Professional Documents
Culture Documents
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
Applied Mathematical Sciences
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
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
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.
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.
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
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
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
References 347
Index 361
1
Normed Spaces
-l
and
b
cp(x) K(x, y)cp(y) dy = f(x), x E [a, b],
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.
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
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
°
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
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.
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
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
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.
Proof. This follows from the Bolzano-Weierstrass theorem using the norm
(1.2). 0
for all k, I 2:: N(e) and all j = 1, ... , i. From the equicontinuity we obtain
we obtain
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
Note that L2[a, b] is a Hilbert space with the scalar product given by
(cp,'ljJ):= lb cp(x)'ljJ(x)dx.
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
(<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.
(1.4)
12 1. Normed Spaces
2 2
<4d
-
2
+-+-
n m
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
IIcpll2 = 2: l(cp,unW·
n=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
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.
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
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
IIA'PII ~ 1: IIAikllll'Plioo
k=l
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
IIAlloo::::
lcr IK(xo,y)1 dy = max r IK(x,y)1 dy.
xEC lc
(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
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.
20 2. Bounded and Compact Operators
with arbitrary <Po E X converge to the unique solution <P of <P - A<p = f·
whence
follows. o
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,
Proof. This is obvious, since relatively compact sets are bounded (see The-
orem 1.15). 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.
f3 := inf
cpEU
Ilf - 'PII > O.
We can choose g E U such that
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 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
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,
[lin
::; Mllcplloowm10 po-m pm-l dp
= Mllcplioo : ; , x E G.
2.4 Compact Operators 25
From this we observe that AnCP -+ Acp, n -+ 00, uniformly, and therefore
Acp E e(O). Furthermore it follows that
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
gij
ax ax
:= ~ . ~, i, j = 1, ... ,m - l.
UUi uUj
26 2. Bounded and Compact Operators
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,
(Acp)(x):= r
laD
K(x,y)cp(y)ds(y), xEaD, (2.5)
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) ,
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
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.
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
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.
and
x = LD(X) ~ Ll(X) ~ ... ~ LT(X) = U+1(X) = .... (3.2)
Furthermore, we have the direct sum
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
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
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
Hence
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<p = Lk+lrp for some rp EX,
because Lk(X) = Lk+1(X). Hence 'ljJ = Lk+2rp E L k+2(X), and therefore
Lk+l(X) c Lk+2(X), i.e., we have proven that
3. Now we show that r = q. Assume that r > 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.
fn
gn := IIL-1 fnll '
n E IN.
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
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
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
(3.5)
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
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
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)
'ljJ(x):= l x
<p(y) dy, xE[a,b],
'ljJ(x) - r
la
Ky(x, y) 'ljJ(y) dy = f(x) ,
K(x,x) K(x,x)
x E [a, b]. (3.8)
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
L
00
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.
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
'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 4.4 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
(B'lj;)(x):= fa K(y,x)'lj;(y)dy, x E G,
are adjoint.
Proof The theorem follows from
(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
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
whence IIFII :S Ilfll follows. On the other hand, inserting f into (4.3) yields
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
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.
44 4. Dual Systems and Fredholm Alternative
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)
(4.7)
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,
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
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
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
46 4. Dual Systems and Fredholm Alternative
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,
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
cP = L L1!iCPi,
i=l
4.3 The Fredholm Alternative 47
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
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
dimN(I - A) = dimN(I - B) E 1N
48 4. Dual Systems and Fredholm Alternative
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.
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
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
i cp(x)g(x) dx = 0
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).
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
(4.16)
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
0= 10 1
cp(y) dy + u'(1) - u'(O).
4.4 Boundary Value Problems 51
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
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
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.,
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
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
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
(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.
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.
(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
58 5. Regularization in Dual Systems
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
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
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),
60 5. Regularization in Dual Systems
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
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
where
n
gi := LPik {Ck - R'(I - B +T'p,)-lK'ck} , i = 1, ... ,n - p.
k=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
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
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
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.
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
Lakbk = TPep = O.
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
K~'l,U = Lf3k'l,U2,k.
k=1
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
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
where
m {)2u
~u:=L {) 2 •
j=l
Xj
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}.
! D
{uL.v+gradu.gradv}dx = r
laD
u ~v
ul/
ds (6.1)
rdiv A
lD
dx = r
laD
1/' A ds
1 aD ul/
av
-;:;- ds = O. (6.3)
u(x) = laD {~~ (y) <t>(x, y) - u(y) a:~(;~)} ds(y), xED. (6.4)
1 aDUr!(x;r)
{
u(y)
a<t>(x, y) au }
a () - -;:;- (y) <t>(x, y) ds(y) = O.
1/ Y ul/
70 6. Potential Theory
v(y)
grady ~(x, y) = wmrm- 1 ' (6.5)
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.
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
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
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,
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.
u(X) = U oo
r {u(y)
+ laD aq,(x, y) au }
av(y) - av (y) q,(x, y) ds(y) (6.8)
r
laD ov
auds=O (6.9)
U oo = -2
1 r
1rr llyl=r
u(y) ds(y) (6.10)
gradu(x)= mm
wmr
r
1B[x;rJ
gradu(y)dy=-
wmr
mm r
In(x;r)
v(y)u(y)ds(y),
~
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)
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) = °
6.1 Harmonic Functions 73
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).
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
[ { u(y) oq>(x,y) ou }
0 () - -;;) (y) q>(x, y) ds(y) --+ u oo , n --+ 00.
f!(x;rn) V Y uV
(6.14)
and
oq>(x,y) _
OXj -
0(_1_)
Ixl m- 1 '
(6.15)
74 6. Potential Theory
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= f on aD,
u= f on aD,
6.2 Boundary Value Problems: Uniqueness 75
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 ]
76 6. Potential Theory
are related by
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
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.
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
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.
78 6. Potential Theory
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)
<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)
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
1=::
XED,
. 1 1
and from
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<I>(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 <I>(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 <I> {x, y)
w{x) := laD 8v(y) ds(y)
and
( 8<I>(x,y)
u{x):= laD {r.p(y) - r.p(z)} 8v{y) ds{y). (6.26)
uniformly on 8D.
From (6.21) we can conclude that
1
Ix - Yl2 ~ "2 {Iz - Yl2 + Ix - Z12}
1
8 <I>(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<I>(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)' }
( 1 I
8 <I>(x,y) _ 8<I>(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
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.
82 6. Potential Theory
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
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
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
84 6. Potential Theory
u(x)
r
= laD cp(y)
a<I>(x, y)
av(y) ds(y), xED, (6.34)
r
u(x) = laD cp(y)
{a<I>(x, y)
av(y)
I} ds(y),
+ Ixl m - 2 (6.36)
r
cp(x) + 2 laD cp(y)
{a<I>(x, y)
av(y) +
I} ds(y)
Ixl m - 2 = 2f(x), x E aD. (6.37)
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
-
(K<p)(x)
r <p(y) {a<I>(x,y)
:= 2 laD av(y) + Ixl
I} ds(y), m- 2 x E aD,
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 <p ds = 0 implies that <p = 0, and the
existence of a unique solution to the integral equation (6.37) follows from
Theorem 3.4. D
u(x) =
laD
r 'lj;(y) <I> (x, y) ds(y), XED, (6.38)
'lj;(x)
r
+ 2 laD 'lj;(y)
a<I>(x,y)
av(x) ds(y) = 2g(x), x E aD. (6.39)
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
86 6. Potential Theory
( aif?(x,y)
1/J(x) - 2 laD 1/J{y) av(x) ds(y) = -2g(x), x E aD, (6.42)
( 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
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
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.
6.5 Nonsmooth Boundaries 89
(KnCP)(x):=
l8D
r h(nlx - yl)k(x, y)cp(y) ds(y), x E 8D.
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"
90 6. Potential Theory
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
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.
J(h):= r
J&Dh
u 2 ds, h > 0,
6.5 Nonsmooth Boundaries 91
we can write
J(h) = r
loD
{I + 2hH(x) + h2K(x)} [u(x - hv(x)))2ds(x)
(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,
92 6. Potential Theory
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
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
.
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
and
1 1
(L<p)(Xi) - (Lrp)(X2) = 2" [<P(Xi) - rp(x)]- 2" [<P(X2) - rp(x)]
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.
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
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 <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
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
1·1m
p-tO
l -d(- = 1·1m
H(z;p) ( - Z p-tO
l H(z;p)
·dfJ·
~ = ~7r,
Hence,
r
lr(z;R)
I<p«) -
( - z
I
<p(z) ds ::; 21<p1", jR-R
10-1",-1 do- = 4R'"
a
l<pl",
7.2 The Cauchy Integral Operator 99
and
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)
::; 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
7.2 The Cauchy Integral Operator 101
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- ,
1
I±(z) = -2.
11"~
Ir cp«() d( =t= -21 cp(z),
-r-
r .. - z
z E r.
1
I(z) := -2.
11"~
Irr .. -
cp«()
-r-
z
de, ZECC\r, (7.11)
J±(z) = -2·
1
11"~
1 cp«()
-r-
r .. - z
1
d( =t= -2 cp(z), zE r, (7.12)
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.
1 1
J± = -2 Aip =f -
2
ip (7.13)
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
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
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
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
f(z) := ~
27r~
r r.p(()
lr (- z
de, 1 ['¢(()
g(z):= -2.
7r~
-;-- de,
r.,-z
z E <C\f.
104 7. Singular Integral Equations
= -2 ( J(z)g(z) dz -+ 0, R -+ 00,
J1z1=R
since J(z), g(z) = O(l/izi), z -+ 00. D
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)
{27r t
Jo cot"2 dt =O,
1 (27r T - t -
ip(t) = - 27r Jo cot -2- 'lj;(T) dT + C, t E [0,27rJ, (7.16)
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.
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
linearly and bijectively onto solutions f of the Riemann problem with the
boundary condition
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
106 7. Singular Integral Equations
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
between the density <p and the function f defined by the Cauchy integral.
This ends the proof. 0
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.
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
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.
(7.21 )
108 7. Singular Integral Equations
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,
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 '
(7.24)
::; 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)
(7.27)
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 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)
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,
_1_fdlna-b
27ri lr a +b .
a(z)cp(z) +~ r,
7rZ lrf k;Z'
. , - ()z cp(() d( = h(z), z E
Ir h(z)'lj;(z) dz = °
is satisfied for all solutions 'lj; of the homogeneous adjoint equation.
7.4 Integral Equations with Cauchy Kernel 113
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.
1
a(t)cp(t) - 27r 10
rrr k(t, r) cot -2-
r-t
cp( r) dr = h(t), t E [0,27rJ,
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,
114 7. Singular Integral Equations
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
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
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
lrr (-
d(
z = lrr dln«( - z) = lrr · r In r . Of)) (0
d(ln r + zf)) = lr ---a;- + z os ds.
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.
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)
116 7. Singular Integral Equations
is given by
1 { (- z
gradu(z) = 271" lr ep«() I( _ Zl2 ds«(),
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
)
7.5 Cauchy Integral and Logarithmic Potential 117
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
1[
= 271" grad ir(d<p 1
ds (() In I( _ zl ds(()
].l. , z E ]R? \ r.
av± 1 d ( dip 1
r,
av (z) = 271" ds(z) ir ds (() In I( _ zl ds((), zE (7.41)
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)
(7.44)
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)
Proof. Let f E c 1,a (f) and assume that ~ E cO,a (r) solves
So~ = f. (7.48)
d
ds So~
1 r
+ Iff lr So~ ds =
df
ds
1
+ 1fT lr f
r ds. (7.49)
df
Rf := ds
1
+ 1fT lr f
r ds
7.5 Cauchy Integral and Logarithmic Potential 119
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)
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
(( - 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
1 r 8 1
u(z) = 21r Jr rp(() 8v(() In I( _ zl ds((), ZElR?\r, (7.51 )
120 7. Singular Integral Equations
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)
h9dS = 0 (7.53)
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.
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
7.6 Logarithmic Single-Layer Potential on an Arc 121
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
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
122 7. Singular Integral Equations
frq z 1 ;1/n)
I(M'IjJ)(() In Ir ~ II ds(() --+ 0,
., z
n --+ 00,
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)
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
1 ('
(Lcp)(t) := 47r io {p(t, r) - 4}cp(r) dr, t E [0,7r],
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].
(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].
124 7. Singular Integral Equations
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
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.
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.
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
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
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
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
L
00
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
m=-oo
128 8. Sobolev Spaces
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
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
8.1 The Sobolev Space HP[O, 211"] 129
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
and Parseval's equality (8.2) for the Fourier coefficients of cp and cp(k) we
obtain 00
m=-oo
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
130 8. Sobolev Spaces
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
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
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
(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
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].
m=-(X)
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
Therefore
Hence,
ior
1 2rr
G(rp):= 27r rp(t)g(t)dt, rp E HP[O, 27r], (8.10)
Theorem 8.12 Let p < q and r = >.p + (1 - >.)q with 0 < >. < 1. Then
L
00
= 1I<pII~A 1I<p1I~-2A,
(8.11)
134 8. Sobolev Spaces
(8.12)
(8.13)
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 0 < ). < 1 and all cp E HP+>.r, i. e., A maps HP+>.r boundedly into
Hq+>.r for 0 < ). < 1.
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 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.
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
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
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
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
138 8. Sobolev Spaces
am(ry) := ~ r
21f 10
27r
u(t, ry)e-imtdt.
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)
8.2 The Sobolev Space HP(r) 139
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
°
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,
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)'
lul£1(r) := Ii udsl·
for some constant C I . Combining this with (8.16) and (8.18), we see that
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)
u(x) = j Xl
-R UXI
OU
-;;- dXI,
i:
with the aid of the Cauchy-Schwarz inequality, we see that
and then apply (8.20) to the first term and (8.21) to the second term on
the right-hand side. Observing that
Ivl grad ul dx 2
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.
8.3 Weak Solutions to Boundary Value Problems 143
L
satisfies
gradu· gradvdx = 0 (8.22)
u =f on r, (8.23)
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).
L l
satisfies
grad u . grad v dx = gv ds (8.24)
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
144 8. Sobolev Spaces
1[
(Kip)(x) := -
a 1
( ) In -I- I ds(y),
ip(Y) - a xEr,
7rr vy x-y
(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,
8.3 Weak Solutions to Boundary Value Problems 145
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
10 ds 10 10
21r
11" Icp(X(T)) -
. 2-
t -
T
~(x(t)W dTdt
sm
2
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
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,
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.)
8.3 Weak Solutions to Boundary Value Problems 147
°
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"
(Sip)(x) := -
7f r
11 x-y
1
ip(y) In -I- I ds(y), x E r,
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)
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
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
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).
8.3 Weak Solutions to Boundary Value Problems 149
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.
1
u(x) = 27r
r
lr 'P(y)
8 1
8v(y) In Ix _ yl ds(y), xED.
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,
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
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
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
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
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].
at = D.u
au in D x (O,Tj, (9.1)
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
and
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
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
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.
G(x,t;y,r):=
v'47f(t-r)
1 m exp{-~~-YI~},
t-r
t > r, (9.5)
u(x, t) := ~m
47ft irlR", exp {_IX ~tY12} w(y) dy (9.6)
~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)
156 9. The Heat Equation
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
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
( ) 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
9.2 Heat Potentials 157
Standard analysis again shows that both the single- and double-layer
heat potentials are infinitely differentiable solutions to the heat equation
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
./,( ) 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
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 <P is uniformly continuous, for any c > 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
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
(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
°
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.
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
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
r1 ds
1:= 10 [s(1 - s)]a .
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.
162 9. The Heat Equation
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).
9.3 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
._ a - b 1 { (a -
h(t,T) .- 2ft ~)3 exp - 4(t _ T)
b?} , 0:::; T < t.
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
Acp = I
(10.1)
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
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
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.
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
1 1
~ '2 Pk + ... + '2 pj-l ~ Pk·
°
Proof. For c > 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
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
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
II( I - An
)-111 <
-
1 + 11(1 - A)-l An II
1 -II(I _ A)-l(An _ A)Anll . (10.5)
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,
and the proof is complete. o
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)
Proof. This follows from Theorem 10.9 by interchanging the roles of A and
An. 0
(10.8)
L
00
k=O
converges in the operator norm and defines a bounded linear operator with
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
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
°
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
We conclude this section with two examples for the application of The-
orem 10.13.
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
uniformly for all directions. Now we can apply Green's Theorem 6.3 to
obtain
(10.10)
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
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.
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
(An<P)(X) = fc Kn(x,y)<p(y)dy, x E G,
(11.5)
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.
180 11. Degenerate Kernel Approximation
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
Then
n
(11. 7)
Lj(x) = 1 (11.8)
h (XHl - x), x E [Xj, XH1], j:::; n -1,
0, otherwise.
(11.9)
Theorem 11.3 Let 9 E C 2[a, b]. Then, for the error in piecewise linear
interpolation we have the estimate
(11.10)
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
-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
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
(11.16)
(11.17)
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)
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
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
186 11. Degenerate Kernel Approximation
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
(11.23)
° ~ q ~p,
1
"2 < p, (11.24)
J.L=n.
Therefore the trigonometric interpolation polynomial is given by
11.4 Degenerate Kernels via Interpolation 187
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
(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
(11.26)
lb a
K(xj, y)Lk(Y) dy ~ L
m=O
n
K(xj, xm)
lb Lm(y)Lk(Y) dy,
a
(11.27)
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)
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
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)
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.
(11.31)
for all c > O. If the kernel is analytic, then from Theorem 11.7 we have
(11.32)
11.4 Degenerate Kernels via Interpolation 191
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.
Example 11.12 Consider the integral equation
ab r 21r <per) dr
o ~ t ~ 211",
<pet) + -; 10 a2 + b2 _ (a 2 _ b2 ) cos(t + r) = J(t), (11.35)
numerically want to solve the case where the unique solution is given by
L
00
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
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)
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
1-1 Vl- x 2
(11.38)
(11.39)
194 11. Degenerate Kernel Approximation
/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
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,
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
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.
196 11. Degenerate Kernel Approximation
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
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.
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.
~ (x - X2j_2)3 - 2~ (x - X2j_2)4,
Ks(x) := {
h (X2j - x )3 - 24
18 1 ( 4
X2j - x) ,
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
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 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
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)
IOn - An<Pn = j,
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
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<P - A<plloo -+ 0, n -+ 00, i.e., we have
pointwise convergence An<P -+ A<p, n -+ 00, for all <P E C(G) (see Problem
12.1).
For e > °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
for all <P E C(G) with 1I<plloo = 1. Using this result, we derive
IIA - Anlloo = sup II(A - An)<Plloo ~ sup sup II (A - An)<P'¢clloo
II'PII =1
00 1I'P1l00=1 1£>0
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
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
12.2 Nystrom's Method 205
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
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.
206 12. Quadrature Methods
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
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)
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.
10r °
27r t - T) K(t,T)CP(T)dT,
1
(Acp)(t):= 27r In ( 4sin 2 -2- ~ t ~ 27r, (12.15)
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)
ft(T) 1
:= -In ( 4sin 2 -
t -- T) , T E ffi,
t-T d71
27r 'F ,
27r 2
12.3 Weakly Singular Kernels 209
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
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
1 1
leimt1 - eimt2 12 ,
00
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
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
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
<I>(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
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
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)
212 12. Quadrature Methods
where
t-T J I ( 2Ksin -2-
KI(t, T) := -Ksin -2- t-T) ,
t-
LI(t,T) := -Jo ( 2Ksin-
2-
T) ,
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,
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
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).
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,
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
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
216 12. Quadrature Methods
Now we are in the position to apply Theorem 10.1 and the error estimate
(10.3) for the approximate solution of <p - A<p = I by <Pn - An<Pn = In
in the Sobolev spaces HP[0,27r] for p ;;::: 1. We note that because of the
imbedding Theorem 8.4, convergence of the approximate solutions in the
Sobolev space HP[O, 27r] for p > 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
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
(f, g) := lim
0'-+8
10
211"
{f(t + iu)g(t + iu) + f(t - iu)g(t - iu)} dt.
f () 1 int n E 7l,
n t = (41Tcosh2ns)1/2 e ,
Problems 217
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
A<p=f (13.1)
(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.
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.
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
and observe that the condition (13.4) is satisfied for S. This completes the
~~ 0
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)
lI<Pn - <p1I :::; M{llPn<P - <p1I + IlPn(An - A)<p1I + IlPn(Jn - J)II} (13.12)
for some constant C. Now (13.12) follows from (13.10) and the uniform
boundedness principle Theorem 10.4. 0
(13.14)
(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
Acp = f, (13.18)
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
(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
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
228 13. Projection Methods
for the linear spline collocation approximate solution 'Pn. Here, M denotes
some constant depending on the kernel K.
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
1 127r
(A~)(t) := 27r 0
t -
In ( 4sin2 -2- r)
K(t, r)~(r) dr, os; t S; 27r, (13.23)
230 13. Projection Methods
where
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
13.3 The Collocation Method 231
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
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].
ior T
21r
1
27r {
cot r t + K(t, r) } <p(r) dr = f(t), o ~ t ~ 27r, (13.25)
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)
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,
13.4 Collocation Methods for Equations of the First Kind 235
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
(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-
236 13. Projection Methods
°
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
(13.31 )
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
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
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
238 13. Projection Methods
for all r.p E Xn and some constant C4. For 0 ~ m ~ n/2, writing
2n
r.p = Lan-jfn-j
j=O
From this, with the aid of the integrals in Lemma 8.21, we find that
1 1 8m
-----, < - (n + j - m)2 ~ n 2 ~ 4(n _ j)2
n-m+j n+m-j - 3n 2 '
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.
13.4 Collocation Methods for Equations of the First Kind 239
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
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
240 13. Projection Methods
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
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
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)
13.5 The Galerkin Method 241
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.
242 13. Projection Methods
(13.39)
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 'P E X n , since orthogonal projections are self-adjoint (see Problem
13.1). Hence,
(13.41 )
13.5 The Galerkin Method 243
Therefore, IIA;:;-l PnAl1 ~ IIAli/c for all n E lN, and the statement follows
from Theorem 13.6. 0
form
(13.42)
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
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
246 13. Projection Methods
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
whence
vi - <P = (A5)-l{/5 - I - (A5 - A)<p}
follows. Now we can estimate
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
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
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
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
(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
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
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
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
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)
254 14. Iterative Solution and Stability
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
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
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
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
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
(14.18)
we can write the recursion for the iteration operator Mn in the form
(14.19)
14.3 Multigrid Methods 257
(14.20)
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
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.,
14.3 Multigrid Methods 259
(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
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)
260 14. Iterative Solution and Stability
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
Proceeding as in the derivation of (14.19), ~ can !?e shown that the iteration
operators satisfy the recurrence relation Ml = Tl and
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].
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
262 14. Iterative Solution and Stability
(14.29)
=L
M
KM(X,y;Z) am(x;z)bm(y;z) (14.30)
m=1
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
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
264 14. Iterative Solution and Stability
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
and n
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
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
bn = -2111" f(y)sinnydy
7r 0
268 15. Equations of the First Kind
iL
00
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
~71" L
00
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
270 15. Equations of the First Kind
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
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.
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
(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
Now let IIcplI = 1 and Acp =f:. O. Then choose 1/J = IIAcpll-l Acp to obtain
IIAcplI = Re(Acp, 1/J) ::; q.
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
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
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
°
(~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.
(15.11)
(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.
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
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
~ (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
k=n+1 k=n+l
with equality holding for <p = <Pn+l' Therefore /-In+l :S An+!'
15.4 Singular Value Decomposition 277
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
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
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.)
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
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
and
Furthermore
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
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)
<p (15.28)
n=1 J.tn
9n ()
x =
V2'sm (2n -2 1)'7I"x .
1
2" To(x)ao(y) + 1: Tm(x)am(y)
n
Kn(x, y) :=
m=l
(15.29)
and
lim q(a, IL)
0<-+0
= 1, 0 < IL ~ IIAII· (15.30)
282 15. Equations of the First Kind
Proof. From
00
ARaf = Lq(o:,f.1,n)(f,gn)gn
n=l
we find
00 00
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
Ra := (aI + A* A)-l A*
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
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
(15.35)
IIARm(o)f - fll :S 8.
15.5 Regularization Schemes 285
and consequently
From this we conclude that there exists a smallest integer m = m(8) such
that IIARm(eI)! - fll :S 8. 0
(15.38)
is satisfied and
(15.39)
(15.40)
and
(15.41 )
286 15. Equations of the First Kind
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)
for all <p EX. But the latter property is obvious from
for all m ~ mo, and therefore m(8) ~ mo. This implies m(8) = mo, because
m(8) ~ mo, as noted above. Observing
°
Theorem 15.27 Let A : X -+ Y be an injective compact linear operator
and let < a < 1/IIAII2. Then the bounded linear operators
m
Rm:= a L(I - aA* A)k A*
k=O
Proof. Using a singular system for A and the singular value decomposition
(15.20) applied to A* f, with the aid of
m
we can write
00 1
Rmf = 2: - {I - (1 - aJL!)m+1 } (J,9n)CPn. (15.44)
n=l JLn
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
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; .
288 15. Equations of the First Kind
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
(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
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
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?
.
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?
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)
(16.2)
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
for all <p with II<p11 :::; p. Obviously, any element <Po with II<Poil = p satisfying
for some a > 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.
294 16. Tikhonov Regularization
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
(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
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)
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
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
(16.12)
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
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
for some a °
> and IIAcpo - III = 8. Then
a Re(cpo, CPo - cp) = Re(A*(f - Acpo), CPo - cp)
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
we get
and
G'(a) = - (dJaa, A* f) -IICPaI1 2 - 2aRe (dJaa 'CPa) ,
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.
Therefore,
As in the proof of Theorem 16.7, this implies cpo -->. A-If, 8 -+ o. Then
using (16.16), we obtain
8 a E q
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.
-lb
'P' E L2[a, b] if
(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.
(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
16.5 Classical Tikhonov Regularization 303
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
By the Cauchy-Schwarz inequality this implies IIc,o - 'ljJII£2 < (b - a)c, and
the proof is complete. 0
(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
(16.24)
304 16. Tikhonov Regularization
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)
't/J(x):= l {a'P~(~)
x
- g(~)} d~, x E [a, b].
for the weights Wmk = fol Lm(y)Lk(Y) dy. Writing the approximate solution
in the form
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
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~
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.
Acp = f (17.1)
(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)
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
(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
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
312 17. Regularization by Discretization
for,,! = ("(1, ... , "!n) E (Cn we can formulate the following error estimate
which replaces Theorem 17.1.
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
(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.
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 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
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.
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
Proof. Define M by (17.15) and set hx := M(x,') and kx := K(x, .). Then
i.e., Akx = hx for all x E [a, b]. Therefore, writing f = Ar,o, we find
(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
(17.20)
with a positive definite matrix for the coefficients 11, ... , In.
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.
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).
A<p = LAj(<p,Uj)Uj,
j=l
Problems 319
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
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].
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)
r
lr cp(y)
aiJ>(x,y)
av(y) ds(y) = f(x), x E aD. (18.4)
(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.
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
For the convergence on the boundary we can state the following results
for an injective compact operator with dense range.
324 18. Inverse Boundary Value Problems
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
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
(18.7)
~ 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
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
(18.9)
L
00
L
00
and, similarly, implies a harmonic continuation into the interior of the con-
centric circle of radius R2. 0
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
'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)
330 18. Inverse Boundary Value Problems
A=SG* (18.15)
-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
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
(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,
._
r.pn·- 1 8 1 / 2 G* gn,
n- n= 1,2, ... , (18.16)
VAn
Theorem 18.11 The point z E B (with Izl ;::: 1) belongs to D if and only
if
1
L
00
L
00
and inserting this into (18.18) yields the convergence of the series (18.17).
18.3 Approximate Solution via Potentials 333
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
(18.22)
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
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
(18.24)
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.,
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
18.3 Approximate Solution via Potentials 337
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
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
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
we can write
and
(18.30)
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
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
with kernel
k(t, Tj r) = 2 [x' (T )]..L . grady <I>(x(t) , x( T)) + 21x' (T) I, t =f. T
342 18. Inverse Boundary Value Problems
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
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
(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
where
v'(z,rjr,q) = [x'(r))l..· <I>~(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 <I>, and grady <I> = - grad x <I> to transform
v'(z,rjr,q) = [x'(r))l..·gradzdivz{<I>(z,x(r))h(r)}
= -divz{[h(r))l.. :r <I>(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 <I>(z,x(r)) [h(r))l..~'(r) dr+ 10
r 27r x'(r) . h'(r)
IX'(r)1 '¢(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
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
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
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).
346 18. Inverse Boundary Value Problems
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] 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.
[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).
[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.
[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.
[29] Dahmen, W.: Wavelet and multiscale methods for operator equations.
Acta Numerica 6, 55-228 (1997).
[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).
[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
[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
[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).
[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).
[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
[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).
[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).
[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).
[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.
[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).
[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).
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)
Applied Mathematical Sciences
(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.