- ajt;)2rnjt;!(cp, CPnW.

n=l
By induction, it can be seen that
for all mE IN. Proceeding as in the proof of Theorem 15.21, with the help
of (m + 1)(1 - ajt2)2rn -t 0, m -+ 00, for all 0 < ajt2 < 1, we can establish
(15.49). 0
Problems 289
Problems
15.1 Prove the Riemann-Lebesgue lemma
11< K(·, y) sin ny dy --+ 0, n --+ 00,
in the mean square norm for a kernel K E L2([0, 11') X [0,11')). How can this result
be used to illustrate the ill-posedness of integral equations of the first kind?
15.2 Let X be a Hilbert space and A : X --+ X be a compact self-adjoint

operator. With the aid of the spectral Theorem 15.12, solve the equation of the
second kind
ACP - Acp = f, A # o.
15.3 The Poisson integral
.
u(pcost,psmt) = -2
11211'1 1- p2
2 (t ) cp(r) dr, 0::; p < 1, 0::; t ::; 211',
11' 0 + p2 - pcos - r
gives the solution to the Dirichlet problem for the Laplace equation lJ.u = 0 in
the unit disk D = {x E lR? : Ixl < 1} with boundary values u(cost,sint) = cp(t),
0::; t ::; 211', on aD in polar coordinates. Therefore, the continuation of a harmonic
function given on a concentric disk of radius r < 1 to a harmonic function on the
unit disk is equivalent to the solution of the integral equation of the first kind
1 121< 1- r2
-
211' 0 1 + r2 - 2r cos( t - r) cp( rdr=ft,
) ( )
with f(t) = u(rcost,rsint), 0::; t::; 211'. Determine the singular values.
15.4 Determine the singular values of the operator A : L2[0, 1)--+ L2[0, 1) given
1'"
by
(Acp)(x):= (x - y)cp(y) dy, 0::; x ::; 1.
What is the inverse of A?
15.5 Consider the central difference quotient
for functions f that are odd with respect to x = 0 and even with respect to x = 1.
Show that Rh, h> 0, can be interpreted as a regularization scheme in the sense
of Theorem 15.21 for the integral operator of Example 15.19.
16
Tikhonov Regularization
This chapter will continue the study of Tikhonov regularization and will
be based on its classical interpretation as a penalized residual minimiza-
tion. We shall explain the concepts of quasi-solutions and minimum norm
solutions as strategies for the selection of the regularization parameter.
The final section of this chapter is devoted to discussion of the classi-
cal regularization of integral equations of the first kind as introduced by
Tikhonov [172] and Phillips [142].
16.1 The Tikhonov Functional

The following theorem presents another aspect of the Tikhonov regular-
ization complementing its introduction in Theorem 15.23. Throughout this
chapter X and Y will always denote Hilbert spaces.
Theorem 16.1 Let A: X -+ Y be a bounded linear operator and let a> O.
Then for each fEY there exists a unique 'POt. E X such that
(16.1)
The minimizer <POt. is given by the unique solution of the equation
(16.2)
and depends continuously on f.

16.2 Weak Convergence 291
Proof. From the equation
IIA<p - 1112 + adl<p1l2 = IIA<Pa - fll2 + adl<Pall 2

+2Re(<p - <Pa, a<pa + A*(A<Pa - I)) + IIA(<p - <Pa)1I2 + all<p - <PaIl 2,
which is valid for all <P E X, we observe that the condition (16.2) is nece~
sary and sufficient for <Pa to minimize the Tikhonov functional defined by
(16.1).
Consider the operator T : X --* X, given by T := a1 + A * A. Since
the operator T is strictly coercive and therefore, by the Lax-Milgram The-

orem 13.26, has a bounded inverse T- 1 : X --* X. 0
The equation (16.2), of course, coincides with the Tikhonov regulariza-

tion introduced in Theorem 15.23. By the interpretation of the Tikhonov
regularization as minimizer of the Tikhonov functional, its solution keeps
the residual IIA<Pa - 1112 small and is stabilized through the penalty term
all<Pa 112. Although Tikhonov regularization itself is not a penalty method,
such a view nevertheless suggests the foUowing constraint optimization
problems:
(a) For given p > 0, minimize the defect IIA 0, minimize the norm 1I<p1I subject to the constraint
that the defect is bounded by IIA<p - III $ O.
The first interpretation leads to the concept of quasi-solutions and the
second to the concept of minimum norm solutions and the discrepancy
principle. For a study of these two stabilization procedures, we need to
introduce the notion of weak convergence.
16.2 Weak Convergence

Definition 16.2 A sequence (<Pn) of elements from a Hilbert space X is
called weakly convergent to an element <P E X il
lor all t/J E X.

For a weakly convergent sequence we will write <Pn ~ <p, n --* 00. Note
that norm convergence <Pn --* <p, n --* 00, always implies weak convergence
<Pn ~ <p, n --* 00, whereas simple examples show that the converse of this
statement is generally false. We leave it to the reader to verify that a se-
quence in a Hilbert space cannot weakly converge to two different elements.
292 16. Tikhonov Regularization
Theorem 16.3 A weakly convergent sequence in a Hilbert space is bounded.

Proof. Let ('Pn) be a weakly convergent sequence. Then the sequence of
bounded linear functionals Fn : X ~ <C, defined by Fn('IjJ) := ('IjJ, 'Pn)
for 'IjJ EX, is pointwise convergent. Therefore, by the uniform boundedness
principle Theorem lOA it is uniformly bounded, i.e., IlPnll ~ C for all n E IN
and some constant C. Hence from II'Pn112 = Fn('Pn) ~ IlFnllll'Pnll ~ CII'Pnll
we have II'Pnll ~ C for all n E IN. 0
Theorem 16.4 Every bounded sequence in a Hilbert space contains a weakly

convergent subsequence.
Proof. Let ('Pn) be a bounded sequence, i.e., II'Pnll ~ C for all n E IN and
some constant C. Then, for each i E IN the sequence ('Pi, 'Pn) is bounded
in <C. Therefore, by the standard diagonalization procedure (see the proof
of Theorem 1.18) we can select a subsequence ('Pn(k)) such that ('Pi, 'Pn(k))
converges in <C as k ~ 00 for each i E IN. Hence, the linear functional F
given by
F('IjJ);= lim ('IjJ,'Pn(k))
k-too
is well defined on U ;= span {'Pi ; i E IN} and by continuity it is also well
defined on the closure U. By decomposing an arbitrary 'IjJ E X into the form
'IjJ = P'IjJ + 'IjJ - P'IjJ where P ; X ~ U denotes the orthogonal projection
operator, we finally find that F is well defined on all of X. Furthermore, F
is bounded by IIPII ~ C. Therefore, by the Riesz representation Theorem
4.8, there exists a unique element 'P E X such that F('IjJ) = ('IjJ,'P) for all
'IjJ EX. Hence,
lim ('IjJ,'Pn(k)) = ('IjJ,'P)
k-too
for all 'IjJ E X, i.e., ('Pn(k)) converges weakly to 'P as k ~ 00. o
In an obvious meaning we may reformulate Theorem 16.4 by saying that
in a Hilbert space each bounded set is relatively weakly sequentially compact
(see Section 1.3).
16.3 Quasi-Solutions
The principal idea underlying the concept of quasi-solutions as introduced
by Ivanov [78] is to stabilize an ill-posed problem by restricting the solution
set to some subset U c X exploiting suitable a priori informations on
the solution of A'P = f. For perturbed right-hand sides, in general, we
cannot expect a solution in U. Therefore, instead of trying to solve the
equation exactly, we minimize the residual. For simplicity we restrict our
presentation to the case where U = B[O; p] is a closed ball of radius p with
some p > O. This choice requires some a priori knowledge on the norm of
the solution.
16.3 Quasi-Solutions 293
Definition 16.5 Let A : X -t Y be a bounded injective linear operator and

let p > O. For a given fEY an element <Po E X is called a quasi-solution
of A<p = f with constraint p if II <Po II :::; p and
IIA<po - fll = IIcpll~p

inf IIA<p - fll·
Note that <Po is a quasi-solution to A<p = f with constraint p if and only if

A<po is a best approximation to f with respect to the set V := A(B[O; pl). It
is obvious how the definition of a quasi-solution can be extended to more
general constraint sets. The injectivity of the operator A is essential for
uniqueness properties of the quasi-solution. For the sake of clarity, in the
following analysis on quasi-solutions, we confine ourselves to the case where
the operator A has dense range in Y.
Theorem 16.6 Let A : X -t Y be a bounded injective linear operator with
dense range and let p > O. Then for each fEY there exists a unique
quasi-solution of A<p = f with constraint p.
Proof. Since A is linear, the set V is clearly convex. By Theorem 1.27 there
exists at most one best approximation to f with respect to V. Since A is
injective, this implies uniqueness of the quasi-solution.
If f E A(B[O; pl), then there exists <Po with II<poil :::; p and A<po = f·
Clearly, <Po is a quasi-solution. Therefore, we only need to be concerned
with the case where f f/. A(B[O; pl). We will establish existence of the quasi-
solution by constructing an element CPo that satisfies the sufficient condition
for the best approximation given in Theorem 1.27. For approximation with
respect to V, this condition reads Re(f - Acpo, Acp - A<po) :::; 0, i.e.,
Re(A*(f - Acpo), cp - CPo) :::; 0 (16.4)
for all 0, fulfills the condition (16.4) and therefore provides a quasi-
solution. For f f/. A(B[O; pl) we will show that a can be chosen such that
the unique solution <Po of (16.5) (see TheQrem 16.1) satisfies IIcpoil = p.
Define a function F : (0,00) -t lR by
where <Po. denotes the unique solution of (16.2). We have to show that F has
a zero. By a Neumann series argument, using Theorem 10.1, the function F
can be seen to be continuous. From (16.3) we obtain that allcpo.II :::; IIA* fll,
whence
<Po. -t 0, a -t 00, (16.6)
follows. This implies F(a) -t _p2 < 0, a -t 00.
IIA'Pc: -
°
Since A(X) is dense in Y, for every c > there exists 'Pc: E X such that
1112 < c/2. Choose tS such that tSll'Pcl1 2:::; c/2. Then, using Theorem
16.1, for all a < tS we have
This implies convergence
(16.7)
Now assume that F(a) :::; 0, i.e., II'Pall :::; p for all a> 0. Then, by Theorem
16.4, we can choose a sequence (an) with an -t 0, n -t 00, such that we
have weak convergence 'Pn := 'Pan -" 'P, n -t 00, with some 'P E X. From
we obtain II'PII :S p. Writing (A'Pn,'¢) = ('Pn,A*'¢) we conclude weak con-

vergence A'Pn -" A'P, n -t 00. Finally, we can use (16.7) to find
IIA'P - 1112 = lim (A'Pn -

n~oo
I, A'P - f):::; lim IIA'Pn -
n~oo
III IIA'P - III = 0,
which is a contradiction to I t/. A(B[O; pl). Therefore there exists a such

that F(a) = II'Pall 2- p2 > 0. Now the continuity of F implies the existence
of a zero of F, and the proof is completed. 0
The quasi-solution can be shown to restore stability in the sense that it

depends weakly continuously on the right-hand side I (see Problem 16.1).
In applications, errors in the data I generally ensure that I t/. A(B[O; pl).
Then the quasi-solution with constraint p can be obtained numerically by
Newton's method for solving F(a) = 0. By writing
it can be seen that the solution <Pa to (16.2) is differentiable with respect
to a and the derivative d'Pa/da satisfies the equation
(16.8)
Then the derivative of F is given by
F'(a) = 2Re (d;; ''Pa).

We may view the quasi-solution as described in Theorem 16.6 as an
a posteriori strategy for the choice of the parameter a in the Tikhonov
regularization: Given a perturbed right-hand side f6 of f E A(X) with

IIf6 - fll ~ 8, we choose a such that
Cf'~ = (aI + A* A)-IA* f6
satisfies 1ICf'~1I = p with some a priori known bound p on the norm of the
exact solution. Then we can deduce that
ap2 = a(Cf'~, Cf'~) = (ACf'~, f6 - ACf'~) ~ P IIAllllf6 - ACf'~1I
~ pllAllllf6 - AA- 1III ~ p IIAII8
provided IIA- 1 III ~ p. Therefore, in this case, we have the estimate
ap~ IIAII8, (16.9)

which may serve as a starting value for the Newton iteration to find the
zero of F. The following theorem answers the question for regularity in the
sense of Definition 15.7 for this strategy.
Theorem 16.7 Let A : X -+ Y be a bounded injective linear operator
with dense range and let I E A(X) and p ~ IIA-l III. For 1 6 E Y with
11/6 - III ~ 8, let cp6 denote the quasi-solution to ACf' = l li with constraint
p. Then we have weak convergence
cpli ->. A-I I, 8 -+ o.
II p = IIA- 1 III, then we have norm convergence

cpli -+ A-I I, 8 -+ o.
Proof Let g E Y be arbitrary. Then, since IIA-l III ~ p, we can estimate
I(ACf'li - l,g)1 ~ {IIACf'li - llill + Il/li - III} IIgll
(16.10)
~ {IIAA-l 1- llill + Il/ li - III} IIgll ~ 2811gll·
Hence, (Cf'li - A- 1 /,A*g) -+ 0,8 -+ 0, for all g E Y. This implies weak
convergence cpli ->. A-I f, 8 -+ 0, because for the injective operator A the
range A * (Y) is dense in X by Theorem 15.8 and because Cf'li is bounded by
lICf'lili ~ p.
When p = IIA-l III, we have
lICf'li - A-I 1112 = IIcpli 112 - 2 Re( cpli, A-I f) + IIA -1/11 2
~2Re(A-1/-Cf'li,A-lf)-+0, 8-+0,
and the proof is complete. 0
Note that we cannot expect weak convergence if p < IIA- 1fll, since then
we would have the contradiction
IIA -1 fl12 = lim (cp8, A-I 1) :S piiA -1 fll < IIA -1 f112.
8-t0
In general, we also cannot expect norm convergence if p > IIA -1 fll because
generically we will have IIcp811 = p for an 8. Thus, for regularity we need an
exact a priori information on the norm of the exact solution.
Under additional conditions on f, which may be interpreted as regularity
conditions, we can obtain results on the order of convergence.
Theorem 16.8 Under the assumptions of Theorem 16.7, let f E AA * (Y)
and p = IIA-l fll. Then
IIcp8 - A-Ifll = 0(8 1/ 2 ), 8 -+ O.
Proof. We can write A-I f = A *9 with some 9 E Y. Therefore, the last
inequality in the proof of Theorem 16.7, together with (16.10), yields
IIcp8 - A- 1 111 2 :S 2Re(f - Acp8,g) :S 4811gll,

and this is the desired result. o
Example 16.9 The following counterexample shows that the result of
Theorem 16.8 is optimal. Let A be a compact injective operator with
dimA(X) = 00 and singular system (/-In,<Pn,gn)' Consider 1 = /-llgl and
18n = /-llgl + 8ngn with 8n = /-l;'. Then A-II = <PI and
cp8 n = (anI + A* A)-1 A*(/-llgI + 6ngn )
(16.11)
where an must satisfy

1/ 4 82 /1 2
...,----'--r'~I--=-oc- + nr'n = 1
(an + /-li)2 (an + /-l~.)2
so that cp8 n is the quasi-solution with constraint p = 1. Assume now that
we have convergence order
Ilcp8 n -A-IIII=0(6;/2), n-+oo.
Then, using 6n = /-l;', from (16.11) we find
,1/2
Un /-In 0
2 -+ , n -+ 00,
an +/-In
whence a n /6 n -+ 00, n -+ 00 follows. But this is a contradiction to the
inequality (16.9). 0
For a compact integral operator A : L2[a, b] -+ L2[a, b], the Tikhonov

residual functional corresponds to
(16.12)
The regularized equation (16.2) is an integral equation of the second kind.

For its numerical solution the methods of Chapters 12 and 13 are available.
O<x~1.
Its unique solution (see Problem 16.4) is given by rp(x) = 1. For the corre-
sponding integral operator A : L2[0, 1] -+ L2[0, 1] elementary calculations
yield
(A* Acp)(x) = 10 1
H(x, y)rp(y) dy, 0~x ~ 1,
where
H(x,y) = r
10
l
e(x+y)z dz = _1_
x+y
(e x + y -1)
for x + y > 0 and H(O,O) = 1. Our numerical results are obtained by dis-
cretizing -the integral equation of the second kind with a > 0 by Nystrom's
method using Simpson's rule with eight equidistant intervals. The integral
for the right-hand side is also evaluated numerically by Simpson's rule. We
have assumed a regular error distribution f! = fi+( -1)i8 at the grid points
and used the norm p = 1 of the exact solution. Table 16.1 gives the values
of the regularization parameter a (obtained by Newton's method) and the
mean square error E := Ilrp~ - rpll£2 between the regularized solution rp~
and the exact solution rp depending on the error level 8. In addition, the
quotient q := Ej8 1 / 2 is listed. 0
8 a E q
0.02 0.000055 0.112 0.795

0.04 0.000076 0.161 0.805
0.06 0.000093 0.198 0.809
0.08 0.000109 0.229 0.812
0.10 0.000123 0.257 0.814
16.4 Minimum Norm Solutions

As already observed at the end of Section 15.2, the principal motivation for
the discrepancy principle as introduced by Morozov [127, 128] is based on
the observation that, in general, for erroneous data it does not make too
much sense to try and make the residual IIAcp - 111 smaller than the error
in 1. Assume that we have some a priori bound 8 on the error in 1. Then
we look for elements cp satisfying IIAcp - 111 ::; 8 and stabilize by making
the norm Ilcpl! small.
Definition 16.11 Let A : X -+ Y be a bounded linear operator and let

8 > O. For a given 1 E Y an element CPo E X is called a minimum norm
solution of Acp = f with discrepancy 8 il IIAcpo - fll ::; 8 and
Note that CPo is a minimum norm solution to Acp = f with discrepancy

8 if and only if CPo is a best approximation to the zero element of X with
respect to Uf := {cp EX: IIAcp - fll ::; 8}. As in the discussion of quasi-
solutions, for the sake of simplicity, in dealing with existence of minimum
norm solutions we confine ourselves to operators with dense range.
Theorem 16.12 Let A: X -+ Y be a bounded linear operator with dense

range and let 8 > O. Then for each fEY there exists a unique minimum
norm solution of Acp = 1 with discrepancy 8.
Proof. From
for all CpI, CP2 E X and all A E (0,1), we observe that Uf is convex. Then,
by Theorem 1.27, there exists at most one best approximation of the zero
element with respect to Ufo
If Ilfll ::; 8, then clearly CPo = 0 is the minimum norm solution with dis-
crepancy 8. Therefore, we only need to consider the case where Ilfll > 8.
Since A has dense range in Y, the set Uf is not empty and we will again es-
tablish existence of the minimum norm solution by constructing an element
CPo that satisfies the sufficient condition for the best approximation given in
Theorem 1.27. For approximation with respect to Uf, this condition reads
(16.13)
for all cp E X with IIAcp - fll ::; 8. Assume that CPo satisfies
acpo + A* Acpo = A* f (16.14)

16.4 Minimum Norm Solutions 299
for some a °
> and IIAcpo - III = 8. Then
a Re(cpo, CPo - cp) = Re(A*(f - Acpo), CPo - cp)
= Re(Acpo - I, Acp - f) - IIAcpo - 1112

:::; 8 (11Acp - III - 8) :::; 0.
Hence CPo satisfies the condition (16.13) and therefore is a minimum norm
solution. We will show that a can be chosen such that the unique solution
CPo of (16.14) (see Theorem 16.1) satisfies IIAcpo - III = 8.
Define a function G : (0,00) -t lR by
where CPa denotes the unique solution of equation (16.2). We have to show
that the continuous function G has a zero. From the convergence (16.6),
we observe that G(a) -t 11/112 - 82 > 0, a -t 00. On the other hand, from
(16.7) we obtain that G(a) -t _82 < 0, a -t 0. This completes the proof. 0
The minimum norm solution can be proven to depend weakly continu-

ously on the right-hand side I (see Problem 16.2).
In general, we will have data satisfying 11111 > 8, i.e., data exceeding
the error level. Then the minimum norm solution with discrepancy 8 can
be obtained numerically by Newton's method for solving G(a) = 0. After
rewriting
we get
and
G'(a) = - (dJaa, A* f) -IICPaI1 2 - 2aRe (dJaa 'CPa) ,
where the derivative dcpajda is given by (16.8).

We may look at the minimum norm solution as described in Theorem
16.12 as an a posteriori strategy for the choice of the parameter a in the
Tikhonov regularization: Given a perturbed right-hand side fli of an ele-
ment f E A(X) with a known error level IIIIi - fll :::; 8 < IIflill, we choose
a such that
cp~ = (aI + A* A)-l A* fli
satisfies IIAcp~ - f"l1 = 8. Then, using acp~ + A* Acp~ = A* f", we find
provided IIfoll > 8. Hence, we have established the estimate

(16.15)
which we may use as a starting value for the Newton iteration to find the
zero of G.
The following theorem answers the question of regularity for this dis-
crepancy principle for the Tikhonov regularization.
Theorem 16.13 Let A : X -+ Y be a bounded injective linear operator

with dense range and let 8> 0 and f E A(X). For fO E Y with lifo - fll ::; 8
and 8 < IIfoll, let 'Ii denote the minimum norm solution with discrepancy
8. Then
cpO -+ A-I f, 8 -+ o.
Proof. Since IIfoll > 8, from the proof of Theorem 16.12 we know that cpo
minimizes the Tikhonov functional
Therefore,
82+ allcpoll2 = IIAcpo _ fOll2 + a II cpo 112
::; IIAA-l f - fOll2 + allA-l fll2

::; 82 + allA-l f1l 2,
whence
(16.16)
follows. This inequality is also trivially satisfied when Ilfoll ::; 8, since in
this case cpo = O.
Now let g E Y be arbitrary. Then we can estimate
As in the proof of Theorem 16.7, this implies cpo -->. A-If, 8 -+ o. Then
using (16.16), we obtain
IIcpO - A-I fll2 = IIcpoll2 - 2 Re( cpo, A-I f) + IIA -1 fl12

(16.17)
::; 2{IIA- 1 /11 2 - Re(cpO,A- 1 f)} -+ 0, 8 -+ 0,
which finishes the proof. 0
Theorem 16.14 Under the assumptions of Theorem 16.13, let f E AA*(Y).

Then
Proof. Writing A-I f = A*g with some 9 E Y from (16.17) we deduce

/lcp.5 - A- I f/l 2 ~ 2 Re(A-I f - cp.5, A-I f) = 2Re (f - Acp.5,g)
~ 2{/lf - f.511 + IIf.5 - Acp.5I1}/lg/l ~ 48/1gll,

and the proof is finished. o
Using the same example as in connection with Theorem 16.8, it can be
shown that the result of Theorem 16.14 is optimal (see Problem 16.3).
Example 16.15 We apply the discrepancy principle to the integral equa-
tion of Example 16.10. Table 16.2 gives the values of the regularization
parameter a, the mean square error E := /lcp~ - cp/lp, and the quotient
q := E/8 1 / 2 depending on the error level 8. 0
8 a E q
0.02 0.0059 0.067 0.479

0.04 0.0148 0.108 0.542
0.06 0.0252 0.131 0.536
0.08 0.0359 0.145 0.514
0.10 0.0466 0.155 0.492
16.5 Classical Tikhonov Regularization

In our examples of integral equations of the first kind, so far, we have
interpreted the integral operator as a mapping A : L2[a, b] --+ L2[a, b] cor-
responding to the Tikhonov residual functional in the form (16.12). Here,
the choice of the data space Y = L2[a, b], in general, is determined by
the need to adequately measure the error in the data. But we have more
flexibility concerning the solution space X, in particular, when additional
regularity properties of the exact solution are a priori known.
In his pioneering papers on integral equations of the first kind in 1963,
Tikhonov [172, 173] suggested damping out highly oscillating parts in the
approximate solution by incorporating the derivative into the penalty term,
i.e., to replace (16.12) by
lb I (Acp) (x) - f(x)1 2 dx + a lb {lcp(xW + Icp'(xW} dx. (16.18)

To include this approach into our general theory we need a Hilbert space
with norm corresponding to the penalty term in (16.18). Since in Chapter 8
we introduced the Sobolev spaces only via Fourier expansion, we briefly
discuss the definition for the space HI [a, b] based on the concept of weak
derivatives.
Definition 16.16 A function 'P E L2[a, b] is said to have a weak derivative
-lb
'P' E L2[a, b] if
lb 'P'Ij;' dx = 'P''Ij;dx (16.19)
for all 'Ij; E GI[a, b] with 'Ij;(a) = 'Ij;(b) = O.

By partial integration, it follows that (16.19) is satisfied for 'P E GI[a, b].
Hence, weak differentiability generalizes classical differentiability.
From the denseness of {7j; E GI[a,b] : 'Ij;(a) = 'Ij;(b) = O} in L2[a,b], or
from the Fourier series for the odd extension for 'P, it can be seen that
the weak derivative, if it exists, is unique. From the denseness of G[a, b] in
L2[a, b], or from the Fourier series for the even extension of 'P, it follows
that each function with vanishing weak derivative must be constant (almost
everywhere). The latter, in particular, implies
(16.20)
for almost all x E [a, b] and some constant c, since by Fubini's theorem
for all 'Ij; E GI[a, b] with 'Ij;(a) = 'Ij;(b). Hence both sides of (16.20) have the
same weak derivative.
Theorem 16.17 The linear space
endowed with the scalar product
(16.21 )
is a Hilbert space.
Proof. It is readily checked that HI [a, b] is a linear space and that (16.21)
defines a scalar product. Let ('Pn) denote a HI Cauchy sequence. Then
('Pn) and ('P~) are both L2 Cauchy sequences. From the completeness of
L2[a, b] we obtain the existence of 'P E L2[a, b] and X E L2[a, b] such that
IIc,on - c,oll£2 0 and IIc,o~ - xll£2 -t 0 as n

-t -t 00. Then for all 'IjJ E C 1 [a, bj
with 'IjJ(a) = 'IjJ(b) = 0, we can estimate
b
l (c,o'IjJ' + x'IjJ) dx = lb {( c,o - c,on )'IjJ' + (X - c,o~)'IjJ } dx
~ IIc,o - c,onll£2IWIIL2+lIx - c,o~II£2I1'IjJII£2 -t 0, n -t 00.
Therefore, c,o E Hl[a,b] with c,o' = X and IIc,o - c,onllHl -t 0, n -t 00, which
completes the proof. 0
Theorem 16.18 C 1 [a,b] is dense in Hl[a,b].

Proof. Since C[a,b] is dense in L2[a, b], for each c,o E HI [a, b] and c > 0 there
exists X E C[a, b] such that IIc,o' - XIl£2 < c. Then we define 'IjJ E Cl [a, bj by
'IjJ(x) := c,o(a) + 1X(~)~,

x
x E [a, bj,
and using (16.20), we have
c,o(x)-'ljJ(x) = lx{c,o'(~)-x(~)}d~, xE [a,bj.
By the Cauchy-Schwarz inequality this implies IIc,o - 'ljJII£2 < (b - a)c, and
the proof is complete. 0
Theorem 16.19 Hl[a, bj is contained in C[a, bj with compact imbedding.

Proof. From (16.20) we have
c,o(x) - c,o(y) = i c,o'(~) d~,

X
(16.22)
whence by the Cauchy-Schwarz inequality,
(16.23)
follows for all x, y E [a, bj. Therefore, every function c,o E HI [a, bj belongs to
O[a, b], or more precisely, it coincides almost everywhere with a continuous
function.
Choose y E [a, bj such that Ic,o(y) I = mina<x<b Ic,o(x)l. Then from
(b - a) a~~~b Ic,o(x) 12 ~ lb Ic,o(xWdx

and (16.22), again by the Cauchy-Schwarz inequality, we find that
(16.24)
for all cP E H 1 [a, b] and some constant C. The latter inequality means that
the Hi norm is stronger than the maximum norm (in one space dimen-
sion!). The inequalities (16.23) and (16.24), in particular, imply that each
bounded set in H 1 [a,bj is a bounded and equicontinuous subset of C[a,bj.
Hence, by the Arzela-Ascoli Theorem 1.18, the imbedding operator from
Hi [a, bj into C[a, bj is compact. 0
Now we return to integral equations of the first kind and interpret the in-
tegral operator with smooth kernel as a mapping from Hl[a, bj into L 2 [a, bj.
Then our complete theory on regularization including convergence and reg-
ularity remains applicable in this setting. Since the imbedding from Hi [a, b]
into L 2 [a, bj clearly is bounded (by Theorem 16.19 it is even compact) the
integral operator A : H 1 [a,b] -+ L 2 [a,b] is compact by Theorem 2.16.
Only the adjoint operator A* : L2 [a, b] -+ Hi [a, bj looks different. We de-
note it by A* to distinguish it from the adjoint A* : L 2 [a,bj -+ L 2 [a,bj
of A : L 2 [a, bj -+ L 2 [a, bj. We avoid its explicit calculation through the
observation that the regularized equation
(16.25)
is equivalent to
(16.26)
for all 'lj; E Hi [a, bj. This follows from the fact that
for all X E L 2 [a, bj and'lj; E H 1 [a, bj. By Theorem 16.1, there exists a unique
solution CPo. to (16.25). In the following theorem we will show that CPo. is the
solution to a boundary value problem for an integro-differential equation.
Theorem 16.20 Assume that the integral operator A has continuous ker-
nel. Then the unique solution t.po. of the regularized equation (16.25) belongs
to C 2 [a, bj and satisfies the integra-differential equation
(16.27)
and the boundary condition
t.p~{a) = t.p~{b) = O. (16.28)
Proof. First we define 9 E C 1 [a, bj by

and then 't/J E HI [a, b] by
't/J(x):= l {a'P~(~)
x
- g(~)} d~, x E [a, b].
Then, by partial integration, since g(b) = 't/J( a) = 0, we get

Ila'P~ - glli2 = lb (a'P~ - g),¢' dx = lb (a'P~'¢' + g',¢) dx
= a('Pa,'t/J)Hl + (A*A'Pa,'t/J)£2 - (A*j,'t/J)L2 = 0.

Note that by the denseness result of Theorem 16.18 partial integration
can be carried over from 0 1 [a, b] into HI [a, b]. Our last equation implies
a'P~ = g. From this, it follows that 'Pa E 02[a, b] and that the integro-
differential equation (16.27) is satisfied. Inserting (16.27) into (16.26) and
performing a partial integration yields
'P~(b)'t/J(b) - 'P~(a)'t/J(a) = °
for all 't/J E Hl[a,b]. This implies that the boundary conditions (16.28) are
fulfilled. 0
Note that by partial integration a solution of (16.27) and (16.28) also

solves (16.26). Hence, the boundary value problem for the integro-differen-
tial equation and the regularized equation (16.25) are equivalent.
Example 16.21 We compare the Tikhonov regularization based on the £2
norm and the HI norm penalty term for the integral equation of Example
16.10. For the discretization of the regularized equation we use the Petrov-
Galerkin method with linear splines. Let Xj = jh, j = 0, ... , n, be an
equidistant grid with step size h = lin and let L j , j = 0, ... ,n, denote the
corresponding Lagrange basis for linear spline interpolation as introduced
by (11.8). Then straightforward calculations yield the tridiagonal matrices
1 -1
-1 2 -1
-1 2 -1
V=~
h
-1 2 -1
-1 1
for the weights Vmk = fol L'm(y)L',.(y) dy and

2 1
1 4 1
1 4 1
W=~
6
1 4 1
1 2
for the weights Wmk = fol Lm(y)Lk(Y) dy. Writing the approximate solution
in the form
we have to solve the linear system
I: 1'k
n {
a (Wjk + Vjk) + I
r r I}
H(x, y)Lj(x)Lk(Y) dxdy
k=O Jo Jo
= 10 1 11 K(y, x)Lj(x)f(y) dxdy, j = 0, ... , n,
in the case of the HI norm penalty term. Here, K and H denote the kernels
of the integral operators A and A* A, respectively (see Example 16.10). In
the case of the L2 norm the matrix V has to be omitted. Note that the
weight matrix V indicates 'how oscillations in the approximate solution are
penalized by the regularization using the derivative. For numerical evalua-
tion of the matrix elements we apply interpolatory quadratures as described
in detail in Chapters 11 and 13.
Table 16.3 compares the mean square error between the exact and ap-
proximate solutions for both regularizations for an error distribution as in
Example 16.10 with 8 = 0.01.
log a -7 -6 -5 -4 -3 -2 -1
L2 penalty 41.83 6.541 3.268 1.567 0.250 0.040 0.107
HI penalty 3.503 2.353 0.550 0.063 0.015 0.015 0.019
In closing this chapter, we wish to point out that in a paper predating

that of Tikhonov, Phillips [142] proposed penalizing only by the L2 norm
of the derivative, i.e., to replace (16.18) by
lb I (A<p) (x) - f(xWdx +a lb 1<p'(xWdx. (16.29)
Note that this differs from the Tikhonov regularization because the penalty
term is only a semi-norm rather than a norm. For an analysis of this ap-
proach we refer to Groetsch [58]. In view of Theorem 16.20 it is no surprise
that minimizing (16.29) is equivalent to the boundary value problem for
the integro-differential equation
+ A* A<p", = A* f
-a<p~
with boundary condition <p~(a) = <p~(b) = o.

Problems 307
Problems
16.1 Show that the quasi-solution given by Theorem 16.6 depends weakly con-
tinuously on the right-hand side, i.e., norm convergence In -+ I, n -+ 00, implies
weak convergence for the corresponding quasi-solutions cpn ->. cP, n -+ 00. Show
that in the complement of A(B[Oj pJ) the quasi-solution depends continuously on
the right-hand side. How are these results related to the regularity Theorem 16.7?
Hint: By Theorem 16.4 the sequence (CPn) contains a subsequence that converges
weakly to some element 1/1 in X. Establish that 111/111 ~ p and IIA1/1- III ~ IIAcp- III,
i.e., 1/1 is a quasi-solution, and by uniqueness it must coincide with cp. Use this re-
sult to show that the sequence (CPn) itself converges weakly to cp. Finally, conclude
norm convergence from IICPnll = p.
16.2 Show that the minimum norm solution given by Theorem 16.12 depends
weakly continuously on the right-hand side. Show that in the complement of
B[Oj oj C Y the minimum norm solution depends continuously on the right-hand
side. How are these results related to the regularity Theorem 16.13?
Hint: Show that there exists an element X E X with X E Ufn for all sufficiently
large n. Therefore the sequence (CPn) is bounded and contains a weakly convergent
subsequence with limit 1/1 E X. Show that IIA1/1 - III ~ 0 and 111/111 ~ IIcplI, i.e.,
1/1 is a minimum norm solution, and by uniqueness it must coincide with cp. Use
this result to show that the sequence (CPn) itself converges weakly to cp. Finally,
conclude norm convergence from IIAcpn - Inll = o.
16.3 Show that the convergence order given in Theorem 16.14 is optimal.
Hint: Show that J:

16.4 Show that the solution ofthe integral equation in Example 16.10 is unique.
yncp(y) dy = 0, n = 0,1,2, ... , for each solution cp of the
homogeneous equation.
16.5 Show that a linear operator A : X -+ Y from a Hilbert space X into a

Hilbert space Y is compact if and only if weak convergence cpn ->. cp, n -+ 00,
implies norm convergence Acp,. -+ Acp, n -+ 00.
17
Regularization by Discretization
We briefly return to the study of projection methods and will consider

their application to ill-posed equations of the first kind. In particular we
will present an exposition of the moment discretization method. For further
studies of regularization through discretization, we refer to Baumeister [13],
Kirsch [87], Louis [116], and Natterer [136].
17.1 Projection Methods for Ill-Posed Equations

Solving an equation of the first kind
Acp = f (17.1)
with an injective compact operator A : X ~ Y from a Banach space X

into a Banach space Y numerically without regularization usually means
approximately solving it by a projection method. The ill-posedness of the
equation (17.1) then will cause the condition number of the discrete linear
system to grow with the dimension n of the subspace used for the projection
method. Increasing n will make the error due to the discretization smaller,
but due to ill-conditioning the computation will be contaminated by errors
in the given data f. Note that in actual numerical computations such errors
will automatically occur because of round-off effects for the data. On the
other hand, if n is small, then the approximation is robust against errors in
f but will be inaccurate due to a large discretization error. This dilemma
calls for a compromise in the choice of the discretization parameter n.

Recalling the projection methods from Chapter 13, let Xn C X and

Yn C Y be two sequences of subspaces with dim Xn = dim Yn = n and let
Pn : Y -+ Yn be projection operators. For given I E A(X) the projection
method approximates the solution cp E X of Acp = I by the solution
CPn E Xn of the projected equation
(17.2)
For the remainder of this chapter we assume that the projection method
is convergent for exact right-hand sides I E A(X). This, by Theorem 13.6,
implies that the finite-dimensional operators PnA : Xn -+ Y n are invertible
and that the sequence of operators (PnA)-1 PnA : X -+ Xn is uniformly
bounded. Then the operators Rn := (PnA)-1 Pn : Y -+ Xn provide a
regularization scheme in the sense of Definition 15.5 with regularization
parameter n. The approximation
(17.3)
corresponding to (15.6) is obtained by numerically performing the projec-

tion method for the inexact right-hand side 16 •
Theorem 17.1 Let A : X -+ Y be an injective compact linear operator,
I E A(X), and 16 E Y satisfy 11/6- III :S o. Then the approximation
cp~ := Rnl6 by the projection method satisfies
IIcp~ - A-I III :S IIRnllo + C 'l/JEX

inf
n
11'!fJ - A-I III (17.4)
lor some constant C.

Proof. This follows from Theorem 13.6 by decomposing
Ilcp~ - A-I III :S IIRnllll/6 - III + IIRnl - A-I III

:S IIRnllo + (1 + M) inf 1I'!fJ - A-I III,
'l/JEX n
where M is a bound on IIRnAIi. o

The error estimate (17.4) corresponds to (15.7) and illustrates that the
total error consists of two parts: the influence of the incorrect data and
the discretization error of the projection method. Recall that for the com-
pact operator A, by Theorem 15.6, the operators Rn cannot be uniformly
bounded.
At this point we wish to emphasize the fact that for the analysis of
the discretization error we may use any Banach space structure for the
range space. In particular, we may choose an image space such that the
equation becomes well-posed. Hence, we can apply the convergence results
310 17. Regularization by Discretization
on projection methods given in Chapter 13. But for the influence of the error
in the right-hand side we have to stick with an image space as dictated by
the needs of the particular problem. We want to make these remarks more
precise through the following theorem.
Theorem 17.2 Let A : X -+ Y be an injective compact operator from a
Banach space X into a Banach space Y and assume an additional Banach
space Z c Y with continuous imbedding such that A : X -+ Z has a bounded
inverse. Let Xn C X and Yn C Z C Y be two sequences of subspaces with
dimXn = dim Yn = n and let Pn : Y -+ Yn be projection operators (bounded
with respect to the norm on Y) and assume that the projection method is
convergent for all f E A(X) = Z. Then the operators Rn : Y -+ X satisfy
IIRnl1 :::; C IIPnll sup Ilfllz

fEYn ,lIflly=l
for some constant C.

Proof. For abbreviation we set
"In := sup Ilfliz.

fEYn ,lIflly=l
Since by Theorem 1.6 all norms on the finite-dimensional subspace Yn are

equivalent, we have "In < 00. Then, from Ilfllz :::; 'YnllfllY for all f E Y n , it
follows that
(17.5)
for all fEZ and some constant M, since the imbedding from Z into Y
is assumed to be bounded. Hence, the operators Pn are also bounded with
respect to the norm on Z and we can interpret the projection scheme for
A: X -+ Y also as a projection scheme in the structure A : X -+ Z. Since
we assume convergence for all f E A(X), by Theorem 13.6, we know that
the operators (PnA)-l PnA : X -+ X are uniformly bounded. Therefore,
for all fEY with IIfllY = 1 we can estimate
IIRnfll = II(Pn A)-lPn fll = II(PnA)-lPnAA-1Pnfll
:::; II(PnA)-l PnAIlIIA-1Il z -+ x IlPnfliz :::; C'YnllPnll

for some constant C. o
We illustrate the use of Theorem 17.2 by considering the following
example.
Example 17.3 Let A: L 2[O,27r]-+ L 2[O,27r] be given by
(Acp)(t)=2~127r{ln4sin2t~T +K(t,T)}CP(T)dT, O:::;t:::;27r, (17.6)

where we assume that K is infinitely differentiable and 211"-periodic with re-

spect to both variables. As a consequence of Theorem 8.22, it can
be seen that A is a bounded operator A : L 2 [0,211"] -t H1 [0, 211"] with
bounded inverse A-1 : H 1[0,211"] -t L 2 [0,211"], provided A is injective. In
this case, from Theorem 13.29 we can conclude that the Petrov-Galerkin
method using trigonometric polynomials (of degree n) converges for
A : L 2 [0,211"] -t H 1 [0, 211"]. In particular, if the right-hand side also is in-
finitely differentiable, we have a convergence order O(n-m) for all mE IN.
On the other hand, the orthogonal projection operator has norm IlPn II = 1
and for the subspace Yn of trigonometric polynomials of degree n we read~
ily see that "In = VI + n 2 • Hence, in this case we have IINnIl = O(n)
and the equation of the first kind with the operator (17.6) is only mildly
ill-posed. In particular, the decay of the approximation error O(n-m) dom-
inates the increase O(n) of the data error. Therefore, provided the error of
the right-hand side is reasonably small, we conclude that it is numerically
safe to apply the Petrov-Galerkin method in this case, because we need
only a comparatively small n to achieve an acceptable approximation error
(see Example 13.23 and Problem 17.1). For a more detailed discussion of
the regularization of integral equations of the first kind with a logarithmic
singularity we refer to [87]. 0
For projection methods the norm on the image space Y might not always
be the most appropriate measure to describe the inaccuracy of the data.
Consider the Petrov-Galerkin method for an operator A : X -t Y between
two Hilbert spaces X and Y. With subspaces Xn = span{Ub ... ' un} and
Yn = span{vb ... , vn} we write <Pn = L:~=1 "IkUk and then have to solve
the Petrov-Galerkin equations
n
L"Ik(Auk,Vj) = (f,Vj), j = 1, ... ,n, (17.7)
k=1
for the coefficients "11, ... , "In. Here, it makes more sense to measure the
error of the discrete data vector P = ((f, V1) , ... , (f, vn )) in the Euclidean
norm on (Cn (compare the stability analysis of Section 14.1 for equations
of the second kind). Let pd E (Cn be a perturbed right-hand side for (17.7)
with
IIpd -P1l2 ~ d
and denote the Petrov-Galerkin solution to these inexact data by <p~. After
introducing the matrix operators En, An : (Cn -t (Cn by
n
(En"l)j := L(Uk,Uj)"tk, j = 1, ... ,n,
k=1
and n
(An"l)j:= L(Auk,Vj)"tk, j = 1, ... ,n,
k=1
for,,! = ("(1, ... , "!n) E (Cn we can formulate the following error estimate
which replaces Theorem 17.1.
Theorem 17.4 The error in the Petrov-Galerkin method can be estimated

by
(17.8)
where An and Vn denote the smallest singular values of An and AnE;;:-l,

respectively, and C is some positive constant.
Proof. Using the singular value decomposition Theorem 15.16, we can
estimate
n
II<pnll 2 = L "!j'Yk (Uj,Uk) = (En"!,,,!) = (EnA;;:-lp,A;;:-lp)
j, k=l
Now we apply this inequality to the difference <p~ - <Pn and obtain (17.8)
by decomposing the error as in the proof of Theorem 17.1. D
For the least squares method as described in Theorem 13.28, i.e., for
Yn = A(Xn ), the matrix operator An assumes the form
n
(An,,!)j := L(Auk, AUjhk, j = 1, ... , n,
k=l
and is self-adjoint and positive definite. In particular, for an orthonormal

basis U1, ... ,Un of X n , we have En = I. Hence, in this case we only need to
be concerned with the eigenvalues of An. Let (/-Lm, <Pm, gm), m = 1,2, ... ,
be a singular system of the compact operator A. Then we can choose an
element <P = 2:7=1 ,,!jUj in Xn such that (<p, <Pm) = 0, m = 1, ... , n - 1,
and 11<p11 = 1. Now, by the singular value decomposition Theorem 15.16, we
obtain
/-L;:2: (A*A<p,<p) = (An"!,"!):2: An,
since Ihll2 = II<pII = 1. Therefore,
1 1
-
An>- 2
/-L '
n
This again exhibits the influence of the singular values on the degree of
ill-posedness.
For the integral equation of Example 16.21, by Theorem 15.20, the singu-
lar values decay at least exponentially. Hence, projection methods cannot
be recommended for this equation.
17.2 The Moment Method 313
17.2 The Moment Method

We proceed by describing the moment method to approximately solve an
equation of the first kind. Let A : X -+ Y be an injective bounded linear
operator from a Hilbert space X into another Hilbert space Y. Choose a
finite-dimensional subspace Y n C Y with dim Y n = n and, given an element
fEY, define the affine linear subspace Un C X by
Un := {cp EX: (Acp,g) = (f,g), 9 E Yn}.
We assume Un is nonempty. By the Riesz representation Theorem 4.8,

bounded linear functionals in a Hilbert space can be equivalently expressed
by scalar products. Therefore, we may interpret Un as the set of all elements
cp E X for which n linearly independent functionals vanish when applied
to the residual Acp - f.
Definition 17.5 An element CPn E Un is called a moment solution to
Acp = f with respect to Y n if
Note that CPn is a moment solution with respect to Yn if and only if it is

a best approximation to the zero element in X with respect to the closed
affine linear subspace Un. Finding a best approximation with respect to
an affine linear subspace can be equivalently reduced to finding a best
approximation with respect to a linear subspace. Therefore, by Theorem
1.26 there exists a unique moment solution, and by Theorem 1.25 it can be
characterized as the unique element CPn E Un satisfying
(17.9)
for all cp E Un. The following theorem gives an interpretation of the moment
method as a projection method.
Theorem 17.6 Let A : X -+ Y be an injective bounded linear operator
with dense range. Then the moment approximate solution of Acp = f with
respect to a subspace Y n C Y coincides with the Petrov-Galerkin approxi-
mation corresponding to the subspaces Xn := A*(Yn) C X and Y n C Y.
Proof. By assumption we have A(X) = Y and therefore, by Theorem 15.8,

the adjoint operator A * : Y -+ X is injective. Hence, dim Xn = dim Y n = n.
Let Pn : Y -+ Yn be the orthogonal projection operator. We will show that
the Petrov-Galerkin operator PnA : Xn -+ Y n is injective. Indeed, assume
that cp E Xn satisfies PnAcp = O. Then we can write cp = A*g for some
9 E Yn and have
IIcpll2 = (A*g,A*g) = (AA*g,g) = (PnAcp,g) = O.

The injectivity of PnA ensures that the Petrov-Galerkin approximation CPn

exists and is unique. It can be written in the form CPn = A*gn with gn E Yn ,
and it satisfies
(AA*gn,g) = (f,g) (17.10)
for all 9 E Yn . This implies CPn E Un and
(CPn - CP,CPn) = (CPn - cp,A*gn) = (ACPn - Acp,gn) =0
for all cP E Un. Therefore, by (17.9), the Petrov-Galerkin approximation
CPn is the moment solution with respect to Un. 0
When t E A(X) we write Acp = f and have (17.10) to be equivalent to

(A*gn,A*g) = (cp,A*g)
for all 9 E Yn . This implies that the moment method corresponds to the
least squares method for the adjoint (or dual) equation A*h = cP with
respect to Y n . Therefore it is also known as the dual least squares method
(for the least squares method, recall Theorem 13.28).
We continue our analysis with a convergence result on the moment
method. Note that the least squares method, in general, will not converge
(see the counterexample in Problem 17.2).
Theorem 17.7 Let A be as in Theorem 17.6 and let f E A(X). Assume
that the sequence of subspaces Yn C Y has the denseness property
inf
hEYn
IIh - gil -+ 0, n -+ 00, (17.11)
for all 9 E Y. Then the moment method converges, i.e., CPn -+ A-If,
n -+ 00.
Proof. Let Pn : Y -+ Yn and Qn : X -+ Xn = A*(Yn) be the orthogonal
projection operators. We will show that (17.11) implies convergence
QnCP -+ cP, n -+ 00,
for all cP EX. Since A is injective, by Theorem 15.8, the range A * (Y) is
dense in X. Therefore, given cP E X and e > 0, there exists 9 E Y such
that IIA*g - cpli < e/2. As a consequence of (17.11) we can choose N E IN
such that IIPng - gil < e1iAIi/2 for all n ~ N. Since A* Png E X n , for all
n ~ N we can estimate
Now the assertion of the theorem follows from the fact that for f E A(X)
the dual least squares solution CPn is given by CPn = QnA-1 f. Indeed, the
orthogonal projection satisfies
17.3 Hilbert Spaces with Reproducing Kernel 315
for all g E Yn . This is equivalent to
for all g E Yn . Hence QnA-I I solves the Petrov-Galerkin equation. 0
For the dual least squares method, from Xn = A*(Yn) we observe that
the matrix operators occurring in Theorem 17.4 are related by En = An.
Therefore, in this case the error estimate (17.8) can be simplified to
IIcp~ - A-I III::; ~ +C inf II'I/J - A-I III. (17.12)

v An ,peA*(Y.. )
17.3 Hilbert Spaces with Reproducing Kernel

For an application of the above results to collocation methods for equations
of the first kind, we obviously need a Hilbert space in which the point
evaluation functionals are bounded. Therefore, we briefly introduce the
concept of Hilbert spaces with reproducing kernel.
Definition 17.8 Let H be a Hilbert space 01 real- or complex-valued func-
tions I defined on an interval [a, b]. A function M on [a, b] x [a, b] is called
a reproducing kernel il hx := M(x,·) belongs to H lor all x E [a,b] and
(f, h x ) = I(x) (17.13)
lor all x E [a, b] and all IE H.

Theorem 17.9 A Hilbert space H has a reproducing kernel il and only il
the evaluation functionals
11-+ I(x)
are bounded lor all x E [a, b].
Prool. If H has a reproducing kernel, then (17.13) implies continuity for
the evaluation functionals. Conversely, if the evaluation functionals are
bounded, then by the Riesz Theorem 4.8 for each x E [a, b] there exists an
element hx E H such that (17.13) holds for all IE H. Hence M(x,·) := hx
is a reproducing kernel. 0
Obviously, L2[a, b] is not a Hilbert space with reproducing kernel, whereas

HI[a, b] is, since by (16.24) the maximum norm is weaker than the HI norm.
For each injective linear operator A : X --+ Y acting between two Hilbert
spaces X and Y we can make the range H:= A(X) a Hilbert space with
the scalar product given by
(17.14)
Indeed, since A is injective, (17.14) defines a scalar product on A(X). As-

sume (fn) is a Cauchy sequence in H. Then we can write fn = ACPn with
<Pn E X and find that (CPn) is a Cauchy sequence in X because
Since X is a Hilbert space, there exists cP E X with CPn -t cP, n -t 00. Then
f = Acp E Hand
Hence, with the scalar product given by (17.14) the range A(X) is a Hilbert
space. By (17.14), the linear operator A: X -t H can be seen to be bounded
with the adjoint operator A * : H -t X given by A * = A-I. In the case
of an integral operator this Hilbert space turns out to have a reproducing
kernel.
Theorem 17.10 Let A : L2[a,b] -t L2[a,b] denote an injective integral

operator
(Acp)(x) = lb K(x, y)<p(y) dy, a:::; x :::; b,
with continuous kernel K. Then the range H = A(L2[a, b]) furnished with
the scalar product (17.14) is a Hilbert space with reproducing kernel
M(x, y) = lb K(x, z)K(y, z) dz, x, Y E [a, b]. (17.15)
Proof. Define M by (17.15) and set hx := M(x,') and kx := K(x, .). Then
(Akx)(Y) = lb K(y, z)K(x, z) dz = M(x, y),
i.e., Akx = hx for all x E [a, b]. Therefore, writing f = Ar,o, we find
(f, hX)H = (A-If, A- I h x )£2 = (<p, kx)£2
= lb K(x, y)<p(y) dy = (Acp)(x) = f(x).
Hence H is a Hilbert space with reproducing kernel M. o

The statement of Theorem 17.10 remains valid for a weakly singular
kernel K with the property that M, defined by (17.15), is continuous.
For a detailed study of Hilbert spaces with reproducing kernels, we refer
to Aronszajn [6] and Meschkowski [122].
17.4 Moment Collocation 317
17.4 Moment Collocation

For the numerical solution of the integral equation of the first kind
lab K(x, y)rp(y) dy = f(x), a::; x ::; b, (17.16)
with continuous kernel K by moment collocation we choose n collocation

points Xj E [a, bj, j = 1, ... , n. We approximate the solution 'P of (17.16)
by a function 'Pn with minimal L2 norm satisfying the integral equation at
the collocation points
(17.17)
With the kernel function given in Theorem 17.10 we can rewrite (17.17) in
the form
(17.18)
Hence, the moment collocation coincides with the moment method applied
to the integral operator A: L 2 [a, bj -7 H with the subspaces Hn C H given
by Hn := span{h Xll ••• ' h xn }. In particular, from the analysis in Sections
17.2 and 17.3, we conclude that a unique solution to (17.17) with minimal
norm exists and is given by the orthogonal projection of the exact solution
'P onto the subspace
We may assume that the kernel functions K(XI,·), ... ,K(xn ,·) are lin-
early independent, since otherwise the equations (17.17) would be linearly
dependent. Then, writing
n
'Pn = L IkK(Xk,·) (17.19)
k=1
we have to solve the linear system
(17.20)
with a positive definite matrix for the coefficients 11, ... , In.
Theorem 17.11 Assume that the sequence (xn) of collocation points is

dense in [a, bj. Then the corresponding moment collocation solutions con-
verge, i.e., \\rpn - A-I f\\L2 -70, n -7 00.
Proof. By Theorem 17.7, applied to A: L2[a, b] --+ H, it suffices to show that

U := span{h xn : n E IN} is dense in H. Note that each 9 E H is continuous,
since A has continuous kernel K. Denote by P : H --+ U the orthogonal
projection operator. For 9 E H, by Theorem 1.25, we have Pg - 9 1.. U.
This implies (Pg - g, hXn)H = 0 for all n E lN, i.e., (Pg)(x n ) = g(xn) for
all n E IN. From this, by the denseness of the collocation points and the
continuity of Pg and g, it follows that 9 = Pg E U. D
For the error we can apply the special case (17.12) of Theorem 17.4.
Let Fd E <en be a perturbation of the data vector F = (J(Xl), ... , f(xn))
with IIFd - FI12 :S d and denote the moment collocation solution to these
inexact data via (17.19) and (17.20) by <p~. Then we have the following
error estimate.
Theorem 17.12 The error in the moment collocation can be estimated by
where An denotes the smallest eigenvalue of the positive definite matrix of

the linear system (17.20) and C is some positive constant.
The quality of the discretization depends on how well the exact solution
can be approximated by the subspace X n , i.e., (Qesides the smoothness of
the exact solution) it depends on the smoothness of the kernel: the smoother
the kernel, the better the approximation. On the other hand, the smooth-
ness of the kernel also controls the asymptotics of the singular values of
the integral operator and, consequently, the behavior of the eigenvalues An
(see Theorem 15.20). In general, the decay of the singular values increases
with the smoothness of the kernel. Therefore, we can expect the moment
collocation to deliver reliable approximations only for kernels with poor
smoothness. For an example see Problem 17.5.
For more details on the moment collocation, we refer to Engl [37],
Groetsch [58], and Nashed and Wahba [134], and the literature therein.
Problems
17.1 Formulate and prove a variant of Theorem 17.4 for the collocation method
of Section 13.4. Apply the result to equation (13.30).
17.2 Let A : X --+ X be a nonnegative self-adjoint compact operator with

one-dimensional nullspace N(A) = span{uo} and spectral decomposition
00
A<p = LAj(<p,Uj)Uj,
j=l
Problems 319
where Aj i- 0 for all j E IN. Consider the compact operator S : X -+ X with

S<p:= A<p+ (<p,uo)g,
where 9 E X with (g,uo) = 1 and (g,Uj) i- 0 for all j E IN. For the equation
S<p = f with right-hand side f = Ag + 9 and exact solution <p = 9 show that the
coefficients 'Yj, j = 0, ... , n, for the least squares approximation (see Theorem
13.28)
n
<pn = L 'YjUj
j=O
with respect to the subspace Xn = span{uo, ... , Un} satisfy
and
Aj{-yj - (g,Uj)} = (1- 'Yo)(g,Uj), j = 1, ... ,n.
Here Pn : X -+ Xn denotes the orthogonal projection. Use these results to
establish the inequalities
Use this to show that A and 9 can be chosen such that the least squares solutions
<pn do not converge to the exact solution (see Seidman [163]).
17.3 Show that in a Hilbert space with a reproducing kernel a weakly conver-
gent sequence is pointwise convergent. Show that a norm convergent sequence
converges uniformly provided the reproducing kernel is continuous.
17.4 Determine the reproducing kernel of the Sobolev space HI[a, bJ.
17.5 Apply the moment collocation to the Volterra integral operator of Example
15.19 with an equidistant mesh Xj = jh, j = 1, ... , n, where h = lin. Show
that the inverse matrix for the linear system (17.20) in this case is given by the
tridiagonal matrix
2 -1
-1 2 -1
-1 2 -1
n
-1 2 -1
-1 1
and that the total error in the spirit of Theorem 17.12 can be estimated by
1I<p~ _ A-I fllL2 < Cln l / 2 d + C2

- n
for some constants CI and C2 provided the exact solution is twice continuously
differentiable.
18
Inverse Boundary Value Problems
To end this book we shall briefly indicate the application of ill-posed in-
tegral equations of the first kind and regularization techniques to inverse
boundary value problems. In the ten years since the first edition of this
book was written, the monograph [25] on inverse acoustic and electromag-
netic scattering has appeared. Therefore, instead of considering an inverse
obstacle scattering problem as in the first edition, in order to introduce
the reader to current research in inverse boundary value problems we shall
consider an inverse Dirichlet problem for the Laplace equation as a model
problem. Of course, in a single chapter it is impossible to give a complete
account of inverse boundary value problems. Hence we shall content our-
selves with developing some of the main principles. For a detailed study
of inverse boundary value problems, we refer to Colton and Kress [25] and
Isakov [77].
18.1 Ill-Posed Equations in Potential Theory

Before proceeding to discuss an inverse boundary value problem, as a moti-
vation for one of the approaches we intend to present, we shall indicate the
use of ill-posed equations of the first kind in the solution of direct boundary
value problems. In our treatment of the potential theoretic boundary value
problems in Chapters 6-8 we have derived integral equations by seeking the
solutions in the form of single- or double-layer potentials over the bound-
ary of the domain in which the boundary value problem is posed. In the

IS.1 Ill-Posed Equations in Potential Theory 321
sequel we shall refer to the integral equations of the first and second kind
obtained by this approach as boundary integral equations. The boundary
integral equations of the second kind are well-posed, whereas the equations
of the first kind, such as the logarithmic single-layer integral equation,
are mildly ill-posed as indicated in Section 17.1. Due to the singularity
of the fundamental solution, the boundary integral equations, in general,
are weakly singular. Since this causes some difficulties with their numerical
solution, it is certainly tempting to try and solve the boundary value prob-
lems by using potentials with densities on curves or surfaces different from
the actual boundary of the underlying domain. We shall illustrate this idea
by looking at the interior two-dimensional Dirichlet problem.
Let D C IR? be a bounded domain with a connected boundary aD of
class 0 2. Consider the Dirichlet problem for a solution u E 02(D) n O(fJ)
of Laplace's equation
!c:,u = 0 in D (18.1)
with boundary condition
u= f on aD, (18.2)
where f is a given continuous function. We choose a closed curve r of class
0 2 such that the closure fJ is contained in the interior of r and denote
its outward unit normal by v. Then we seek the solution of the Dirichlet
problem (18.1), (18.2) in the form of a double-layer potential
u(x) =
r
lr cp(y)
aiJ>(x,y)
av(y) ds(y), xED, (18.3)
where we recall the fundamental solution to Laplace's equation

1 1
iJ>(x,y) := - In -I- I ' xi:- y,
27r X - Y
in IR? and cp E £2 (r) is an unknown density. The harmonic function u

solves the Dirichlet problem provided the density cp is a solution to the
integral equation of the first kind
r
lr cp(y)
aiJ>(x,y)
av(y) ds(y) = f(x), x E aD. (18.4)
We introduce the double-layer operator A : L2(r) ---+ L2(aD) by
(Acp)(x) :=
r
lr cp(y)
aiJ>(x, y)
av(y) ds(y), x E aD, (18.5)
and rewrite the integral equation (18.4) in the short form Acp = J. Since
the integral operator A has a continuous kernel, it is compact, and there-
fore the integral equation (18.4) is ill-posed. Furthermore, the kernel is
analytic with respect to x, and therefore (18.4) is severely ill-posed. This
322 18. Inverse Boundary Value Problems
ill-posedness can also be seen from the observation that the double-layer
potential (18.3) defines a harmonic function in all of the interior of r, and
not only in D. Hence, for a given function f, the integral equation (18.4)
can have a solution only if the solution to the Dirichlet problem in D with
boundary values f can be continued across the boundary aD as a harmonic
function into the interior of r. A closer examination shows that this condi-
tion, together with the property that the boundary data of the continued
function belong to L 2 (r), is also sufficient for the solvability. In general,
however, there are no means available to decide for given boundary values
whether such a continuation is possible.
In order to apply the regularization results of preceding chapters we
establish the following theorem.
Theorem 18.1 The compact integral operator A : L2(r) -+ L 2(aD), de-

fined by (18.5), is injective and has dense range.
Proof. Let <p satisfy A<p = 0 and define u in lR? \ r by (18.3). Then u = 0
on aD and, by the uniqueness Theorem 6.11, it follows that u = 0 in D.
The analyticity of harmonic functions from Theorem 6.6 then yields u = 0
in the interior of r. From the jump relation (6.47) for the double-layer
potential with L2 densities we derive that <p - K <p = 0, where K denotes
the double-layer boundary operator on the curve r. Since the kernel of K is
continuous (see Problem 6.1), from <p-K<p = 0 we conclude that <p E G(r).
Now Theorem 6.20 implies that <p = 0, hence A is injective.
The adjoint operator A* : L2(aD) -+ L2(r) of A is given by
When A*'¢ = 0, then, with the aid of Example 6.16, we have
(1, '¢)£2(aD) = -(AI, '¢)£2(aD) = -(1, A*'¢)£2(r) = 0,
i.e., faD '¢ ds = O. Hence, the single-layer potential

v(x) = { '¢(y)iP(x,y)ds(y), x ElR?,
laD
vanishes at infinity. Because A*'¢ = 0 implies avjav = °
on r, by the
uniqueness Theorem 6.12, we have v = 0 in the exterior of r, and by
analyticity it follows that v = 0 in the exterior of aD. Since logarithmic
single-layer potentials with L2 densities are continuous in all of lR? (see
Problem 18.1), the uniqueness for the interior Dirichlet problem tells us
that v = 0 everywhere. Then the jump relations imply '¢ = 0 (use Prob-
lems 6.1 and 8.5). Therefore A* is injective, whence, by Theorem 15.8, the
range of A is dense in L2(aD). 0
We now apply Tikhonov regularization, i.e., we approximate (18.4) by

the regularized equation
acpoe + A* Acpoe = A* f (18.6)
with some regularization parameter a > O. Note that for the solution of
(18.6) we have CPoe E C(r), since A* maps L2(8D) into C(r). As is always
the case with Tikhonov regularization, we can expect convergence of the
solution CPoe to (18.6) as a -+ 0 only when the original equation (18.4) has
a solution. Assume that CPoe -+ cP E L2(r) for a -+ o. Then, since A and A*
are continuous, it follows that A*(Acp - f) = o. Therefore Acp = f, since
A * is injective. Fortunately, in our current situation, in order to obtain
approximate solutions to the boundary value problem by
( 8iP(x,y)
uoe(x) := lr CPoe(Y) 8v(y) ds(y), x E fJ,
we do not need convergence of the densities CPoe. What we actually want

is convergence Acpoe -+ f as a -+ 0 on 8D in order to satisfy the bound-
ary condition approximately. Then, since the Dirichlet problem itself is
well-posed, small deviations U oe - f in the boundary values ensure small
deviations U oe - u in the solution in all of D. The following remark makes
this comment more precise.
Remark 18.2 For each compact set WeD there exists a constant C
(depending on D and W) such that
Ii u lic(w) :::; Cli u IiL2(8D)
for all harmonic functions u E C(fJ).
Proof. We define the double-layer potential operator V : C(8D) -+ C(D)
by
( 8iP(x,y)
(V'¢)(x) := l8D '¢(y) 8v(y) ds(y), xED.
Then, by Theorems 6.21 and -6.22, we can write
u = -2V(/ - K)-lUI8D
with the double-layer integral operator K for the boundary 8D. By the
Cauchy-Schwarz inequality, there exists a constant c depending on D and
W such that
Ii V '¢IiC(W) :::; cli'¢Ii£2(8D)
for all '¢ E C(8D). Since K has a continuous kernel, from Problem 2.5,
Theorem 6.20, and the Riesz Theorem 3.4 it follows that (/ - K)-l is
bounded with respect to the L2 norm. 0
For the convergence on the boundary we can state the following results
for an injective compact operator with dense range.
Theorem 18.3 For the unique solution CPa to the regularized equation
(18.6), we have convergence IIACPa - fll -? 0, 0 -? O.
Proof. By (16.7) this convergence is generally true for injective bounded
linear operators with dense range. 0
In general, convergence of ACPa -? f, 0 -? 0, will be slow. Therefore in

order to obtain acceptable approximations for the solution to the boundary
value problem we might have to choose the regularization parameter very
small, leading to an unsatisfactory condition number for the regularized
equation. However, the convergence rate improves in the case when Acp = f
has a solution.
Theorem 18.4 Assume that f E A(L 2(f)). Then IIACPa - fll = o(fo),
o -? O.
Proof. We write f = Acp and, by Theorem 15.23, we have CPa -? cP, 0 -? O.
Then
IIACPa - fl12 = (CPa - cP, A * [A CPa - f]) = o( cP - CPa, CPa),

which implies IIACPa - fll2 = 0(0), and the proof is complete. o
The optimal rate of convergence is described in the following theorem.

Theorem 18.5 The condition f E AA*(L2(8D)) is necessary and suffi-
cient for IIACPa - fll = 0(0), 0 -? O.
Proof. Let (J-ln, CPn, gn), n = 1,2, ... , be a singular system for the operator
A. Since A * is injective, by the singular value decomposition Theorem 15.16,
the {gn : n E IN} form a complete orthonormal system in L2(8D), i.e., we
can expand f = 2::=1 (j, gn)gn. Further from Theorem 15.23 we recall that
The expansions for f and CPa together imply that
(18.7)
Now let f = AA*g for some 9 E L2(8D). Then
(j,gn) = (AA*g,gn) = (g,AA*gn) = J-l;'(g,gn),

and from (18.7) it follows that
Conversely, assume that IIAcpQ - III = 0(0:), 0: -t O. Then from (18.7)

we observe that there exists a constant M > 0 such that
~ 1(f,gn)j2 <M
~ (0:+ J..L2)2 -
n=l n
for all 0: > O. Passing to the limit 0: -t 0, this implies that
1
L
00
4"1(f,gn)1 2 :::; M.
n=l J..Ln
Hence,
1
L
00
g := 2 (f, gn)gn
n=l J..Ln
is well defined and AA * g = f. o
The conditions IE A(L2(r)) and I E AA*(L2(aD)) ofthe two preceding

theorems can be interpreted as regularity assumptions on I because they
control how fast the Fourier coefficients (f, gn) tend to zero as n -t 00.
They may also be considered as conditions on how far the solution to the
Dirichlet problem can be continued across the boundary aD. The following
example will shed some light on these remarks.
Example 18.6 Consider the simple situation where D is the unit disk and
r a concentric circle with radius R> 1. We parameterize x E aD by
x(t) = (cost,sint), 0:::; t:::; 211",
andxErby
x(t) = (Rcost,Rsint), 0:::; t:::; 211",
and write 'I/;(t) := cp(x(t)) and g(t) .- I(x(t)). Then we transform the
integral equation (18.4) into
R (27r cos(t - 1') - R
211" 101 - 2Rcos(t _ 1') + R2 '1/;(1') d1' = g(t), (18.8)
which we rewrite in operator notation 1'1/; = g with an obvious meaning

for the self-adjoint operator 1: L2[0, 211"] -t L2[0, 211"]. By decomposing
2( cos t - R)eint eint eint
1 - 2R cos t + R2 = eit - R + e- it - R '
we derive the integrals
-1
int n = ±1,±2, ... ,
R
--
127r (cos t - R)e &=
{ 2Rl n l'
211" 0 1- 2Rcost + R2
-1, n=O.
Hence, for 'ljJn(t) = eint we have
Therefore, in polar coordinates x = (rcosO,rsinO), a singular system

(J.Ln, 'Pn, gn) of A is given by
1
J.Ln = 2Rlnl+1/2 (1 + 8no ),
for n = 0, ±1, ±2, .... The approximate solution to the Dirichlet problem
obtained by (18.6) can now be written in the form
1 00 2
U (r
0:,
0) = - " ~
'21T L....J a+ J.Ln2 (f ,9n )rlnlein8
V 4If n=-oo
compared with the exact separation of variables solution
(18.9)
The solvability condition f E A(L2(r)) now corresponds to
L
00
R 2Inl l(J,gnW < 00,

n=-oo
and therefore, by the Cauchy-Schwarz inequality, the series (18.9) can be

seen to converge uniformly in each disk with radius p < R centered at
the origin. Hence, in this case the solution to the Dirichlet problem in
the unit disk with boundary values f can be harmonically extended into
the interior of the concentric circle of radius R. The stronger condition
f E AA*(L2(8D)) corresponds to
L
00
R 4In 'l(J,gnW < 00

n=-oo
and, similarly, implies a harmonic continuation into the interior of the con-
centric circle of radius R2. 0
To illustrate our results numerically we consider the boundary condition

given by f(x) = In Ix-xol, where xo = (q,O) with q > 1. Here, the solution
to the boundary value problem clearly allows a harmonic continuation into
the open disk of radius q centered at the origin. Table 18.1 gives the error
IIA'Pa - fll between the exact solution and the Tikhonov regularization
for (18.8), approximated by Nystrom's method using the trapezoidal rule
with 32 knots. The radius of r is chosen to be R = 2. The dependence

of the error on the parameters a and q confirms the convergence behavior
predicted by our theoretical results.
This approach, of course, can be extended to cover the interior and ex-
terior Dirichlet and Neumann problem in two and three space dimensions.
Also, instead of the double-layer potential, a single-layer potential can be
used (see Problem 18.2). For an application of these ideas to the Helmholtz
equation, see Kress and Mohsen [108].
log a q= 10 q=4 q=3 q=2 q = 1.5
-5 0.00006 0.00009 0.0001 0.001 0.01

-4 0.00060 0.00091 0.0014 0.006 0.02
-3 0.00642 0.00873 0.0125 0.030 0.08
-2 0.06255 0.07370 0.0955 0.162 0.27
By Green's Theorem 6.3, for each harmonic function u in C 2 (D) nC (b)

1
we have that
f
laD
{au aip(x,y)}
ay (y) ip(x, y) - u(y) ay(y) ds(y) =0
for all x E IR? \ D. From this, we derive the ill-posed integral equation of
the first kind
fa 'IjJ(y)ip(x, y) ds(y) = faaD f(y)

aip(x, y)
a () ds(y), xEr,
aD yY
for the unknown normal derivative 'IjJ = au/ay on aD. This method to solve
the potential theoretic boundary value problems is due to Kupradze [111].
A f~rly complete description and extensive references are given by Chris-
tiansen [23].
The integral equations described in this section are simple, since as op-
posed to the boundary integral equations they possess smooth kernels.
Hence, they allow the application of simple quadrature rules without the
need to treat singular integrals. On the other hand, in general, we can ex-
pect only poor convergence as the regularization parameter tends to zero.
Therefore, this approach may serve as a very instructive example for the
use of Tikhonov regularization but will not be a serious competitor for the
numerically well-established boundary integral equations. However, in Sec-
tion 18.3 we will illustrate that integral equations of this nature have some
merits in solving inverse problems where the boundary is not known.
18.2 An Inverse Problem in Potential Theory

In this book, so far, we have considered only so-called direct boundary value
problems where, given a differential equation, its domain, and a boundary
condition, we want to determine its solution. From Chapters 6 and 9 we
know that the classical boundary value problems for the Laplace equation
and the heat equation are well-posed. However, if for a given differential
equation the problem consists of determining its domain, i.e., the boundary
of the domain, from a knowledge of the type of boundary condition and
information on one (or several) of its solutions, then this inverse boundary
value problem will often be ill-posed in any reasonable setting. Hence, the
study of inverse boundary value problems is closely interlinked with the
investigation of ill-posed problems.
To introduce the reader to inverse boundary value problems, for the
sake of simplicity, we will confine ourselves to the presentation of an in-
verse boundary value problem for the Laplace equation. Let D C IR2 be a
bounded domain with a connected boundary {JD of class C 2 • Then we con-
sider the two-dimensional exterior Dirichlet problem for a bounded function
u E C 2 (IR2 \ 15) n C(IR2 \ D) satisfying Laplace's equation
6.u =0 in IR2 \ 15 (18.10)

and the Dirichlet boundary condition
u = -if>(., z) on {JD (18.11 )
for z E IR2 \ D. Let B C IR2 be an additional bounded domain with con-
nected boundary {JB of class C2 containing the closure 15. Then the inverse
problem we want to consider is, given ujaB on {JB, determine the shape of
D, i.e., we want to construct the shape of a perfectly conducting infinite
cylinder with cross section D from the response to an incident static electric
field of a line source parallel to the cylinder axis passing through a point z
in the exterior of D. We may view this inverse problem as a model problem
for inverse boundary value problems of similar structure occurring in vari-
ous applications such as remote sensing, nondestructive testing, ultrasound
medicine and seismic imaging. As we shall illustrate in the sequel, this in-
verse boundary value problem associated with (18.10), (18.11) is nonlinear
and improperly posed. It is nonlinear, since the solution of (18.10), (18.11)
depends nonlinearly on the boundary {JD. Roughly speaking, the inverse
problem is ill-posed, since perturbations of ujaB in the maximum norm (or
L2 norm) may lead to functions that are no longer analytic and therefore
fail to be the restriction of a harmonic function on {JB. Hence, for small
perturbations of ujaB, in general, the inverse problem is not solvable, and
if it is solvable the solution cannot depend continuously on the data.
An important question to ask about the inverse problem is uniqueness,
i.e., is the boundary curve {JD uniquely determined by the values of the
harmonic function u on the curve {JB.
Theorem 18.7 Let Dl and D2 be two bounded domains with connected

boundaries aD l and aD 2 of class C2, respectively, such that th c Band
th c B. Denote by Ul and U2 the solutions to the exterior Dirichlet problem
(18.10)-{18.11) for the domains Dl and D 2, respectively, and assume that
Ul = U2 on aB. Then Dl = D 2.
Proof. Assume that Dl f:. D 2. From Ul = U2 on aB, by the uniqueness

of the solution to the two-dimensional exterior Dirichlet problem (Theo-
rem 6.11), it follows that Ul = U2 in lR? \ B. From this, since harmonic
functions are analytic (Theorem 6.6), we can conclude that Ul = U2 in the
unbounded component W of the complement of th U D2 • Without loss of
generality, we can assume that W* := (IR? \ W) \ D2 is nonempty. Then
U2 is defined in W*, since it describes the solution of (18.10), (18.11) for
D 2. The function v := cI>(., z) + U2 solves the Laplace equation in W*,
is continuous in W*, and satisfies the homogeneous boundary condition
v = 0 on aW*. This boundary condition follows from the observation that
each boundary point of W* either belongs to aD 2 or to aW n aDl. For
x E aD 2 we have cI>(x, z) + U2(X) = 0 as a consequence of the boundary
condition for U2, and for x E aW n aD l we have U2(X) = Ul(X) and there-
fore cI>(x, z) +U2(X) = (x, z) +Ul(X) = 0 as a consequence of the boundary
condition for Ul. Now, by the maximum-minimum principle for harmonic
functions we can conclude that cI>(- ,z) + U2 = 0 in W* and consequently,
by analyticity, ( • , z) + U2 = 0 in lR? \ (D2 U {z } ). Since U2 is harmonic in
all of IR2 \ th, this contradicts the singularity of the fundamental solution
at the point z, and the proof is complete. 0
Following Hahner [65), we now describe an explicit characterization of the

unknown domain D in terms of spectral data related to the known values
of U on aB. For this, in order to take care of the logarithmic behavior of the
fundamental solution at infinity, we need to slightly modify the boundary
value problem (18.10), (18.11). Without loss of generality we assume that
the closed unit disk is contained in D, i.e., we require some weak a priori
knowledge on the unknown domain D. Then we introduce the function
'T'(X,
'I'
y) . = ~ I
In Iyllx - y*l, Ix I, Iy I >_ 1,x ...I.f y,
21f Ix - y
where y* := ly(2y denotes the reflection of y across the unit circle. From
IYI2Ix-y*12 = IxI 2IyI2-2x·y+1 we observe the symmetry w(x,y) = w(y,x)
for Ixl,lyl > 1 and the boundary condition w(x, y) = 0 for Ixl = 1 and
Iyl > 1. (Note that W is the so-called Green's function for the exterior of the
unit disk.) Instead of (18.10), (18.11) we consider the Dirichlet problem for
a bounded function w(·, y) E C 2(IR2 \ D) n C(IR2 \ D) satisfying Laplace's
equation
~w(·,y)=o inIR2\D (18.12)
and the boundary condition

w(· ,y) = -w(· ,y) on aD (18.13)
for y E aBo We now want to investigate the inverse boundary value problem
of determining aD from a knowledge of w(x, y) for all x, y E aBo We note
that, analogous to Theorem 18.7, the knowledge of w(x, y) for all x E aB
and one y E aB already uniquely determines aD, i.e., the inverse problem
with data for all x, y E aB is overdetermined.
From the well-posedness of the exterior Dirichlet problem (Theorem 6.29)
we deduce that the solution w of (18.12), (18.13) is continuous on aB x aB.
Hence, the integral operator F : L2(aB) -t L2(aB) defined by
(Fg)(x) := - ( w(x, y)g(y) ds(y), x E aB,

laB
has a continuous kernel and is therefore compact. In addition, we introduce
operators S : L2(aD) -t L2(aD) and A : L2(aB) -t L2(aD) by
(S<p)(x):= ( w(x,y)<p(y)ds(y), x E aD,

laD
and
(Ag)(x):= ( w(x, y)g(y) ds(y), x E aD.
laB
Then S is compact and self-adjoint, since it has a weakly singular, real,
and symmetric kernel. The adjoint A* : L2(aD) -t L2(aB) of A is given
by
(A*<p)(x):= ( w(x,y)<p(y)ds(y), x E aB.
laD
Finally, by G : UlaD H- UlaB we denote the operator that maps the bound-
ary values on aD of bounded harmonic functions u E C(IR? \ D) onto their
a
restriction on B .
Remark 18.8 The operator G can be extended to a bounded and injective
operator G : L2(aD) -t L2(aB).
Proof. Analogous to Remark 18.2 we write G = 2V(I + K)-l, where V
represents the modified double-layer potential (6.36) on aD and K is the
modified double-layer integral operator as used in Theorem 6.24. From this,
the boundedness of G can be concluded as in Remark 18.2, and injectivity of
its extension (see Problem 2.1) follows from the uniqueness for the exterior
Dirichlet problem and analyticity with the help of the jump relation (6.47)
for the double-layer potential with L2 densities. D
Theorem 18.9 The operators F, G, and S are related through

F = GSG*, (18.14)
where G* : L2(aB) -t L2(aD) is the adjoint of G.
Proof. By definition of the operators we have A* = GS, whence
A=SG* (18.15)
follows. For the solution of (18.12), (18.13) we observe that
-w(',y)laB = GW(',y)laD
for all y E aB. From this, multiplying by 9 E L2(aB) and integrating over
aB, in view of the well-posedness of the exterior Dirichlet problem and the
continuity of G we deduce that Fg = GAg. Hence F = GA, and now using
(18.15) completes the proof of (18.14). 0
Theorem IS.10 The operator F is compact and positive definite. There

exists a complete orthonormal system gn E L2(aB) of eigenelements of F
with positive eigenvalues An, i.e.,
Fgn = Angn , n = 1,2, ....
Proof. Let rp E L2(aD) and consider the single-layer potential
v(x):= r w(x,y)rp(y)ds(y),
laD
Ixl 2: 1.
Then we apply Green's theorem and use the jump relations of Theorems
6.14 and 6.18 for L2 densities and the boundary condition v(x) = 0 for
Ixl = 1 to obtain
(Srp, rp) = r rpSrp ds laD

laD
r v {aaiL -
=
1/
aaV+ } ds =
1/
r
llxl?l
Igrad vl2dx
(compare the proof of Theorem 6.20). Hence, (Srp, <p) 2: 0 for rp E L2(aD).
Equality (Srp, rp) = 0 implies grad v(x) = 0 for alllxl 2: 1 and consequently
<p = O. Therefore, S is positive definite.
From the uniqueness results for the exterior and interior Dirichlet prob-
lem, using W(x, y) --+ 0 as Ixl --+ 00 and w(x, y) = 0 for Ixl = 1, analogous
to the proof for the injectivity of the adjoint operator A* in Theorem 18.1,
it can be seen that A is injective. Then the relation (18.15) implies that
G* is injective. Now, let 9 E L2(aB). Then
(Fg,g) = (SG*g,G*g) 2: 0,
and the positive definiteness of S and the injectivity of G* imply that F is

positive definite.
Finally, the statement on the eigenelements and eigenvalues of F is a
consequence of the spectral Theorem 15.12 for self-adjoint compact opera-
tors and the positive definiteness of F. 0
Since 8 is compact and positive definite there exists a uniquely defined

compact and positive definite operator 8 1 / 2 : L2(aD) -+ L2(aD) satisfying
8 1/ 2 8 1/ 2 = 8 (see Problem 18.4). Then the functions
._
r.pn·- 1 8 1 / 2 G* gn,
n- n= 1,2, ... , (18.16)
VAn
form a complete orthonormal system for L2(aD). The orthonormality of

the r.pn is an immediate consequence of the orthonormality of the gn. To
establish completeness, let r.p E L2 (aD) satisfy (r.p, r.pn) = 0, n = 1,2, ....
Then
and, by Theorem 1.28, the completeness of the gn implies that G8 1/ 2 r.p = 0.

From this we obtain r.p = 0, since G and 8 1 / 2 are both injective. Hence,
again by Theorem 1.28, the r.pn are complete.
Now we characterize the unknown domain D in terms of the spectral
data of the operator F, which can be deduced from a knowledge of w(x, y)
for all x, y E aB.
Theorem 18.11 The point z E B (with Izl ;::: 1) belongs to D if and only
if
1
L
00
An 1(\[1(·, z), gnW < 00. (18.17)

n=l
Proof. Let zED. Then \[1(., z) is harmonic in the exterior of D with con-
tinuous boundary values f = \[1(., z)lcw and vanishes at infinity. Therefore
G f = 9 where 9 = \[1 ( . ,z) laB. The operator S may be considered a com-
pact perturbation of the bijective operator So : CO,CX(aD) -+ C1,CX(aD)
from Theorem 7.30. Therefore, by the injectivity of S and Corollary 3.6
of the Riesz theory, we can conclude that there exists 'IjJ E CO,CX(aD) such
that S'IjJ = j, since j E C1,CX(aD). Hence we have 9 = Gf = GS'IjJ and con-
sequently 9 = GS 1/ 2 r.p, where r.p = Sl/2'IjJ E L2(aD). Using the complete
orthonormal system (18.16), we can expand r.p = L~=1(r.p,r.pn)r.pn and, by
Parseval's equality, we have
L
00
1(r.p,r.pn)1 2 = 1Ir.p1l2. (18.18)

n=l
Then
and inserting this into (18.18) yields the convergence of the series (18.17).
18.3 Approximate Solution via Potentials 333
Conversely, assume that (18.17) is satisfied. Then, by

00 1
<P:= ~~ (w(·,z),gn)<Pn
we define a function <P E L2(8D) satisfying GS1/2<p = g, where again

9 = w(·, Z)laB (compare the proof of Picard's Theorem 15.18). By the
uniqueness to the Dirichlet problem in the exterior of B and analytic-
ity, this implies that the solution to the Dirichlet problem in the exterior
of D with boundary values Sl/2<p is given by w(., z) and, in particular,
Sl/2<p = f, where again f = w(., Z)18D. Hence, z cannot belong to B \ D.
From the integrals of Lemma 8.21 it can be deduced that for z in 8D we
have w(. , Z)18D ¢. Hl/2(8D), whereas Sl/2<p E Hl/ 2(8D) for <p E L2(8D).
Therefore z must belong to D. 0
We note that in the three-dimensional case the modification (18.12),

(18.13) is not necessary (see [65]). We did not present this case here, since
its analysis requires the three-dimensional analog of Theorem 7.30 for the
single-layer integral equation of the first kind, which we did not develop in
this book.
Theorem 18.11 can be used for an approximate solution of the inverse
problem by first solving the eigenvalue problem for the operator F and
then numerically checking for a grid of points z whether the series (18.17)
converges or diverges. For numerical examples, we refer to Hahner [65]. The
corresponding analysis for the inverse obstacle scattering problem for time-
harmonic waves, i.e., for the inverse Dirichlet problem for the Helmholtz
equation, is due to Kirsch [88].
18.3 Approximate Solution via Potentials

Roughly speaking, most approaches for approximately solving an inverse
boundary value problem like our model problem (18.10), (18.11) belong to
one of the following two groups. In the first group of methods, the inverse
problem is separated into a linear ill-posed part for the construction of the
harmonic function from the given data and a nonlinear well-posed part
for finding the location of the boundary from the boundary condition. In
a second group of methods the inverse boundary value problem is either
considered as an ill-posed nonlinear operator equation or reformulated as a
nonlinear optimization problem. It is the purpose of the final two sections
of this book to outline some basic ideas of an example for each of these two
groups.
We begin with a method belonging to the first group. In its first part, we
try to represent the solution u of (18.10), (18.11) as a double-layer potential
with a density <p E L2(r) on some auxiliary closed curve r contained in the
unknown domain D. The knowledge of such an internal curve r requires

weak a priori information about D. Since we are dealing with an exterior
problem, we need to modify the double-layer potential as in Theorem 6.23,
i.e.,
u(x) = Ir <p(y) { 8:~~~~) + I} ds(y), x E IR2 \ r, (18.19)
where v denotes the outward unit normal to r. If we define an integral op-

erator A : L2(r) -+ L2(8B) by the restriction of the double-layer potential
(18.19) on 8B, i.e., by
(A<p)(x):= Ir <p(y) { 8:~~~~) + I} ds(y), x E 8B, (18.20)
then, given the values UlaB on 8B, we are looking for a density <p satisfying
the equation
A<p = UlaB. (18.21)
The integral operator A has a smooth kernel, and therefore equation (18.21)
is a severely ill-posed equation of the first kind. As for the integral equation
(18.4), the solvability can again be related to the analytic continuation of
the potential u across the boundary aD. The integral equation (18.21) is
solvable if and only if u can be extended as a harmonic function into the
exterior of r with boundary data in L2(r). As in Example 18.6 it can be
shown that the singular values of A for the special case where r is the unit
circle and B a concentric disk of radius R decay exponentially (see Problem
18.3).
Theorem 18.12 The integral operator A, defined by (18.20), is injective
and has dense range.
Proof. The proof is analogous to that of Theorem 18.1. For the injectivity
of A, the uniqueness result of the proof of Theorem 6.23 has to be used.
For the injectivity of A*, we observe that from A*'IjJ = 0, with the aid of
Example 6.16, it follows that
Hence, the single-layer potential v with density 'IjJ on aB satisfies 8v / 8v = 0

on r. Now the interior uniqueness Theorem 6.12 and analyticity imply that
v is constant in B. From the jump relation corresponding to (6.47) for the
normal derivative of the single-layer potential with L2 densities we derive
'IjJ+K''IjJ = 0, where K' denotes the normal derivative boundary operator on
the curve 8B. Now the continuity of the kernel of K' (see Problem 6.1) and
Theorem 6.20 imply that 'IjJ = 0, since we already know that JaB 'IjJ ds = O.
Hence A * is injective. 0
The Tikhonov regularization technique may be applied for the numer-

ical treatment of the ill-posed equation (18.21). Recall that solving the
regularized equation
(18.22)
with regularization parameter a > 0 is equivalent to minimizing the penal-

ized residual
/Ll(cp; a) := IIAcp - uI8Blli2(8B) + allcplli2(r)

over all cP E L2(r).
After we have determined CPo as the solution of (18.22), we use the cor-
responding double-layer potential U o = V CPo as an approximation for the
solution u of (18.10)-(18.11). Here, for convenience, we use the abbreviation
r
Vcp(x):= Jr cp(y) {a~(X'Y)
av(y) + 1} ds(y),
for the double-layer potential with density cp E L2(r) as defined through
(18.19). Then, in the second part of our method, we seek the boundary of
D as the location ofthe zeros of ~(., z)+uo in a minimum norm sense, Le.,
we approximate aD by minimizing the defect II ~(. , z) +U o II £2 (A) over some
suitable class U of admissible curves A. In the following, we will choose U
to be a compact subset (with respect to the ci?r
norm) of the set of all
starlike closed C 2 curves, described by
x(t) = r(t) (cost,sint), t E [0, 27r], (18.23)
where r E Ci?r satisfies the a priori assumption
Here, ri and r e are given functions representing curves Ai and Ae such that
the internal auxiliary curve r is contained in the interior of Ai, the boundary
aD of the unknown scatterer D is contained in the annulus between Ai
and Ae, and Ae is contained in B. For a sequence of curves we understand
convergence An --+ A, n --+ 00, in the sense that IIrn - rllc2 --+ 0, n --+ 00,
2"
for the functions rn and r representing An and A via (18.23). We say that
a sequence of functions In from C(An) is L2 convergent to a function I in
C(A) if
Jor ?r lin (rn(t) (cost,sint)) -

2
I(r(t) (cost,sint))1 2 dt --+ 0, n --+ 00.
Now we introduce the functional

Then, given CPo" we seek an approximation to 8D by minimizing the de-

fect f.l2(A, CPcJ over all A E U. Due to the continuity of f.l2(·, cp) and the
compactness of U, this minimization problem has a solution (see Problem
1.3).
Since, in general, we do not have convergence of the densities CPo as
a -* 0, for a theoretically satisfactory reformulation of the inverse boundary
value problem as an optimization problem we combine the two steps and
minimize the sum
(18.24)
simultaneously over all cP E L2(r) and A E U, i.e., we call Ao E U optimal

if there exists CPo E L2 (r) such that
f.l(CPo, Ao; a) = M(a),
where
M(a):= inf f.l(cp, A; a).
<pE£2(r),AEU
°
Here, "y > denotes a coupling parameter, which has to be chosen appro-
priately for the numerical implementation to make the two terms in (18.24)
the same magnitude. In the sequel, for theoretical purposes we may always
assume that "Y = 1. For our reformulation of the inverse boundary value
problem into a nonlinear optimization problem we can state the following
results. Note that in the existence Theorem 18.13 we need not assume that
UlaB is the exact restriction of a harmonic function.
Theorem 18.13 For each a > 0, the optimization formulation of the in-
verse boundary value problem given through (18.24) has a solution.
Proof. Let (CPn, An) be a minimizing sequence in L2(r) x U, i.e.,
Since U is compact, we can assume that An -* A E U, n -* 00. From a > 0

and
aIICPnIl12(r) :::; f.l(CPn, An; a) -* M(a), n -* 00,
we conclude that the sequence (CPn) is bounded, i.e., IICPnll£2(r) :::; c for all
n and some constant c. Hence, by Theorem 16.4, we can assume that it
converges weakly CPn -' cP E L2(r) as n -* 00. Since A : L2(r) -* L2(8B)
and V : L2(r) -* L2(A) represent compact operators, by Problem 16.5 it
follows that ACPn -* Acp and V CPn -* V cP as n -* 00. With functions rn and
r representing An and A via (18.23), by Taylor's formula we can estimate
1
8<l>(rn(t) (cost,sint),y) _ 8<l>(r(t)(cost,sint),y)
8v(y) 8v(y):::; rn t
I
LI ()- ()I
r t
for t E [0,211"] and y E f. Here, L denotes a bound on grad x {a<1>(x, y)Jav(y)}

on W x r, where W is the closed annular domain between the two curves
Ai and Ae. Then, using the Cauchy-Schwarz inequality, we find that
I(V<pn)(rn(t) (cost, sin t))-(V<pn)(r(t) (cost,sint))1 :::; cLJffilrn(t)-r(t)1
for t E [0,211"]. Therefore, from IIV<Pn - V <plli2(A) -+ 0, n -+ 00, we can
deduce that
This now implies
for n -+ 00. Since we already know weak convergence <Pn .....>. <p, n -+ 00, it
follows that
i.e., we also have norm convergence <Pn -+ <p, n -+ 00. Finally, by continuity
we obtain that
and this completes the proof. o
In order to prove regularity (in the sense of Definition 15.7) of the regu-
larization included in the above optimization reformulation of the inverse
boundary value problem, we need the following two lemmas.
Lemma 18.14 Let Ul8B be the restriction to aB of the solution u to
(18.10), (18.11) for a domain D such that aD is contained in U. Then
lim M(a) = O. (18.25)

0-+0
Proof. By Theorem 18.12, applied to the double-layer operator from L2(r)

into L2(8D), given c > 0 there exists (., z) + u = 0 on 8D we then deduce that
J.L(<p,8Dja) :::; (1 + C2 )e2 + all<plli2(r) -+ (1 + c2 )c2 , a -+ O.
Since c is arbitrary, (18.25) follows. 0
Lemma 18.15 Let (An) be a sequence of starlike curves of class C 2 that

converges to a starlike curve A of class C 2 and let Un and U be bounded
harmonic functions in the exterior of An and A, respectively. Assume that
Un and U are continuous in the closure of the exterior of An and A, respec-
tively, and that the boundary values of Un on An are L2 convergent to the
boundary values of U on A. Then Ilun - UIlC(aB) -+ 0, n -+ 00.
Proof. We use the modified double-layer potential of Theorem 6.24, i.e., we
represent u in the form
with a density cp E C(A) and, analogously, we write Un as a double-layer

potential with density CPn E C(An). With the aid of the parametric rep-
resentation (18.23) of A we transform the integral equation (6.37) into an
integral equation of the form 'ljJ + L'ljJ = f, where 'ljJ(t) = cp(x(t)) and
f(t) = 2u(x(t)) for t E [0, 21f]. Analogously, we have an integral equation
'ljJn + Ln'ljJn = fn corresponding to Un· By assumption, Ilfn - fll£2[0,211"] -+ 0,
n -+ 00.
We denote the kernel functions of Land Ln by k and k n , respectively.
By Problem 6.1, the kernel k is given by
k( ) = ~ [x'(1')]-L· [x(t) - x(1')] 21 '( )1 t =F 1',

t,1' 1f Ix(t)-x(1')1 2 +x1',
and k n is given by the analogous expression in terms of the parametric
representation xn(t) = rn(t) (cos t, sin t). As before, for a vector a = (al, a2)
in m? we denote a-L = (a2, -al). From Taylor's formula
x(t) - x(1') = (t - 1') 11 x'(1' + A(t - 1')) dA (18.26)
we observe that the function

Ix(t) - x(1')1
t =F 1',
g(t,1'):= { It-1'1
Ix'(t)l, t=1',
is continuous in m? Therefore, in view of x(t) =F x(1') for 0 < It - 1'1 ::; 1f

and Ix'(t)12 = [r(tW + [r'(t)J2 > 0 for all t, there exists a constant Cl,
depending on r, such that
Ix(t) - x(1')1 ~ cdt - 1'1, It - 1'1::; 1f. (18.27)
From this, using (18.26) applied to Xn - x, we deduce that
I[xn(t) - xn(1')]- [x(t) - x(r)JI ::; c211rn - rllqJx(t) - x(r)1 (18.28)

for It - rl ::; 7f and some constant C2 depending on 1'. By Taylor's formula

we have
1 I 16
IIYnl2
1
- iYf2 ::; iYf3 IYn - YI, Y ~ 0, 21Yn - YI ::; IYI·
From this, setting Y = x(t) - x(r), Yn = Xn(t) - Xn(r), and using (18.27)
and (18.28) we find the estimate
I______ 1 _ 1 I< C II rn- rllc 1

2"
(18.29)
IXn(t) - Xn(r)12 Ix(t) - x(r)12 - 3 (t - r)2 ,
which is valid for sufficiently small IIrn - l' II Ci" , for It - rl ::; 7f, and for
some constant C3 depending on r. Using Taylor's formula
X(t) - x(r) = (t - r)x'(r) + (t - r)21 (1 - A)x"(r + A(t - r)) dA,

1
we can write
[x'(r)]J. . [x(t) - x(r)] = (t - r)2 [x'(r)]J. .1 1

(1- A)x"(r + A(t - r)) dA.
Hence, we can estimate
and
for all It - r I ::; 7f and some constants C4 and C5 depending on r. Combining

these two estimates with (18.27)-(18.29), by decomposing
kn(t,r) - k(t,r) = 2Ix~(r)l- 2Ix'{r)1
1 [x~(r)V' [xn(t) - xn(r)]- [x'(r)]J.· [x(t) - x(r)]

+; Ix(t) - x(r)12
+~ [x~(r)]J. . [xn{t) - xn(r)] {Ixn{t) ! xn{r)12 - Ix{t) -\(r) 12 },
we finally obtain that
(18.30)
for sufficiently smallll1'n - 1'llc2 and some constant C depending on r.

2"
The estimate (18.30) now implies norm convergence
Therefore, using Theorem 10.1 we can conclude that II'I/'n - '1/'11£2[0,271"] --70,
n --7 00. From this, the statement of the lemma follows by parameterizing
the double-layer potential representation for Un - u and then using the
Cauchy-Schwarz inequality. 0
Now we are ready to establish regularity through the following conver-

gence result.
Theorem 18.16 Let (an) be a null sequence and let (An) be a correspond-
ing sequence of solutions to the optimization problem with regularization
parameter an. Assume that UloB is the restriction to aB of the solution
u to (18.10), (18.11) for a domain D such that aD belongs to U. Then
An --7 aD, n --7 00.
Proof. Since by assumption U is compact, there exists a convergent sub-

sequence An(k) --7 A * E U, k -+ 00. Let u* denote the unique solution to
the boundary value problem (18.10), (18.11) for the exterior domain with
boundary A*, i.e., the boundary condition reads
(·,z)+u*=o onA*. (18.31)
Since An(k) is optimal for the parameter an(k), there exists IPn(k) E L2(r)
such that J1(IPn(k) , An(k); an(k») = M(an(k») for k = 1,2, .... By Lemma
18.14, we have
11<1>(· ,z) + VIPn(k)1112(A n (k) ~ M(an(k») --70, k --7 00. (18.32)
Now, by Lemma 18.15, from (18.31) and (18.32), we deduce that
IIAIPn(k) - u* 11£2(oB) --70, k --700.
By Lemma 18.14 we also have
Therefore, we conclude u = u* on aB, and by Theorem 18.7 we have

aD = A*.
Now, by a standard argument, assuming that the sequence (An) does not
converge to aD leads to a contradiction, since by the above argument any
subsequence An contains a subsequence that converges to aD. 0
The extension of the above method to the inverse obstacle scattering

problem for time-harmonic waves, i.e., for the inverse Dirichlet boundary
value problem for the Helmholtz equation including numerical examples is
due to Kirsch and Kress. For further discussion of the method including
numerical examples, we refer to Kirsch and Kress [89, 90, 91], Kirsch et
al. [92), Kress [103), and the monograph [25).
18.4 Differentiability with Respect

to the Boundary
We now turn to an approximation method of the second group as mentioned
at the beginning of Section 18.3. The solution to the boundary value prob-
lem (18.10), (18.11) defines an operator F : 8D I-t ulc/B that maps the
boundary 8D onto the restriction of the solution u to 8B. In terms of
this operator, given Uli7B, the inverse boundary value problem consists of
solving the equation
F(8D) = Ul8B (18.33)
for the unknown boundary curve 8D. Hence, it is, quite natural to try some
of the iterative methods developed for the solution of nonlinear equations
such as Newton's method. For this, one needs to establish differentiability
of the operator F and to characterize the derivative.
A mapping M : U C X -+ Y from an open subset of a normed space
X into a normed space Y is said to be Frechet differentiable at the point
rp E U if there exists a bounded linear operator M' (rp, .) : X -+ Y such
that
~~ "!,, "M(rp + h) - M(rp) - M'(rp, h)" = O.
The bounded linear operator M'(rp,,) is called the Frechet derivative of M
at the point rp. The mapping M is called Frechet differentiable on U if it
is Frechet differentiable for all rp E U. For an introduction to the concept
of the Frechet derivative and its basic properties we refer to Berger [15].
So far we have not specified a domain of definition for the operator F.
For this, a parameterization of the boundary curve is required. For the sake
of simplicity, as in Section 18.3, we will restrict our analysis to the case of
starlike domains of the form (18.23) and consider F as a mapping from the
open set U := {r E Ci7r : r > O} of positive functions in Ci7r into C(8B).
We will also write F(r) synonymously for F(8D).
We now establish differentiability of F with respect to the boundary.
For this we first need to examine the corresponding differentiability of the
boundary integral operators used in the existence proof for the solution of
the Dirichlet problem.
We recall the double-layer integral operator K : C(8D) -+ C(8D) from
Theorem 6.24 and, as in the proof of Lemma 18.15, via
x(t) = r(t) (cost,sint), t E [0,271'],
we introduce its parameterized form
(Lr'IjJ)(t) = ior27r k(t,Tjr)'IjJ(T)dT, t E [0,271'],
with kernel
k(t, Tj r) = 2 [x' (T )]..L . grady (x(t) , x( T)) + 21x' (T) I, t =f. T
(see Problem 6.1). The subscript indicates the dependence of Lr on the

boundary curve.
Theorem 18.11 The mapping r t-+ Lr is Frechet differentiable from
U C 0§1I" into the space L(0211" , 0 211") of bounded linear operators from 0 211"
into 0 211".
Proof. For t -I- r, the Fnkhet derivative of r t-+ k(·,·; r) is given by
k '( t,r;r,q ) =k h r ))~ ·grady<p (x ()

-,( t,r;r,q ) +2 ['( r +2 x'(r)
t ,x ()) . h'(r) '
IX'(r)1
where
k'(t,r;r,q) = 2[x'(r))~. <p~(x(t),x(r)) [h(r) - h(t))
with <p~(x, y) denoting the Hessian of <p(x, y) with respect to y. The func-
tion q E 0§1I" is such that r + q > 0 and we have set
h(t) = q(t) (cost,sint), t E [0, 21f).

To establish this, proceeding as in the proof of the estimate (18.30) in
Lemma 18.15, straightforward calculations using Taylor's formula show
that k' (. , . ; r, q) is continuous on IR? and 21f-periodic with respect to both
variables and satisfies
Ilk(·,·; r + q) - k(·,·; r) - k'(·,·; r, q)lloo ~ c Ilqll~22.- (18.34)
for all sufficiently small Ilqll~2 and some constant c depending on r (see
2.-
Problem 18.5). From this, it follows that the integral operator L~,q with
kernel k' (. , . ; r, q) satisfies
IILr+q - Lr - L~,qlloo ~ 21fC Ilqll~?.-,

and therefore the linear mapping q t-+ L~,q is the Fnkhet derivative of
r t-+ L r . 0
The parameterized form of the double-layer operator is given by the

integral operator
(Vr'lj;)(Z) r 27r
= 10 v(z,r;r)'Ij;(r) dr, z E IR2 \ fJ,
with kernel
2 -
v(z,r;r) := [x'(r)]~. grady <P(z,x(r)) + Ix'(r)l, z E IR \ D, r E [0,271-]'
We introduce the integral operator V:,q with kernel
~ x' (r) . h' (r )

v' (z, r; r, q) = Vi (z, r; r, q) + [h'(r)] . grady <p(z, x(r)) + IX'( r)1
where
v'(z,rjr,q) = [x'(r))l..· ~(z,x(r)) her).
Then, since the kernel of Vr has no singularities, straightforward differen-
tiation shows that for closed subsets W of JR2 \ jj the Frechet derivative
of the mapping r M Vr from U C C?7r into the space L(C27r , C(W)) of
bounded linear operators from C27r into C(W) is given by q MV;,q.
Theorem 18.18 Assume that '¢ E ci~"'. Then V;,q'¢ is a bounded har-
monic function in JR2 \ jj that is continuous in JR2 \ D with boundary
values
,
v;.,q~ + h· grad ( ) = 2"1 Lr,q'¢
Vr,¢ , on aD. (18.35)
Proof. From the form of the kernel v'(·,· j r, q), it is obvious that V;,q'¢
is bounded at infinity and harmonic. To show that it is continuous up to
the boundary aD and to compute its boundary values we use the vector
identity
grad div w = c:"w + [graddivwl..)l..,
Laplace's equation for , and grady = - grad x to transform
v'(z,rjr,q) = [x'(r))l..·gradzdivz{(z,x(r))h(r)}
= -divz{[h(r))l.. :r (z,x(r))}
(compare the proof of Theorem 7.29). From this, by partial integration, for
z E JR2 \ jj we find that
r 27r
(V:,q~)(z) = div z 10 (z,x(r)) [h(r))l..~'(r) dr+ 10
r 27r x'(r) . h'(r)
IX'(r)1 '¢(r) dr,
i.e., we can interpret V;,q'¢ as a derivative of a single-layer potential plus a

constant. Therefore, from Theorem 7.28 it follows that V;,q'¢ is continuous
up to the boundary with boundary values
(V;,q~)(z) = - fo27r grady (x(t),x(r)) . [h(r))l..~'(r)dr
+ 10
r27r x'(r) . h'(r)
IX'(r)1
,¢'(t),
~(r) dr - 2Ix'(t)1 x (t) . h(t)
for z = x(t) E aD. Proceeding analogously for the gradient of the double-
layer potential, i.e., from Theorem 7.29, we find that
h(t)· grad(Vr~)(z(t)) = fo27r grady (x(t),x(r))· [h(t))l..'¢'(r) dr

'¢' (t) ,
+ 2Ix'(t)1 x (t)· h(t), z = x(t) E aD.
Adding the last two equations and performing an analogous partial inte-
gration in the integral representing L~,q'¢ now yields (18.35). For the latter
partial integration one has to take proper care of the singularity at t = T. 0
Note that (18.35) can be interpreted as formal differentiation of the jump

relation 2Vr ,¢ = (1 + Lr)'¢ with respect to r. Analogously, the boundary
condition (18.36) in the following theorem can be obtained by formally
differentiating the boundary condition (18.11) with respect to r.
Theorem 18.19 The mapping F : r H- UlaB is Prechet differentiable from
U C C~7r into C(8B). The Prechet derivative is given by F'(r,q) = WlaB,
where wE C 2 (rn?\D)nC(rn?\D) denotes the uniquely determined bounded
harmonic function satisfying the Dirichlet boundary condition
w=-grad{<P(·,z)+u}·h on8D. (18.36)
Proof. In view of Theorem 6.24, we write the solution to (18.10), (18.11)

in the form
u = -2Vr (I + Lr)-l fr in rn? \ D,
where fr(t) := <P(x(t), z). We note that differentiability of r H- Lr im-
plies differentiability of r H- . (1 + Lr)-l with the derivative given by
-(1 + Lr)-lL~,q(1 + Lr)-l. Then, using the chain rule for the Frechet
derivative, we obtain
u' = -2V:,q(I+Lr )-1 fr+2Vr(I+Lr)-1 L~,q(I+Lr)-lfr-2Vr(I+Lr)-lbr,q
in rn? \ D, where br,q(t) = h(t) . grad <P(x(t), z) represents the Frechet

derivative of the boundary values with respect to the boundary, Le., of
r H- fro From this, it is obvious that u' is bounded and harmonic. To derive
the boundary condition (18.36) we use the jump relation 2Vr (I +Lr)-l = 1
on 8D for the double-layer potential. Now (18.36) follows, since (18.35)
implies that
-2V:,q(I + Lr)-l frlaD = -L~,q(I + Lr)-l fr - h· gradulaD'
This completes the proof, provided we justify the application of (18.35) by

showing that (I + Lr)-l fr E C~;."'.
Proceeding analogously to the proof of Theorem 7.5 it can be shown
that the double-layer integral operator K maps C(8D) boundedly into
CO''''(8D) and that it maps CO''''(8D) boundedly into C 1''''(8D) (see also
Theorems 7.27 and 7.28). For a proof in the three-dimensional case we refer
to [24]. Then, from <P(. ,Z)18D E C 1''''(8D), it follows that the solution of
the integral equation (6.37) is in C 1 ''''(8D). 0
For the extension of the above analysis on Frechet differentiability to the

Helmholtz equation, we refer to Potthast [146] and the monograph [25].
Problems 345
Theorem 18.19 may be considered the theoretical foundation for the ap-
plication of Newton's method and related iteration methods for the ap-
proximate solution of (18.33). In this method, given ulcm, the nonlinear
equation
F(r) = UlaB
is replaced by its linearization
F(r) + F'(r, q) = UlaB, (18.37)

which has to solved for q in order to improve an approximate boundary
r
given by r into a new approximation given by = r + q. As usual, New-
ton's method consists of iterating this procedure. From the form of the
Frechet derivative given in Theorem 18.19 it can be concluded that the
linear operator q 1-+ F'(r,q) is compact from L2[0,27r] into C(8B). There-
fore, for the solution of the linear equation (18.37) a regularization has to
be incorporated, such as Tikhonov regularization. For a description of the
numerical implementation of Newton's method for the case of inverse ob-
stacle scattering, i.e., for the inverse Dirichlet problem for the Helmholtz
equation we refer to Kress [102, 103, 105] and the monograph [25].
When this was written, no satisfactory convergence analysis for regu-
larized Newton iteration methods for the solution of (18.33) for inverse
boundary value problems had been completed. However, for preliminary
results in the case of inverse obstacle scattering, we refer to Hohage [72].
Comparing the approaches of Section 18.3 and 18.4, methods for solving
the operator equation (18.33) directly by a Newton-type method in general
require the solution of the direct problem for the evaluation of F and its
derivative in each iteration step. Hence, these methods are more costly than,
for example, the method of Section 18.3, because the latter requires only
computation of a potential for the evaluation of the cost functional in each
iteration step. However, a comparison of the numerical results obtained by
methods of the two groups reveals that the direct approach using the opera-
tor equation (18.33) leads to more satisfactory results for the reconstruction
of the unknown boundary. Although more research has to be carried out on
the approximate solution of inverse boundary value problems, the author
expects that the direct approach to solve (18.33), in general, will turn out
to be the most effective.
Problems
18.1 Show that the logarithmic single-layer potential with L2 density is contin-
uous in all of IR? .
18.2 Analogous to Theorem 18.1, construct a single-layer potential operator
that is injective and has dense range.
Hint: Take care of the logarithmic behavior at infinity as in (7.45).
18.3 Analogous to Example 18.6, determine a singular value decomposition of

the operator A defined by (18.20).
18.4 Let A : X -+ X be a compact and positive definite operator in a Hilbert

space X. Show that for each r > 0 there exists a uniquely defined compact and
positive operator AT: X -+ X such that AT = A for r = 1 and A T+ s = AT AS for
all r, s ~ o.
Hint: Define AT tp := 2::::'=1
)t (tp, tpn)tpn in terms of the spectral decomposition
from Theorem 15.12.
18.5 For the kernel of the Fn§chet derivative of the double-layer integral oper-
ator show that
k' (t, T; r, q) =-~ [x' (T)].L . {x(t) - x(T)} {x(t) _ x( T)} . {h(t) - h(T)}
1'i Ix(t)-x(r)1 4
1 [h'(T)].L· {x(t) - X(T)} + [X'(T)].L. {h(t) - h(T)} 2 X'(T)· h'(T)

+;;: Ix(t) - X(TW + IX'(T)I
for t # T. Use this representation to show that k'(·,·;r,q) is continuous and to

verify the inequality (18.34).
References
[1] Adams, R.A.: Sobolev Spaces. Academic Press, New York 1975.
[2] Anderssen, R.S., de Hoog, F.R., and Lukas, M.A.: The Application
and Numerical Solution of Integral Equations. Sijthoff and Noordhoff,
Alphen aan den Rijn 1980.
[3] Anselone, P.M.: Collectively Compact Operator Approximation The-
ory and Applications to Integral Equations. Prentice-Hall, Englewood
Cliffs 1971.
[4] Anselone, P.M., and Moore, R.H.: Approximate solutions of integral
and operator equations. J. Math. Anal. Appl. 9, 268-277 (1964).
[5] Arnold, D.N., and Noon, D.N.: Coercitivity of the single layer heat
potential. J. Comput. Math. 7, 100-104 (1989).
[6] Aronszajn, N.: Theory of reproducing kernels. Trans. Amer. Math.
Soc. 68, 337-404 (1950).
[7] Atkinson, F.V.: Normal solvability of linear equations in normalized
spaces. Matern. Sbornik 28, 3-14 (1951) (In Russian).
[8] Atkinson, K.E.: Iterative variants of the Nystrom method for the
numerical solution of integral equations. Numer. Math. 22, 17-31
(1973).
[9] Atkinson, K.E.: The Numerical Solution of Integral Equations of the
Second Kind. Cambridge Univ. Press, Cambridge 1997.
348 References
[10] Aubin, J.P.: Applied Functional Analysis. John Wiley & Sons, New
York 1979.
[11] Baderko, E.A.: Parabolic problems and boundary integral equations.

Math. Meth. in the Appl. Sci. 20, 449-459 (1997).
[12] Baker, C.T.H.: The Numerical Treatment of Integral Equations.

Clarendon Press, Oxford 1977.
[13] Baumeister, J.: Stable Solution of Inverse Problems. Vieweg, Braun-

schweig 1986.
[14] Beer, A.: Einleitung in die Elektrostatik, die Lehre vom Magnetismus
und die Elektrodynamik. Vieweg, Braunschweig 1865.
[15] Berger, M.S.: Nonlinearity and Functional Analysis. Academic Press,
New York 1977.
[16] Beylkin, G., Coifman, R.R., and Rokhlin, V.: Fast wavelet transforms
and numerical algorithms. Comm. Pure Appl. Math. 44, 141-183
(1991).
[17] Brakhage, H.: Uber die numerische Behandlung von lntegralgleichun-
gen nach der Quadraturformelmethode. Numer. Math. 2, 183-196
(1960).
[18] Brebbia, C.A., Telles, J.C.F., and Wrobel, L.C.: Boundary Element
Techniques. Springer-Verlag, Berlin 1984.
[19] Calderon, A.P.: The multipole expansion of radiation fields. J. Rat.
Mech. Anal. 3, 523-537 (1954).
[20] Cannon, J.R.: The One-Dimensional Heat Equation. Addison-Wesley,

Menlo Park 1984.
[21] Cea, J.: Approximation variationelle des problemes aux limites. Ann.
lnst. Fourier 14, 345-444 (1964).
[22] Chen, G., and Zhou, J.: Boundary Element Methods. Academic Press,
London 1992.
[23] Christiansen, S.: Condition number of matrices derived from two

classes of integral equations. Math. Meth. in the Appl. Sci. 3, 364-392
(1981).
[24] Colton, D., and Kress, R.: Integral Equation Methods in Scattering
Theory. Wiley-Interscience Publication, New York 1983.
[25] Colton, D., and Kress, R.: Inverse Acoustic and Electromagnetic Scat-
tering Theory, 2nd ed. Springer-Verlag, Berlin 1998.
References 349
[26] Costabel, M.: Boundary integral operators for the heat equation. In-
tegral Eq. and Operator Theory 13, 498-552 (1990).
[27] Courant, R., and Hilbert, D.: Methods of Mathematical Physics II.
Wiley-Interscience Publication, New York 1966.
[28] Cryer, C.W.: Numerical Functional Analysis. Clarendon Press, Ox-

ford 1982.
[29] Dahmen, W.: Wavelet and multiscale methods for operator equations.
Acta Numerica 6, 55-228 (1997).
[30] Davis, P.J.: On the numerical integration of periodic analytic func-

tions. In: Symposium on Numerical Approximation (Langer, ed.).
Univ. of Wisconsin Press, Madison, 45-59 (1959).
[31] Davis, P.J.: Interpolation and Approximation. Blaisdell Publishing

Company, Waltham 1963.
[32] Davis, P.J., and Rabinowitz, P.: Methods of Numerical Integration,

2nd ed. Academic Press, San Diego 1984.
[33] Delves, L.M., and Mohamed, J.L.: Computational Methods for Inte-
gral Equations. Cambridge Univ. Press, Cambridge 1985.
[34] Dieudonne, J.: Sur les homomorphismes d'espaces normes. Bull. Sci.
Math. 67, 72-84 (1943).
[35] Dieudonne, J.: The index of operators in Banach spaces. Integral Eq.
and Operator Theory 8, 580-589 (1985).
[36] Engels, H.: Numerical Quadrature and Cubature. Academic Press,

New York 1980.
[37] Engl, H.W.: On the convergence of regularization methods for iU-

posed linear operator equations. In: Improperly Posed Problems and
Their Numerical Treatment (Hammerlin and Hoffmann, eds.). Inter-
national Series of Numerical Mathematics. Birkhauser-Verlag, Basel
63, 81-96 (1983).
[38] Engl, H.W., Hanke, M., and Neubauer, A.: Regularization of Inverse
Problems. Kluwer Academic Publisher, Dordrecht 1996.
[39] Fenyo, 1., and Stolle, H.: Theorie und Praxis der linearen Integral-
gleichungen, I-IV. Birkhauser-Verlag, Basel 1981-1984.
[40] Fichera, G.: Linear elliptic equations of higher order in two indepen-
dent variables and singular integral equations. In: Proc. Conf. Partial
Diff. Equations. Univ. of Wisconsin Press 1961.
350 References
[41] Folland, G.: Introduction to Partial Differential Equations. Princeton

Univ. Press, Princeton 1976.
[42] Fredholm, 1.: Sur une classe d'equations fonctionelles. Acta Math.
27, 365-390 (1903).
[43] Freud, R., Golub, F., and Nachtigal, N.: Iterative solutions of linear
systems. Acta Numerica 1, 57-100 (1992).
[44] Fridman, V.: Method of successive approximation for a Fredholm
integral equation of the first kind. Uspeki Mat. Nauk. 11, 233-234
(1956) (In Russian). .
[45] Friedman, A.: Partial Differential Equations of Parabolic Type.
Prentice-Hall, Englewood Cliffs 1964.
[46] Gakhov, F.D.: Boundary Value Problems. Pergamon Press, Oxford
1966.
[47] Galerkin, B.G.: Expansions in stability problems for elastic rods and
plates. Vestnik inzkenorov 19, 897-908 (1915) (In Russian).
[48] Gevrey, M.: Sur les equations aux derivees partielles du type
parabolique. J. de Math. 9,305-471 (1913).
[49] Gilbarg, D., and Trudinger, N.8.: Elliptic Partial Differential Equa-
tions of Second Order. Springer-Verlag, Berlin 1977.
[50] Gilbert, R.P.: Function Theoretic Methods in Partial Differential
Equations. Academic Press, New York 1969.
[51] Golberg, M.A., and Chen, C.S.: Discrete Projection Methods for In-
tegral Equations. Computational Mechanics Publications, Southamp-
ton 1997.
[52] Golub, G., and van Loan, C.: Matri4; Computations. Johns Hopkins
University Press, Baltimore 1989.
[53] Greenbaum, A.: Iterative Methods for Solving Linear Systems. SIAM,
Philadelphia 1997.
[54] Greengard, L., and Moura, M.: On the numerical evaluation of elec-
trostatic fields in composite materials. Acta Numerica 3, 379-410
(1994).
[55] Greengard, L., and Rokhlin, V.: A fast algorithm for particle simu-
lations. J. Comput. Phys. 73, 187-207 (1987).
[56] Greengard, L., and Rokhlin, V.: A new version of the fast multipole
method for the Laplace equation in three dimensions. Acta N umerica
6, 229-269 (1997).
References 351
[57] Grisvard, P.: Elliptic Problems in Nonsmooth Domains. Pitman,

Boston 1985.
[58] Groetsch, C.W.: The Theory of Tikhonov Regularization for Fredholm
Equations of the First Kind. Pitman, Boston 1984.
[59] Hackbusch, W.: Multi-Grid Methods and Applications. Springer-
Verlag, Berlin 1985.
[60] Hackbusch, W.: Integral Equations: Theory and Numerical Treat-
ment. Birkhauser-Verlag, Basel 1994.
[61] Hackbusch, W., and Nowak, Z.P.: On the fast matrix multiplication
in the boundary element method by panel clustering. Numer. Math.
56, 229-245 (1989).
[62] Hackbusch, W., and Sauter, S.: On the efficient use of the Galerkin
method to solve Fredholm integral equations. Appl. Math. 38, 301-
322 (1993).
[63] Hadamard, J.: Lectures on Cauchy's Problem in Linear Partial Dif-
ferential Equations. Yale University Press, New Haven 1923.
[64] Hahner, P.: Eindeutigkeits- und Regularitatssatze fur Randwertprob-
Ierne bei der skalaren und vektoriellen Helmholtzgleichung. Disserta-
tion, G6ttingen 1990.
[65] Hahner, P.: An inverse problem in electrostatic. Inverse Problems (to
appear).
[66] Helms, L.: Introduction to Potential Theory. Wiley-Interscience, New
York 1969.
[67] Hemker, P.W., and Schippers, H.: Multiple grid methods for the solu-
tion of Fredholm integral equations of the second kind. Math. Compo
36, 215-232 (1981).
[68] Henrici, P.: Applied and Computational Complex Analysis. Vol 3.
Wiley-Interscience Publication, New York 1986.
[69] Hestenes, M.R., and Stiefel, E.: Methods of conjugate gradients for
solving linear systems. J. Res. Nat. Bur. Standards 49, 409-436
(1952).
[70] Heuser, H.: Funktionalanalysis, 2. Auff. Teubner-Verlag, Stuttgart
1986.
[71] Hilbert, D.: Uber eine Anwendung der Integralgleichungen auf ein
Problem der Funktionentheorie. In: Verhandlungen des dritten inter-
nationalen Mathematiker Kongresses in Heidelberg 1904. Teubner,
Leipzig, 233-240 (1905).
352 References
[72] Hohage, T.: Logarithmic convergence rates of the iteratively regular-

ized Gauss-Newton method for an inverse potential and an inverse
scattering problem. Inverse Problems 13, 1279-1299 (1997).
[73] Holmgren, E.: Sur l'equation de la propagation de la chaleur. Arkiv

Mat. Fysik. 4, 14, 1-11 (1908).
[74] Holmgren, E.: Sur l'equation de la propagation de la chaleur II. Arkiv

Mat. Fysik. 4, 18, 1-28 (1908).
[75] Hsiao, G.C., and MacCamy, R.C.: Solution of boundary value prob-
lems by integral equations of the first kind. SIAM Review 15, 687-705
(1973).
[76] Hsiao, G.C., and Saranen, J.: Boundary integral solution of the two-
dimensional heat equation. Math. Meth. in the Appl. Sciences 16,
87-114 (1993).
[77] Isakov, V.: Inverse Problems for Partial Differential Equations.

Springer-Verlag, New York 1998.
[78] Ivanov, V.K.: Integral equations of the first kind and an approximate
solution for the inverse problem of potential. Soviet Math. Doklady
3, 210-212 (1962) (English translation).
[79] J6rgens, K.: Lineare Integraloperatoren. Teubner-Verlag, Stuttgart

1970.
[80] Kantorovic, L.V., and Akilov, G.P.: Functional Analysis in Normed
Spaces. Pergamon Press, Oxford 1964.
[81] Kantorovic, L.V., and Krylov, V.I.: Approximate Methods of Higher
Analysis. Noordhoff, Groningen 1964.
[82] Kellogg, 0.: Foundations of Potential Theory. Dover, New York 1954.
[83] Kersten, H.: Grenz- und Sprungrelationen fur Potentiale mit quadrat-
summierbarer Dichte. Resultate d. Math. 3, 17-24 (1980).
[84] Kirsch, A.: Generalized boundary value- and control problems for the
Helmholtz equation. Habilitationthesis, G6ttingen 1984.
[85] Kirsch, A.: Private communication, 1986.
[86] Kirsch, A.: Surface gradients and continuity properties for some inte-
gral operators in classical scattering theory. Math. Meth. in the Appl.
Sci. 11, 789-804 (1989).
[87] Kirsch, A.: An Introduction to the Mathematical Theory of Inverse

Problems. Springer-Verlag, New York 1996.
References 353
[88) Kirsch, A.: Characterization of the shape of the scattering obstacle

by the spectral data of the far field operator. Inverse Problems 14,
1489-1512 (1998).
[89) Kirsch, A., and Kress, R.: On an integral equation of the first
kind in inverse acoustic scattering. In: Inverse Problems (Cannon
and Hornung, eds.). International Series of Numerical Mathematics.
Birkhauser-Verlag Basel 77, 93-102 (1986).
[90) Kirsch, A., and Kress, R.: A numerical method for an inverse scat-
tering problem. In: Inverse Problems (Engl and Groetsch, eds.). Aca-
demic Press, Orlando, 279-290 (1987).
[91) Kirsch, A., and Kress, R.: An optimization method in inverse acous-
tic scattering. In: Boundary Elements IX, Vol. 3. Fluid Flow and
Potential Applications (Brebbia et al., eds.). Springer-Verlag, Berlin
Heidelberg New York, 3-18 (1987).
[92) Kirsch, A., Kress, R., Monk, P., and Zinn, A.: Two methods for solv-
ing the inverse acoustic scattering problem. Inverse Problems 4,749-
770 (1988).
[93) Kirsch, A., and Ritter, S.: The Nystrom method for solving a class
of singular integral equations and applications in 3D-plate elasticity.
Math. Meth. in the Appl. Sci. (to appear).
[94) Kral, J.: The Fredholm method in potential theory. Trans. Amer.
Math. Soc. 125,511-547 (1966).
[95) Krasnoselski, M.A., Vainikko, G.M., Zabreko, P.P., and Rutitskii, Y.:
Approximate Solution of Operator Equations. Wolters-Noordhoff,
Groningen (1972).
[96) Kress, R.: Ein ableitungsfreies Restglied fur die trigonometrische In-
terpolation periodischer analytischer Funktionen. Numer. Math. 16,
389-396 (1971).
[97) Kress, R.: Zur numerischen Integration periodischer Funktionen nach
der Rechteckregel. Numer. Math 20,87-92 (1972).
[98) Kress, R.: A boundary integral equation method for a Neumann
boundary value problem for force-free fields. J. of Eng. Math. 15,
29-48 (1981).
[99) Kress, R.: On the Fredholm alternative. Integral Equations and Op-
erator Theory 6, 453-457 (1983).
[100) Kress, R.: A Nystrom method for boundary integral equations in
domains with corners. Numer. Math. 58, 145-161 (1990).
354 References
[101] Kress, R.: Fast 100 Jahre Fredholmsche Alternative. In: Uberblicke
Mathematik. Jahrbuch 1994 (Chatteri et al., eds.). Vieweg, Braun-
schweig, 14-27 (1994).
[102] Kress, R.: A Newton method in inverse obstacle scattering. In: In-
verse Problems in Engineering Mechanics (Bui et al., eds.). Balkema,
Rotterdam 425-432 (1994).
[103] Kress, R.: Integral equation methods in inverse obstacle scattering.

Eng. Anal. with Boundary Elements 15, 171-179 (1995).
[104] Kress, R.: On the numerical solution of a hypersingular integral equa-

tion in scattering theory. J. Compo Appl. Math. 61, 345-360 (1995).
[105] Kress, R.: Integral equation methods in inverse acoustic and electro-
magnetic scattering. In: Boundary Integral Formulations for Inverse
Analysis (Ingham and Wrobel, eds.). Computational Mechanics Pub-
lications, Southampton, 67-92 (1997).
[106] Kress, R.: Numerical Analysis. Springer-Verlag, New York 1998.
[107] Kress, R.: On the low wave number behavior of two-dimensional scat-
tering problems for an open arc. J. for Anal. and Its Appl. (to appear).
[108] Kress, R., and Mohsen, A.: On the simulation source technique for
exterior problems in acoustics. Math. Meth. in the Appl. Sci. 8, 585-
597 (1986).
[109] Kress, R., and Sloan, LH.: On the numerical solution of a logarithmic
integral equation of the first kind for the Helmholtz equation. Numer.
Math 66, 193-214 (1993).
[110] Kress, R., Sloan, LH., and Stenger, F.: A sinc quadrature method for
the double-layer integral equation in planar domains with corners. J.
for Integral Eq. and Appl. (to appear).
[111] Kupradze, V.D.: Dynamical problems in elasticity. In: Progress in
Solid Mechanics, Vol. III (Sneddon and Hill, eds.). North Holland,
Amsterdam, 1-259 (1963).
[112] Kussmaul, R.: Ein numerisches Verfahren zur Lasung des Neu-
mannschen Aussenraumproblems fur die Helmholtzsche Schwin-
gungsgleichung. Computing 4, 246-273 (1969).
[113] Landweber, L.: An iteration formula for Fredholm integral equations
of the first kind. Amer. J. Math. 73, 615-624 (1951).
[114] Lax, P.D.: Symmetrizable linear transformations. Comm. Pure Appl.
Math. 7, 633-647 (1954).
References 355
[115] Little, S., and Reade, J.B.: Eigenvalues of analytic kernels. SIAM J.
Math. Anal. 15, 133-136 {1984}.
[116] Louis, A.K.: Inverse und schlecht gestellte Probleme. Teubner, Stutt-
gart 1989.
[117] Martensen, E.: Uber eine Methode zum raumlichen Neumannschen
Problem mit einer Anwendung fiir torusartige Berandungen. Acta
Math. 109, 75-135 {1963}.
[118] Martensen, E.: Potentialtheorie. Teubner-Verlag, Stuttgart 1968.
[119] Martensen, E.: Riesz theory without axiom of choice. Proc. Amer.
Math. Soc. 99, 496-500 {1987}.
[120] Meinardus, G: Approximation of Functions: Theory and Numerical
Methods. Springer-Verlag, New York 1967.
[121] Meister, E.: Randwertaufgaben der Funktionentheorie. Teubner-Ver-
lag, Stuttgart 1983.
[122] Meschkowski, H.: Hilbertsche Riiume mit Kernfunktion. Springer-
Verlag, Berlin 1962.
[123] Mikhlin, S.G.: Mathematical Physics, an Advanced Course. North-
Holland, Amsterdam 1970.
[124] Mikhlin, S.G., and Prassdorf, S.: Singular Integral Operators. Sprin-
ger-Verlag, Berlin 1986.
[125] Miranda, C.: Partial Differential Equations of Elliptic Type, 2nd ed.
Springer-Verlag, Berlin 1970.
[126] Manch, L.: On the numerical solution of the direct scattering prob-
lem for an open sound-hard arc. J. Compo Appl. Math. 71,343-356
(1996).
[127] Morozov, V .A.: On the solution of functional equations by the method
of regularization. Soviet Math. Doklady 7, 414-417 {1966} {English
translation}.
[128] Morozov, V.A.: Choice of parameter for the solution of functional
equations by the regularization method. Soviet Math. Doklady 8,
1000-1003 {1967} {English translation}.
[129] Morozov, V.A.: Methods for Solving Incorrectly Posed Problems.
Springer-Verlag, Berlin 1984.
[130] Miiller, C.: Foundations of the Mathematical Theory of Electromag-
netic Waves. Springer-Verlag, Berlin 1970.
356 References
[131] Multhopp, H.: Die Berechnung der Auftriebsverteilung von Tragflu-

geln. Luftfahrt-Forschung 4,153-169 (1938).
[132] Muntz, C.H.: Zum dynamischen Wiirmeleitungsproblem. Math. Zeit.

38, 323-337 (1934).
[133] Muskhelishvili, N.I.: Singular Integral Equations. Noordhoff, Gronin-

gen 1953.
[134] Nashed, M.Z., and Wahba, G.: Convergence rates of approximate

least squares solutions of linear integral and operator equations of
the first kind. Math. Compo 28, 69-80 (1974).
[135] Natanson, I.P.: Konstruktive Funktionentheorie. Akademie-Verlag,

Berlin 1955.
[136] Natterer, F.: Regularisierung schlecht gestellter Probleme durch Pro-

jektionsverfahren. Numer. Math. 28, 329-341 (1977).
[137] Neumann, C.: Untersuchungen iiber das logarithmische und Newton-

sche Potential. Teubner-Verlag, Leipzig 1877.
[138] Noether, F.: Uber eine Klasse singuliirer Integralgleichungen. Math.

Ann. 82, 42-63 (1921).
[139] Nussbaumer, H.J.: Fast Fourier Transform and Convolution Algo-

rithms. Springer-Verlag, Berlin 1982.
[140] Nystrom, E.J.: Uber die praktische Auflosung von Integralgleichun-

gen mit Anwendungen auf Randwertaufgaben. Acta Math. 54, 185-
204 (1930).
[141] Petrov, G.I.: Application of Galerkin's method to a problem of the

stability of the flow of a viscous fluid. Priklad. Matern. i Mekh. 4,
3-12 (1940) (In Russian).
[142] Phillips, D.L.: A technique for the numerical solution of certain inte-
gral equations of the first kind. J. Ass. Compo Math. 9, 84-97 (1962).
[143] Plemelj, J.: Ein Ergiinzungssatz zur Cauchyschen Integraldarstellung

analytischer Funktionen, Randwerte betreffend. Monatshefte f. Math.
u. Phys. 19, 205-210 (1908).
[144] Plemelj, J.: Potentialtheoretische Untersuchungen. Preisschriften der

Fiirstlich Jablonowskischen Gesellschaft. Teubner, Leipzig 1911.
[145] Pogorzelski, W.: Integral equations and their applications. Pergamon

Press, Oxford 1966.
References 357
[146] Potthast, R.: Frechet differentiability of boundary integral operators

in inverse acoustic scattering. Inverse Problems 10, 431-447 (1994).
[147] Prossdorf, S.: Einige Klassen singularer Gleichungen. Birkhauser-

Verlag, Basel 1974.
[148] Prossdorf, S., and Saranen, J.: A fully discrete approximation method
for the exterior Neumann boundary value problem for the Helmholtz
equation. Z. Anal. Anwendungen 13, 683-695 (1994).
[149] Prossdorf, S., and Silbermann, B.: Projektionsverfahren und die
naherungsweise Losung singularer Gleichungen. Teubner, Leipzig
1977.
[150] Prossdorf, S., and Silbermann, B.: Numerical Analysis for Integral
and Related Operator Equations. Akademie-Verlag, Berlin 1991, and
Birkhauser-Verlag, Basel 1991.
[151] Radon, J.: Uber die Randwertaufgaben beim logarithmischen Poten-
tial. Sitzungsberichte der Akadademie der Wissenschaften Wien. I1a
128, 1123-1167 (1919).
[152] Lord Rayleigh: The Theory of Sound. London 1896.
[153] Riesz, F.: Uber lineare Funktionalgleichungen. Acta Math. 41, 71-98
(1918).
[154] Ritz, W.: Uber eine neue Methode zur Losung gewisser Variations-
probleme der mathematischen Physik. J. f. reine u. angew. Math.
135, 1-61 (1908).
[155] Rokhlin, V.: Rapid solution of integral equations of classical potential

theory. J. Comput. Phys. 60, 187-207 (1985).
[156] Rudin, W.: Functional Analysis. McGraw-Hill, New York 1973.
[157] Saad, Y.: Iterative Methods for Sparse Linear Systems. PWS Pub-
lishing Company, Boston 1995.
[158] Saad, Y., and Schultz, M.H.: GMRES: A generalized minimum resid-
ual algorithm for solving nonsymmetric linear systems. SIAM J. Sci.
Stat. Comput. 7, 856-869 (1986).
[159] Saranen, J., and Vainikko, G.: Trigonometric collocation methods
with product integration for boundary integral equations on closed
curves. SIAM J. Numer. Anal. 33, 1577-1596 (1996).
[160] Schauder, J.: Uber lineare, vollstetige Funktionaloperationen. Studia
Math. 2, 183-196 (1930).
358 References
[161] Schippers, H.: Multiple Grid Methods for Equations of the Second
Kind with Applications in Fluid Mechanics. Mathematical Centre
Tracts 163, Amsterdam 1983.
[162] Schonhage, A.: Approximationstheorie. de Gruyter, Berlin 1971.
[163] Seidman, E.: Nonconvergence results for the application of least-

squares estimation to ill-posed problems. J. Optimization Theory and
Appl. 30, 535~547 (1980).
[164] Sloan, LH.: Error analysis for a class of degenerate-kernel methods.
Numer. Math. 25, 231~238 (1976).
[165] Sloan, I.H.: Analysis of general quadrature methods for integral equa-
tions of the second kind. Numer. Math. 38, 263~278 (1981).
[166] Sloan, I.H.: Error analysis of boundary integral methods. Acta Nu-
merica 1, 287~339 (1992).
[167] Sokhotski, Y.V.: On definite integrals and functions utilized for ex-
pansions into series. Doctor thesis, St. Petersburg (1873) (In Rus-
sian).
[168] Stetter, H.J.: The defect correction principle and discretization meth-
ods. Numer. Math. 29, 425~443 (1978).
[169] Taylor, A.E.: Introduction to Functional Analysis. John Wiley &

Sons, New York 1967.
[170] Taylor, M.E.: Pseudodifferential Operators. Princeton Univ. Press,
Princeton 1981.
[171] Taylor, M.E.: Partial Differential Equations. Springer-Verlag, New
York 1996.
[172] Tikhonov, A.N.: On the solution of incorrectly formulated problems

and the regularization method. Soviet Math. Doklady 4, 1035~ 1038
(1963) (English translation).
[173] Tikhonov, A.N.: Regularization of incorrectly posed problems. Soviet

Math. Doklady 4, 1624~1627 (1963) (English translation).
[174] Tikhonov, A.N., and Arsenin, V.Y.: Solutions of Ill-Posed Problems.

Winston and Sons, Washington 1977.
[175] Trefethen, L.N., and Bau, D.: Numerical Linear Algebra. SIAM,
Philadelphia 1997.
[176] Treves, F.: Basic Linear Partial Differential Equations. Academic

Press, New York 1975.
References 359
[177] Treves, F.: Introduction to PseudodifJerential Operators and Fourier

Integral Operators. Plenum, New York 1980.
[178] Vekua, N.P.: Systems of Singular Integral Equations. Noordhoff,
Groningen 1967.
[179] Verchota, G.: Layer potentials and regularity for the Dirichlet prob-
lem for Laplace's equation in Lipschitz domains. J. Func. Anal. 39,
572-611 (1984).
[180] Volterra, V.: Sulla inversione degli integrale definiti. Rom. Acc. Linc.
Rend. 5, 177-185 (1896).
[181] Wendland, W.L.: Die Fredholmsche Alternative fur Operatoren, die
bezuglich eines bilinearen Funktionals adjungiert sind. Math. Zeit.
101, 61-64 (1967).
[182] Wendland, W.L.: Die Behandlung von Randwertaufgaben im IR3 mit

Hilfe von Einfach- und Doppelschichtpotentialen. Numer. Math. 11,
380-404 (1968).
[183] Wendland, W.L.: Bemerkungen uber die Fredholmschen Siitze. Meth.

Verf. Math. Phys. 3, 141-176 (1970).
[184] Wendland, W.L.: Elliptic Systems in the Plane. Pitman, London
1979.
[185] Wendland, W.L.: Boundary element methods and their asymptotic

convergence. In: Theoretical Acoustics and Numerical Techniques
(Fillipi, ed.). CISM Courses 277, Springer-Verlag, Wien, 135-216
(1983).
[186] Wendland, W.L.: On some mathematical aspects of boundary ele-

ment methods for elliptic problems. In: The Mathematics of Finite
Elements and Applications V. Mafelap 1984 (Whiteman, ed.). Aca-
demic Press, London, 193-227 (1985).
[187] Wendland, W.L.: Boundary element methods for elliptic problems.
In: Mathematical Theory of Finite and Boundary Elements (Schatz
et al., eds.). Birkhiiuser, Boston, 219-276 {1990).
Index
adjoint operator, 40 closed ball, 5

Arzela-Ascoli theorem, 8 closed set, 5
closure, 5
Banach space, 5 codimension, 62
Banach-Steinhaus theorem, 175 collocation method, 225
best approximation, 11 compact, 6
bijective, 2 relatively, 7
bilinear form, 39 sequentially, 6
nondegenerate, 39 compact operator, 21
boundary element method, 232 complete orthonormal system, 13
bounded linear functionals, 49 complete set, 5
bounded operator, 16 completion, 6
bounded set, 5 condition number, 248
totally, 6 conjugate gradient method, 245
Bubnov-Galerkin method, 240 constraint optimization, 291
continuous extension, 27
Cauchy data, 74 continuous function, 3
Cauchy integral, 97 continuous operator, 3
Cauchy integral operator, 102 contraction, 18
Cauchy kernel, 108 convergence, 2
Cauchy principal value, 98 mean square, 3
Cauchy sequence, 5 norm, 16
Cauchy-Schwarz inequality, 10 pointwise, 16, 165
Cea's lemma, 221 uniform, 3
Chebyshev polynomial, 193 weak, 291
362 Index
convex set, 12 fundamental solution

heat equation, 155
defect correction iteration, 252 Helmholtz equation, 210
dense set, 5 Laplace equation, 68
direct sum, 30
Dirichlet integral, 142 Galerkin equation, 240
Dirichlet problem, 74 Galerkin method, 240
weak,143 Green's formula, 69
discrepancy, 298 Green's theorem, 69
principle, 271, 284, 300
distance, 2 Holder continuous, 94
domain of class C n , 25 Holder space Co,,,, (G), 95
dot product, 26 harmonic function, 68
double-layer heat potential, 156 hat functions, 181
double-layer potential, 78 heat equation, 152
dual least squares method, 314 heat potential, 155
dual space, 49, 131 Helmholtz equation, 210
dual system, 40 Hilbert inversion formula, 105
duality pairing, 132 Hilbert kernel, 105
Hilbert scale, 133
eigenelement, 35 Hilbert space, 10
eigenvalue, 35 holomorphic
multiplicity, 275 sectionally, 97
equicontinuous, 8 homeomorphic, 130
equivalent norm, 3
ill-posed problem, 266
fast multipole methods, 261 mildly, 280
finite defect, 63 severely, 280
Fourier series, 12, 126 imbedding operator, 96
Fn§Chet derivative, 339 improperly posed problem, 266
Fredholm alternative, 47 index, 63, 106
Fredholm integral equations, 1 injective, 2
Fredholm theorems, 46 inner product, 9
function, 1 integral equations
continuous, 3 of the first kind, 1
function space of the second kind, 1
C~T" 129 integral operator, 18
cn(D),68 integro differential equation, 304
cn(tJ),68 interpolation, 179
C O''''(G),95 piecewise linear, 181
HI(D),137 trigonometric, 182
HP(r),135 interpolation operator, 180
HP[0,27rj, 126 interpolation space, 133
L2[a, b], 6 inverse operator, 1
C[a,bj,3 inverse problem, 328
Index 363
involution, 41 Neumann series, 19, 171

isometric, 6 Noether theorems, 112
isomorphic, 6 norm, 2
iterated operator, 19 equivalent, 3
maximum, 3
Jordan measurable, 17 mean square, 3
jump relation, 78 stronger, 13
weaker, 13
kernel,18
norm convergence, 16
degenerate, 179
normal solvability, 61
weakly singular, 24, 26
normal vector, 25
Krylov subspaces, 245
normed space, 2
Lagrange basis, 180 nullspace, 28
Landweber-Fridman iteration, 287 numerical integration operator, 201
Laplace equation, 68 Nystrom method, 201
Lax's theorem, 44
open ball, 5
Lax-Milgram theorem, 242
open set, 5
least squares solution, 243
operator, 1
lemma
adjoint, 40
Cea, 221
bijective, 2
Riemann-Lebesgue, 289
bounded, 16
Riesz, 22
collectively compact, 167
limit, 2
compact, 21
linear operator, 15
continuous, 3
logarithmic double-layer potential,
imbedding, 96
78
injective, 2
logarithmic single-layer potential,
inverse, 1
78
linear, 15
Lyapunov boundary, 87
nonnegative, 276
mapping, 1 positive definite, 240
maximum norm, 3 projection, 34
maximum-minimum principle, 71, self-adjoint, 240, 272
153 surjective, 2
mean square norm, 3 operator equation
mean value theorem, 70 of the first kind, 23, 265
minimum norm solution, 298 of the second kind, 18, 23
minimum-maximum principle, 276 operator norm, 16
moment collocation, 317 orthogonal, 10
moment solution, 313 complelement, 10
multigrid iteration, 256 system, 10
full,258 orthogonal projection, 219
orthonormal basis, 275
Neumann problem, 74 orthonormal system, 10
weak, 143 complete, 13
364 Index
panel clustering methods, 261 Riemann-Lebesgue lemma, 289

parallel surfaces, 75 Riesz lemma, 22
parametric representation, 25 Riesz number, 30
regular, 25 Riesz theorem, 28, 42
Parseval equality, 13
Petrov-Galerkin method, 240 scalar product, 9
Picard theorem, 279 Schauder theory, 49
piecewise linear interpolation, 181 semi-norm, 95
pointwise convergence, 16 sequentially compact, 6
Poisson integral, 155, 289 series, 14
polynomial sesquilinear form, 41
Chebyshev, 193 positive definite, 42
pre-Hilbert space, 10 symmetric, 42
product integration method, 207 Simpson's rule, 199
projection method, 219 single-layer heat potential, 156
projection operator, 34, 219 single-layer potential, 78
properly posed problem, 266 singular integral operator, 108
singular system, 278
quadrature, 198
singular value, 277
convergent, 199
decomposition, 277
interpolatory, 198
Sobolev space
points, 198
weights, 198 H 1 (D),137
quasi-solution, 293 HP(r),135
HP[O, 21l"j, 126
range, 1 Sokhotski-Plemelj theorem, 101
Rayleigh-Ritz method, 240 space
reduced wave equation, 210 Banach, 5
regular value, 35 Hilbert, 10
regularization normed,2
method,269 pre-Hilbert, 10
parameter, 269 Sobolev, 126
scheme, 269 span, 12
strategy, 270 spectral cut-off, 284
regularizer, 55 spectral radius, 35
equivalent, 56 spectral theorem, 273
left, 55 spectrum, 35
right, 55 splines, 181
relatively compact, 7 Steklov theorem, 200
reproducing kernel, 315 successive approximations, 20, 172
residual, 252 superconvergence, 246
residual correction, 252 surface element, 25
resolvent, 35 surface patch, 25
resolvent set, 35 surjective, 2
Riemann problem, 105 Szeg6 theorem, 200
Index 365
tangent vector, 25 trapezoidal rule, 198

theorem trial and error, 271
ArzeUt-Ascoli, 8 triangle inequality, 2
Banach-Steinhaus, 175 trigonometric interpolation, 182
Fredholm, 46 two-grid iteration, 252
Green, 69
Lax, 44 uniform boundedness principle, 165
Lax-Milgram, 242 uniformly continuous, 3
Noether, 112 unisolvent, 180
Picard,279
Riesz, 28, 42 Volterra integral equations, 36, 174
Sokhotski-Plemelj, 101
wavelets, 261
Steklov, 200
weak derivative, 302
Szego, 200
weakly convergent, 291
Weierstrass, 5
weakly singular kernel, 24
Tikhonov functional, 291
Weierstrass approximation theo-
Tikhonov regularization, 284
rem, 5
totally bounded, 6
weight function, 198
trace operator, 138
well-posed problem, 266
(continued from page ii)
61. SattingerlWeaver: Lie Groups and Algebras with 89. O'MaUey: Singular Perturbation Methods for
Applications to Physics, Geometry, and Ordinary Differential Equations.
Mechanics. 90. Meyer/HaU: Introduction to Hamiltonian
62. LaSalle: The Stability and Control of Discrete Dynamical Systems and the N-body Problem.
Processes. 91. Straughan: The Energy Method, Stability, and
63. Grasman: Asymptotic Methods of Relaxation Nonlinear Convection.
Oscillations and Applications. 92. Naber: The Geometry of Minkowski Spacetime.
64. Hsu: Cell-to-Cell Mapping: A Method of Global 93. Colton/Kress: Inverse Acoustic and
Analysis for Nonlinear Systems. Electromagnetic Scattering Theory. 2nd ed.
6S. Rand/Armbruster: Perturbation Methods, 94. Hoppensteadt: Analysis and Simulation of
Bifurcation Theory and Computer Algebra. Chaotic Systems.
66. HlawiceklHaslinger/NecasIlLov{sek: Solution of 95. Hackhusch: Iterative Solution of Large Sparse
Variational Inequalities in Mechanics. Systems of Equations.
67. Cercignani: The Boltzmann Equation and Its 96. Marchioro/Pulvirenti: Mathematical Theory of
Applications. Incompressible Nonviscous Fluids.
68. Temam: Infinite-Dimensional Dynamical 97. Lasota/Mackey: Chaos, Fractals, and Noise:
Systems in Mechanics and Physics, 2nd ed. Stochastic Aspects of Dynamics, 2nd ed.
69. Golubitsky/Stewart/Schaeffer: Singularities and 98. de Boor/Hollig/Riemenschneider: Box Splines.
Groups in Bifurcation Theory, Vol. II. 99. Hale/Lunet: Introduction to Functional
70. Constantin/FoiasiNicolaenkolTemam: Integral Differential Equations.
Manifolds and Inertial Manifolds for Dissipative 100. Sirovich (ed): Trends and Perspectives in
Partial Differential Equations. Applied Mathematics.
71. Catlin: Estimation, Control, and the Discrete 101. NusselYorke: Dynamics: Numerical
Kalman Filter. Explorations, 2nd ed.
72. LochakiMeunier: Multiphase Averaging for 102. Chossatl/o{)ss: The Couette-Taylor Problem.
Classical Systems. 103. Chorin: Vorticity and Turbulence.
73. Wiggins: Global Bifurcations and Chaos. 104. Farkas: Periodic Motions.
74. MawhinIWillem: Critical Point Theory and 105. Wiggins: Normally Hyperbolic Invariant
Hamiltonian Systems. Manifolds in Dynamical Systems.
75. Ahraham/Marsden/Ratiu: Manifolds, Tensor 106. Cercignanilltlner/Pulvirenti: The Mathematical
Analysis, and Applications, 2nd ed. Theory of Dilute Gases.
76. Lagerstrom: Matched Asymptotic Expansions: 107. Antman: Nonlinear Problems of Elasticity.
Ideas and Techniques. 108. Zeidler: Applied Functional Analysis:
77. Aldous: Probability Approximations via the Applications to Mathematical Physics.
Poisson Clumping Heuristic. 109. Zeidler: Applied Functional Analysis: Main
78. Dacorogna: Direct Methods in the Calculus of Principles and Their Applications.
Variations. 110. Diekmann/van GilslVerduyn LunellWalther:
79. Hernandez-Lerma: Adaptive Markov Processes. Delay Equations: Functional-, Complex-, and
80. Lawden: Elliptic Functions and Applications. Nonlinear Analysis.
81. Bluman/Kumei: Symmetries and Differential 111. Visintin: Differential Models of Hysteresis.
Equations. 112. Kuznetsov: Elements of Applied Bifurcation
82. Kress: Linear Integral Equations, 2nd ed. Theory, 2nd ed.
83. BebernesiEberly: Mathematical Problems from 113. Hislop/Sigal: Introduction to Spectral Theory:
Combustion Theory. With Applications to Schrodinger Operators.
84. Joseph: Fluid Dynamics of Viscoelastic Fluids. 114. Kevorkian/Cole: Multiple Scale and Singular
85. Yang: Wave Packets and Their Bifurcations in Perturbation Methods.
Geophysical Fluid Dynamics. lIS. Taylor: Partial Differential Equations I, Basic
86. Dendrinos/Sonis: Chaos and Socio-Spatial Theory.
Dynamics. 116. Taylor: Partial Differential Equations II,
87. Weder.· Spectral and Scattering Theory for Wave Qualitative Studies of Linear Equations.
Propagation in Perturbed Stratified Media. 117. Taylor: Partial Differential Equations III,
88. BogaevskilPovzner: Algebraic Methods in Nonlinear Equations.
Nonlinear Perturbation Theory.
(continued on next page)
(continued from previous page)
118. Godlewski/Raviart: Numerical Approximation of 127. Isakov: Inverse Problems for Partial Differential
Hyperbolic Systems of Conservation Laws. Equations.
119. Wu: Theory and Applications of Partial 128. LilWiggins: Invariant Manifolds and Fibrations
Functional Differential Equations. for Perturbed Nonlinear SchrMinger Equations.
120. Kirsch: An Introduction to the Mathematical 129. Muller: Analysis of Spherical Symmetries in
Theory of Inverse Problems. Euclidean Spaces.
121. BrokateiSprekels: Hysteresis and Phase 130. Feintuch: Robust Control Theory in Hilbert
Transitions. Space.
122. Gliklikh: Global Analysis in Mathematical 131. Ericksen: Introduction to the Thermodynamics
Physics: Geometric and Stochastic Methods. of Solids, Revised ed.
123. Le/Schmitt: Global Bifurcation in Variational 132. Ihlenburg: Finite Element Analysis of Acoustic
Inequalities: Applications to Obstacle and Scattering.
Unilateral Problems. 133. Vorovich: Nonlinear Theory of Shallow Shells.
124. Polak: Optimization: Algorithms and Consistent 134. Vein/Dale: Determinants and Their Applications
Approximations. in Mathematical Physics.
125. Arnold/Khesin: Topological Methods in 135. Drew/Passman: Theory of Multicomponent
Hydrodynamics. Fluids.
126. Hoppensteadtllzhikevich: Weakly Connected
Neural Networks.

Kress Linear Integral Equations, 1999

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Kress Linear Integral Equations, 1999

Uploaded by

Copyright:

Available Formats

Applied Mathematical Sciences

Linear Integral Equations

Library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

ISBN 978-1-4612-6817-8 SPIN 10707345

G6ttingen, October 1998 Rainer Kress

Gottingen, September 1988 Rainer Kress

Preface to the Second Edition vii

Preface to the First Edition ix

2 Bounded and Compact Operators 15

4 Dual Systems and Fredholm Alternative 39

5 Regularization in Dual Systems 55

7 Singular Integral Equations 94

8 Sobolev Spaces 125

9 The Heat Equation 152

10 Operator Approximations 163

10.3 Collectively Compact Operators. . . . . . . 167

11 Degenerate Kernel Approximation 177

12 Quadrature Methods 197

13 Projection Methods 218

14 Iterative Solution and Stability 247

15 Equations of the First Kind 265

16 Tikhonov Regularization 290

16.2 Weak Convergence 291

17 Regularization by Discretization 308

18 Inverse Boundary Value Problems 320

The topic of this book is linear integral equations of which

lb K(x, y)cp(y) dy = f(x), x E [a, b],

are typical examples. In these equations the function cp is the unknown,

R. Kress, Linear Integral Equations

1.1 Convergence and Continuity

(N2) II'P II =o if and only if 'P = 0, (definiteness)

(N3) II a'P II = lalll'PlI, (homogeneity)

(N4) II'P + 1/111 < II'PII + 111/111, (triangle inequality)

I II'PII - 111/111 I ~ II'P - 1/111· (1.1)

Note that by (N4) the limit of a convergent sequence is uniquely deter-

Definition 1.3 A function A : U C X --+ Y mapping a subset U of a

An equivalent definition is the following: A function A : U C X --+ Y

or the mean square norm

Convergence of a sequence of continuous functions in the maximum norm

Theorem 1.6 On a finite-dimensional linear space all norms are equiva-

for all cP E X, where

Because of II'l/Jnlloo = 1 each of the sequences (akn), k = 1, ... , m, is

Obviously, open balls are open.

A subset U of a normed space X is closed if and only if its complement

A subset U is closed if and only if it coincides with its closure. By the

Convergent sequences are bounded.

complex numbers, using the norm (1.2), we observe that finite-dimensional

A subset U is called sequentially compact if every sequence of elements

Theorem 1.15 A subset of a normed space is compact if and only if it is

Proof. Let U be compact and assume that it is not sequentially compact.

In particular, from Theorem 1.15 we observe that compact sets are

Definition 1.16 A subset of a normed space is called relatively compact

As a consequence of Theorem 1.15, a set U is relatively compact if and

Theorem 1.17 A bounded and finite-dimensional subset of a normed space

For the compactness in C(G), the space of continuous real- or complex-

we have the following criterion.

Theorem 1.18 (Arzela-Ascoli) A set U c C(G) is relatively compact if

1'P(x) - 'P(y) I < E:

Proof. Let U be bounded and equicontinuous and let ('Pn) be a sequence in

+ICPn(I)(Xj) - CPn(I)(x)1 < 3e

. min .lIcp - cpjlloo < :.

for all x, y E G with Ix - yl < 6. Therefore U is equicontinuous. Finally, U