Author: Moiseiwitsch B.L.  

Tags: mathematics  

ISBN: 0-582-44288-5

Year: 1977

Text
                    Longman Mathematical Texts
Integral equations





Longman Mathematical Texts Edited by Alan Jeffrey and lain Adamson Elementary mathematical analysis I. T. Adamson Elasticity R. J. Atkin and N. Fox Set theory and abstract algebra T. S. Blyth The theory of ordinary differential equations J. C. Burkill Random variables L. E. Clarke Electricity C. A. Coulson and T. J. M. Boyd Waves C. A. Coulson and A. Jeffrey Optimization A. H. England and W. A. Green Integral equations B. L. Moiseiwitsch Functions of a complex variable E. G. Phillips Continuum mechanics A. J. M. Spencer 
Longman Mathematical Texts Integral equations B. L. Moiseiwitsch Department of Applied Mathematics and Theoretical Physics, The Queen's University of Belfast [> r::J [> [:][:J[:] [:][:][:] \ / Longman London and New York 
Lonwnan Group Lmuted London Associated companies, branches and representatives throughout the world Published in the United States of America by Longman Inc., New York @ Longman Group Limited 1977 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by. any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior permission of the Copyright owner. First published 1977 Library of Congress Cataloging in Publication Data Moiseiwitsch, Benjamin Lawrence. Integral equations. (Longman mathematical texts) 1. Integral equations. 2. Hilbert space. 3. Linear operators. I. Title. QA431.M57 515'.45 76-10282 ISBN 0-582-44288-5 Printed at The Pitman Press, Bath Ltd. 
Preface Many problems arising in mathematics, and in particular in applied mathematics or mathematical physics, can be formulated in two distinct but connected ways, namely as differential equations or as integral equations. In the former approach the boundary conditions have to be imposed externally, whereas in the case of integral equations the boundary conditions are incorporated within the formulation and this confers a valuable advantage to the latter method. Moreover the integral equation approach leads quite natur- ally to the solution of the problem as an infinite series, known as the Neumann expansion, in which the successive terms arise from the application of an iterative procedure. The proof of the convergence of this series under appropriate conditions presents an interesting exercise in elementary analysis. Integral equations have been of considerable significance in the history of mathematics. Thus Laplace and Fourier transforms are examples of integral equations of the first kind, while another interesting example is Abel's integral equation which is associated with Huygens' tautochrone problem and has a singular kernel. The Hilbert transform also possesses a singular kernel. It arises from a boundary value problem in a plane and enables the solution of the Hilbert type of singular integral equation to be derived. Integral transforms in general often provide a convenient method for finding the solution to the class of integral equations having kernels of the difference, or convolution, type. This book is mainly concerned with linear integral equations although a brief discussion of a simple type of non-linear equation is given at the end of the first chapter. The theory of linear integral equations of the second kind was developed originally by Volterra and Fredholm. In its more general form the analysis is carried out for Lebesgue square integrable functions since they form a Hilbert space. In this volume I have attempted to avoid unnecessary complications wherever possible by proving results for square integrable functions without usually specifying the sense in which the integration 
vi Preface is to be carried out. Thus the book can be followed, for the most part, without distinguishing between Riemann and Lebesgue inte- gration and this, I hope, will make it suitable for a wider range of mathematics students. I have devoted two chapters to Hilbert space and linear operators in Hilbert space, the integral occurring in linear integral equations being an example of a linear operator. Hilbert space is of fundamen- tal importance in mathematical physics since it provides the founda- tion for an axiomatic formulation of quantum mechanics. For this reason I have thought it worthwhile to discuss, even if rather briefly, some more general situations in which we are concerned with elements or vectors in an abstract Hilbert space acted upon by completely continuous linear operators, an example of which is the linear integral operator with square interable kernel. The final chapter is concerned with the theory of Hilbert and Schmidt on Hermitian integral operators with square integrable kernels. In this theory the solution of linear integral equations of the second kind is expanded in terms of the characteristic functions and values of the kernel. This is a common procedure in theoretical physics although its mathematical justification is often disregarded. The book concludes with a short discussion of variational principles and methods. The equations are numbered consecutively in each chapter. When referring to an equation in another chapter, the number of the chapter is inserted as a prefix but if the referenced equation is in the same chapter the prefix is omitted. The mathematical knowledge required to work through this book, and to do the problems at the end of each capter, is that which an undergraduate student should possess as a result of attending elementary courses in analysis, complex variable and linear algebra. Thus the book should be suitable for students in their final year of an honours mathematics or mathematical physics course. The problems have been chosen so as to illuminate the theory given in the main text. They are not too difficult and to gain full advantage from the book the student is strongly advised to tackle them. 
Contents Preface 1: Classification of integral equations 1.1 Historical introduction 1.2 Linear integral equations 1.3 Special types of kernel 1.3.1 Symmetric kernels 1.3.2 Kernels producing convolution integrals 1.3.3 Separable kernels 1.4 Square integrable functions and kernels 1.5 Singular integral equations 1.6 Non-linear equations Problems v 1 3 4 4 5 6 8 9 11 12 2: Connection with differential equations 2.1 Linear differential equations 14 2.2 Green's function 18 2.3 Influence function 20 Problems 22 3: Integral equations of the convolution type 3.1 Integral transforms 24 3.2 Fredholm equation of the second kind 26 3.3 Volterra equation of the second kind 31 3.4 Fredholm equation of the first kind 34 3.4.1 Stieltjes integral equation 34 3.5 Volterra equation of the first kind 36 3.5.1 Abel's integral equation 37 3.6 Fox's integral equation 39 Problems 40 4: Method of successive approximatioDS 4.1 Neumann series 4.2 Iterates and the resolvent kernel Problems 43 46 51 
Vlll Contents s: Integral equations with singular kernels 5.1 Generalization to higher dimensions 5.2 Green's functions in two and three dimensions 5.3 Dirichlet's problem 5.3.1 Poisson's formula for the unit disc 5.3.2 Poisson's formula for the half plane 5.3.3 Hilbert kernel 5.3.4 Hilbert transforms 5.4 Singular integral equation of Hilbert type Problems 53 54 55 59 60 61 63 65 67 6: HDbert space 6.1 Euclidean space 6.2 Hilbert space of sequences 6.3 Function space 6.3.1 Orthonormal system of functions 6.3.2 Gram-Schmidt orthogonalization 6.3.3 Mean square convergence 6.3.4 Riesz-Fischer theorem 6.4 Abstract Hilbert space 6.4.1 Dimension of Hilbert space 6.4.2 Complete orthonormal system Problems 69 71 74 75 76 77 79 80 82 82 83 7: Linear operators in HUbert space 7.1 Linear integral operators 85 7.1.1 Norm of an integral operator 87 7.1.2 Hermitian adjoint 88 7.2 Bounded linear operatbrs 89 7.2.1 Matrix representation 91 7.3 Completely continuous operators 92 7.3.1 Integral operator with square integrable kernel 93 Problems 95 8: e resolvent 8.1 Resolvent equation 8.2 Uniqueness theorem 8.3 Characteristic values and functions 98 99 101 
Contents ix 8.4 Neumann series 102 8.4.1 Volterra integral equation of the second kind 105 8.4.2 Bacher's example 109 8.5 Fredholm equation in abstract Hilbert space 109 Problems 111 9: Fredholm theory 9.1 Degenerate kernels 114 9.2 Approximation by degenerate kernels 120 9.3 Fredholm';, theorems 121 9.3.1 Fredholm theorems for completely continuous operators . 125 9.4 Fredholm formulae for continuous kernels 126 Problems 135 10: HUbert-Schmidt theory 10.1 Hermitian kernels 10.2 Spectrum of a Hilbert-Schmidt kernel 103 Expansion theorems 10.3.1 Hilbert-Schmidt theorem 10.3.2 Hilbert's formula 10.3.3 Expansion theorem for iterated kernels 10.4 Solution of Fredholm equation of second kind 10.5 Bounds on characteristic values 10.6 Positive kernels 10.7 Mercer's theorem 10.8 Variational principles 10.8.1 Rayleigh-Ritz variational method. Problems Bibliography Index 136 136 139 141 143 143 144 146 147 148 150 152 154 157 158 
1 Classification of integral equations 1.1 Historical introduction The name integral equation for any equation involving the unknown function cf>(x) under the integral sign was introduced by du Bois- Reymond in 1888. However, the early history of integral equations goes back a considerable time before that to Laplace who, in 1782, used the integral transform f( x) = {OO e -xs <I> (s) ds ( 1 ) to solve linear difference equations and differential equations. In connection with the use of trigonometric series for the solution of heat conduction problems, Fourier in 1822 found the reciprocal formulae ,- 00 f(x) = \j  { sin xs <I>(s) ds, (2) <I>(s) =   {OOSin xs f(x) dx (3) and f(x) =   roo cos xs <I>(s) ds, <I>(s) =  roo cos xs f(x) dx (4) (5) where the Fourier sine transform (3) and the cosine transform (5) provide the solutions cf>(s) of the integral equations (2) and (4) respectively in terms of the known function f(x). In 1826 Abel solved the integral equation named after him having the form f(x) = f(x-s)-a<l>(s) ds (6) where f(x) is a continuous function satisfying f(a) = 0, and 0 < a < 1. 
2 Classification of integral equations For a =! Abel's integral equation corresponds to the famous tautochrone problem first solved by Huygens, namely the determi- nation of the shape of the curve with a given end point along which a particle slides under gravity in an interval of time which is independent of the starting position on the curve. Huygens showed that this curve is a cycloid. An integral equation of the type cf>(x) = f(x) + A r k'(x - s)cf>(s) ds (7) in which the unknown function cf>(s) occurs outside as well as before the integral sign and the variable x appears as one of the limits of the integral, was obtained by Poisson in 1826 in a memoir on the theory of magnetism. He solved the integral equation by expanding cf>(s) in powers of the parameter A but without establishing the convergence of this series. Proof of the convergence of such a series was produced later by Liouville in 1837. Dirichlet's problem, which is the determination of a function t/1 having prescribed values over a certain boundary surface Sand satisfying Laplace's equation V 2 t/1 = 0 within the region enclosed by S, was shown by Neumann in 1870 to be equivalent to the solution of an integral equation. He solved the integral equation by an expansion in powers of a certain parameter A. This is similar to the procedure used earlier by Poisson and Liouville, and corresponds to a method of successive approximations. In 1896 Volterra gave the first general treatment of the solution of the class of linear integral equations bearing his name and characterized by the variable x appearing as the upper limit of the integral. A more general class of linear integral equations having the form cf>(x) = f(x) + r K(x, s)cf>(s) ds (8) which includes Volterra's class of integral equations as the special case given by K(x, s) = 0 for s> x, was first discussed by Fredholm in 1900 and subsequently, in a classic article by him, in 1903. He employed a similar approach to that introduced by Volterra in 1884. In this method the Fredholm equation (8) is regarded as the limiting form as n  00 of a set of n linear algebraic equations n cf>(x r ) = f(x r ) + L K(x r , x s )cf>(x s )5n s=1 (r = 1, . . . , n) (9) 
Linear integral equations 3 where 5n = (b - a)/n and X r = a + r5n. The solution of these equa- tions can be readily obtained and Fredholm verified by direct substitution in the integral equation (8) that his limiting formula for n  00 gave the true solution. 1.2 Linear integral equations Integral equations which are linear involve the integral operator L = r K(x, s) ds (10) having the kernel K(x, s). It satisfies the linearity condition L [A 1 cP 1 (s ) + A 2 cP2 (s )] = AIL [ cP 1 (s ) ] + A 2 L [ cP2 (s ) ] ( 11 ) where L[<f>(s)] = r K(x, s)<f>(s) ds (12) and Al and A2 are constants. Linear integral equations are named after Volterra and Fredholm as follows: The Fredholm equation of the first kind has the form f(x) = rK(X,S)<f>(S)dS (axb) (13) Examples are given by Laplace's integral (1) and the Fourier integrals (2) and (4). The Fredholm equation of the second kind has the form <f>(x) = f(x) + r K(x, s)<f>(s) ds (a  x  b) (14) while its corresponding homogeneous equation is <f>(x) = r K(x, s)<f>(s) ds (axb) (15) The Volterra equation of the first kind has the form f(x) = fK(X,S)<f>(S)dS (axb) (16) An example of such an equation is Abel's equation (6) which, 
4 Classification of integral equations however, has a special feature arising from the presence of a singular kernel K(x, s)=(x-s)-a (0 < a < 1) with a singularity at s = x. The Volterra equation of the second kind has the form <f>(x) = f(x) + r K(x, s)<f>(s) ds (axb) (17) We see that the Volterra equations can be obtained from the corresponding Fredholm equations by setting K(x, s) = 0 for a  x < s  b. It can be readily seen also that the Volterra equation (16) of the first kind can be transformed into one of the second kind by differentiation, for we have f x a f'(x) = K(x, x)cf>(x) + - K(x, s)cf>(s) ds a ax so that provided K(x, x) does not vanish in a  x  b we obtain f'(x) f X [ a ] <f>(x) = K(x, x) a ax K(x, s)/ K(x, x) <f>(s) ds 1.3 Special types of kernel 1.3.1 Symmetric kernels Integral equations with symmetric kernels satisfying K(s, x) = K(x, s) (18) possess certain simplifying features. For this reason it is valuable to know that the integral equation o/(x) = g(x) + r P(x, s)/L(s)o/(s) ds (19) with the unsymmetrical kernel P(x, s)f.L(s), where however P(s, x) = P(x, s), can be transformed into the integral equation (14) with symmetric kernel K(x, s) =  f.L(x) P(x, s)  f.L(s) (20) 
Special types of kernel 5 by multiplying (19) across by .J I-L(x) and setting cf>(x) = .J I-L(x) t/1(x) (21) and t(x) = .J I-L(x) g(x). (22) Real symmetric kernels are members of a more general class known as Hermitian kernels which are not necessarily real and are characterized by the relation K(s, x) = K(x, s), (23) the bar denoting complex conjugate. We shall investigate the prop- erties of integral equations with Hermitian kernels in chapter 10. 1.3.2 Kernels producing convolution integrals A class of integral equation which is of particular interest has a kernel of the form K(x, s) = k(x - s) (24) depending only on the difference between the two coordinates x and s. The type of integral which arises is L: k(x - s )cf>(s) ds (25) called a convolution or folding. This nomenclature comes from the situation which occurs in Volterra equations where the integral (25) becomes r k(x - s)cf>(s) ds (26) and the range of integration can be regarded as if it were folded at s = x/2 so that the point s corresponds to the point x - s as shown in Fig. 1. o : s : ); x x-s Fig. 1. Convolution or folding 
6 Classification of integral equations Integral equations of the convolution type can be solved by using various kinds of integral transform such as the Laplace and Fourier transforms and will be discussed in detail in chapter 3. 1.3.3 Separable kernels Another interesting type of integral equation has a kernel possessing the separable form K(x, s) = AU(X)V(s) (27) The Fredholm integral equation of the second kind (14) now becomes <f>(x) = f(x)+ AU(X) r v(s)<f> (s) ds and can be readily solved exactly. Thus we have, on multiplying (28) (28) across by v(x) and integrating: r v(x)<f> (x) dx = r v(x) f(x) dx + A r v(x) u(x) dx r v(s)<f> (s) ds which gives J b v(x) <f>(x) dx = S vx) dx a 1- AS v(x)u(x) dx (29) so that <f>(x) = f(x) + AU(X)J f(s) ds (30) 1- AS v(t)u(t) dt We see that the solution (30) can be expressed in the form <f>(x) = f(x) + A r R (x, s; A)f(s) ds (31) where R( . \ ) = u(x)v(s) X,S,I\ - 1- AS v(t)u(t) dt and is called the solving kernel or resolvent kernel. (32) 
Special types of kernel 7 The homogeneous equation (15) ,corresponding to (14) becomes in the present case cf>(X) = AU(X) r V(S ) cf>(S) ds. The solution cP(x) of equation (33) must satisfy (33) r v(x) cf>(x) dx = A r v(x) u(x) dx r v(s)cf> (s) ds The values of A for which the homogeneous equation has solutions are called characteristic values. There exists just one value of A for which (33) possesses a solution. This characteristic value Al is given by 1 = Ai r v(x) u(x) dx, the associated characteristic solution of (33) being cPl(X) = cu(x) where c is an arbitrary constant. The simple separable form (27) is a special case of the class of degenerate kernels (34) n K(x, s) = A L Ui(X) Vi(S) i=l (35) gIvIng rise to integral equations which can be solved In closed analytical form as we shall show in chapter 9. Example 1. As a simple illustration of an integral equation with separable kernel (27) we consider cf>(x) = 1 + A r xscf>(s) ds (36) Here f(x) = 1, K(x, s) = AXS and so u(x) = x and v(s) = s. It follows that the resolvent kernel is xs R (x, s; A) = 1 _ A S5 t 2 d t xs (37) 1 - A/3 
8 Classification of integral equations and hence the solution to (36) is AX [1 cf>(X) = 1 + 1 _ A/3 Jo s ds =1+ 3h (A3) 2(3 - A) (38) Example 2. Another example of an integral equation with a separ- able kernel is cf>(X) = eX + A r eia(x-s)cf>(s) ds (39) where f(x) = eX, K(x, s) = Aeia(x-s) so that u(x) = e iax and v(s) = e ias Then the resolvent kernel is e ia(x-s) R(X,S;A)= I-A (40) and hence the solution of (39) is Ae iax i 1 . cf>(x) = eX + e(1-1a)s ds I-A 0 AeiaX(el-ia -1) =e x + (1- A)(I- ia) (41) 1.4 Square integrable functions and kernels Functions cf>(x) which are square integrable in the interval a  x  b satisfy the condition flcf>(XW dx <00 (42) where the integral is taken to be Riemann, or Lebesgue for greater generality. In the former case it is said that cf>(x) is an R 2 function and in the latter case that it is an L 2 function. Continuous functions are square integrable over a finite interval since they are bounded. However the converse is not necessarily true, that is square integra- ble functions need not be continuous or bounded. Kernels K(x, s) defined in a  x  b, a  s  b are said to be square integrable if they satisfy. r fIK(X, sW dx ds<oo (43) 
Singular integral equations 9 together with fIK(x, sW ds <00 (axb) (44) and fIK(x, sWdx<oo (asb) (45) where the integrals are taken to be Riemann or Lebesgue. The kernel is then called R 2 or L 2 respectively. Kernels of particular interest are those which are singular. Thus consider a Volterra kernel of the form F(x, s) K(x,s)= (x-s)a o (x < s) where F(x, s) is a continuous function and 0< a < 1. Then IF(x, s)1  M where M is a constant and so (x> s) (46) f b f b 2 f b f x 1 F( x, s) 1 2 a a IK(x, s)1 dx ds = a dx a ds (X _ S)201 :so: M 2 f dx r ds(x - S)-201 M 2 f b = dx(x - a)1-2a 1-2a a M 2 (b _ a)2-2a 2(1- 2a)(1- a) (a < !) (47) and so the double integral is finite if 0 < a <!. However if a ! the singular kernel (46) is not square integrable. 1.5 Singular integral equations An integral equation of the type cf>(x) = f(x) + A r K(x, s )cf>(s) ds (48) is said to be singular if the range of definition is infinite e.g., 0< x < 00 or -00 < x < 00, or if the kernel is not square integrable. 
10 Classification of integral equations Non-singular equations have a discrete spectrum, that IS the associated homogeneous equation cf>(X) = A r K(x, s )cf>(s) ds (49) has non-trivial solutions cPv(x) for a finite or at most a countable, although infinite, set of characteristic values Av of the parameter A. Also each characteristic value Av has a finite rank (or index), that is it has a finite number of linearly independent characteristic func- tions cPl)(X),..., cP)(x). However if the integral equation is singular by virtue of having an infinite range of definition, the spectrum of values of A may include a continuous segment. For example it may be readily verified that the homogeneous equation cf>(x) = A L e-1x-s1cf>(s) ds (50) has solutions of the form cP(x) = cle- v'1-2,\, X+ c2e v'1-2,\, X (51) for the continuous spectrum of values 0 < A < 00. Equation (50) is of the convolution or difference type and will be discussed further in section 3.2 of the chapter on integral equations of the convolution type. Also if the integral equation has an infinite range of definition the characteristic values may have an infinite rank. Thus consider the equation cf>(x) = A   [00 cos xs cf>(s) ds (52) with kernel K(x, s) =   cos xs Using the Fourier cosine transform (5) and its reciprocal formula (4) we have (53) Acf>(x) =   [00 cos xs cf>(s) ds and so cf>(X) = A 2cf>(X) 
Non-linear equations 11 giving A 2 = 1, or A = :f: 1 as characteristic values. Their ranks are infinite since it can be shown without difficulty that ( ) 1'7T -ax a ( ) cf> x = -V 2 e :f: a 2 +x2 x >0 (54) are characteristic functions corresponding to the characteristic val- ues A = :f: 1 for all values of a> O. The kernel (53) belongs to a class of singular kernels known as Weyl kernels. 1.6 Non-linear equations In the case of non-linear equations, the spectrum of characteristic values A may have the interesting property that the number of solutions of the integral equation changes as we pass through particular values of A known as bifurcation points. As an example we consider the simple non-linear equation cf>(x) - A[ s{ cf>(s)Y ds = 1 (55) and investigate its real solutions cf>(x). We put [S{cf>(S)}2 ds = a (56) in which case we have cf>(x) = 1 + Aa (57) so that a = (1 + Aa)2 [s ds = t(1 + Aa)2 This yields A 2 a 2 +2(A-1)a + 1 = 0 (58) which gives 1- A :f:v'(A -1)2- A 2 a= 2 A (59) Hence 1 :f: v' 1- 2A cf>(x) = A . (60) 
12 Classification of integral equations Thus there are two real solutions of the non-linear integral equation (55) if A <!, one real solution cP(x) = 2 if A =! and no real solutions if A >!. This means that A =! is a bifurcation point. When A = 0 one of the solutions is cP(x) = 1 while the other solution is infinite. Thus A = 0 is a singular point. We now look at the associated homogeneous equation cp(x) = A[ s{cp(S)}2 ds (61) This gives cP(x) = Aa (62) so that 1 = A 2 a {\ ds =!A 2a yielding 2 a=- A 2 (63) and the solution 2 cP(x) =-. A (64) In general non-linear equations present considerable difficulties and we shall not consider them again after this chapter. Problems 1. Transform the Volterra integral equation of the first kind x= f(ex+eS)cp(S)dS into a Volterra equation of the second kind. Show that the solution cP satisfies a first order differential equation and hence solve the integral equation. 2. Solve the Fredholm equation of the second kind cp(x) = x + A[ cp(s) ds and show that the characteristic value of the associated homogene- ous equation is A 1 = 1. 
Problems 13 3. Solve the Fredholm equation cp(x) = 1 + A[ eX+scp(s) ds and show that the characteristic value of the associated homogene- ous equation is At = 2/( e 2 -1). 4. Solve the Fredholm equation cp(x) = 1 + A[xnsmcp(s) ds and show that the characteristic value of the associated homogene- ous equation is At = m + n + 1. 5. Solve the Fredholm equation cp(x) = x + AfT sin nx sin ns cp(s) ds where n is an integer, and show that the characteristic value of the associated homogeneous equation is At = 2/11'. 6. Show that the singular integral equation cp(x) = A -J! rOSin xs cp(s) ds has characteristic solutions ( ) r;. -ax X cf> X = -V "2 e :f: a 2 + X2 associated with characteristic values A = :f: 1 for all a > O. (X> 0) 7. Solve the non-linear equation cp(x) = 1 + A[{cp(S)}2 ds and show that A = * is a bifurcation point while A = 0 is a singular point of the spectrum of characteristic values. 
2 Connection with differential equations 2.1 Linear differential equations We first observe that the first order differential equation dcf> dx = F(x, cf» (axb) (1) can be written immediately as the Volterra integral equation of the second kind cp(x) = cp(a)+ r F[s, cp(s)] ds. (2) As an interesting but simple example of the above we consider the following problem, solved by Johannes Bernoulli for the n = 3 case: To find the equation y = cf>(x) of the curve joining a fixed origin o to a point P such that the area under the curve is l/nth of the area of the rectangle OXPY having OP as diagonal for all points P on the curve (see Fig. 2). Evidently this problem is equivalent to solving the homogeneous V olterra integral equation of the second kind I x 1 cf>(s) ds = - xcf>(x) o n (3) for the unknown function cf>(x) where c/>(O) = O. This equation can be solved by converting it into the differential equation 1 <!>(x) = - {<!>(x) + x<!>'(x)} n 
Linear differential equations 15 y p V o <P (x) <P (s) s x x Fig. 2. Bernoulli's problem for n = 3; y = x 2 . or more simply n-1 cf>'(x) = cf>(x) x (4) which has the solution <f>(x) = Ax n - 1 (5) For the case n = 3 solved by Bernoulli the curve is a parabola y = Ax 2 . We next consider the second order differential equation d 2 <f> dxz =F(x,cP) (axb) (6) 
16 Connection with differential equations which can be expressed as 4>(x) = 4>(a) + (x - a)4>'(a) + r (x - s)F[s, 4>(s)] ds, (7) this particular form being obtained on carrying out an integration by parts. Taking a = 0 and b - a = I this may be rewritten as 4>(x) = 4>(0) + 4>( I) - 4>(0) x I + f(X - s)F[s, 4>(s)] ds +} r (s -l)F[s, 4>(s)] ds (8) sInce 14>'(0)= 4>(1) - 4>(0) + r (s -1)F[s, 4>(s)] ds. If the differential equation is linear and we write F[x, cf>(x)] = r(x) - q(x)cP(x) (9) we obtain the Fredholm integral equation of the second kind 4>(x) = f(x)+ r K(x, s)q(s)4>(s) ds (10) where cf> ( I) - cf> ( 0 ) i I f(x) = 4>(0) + 1 x - 0 K(x, s)r(s) ds (11) and K(x, s) = s(l-x) 1 x(l- s) I (s  x) (s  x) (12) In the particular case of a flexible string having mass p (x) per unit length, stretched at tension T and vibrating with angular frequency w, we have q(x) = w 2 p(x)/T and r(x) = O. If the string has fixed ends at x = 0 and I, the transverse displacement cP(x) satisfies cP(O) = cP(l) = 0 so that f(x) = O. Thus we arrive at the homogeneous Fredholm 
Linear differential equations 17 integral equation cf>(x) = r K(x, s)q(s)cf>(s) ds (13) An important distinction between the differential equation and the equivalent integral equation approach should be observed. In the case of the differential equation formulation of a physical problem the boundary conditions are imposed separately whereas the integral equation formulation contains the boundary conditions implicitly. Thus, for example, we see at once from (13) and (12) that cf>(x) vanishes at x = 0 and x = I. Let us now consider the second order linear differential equation Lv(x) = r(x) (0  x  I) (14) where L is the linear differential operator given by L= d {P(X) ddJ +q(x) (15) and p(x) has no zeros in the range of definition 0  x  I. We shall treat this equation somewhat differently from the previ- ous equations by setting dZv cf> (x) = dx z . (16) Then we have dv r x dx = v'(O) + Jo cf>(s) ds (17) and v (x) = v (0) + xv' (0) + r (x - s) cf> (s) ds, from which it follows that the differential equation may be ex- pressed as the Volterra integral equation of the second kind (18) cf>(x) = f(x) + r K(x, s)cf>(s) ds (19) where f(x) = {p(x)}-l[r(x) - v'(O)p'(x) -{v(O) + xv'(O)}q(x)] (20) 
18 Connection with differential equations and K(x, s) = {p(X)}-l[(S - X)q(X) - p'(X)]. (21) 2.2 Green's function We consider the second order linear differential equation Lv(x) = r(x) (axb), (22) where L is the linear differential operator (15), and suppose that its solution v(x) satisfies homogeneous boundary conditions at x = a and x = b, namely av(a) + {3v'(a) = 0 av(b) + {3v'(b) = 0 (23) where a and {3 are prescribed constants. Let LVi(X)=O (i=1,2) (24) where Vl(X) and V2(X) are two linearly independent functions which satisfy the boundary conditions (23) prescribed at x = a and b respectively. We shall show that the solution of (22) may be expressed in the form v(x) = r G(x, s)r(s) ds (25) where G(x, s) = 1 A Vl(X)V2(S) 1 A V2(X)Vl(S) (x  s) (x  s) (26) and A is a certain constant. G(x, s) is called the Green's function. The method we shall use involves the Dirac delta function 8(x) which vanishes for x  0 and satisfies r 8(x-s)r(s) ds = r(x) (27) Clearly 5(x) is not a function as defined in the usual sense but 
Green's function 19 belongs to a class known as generalized functions. A useful represen- tation of the Dirac function is 8(x) = lim I(h, x) hO (28) where I(h, x) is the impulse function defined as 1 I(h, x) = h o (Oxh) otherwise. (29) Thus the representation (28) vanishes for x ¥= 0 and has infinite magnitude at x = O. Further lim f oo I(h, x) dx = 1, (30) h  0 -00 since the integral is unity for all values of h, and we write this as L 8(x)dx=1 (31) although care has to be exercised when interchanging the order of the limiting operation and the integration. Now (25) has to provide a solution of (22) and so LG(x, s) = 8(x - s). (32) It follows that r S + B ! Js-e LG(x, s) dx = 1 (33) which leads to [ a ] X=S+B 1 lim - G(x, s) =-, BO ax X=S-B p(s) (34) that is, Vl(S)V(S)-V2(S)vHs) = p) (35) or Vl (s) V  (s) V2(S) v(s ) A - p(s) (36) 
20 Connection with differential equations where the determinant on the left-hand side is called the Wronskian. For A to be non-vanishing it is evident that Vl and V2 must be linearly independent. As a simple example we consider the operator d 2 L=-c- (Oxl) (37) dx 2 and the end conditions v(O):c v(l) = O. Then p(x) = -c and Vl(X) = x, V2(X) = 1- x so that A = cl giving for the Green's function x(l- s) cl G(x, s) = (1- x)s cl (x  s) (xs) (38) 2.3 Influence function . We consider a string of length 1 at tension T. The displacement G(x, s) of the string at the point with coordinate x, resulting from the application of a unit force perpendicular to the string at the point with coordinate s, is called the influence function. Now the displaement G(s, s) of the string at the point s, assumed to be small, satisfies the equilibrium equation (see Fig. 3): TG(S,S) e + ,1S )=1 (39) which yields G( ) = s(l-s) s, s Tl and so it follows that the influence function is given by x(l- s) TI s(l- x) Tl (40) G(x, s) = (x  s) (x  s) (41) o s L Fig. 3. Displaced string. 
Influence function 21 Comparison with (38) shows that the influence function 'derived above is just the Green's function for the linear differential operator d 2 L=-T- dx 2 (42) Now suppose that we apply a loading force F(x) per unit length to the string. Then the displacement of the string at the point with coordinate x is given by cf>(x) = r G(x, s)F(s) ds (43) using the principle of superposition. In the case of a horizontal wire at tension T, possessing mass p(x) per unit length, the loading due to the force of gravity is F(x) = -gp(x) and so the displacement of the wire becomes cf>(x) = - g r G(x, s )p(s) ds. (44) As a further example we consider a shaft of length I and mass p(x) per unit length. If the loading at the point with coordinate s is given by F(s) per unit length and G(x, s) is the influence function of the shaft, then the displacement ck(x) at the point x is again given by (43). Let us now suppose that the shaft is rotating with angular velocity w. Then the loaqiag due to the rotation is given by the centrifugal force - F(s) = w 2 p(s)cp(s), (45) and so we obtain the following homogeneous integral equation for the displacement of the shaft cf>(x) = w 2 r G(x, s)p(s)cf>(s) ds (46) The balancing of the elastic and centrifugal forces can occur only for certain discrete values of w known as the critical speeds which correspond to values of w 2 for which the integral equation (46) has non-vanishing solutions, that is its characteristic values. 
22 Connection with differential equations Problems 1. Solve the Volterra equations of the second kind (i) cf>(x) = 1 + r cf>(s) ds (ii) cf>(x) = 1 + 2 r scf>(s) ds by finding the solutions of the equivalent differential equations. 2. Solve the Volterra equation xcf>(x) = x n + r cf>(s) ds 3. Solve the Volterra equation cf>(x) = f(x) +,\ r (x - s)cf>(s) ds for (i) f(x) = 1, (ii) f(x) = x and A = :t: 1. 4. Solve the Volterra equation cf>(x) = f(x) +,\ r sin (x - s)cf>(s) ds for (i) f(x) = x and A = 1, (ii) f{x) = e- x and A = 2. 5. Show that the Volterra equation of the first kind fX(x-s)n-l f(x) = Jo (n -1)! cf>(s) ds has the continuous solution cP(s) = f(n)(s) when f(x) has continuous derivatives f'(x), f"(x),..., f(n)(x), and f(x) and its first n-1 derivatives vanish at x = o. 6. Show that the nth order linear differential equation dny n dn-ry d n -'\L a,.(x) d n-r - f(x) X r=l X where y(O) = y'(O) = · · · = y(n-l)(o) = 0, is equivalent to the Volterra 
Problems 23 equation of the second kind r x n a (x)(x-s)r-l «f>(x) = f(x) + A Jo rl r (r-1)! «f>(s) ds where cP(x) = dny/dx n . 7. Show that the solution of the differential equation d2cP + W2cP = F[x, cP(x)] (O x  l) dx obeying the end conditions cP(O) = cf>(l) = 0, satisfies the integral equation «f>(x) = r G(x, s)F[s, «f>(s)] ds where G( ) = { -(w sin wl)-l sin wx sin w.(l- s) (x  s) x, s -(w sin wl)-l sin w(l- x) sin ws (x  s) is the Green's function for the linear differential operator d 2 /dx 2 + w 2 . 
3 Integral equations of the convolution type In the present chapter we shall be concerned with integral equations whose kernels have the form K(x,s)=k(x-s) (1) which is a function only of the difference between the two coordi- nates x and s. The method of solution generally involves the use of integral transforms. Hence before proceeding to consider the vari- ous kinds of convolution type integral equations we shall briefly discuss the more important forms of integral transform. 3.1 Integral transforms We have already mentioned Laplace and Fourier transforms in the historical introduction given in chapter 1. The Laplace transform of a function cP(x) is given by <I>(u) = l°Oe-uxq,(X)dX (2) and the basic result which enables us to solve integral equations of the convolution type is the convolution theorem which states that if K(u) is the Laplace transform of k(x) then K(u)<I>(u) is the Laplace transform of r k(x - s)q,(s) ds (3) To verify that this is correct, without any attempt at a rigorous argument, we see that the Laplace transform of (3) is 1 00 e- UX dx r k(x - s)q,(s) ds = 1 00 e-utk(t) d{OO e-USq,(s) ds = K(u)<I>(u) on setting t = x - s and noting that we may take k(x) to vanish for x<O. 
Integral transforms 25 In addition to the sine and cosine transforms (2) and (4) referred to in section 1.1, we have the exponential Fourier transform of a function cP (x) defined as <I>(u) =  foo e iux 4>(x) dx' (4) v2Tr -00 and the reciprocal formula 1 f oo . cP(x) = r;:- e-1UX<I>(u) duo v 2Tr -00 Then, if K(u) is the Fourier transform of k(x), the convolution theorem for Fourier transforms states that .J2Tr K(u)<I>(u) is the Fourier transform of (5) L k(x - s)4>(s) ds (6) which can be easily verified as for the Laplace transform. Another valuable integral transform is the Mellin transform <1>( u) = [00 x U - l 4>(x) dx (7) and the corresponding convolution theorem states that K(u)<I>(l- u) is the Mellin transform of [00 k(xs)4>(s) ds where K(u) is the Mellin transform of k(x). In fact, again without any attempt at rigour, we have that the Mellin transform of (8) is [00 x u - 1 dx [00 k(xs)4>(s) ds = [00 t u - 1 k(t) d{OO s-u4>(s) ds = K(u)<I>(l- u) (8) on setting t = xs. Throughout the following sections we shall assume that all the functions arising in the integral equations satisfy suitable conditions which permit all the transformations to be performed with validity.* * For details see E. C. Titchmarsh, Introduction to the Theory of Fourier Integrals, 2nd ed., Oxford University Press, 1948. 
26 Integral equations of the convolution type 3.2 Fredholm equation of the second kind We consider firstly Fredholm equations of the type cf>(x) = f(x)+ f k(x - s)cf>(s) ds (-00< X <00) (9) A formal solution of this equation can be obtained by introducing the Fourier transforms 1 f oo <I>(u) = - eiUX<f>(x) dx, J2; -00 1 f oo F(u) = r:::- eiUXf(x) dx, "2,,, -00 1 f oo K(u) = r:::- eiUXk(x) dx. "2,,, -00 Using the convolution theorem (10) (11) (12) 1 f oo K(u)<I>(u) = T r;:- k(x - s)c/>(s) ds "2,,, -00 (13) where T denotes the Fourier integral operator 1 f oo T=- e iux dx J2; -00 ' (14) we obtain <I>(u) = F(u) + 2", K(u)<I>(u) (15) as a result of operating with T on both sides of (9). This gives <I>(u) = F(u) 1- ../2", K(u) (16) which provides the solution to the integral equation (9) in the form <f>(x) = T- 1 [<I>(u)] 1 f oo e-iXUP(u) = J2Tr -00 1- ../2Tr K(u) du (17) on using the reciprocal formula (5). 
Fredholm equation of the second kind 27 Let us now consider the homogeneous equation 4>(x) = L k(x - s)4>(s) ds. (18) The solution to this can be written n cP(x) = L L cv,pxP-le-iwvX v p=l (19) where the cv,p are constants, the W v are the zeros of 1- .J2'Tr K(u) and n is the order of the multiplicity of the zero W v . Thus we have that 1 = L k(t)eu"pt dt (20) on setting u = W v in (12), and o = L k(t)tP-1ei"'pt dt (p = 2, . . . , n) (21) on differentiating both sides of (12) p -1 times with respect to u and setting u = W v . Hence f 00 ( t ) P-l 1 = Lx> k(t) 1- x eu"pt dt (22) and so, setting t = x - s, we obtain xp-1e-i"'px = L k(x - s)sP-1e-u"ps ds (23) which shows that each term in the sum on the right-hand side of (19) is a solution of (18). Example 1. As a first example we consider the case f(x) = e- 1xl k(x) = {,\x (x < 0) (x >0) (24) (25) Then F(u) = fco e-lxl+iUX dx J2; -00 12 1 =V ; 1+u 2 (26) 
28 Integral equations of the convolution type while K(u) =  r e X + iUX dx 5;, -r¥J A 1 - 5;, 1 + iu (27) and so it follows that the solution is given by 1 f r¥J e-iXu du 4>(x) = 7T -co (1- iu)(1-'\ + iu) Then if 0 < A < 1 we may apply Cauchy's residue theorem to evaluate this integral obtaining the particular solution (28) 2e- x cP(x)= 2-'\ 2e(l-A)X 2-A (x > 0) (x < 0) (29) Since there is just one zero, having order of multiplicity 1, of 1- v'21T K(u) = (1- A + iu)/(l + iu) at u = i(l- A), it follows that the solution of the associated homogeneous equation 4>(x) = ,\ l co e x - s 4>(s) ds (30) is just cP(x) = Ce(l-A)X as can be readily verified by inspection. Hence the general solution is given by 2e- x C (l-A)X + e cP(x)= 2-A ( 2 + C ) e(l-A)X 2-A (31) (x > 0) (x < 0) (32) where A is confined to the range of values 0 < A < 1. Example 2. As a second example we discuss the Lalesco - Picard equation for which k(x) = Ae- ixi . (33) 
Fredholm equation of the second kind 29 We shall suppose that -f(x) and hence F(u) are unknown, in which case it is convenient to rewrite the solution (17) in the form cf>(x) = f(x)+ t: e-iXUP(u)M(u) du (34) where M(u) = K(u) 1-fu K(u) (35) and 1 J oo . f(x) = r:::- e-JXUP(u) duo v 2", -00 (36) For then we have, using the convolution theorem for Fourier transforms t: e-ixUP(u)M(u) du = t: f(u)m(x - u) du (37) where 1 J oo . m(x) = r;:- e-JXUM(u) du, v 2", -00 (38) that the solution (34) is given by cf>(x) = f(x)+ t: f(u)m(x - u) duo (39) Now in the present example K(u)= J'" Ae-lxl+iUX dx fu -00 =   1:u 2 ' (40) and so  - 2 A M(u)= 1T1+u 2 -2A (41) 
30 Integral equations of the convolution type from which it follows that 1 roo Ae -ixu m(x) = 1T Lo 1 + u2-2'\ du A - v'1-2A Ixl - 1-2A e (42) provided A <! and using Cauchy's residue theorem. Hence a par- ticular solution of the integral equation is 4>(x) = !(x) + ,\ foo !(u)e- v'1-2,\ lx-u l duo (43)  1-2A -00 Turning our attention to the homogeneous equation 4>(x) =,\ L e- 1x - sl 4>(s) ds (44) we note that there are two zeros, both having multiplicity of order 1, of 1+u 2 -2A 1-.J2; K(u) = 1 2 +u (45) at u = :t:i  l- 2A, so that the solution of (44) takes the form cP(x) = Cl e - v'1-2A X + c2e v'1-2A X (46) where, fo r the i ntegral occurring in (44) to have a meaning, the real part of  1 - 2A must satisfy Re  1-2A< 1. (47) Thus all values of A in the range 0 < A < 00 are allowed, but if A >! the solution may be more suitably written as cP(x) = alsin(  2A-1 x)+a2cos(  2A-1 x). (48) However when A =!, 1- 2Tr K(u) has a single zero at u = 0 with multiplicity of order 2 in which case the general solution of the homogeneous equation is cP(x) = Cl + C2 X . (49) 
V olte"a equation of the second kind 31 3 .3 Volterra equation of the second kind If we set f(x)=O, k(x)=O, and cf>(x) =0 for x<O in the Fredholm equation (9) of the second kind considered in section 3.2 we obtain the integral equation </>(x) = f(x)+ r k(x -s)</>(s) ds (x > 0) (50) which is a V olterra equation of the second kind with a convolution type integral. Equation (50) can be solved most conveniently by introducing the Laplace transforms cI>(u) = ro e- UX </> (x) dx, F( u) = {'" e -UXf( x) dx, K(u) = r"" e-UXk(x) dx. Jo . (51) (52) (53) Applying the convolution theorem for Laplace transforms K(u)cI>(u) = L r k(x - s)</>(s) ds . 0 . . (54) where L denotes the Laplace integral operator L = L"" e- ux dx, (55) we find that cI>(u) = F(u) + K(u)cI>(u) (56) which yields F(u) cI>(u) = 1- K(u) = F(u)+ F(u)M(u) (57) where M(u) = K(u) 1- K(u)" (58) 
32 Integral equations of the convolution type Hence the solution of (50) can be written as «(>(x) = f(x) + r m(x - s)f(s) ds (59) where M(u) = I'O e-UXm(x) dx. (60) Example 1. We examine first the simple case where the kernel is k(x) = {x (x > 0) (x < 0) (61) Then we have 1 00 A- K(u)=A- e- ux xdx=2 o u (62) so that A- M(u)= 2 A- u -  ( 1 1 ) =2 u--u+ (63) and hence ( )  ( .J>...x -X ) mx=-e -e . 2 (64) Thus the solution to the integral equation (50), as given by (59), is «(>(x)=f(x)+  fXf(s){e(X-S)_e-(X-S)}ds (65) 2 Jo Example 2. Our next example has the kernel k(x) = {AX (x >0) (x < 0) (66) Then 1 00 A K(u) = A- e(1-u)x dx =- o u-l (67) 
Volterra equation of the second kind 33 so that A M ( u ) - . u-(A+1) (68) and hence m(x) = Ae(A+l)x. (69) Thus, using (59) the solution of the integral equation (50) is «(>(x) = f(x) + Af f(s)e(Hl)(X-S) ds (70) Example 3. Our last example has the kernel k(x) = { A s o in x (x > 0) (71) (x < 0) Then 1 00 A K(u) = A e- ux sin x dx = 2 1 o u + (72) and so A M ( u ) - - u 2 +1-A (73) which yields A sin ( .J 1- A x) .J 1-A (A < 1) m(x) = AX (A = 1) A ../ sinh ( ../ A -1 x) (A > 1) A-I (74) on using the appropriate Laplace transform formulae. Hence the solution of the integral equation (50) is given by A J x ../ f(s) sin { ../ 1- A (x-s)} ds I-A 0 «(>(x)-f(x)= ff(S)(X-S)dS A J x ../A-1 0 f(s)sinh{ ../ A-1 (x-s)}ds (A < 1) (A = 1) (A > 1) (75) 
34 Integral equations of the convolution type 3.4 Fredholm equation of the first kind Here we are concerned with Fredholm equations of the form f(x) = t: k(x - s )cf>(s) ds (76) with an integral of the convolution type on the right-hand side. To solve this equation we take Fourier transforms. On applying the convolution theorem (13) this yields F(u) =.J2; K(u)cI>(u) (77) which leads to the formal solution of (76): 1 f oo . cf>(x) = r;:- e -IXUcI>(u) du "" 21T -(X) = roo e- ixu F(u) duo 21T J-oo K(u) (78) 3.4.1 Stieltjes integral equation An interesting example of a Fredholm equation of the first kind possessing a convolution type integral is obtained by applying Laplace's integral operator LOO e- xs ds twice. Thus consider the equation g(x) = A Loo e- xt d{OO e-tsl{I(s) ds. (79) Then, on reversing the order of the integrations, we see that g(x) = A Loo l{I(s) ds Loo e -(x+s)t dt = A L oo l{I(s) ds (80) x+s which is known as Stieltjes integral equation. Now setting x- = e-, s = eO", e1/2g(e) = f(), el/2t/J(e) = cf>() (81) 
Fredholm equation of the first kind 35 we obtain A f oo cf>(u) f() = 2 h .!( _ ) du. -00 COS 2 U (82) This is a Fredholm equation of the first kind with kernel A k()= 2 hl ' cos 2: (83) To solve (82) we require to find the Fourier transforn1 of k() which is given by A f oo eiu K(u)=- d J2; -00 e 1/2 + e -1/2 A 1 00 Xiu-(1/2) =- dx J2; 0 l+x (84) on putting x = e. But it is well known that 1 00 a-1 X dx = '1T o 1 + x sin '1Ta (85) and so we obtain A ; A ; A ; K(u) = = = (86) sin {'1T(iu + !)} cos i'1TU cosh '1TU Now, making use of the solution (78) to the equation (76), we see that the solution of our integral equation (82) may be written as 4>() =  f"" F(u) cosh 1TU e-il;u du A'1T 2 '1T -00 = 1 f"" F(u){e-i(I;HJT)U + e-i(!;-i....)U} du (87) 2 '1T A J2; -00 where F(u) is the Fourier transform of f(). Thus 1 cf>() = 2'1TA {f( + i'1T) + f( - i'1T)} (88) and hence the solution to Stieltjes integral equation (80) is l . . tfJ(x) = 2'1TA {g(xel'IT) - g(xe-I'IT)}. (89) 
36 Integral equations of the convolution type Example. Suppose that x In- a g(x)= x-a 1 (x;e a) (x = a) (90) a Then (89) provides the solution 1 l(1(x) = '\'(x + a) (91) to Stieltjes equation (80). 3.5 Volterra equation of the first kind We now discuss Volterra equations of the form f(x) = f k(x - s )«(>(s) ds (x > 0), (92) where f(O) = 0 and the integral on the right-hand side is of the convolution type again. This can be obtained from the Fredholm equation (76) of the first kind, examined in the previous section, by taking f(x) = 0, k(x) = 0 and cf>(x) = 0 for x <0. To solve equation (92) we introduce the Laplace transforms F(u), K(u) and <I>(u) of f(x), k(x) and cf>(x). Then employing the convolution theorem for Laplace transforms we obtain F(u) = K(u)<I>(u) (93) and so «(>(x) = L-l{ :: } where L -1 denotes the inverse of the Laplace transform. Example. Consider the integral equation f(x) = fJo(X-S)«(>(S)dS (94) (95) with a kernel of the convolution type given by k(x) = Jo(x) (96) 
Volterra equation of the first kind 37 where Jo(x) is the zero order Bessel function satisfying 1 0 (0) = 1. We can solve this equation by taking Laplace transforms. Since 1 L{Jo(x)} = .J 2 U +1 (97) it follows that cI>(u) =  U2 + 1 F(u) using (93). For the special case f(x) = sin x (98) (99) we have 1 F( u) = 2 1 u + (100 ) and then 1 <1>( u) = / 2 · '\/u +1 We see at once from (97) that for this case cP(x) = Jo(x) (102) (101) which produces the interesting formula [Jo(X - s)Jo(s) ds = sin x. (103) 3.5.1 Abel's integral equation The equation solved by Abel has the form f(x) = [(x-s)-aq,(S)dS (0<a<1) (104) with f(O) = 0, which is a Volterra integral equation of the first kind with a singular kernel of the convolution type given by k(x) = x-a. (105) This equation is of c<?nsiderable historical importance since for a =! it describes the tautochroRe problem introduced briefly in section 1.1. 
38 Integral equations of the convolution type Now we have K(u) = [00 e-uxx- a dx = r(l- a)u a - 1 where r(p) is the gamma function defined by f(p) = [00 e -xx p - 1 dx (p > 0), (106) (107) so that F(u)u 1 - a L{«f>(x)}= r(1-a) (108) using (93). Let us set cf>(x) = t/J'(x) (109) where t/J(O) = 0, so that L{ cf>(x)} = uL{ t/J(x)} (110) on performing a single integration by parts. Then F(u)u- a L{tf1(x)} = f(l-a) (111) and since we may write -a L(x a - 1 ) u = r(a) , it follows from the convolution theorem for Laplace transforms that tf1(x) = {f(a)r(I- a)}-l r (x - s)"'-lf(s) ds. (112) U sing the well known result '1T r(a)r(l- a) = . sIn '1Ta (113) we see that the solution of Abel's integral equation (104) is «f>(x) = sin 'ITa  r x (x _ s)"'-lf(s) ds. (114) '1T dx Jo 
Fox's integral equation 39 For the particular case a =! corresponding to the tautochrone problem this solution simplifies to 1 d I x f(s) cf>(x) = - - d ( )1/2 ds. '1T X 0 x-s (115) 3.6 Fox's integral equation We seek the solution of the Fredholm integral equation of the second kind having the form <f>(x) = f(x) + L"" k(xs)<f>(s) ds (O<x<oo) (116) named after Fox. This can be achieved by employing the Mellin transforms <I>(u), F(u) and K(u) of cf>(x), f(x) and k(x) respectively defined by (7). Now we have shown in section 3.1 that K(u)<I>(l- u) is the Mellin transform of L"" k(xs)<f>(s) ds and so we see that <I>(u) = F(u) + K(u)<I>(l- u) (11 7) on taking the Mellin transforms of both sides of (116). Also, replacing u by 1 - u, we have <1>(1- u) = F(l- u) + K(l- u)<I>(u) (118) which enables us to write cI>(u) = F(u) + K(u)F(l- u) 1- K(u)K(l- u) (119) Thus we have obtained a solution of Fox's integral equation (116) provided we can derive cf>(x) from its Mellin transform <I>(u). Example. Let us take as an example k(x) = A .Jl: sin x (120) 
40 Integral equations of the convolution type Then the Mellin transform of k(x) is given by K(u) = A  IX> x u - 1 sin x dx = A  f(u) sin u (121) and 2A 2 . u u K(u)K(l- u) = --:;; f(u)f(l- u) SIn 2 cos 2 = A 2f(u)f(1- u) sIn u  =A 2 using (113). Hence, provided A 2  1, we see that (122) F(u) A /2 . u <I>(u) = 1- A 2 + 1- A 2 -y  f(u) SIn 2 F(l- u) making use of (119). But by the convolution theorem for Mellin transforms established in section 3.1 we know that f( u) sin u F( 1 - u) is the Mellin transform of IX> sin xsf(s) ds, and so the solution of Fox's integral equation with kernel given by (120) is (123)  - (x) A 2 00. <fJ(x) = / 2 + 1 2 _ 1 SIn xsf(s) ds -A -A  0 (O<x<oo) (124 ) This result can be verified directly by using Fourier's reciprocal sine formulae (1.2) and (1.3) in the form 2 1 00 1 00 f(x) = - sin xs ds sin stf(t) dt  0 0 Problems (O<x<oo) (125) 1. Solve the Lalesco- Picard equation <fJ(x) = cas JLX + A L e-lx-sl<fJ(s) ds (A <!) 
Problems 41 2. Find the solutions to the Volterra equations of the second kind <fJ(x) = f(x) =i: r (x - s)<fJ(s) ds when (i) f(x) = 1, (ii) f(x) = x using the general solution (65), and verify that they agree with the solutions obtained by solving the equivalent differential equations (see problem 3 at the end of chapter 2). 3. Find the solutions to the Volterra equation of the second kind <fJ(x) = f(x) +,\ r sin (x - s)<fJ(s) ds when (i) f(x) = x and A = 1, (ii) f(x) = e- x and A = 2, using the general solution (75) and verify that they agree with the solutions obtained by solving the equivalent differential equations (see problem 4 at the end of chapter 2). 4. If F(u) is the Fourier transform of f(x) show that the transform of f"(x) is -u 2 F(u) provided f, f'  0 as x  :t:oo. Hence show that f(x) = f e-lx-sl<fJ(s)ds (-00< x <(0) has the solution cf>(x) = !{f(x) - f"(x)} 5. If F(u) is the Laplace transform of f(x) show that the transform of f'(x) is uF(u)-f(O) for u>a provided f(x)e-axo as xoo. Hence find the solution of the Volterra equation of the first kind f(x) = r e-s<fJ(s) ds where f(O) = o. 6. Use Fourier's sine formulae to verify that the solution of Fox's integral equation <fJ(x) = f(x) + ,\   1"" sin xs<fJ(s) ds is given by (124). (O<x<oo) 
42 Integral equations of the convolution type 7. Use the method of Mellin transforms to solve Fox's integral equation when the kernel is k(X)=A  cosx. Verify the correctness of your solution by using Fourier's cosine formulae. 
4 Method of successive approximations 4.1 Neumann series A valuable method for solving integral equations of the second kind is based on an iterative procedure which yields a sequence of approximations leading to an infinite series solution associated with the names of Liouville and Neumann. It is sometimes called the Liouville-Neumann series but more often it is called the Neumann . serIes. Let us first examine the Fredholm equation <fJ(x) = f(x) + A r K(x, s)<fJ(s) ds (axb) (1) and consider the set of successive approximations to the solution cf> (x) given by cf>O(x) = cf> (O)(x) cf>1 (x) = cf> (O)(x) + Acf> (1\X) cf>2(X) = cf> (O\x) + Acf> (l)(x) + A 2 cf> (2)(X) (2) and so on, the Nth approximation being the sum N cf>N(X) = L A ncf>(n)(x) n=O (N = 0, 1, 2, . . .) (3) where cf> (O\x) = f(x), <fJ(l)(X) = r K(x, s)<fJ(O)(s) ds, <fJ(2)(X) = r K(x, S)<fJ(l)(S) ds, (4) and in general <fJ(n)(x) = r K(x, s)<fJ(n-l)(s) ds (n;?; 1). (5) 
44 Method of successive approximations We see that the sequence of approximations is generated by an iterative process of successive substitutions on the right-hand side of (1). Suppose that f(x) and K(x, s) are continuous functions in the range of definition so that they are bounded and we may write If(x)1  m IK(x, s)IM (axb) (a  x  b, a  s  b) (6) when m and M are positive constants. Then 1cf>(O)(x)1 = If(x)1  m, IcfJ(l)(x)l:s;;; flcfJ(O)(s)IIK(X, s)1 ds:S;;; mM(b - a), IcfJ(2)(x)l:s;;; flcfJ(l)(s)IIK(X, s)1 ds:S;;; mM2(b - a)2, and in general 1cf>(n)(x)1  mMn(b - a)n. (7) Hence N N 00 L A ncf>(n)(x)  L IAlnlcf>(n)(x)1  m L IAlnMn(b - a)n (8) n=O n=O n=O and so L=o A ncf>(n\x) is absolutely and uniformly convergent in axb if p=IAIM(b-a)<l, that is if 1 IAI< M(b-a) (9) since then the series is dominated by the convergent geometric serIes f n m mi.Jp = 1 . n=O - P (10) When condition (9) is satisfied, 00 cf>(x) = LAn cf> (n)(x) n=O (11) is a continuous solution of the integral equation (1). 
Neumann series 45 The error made in replacing the exact solution cf> by the Nth approximation cf>N is given by 00 rN(x) = cf>(x) - cf>N(X) = L A ncf>(n)(x). (12) n=N+l It is readily seen that 00 IrN(x)lm L pn_ n=N+l N+l mp 1-p (13) and so IrN(x)1 o as Noo if p<l. Now let us turn our attention to the Volterra equation cfJ(x) =f(x)+ A f K(x, s)cfJ(s) ds. (14) Then the previous analysis holds with K(x, s)=O for s>x. But we have further that 1 cf> (O)(x)1 = If(x)1  m, 14>(1)(x)I";;;; flcfJ(O)(s)IIK(x, s)1 ds";;;; mMf ds = mM(x - a), fX fX mM 2 (x - a)2 IcfJ (2)(x)I";;;; la IcfJ (l)(S )IIK(x, s)1 ds";;;; mM2la (s - a) ds = 2! ' fX mM 3 fX mM 3 (x - a)3 IcfJ(3)(x)l,,;;;; la IcfJ(2)(s)IIK(x, s)1 ds";;;; 2! la (s - a)2 ds = 3! and in general, assuming that IcfJ(n-1)(x)l,,;;;; mMn-l(x - at- 1 (n -I)! and using the principle of induction, we have IcfJ(n)(x)l,,;;;; (m, r x (s _ at --I ds = mMn( - at . (15) n . Ja n. Hence f A ncfJ(n)(x) ,,;;;; m f IAlnMn(b - at (16) n=O n=O n! and so L=o A ncf>(n)(x) is absolutely and uniformly convergent in a  x  b for all values of A since it is dominated by 00 IAlnMn(b _ a)n m L , =mexp{IAIM(b-a)}. (17) n=O n. 
46 Method of successive approximations Thus the Neumann series (11) converges for all values of A in the case of the Volterra equation (14) whereas it converges only for sufficiently small values of A in the case of the Fredholm equation (1). 4.2 Iterates and the resolvent kernel In the previous section we obtained solutions to the Fredholm and Volterra equations of the second kind in the form of the infinite series (11). It is convenient to express these solutions in terms of iterated kernels defined by K1(x, s) = K(x, s) Kn(x, s) = r K(x, t)Kn-1(t, s) dt (n  2) (18) (19) so that K 2 (x, s) = r K(x, t1)K(ti> s) dtl, K 3 (x, s) = r K(x, tt)K 2 (tt, s) dt 1 (20) = r r K(x, t1)K(ti> t 2 )K(t 2 , s) dt 1 dt 2 (21) and in general Kn(x, s) = r. · · r K(x, t1)K(tl, t 2 ) · · · K(t n -2, tn-1)K(t n - b s) dt 1 · · · dtn-l (22) It follows at once that the iterated kernels satisfy Kn(x, s) = r Kp(x, t)Kq(t, s) dt for any p and q with p+q = n. Now (23) </J(n)(x) = r Kn(x, s)[(s) ds (24) 
Iterates and the resolvent kernel 47 and hence we may write the solution (11) of the Fredholm 'equation (1) in the form 00 f b <fJ(X) = f(x)+ nl A n la Kn(x, s)f(s) ds (25) or equivalently <fJ(x) = f(x) + A r R(x, s; A)f(s) ds (26) where 00 R(x, s; A) = L A nK n + 1 (x, s) n=O (27) is the solving kernel or resolvent kernel already introduced in section 1.3.3 during the discussion of separable kernels. If f(x) and K(x, s) are continuous functions satisfying (6) and p = IAI M(b - a) < 1, the infinite series (27) for R(x, s;'\) is abso- lutely and uniformly convergent in a  x  b, a  s  b since N N L A nK n + 1 (x, s)  L lAin IKn+t(x, s)1 n=O n=O f lAin M n +1(b-at= M . n=O 1 - P Moreover for Volterra kernels satisfying K(x, s) = 0 for x < s we have, using (6) and the knowledge that the integrand on the right-hand side of (22) vanishes except for x  t 1  t 2 . · ·  tn-t  s: M n (x-s)n-1 Mn(b-a)n-1 IKn(x, s)1  (n -1)!  (n -1)! and so the infinite series (27) for R(x, s; A) is absolutely and uniformly convergent in a  x  b, a  s  b for all values of A since N N L A nKn+t(x, s)  L lAin IKn+t(x, s)1 n=O n=O 00 lAin Mn(b - a)n ML n=O n! = M exp {IAI M(b - a)}. It can also be shown that the resolvent kernel is continuous in the same region for IAI<{M(b-a)}-1 in the case of continuous 
48 Method of successive approximations Fredholm kernels, and for all A in the case of continuous Volterra kernels. Example 1. As a first example of the Neumann series we consider the Fredholm equation of the second kind cf>(x) = f(x) + ,\ r u(x) v(s) cf>(s) ds (28) with separable kernel K(x, s) = u(x)v(s). This equation was dis- cussed previously in section 1.3.3 where an exact solution (1.30) in closed analytical form was obtained. Now we have K1(x, s) = K(x, s) = u(x)v(s) K 2 (x, s) = r K1(x, t)K1(t, s) dt b = u(x) v(s) I v(t) u(t) dt = au(x)v(s), where a = r v(t) u(t) dt, and if we assume that Kn(x, s) = an-1u(x) v(s) then Kn+i(x, s) = r K1(x, t)Kn(t, s) dt = an-1u(x) V(S)f v(t)u (t) dt = anu(x)v(s) (29) establishing the general form of the interated kernel Kn+1(x, s) by the principle of induction. 
Iterates and the resolvent kernel 49 Hence the resolvent kernel is 00 R(x,S;A)= L A n K n + 1 (x,S) n=O 00 = U(X) V(S) L (Aa)n n=O (30) u(x)V(S) - 1-Aa (31) provided IAal < 1. This is in accordance with the exact solution (1.32) derived previously for all A ¥= a-i. The homogeneous equation corresponding to (28) is cf>(x) = A r u(x) v(s) cf>(s) ds It possesses one characteristic value Ai = a -1 and we see that the Neumann expansion (30) converges to (31) for IAI<IAll. Example 2. We now consider the Fredholm equation cf>(x) = 1 + A r xscf>(s) ds (O::S:; x::S:; 1) (32) with f(x) = 1 and separable kernel K(x, s) = xs. This was examined earlier as example 1 of section 1.3.3. We have a= {\2dt=i (33) so that xs K n + 1 (x, s) = 3 n . Hence the resolvent kernel is (34) 00 R(x,S;A)= L A n K n + 1 (x,s) . n=O 00 ( A ) n = xs L - n=O 3 xs (35) 1 - A/3 
50 Method of successive approximations provided IA 1< 3, this being less severe that the condition given by p < 1. Thus, using (26) we see that <fJ(x) = f(x)+ A r R(x, s; A)f(s) ds = 1 + 1 :/3 r s ds 3Ax -1+ 2(3-A) (36) which is in accordance with the exact solution (1.38) derived previ- ously for all A ¥= 3. Example 3. Lastly we discuss the Volterra equation with kernel K(x, s) = { e-s (x>s) (s > x) (37) This kernel is of the convolution type and was discussed before as example 2 of section 3.3 where an exact solution in closed analytical form was derived. Now K 2 (x, s) = f' K(x, t)K(t, s) dt = IX e x - t . e t - s dt = e X - S I Xd't = eX-S(x - s) (x> s), K 3 (x, s) = f'K(X, t)K 2 (t, s) dt = IX e x - t . et-S(t - s) dt = ex-sf'(t- s) dt x-s (x - S )2 ( X > S ) :; e 2! 
Problems 51 and in general, assuming that (x - S)n-l Kn(x, s) = e X -' (n -1)! (x>s) we have K n + 1 (x, s) = r K(x, t)Kn(t, s) dt I x ( ) n-l _ x-t. t-s t - s d - see (n -I)! t _ x_srX (t-s)n-l - e J. (n -1)! dt ( X -- S ) n x-s =e n! (x > s) (38) which establishes the general form of the iterated kernel by the principle of induction. Hence the resolvent kernel given' by (27) is R( . \ ) = f ,\n(x-s)n x-s x, S , 1\ i..J , e n=O n. = e(A+l)(x-s) (x> s) (39) for all '\, which is in agreement with the solution (70) obtained in section 3.3 since the resolvent kernel vanishes for x < s. Problems 1. Use the method of successive approximations to solve <f>(x) = 1 + A r eX+S <f>(s) ds and verify that your solution agrees with the solution to problem 3 at the end of chapter 1 for ,\ < 2/(e 2 -1). 2. Use the method of successive approximations to solve <f>(x) = x + A ["Sin nx sin ns<f>(s) ds where n i. an integer, and verify that your solution agrees with the solution to problem 5 at the end of chapter 1 for ,\ < 2/ 'IT. 
52 Method of successive approximations 3. Obtain the solution in closed analytical form of <f>(x) = f(x) + f1T K(x, s)<f>(s) ds where 00 K(x, s) = L an SIn nx cos ns n=l and 00 L lanl<oo, n=l using the method of successive approximations. 4. Use the method of successive approximations to solve <f>(x) = 1 + A r <f>(s) ds, verifying that your solution for A = 1 agrees with that obtained to problem l(i) at the end of chapter 2. 5. Use the method of successive approximations to obtain the resolvent kernel for <f>(x) = f(x) + A r (x - s)<f>(s) ds Verify that this agrees with the solutions to problem 3 at the end of chapter 2. 6. Use the method of successive approximations to show that the resolvent kernel for <f>(x) = 1 + A r xs<f>(s) ds is given by 00 ( A ) n (X3 _ s3)n R(X,S;A)=XSno 3 n! and hence solve the integral equation. 
5 Integral equations with singular kernels 5.1 Generalization to higher dimensions So far in this book we have been concerned solely with integral equations which involve an unknown function cf>(x) of a single real variable x. However it is interesting to consider integral equations in higher dimensions and to this end we suppose that V is a region of an n-dimensional space and let M, N de-note points in the region V. Then an integral equation of the second kind takes the form cP(M) = f(M) + L K(M, N)cP(N) dUN (1) where K(M, N) is the kernel, f(M) is a given function, cf>(M) is the function which we require to find, and the integration in (1) is over all the points N of the n-dimensional region V. Also an integral equation of the first kind takes the form f(M) = L K(M, N)cP(N) dUN. (2) A common type of singular kernel is given by K(M, N) = F( N) (3) r where F(M, N) is a bounded function and r is the distance between the points M and N in the n-dimensional space. This type of kernel is called polar and as r  0 it approaches an infinite value. It is said to give rise to an integral equation with a weak singularity if 0< a < n. The polar kernel (3) is square integrable, Le. L L IK(M, NW dUM dUN < co, (4) when a is restricted further to the range 0 < a < n12, as we have shown for the one-dimensional case (n = 1) in section 1.4. 
54 Integral equations with singular kernels 5.2 Green's functions in two and three dimen- . stons We now consider the linear second order partial differential equa- tion LtfJ = n (5) and begin by investigating the two-dimensional case for which iJ2 iJ2 L =++q(x, y) (6) iJx iJy Then the corresponding Green's function G(x, y; s, t) satisfies LG(x, y; s, t) = S(x - s )S(y - t). (7) Integrating over the interior of a circle l' of radius e centered at the point with coordinates s, t we have L LG(x, y; s, t) dx dy = 1 (8) since the integral of the S functions amounts to unity. This leads to lim f iJG du= 1 BO 'Y iJp (9) where du 'represents an .element of arc length and p2 = (x - S)2 + (y - t)2. (10) We now see that iJG 27Tp- 1 iJp (11) as p  0 and so 1 G (x, y; s, t) = 27T In p + g (x, y; s, t) (12) where 1 L g( x, y; s, t) = - 27T q (x, y) In p (13) and g is chosen so that G satisfies prescribed boundary conditions. 
Dirichlet's problem 55 In the three-dimensional case L = V 2 +q(r) and the Green's function satisfies (14) LG(r, s) = 5(r-s) (15) where rand s are the position vectors of the points with coordinates (x, y, z) and (s, t, u) respectively. Integrating over the region contained by a sphere S of radius e centered at the point with position vector s we obtain Is LG(r, s) dr= 1 (16) gIvIng lim! aG dS = 1 8-+0 k aR (17) where the integration in (17) is over the surface of the sphere Sand R 2 = Ir-sl 2 = (x - S)2+(y - t)2+ (z - U)2. (18) It follows that 41TR 2 aG  1 aR (19) as R  0 and hence the Green's function has the form 1 G(r, s) = - 41TR + g(r, s) (20) where q(r) Lg(r, s) = 41TR (21) and we choose g so that G satisfies given boundary conditions. 5.3 Dirichlet's problem The problem named after Dirichlet is concerned with determining a function tf1 which has prescribed values f over the boundary r of a certain simply connected region R and which is harmonic, Le., satisfies Laplace's equation, at all interior points of R. 
56 Integral equations with singular kernels We begin by considering the two-dimensional Dirichlet problem in which we have a plane region R bounded by a closed contour l' with a continuously turning tangent. Let us suppose that tf/(x, y) is the harmonic function satisfying a 2 tf/ a 2 tf/ Vitf/ = -+-= 0 ax 2 ay2 (22) in R and having the prescribed values given by f(s, t) at all points (s, t) of 1'. We now express tf/ as the real part of an analytic function F(z) of the complex variable z = x + iy by putting tf/(x, y) = Re F(z) (23) and look for a solution in the form of the Cauchy-type integral F(Z)= f p,(C) dC 21Tl 'Y l- z (24) where the density IL«() is a real function of the complex variable , and the contour l' is directed in the anticlockwise sense. Next we allow z to approach a point w on the boundary curve l' from the interior of R. In the limit we obtain F(w) =!IL(W)+ 2 1 · P f ;(C) dC 1Tl - W 'Y (25) where the integral occurring on the right-hand side of (25) is its Cauchy principal value. (indicated by the prefix P): 'Urn f. p,(C) dC BO 'Ye l- w (26) where I'B is the part of l' outside a circle of radius e centered at the point w, and the term !IL(W) is the contribution arising from the singularity at ,= w. Now taking the real parts of (25) we find that f(x, y) =tp,(x, y)+ 2 plp,(C) Im( CCw ) (27) where (x, y) are the coordinates of the point w on 1'. Let us write ,- w = re i8 . Then the imaginary part of d'/(' - w) 
Dirichlet's problem 57 becomes Im ( d' ) =Im[d{ln(,-W)}] '-w = Im[d(In r+ i8)] =d8 = a8 du au where du is an element of arc of 1'. Using one of the Cauchy- R . . at/J ae/> at/J ae/> · h I d .. Iemann equatIons - = -, - = -- connectIng t e rea an Imagl- ax ay ay ax nary parts t/J, e/> of an analytic function we get (28) a8 alar -=- (In r) =--' au an r an where n is measured in the direction of the outward nQrmal to l' (see Fig. 4). Now (29) ar A A -=r.R an (30) here " denotes a unit vector, so that we obtain 1 f " " 1 f.n f(X,y)=21L(x,y)+- 2 P lL(s,t)-du. 7T 'Y r (31) Since we can characterize the point , by its arc distance 00 along l' from the point w, we may rewrite (31) in the one-dimensional form 1 f. "" f.n IL(U) = 2f(u) -- P IL(U) - du 7T 'Y r which is a Fredholm integral equation of the second kind with a polar kernel (3) for a = 1. I We next turn to the three-dimensional interior Dirichlet problem. Let S be the boundary surface of the three-dimensional region R and suppose that the solution t/J of Laplace's equation V 2 t/J = 0 possesses the prescribed values at the boundary surface given by the function f. If M is an interior point of R we may express the solution as the potential of a double layer, which in electricity is a distribution of (32) 
58 Integral equations with singular kernels w Fig. 4. Dirichlet's problem. electric dipoles spread over the surface S with their axes in the directions of the normals. Thus we have -f/(M)=- 4 1 f p,(N) r .2° dS N 1T Js r (33) where n is a unit vector along the outward normal at the poinN of the surface S, IL(N) is the density of the double layer and r = MN. Then allowing M to tend to a point of S from the interior of R yields the two-dimensional Fredholm integral equation of the sec- ond kind for IL: IL(M) = 2f(M) -- 2 1 pf p,(N) r .2° dS N (34) 1T Js r with a polar kernel (3) for a = 2. 
Dirichlet's problem 59 5.3.1 Poisson's formula for the unit disc We now search for the solution .p(r, 8) of Dirichlet's problem for a unit circle 1', the polar coordinates (r, 8) being referred to the centre of the circle. To this end we consider Laplace's equation in a.plane vi.p = 0 (35) where .p(r, 8)  f( 8) as r  1- o. If G(r, s) is the Green's function satisfying ViG(r, s) = 5(r-s) (36) with G(r, s) = 0 when s is on the boundary given by s = 1, it follows from (12) that 1 G(r, s) = 2'1T {In Ir-sl-in (r Ir' -sl)} (37) where r' is the inverse point of r with respect to the unit circle l' so that rr' = 1 and thus, using similar triangles (see Fig. 5), Ir - sl = r Ir' - sl if s is on the boundary. p' o 'Y Fig. 5. Poisson's formula: P' is the inverse point of P with respect to the unit circle l' centered at 0 so that OP.OP' = OQ2 = 1, and OQP and OP'Q are similar triangles. 
60 Integral equations with singular kernels Green's theorem states that i 2 2 f ( a.p a cf» (cf>V1.p- .pV1cf» dS = cf>--.p- dO" R 'Y an an where n is measured in the direction of the outward normal to the boundary curve l' and so, setting cf> = G(r, s), we obtain (38) .p(r) = f l/1(s)  G(r, s) duo 'Y an (39) Now Ir -S12 = r 2 + S2 - 2rs cos (8 - x) and ,2 Ir' - Sl2 = r 2 {,,2 + s 2 - 2,' s cos (8 - X)} where (s, X) are the polar coordinates of the point s on the unit circle 1'. But r' = r- 1 and so r 2 lr' -S12 = 1 + ,2S2_ 2rs cos (8 - X). Hence 1 f a .p(r) =- 2 .p(s) - [! In {r 2 + s2-2rs cos (8- X)} 7T 'Y as -! In {I +r 2 s 2 -2rs cos (8 - X)}] dO" and since s = 1 and .p(s) = f(x) over the unit circle l' we find that 1 _,2 r 21T l/1(r)= 2'7T Jo f(x)[1+r 2 -2rcos(O-X)]-ldX (40) which is Poisson's formula for the unit disc. 5.3.2 Poisson's formula for the half plane We consider the half plane y > 0 bounded by the straight line y = 0 and suppose that .p(x, y)  f(x) as y  +0. Then the Green's func- tion (12) which vanishes when the point s with coordinates (s, t) lies on the boundary line t = 0, takes the form 1 G(r, s) = 2'7T {In Ir-sl-In Ir' -sl} (41) where the coordinates of rand r' are (x, y) and (x, -y) respectively. 
Dirichlet's problem 61 Now using (39) we obtain 1 f oo a tfJ(x, Y)=-- 2 f(s)-[!ln{(s-x)2+(t-y)2} 1T -00 at -!In {(s - X)2 + (t + y)2}]t=ods = ; Lf(S)[(S-X)2+y2rldS (42) which is Poisson's formula for the half plane. 5.3.3 Hilbert kernel An integral equation with the singular kernel ( x - 8 ) cot 2 named after Hilbert can be obtained by using Poisson's formula (40) for a unit disc and setting z = re iB and' = eix. Then d'/' = i dX and (43) , + z _ 1 + r cos ( 8 - X) + ir sin (8 - X) , - z 1 - r cos (8 - X) - ir sin (8 - X) gIvIng ( ' + Z ) 1 - r 2 Re = 2 · , - z 1 + r - 2 r cos (8 - X) Hence 1 f ' + z d' I/1(r, 0) = Re 27Ti ./(X) , - z T where l' is the circle 1'1 = 1. Let cf>(r,8) be the harmonic function which is conjugate to tfJ(r, 8). Then cf>(r, 8) is determined apart from an additive arbitrary constant which we choose so as to make cf>(r, 8) vanish at r = O. We have then (44) . 1 I ' + z d' I/1(r, 0) + 1<I>(r, 0) = 27Ti ./(X) , - z T' Now .letting r  1- 0 so that z approaches a point of l' from the interior of the unit disc, we obtain (45) tfJ(l, 8) + icf>(l, 8) = f( 8) + 2 1 . P I f(x) ,+ z d' . 1Tl y , - z , 
62 Integral equations with singular kernels We already know that tf/(1, 8) = f(6). Putting cf>(1, 8) = g(8) we find that 1 1 211 eix + e iO g(8)=- 2 . P f(x) ix io dx 1Tl 0 e - e i ,r 211 e i[(x- O )/2] + e -i[(x- O )/2] = - 2'7T PJo f(x} e i [(x- IJ )/2]_ e- i [(x- IJ )/2] dX and so 1 J 211 ( 8 ) g( O} = - 2'7T PJo f(x} cot X  dX (46) which is an integral equation of the first kind for f(x) having the Hilbert singular kernel (43). Now we turn to the Poisson formula for the half plane and set z = x + iy and ,= s + it. On the t = 0 axis we have 1m ( 1 ) = 1m ( 1 ) , - z s - x - iy Y - (S-X)2+ y 2 and so 1 f tf/(,) tf/(x, y) = Re --: y d' 1Tl 'Y  - z (47) where l' is the straight line t = 0 together with the infinite semi- circle in the upper half plane traced in the anticlockwise sense. We let cf>(x, y) be the harmonic function conjugate to tf/(x, y). Then . 1 f t/!(,) I/1(x, y} + 1cf>(X, y} = -: , d' 1Tl - Z 'Y (48) and so, letting y  + 0, we obtain . 1 1 00 f(s) tf/(x, +0) + lcf>(X, +0) = f(x) +--: P - ds 1Tl -00 S - x . . gIvIng g(x} = -1. pf'" f(s} ds 1T J-oo s - x (49) 
Dirichlet's problem 63 where [(x) = t/1(x, +0), g(x) = cf>(x, +0) and the integral on the right- hand side is the principal value defined by lim [f x-e [(s) ds + 1 00 [(s) dS ] . (50) e-+O -00 S - X x+e S - X Equation (49) is an integral equation of the first kind having the singular kernel (s - X)-I. 5.3.4 Hilbert transforms Suppose that t/1(x, y), t/11 (x, y), t/12(X, y) are functions which are harmonic in a plane region R bounded by a closed curve 1', and t/11 is conjugate to t/1 while t/12 is conjugate to t/11. Then using the Cauchy- Riemann equations for the analytic functions t/1 + it/11 and t/11 + it/12 respectively we obtain at/1 _ at/11 - , ax ay and at/11 = at/12 ax ay' so that a ax (t/1 + t/12) = 0, Hence at/1 _ at/11 -- -...-- ay ax (51) at/11 _ at/12 ---- ay ax (52) a -( t/1 + t/12) = 0 ay (53) t/1 = -t/12 + C (54) where C is a constant. Let us now apply this result to the case of the unit disc for which l' is the circle r = 1. We take the values of t/1, t/1b t/12 over l' to be [(8), [1(8), [2(8) respectively and choose t/12 to vanish at r = 0 so that C = t/1(r = 0). But Poisson's formula (40) for the unit disc informs us that 1 r 21T "'(r = 0) = 27T Jo f(x) dX and so (55) 1 i 21T t/1 = -t/12 +- [(X) dX 27T 0 (56) 
64 Integral equations with singular kernels giving in particular 1 i 2'71" f(8)=-fz(8)+ 27T 0 f(X)dX (57) Hence, if f( 8) and g( 8) are the real and imaginary parts of the analytic function <1>( z) = t/J + icf> over the boundary circle r = 1, we see that 1 1[2'71" ( 8 ) g(8) = - 27T PJo f(x) cot x dX (58) and 1 J 2'71" ( X 8 ) 1 i 21't f(8)= 27T P Jo g(x) cot 2 dX+ 27T 0 f(X)dX These are the reciprocal formulae deduced by Hilbert in 1904. Equation (59) provides a solution of the integral equation (58) of the first kind with the singular kernel (43). Combining (58) and (59) together gives the Hilbert formula 1 i 2'71" f(8)- 27T 0 f(X)dX 1 J2'71" ( X - 8 ) J2'71" ( , - ) = - (27T)2 PJo cot 2 dX · PJo f(x') cot X 2 X dX' (60) (59) For the case of the half plane, we have likewise 1 i oo f(s) g(x)=--P -ds 1T -00 s - x f(x) =1. P i 00 g(s) ds 1T -00 S - x (61) (62) where f and g are the real and imaginary parts of the values <I>(x + iO) of an analytic function <I>(z) = t/J + icf> over the boundary line y = O. The function g(x) is the Hilbert transform of f(s). Denoting the integral operator -1. P I '"  (63) 1T -00 s - x by H we obtain H{H[f]} = -f (64) 
Singular integral equation of Hilbert type 65 Whereas the reciprocal forulae (61) and (62) involve integra- tions over the infinite range -00 < s < 00, the range of integration in the pair of formulae (58) and (59) is 0  X  27T and are conse- quently referred to as finite Hilbert transforms. 5.4 Singular integral equation of Hilbert type Let us now consider the singular integral equation aq,(x) + 2 pf1T q,(s) cot (s 2 x) ds = f(x) having a singular kernel of the Hilbert type (43). Operating on both sides of (65) with af1T s (s-X) dx+ 2 pf 1T cot e 2 x ) dx (65) and setting b J21T ( s X ) F(x) = af(x)- 27T PJo f(s) cot 2 ds, (66) we obtain the simple result b 2 i 21T (a 2 + b 2 )q,(x) - 27T 0 q,(s) ds = F(x) on using the Hilbert formula (60). To solve this equation we note that (a 2 + b 2 ) f1T q,(x) dx - b 2 f1T q,(s) ds = f1T F(x) dx and hence (67) i 21T 1 i 21T 1 i 21T <f>(x)dx=2 F(x)dx=- f(x)dx o a 0 a 0 (68) sInce f 1T cot( s 2 x )dX=O. Thus the solution of (65) is F(x) b 2 r 21T q,(x) = a 2 +b 2 + 27Ta(a 2 +b 2 ) Jo f(s)ds (69) 
66 Integral equations with singular kernels If a = 0 (and we put b = 1) we obtain an integral equation of the first kind having the form 1 f 2'71" ( s X ) f(x) = 27T P 0 q,(s) cot 2 ds. (70) Using (59) we arrive at the solution 1 1 2'71" 1 f 2'71" ( s X ) q,(X) = 27T 0 q,(S) ds - 27T P 0 f(s) cot 2 ds (71) If we now put 1 1 2'71" 27T 0 q,(S) ds = C and substitute 1 f 2'71" ( s X ) q,(x) = C- 27T P 0 f(s) cot 2 ds (72) back into (70), we see that (72) provides a solution for any value of the constant C if and only if 1 2'71" o f(x)dx=O. (73) Finally we consider the singular integral equation of the second kind A 1 00 cf>(s) cf>(x)--P -ds=f(x) 1T -00 s - X (74) Operating on both sides of this equation with fOO A foo d Lx> 8(x - s) dx + 7T PLoo x X s and setting A foo f(s) F(x)=f(x)+ 1T P J-oo s_ xds (75) we find that (1 + A 2)cf>(X) = F(x) 
Problems 67 on using (64). Hence the solution of (74) is cf>(x) = 1 2 {f(x)+ pr'" f(s) dS } (76) 1 + A 1T J-oo S - x Problems 1. By taking f(21T-X)=f(x) and g(21T-X)=-g(X) in the pair of finite Hilbert transforms (58) and (59), show that they may be rewritten in the form . 1 i 7T sin 0 (1) g( 0) = -- P f( cf> ) dcf> 1T 0 cos 0 - cos cf> 1 i 7T sin cf> 1 i 7T (ii) f( 0) = - P g( cf» 0 dcf> + - f( cf> ) dcf>. 1T 0 COS - cos cf> 1T 0 [ . sin 0 ] HInt: use !{cot !( 0 + cf» + cot !( 0 - cf»} = cos cf> - cos 0 2. By putting x = cos 0, y = cos cf> and -u(x) = f(O)/sin 0, v(x) = g( O)/sin 0 show that the pair of reciprocal formulae (i) and (ii) in problem 1 may be rewritten as (i) v(x)=_!pr l u(y) dy 1T J-1 X - Y (ii) U(X)=!p f l  v(y) dy+! 1 f l u(y)dy 1T -1 l-x 2 x-y 1T .J 1-x 2 -1 3. Find the solution of pr l u(y) dy = 0 J-1 X - Y 4. Find the solution of Foppl's integral equation ! pr I :(t) 2 dt = f(s) 1T J-1 t - S where cf>(t) and f(s) are even functions. 
68 Integral equations with singular kernels 5. Find the solution of 1 f a -p 1T -a tg(t) d _ 2 2 2 t-s. s -t 1 f 00 is 6. Evaluate - P  ds using Cauchy's residue theorem and 'iT -00 S - x hence find the solutions of ( . ) . 1 pf oo f(s) d 1 slnx=-- - s 1T -00 S - x ( .. ) 1 pf oo g(s) d II COS X = - - S 1T -00 S - x Verify that (i) and (ii) form a pair of reciprocal Hilbert trans- forms. 7. Find the solution of 1 1+x2 p  f oo f(s) ds. 1T -00 S - x 
6 Hilbert space In the remaInIng chapters of this book we shall be gIvIng an introduction to the general theory of linear integral equations as developed by Volterra, Fredholm, Hilbert, and Schmidt. For this purpose we need to introduce the concept of a Hilbert space. This is a suitable generalization of ordinary three-dimensional Euclidean space to a linear vector space of infinite dimensions which, for the subject of integral equations, is chosen to be a complete linear space composed of square integrable functions having a distance property defined in terms of an inner product. To explain the meanings of these terms we consider Euclidean space first and then the Hilbert space of sequences. 6.1 Euclidean space In three-dimensional Euclidean space each point is specified by an ordered set of three real numbers or coordinates (Xh X2, X3) forming the components of the position vector x. The vector Ax has coordi- nates (AXh AX2, AX3) and the vector sum x+y of two vectors x, y has coordinates (Xl + Yh X2 +,Y2, X3 + Y3). These are properties of a linear vector space. The scalar product or inner product of two vectors x, y is defined as (x, y) = XIYI + X2Y2 + X3Y3 (1) and the vectors are orthogonal, Le. at right angles, if (x, y) = o. We have (x, x) = xi + x + x  0, where (x, x) = 0 if and only if x is the zero vector 0 with coordinates (0, 0, 0). The magnitude or norm Ilxll of a vector x is given by Ilxll =  (x, x) =  xi+ x+ x <00. (2) 
70 Hilbert space The vector is said to be normalized if Ilx = 1 and then x is a unit vector. The distance between two points specified by the vectors x, y is given by Ilx - yll. Evidently Ilxll is the distance of the point x from the origin specified by the zero vector o. Since the length of a side of a triangle is less than, or equal to when the triangle collapses into a straight line, the sum of the lengths of the other two sides, we have the triangle inequality Ilx - yll  Ilxll + Ilyli. (3) Suppose that x, y, z are linearly independent vectors so that AX+#LY+VZ is not the zero vector 0 except when A=#L=V=O. Then we can use them to construct an orthogonal and normalized, Le. orthonormal set of three vectors e1, e2, e3. Thus let us put e1 = x/llxli. This vector is clearly normalized. Then y' = y - (y, e1)e1 is orthogonal to e1 and e2 = y'/lly'li is normalized. Further z' = z - (z, e1)e1 - (z, e2)e2 is orthogonal to e1' and e2 while e3 = z'/llz'll is also normalized. The three mutually orthogonal unit vectors eb e2, e3 are said to form a basis since any vector a of the three-dimensional space can be expressed as the linear combination a = (a, e1)e1 + (a, e2)e2 + (a, e3)e3. The foregoing vector algebra can be extended to an n- dimensional space whose points are specified by an ordered set of n complex numbers (XI, X2, . . . , x n ) denoted by the vector x. The inner product of two vectors x, y is now defined as n (x, y) = L xrYr = (y, x) r=l (4) while the norm of the vector x is given by .. IIxll = v'(x, x) =  t IXrl2 < 00, r=l (5) An orthonormal set of n vectors e1, e2, . . . , en form a basis of the n-dimensional space, and an arbitrary vector a in the space can be 
Hilbert space of sequences 71 expressed as the linear combination n a = L (a, er)e r r=l where the a r = (a, er)(r = 1, . . . , n) are the components of the vector a with respect to the basis vectors. The only vector which is orthogonal to every vector of the basis is clearly the zero vector 0 and so the orthonormal set eh e2,..., en is said to span the n-dimensional space. If we take el = (1, 0, 0, . . . , 0), e2 = (0, i, 0, . . . , 0), . . . , en = (0, 0, 0, . . . , 1) we see that a is the vector specified by (ah a2, . . . , an). 6.2 Hilbert space of sequences By a natural generalization of a finite dimensional space, we can consider an infinite dimensional space whose points are represented by vectors x having components, or coordinates, given by the infinite sequence of complex numbers {xr} = (Xl, X2, . . . , X r , . . .) satisfying 00 L Ix r I 2 <00. r=l (6) We now introduce the scalar product or inner product of two vectors x and y given by 00 (x, y) = L xrYr= (Y, x) r=l (7) Then we have 0  (x, x) < 00, where (x, x) = 0 if and only if x is the zero vector 0 whose components all vanish. Also we define the norm Ilxll of a vector x by the formula Ilxll = v' (x, x) =  rtl I X rI 2 . Thus Ilxll = 0 if and only if x = o. Further we let Ax be the vector with components {AX r } so that IIAxl1 = IAlllxrl, and let the sum x+y of two vectors x, y be the vector having components {x r + Yr} as in the case of a finite dimensional space. Now, by the Cauchy inequality (8) rlIXrIIYrl  C1IXrI2)C1IYrI2) (9) 
72 Hilbert space and the inequality 00 00 L XrYr  L IXrIIYrl, r=l r=l (10) we obtain Schwarz's inequality I(x, y)1  Ilxllllyll. (11) This is the generalization to infinite sequences of the corresponding result in three-dimensional Euclidean space which follows im- mediately from (x, y) = Ilxllllyll cos a where a is the angle between the vectors x and y. Hence 00 Ilx+yI12= L I X r+YrI 2 r=l 00 00 00 = L IXrI 2 + L IYrI 2 + L (xrYr+xrYr) r=l r=l r=l  (11xll + Ilyll)2 < 00. This shows that the sum x + y satisfies the condition 00 L IX r +YrI 2 <00 r=l (12) and also yields the triangle inequality Ilx + yll  Ilxll + Ilyll (13) which can be rewritten in the form (3) by reversing the sign of y. The real number d(x,y)=llx-yll=  rtlIXr-YrI2 (14) represents the distance between two points characterized by vectors x, y and is an obvious generalizatio of the idea of distance in three-dimensional Euclidean space. Clearly Ilxll is the distance of the point x from the origin given by the zero vector O. A sequence of vectors {Xn} converges strongly to a limit vector x if, given any B > 0, there exists N such that for n > N we have Ilx n -xii < B. Strong convergence is denoted by X n  x. If X n  x we have, using the triangle inequality, Ilxn -xmll = Ilx n -x+x-xmll llxn -xii + Ilxm -xll< B 
Hilbert space of sequences 73 for sufficiently large nand m. A sequence {xn} satisfying Ilxn - xmll < B for sufficiently large n, m is known as a Cauchy sequence. We shall now demonstrate the converse of the above result, namely that every Cauchy sequence has a limit vector x in the space. Suppose that el = (1, 0, 0, . . .), e2 = (0, 1, 0, . . .), . .. are unit vectors in the infinite dimensional space. They form a basis which spans the space, and an arbitrary vector a = ah a2, . . . , a r , . . .) can be expressed as 00 a = L are r r=l where a r = (a, e r ) and e r is the rth unit vector satisfying \\e r \\ = 1. Then we have, using Schwarz's inequality, I(x n , e r ) - (x m , e r )\ = I(x n -X m , er)1  \\Xn -xm\1 < B for all sufficiently large nand m. It follows that the sequence of numbers (x n , e r ) = xn) is a Cauchy sequence and approaches a limiting value xr(r = 1, 2, . ..) as n' 00. But for sufficiently large n, m we have 00 Ilxn-xm\\= L \xn)_xm)\2<B r=l and so for every k k L Ixn) - xm)\2 < B. r=l Hence, in the limit as m  00 we obtain k L \xn) - xrl 2 < B r=l and since this is true for every k we get 00 \\xn-x\\= L \xn)-XrI2<B. r=l But  rl I X rl 2 = IIxll = II(I -In) +Inll  Ilx-xnl\ + \\Xn\! < B + IIXnl1 
74 Hilbert space and so 00 L Ix r I 2 <00. r=l Thus I = (Xl, X2, . . . , X r , . ..) belongs to the space and In  I as n  00. A space in which In  I when {Xn} is a Cauchy sequence is called complete. The space of sequences described above is an example of a Hilbert space. We denote this space by 1 2 . 6.3 Function space We consider the function space composed of all sectionally continu- ous complex functions f(x) of a real variable x, defined in the interval a  x  b, which are square integrable and thus satisfy the condition flf(XW dx <00. Introducing the inner product of two such functions f(x) and g(x) given by (15) (f, g) = r f(x) g(x) dx, (16) we define the norm of the function f(x) as 11111 =  (f, f) . (17) Next we establish the important inequality named after Schwarz. We have f b (f) 2 f(x)- ( ,g) g(x) dx;;o:Q a g, g (18) so that (f, f) - 2 I(f, gW + I(f, g)1 2 ;;0: Q (g, g) (g, g) . I.e. ({, f)(g, g)  I(f, g)12 (19) 
Function space 75 Hence 11111 Ilgll  I({,. g)1 (20) which is Schwarz's inequality for square integrable functions. Also (11111 + IIgll)2 = 11111 2 + IIgl12 + 21111111g11  ({, f) + (g, g) + 21(f, g)1 by Schwarz's inequality. But 21(f, g)1  (f, g) + (f, g) and so (11fll + IIgI1)2  (f, f) + (g, g) + (f, g) + (g, f) = (f + g, f + g). Hence we have 11111 + II gll  Ilf + gll (21) which is the triangle inequality for functions, sometimes known as Minkowski's inequality. 6.3.1 Orthonormal system of functions Two functions f(x) and g(x) belonging to the function space are said to be orthogonal if . (f, g) = r f(x) g(x) dx = 0 (22) and the function f(x) is normalized if 11111 = 1. (23) We consider a set of sectionally continuous complex functions cP1 (x), cP2(X),. . . , cPr(x), . . . satisfying the orthonormality condition (<Pn <Ps) ':'" r <Pr(x) <ps(x) dx = 8rs (24) where 8rs is the Kronecker delta symbol 8rs = { (r = s) (r;es). (25) Such a set of functions is called orthonormal. 
76 Hilbert space An orthonormal system of functions is said to form a basis or a complete system * if and only if the sole function which is orthogonal to every member cPr(x) of the system is the null function which vanishes throughout the interval a  x  b except at a finite number of points. It can be shown that every orthonormal system is either finite or denumerably infinite, Le., the members of the system can be placed in 1-1 correspondence with the natural numbers, and that an incomplete system can always be completed to form a basis by adding a finite or denumerable set of functions. 6.3.2 Gram-Schmidt orthogonalization Now suppose that we have a finite or denumerable system of functions Xl (x), X2(X), . . . , Xr(X), . . . which is not orthonormal. We assume that the functions are linearly independent and that none of them is null. Then if AIXI + A2X2 +. . . + AnXn is the null function we must have Al = A2 = . . . = An = 0 and this holds for all n. We aim to construct out of this system of functions, a new system which is orthonormal, just as we were able to do in section 6.1 for three- dimensional Euclidean space. We proceed by using the principle of induction. Clearlv XI(X) cf>l (x) = Ilxlll (26) is normalized. Further tP2(X) cf>2(X) = 111/1211 ' (27) where tP2(X) = X2(X) - (X2, cPI)cPI (x), (28) is normalized and also orthogonal to cP I (X). Now let us suppose that we have constructed n such functions cPI(X), cP2(X),..., cPn(x) which are normalized and mutually or- thogonal. Then ( ) tP+I(X) cf>n+l X = IIl/1n+111 ' (29) * A complete system of functions should not be confused with a complete space, discussed in section 6.3.4. 
Function space 77 where n tfln+1(X) = Xn+1(X) - L (Xn+h cPr)cPr(X), r=1 (30) is normalized and orthogonal to all the functions cP1 (x), . . . , cPn (x). Thus, using the principle of induction, we have shown that it is possible to construct an orthonormal set of functions from the original set. The above process is called Gram-Schmidt orthogonalization. 6.3.3 Mean square convergence Let f(x) be any function belonging to the function space and suppose that we wish to obtain a best approximation to f(x) in the form of the sum n L crcPr(X) r=1 (31) where the C r are parameters to be detemined. We can achieve this by requiring that fb n 2 In = 1 f(x)- rl cr<p,(x) dx (32) be chosen as small as possible. This will then provide us with the best mean square approximation to f(x). We have that n n In = (f, f) - L {cr(f, cPr) + cr(f, cPr)} + L Ic r l 2 r=1 r=1 n n =(f,f)+ L Ic r -a r I 2 - L 1a,12 r= 1 r= 1 (33) where the a r = ([, cPr) are known as the Fourier coefficients of f(x). Evidently In attains its least value when C r = ar(r = 1, . . . , n) and then we have n In = (f,f)- L la r l 2 . r=1 (34) Since In  0 it follows at once that n (f, f)  L la r l 2 r=1 (35) for all n. This is called Bessel's inequality. 
78 Hilbert space If now lim In = 0 (36) n-+ oo we say that f(x) is the mean square limit of the sequence of functions {fn(x)} given by n fn (x) = L arcPr(x), r=l (37) which we may write Ilf- fnll  0 as n  00, or that {fn(x)} is strongly convergent to f(x), written fn(x)  f(x) as n  00. Then it follows that 00 IIfl1 2 = L I(f, cPr )1 2 r=l (38) which is known as Parseval's formula or the completeness relation. If this formula holds, the orthonormal system of functions cPr(x) forms a basis or a complete system. For suppose that, on the contrary, there exists a function f(x) with non-vanishing norm in the function space satisfying (f, cPr) = 0 for all values of r. Then it follows from Parseval's formula that Ilfll = 0 which provides a contradiction. Now consider two functions f(x) and g(x) belonging to the function space. Then 00 Ilf + Agl1 2 = L I(f + Ag, cPr)12 r=l supposing that the system of functions cPr forms a basis, and so 11111 2 + IA 1211g112 + X (f, g) + A (g, f) 00 00 = L I (f, cPr) 1 2 + I A 1 2 L I (g, cPr) 1 2 r=l r=l 00 + L {X (f, cPr)( cPr' g) + A (g, cPr)( cPr' f)}. r=l Since A is an arbitrary parameter we see that 00 (f, g) = L (f, cPr)( cPr' g) r=l (39) which is known as the generalized Parseval formula. 
Function space 79 6.3.4 Riesz-Fischer theorem A sequence of functions {fn(x)} is called a Cauchy sequence if lim Ilfn - fmll = O. (40) n,m-+ oo If a sequence {fn(x)} is mean sqare convergent to f(x) it is a Cauchy sequence, for using the triangle inequality \\te see that Ilfm - fnll = II(fm - f) + (f - fn)11  Ilfm - fll + Ilf - fnll  0 as m, n  00. A function space is said to be complete if every Cauchy sequence {fn(x)} is mean square convergent to a function f(x) belonging to the space, i.e., fn(x) is strongly convergent to f(x) or fn(x)  f(x). Now, so far, we have assumed that our function space is com- posed of sectionally continuous functions and such a space is not complete. However a function space composed of functions that are square integrable in the Lebesgue sense, i.e., L 2 functions, is complete. This result is Fischer's form of the Riesz-Fischer theorem which we assert without proof. It leads directly to Riesz's form of the Riesz-Fischer theorem: If cP1 (x), cP2(X), . . . , cPr(x), . .. is an orthonormal system of L 2 functions and {a r } is a sequence of complex numbers, then n L arcPr(x) r=1 is mean square convergent to a L 2 function f(x) whose Fourier coefficients are {a r } if and only if 00 L la r l 2 < 00. r=1 The space of L 2 functions f(x) defined over the interval a  x  b is another example of a Hilbert space, usually denoted by L 2 . In this space f = g if the functions f and g are equal 'almost everywhere', that is everywhere except at a finite number or a de- numerable infinity of points of the interval a  x  b. Further f = 0 in this space if f vanishes 'almost everywhere'. 
80 Hilbert space 6.4 Abstract Hilbert space H We have described above two examples of Hilbert spaces, the space of sequences 1 2 and the space of L 2 functions. We shall conclude this chapter by abstracting the common axioms that all Hilbert spaces must satisfy. A Hilbert space H is a complete linear vector space possessing a distance function or metric which is given by an inner product. A linear vector space is a set of elements, sometimes called points or vectors, f, g, h,... forming an Abelian group and permitting multiplication by the field of complex numbers A. An Abelian group has an internal law of" composition denoted by the addition sign + satisfying the commutative law {+g=g+f (41) and the associative law f+(g+h)=(f+g)+h, (42) , having a zero element 0 such that O+f= {+O= f, (43) and an inverse element - f corresponding to each element f of the set such that f+(-f) = (-f) + f= o. (44) The multiplication by the field of complex numbers satisfies 1 · f= f, o . f = 0, (AIL)f = A (lLf), (45) (46) (47) and satisfies the distributive law with respect to the elements f, g A(f + g) = Af + Ag and the distributive law with respect to the numbers A, IL (48) (A + lL)f = Af + ILf. (49) The inner product of two elements f, g is a complex number denoted by ({, g) satisfying the conditions (f, g) = (g, f), (Af, g) = A (f, g), ({1 + {2, g) = ({1, g) + ({2, g). (50) (51) (52) 
Abstract Hilbert space H 81 It follows at once that (f, f) = (f, f) so that (f, f) is a real number. We shall assume that (f, f) o and further that (f, f) = 0 if and only if f = O. Then we have also that (53) (f, Ag) = (Ag, f) = X(g, f) = X(f, g) (54) and (f, gl + g2) = (gl + g2, f) = (gt, f) + (g2, f) = (f, gl) + (f, g2). (55) The norm of the element f is defined as Ilfll =  (f, f) . (56) We see that 11111 = 0 if and only if f = 0, and IIAfl1 = IA 111111. Now (f+ Ag, f+ Ag)  0 and so taking A = -(f, g)/(g, g) we obtain Schwarz's inequality 1II1111gli  I(f, g)l. (57) Again following the proof given in section 6.3 we. arrive at the triangle inequality 11111 + Ilgll  Ilf + gll. (58) We now define a distance function d(f, g) in terms of the norm according to the formula d(f, g) = Ilf- gll. (59) This satisfies the conditions required of a distance between two points f and g, namely (i) d(f, g) = d(g, f), (ii) d(f, g)  0, (iii) d(f, g) = 0 if and only if f = g, (iv) d(f, g)  d(f, h) + d(h, g). (60) This last condition follows from the triangle inequality: Ilf - gll = lI(f - h) + (h - g)1I  IIf - hll + IIh - gll. 
82 Hilbert space A sequence of elements {fn} converges strongly to a limit element f if, given any E > 0, there exists a N such that for n > N we have Ilfn - fll < E. Strong convergence is denoted by fn  f. If fn  f we have Ilfn - fmll = II(fn - f) + (f - fm)11 llfn - fll+ Ilfm - fll< E for sufficiently large n, m. A sequence {fn} of elements satisfying Ilfn - fmll < E for sufficiently large n, m is known as a Cauchy se- quence. A Hilbert space is complete, that is every Cauchy sequence converges to a limit vector in the space. 6.4.1 Dimension of HUbert space To define the dimension of a space R we introduce the following concepts. A subset T of R is said to be dense in R if, given any E > 0 and any element fER, there exists an element gET such that Ilf - gll < E. S is called a fundamental set in R if the set T of all linear combinations L;=1 crfr with fr E S is dense in R. Then the dimension of the space R is the least possible cardinal number of a fundamen- tal set S in R. The dimension of Hilbert space H is required to be denumerably infinite. 6.4.2 Complete orthonormal system Any set of linearly independent elements belonging to H can be combined together to form an orthonormal system cPt, cP2, · · · , cP" · · · satisfying (cPr' cPs) = 5rs by applying the Gram-Schmidt orthogonalization process described in section 6.3.2. Because the dimension of H is denumerably infinite it follows that any orthonormal system of elements belonging to H has not more than a denumerable infinity of elements which we may denote by cPt, cP2,. · · , cPr' · · · Given any element f E H and putting n fn = L (f, cPr)cPr r=1 (61) 
Abstract HHbert space H 83 where the (f, cPr) are the Fourier coefficients of f, we can show that n IIf - fnl1 2 = 11111 2 - L I(f, cPr)12 r=1 (62) by following the analysis given in section 6.3.3. Since Ilf- fnllO, we obtain Bessel's inequality n 11111 2  L I(f, cPr)12, r=1 (63) valid fori all n, from which it follows that the series on the right-side converges as n  00. Hence, given any e > 0, we have Ilfn - fml1 2 = 'r=+l (f, <Pr)<Pr 2 (n> m) n - L l(f,cPr)12<e r=m+1 provided n, m are sufficiently large. Therefore {fn} is a Cauchy sequence and so there exiss an element f' E H such that fn .....:; f' as n  00, using the fact that H is complete. Thus 00 f' = L (f, cPr)cPr r=1 (64) and 11111 2  Ilf'11 2 . (65) If the system cP1, cP2,. . . , cPr' . . . is a fundamental set in H then f' = f and we have 00 IIfl1 2 = L I(f, cPr)12. r=1 (66) Then cP1, cP2, · · · , cPr' · . . is called a complete system or basis and (66) is known as Parseval's formula or the completeness relation. Following the analysis at the end of section 6.3.3 we can also show that the generalized Parseval formula (39) holds for f, g E H. Problems 1. Construct an orthonormal set of functions defined in the interval -1  x  1 from 1, x, x 2 , x 3 using the Gram-Schmidt orthogonaliza- tion process. Verify that the functions are proportional to the first four Legendre polynomials. 2. Show that the set of orthogonal functions {cos nx} (n = 0, 1, 2,. . .) do not form a basis spanning the space of continuous functions defined over the interval - Tr  X  Tr. Likewise show that the set of 
84 Hilbert space orthogonal functions {sin nx} (n = 1, 2, . . .) do not form a basis of the space. 3. Show that the sequence of functions {fn(X)} where o (ox  ) fn();)=  C <x<  ) o (   x  1 ), converges to zero at all points of the interval 0  x  1 but that the sequence is not mean square convergent to zero. 4. Establish that the space of continuous functions is not complete by showing that the sequence of continuous functions {fn(x)} where o In(x) = !(nx + 1) 1 (-lX-  ) ( -  < x <  ) (  Xl), converges strongly to the discontinuous function o I(x) = ! 1 (-1 x <0) (x=O) (O<xl). 5. Show that if a Cauchy sequence {In} converges to an element I then I is unique. 6. Verify that the space of infinite sequences 1 2 discussed in section 6.2 satisfies the axioms for abstract Hilbert space. 7. By considering the Cauchy sequence {xn} where X n = (1, !, l, . . . , 1 / n, 0, 0, . . .), show that the linear vector space of infinite sequences (x h X2, · . . , x" . . .) in which only a finite number of the coordinates x, do not vanish, is not complete. 
7 Linear operators in Hilbert space In an integral equation the unknown function occurs under the integral sign and thus, if the functions involved belong to a Hilbert space, it is clear that we have to deal with integral operators acting on a Hilbert space of functions. We pointed out in section 6.3.4 that square integrable functions in the Lebesgue sense, that is L z functions, form a Hilbert space and consequently the appropriate integral operators have L Z kernels introduced in section 1.4. However, in order to avoid unduly difficult concepts, we shall suppose that ou functions and kernels are square integrable without usually specifying the sense in which the integrals are to be performed. It is worth while placing linear integral operators in a more general context and so we shall conclude this chapter by giving an introduction to the theory of linear operators in an abstract Hilbert space. 7 .1 Linear integral operators We consider the linear integral operator K=fK(X,S)dS where K(x, s) is a square integrable kernel, and write tf1(x) = f K(x, s)cf>(s) ds, (1) (2) where cP(s) is a square integrable function, in the symbolic form t/J = K cP. (3) It is evident that the operator K is linear since K(At cPt + AzcPz) = AtK cPt + AzK cPz (4) 
86 Linear operators in Hilbert space where AI, A2 are constants and cf>I, cf>2 are square integrable func- tions. We also introduce the identity operator I satisfying Icf> = cf> (5) for every square integrable function cf>(s), which may be expressed in the form of the integral operator 1= f8(X-S)dS (6) where 5 is the Dirac delta function defined in section 2.2. If L= fL(X,S)dS (7) is a second integral operator we have x = LtfJ = L(Kcf> ) (8) where x(x) = f L(x, t) dtf K(t, s)<I>(s) ds = f P(x, s)<I>(s) ds (9) and P(x, s) = f L(x, t)K(t, s) dt, (10) that is x = Pcf> where P = LK is the integral operator with kernel P(x, s). Integral operators satisfy the associative law (11) M(LK) = (ML}K (12) and the distributive laws M(L+K)=ML+MK (L + K)M = LM + KM but, in general, do not satisfy the commutative law. (13) (14) 
Linear integral operators 87 Using the associative law we see that KmK n = K m + n , (Km)n = K mn (m, n 1) (15) where K n = f Kn(x, s) ds (16) and Kn(x, s) is the iterated kernel defined by (4.22). 7.1.1 Norm of an integral operator If K(x, s) is a square integrable kernel its norm is defined by [ r b r b ] 1/2 IIKIIz = Ja lIK(x, sW dx ds · (17) Then if cP(s) is a square integrable function and tf/(x) is given by (2) _ we have, using Schwarz's inequality (6.20), that 1t/1(xW..;; fIK(x, sW dsfl<P(SW ds which yields ,fJt/1(XW dx";; f fIK(x, sW dx dsfl<p(sW ds so that Iitf/II::;; IIK11211ef>11 < 00 (18) and thus tf/(x) is square integrable. When IIKI12 = 0 then Iitf/il = 0 so that Kef> vanishes 'alm,ost. everywhere' for all square integrable functions ef>, and K is called a null operator. Also if L(x, t) and K(t, s) are square integrable kernels, then P(x, s) given by (10) is square integrable since, by Schwarz's in- equality, IP(x, sW..;; l"IL(X, tW dtfIK(t, sW dt 
88 Linear operators in Hilbert space so that f fIP(x, sW dx ds  f fIL(x, tW dx dtf fIK(t, sW dt ds giving IILK1b::;; IILlb IIKlb < 00. (19) Putting L = K we see that IIK 2 Ib::;; IIKII and in general IIKnlb ::;; IIKII (n = 1, 2, 3, . . .) (20) 7.1.2 Hermitian adjoint The Hermitian adjoint of a kernel K(x, s) is defined to be the kernel K*(x, s) = K(s, x). (21) We see at once that (AK)* = AK*, K** = K. (22) Also (LK)*(x, s) = LK(s, x) = r L(s, t ) K(t, x ) dt = f K*(x, t)L *(t, s) dt = K*L *(x, s) and so (LK)* = K*L*. (23) Further, using the definition of the inner product given by (6.16), we have (K t/>, .,,) = r "'(x) dx f K (x, s )t/>(s) ds = J: t/>(s) dSJ: K(x, s) "'(x) dx = (cP, K*.p). (24) If K* = K the kernel is called Hermitian or self adjoint. 
Bounded linear operators 89 7.2 Bounded linear operators So far we have been concerned with linear integral operators acting on a space of square integrable functions which, if chosen to be all the functions which are integrable in the Lebesgue sense, would form a Hilbert space of functions L 2 . We shall now turn our attention to the case of an abstract Hilbert space, defined in section 6.4, acted on by bounded linear operators. Consider a subset D of an abstract Hilbert space H such that if f, g E D then Af + ILg E D where A, IL are arbitrary complex numbers. Then D is called a linear manifold. We note that a linear manifold must contain the zero element since f+(-l)f=O. Suppose that corresponding to any element fED we assign an element Kf E  where  is also a linear manifold. Then K maps D onto  and is called a linear operator with domain D and range  if K(Af+ ILg) = AKf+ ILKg (25) for f, g E D. A linear operator K having the Hilbert space H as domain is bounded if there exists a constant C  0 such that IIKf1I C IIfll for all IE H. The norm IIKII of the bounded linear operator K is defined as the smallest possible value of C. Thus IIKII is the least upper bound or supremum of IIKf1I/IIf1I, that is IIKf11 IIKII = E lfiI (26) and so IIKfll  IIKIlIlf1I. (27) This is a generalization of the inequality (18) for the linear integral operator (1) with square integrable kernel. Clearly (1) is a bounded linear operator. The linear operator K has an adjoint operator K* defined so that (KI, .g) = (f, K*g) (28) for all I, g E H. This is in accordance with the definition (21) for linear integral operators, as shown in section 7.1.2. If K* = K the operator is self adjoint or Hermitian. 
90 Linear operators in Hilbert space Now by Schwarz's inequality (6.57), and (27), we have I(Kf, g)1  IIKt1ll1gll  IIKII I If II Ilgll (29) and so (Kf, g) is bounded. Likewise I(f, K* g)1  IIK*lIlIfllllgll (30) so that IIK*II = IIKII (31) and K* is also a bounded linear operator. Further, any bounded operator K is a continuous linear operator which transforms a strongly convergent sequence {fn} into a strongly convergent sequence {Kfn}. For we have IIKfn - Kt11 = IIK(fn - f)1I  IIKllllfn - fll so that if fn  f as n  00 then IIKfn - Kfll  0, that is Kfn is strongly convergent to Kf Actually every continuous linear operator K with domain H is bounded. For otherwise there would exist a sequence {fn} such that IIKfnll  nllfnll so that IIKgnll  1, but gn = fnln IIfnll  0 as n  00, which is contrary to the hypothesis that K is continuous. We see that the linear integral operator (1), with a square integrable kernel, is continuous. If K and L are two bounded linear operators which map H onto itself, then the product operator LK corresponds to (LK)f = L(Kn. Then IILKfll  ilL IIIIKt11  ilL II IIKII 1It11 so that IILKII  IILIIIIKIl (32) and therefore LK is a bounded operator. Also (LKf, g) = (Kf, L * g) = (f, K*L*g) 
Bounded linear operators 91 and hence (LK)* = K*L*. (33) Lastly it is interesting to observe that the norms of bounded linear operators K and L satisfy the triangle inequality. For, using the triangle inequality (6.58) for the elements of Hilbert space, we have IIKf + Lf11  IIKf11 + IILf11  (11K11 + 111)11f11 so that IlK + LII  IIKII + IILII. (34) 7.2.1 Matrix representation Suppose that CPt, CP2, . . . , CPr' . . . is a complete orthonormal system or basis in H. Then the matrix with elemnts k rs = (K CPr, CPs) (35) is called the kernel matrix of the bounded operator K. Introducing the Fourier coefficients X r = (f, CPr), Ys = (g, CPs) (36) where f, g E H, we have 00 f = 1 XrCPr, r=1 00 g = 1 YsCPs s=1 (37) and 00 00 00 Kf = 1 xrK CPr = L 1 XrkrsCPs r=1 r=1s=1 so that (Kf, g) has the bilinear form 00 00 (Kf, g) = 1 1 XrkrsYs. r=ls=1 (38) Since K is bounded it follows that I(Kf, g)1  IIKIlIlf1lllgll = IIKII  Ctl I X rl 2 )Ctl IY s I2). 
92 Linear operators in Hilbert space Hence rl Sl xrkrsys .;;; IIKII Cl I X rl 2 )Cl IY s I2) (39) and thus the bilinear form (38) is also bounded. 7.3 Completely continuous operators We now introduce the concept of weak convergence. A sequence of elements {fn} in Hilbert space is said to converge weakly to a limit element f if (fn, g)  (f, g) as n  00 for all g E H. Weak convergence is written fn  f as n  00. If a sequence is strongly convergent it is also weakly convergent. For, by Schwarz's inequality (6.57), we have I(fn - f, g)1  IIfn - fll !tgll and thus if IIfn - fll  0 as n  00 then (fn, g)  (f, g) as n  00. However, the converse need not be true. Thus we may have fn  f but fn  f as n  00. Let us suppose that K is a bounded linear operator in H. Then I(Kln - Kf, g)1  IIK(ln - 1)11 IIgil  IIKII IIf n - fll II gll and so if fn  f as n  00 it follows that Kfn  Kf, that is Kfn is weakly convergent to Kf. Now suppose that Kfn  Kf when fn  f as n  00. Then K is called a completely continuous (or compact) linear operator in H. Every completely continuous operator is bounded. For if fn  f as n  00 we have shown above that fn  f and so Kfn  Kf. Thus a completely continuous operator is a continuous operator and this means that it is bounded as shown in section 7.2. However a continuous operator is not necessarily completely continuous. For example the identity operator I is continuous but it s not completely continuous for if fn  f but fn  f as n  00 then obviously Ifn  If but Ifn  If. If the kernel matrix (k rs ) of a bounded linear operator K satisfies 00 00 L L Ik rs l 2 <00 r=ls=l (40) 
Completely continuous operators 93 then K is completely continuous. For we have, using (38), that 00 (K(fn - f), cPs) = L (xn) - xr)k rs r=1 (41) where xn) and X r are the Fourier coefficients of fn and f respec- tively, which gives m IIK(fn - f)II2 L I(K(fn - I), cPs)12 s=1 00 00 + L L Ik rs l 2 IIfn - fll 2 s=m+1 r=1 (42) using Cauchy's inequality. On the right-hand side of (42) the second sum approaches zero as m  00 uniformly in n since IIfn - fll is bounded and the operator K satisfies the condition (40). The first sum approaches zero as n  00 since (K(fn - f), cPs) = (fn - f, K*cPs)  0 if'fn  f. Thus Kfn  Kf if fn  f as n  00 and thus K is completely continuous. Any finite dimensional linear operator is copletely continuous since for this case the double sum on the left-hand side of (40) contains a finite number of terms only. 7.3.1 Integral operator with square integrable kernel As an example of a completely continuous operator we consider the integral operator K over the space of square integrable functions given by g(x) = f K(x, s)f(s) ds (axb) (43) where the kernel is square integrable and satisfies fIK(x, sW ds <00 (axb), fIK(x, sW dx <00 f fIK(X, sW dx ds <00. (asb), (44) 
94 Linear operators in Hilbert space By Schwarz's inequality we have Ig(xW.;:;; fIK(x, sW dSflf(sW ds and so flg(x W dx';:;; f fIK(x, sW ds dx flf(SW ds, that is Ilg112.;:;; f fIK(x, sW ds dx Ilf11 2 . (45) Hence IIKII.;:;;  f fIK(x, sW ds dx= IIKlb (46) since the norm of K is the least upper bound of Ilg"'"fll. Now consider any complete orthonormal system cPl(X), cP2(X), . . . , cPr(x), . .. in the Hilbert space of L 2 functions defined over a  x  b. We let krs = r r cf>s(x) K(x, t)cf>r(t) dxdt, x r = r f(t) cf>r(t) dt (47) (48) and Ys = r g(x) cf>s(x) dx. (49) Then Ys = r r K(x, t)f(t) cf>s(x) dx dt (50) But f b 00 a K(x, t)cf>s(X) dx = rl k,.scf>r(t) (51) and 00 f(t) = L XrcPr(t) r=l (52) 
Problems 95 for almost all values of t, that is except for a set of values of t of Lebesgue measure zero, and hence 00 Ys = L xrkrs. r=l (53) Also r r K(x, t) <p.(x) dx 2 dt= rt 1 1k,.1 2 and, using Parseval's formula (6.66), fIK(X, tW dx = 'l f K(x, t) <p.(x) dx 2, (54) (55) Consequently ffIK(x, tWdx dt= rt.J 1 1k,.1 2 (56) and so the matrix representation of the integral operator K with square integrable kernel satisfies the condition (40) which ensures that K is completely continuous. Problems 1. If 1 2 is the Hilbert space of infinite squences x = (Xh X2, . . . , X r , . . .) and K is the operator defined by Kx = (0, X2, X3, . . . , x,., . . .), show that K is a linear operator in 1 2 having unit norm. 2. If K is the operator in 1 2 defined by Kx=y where Yr =x r +l/(r+ l)(r  1), show that K IS a bounded linear operator and find IIKII. 3. Prove that the linear differential operator L = d/dx, acting on the space of continuous functions f(x) defined in-the interval Oxl, is unbounded by showing that IILfnl1 = n7T where fn(x) = J2 sin n7TX. 
96 Linear operators in HUbert space 4. Show that (i) the integral operator K 1 = i 1 sin xs ds o S (Oxl) is bounded, but that (ii) the integral operator i 1 sin xS d K 2 = 2 S o S (Ox 1) is unbounded. 5. A normal operator K satisfies KK* = K*K. Show that the normal operator K also satisfies IIKf11 = IIKfll for every element f E H. If M = (K + K*)/2, N = (K - K*)/2i show that the operator K is normal if and only if M and N commute, that is MN = NM. 6. Find the kernel matrix (k rs ) for K(x, t) = cos (x - t) (-1TX1T, -1Tt1T) using the orthonormal basis composed of the trigonometric func- tions { Ill. } .[2;, ' .J; cos rx, .J; sIn rx (r=1,2,...) defined in the interval -1T  X  1T. Show that t: t: IK(x, tW dx dt =   11<,.1 2 = 271"2. 7. Find the kernel matrix (k rs ) for K(x,t)=X 2 +t 2 (-lxl,-ltl) 
Problems 97 using the orthonormal basis composed of the functions /2r+ 1 cPr(X) = V 2 Pr(X) (r=O, 1,2,...) where the Pr(x) are Legendre polynomials defined in the interval -lxl. Show that i 1 i l 00 00 112 -1 -1 IK(x, tW dx dt = ro so Ik,.l = 45 · 
8 The resolvent We have already indicated in previous chapters that the solution of an integral equation of the second kind can be expressed in terms of a resolvent kernel. The purpose of the present chapter is to examine the resolvent kernel and the resolvent operator in some detail and to derive results of greater generality than before. 8.1 Resolvent equation Let us turn our attention again to the Fredholm linear integral equation of the second kind cP(x) = f(x) + A r K(x, s)cP(s) ds (axb) (1) where A is a parameter. Introducing the integral operator K= rK(X, s)ds (2) we may rewrite this equation in the form cP = f+ AKcP (3) or (1 - AK)cP = f. (4) We now seek an integral operator R given by R= r R(x, s; A) ds (5) such that cP=f+ARf, (6) that is cP = (1 + AR)f. (7) The operator R depends on the parameter A and since it provides 
Uniqueness theorem 99 the solution to the integral equation (1), R is called the resolvent and R(x, s; A) the resolvent kernel. We have seen in section 1.3.3 how the resolvent kernel arises in a natural way for integral equa- tions possessing separable kernels. Substituting (7) into (4) we find that, provided R exists, it must satisfy (I - AK)(I + AR)f = f and so we anticipate that (8) R-K=AKR. Likewise substituting (4) into (7) we find that cP = (I + AR)(I - AK)cP (9) (10) which yields R-K=ARK. (11) The oprator equation R- K = AKR = ARK (12) is called the resolvent equation. If there exists an operator R with a square integrable kernel R(x, s; A) satisfying the resolvent equation (12) for a given value of A, then this value of A is said to be a regular value of the kernel K(x, s). The set of all regular values of an operator K is known as the resolvent set A. The adjoint equation of (3) is defined as t/J = g + AK*t/J (13) where K* is the adjoint operator of K and g is a given square integrable function while t/J is a square integrable solution. Taking adjoints of the operators occurring in the resolvent equa- tion (12) we obtain R* - K* = AR* K* = AK* R* (14) and so R* is the resolvent for the adjoint equation (13). Evidently A is a regular value of K* if and only if A is a regular value of K. 8.2 Uniqueness theorem We shall show first in this section that if there exists a resolvent kernel R(x, s; A) of the kernel K(x, s) for a given value of the parameter A, then it is unique. 
100 The resolvent For suppose that there are two such kernels R 1 (x, s; A) and R 2 (x, s; A). Then R 1 - K = AR 1 K = AKR 1 R 2 -K=AR 2 K=AKR 2 and so, putting r=R 1 -R 2 (15) we see that r = AKr. (16) Hence R 1 r = AR 1 Kr = (R 1 -K)r =R 1 r-Kr yielding Kr = o. (17) It follows from (16) that r = 0 giving R 1 = R 2 . (18) Next we show that if f(x) is a square integrable function and if A is a regular value of the square integrable kernel K(x, s) possessing the square integrable resolvent kernel R(x, s; A), then the integral equation (3) has the unique square integrable solution (6). Suppose that the function cP(x) is given by (6). Then we have f+ AKcP = f+ AK(f+ ARt) = f+AKf+A2KRf = f+ AKf+ A(R- K)f =f+ARf =cP and so cP is a solution of the integral equation (3). Conversely, if the square integrable function cP satisfies (3) we have f = cP - AK cJ> 
Characteristic values and functions 101 so that f+ ARf= cp - AKcp + AR(cp - AKcp) = cp+A(R-K-ARK)cp =cp using the resolvent equation (12), which proves the uniqueness of the solution. However it should be noted that there may exist other solutions of (3) which are not square integrable. 8.3 Characteristic values and functions We next consider the homogeneous linear integral equation of the second kind <I> (x) = Af K(x, S)<I>(S) ds (19) which we may rewrite in the form cp = AKcp (20) where K is the linear integral operator (2). This equation clearly has the solution cp = 0 for all values of A. In addition to the solution cp = 0, the integral equation (20) may have other square integrable solutions CPl (x), CP2(X), . . . , CPv(x), . .. for certain values A b A2, . . . , Av, . . . respectively of the parameter A. We shall call these values of A characteristic values. However they are sometimes called eigenvalues although the term eigenvalue is usu- ally reserved for the values of A -1. The corresponding solutions of (20) are called characteristic functions or eigenfunctions. A regular value A of a square integrable kernel cannot be a characteristic value. For then we would have cp = f+AKcp with f = O. But for a regular value of A the integral equation has the unique square integrable solution cp=f+ARf from which it follows immediately that cp = O. The complement of the resolvent set A is called the spectrum L of the operator K. Thus L contains the set of characteristic values of K. 
102 The resolvent If cp(x) is a square integrable characteristic function of a continu- ous kernel K(x, s) we have I cfJ (x) - cfJ (x') I = IA I f {K (x, s) - K (x', s)}cfJ (s) ds  IAlllcfJl1 {fIK(X, s) - K(x', sW dSY'2 using Schwarz's inequality. Hence, given e > 0, there exists S > 0 such that I cp (x) - cp (x')1 < e if Ix - x'I < S by virtue of the continuity of K(x, s), and thus cp(x) is also continuous. However if A-I = 0 corresponding to a zero eigenvalue the above argument breaks down and the eigenfunction can be discontinuous (see problem 5, p. 112). Now suppose that cp is a characteristic function of the kernel K for the characteristic value A, and .p is a characteristic function of the adjoint kernel K* for the characteristic value  where IL  A. Then (cp, .p) = (AK cp, .p) = A (K cp, .p) and (cp, .p) = (cp, iiK*.p) = IL (K cp, .p) using (7.24), so that (A -IL)(K cp, .p) = o. Since IL  A it .,follows that (K cp, .p) = 0 and hence (cp, .p) = 0, that is, cp and .p are orthogonal. 8 .4 Neumann series We have already shown in chapter 4 that the method of successive substitutions leads to the solution of the Fredholm linear integral equation of the second kind (1) in the form of an infinite series, named after Neumann, given by cfJ(x) = f(x) + Af R(x, s; A)f(s) ds (21) 
Neumann series 103 where ex> R(x, s; A) = L A nK n + 1 (x, s) n=O (22) is the resolvent kernel. In terms of the integral operator (2) it can be readily seen that the solution cp to the integral equation (1) can be written ex> cp = f + LAn Knf n=l (23) while the resolvent R can be expressed in the form ex> R= L A n K n + 1 . n=O (24) The Neumann series for the solution to the Fredholm equation (1) and for the resolvent kernel (22) can be shown to be convergent under considerably less stringent conditions than those derived in sections 4.1 and 4.2. Thus let us suppose that f(x) is a square integrable function and that K(x, s) is a square integrable kernel satisfying fiK(X, tW dt < A 2 (axb) (25) where A is a finite positive constant. Since K n +1(x, s) = f Kn(x, t)K(t, s) dt where Kn(x, s) is the iterated kernel defined in section 4.2, we have by Schwarz's inequality IK n +1(x, sW.;;; fIKn(X, tW dtfIK(t, sW dt (26) so that, on integrating over s, we obtain fI Kn +1(X, sW ds.;;; fIKn(X, tW dt"K" (27) where IIKII2 is the norm of K(t, s) given by (7.17). Repeated application of (27) gives fIKn+1(X, sW ds.;;; A 2"KIIn (28) 
104 The resolvent from which it follows that f R(x, s; A)f(s) ds  no lAin f K n +1(x, s)f(s) ds  no lAin Ilfll{fI Kn +1(X, sW dS} 1/2 ex> A Ilfll L IAlnllII. (29) n=O Hence the Neumann series for cp(x) converges absolutely and uniformly for IA IIIKII2 < 1. If we assume further that fIK(t, sW dt<B 2 (asb) (30) where B is a finite positive constant, then ex> IR(x, s; A)-K(x, s)l L IAl n IK n + 1 (x, s)1 n=l  nl IAln{fIK(X, tW dtr21IKII-1{fIK(t, sW dtr2 ex> AB L IAlnIlKII-l (31) n=l and so the Neumann series for the resolvent kernel converges absolutely and uniformly if IA IIIKlb < 1. We have shown in section 8.2 that the solution cp and the resolvent R are unique. Hence if IA IIIKII2 < 1 the corresponding homogeneous equation (19) has the unique solution cp = O. However this can be verified directly for, by Schwarz's inequality, we have IcP(xW IAI211cPI12 fIK(X, sW ds giving, on performing an integration over x, IIcpl12  IA 1211cp11211KII I.e. (1-1'\121IKII)lIcpI12 O. But IA IIIKI12 < 1 and so it follows that Ilcpll = 0 which implies cp = O. 
Neumann series 105 This also demonstrates that the solution of the integral equation (1) is unique when IAIIIKII2 < 1. For if there exist two solutions CPl (x), CP2(X) satisfying <l>l(X) = f(x) + Af K(x, S)<I>l(S) ds <l>2(X) = f(x) + Af K(x, S)<I>2(S) ds, we see at once that cp(x) = CPl(X) - CP2(X) is a solution of the homogeneous equation (19) from which it follows from the preced- ing result that cp(x) =,0, giving CPl(X) = CP2(X). 8.4.1 V olterra integral equation of the second kind We now examine the convergence of the Neumann series solution of the Volterra linear integral equation of the second kind <I> (x) = f(x)+ Af K(x, s)<I>(s) ds (axb) (32) where f(x) is a square integrable function and K(x, s) is a square integrable kernel satisfying K(x, s) = 0 for a  x < s  b, and a 2 (x) = fIK(x, sWds<A 2 2(S) = fIK(x, sW dx < B 2 (axb) (33) (asb) (34) where A and B are positive constants. We shall show that the Neumann series for the resolvent kernel is absolutely and uniformly convergent for all A. We begin by noting that, from Schwarz's inequality, IK 2 (x,sW= fK(X,t)K(t,S)dt 2  fIK(X, tW dtfIK(t, sW dt  a2(x)2(s) (35) 
106 The resolvent and J x 2 IK 3 (x, sW = s K(x, t)K 2 (t, s) dt  fIK(X, tW dtfIK2(t, sW dt  a?(x)2(s) f a 2 (t) dt, I.e. , 1 K 3 (x, s) 1 2  a 2 (x) {3 2 (s ) k 1 (x, s) (36) where k 1 (x, s) = f a 2 (t) dt. (37) Next we use the principle of induction to establish the inequality I K n+2(x, s)1 2  a 2 (x){32(s)k n (x, s) (n = 1, 2, . . .) (38) where kn(x, s) = f a 2 (t)k n _ 1 (t, s) dt We have already proved that the inequality (38) holds for n = 1. Now assuming the truth of the inequality for a positive integer n and using Schwarz's inequality again, we see that (n = 2, 3, . . .) (39) IK n + 3 (x, sW= fK(X, t)K n + 2 (t, s) dt 2  fIK(X, tW dtfIK n + 2 (t, sW dt  a2(x)2(s) f a 2 (t)k n (t, s) dt = a 2 (x){32(s )k n + 1 (x, s) which establishes the inequality for integer n + 1. Also, by using the principle of induction, we can show that k ( ) = {k 1 (x, s)} n n X, S ,. n. (40) 
Neumann series 107 Obviously this holds for n = 1. Assuming its validity for a positive integer n we have, using the definition (39), that k n + 1 (x, s) = f Xa2(t){k1(t, sW dt n. s =  J x { i. kl(t, S) } {k 1 (t, s)}n dt n. s at = 1- [ {k 1 (t, s)}n+l ] x n! n+1 s { k 1 (x, s)} n + 1 (n + I)! and so the formula (40) holds for integer n + 1. Now k 1 (x, s) = fdt{' IK(t, sW ds  IIKII (41) and so, using (38), (33), (34), (40) and (41) we obtain AB IIKII IKn+2(x,s)l J;! · n! (42) Hence ex> IR(x, s; A)-K(x, s)l L IAln+lIK n + 2 (x, s)1 n=O ,c:: f I A In +1 AB IIKII n=o  = IAI AB f IAlnIlKII (43) n=O  which shows that the Neumann series for the resolvent kernel is absolutely and uniformly convergent for all A since the infinite series on the right-hand side of (43) is of the type L:=O zn rJn! which converges for Izl < 00 by the ratio test. Thus every value of A is a regular value for the Volterra equation of the second kind. Since a regular value of A cannot be a characteristic value, it 
108 The resolvent follows that the homogeneous equation cf>(x) = Af K(x, s)cf>(s) ds (axb) (44) has the unique square integrable solution cp = 0 for all A. The Neumann series solution of the Volterra equation of the second kind (32) takes the form ex> cp(x) = f(x) + A L A nfn+l(X) n=O (45) where fn(x) = f Kn(x, s)f(s) ds. (46) We have, by Schwarz's inequality, that Ifl(XW fIK(X, sW dsflf(sW ds  A 211fll 2 (47) and, using the inequalities (38) and (41) together with (33) and (40), that Ifn+l(xW,;; fIKn+1(X, sW ds{'lf(SW ds  A 21IKII(n-1) f x 13 2 (s) ds 11ft (n-1)! a  A 211f1l211KIIn -.;:: . (n -I)! (48) Hence ex> Icp(x)-f(x)IIAI L IAl n lfn+l(x)1 n==O  IAI A Ilfll { I + ! IAnKII; } (49) n=l  (n -I)! and thus the Neumann series for cp(x) is absolutely and uniformly convergent for IA 1< 00. This demonstrates that the Volterra equation of the second kind (32) has a unique square integrable solution cp(x) for any square integrable function f(x) and all values of A. 
Fredholm equation in abstract Hilbert space 109 8.4.2 Bocher's example Although we have shown that the homogeneous Volterra equation (44) with a square integrable kernel has the unique square integra- ble solution cp = 0 for all values of the parameter A it may have other solutions which are not square integrable. Thus consider the homogeneous Volterra equation cf>(x) = r sx-scf>(s) ds (Oxl) (50) with the bounded kernel K(x, s) = { S-S (0< s x  1) (s = 0) (51) given by Bacher in 1909. It can be readily verified that it has the discontinuous solution { cx x-I cpo(x) = 0 (O<xl) (x=O) (52) where c is an arbitrary constant. This solution is not square integra- ble since r cf>(x) dx evidently diverges owing to the presence of the x- 2 singularity near the origin. We now see that if cp(x) is a solution of cf>(x) = f(x) + r sx-scf>(s) ds (0  x  1), (53) then cp(x) + CPo(x) is also a solution of (53). Thus the solution of (53) is unique if and only if we restrict the solution to be square integrable. 8.5 Fredholm space equation . In abstract Hilbert In the previous sections of this chapter we discussed Fredholm's integral equation (1) in terms of the integral operator K given by 
110 The resolvent (2). However we may consider a more general form of Fredholm equation cp=f+AKcp (54) where K is a completely continuous linear operator which maps abstract Hilbert space H onto itself, and the given element f and the solution cp belong to H. If there exists a continuous linear operator R called the resolvent satisfying the resolvent equation R-K=AKR=ARK (55) for a given value of A, this value is called a regular value. For such a value of A the resolvent operator is unique as can be verified by using the method described in section 8.2. Moreover (54) has the unique solution cp = f+ARf (56) as can be seen by direct substitution. The adjoint equation of (54) is .p = g + AK*.p (57) where K* is the adjoint of K, and g and .p belong to H. The resolvent R* of (57) satisfies the resolvent equation R*-K*=AR*K*=AK*R*. (58) The homogeneous equation cp = AKcp (59) possesses the trivial solution cp = 0 for all A. It may also have other solutions belonging to H called characteristic vectors CPb CP2, · · · , CPv, . .. for certain values AI, A2, . . . , Av, . . . of A called characteristic values. Following the analysis given in section 8.3 we can show that a regular value of A cannot be a characteristic value. Also (<f>,\fI)=O (60) where cp is a characteristic vector of K for the characteristic value A and .p is a characteristic vector of K* for the characteristic value jl where IL  A. We see from (55) that the resolvent is given by R=K(I-AK)-l (61) 
Problems 111 where the inverse L -1 of an operator L satisfies L- 1 L=LL- 1 =1 , (62) I being the identity operator. The resolvent operator may be expressed In the form of a Neumann expansion 00 R= L A n K n + 1 n=O (63) where, using the triangle inequality (7.34) for bounded linear operators, 00 IIRII L lAin IIK n + 1 11 n=O 00  L lAin IIKlln+1 n=O since IIK n + 1 11  IIKlln+1 by successive application of (7.32). Hence R is bounded if IA IIIKII < 1. We also see that R is completely continuous if IAIIIKII < 1, for we have 00 IIRfm - RfJl = L A nK n + 1 (fm - f) n=O 00  L A nK n IIK(fm - f)11 n=O o as fm  f SInce K IS completely continuous and L=o A n K n IS bounded. Problems 1. If R A , RIJ- are the resolvents for a given kernel K corresponding to parameter values A, IL respectively, show that R A - RIJ- = (A -IL)RARIJ- 2. If the linearly independent functions <Ph <P2, · · · , <Pn; .ph .p2, · · · , .pn form a biorthogonal series so that (<Pi, .pj) = 8 ij (i, j = 1, 2, . . . , n) , show that the kernel n K(x, s) = L Ui(X) Vi(S) i=l (a  x  b, a  S  b) (1) 
112 The resolvent has the characteristic functions cf>i(X) with characteristic values ail (i = 1, . . . ,n). Further, use the Neumann expansion to show that the resolvent kernel of K(x, s) converges to R(x, s; A) = ! ai<f>;(x)t/1i(S) i = I 1 - ai A if laiAI < 1 for all i. Verify that the relation obtained in problem 1 is satisfied. 3. Show that the kernel 00 K(x, s) = L a v cos vx cos vs v=o (0  x  7T, 0  S  7T), where L:=o I a v 1< 00, has the characteristic functions 1/.[;, -J2/7T COS vx (v = 1, 2, . ..) with characteristic values 1/7Tao,2/7Tav respectively. Use the Neumann expansion to show that for sufficiently small A the resolvent kernel for K(x, s) is given by 00 R ( . ) - ao  a v cos vx cos vs x, s , A - + i..J . 1- 7TaoA v=l 1- 7T/2 avA 4. Show that the Poisson kernel 1 1- a 2 K(x, s) =- 2 (O X 27T, O S 27T), 27T 1 +a -2a cos (x-s) where lal < 1, can be expressed as the Fourier series 11 00 -+- L a V cos v(x-s). 27T 7T v=l Obtain the Neumann expansion for the resolvent kernel of K(x, s) and show that it converges to 1 1 1 f a V R(x, s; A)=- 2 1 A +- i..J 1 v cos v(x-s) 7T - 7T v=l - a A if I A I < 1. 5. Show that the continuous kernel K(x, s) = x{n -1- (2n -l)s} (Oxl, Osl) 
Problems 113 has the discontinuous square integrable eigenfunction when n > 2: { -l/n <f>(x) = X 0 (O<xl) (x=O) for the eigenvalue 0, that is r K(x, s)<f>(s) ds = O. 6. Show that the discontinuous kernel K(x, s) = { (Ox<!,Osl) (!xl,Osl) has the discontinuous characteristic function <f>(x) = { (O x <!) (! x  1) for the characteristic value 2. Use the Neumann expansion to show that the resolvent kernel of K(x, s) is given by R(X,S;A)= { O 2 (! x  1) 2-A (O x <!) forIAI<2. 7. Show that the discontinuous kernel {( S ) 1/2 K(x, s)=  (O<xl,Osl) (x = 0, 0  S  1) and all its iterates Kn (x, s) are not square integrable. Establish that its resolvent kernel is given by R ( . \ ) = K (x, s) x, S, 1\ 1 _ A · Further show that K(x, s) has the discontinuous characteristic function { -1/2 <f>(x)= Xo for the characteristic value 1. (0 < x  1) (x=O) 
9 Fredholm theory We are now in a suitable position to discuss the theory originally developed by Fredholm in 1903 and by Schmidt in 1907 concerning the solution of linear integral equations. We shall approach this by first considering degenerate kernels for which it is possible to express the resolvent kernel in a closed analytical form. This will enable us to use the method introduced by Schmidt to treat a general square integrable kernel and to establish the Fredholm theorems. We conclude this chapter by obtaining the Fredholm solution for the case of a continuous kernel. This solution expresses the resolvent kernel as the ratio of two power series which are convergent for all values of the parameter A. 9.1 Degenerate kernels Let us examine the special case of a degenerate kernel having the separable form n K(x, s) = L Ui(X) Vi(S) (a  x  b, a  s  b) (1) i=l where Ul(X),. . . , un(x) and Vl(S), . . . , vn(s) are two sets of linearly independent square integrable functions. The least integer n for which a degenerate kernel can be expressed in the form (1) is called the rank of the kernel. Thus a degenerate kernel is often called a kernel of finite rank. Sometimes such a kernel (1) is referred to as a Pincherle-Goursat kernel. The corresponding Fredholm linear integral equation of the second kind takes the form n ! b cp(x) = f(x)+ '\1 Ui(X) a Vi(S) CP(S) ds. (2) 
Degenerate kernels 115 To solve this equation we introduce the unknown constants c; = f v;(s) <f>(s) ds (3) depending on the solution 4>(x), so that we have n 4>(x) = f(x) + A L CiUi(X). i=1 (4) On substituting back into (2) we obtain .t U;(X) [ Ci - i :;(S) { f(S) + A .t CjUj(S) } dS ] = 0 l-l a J-l and since the Ui(X) are linearly independent functions it follows that Ci - r :;(S) { f(S) + A .! CjUj(S) } ds = O. J a J = 1 (5) Writing a;j = f v;(s) Uj(s) ds (6) and f; = r v;(s)f (s) ds (7) we obtain the system of linear algebraic equations n Ci - A L aijCj = fi (i = 1, . . . , n) (8) j=l characterized by the determinant d(A) = det (1- AA) 1- Aal1 - Aa 12 - Aa ln - Aa 21 1 - Aa22 - Aa2n - (9) -Aa n l - Aa n 2 1 - Aa nn where A is the matrix with elements (aij) and 1 is the unit matrix. The determinant d(A) is a polynomial of degree n in A. If d(A)  0 for a given value of A, the system of equations (8) has the 
116 Fredholm theory unique solution given by Cramer's rule 1 n c; = d(A) jl d;j(A}h (i = 1, . . . , n) (10) where (d ij ) are the elements of the adjugate matrix D of 1- AA, that is the transposed matrix of its cofactors. Using (4) we see that Ann <f>(x) = f(x) + d(A) il jl ui(x)dij(A}h (11) from which it follows that the resolvent kernel is given by 1 n n _ R(x, slA) = d(A) ;l jl u;(x)dij(A)Vj(s). (12) Further, the corresponding homogeneous equation <f>(x) = Af K(x, s)<f>(s) ds (13) evidently has the unique solution 4>(x) = 0 if d(A)  0, that is if A is a regular value of K(x, s). Now suppose that d(A) = O. Then A must coincide with one of the characteristic values Av of the homogeneous equation (13). If r is the rank of the matrix 1- AA and p = n - r so that 1  P  n -1, the system of n homogeneous equations n Ci - A L aijCj = 0 j=l (i = 1, . . . , n) (14) has p linearly independent solutions {ck)}(k = 1, . . . , p). Thus the homogeneo,us equation (13) has the general solution p 4>(x) = L ak4>(k)(x) k=l (15) where p is called the rank (or sometimes the index to avoid confusion with the previous terminology) of the characteristic value A, and n 4>(k)(X) = A L Ck)Ui(X). i=l (16) 
Degenerate kernels 117 Since 1- AA * = (1- AA)* it follows that the characteristic value A of the adjoint homogeneous equation t/1(x) = if K(s, x) «fi(s) ds (17) possesses the same rank p, and there.fore has the same number of linearly independent solutions, as the homogeneous equation (13).. We see also that if A is a characteristic .value for which the inhomogeneous equation (2) has a particular square integrable solution ef>o(x), then its general square integable solution is 4>(x) = 4>o(x) + t Ctk4>(k)(X). k=l (18) Lastly we shall show that there exists a square integrable solu- tion ef>(x) of the inhomogeneous equation (2) for a given value of A if and only if f(x) is orthogonal to every square integrable solution of the adjoint homogeneous equation (17). If A is a regular value of K (x, s) then «/1 = 0 and the result is obvious. However if A is a characteristic value of K(x, s), and writing the inhomogeneous equation (2) in the operator form <f> = f + AKef>, (19) we see that (f, t/J) = (</>, «/1) - A (Kef>, «/1) = (ef>, «/1 - AK* «/1) = 0 (20) since the adjoint homogeneous equation (17) may be written «/1 = AK* t/J (21) It follows that a necessary condition for <f> to be a solution of (19) is that f should be orthogonal to every square integrable solution of (21). . This is also a sufficient condition. Thus if (f, t/J) = 0 for all the p linearly independent solutions of the adjoint homogeneous equa- tion, then the n equations (8) _can be reduced to r = n - p indepen- dent equations possessing r independent solutions where 1 =s.; r  n -1. We now discuss two examples of degenerate kernels. 
118 Fredholm theory Example 1. Consider the degenerate kernel of rank 2 K(x, s) = sin (x + s) = sin x cos s + cos x sin s (22) with 0  x  2'7T, 0  S  2'7T. We have Ul(X) = sin x, Vt(s) = cos s, U2(X) = cos x, V2(S) = sin s, (23) and thus r2 all= a22= Jo sin s cos s ds = 0, i 2 a12= 0 cos 2 s ds = 7T, (24) i 2 a21 = 0 sin 2 s ds = 7T. Hence 1 d(A) = -'7TA -'7TA = 1- '7T 2 A 2 1 (25) yielding the characteristic values Al = 1/1T, A 2 = -1/1T and the corre- sponding orthonormalized characteristic functions 4>l(X) = / (sin x +cos x), 4>2(X) = / (sin x -cos x) (26) V21T v2'7T of the homogeneous Fredholm equation of the second kind. Sym- metric real kernels such as (22), satisfying K(s, x) = K(x, s) always have real characteristic values as we shall show in section 10.2. The elements of the adjugate matrix Dare d ll = d 22 = 1, d 12 = d 21 = 1TA (27) giving for the resolvent kernel ( . _ sin (x + s) + '7TA cos (x - s) R x, s, A) - 1 _ '7T 2 A 2 (28) 
Degenerate kernels 119 and for the unique solution to the Fredholm equation of the second kind (2) when A =fi :f: 1/ 'Tr: <f>(x) = I(x) + 1- z A Z {(It + 1TAfz) sin x + (1TA/ I + fz) cos x} (29) where r2 r2 II =.10 I(s) cos s ds, Iz = Jo I(s) sin s ds, (30) If f(s) = 1 we have {I = f2 = 0 and thus ({, 4>1) = ({, 4>2) = 0 so that f(x) is orthogonal to the solutions 4>1(X), 4>2(X) of the adjoint homogeneous equation (17) for the characteristic values Al = 1/'Tr, A2 = -1/ 'Tr. Hence the inhomogeneous equation (2) has the general solutions 4>(x) = 1 + a4>l(x), 1 + a4>2(x) corresponding to the charac- teristic values A b A2 respectively. If f(s) = s we have fl = 0, f2 = -2'Tr and thus (f, 4>1) = ({, 4>2) = - .J2'Tr . Since the orthogonality condition is not satisfied in this case the inhomogeneous equation (2) has no solutions for A = :i: 1/ 'Tr. Example 2. Next we consider the degenerate kernel of rank 2 K(x,s)=x-s (Oxl,Osl) (31) for which Ul(X) = x, U2(X) = -1, Vl(S) = 1, V2(S) = s (32) so that i 1 1 all = -a22 = s ds =- o 2 r 1 r 1 1 a12=- Jods=-1,azI= JoszdS= 3 ' (33) Then d(A) = i A 1 :!A = 1 + lz A z and so the characteristic values given by d(A) = 0 are the pure imaginary numbers :f:i2J3. Skew symmetric real kernels such as (34) 
120 Fredholm theory (31), satisfying K(s, x) = - K(x, s), always have pure imaginary characteristic values. The elements of the adjugate matrix D of 1- AA are d 11 = 1 +!A, d 22 = 1-!'\, d 12 = -A, d 21 = lA (35) and so the resolvent kernel is R ( · _ (l+!A)x-(l-!A)s-Axs-lA (36) x, s , A) - 1 + l2 A 2 giving for the unique solution of (2) <f>(x) == f(x) + I A 2 {(I +tA)x/t - (I-!A) f2 - hfz -tAfl} +12 (37) valid for all values of A  :i: i2J3 and in particular for all real A. 9.2 Approximation by degenerate kernels In this section our aim is to show that we can approximate a square integrable kernel as closely as we please, in the mean square sense, by a degenerate kernel of sufficiently high rank. We suppose that K(x, s) is a square integrable kernel satisfying fIK(x, s) 1 2 ds <00 DK(X,SWdX<oo (asb), II KII = r D K(x, s)1 2 dx ds <00, (axb), (38) and we shall prove that given any e > 0 there exists a degenerate kernel P(x, s) such that IIK-Plb E. (39) Let Vl(S), V2(S),. .., Vi(S)". . . be a complete orthonormal system of square integrable functions and let fK(X, S)Vi(S) ds = Uj(x). a (40) 
Fredholm theorems 121 Using Parseval's formula we see that 00 I b jl ' Uj(X) 1 2 = a' K(x, S) 1 2 ds (41) and so il fi U;(X) 1 2 dx = II KII (42) from which it follows that there exists a N such that for n > N 00 I b j=+l a IUj(xWdx<e 2 . (43) Now, as a consequence of the completeness of the orthonormal system {Vi(S)}, we have fb n 00 L IK(x, s)- jl U;(X) Vi(SW ds = j=+l IUj(xW and so i b i b n 2 a a K(X,S)-il Uj(X) Vj(S) dxds<e 2 , (44) Hence taking n P(X, S) = L Ui(X) Vi(S) i=l (45) where n > N, we obtain the desired result (39). 9.3 Fredholm theorems We consider the Fredholm integral equation of the second kind <I> (X) = f(x) + A r K(x, s)<I>(s) ds (46) where K(x, s) is a square integrable kernel and A is chosen so that I A 1< w where w is a positive number. From the discussion given in the previous section we know that there exists a square integrable 
122 Fredholm theory degenerate kernel (45) such that K(x, s) = P(x, s) + Q(x, s) (47) and 1 IIQI\z<-. w (48) Such a dissection of the kernel K(x, s) was introduced by Schmidt in 1907 and it enables us to write the integral equation (46) as cf>(x) = fl(X;..\) +..\ r Q(x, s)cf>(s) ds (49) where fl(X;..\) = f(x)+..\ fp(X, s)cf>(s) ds. (50) Since IAIIIQI\z< 1 it follows that the Neumann series for the resolvent kernel RQ(x, s; A) of Q(x, s) is convergent, so that the solution of the integral equation (49) can be expressed in the form cf>(x ). fl(X;..\) +..\ f Ro(x, s; ..\)fl(S; ..\) ds. (51) This may be rewritten cf>(x) = fz(x; ..\) +..\ r T(x, s; ..\)cf>(s) ds (52) where fz(x; ..\) = f(x) +..\ f Ro(x, s; ..\)f(s) ds (53) and T(x, s; ..\) = P(x, s) +..\ f Ro(x, t; ..\)P(t, s) dt. (54) Here T(x, s; A) is a degenerate kernel since it can be expressed in the form n T (x, s; A) = L Y i (x) Vi (s ) i=l (55) 
Fredholm theorems 123 where Yi(X) = Ui(X) +..\ r Ro(x, t; ..\)Ui(t) dt. (56) Thus we have converted our original integral equation (46) posses- sing a general square integrable kernel K(x, s) into an integral equation (52) with a degenerate kernel T(x, s; A). In particular we note that if our original equation is homogeneous so that f(x) = 0, the new equation with degenerate kernel T(x, s; A) is also homogeneous since f2(X; A) = O. This means that the results obtained in section 9.1 on degenerate kernels can now be applied to the present case of a general kernel. Using the expression (12) for the resolvent kernel corresponding to a degenerate kernel we obtain A i b n n _ cf>(x) = f2(x, ..\)+ d(..\) a il jl Yi (x)dij(..\)Vj (s)f2(s; ..\) ds (57) w here d(A) = det (1- AA) and the matrix A has elements aij = J Vi(S)Yj(s) ds while the dij(A) are the elements of the adjugate matrix D of I-AA. Hence the resolvent kernel of K(x, s) can be seen to have the form 1 n n R(x, s;..\) = Ro(x, s; ..\)+ d(..\) il jl Yi(x)dij(..\) zj(s) (58) where Zj(s) = Vj(s) + A r vj(t) Ro(t, s; ..\) dt. (59) The resolvent kernel R(x, s; A) is therefore an analytic function of A for IA 1< w except for poles occurring at the zeros of d(A). For a given w these poles are finite in number and so as w  00 it is clear that the number of poles becomes at most denumerably infinite with no finite points of accumulation. This means that R (x, s; A) is a meromorphic function of A whose poles coincide with the charac- teristic values of K. 
124 Fredholm theory Our results can be expressed in the form of a set of theorems known as the Fredholm alternatives: Theorem 1. If K is an integral. operator with a square integrable kernel K(x, s), and cJ> = f + AK cf.> (60) is a Fredholm integral equation of the second kind, then A is either a regular value or a characteristic value of K. If A is a regular value of K, the integral equation (60) has the unique square integrable solution cf.> = f + ARf (61) for any square integrable function f, where the resolvent operator R with kernel R(x, s; A) is a meromorphic function of A. The characteristic values of K are at most denumerably' infinite with no finite accumulation point. For each characteristic value A of rank p, the homogeneous equation cf.> = AK cf.> (62) has p linearly independent solutions cf.> (k) (k = 1, . . . , p). Theorem 2. If A is a characteristic value of K then A is a characteristic value of K*. The number of linearly independent solutions of (62) and the adjoint homogeneous equation tf/= AK* tf/ (63) are the same. Theorem 3. If A is a characteristic value of K and f is a square integrable function, the integral equation (60) has a square integra- ble. solution if and only if f is orthogonal to every square integrable solution of the adjoint homogeneous equation (63). Then the gen- eral square integrable solution of (60) is given by p cf.> = cf.>o  L akcf.> (k) k=l (64) where CPo is a particular square integrable solution of (60). 
Fredholm theorems 125 9.3.1 Fredholm theorems for completely continuODs operators The Fredholm theorems stated above were generalized by Hilbert to hold for completely continuous operators. As for integral operators, the demonstration depends upon the introduction of a degenerate, that is finite dimensional, linear operator P. Suppose K is a bounded linear operator and let Vb V2, . . . , Vi, . . . be a complete orthonormal system in abstract Hilbert space H. Then the kernel matrix of K has elements k .. = (K v. v. ) IJ I' J (65) Putting K Vi = Ui (i = 1, 2, . . .) and expanding Ui in the form (66) 00 _  (i) Ui -  x j Vj j=l (67) where 00 IIudl 2 = L IxJi}12 < 00, j=l (68) we have 00 00 00 00 00 00 L L Ik ij l 2 = L L I(ui, vj)1 2 = L L IxJi}12 i=l j=l i=l j=l i=l j=l 00 = L Iludl 2 . i=l (69) Let us now assu.me that 00 00 L L Ik ij l 2 <00 i=lj=l (70) so that K is completely continuous as shown in section 7.3. Then it follows from (69) that, given any e > 0, there exists N such that if n>N 00 L Il u iI1 2 <e 2 . i=n+l (71) 
126 Fredholm theory Next we introduce a degenerate operator P such that PVi = Ui =0 (i = 1, . . . , n) (i > n). (72) Suppose that v is any element in Hand 00 v = L YiVi i=l (73) where 00 I/vII2 = L I yd 2 < 00. i=l (74) Then we have 00 II(K - p)vll = L Yi(K - P)vdl i=l 00  L IYd II(K - P)vdl i=l 00 = L !Yi !lIud! i=n+l   CJ+1 lyd 2 )CJ+l II U ill 2 ) (75) < Ilvlle using the triangle inequality and Cauchy's inequality together with (71) and (74). Thf by choosing n sufficiently large, we can ensure that IlK - PII < e /provided the kernel matrix of K satisfies (70) so that K is completely continuous. This result allows us to approxi- mate a completely continuous operator K as closely as we please by a suitable degenerate operator P. Setting Q = K - P and choosing n so large that IAIIIQII< 1, the resolvent operator R Q of Q is bounded and given by a convergent Neumann series, as shown in section 8.5, which enables us to carry through a similar analysis to that pre- sented in section 9.3. 9.4 Fredholm formulae for continuous kernels To conclude this chapter we shall derive the Fredholm solution of the integral equation cf>(x) = f(x) + A r K(x, s)cf>(s) ds (a  x  b) (76) 
Fredholm formulae for continuous kernels 127 assuming that the kernel K{x, s) is continuous in the square domain a  x  b, a  s  b, that f{x) is continuous in the interval a  x  b, and that the integration is performed in the Riemann sense. We first note the solution of (76) must be continuous for we have { f b } 1/2 Icf>(x) - cf>(x')J.s; If(x) - f(x')IIA J t JK(x, s) - K(x', sW ds using the triangle and Schwarz inequalities, so that the continuity of K{x, s) and f{x) ensures the continuity of the square integrable solution cf.>{x). We shall see that the solution obtained by Fredholm involves a resolvent kernel given by the ratio of two power series in A which are convergent for all A. The method introduced by Fredholm in 1903 to establish his formulae treats the integral equation (76) as the limiting form of a finite system of linear algebraic equations. These equations are obtained by choosing a net of n equal sub-intervals having length 8n = (b - a)/n given by a = Xo < Xl < X2 < . . . < X r < . . . < X n = b (77) where X r = a + r5n. Approximating the Riemann integral on the right-hand side of (76) by the finite sum n 8n L K{xr, xs)cf.>{xs) s=l where x = X r , we may replace the integral equation by the system of n algebraic equations n cf.>{Xr) - A8n L K{xr, xs)cf.>{xs) = f{x r ). s=l (78) Provided the matrix 1- A8nK= 1- A8nK{Xt, Xl) - A8 n K {Xt, X2) -A8 n K{X2' Xl) 1- A8 n K{X2' X2) . .. -A8 n K{xt, X n ) -A8 n K{X2' X n ) - A8 n K {xm Xl) -A5 n K{x n , X2) 1- A8nK{xn, X n ) (79) 
128 Fredholm theory has the non-vanishing determinant dn(A) = det(l- A5nK), the system of equations has the unique solution 1 n cfJ(xr) = dn(A) Sl Dn(x" xs)f(xs) (80) where Dn(x" xs) is the r, s element of the adjugate matrix of 1- A8nK, that is Dn(xr, xs) is the cofactor of the element involving K(xs, x r ) in 1- A8nK. Expanding dn(A) in powers of A we obtain A8n A 28 (-l)n A n8: d n (A)=I- lT S I +2!S2-".+ n! Sn (81) where n 51 = L K(x r , x r ), r=l n K(x rt , x rt ) K(x rt , x r2 ) 5 2 = L K (X r2 , X rt ) K (X r2 , X r2 ) , rJ, r 2=1 and generally K (x rt , x rt ) K ( X rt , Xr'2) K(x rt , x rm ) K (x r2 , x rt ) K(x r2 , x r2 ) K(x r2 , x rm ) n 5m= L (82) rJ,r2,...,rm = 1 K(x rm , X rt ) K(x rm , X r2 ) K(x rm , X rm ) Now letting n  00 and 8n  0 we obtain dn(A)  d(A) where d(A) = 1- A f b K(xl, Xl) dXI + A: f b f b K(Xl, Xl) K(Xl, X2) dXI dX2 a 2. a a K(X2' Xl) K(X2, X2) A 3 f b f b f b I K(Xb Xl) K(Xb X2) K(Xb X3) - 3! a a a K(X2' Xl) K(X2' X2) K(X2, X3) dXI dX2 dX3 K(X3, Xl) K(X3, X2) K(X3, X3) +. . . (83) 
Fredholm formulae for continuous kernels 129 Introducing the notation K ( Xh X2, · · . , Xm ) = SI, S2, . . · , Sm K(Xh S1) K(Xh S2) K(X2, S1) K(X2, 82) K(Xh Sm) K(X2, Sm) K(xm, S1) K(xm, S2) K(x m , Sm) (84) we see that 00 d(A)= L dmAm m=O (85) where do = 1 and _(-l)m J b J b J b K( Xt,X2,...,Xm )d d d (86) d m - . . . X 1 X2... Xm. m! a a a Xl, X2, · · · , X m d(A) is known as the Fredholm determinant. Now, Dn(xr, xs) being the cofactor of the (s, r) element in dn(A), we have for r S [ f K(x" xs) K (x,., X St ) Dn (x" xs) = ABn K(x" xs) - ..\8 n K( ) K(x st , X St ) St=1 XS1,X S A 2 8 2 n K(x r , xs) K (x r , X St ) K(x" X S2 -. . .] + n L K(x st , xs) K(x st , X St ) K(xst, X S2 ) 2! Sts2=1 K(x S2 , xs) K (X S2 , X St ) K (X S2 , X S2 ) (87) and so, letting n  00 and X r  x, Xs  s, we see that 81 Dn(x" xs)  AD(x, s; A) where D(x,s;..\)=K(x,s)-..\ r b K(x,s) K(X,Sl) dS 1 J a K(st, s) K(sl, S1) A 2 1 b I b K(x, s) K(x, S1) K(x, S2) + 2! a a K(st, s) K(st, Sl) K(st, S2) dS 1 ds 2 -. · · K(S2, s) K(S2, S1) K(S2, S2) (88) 
130 Fredholm theory However for r = s, Dn has a similar expansion to dn(A) and in fact it can be readily verified that Dn(x" x r )  d(A) as n  00. Hence in the limit as n  00 we obtain <f>(x) = f(x) + A r R(x, s; A)f(s) ds (89) where the resolvent kernel R(x, s; A) is given by D (x, s; A) R(x, s; A) = d(A) (90) with 00 D(x, s; A) = L Dm(x, S)A m m=O (91) and Do(x, s) = K(x, s) while generally ( l) m f b f b i b ( ) - X, Sl, S2, . . . , Sm Dm (x, s) = . . . K dS l dS 2 . . . ds m . m! a a a S,Sl, S 2,...,Sm (92) D(x,s; A) is known as the first Fredholm minor. To establish the convergence of the series for d(A) and D(x, s; A) we employ Hadamard's inequality " " Idet AI2  n L tars 1 2 r= 1 s = 1 (93) where A is a n x n matrix whose (r, s) element is the complex number a rs . This inequality has a simple interpretation in three-dimensional real Euclidean space. If 81 = (all, a12, a13), 82 = (a2l, a22, a23), 83 = (a3l, a32, a33) are three vectors determining the sides of a parallelepiped, its volume is given by the scalar triple product all a12 a13 .81 · (82 X 83) = a2l a22 a23. a3l a32 a33 This is less than or equal to the volume of the corresponding rectangular parallelepiped which is 118111118211118311 and thus (93) follows for n = 3 and real numbers a rs . If larsl  M for all r, s we have IdetAl2 n"M 2n . (94) Since the kernel K(x, s) is continuous it is also bounded and so there 
Fredholm formulae for continuous kernels 131 exists a positive number M such that IK(x, s)IM. Hence mm/2M m (b-a)m \dml  , = am (95) m. and so 00 00 Id(A)I L IdmllAlm L amlAlm. m=O m=O (96) Now ( 1 + ) m/2 M(b - a) a m +1 m am = (m + 1)1/2  0 (97) as moo since (l+l/m)me and so L;=o amlAlm is convergent for all A. Hence (96) implies the absolute convergence of d(A) for all A. Also, by the inequality (94), we have (m + 1)(m+1)/2M m + 1 (b - a)m IDm(x, s)1  , = b m (98) m. and so 00 00 ID(x, s; A)I  L IDm(x, s)IIAlm  L bmlAlm. (99) m=O m=O Now b m (m + 1) 1/2 ( 1 ) m/2 -= 1+- M(b-a)O b m - 1 m m (100) as m  00, from which it follows that I;=o b m IAlm is convergent so that D(x, s; A) is absolutely and uniformly convergent for all A. Our next step is to verify that the Fredholm solution (89) is indeed correct by showing that the resolvent kernel (90) satisfies the resolvent equation. To this end we note that Dm(x, s) = dmK(x, s) + Qm(x, s) (101) where (-l)m i b i b Q m (x, s) = , . . . m. a a o K(x, S1) K(S1, s) K(st, S1) K(x, sm) K(st, sm) dS 1 . . . dS m (102) K(sm, s) K(sm, S1) K(sm, sm) 
132 Fredholm theory If we now expand in terms of the minors of the first column we obtain Om (X, s) = (-1  m f (-1 )i 1 b. . . i b J x, S h.. · · , Si- h Si + h.. · , Sm ) m. i = 1 a a A\ S h S2, . . . , St, Si + h . . . , Sm XK(Sh S) d'St . . . dSm = (_l)m-l J b.. . 1 b f b K ( x, Sh... ,.Sm-t ) K(t, S) ds 1 ... dS m - 1 dt (m - 1)! a a a t, s 1, . . · , Sm-1 = fDm-t(X, t)K(t, s) dt (103) which yields the recurrence relation Dm(x, s) = dmK(x, s) + r Dm-t(x, t)K(t, s) dt. (104) Hence D(x, s; A) = d(A)K(x, s)+ AfD(X, t; A)K(t, s) dt. (105) Similarly by expanding in terms of the minors of the first row of (102), we can show that Dm(x, s) = dmK(x,s) + l(X, t)Dm-t(t, s) dt (106) and so we have also D(x, s; A) = d(A)K(x, s)+ AJ:K(X' t)D(t, s; A) dt. (107) It follows that the resolvent kernel (90) satisfies the resolvent equation R(x, s; A)- K(x, s) = Ai(X, t; A)K(t, s) dt = AfK(X' t)R(t, s;) dt (108) and thus the Fredholm solution is verified. Since D ( ) ( -l)m-l i b 1 b ( x,St,...,Sm-t ) d m -1 X, S = ( _ 1 ) ' · · .' K ds t . .. Sm-1 m . a a S, St, · · · , Sm-1 
Fredholm formulae for continuous kernels 133 it follows that f b D ( ) d = (-1) m-1 f b f b f b K( S, S 1 , · . . , Sm -1 ) d d m-1 S, S S ,. · · S S1 a (m - 1). a a a S, S h . . . , Sm-1 . . . dS m - 1 and so we obtain the relation 1 f b d m = -- D m - 1 (s, s) ds. m a (109) Also, using (109), we have that 00 d'(A) = L mdmA m-1 m=1 00 f b =- mo Am a Dm(s,s)ds = -l(S' s; A) ds (110) and so d'(A) f b d(A) = - aR(s, s; A) ds = - no A n ln+1(S, s) ds (111) employing the Neumann expansion (8.22) for the resolvent kernel. The trace of the kernel K(x, s) is defined to be the quantity Kl = fK(S, s) ds (112) while the trace of the iterate Kn (x, s) is given by Kn = r · · · r r K(s, t1)K(tt, tz). · .K(tn-t, s) dtl. · · dt n - 1 ds. (113) Hence we obtain the formula d'(A) 00 d(A) = - no K n +1 A n (114) 
134 Fredholm theory from which it follows that the radius of convergence of the power series on the right-hand side is IA11 where Al is the characteristic value of the kernel K(x, s) having the least absolute magnitude. Example. To illustrate the application of the Fredholm recurrence formulae (104) and (109) obtained above we consider the kernel K(x,s)=x+s (Oxl,Osl). (115) We have do = 1, Do(x, s) = K(x, s) = x + s, d 1 = - Lo(S, s) ds = - fs ds =-1 and D1(x, s) = -K(x, s)+ Lo(X, t)K(t, s) dt =-(x+s)+ fo<x+t)(t+S)dt 1 1 = 3 - 2 (x + s) + xs. Hence 1 f 1 d 2 = - 2 Jo D1(s, s) ds. 1 1 1 1 2 = -- (3- s + s ) ds 2 0 1 = - 12 . Further D 2 (x, s) = - A K(x, s)+ Ll(X, t)K(t, s) dt = 0 
Problems 135 from which it follows that d 3 = O. Repeated application of the recurrence relations leads to Dm(x,s)=.O (m2), dm=O (m3). Thus d(A)=l-A- lz A Z and D(x, s; A) = x + s +{t-!(x + s) + XS}A. Problems 1. Obtain the resolvent kernels for the symmetric kernels (i) K(x, s) = 1 + xs (O:S:;; x:S:;; 1, O:s:;; s:S:;; 1), (ii) K (x, s) = X z + s 2 (0 :s:;; x :s:;; 1, 0 :s:;; s :s:;; 1), (iii) K (x, s) = xs + x 2 S 2 ( - 1 :s:;; x  1, - 1 :s:;; s :s:;; 1). Verify that their characteristic values are real numbers. 2. Obtain the resolvent kernels for the skew-symmetric kernels (i) K(x, s) = sin (x - s) (O:S:;; x :s:;; 27T, O:s:;; s :s:;; 27T), (ii) K (x, s) = X 2 - S 2 (0 :s:;; x :s:;; 1, 0 :s:;; s :s:;; 1), (iii) K (x, s) = x 2 S - xs 2 (O:S:;; X :s:;; 1, O:s:;; s :s:;; 1). Verify that their characteristic values are pure imaginary num- bers. 3. Show that the non symmetric kernels (i) K (x, s) = sin x cos s (0 :s:;; x :s:;; 7T, O:s:;; s :s:;; 7T), (ii) K (x, s) = sin x sin 2 s (0 :s:;; x :s:;; 7T, O:s:;; s :s:;; 7T), have no characteristic values. 4. Use the theory on degenerate kernels given in section 9.1 to solve the Fredholm integral equations of the second kind possessing the degenerate kernels ( i) K (x, s) = x + s (0 :s:;; x :s:;; 1, O:s:;; s :s:;; 1), (ii) K (x, s) = cos (x + s) (0 :s:;; x :s:;; 2 7T, 0 :s:;; s :s:;; 2 7T) . 5. Use the Fredholm formulae given in section 9.4 to solve the Fredholm integral equations of the second kind possessing the kernels (i) K(x, s) = sin (x + s) (O:S:;; x:S:;; 27T, O:s:;; s:S:;; 27T), (ii) K (x, s) = x - s (0 :s:;; x :s:;; 1, 0 :s:;; s :s:;; 1), (iii) K (x, s) = 1 - 3 xs (0 :s:;; x :s:;; 1, 0 :s:;; s :s:;; 1). Verify that your solutions to (i) and (ii) agree with examples 1 and 2 of section 9.1. 
10 Hilbert-Schmidt theory In this final chapter we shall direct our attention to self-adjoint or Hermitian integral operators K satisfying (K 4>, tfJ) = (4), KtfJ) (1) for all 4>, tfJ belonging to the Hilbert space of L 2 functions. Our objective will be to develop the theory originated by Hilbert and Schmidt, in which the resolvent kernel R(x, s; A) was expanded in terms of the characteristic functions 4>v (x) and characteristic values Av of the Hermitian kernel K(x, s). 10.1 Hermitian kernels The original theory of Hilbert and Schmidt dealt with real symmet- ric kernels satisfying the condition K(s, x) = K(x, s) but we shall be studying the more general case of Hermitian kernels for which K(s, x) = K(x, s) (2) where K(x, s) is not necessarily real. Then the integral operator K= r K(x, s) ds (3) is Hermitian since it satisfies the relation (1). We see from (1) that (KtfJ, tfJ)=(tfJ,KtfJ)=(KtfJ, tfJ) and so (KtfJ, tfJ) is real for all L 2 functions tfJ. A Hermitian kernel which is square integrable is called a Hilbert- Schmidt kernel. 10.2 Spectrum of a Hilbert-Schmidt kernel Let us begin by investigating the properties of the set of characteris- tic values or spectrum of a Hermitian square integrable kernel, that is a Hilbert-Schmidt kernel K(x, s). 
Spectrum of a Hilbert-Schmidt kernel 137 We shall show first that every non-null operator K possessing a Hilbert-Schmidt kernel has at least one characteristic value A1. The method of proof is due to Kneser. Since K is Hermitian it follows from (1) that K n is also Hermi- tian and hence, using the definition (9.112) of the trace of an integral operator with kernel K(x, s), we have K2n = trace (K n K n ) = f f Kn(x, s)Kn(s, x) ds dx = r r Kn(x, s) Kn(x, s) ds dx =IIKnll (4) and so K2n  O. Also K2n< 00 SInce IIK n Il 2 :s:;; IIKII and K(x, s) IS square integrable.. Using Schwarz's inequality for double integrals we have Kn = {trace (K n - 1 Kn+l)}2 = {f f K n - 1 (x, s)K n +1(s, x) ds dXY  {f fI Kn - 1 (X, sW ds dxHf fI Kn +1(x, sW ds dX} (n  2). =K2n-2 K 2n+2 (5) Now IIK112> 0 since K is non-null and so K2>0. Further K4>O, for if we were to suppose that K4 = 0 it would imply that K 2 (x, s) = 0 almost everywhere and in particular we would have K 2 (s, s) = 0 so that K2 = O. Since K2 > 0 and K4 > 0 it follows from (5) that K2n > 0 for all nand K2n+2 K2n K4  ...-. K2n K2n-2 K2 (6) Putting Un = K2n IAl z n-1 we see that U n +1 = IAl z K2n+2  IAI2 K4 Un K2n Kz and hence I:=lUn is divergent if IAI> .J K2/K4. But I:=oK n +1 An is convergent for sufficiently small IAI and indeed we showed at the end of section 9.4 that the radius of convergence of this series is IAII 
138 Hilbert-Schmidt theory where Al is the characteristic value of K having least absolute magnitude. Hence L:=l Un is also convergent for IAI<IAtl and si nce we already know that this series is divergent for IA 1 >  K2/ K4, it follows that there exists at least one characteristic value A 1 and that IA11:s:;;  K2/ K4. Next we shall show that all the characteristic values of a Hermi- tian kernel K(x, s) are real. Thus let AKcP = cP where cP is a non-null characteristic function associated with the characteristic value A. Then (KcP, cP) = (A -lcP, cP) = A -l(cP, cP) and since (KcP, cP) is real, and (cP, cP) is real and non-zero, it follows that A is real. We also see that the characteristic functions cP1, cP2 of a Hermi- tian kernel K(x, s) for different characteristic values AI, A2 are orthogonal. For we have (cP1, cP2) = (A 1 K cPt, cP2) = Al (K cP1, cP2) and (cP1, cP2) = (cP1, A 2 K cP2) = A2(K cP1, cP2) since K is Hermitian and A2 is real. This yields (A}l- A2 1 )(cP1, cP2) = 0 and as Al '\2 it follows that (cPt, cP2)= o. We proved above that every non-null Hermitian kernel K(x, s) possesses at least one characteristic value Al and associated nor- malized characteristic function cPl. Now consider the Hermitian kernel K (2) ( ) = K( ) _ cP1(X)cP1(S) X, S x, s AI. (7) If K(2)(X, s) is non-null it also must have at least one characteristic value A2 and associated characteristic function cP2. Since f b f b cP (x) K(2)(x, S)«>1(S) ds = K(x, S)«>1(S) ds -  = 0 a a 1 
Expansion theorems 139 we see that CPl cannot be a characteristic function of K(2)(X, s) and so CPl  CP2 although Al may equal A2. Continuing this procedure we shall find that either there exists an n such that K(n+1)(x, s) = K(x, s)- ! <f>,,(x) <f>,,(s) ==0 v=l Av in which case we have the bilinear formula K(x, s) = ! <f>,,(x)<f>,,(s) . (8) v=l Av or there exists an infinite number of characteristic values Av and their associated characteristic functions CPv. By Bessel's inequality (6.35) applied to K(x, s) we have fIK(x, sW ds;;. "1 f K(x, s)<f>,,(s) ds 2 = ! 1<f>,,(;W (9) v=l Av so that, remembering that the CPv are normalized, n 1 IIKII ;;. "1 A; " (10) Hence if IAvl < c for v = 1, . . . , n then n:S:;; c211KII and thus there are only a finite number of characteristic values in the interval (-c, c). Consequently the spectrum of characteristic values is countable and has no finite point of accumulation. We shall suppose that IAll:s:;; IA21:s:;; . . . and that if Av is a charac- teristic value of rank p it will be repeated p times in this series and that the associated p linearly independent characteristic functions are orthogonal and normalized. We shall refer to these characteristic functions CPl (x), CP2(X), . . . and their associated characteristic values AI, A2,. . . as a full orthonormal system. We note however that this system of functions need not be complete. 10.3 Expansion theorems Although the series Sn(x, s) = ! <f>,,(x)<f>,,(s) v=l Av (11) 
140 Hilbert-Schmidt theory need not converge as n  00, we can show that it is mean square convergent to K(x, s), that is ! f fIK(X, s) - Sn(x, S W dx ds = O. (12) Thus, by the Riesz form of the Riesz-Fischer theorem stated in section 6.3.4, there exists a L 2 Hermitian kernel Q(x, s) such that ! fIQ(X, s)- Sn(x, sW ds = 0 (axb) (13) and f b cPv(x) t Q(x, s)<I>v(s) ds = Av (v= 1,2,.. .), (14) sInce l l<1>vW  fIK(x, sWds<oo using (9) and that K (x, s) is square integrable. Hence, setting (15) P(x, s) = K(x, s) - Q(x, s) (16) we see that f P(x, s)<I>v(s) ds = 0 (v=1,2,...) (17) since Q(x, s) has the same Fourier coefficients as K(x, s). In order to establish (12) we need to show that P(x, s) is null. Let us suppose that, on the contrary, P(x, s) is non-null. Then there exists a normalized characteristic function cPo( x) with characteristic value Ao such that AOPcPO = cPo (18) which gives (cPo, cPv) = Ao(PcPo, cPv) = Ao( cPo, PcPv) = 0 since P is Hermitian and using (17). Now for any positive integer n we have f Q(x, s)<I>o(s) ds = f{ Q(x, s)- Sn(x, s) } <l>o(s) ds (20) (v=1,2,...) (19) 
Expansion theorems 141 since (c/>o, c/>v) = 0 for v = 1, 2, . . . and so, given any e > 0 there exists N such that for n > N IQc/>oI2 fIQ(X, s) - Sn(x, sW ds < e (21) using Schwarz's inequality and (13). As QC/>o is independent of n it follows that QC/>o = 0 and so AoK <Po = AoPc/>o = <Po (22) which means that C/>O is a characteristic function of K but is orthogonal to all the c/>v. This provides a contradiction since the C/>V form a full orthonormal system. Hence P(x, s) must be a null kernel and so Q(x, s) = K(x, s) (23) which means that (13) holds with Q replaced by K, yielding the result (12) we wished to establish. 10.3.1 Hilbert-Schmidt theorem If f(x) is a square integrable function given by f(x) = f K(x, s)g(s) ds where K(x, s) is a Hilbert-Schmidt kernel and g(s) is a square integrable function, then the Hilbert-Schmidt theorem states that (24) 00 f(x) = L avc/>v(x) v=l (25) where 1 a v = (f, c/>v) = Av (g, c/>v) (26) and c/>v(x), Av (v'= 1, 2, . ..) are the characteristic functions and values of K(x, s). The Hilbert-Schmidt series (25) for f(x) converges absolutely and uniformly if, in addition, the kernel satisfies fIK(x, s W ds < A 2 (27) where A is a constant. 
142 Hilbert-Schmidt theory It is not difficult to establish (26) for we have a v = (f, c/>v) f b b = a cf>,,(X) dx L K(x, s)g(s) ds b b = L g(s) ds 1 K(s, x) cf>,,(x) dx = (g, K<pv) 1 = A" (g, cf>,,). Hence f b n <Pv (x) f b f(x) = a {K(x, s) - Sn(x, s)}g(s) ds + "1 A" a cf>,,(s) g(s) ds f b n = a {K(x, s) - Sn(x, s)}g(s) ds + "1 a"cf>,,(x) (28) and so, employing Schwarz's inequality, we obtain f(x)- "1 a"cf>,,(x) 2  fIK(x, S)-Sn(X, sW ds flg(sW ds. (29) Since the right-hand side of (29) can be made as small as we wish by letting n become sufficiently large, the result (25) is proved. Setting b v = (g, cf>v) (30) we have { 00 } 2 { 00 b } 2 "-+1 1a"cf>,,(x)1 = "=+1 A: cf>,,(x) L=+1Ib,,12}L=+1 Icf>,,W } (31) using Cauchy's inequality (6.9). Now g(x) is square integrable and so L':=1Ib v I 2 <oo. Hence, given any £ >0 there exists N such that for n>N 00 L Ib v l 2 < £2. v=n+l (32) 
Expansion theorems 143 Also (15) with the condition (27) gives "=+1 Icp,,W  fIK(X, sW ds < A 2. (33) Thus for n > N we have 00 L lav<pv(x)1 < eA v=n+l (34) from which it follows that the Hilbert-Schmidt series (25) is abso- lutely and uniformly convergent. 10.3.2 Hllbert's formula As a corollary to the Hilbert-Schmidt theorem proved above, we have that if g(x) and h(x) are both square integrable functions then 00 1 (Kg, h) = "1 A" (g, cp")(cp,,, h) (35) which is Hilbert's formula. It follows from the Hilbert-Schmidt series (25) for f = Kg by taking the inner product with h. This can be done term by term on the right-hand side since the series is uniformly convergent. For h = g Hilbert's formula reduces to 00 1 (Kg, g) = "1 A" I(g, cp"W. (36) 10.3.3 Expansion theorem for iterated kernels Since the iterated kernel Kn(x, s) for n;=:: 2 is given by . Kn(x, s) = f K(x, t)K n - 1 (t, s) dt and so is of the form (24) with g = Kn-b it follows from the Hilbert-Schmidt theorem that if K(x, s) is a Hilbert-Schmidt kernel then 00 Kn(x, s) = L av(s)q,v(x) v=l where f b 1 av(s) = Kn(x, s)c/>v(x) dx = Pi c/>v(s). a Av 
144 Hilbert-Schmidt theory Hence Kn(x, s) = vt 1 CPv(xlt(s) (n  2) (37) where the series is absolutely and uniformly convergent if the kernel K(x, s) satisfies the condition (27). We deduce that Kn=trace(K n )= fKn(X,X)dX 00 1 = L --;; (n,2) v=l Av (38) and in particular we have 00 1 K2=IIKI@= L 2<00. v=1 A v (39) 10.4 Solution of Fredholm equation of second kind Consider the Fredholm equation of the second kind cp(x) - f(x) = A r K(x. s)cp(s) ds (40) where K(x, s) is a Hilbert-Schmidt kernel satisfying the condition (27), the function f(x) is square integrable and A is a regular value. If the solution c/>(x) of (40) is square integrable then the Hilbert- Schmidt theorem gives 00 c/>(x) - f(x) = L avc/>v(x) v=l where a v . (cf> - f, c/>v) = (c/>, c/>v) - (f, c/>v) and also Hence A a v = A(K cp, CPv) = Av (cp, CPv) (cp, CPv) = AvA A (f, CPv) 
Solution of Fredholm equation of second kind 145 and A a" = A" _ A (f, cf>,,). Consequently the solution can be expressed as the absolutely and uniformly convergent series cf>(x) = f(x) + A f (f, cf>,,) cf>" (x ) v=l Av-A = f(x) + A "1 r cf>,,:>!.(s) f(s) ds (41) and so the resolvent kernel takes the form R(x, s; A) = f cf>,,(x)cf>,,(s) v=l Av - A (42) if this series is uniformly convergent, for then it is permissible to reverse the order of the integration and the summation in (41). Now we have shown in section 10.3 that L;=11<Pv(x)1 2 /A; is convergent, and since Av/(Av - A) -+ 1 as v -+ 00 it follows that f Icf>,,(x W 2 < 00. v=l (A v - A) (43) Hence, by the Riesz-Fischer theorem, the series (42) is mean square convergent to a square integrable kernel R(x, s; A). Then using the Hilbert-Schmidt theorem and the resolvent equation (8.12) written in the form R(x, s; A) = K(x, s)+ A r K(x, t)R(t, s; A) dt, (44) we obtain the absolutely and uniformly convergent series for the resolvent kernel R ( · A ) = K ( x S ) + A f cf>v(x)cf>v(s) x, s , , I..J \ ( \ _ \ ) v= 1 I\v I\v 1\ (45) which enables us to express the solution of (40) as cf>(x) = f(x) + A I b K(x, s)f(s) ds + A 2 f (f, cf>,,) cf>" (x) . (46) a v=l Av(Av - A) 
146 Hilbert - Schmidt theory It can be seen from (45) that the singularities of the resolvent kernel are simple poles occurring at the characteristic values Av of the Hermitian kernel K(x, s). 10.5 Bounds on characteristic values Using Hilbert's formula (36) for a square integrable function tfJ(x) we have (K t{1, t(1) = f I( t{1, <P" W v=l '\V (47) and clearly 1 (K <p", <p,,) = A" (v = 1, 2, . . .) (48) Let A, A (v = 1, 2, . . .) denote the positive and negative charac- teristic values associated with the characteristic functions <p(x), <p(x) respectively, and arranged so that O< \+\+ \+ 1\ 1  1\2  · · ·  1\ v  · · ., (49) O>'\1";=:A2";=:.. .;=:A;=:.... Hence (Kt{1, t(1) f 1(t{1, W -; f 1(t{1, <PW v=l Av Al v=l 1  At (t{1,t{1) (50) and so, introducing the functional I[ t{1 ] = (K t{1, t(1) (tfJ, tfJ) (51) we obtain 1 I[ t{1 ]  At' (52) the equality being attained for tfJ = cf> . Now suppose that tfJ(x) is orthogonal to the characteristic func- tions cf> (x), cf>(x), · · · , cf> -1 (x) so that (t/1, cf> t) = (t/1, q,) = · · · = (t/1, cf> -1) = o. 
Positive kernels 147 Then (K t/1, t/1).s;; f I (t/1, W .s;;  f I (t/1, <f> w v=n Av An v=n 1 .s;; A  (t/1, t/1) (53) which gives 1 I[ t/1 ].s;; A ' Similarly we can show that (54) 1 I[ t/1 ];;;.: A 1" and that if .p(x) is orthogonal to <PI, <P2, . . . , c/>;;-1 then 1 I[ t/1 ];;;.: A;; ' (55) (56) 10.6 Positive kernels A Hermitian kernel K(x, s) is said to be positive or non-negative definite if (K.p, .p)0 (57) for every square integrable function .p(x); and said to be positive definite if (K.p, .p»0 (58) for every square integrable function satisfying 11.p11 > O. We see at once from (47) and (48) that K(x, s) is a positive kernel if and only if all its characteristic values are positive, that is 0<A1 A2.... Further K(x, s) is positive definite if and only if the full orthonor- mal system of characteristic functions c/>v(x) (v = 1, 2, . ..) is also complete. For (K.p, .p) = 0 if and only if (.p, <Pv) = 0 for all v. Now let us suppose that the positive kernel K(x, s) is continuous. Since K(x, s) is Hermitian it follows that K(x, x) is real. Then we can show that K(x,x)O for axb. To this end we suppose that, on the contrary, there exists Xo with a<xo<b such that K(xo,xo)=-e where e>O. We are given that 
148 Hilbert-Schmidt theory K(x, s) is continuous. Hence the real part of the kernel must also be continuous and so we can find 5 > 0 such that for Xo - 5  x  Xo + 5, Xo- 5  s  xo+ 5 we have Re {K(x, s)}< -!e. Let .po(x) = { I (xo- 5  x  xo+ 5) o elsewhere and then, since (K .po, .po) is real, we see that (Kt{1o, t{1o) = r r Re {K(x, s)}t{1o(s)t{1o(x) ds dx = [:8 [:8Re{K( s)}ds dx < -!e(25)2 < 0 which contradicts our supposition that the kernel K(x, s) is positive. Hence K(x, x);=:: 0 for a < x < b and also at the end points since the kernel is continuous. 10.7 Mercer's theorem Mercer showed in 1909 that the expansion theorem given in section 10.3 can be strengthened if K(x, s) is a continuous positive kernel. Thus, supposing that K(x, s) is such a kernel, we see that K(n+1)(x, s) = K(x, s) - vt 1 cf>v(Xlv(S) (59) is also positive, and continuous for all n since the characteristic functions c/>v(x) of a continuous kernel are continuous. Hence, using the result proved in the previous section, we have K(x, x) - ! cf>v (x)cf>v (x) = K(n+1)(x, x)  0 (60) v=l Av . . gIvIng ! Icf>v(xW  K(x, x) v=l Av (61) for all n. Integrating we obtain n 1 [b Vl Av  Ja K(x, x) dx = trace K = Kl (62) 
Mercer's theorem 149 and so, as K(x, x) is continuous and therefore bounded, it follows that 00 1 L -<00 v = 1 Av (63) which is a much stronger result than (39). Now, by Cauchy's inequality (6.9), we have L=+1 lcf>v(Xlv(S)r  L=+1 lcf>vW } L=+1 lcf>v:W }. (64) Al (61) . I . h oo l<Pv(x)1 2 . f . so Imp Ies t at L,v=l IS convergent or all x. SInce K(x, x) is bounded Av f I cf>v (xW < C 2 v=n+l Av (axb) (65) where C is a positive constant, and given any e > 0 and a fixed value of s there exists N such that f I cf>v (sW < e 2 v=n+l Av (66) for n, m > N. Thus we have f lcf>v(x)cf>v(s)1 < eC v=n+l Av (67) and so we see that f cf>v(x )cf>v(s) v=l Av (68) is absolutely and uniformly convergent in x for each s. Similarly we can show that (68) is absolutely and uniformly convergent in s for each x. Since we know that (68) is mean square convergent to K(x, s) it follows that we must have K(x, s) = f cf>v(x)cf>v(s) v=l Av (69) In particular K(x, x) = f lcf>v(xW v=l Av (70) 
150 Hilbert-Schmidt theory where the partial sums of the infinite series form a monotonic se- quence of continuous functions which is convergent to a continuous function. Now Dini's theorem states that if a monotonic sequence of real continuous functions is convergent in the closed interval a  x  b to a continuous function, then the sequence converges uniformly in this interval. It follows that the series (70) is uniformly convergent in x and hence, by Cauchy's inequality, that (69) is uniformly convergent in x, s. 10.8 Variational principles We showed in section 10.5 that if K is an integral operator possessing a square integrable Hermitian kernel and .p is a square integrable function, then the functional I[ 1/1 ] = (K 1/1, 1/1) (.p, .p) (71) has the maximum value 1/ A  when .p = cP  and the minimum value l/Al when .p = q,1; and moreover if .p is orthogonal to the charac- teristic functions <p, <p;,..., <P-1 then I[.p] has the maximum value l/A when .p = <p, while if .p is orthogonal to CPl, c/>"2,. .., CP-l then I[.p] has the minimum value l/A when .p = cp. These properties lead us to expect that I[.p] is stationary whenever .p is one of the characteristic functions <Pv of K, thus giving rise to a variational principle. To establish that this is the case we examine the variation 5I[ <Pv] = I[ <Pv + 5<pv] - I[ <Pv] (72) arising from an arbitrary variation 5<pv in the characteristic function cf>v. We have I[ </> + l></> ] = (K </>" + K l></>", </>" + l></>,,) v v ( <Pv + 8cpv, <Pv + 5<pv) = I[ </> ] + (K </>'" l></>,,) + (K l></>", </>,,) v (<Pv, <Pv) _ (K </>'" </>,,){ (</>'" l></>,,) 2+ (l></>", </>,,)} + 0 {Il></>" 12} (<Pv, c/>v) = I[ c/>v] + O{15c/>vI 2 } (73) 
Variational principles 151 since AvKcpv = CPv. Neglecting the quantity O{I«SCPvI 2 } of the second order of smallness, we may write this result as the variational principle «SI[cpv] = 0 (74) which establishes that I[ t/J] is stationary when t/J = CPv. We can also derive a variational principle for ({, cp) where cp is the solution of the inhomogeneous equation cp = { + AK cf> (75) choosing A to be real so that ({, cf» = (cf>, cp) - A (K cp, cp) = (cf>, n (76) and thus is a real quantity also. We introduce the functional J[ t/J] = (f, t/J) + (t/J, n - (t/J, t/J) + A (K t/J, t/J) (77) and so we have J[ cf> ] = ({, cf». Then J[ cp + «Scp ] = ({, cf> + 8cp ) + ( cp + «Scf>, f) - ( cf> + «Scp, cf> + «Scf> ) + A (K cf> + K «Scp, cp + «Scp) = J[ cp ] + (f - cp + AK cf>, «Scf» + (<<Scf>, {- cp + AK cf» - ({1- AK} 8cf>, «Scf» = J[ cf>] - ({1- AK} «Scf>, «Scp) (78) since cp satisfies (75). Neglecting the second order quantity on the right-hand side of (78) yields the variational principle «SJ[cf>] = O. (79) Moreover if AK is a negative operator we see that «SJ[cp]O (80) and this means that J[ t/J] attains a maximum value for t/J = cf> so that J[ t/J] provides a lower bound to ({, cp). An alternative expression for J[ t/J] may be obtained by setting t/J = ax and optimizing with respect to the parameter a. We have J[ax] = a({, X) + a (X, n - aa(x, X) + Aaa(Kx, X), 
152 Hilbert-Schmidt theory and taking aJ _ aJ - 0 ---- aa aa we find that (f, X) a= (X, X) - A (KX, X). This results in the homogeneous formula J[x] = (t, X)(X, f) (X, X)-A(KX, X) and clearly J[cp]=(f, cp) using (76). (81) 10.8.1 Rayleigh-Ritz variational method We may use the variational principle (74) to devise a method of obtaining approximations to the characteristic values. This method is associated with the original work of Rayleigh and Ritz in connec- tion with the determination of the natural frequencies of vibration of mechanical systems. Let t/11, t/12, . . .., t/1n be a set of n linearly independent square integrable functions and put n t/1 = L c r t/1r r=1 (82) where the C r (r = 1, . . . , n) are n adjustable parameters. Then we have 1[t/1] = L:=l L=l cr.krs r=1 s=1 crcsa rs (83) where k rs = (Kt/1r, t/1s) and a rs = (t/1r, t/1s). We now optimize with re- spect to the parameters C r by taking aI _ aI _ 0 ---- a C r a c r (r=l,...,n) (84) and these yield the set of n linear homogeneous equations n n A L krsc s - L arsc s = 0 s=1 s=1 (r= 1,..., n) (85) 
Variational principles 153 where 1/ A is the value of I[ tf1]. These equations have a non-trivial solution if and only if the determinant of the coefficients vanishes: Ak 11 - all Ak 12 - a12 Ak 21 - a21 Ak 22 - a22 Ak 1n - a1n Ak 2n - a2n = 0 (86) Ak n1 -a n 1 Ak n2 -a n 2 Ak nn - ann The n roots A (1), A (2), . . ., A (n) of this secular equation provide variational approximations to the characteristic values of K. In particular if K is positive and A (1) is the least of these roots then o < A 1  A (1), thus giving an upper bound to the smallest characteris- tic value. Example. As a simple illustration of the Rayleigh-Ritz method we consider a vibrating flexible string having uniform density p and length I which is stretched at tension T and fixed at its two ends x = 0, I. Referring to section 2.1 we see that the homogeneous Fredholm linear integral equation corresponding to this physical problem has the positive symmetric kernel s(l- x) (0  s  x  I) I K(x, s) = x(l- s) (87) (O x  s  I) I Let us take t/J(x) = C1X + C2 X2 (88) as our trial function. Then 1 3 all = 3 ' 1 4 a12 = a21 =4' 1 5 a22 = - 5 (89) and 1 5 k 11 = 45 ' 1 6 k 12 = k 21 = 72 ' 1 7 k 22 = 112 (90) 
154 Hilbert-Schmidt theory The secular equation is >..1 5 1 3 --- 45 3 Al 6 1 4 --- 72 4 >.. 1 6 1 4 --- 72 4 AI? 1 5 --- 112 5 =0 (91) which gives the quadratic equation 5(AI 2 )2-432AI 2 +3780=0 (92) whose least root is A (1) = 9.880/1 2 . Then the approximate least angular frequency of vibration resulting from the application of the Rayleigh-Ritz method to the function (88) is given by {W(1)}2 = T A (1) = 9.880  (93) p pi which is only slightly greater than the exact value 2 2 T T Wi = 17' pzZ = 9.870 p12 ' (94) Problems 1. Show that the characteristic values of a skew-Hermitian kernel (called skew-symmetric if real) satisfying K(x,s)=-K(s,x) are pure imaginary numbers. 2. Show that the Poisson kernel 1 1- a 2 K(x, s) = 217' 1 + a 2 - 2a CDS (x.- s) where lal < 1, has the characteristic functions 1 .J27r ' 1 . J;. SIn vx, (0x27r,0s27r) 1 J;. cos vx (v = 1, 2, . . . ) with characteristic values 1, a -v, a -v respectively. Use the expansion theorem (42) for the resolvent kernel to show that it is given by the series obtained in problem 4 at the end of chapter 8. 
Problems 155 3. Show that the homogeneous equation «J(x) = A r K(x, s)«J(s) ds with the symmetric kernel { X(l- s) K(x, s) = s(1- x) (x  s) (x  s) is equivalent to the differential equation cp" + Acp = 0 with cp(O) = cp(l) = O. Hence show that the kernel has characteristic functions J2 sin V'1TX (v = 1, 2, . ..) with characteristic values (V'1T)2 respec- tively. Use the expansion theorem (69) to show that the kernel K(x, s) is given by the uniformly convergent series  f sin V'1TX sin V'1TS 2 i.J 2 '1T v=l V and also obtain an expansion for the resolvent kernel.  4. Show that the homogeneous equation with the symmetric kernel { X (Oxsl) K(x, s) = s (Osxl) is equivalent to the differential equation cp" + Acp = 0 with cp(O) = cp'(l) = O. Hence show that the kernel has characteristic functions J2 sin {[(2v-1)/2]'1Tx} (v = 1, 2,...) with characteristic values [(2v-1)/2]2'1T 2 . Use the expansion theorems to obtain uniformly convergent series for K(x, s) and the resolvent kernel. 5. Use the Rayleigh-Ritz variational method together with the trial function t/1(x) = Cl + C2 X to obtain an upper bound to the characteristic value for the separ- able kernel ( ) x+s K x,s =e (Ox  1, O s 1) 
156 Hilbert-Schmidt theory Verify that the result is slightly greater than the exact value ob- tained in problem 3 at the end of chapter 1. 6. Use the Rayleigh-Ritz variational method and the trial function t/J(x) = CIX + C2 X2 to obtain an upper bound to the least characteristic value for the symmetric kernel in problem 4 and verify that it is slightly greater than the exact value (7r/2)2. 7. By using the variational principle (79) together with the trial function 00 t/J(x) = [(x) + L cvCPv(x), v=l where the CPv(x) are the characteristic functions of the square integrable Hermitian kernel K(x, s) associated with the characteris- tic values Av, show that A([, CPv) C v = Av - A in agreement with the solution of the Fredholm equation of the second kind obtained in section 10.4. Show also that for real A (t, ef» = (t, f) + A f (t, cf>v)( cf>", f) = (cf>, f). v = 1 Av - ,\ 
Bibliography Bocher, M., An Introduction to the Study of Integral Equations, Cambridge Tracts in Mathematics and Mathematical Physics, No. 10, Cambridge University Press, 1913. Cochran, J. A., The Analysis of Linear Integral Equations, McGraw-Hill, New York, 1972.. Courant, R. and Hilbert, D., Methods of Mathematical Physics, Vol. 1, Ch. III, Interscience, New York, 1953. Goursat, E., A Course in Mathematical Analysis, Vol. III, Part 2, Chs. VIII-XI, Dover, New York, 1964. Green, C. D., Integral Equation Methods, Nelson, London, 1969. Hildebrand, F. B., Methods of Applied Mathematics, 2nd edn., Ch. 3, Prentice-Hall, Englewood Cliffs, New Jersey, 1965. Hochstadt, H., Integral Equations, Wiley, New York, 1973. Hoheisel, G., Integral Equations, Nelson, London, 1967. Kanwal, R.P., Linear Integral Equations, Academic Press, New York, 1971. Lovitt, W. V., Linear Integral Equations, Dover, New York, 1950. Mikhlin , S. G., Integral Equations, 2nd edn., Pergamon Press, London, 1964. Morse, P. M. and Feshbach, H., Methods of Theoretical Physics, McGraw- . Hill, New York, 1953. Schmeidler, W., Linear Operators in Hilbert Space, Academic Press, New York, 1965. Smirnov, V. I., A Course of Higher Mathematics, Vol. IV, Integral Equations and Partial Differential Equations, Pergamon Press, London, 1964. Smithies, F., Integral Equations, Cambridge Tracts in Mathematics and Mathematical Physics, No. 49, Cambridge University Press, London, 1958. Titchmarsh, E. C., Introduction to the Theory of Fourier Integrals, 2nd edn., Ch. 11, Oxford University Press, 1948. Tricomi, F. G., Integral Equations, Interscience, New York, 1957. Whittaker, E. T. and Watson, G. N., A Course of Modern Analysis, 4th edn., Ch. XI, Cambridge University Press, London, 1927. Zabreyko, P. P., Koshelev, A. I., Krasnosel'skii, M. A., Mikhlin , S. G., Rakovshchik, L. S. and Stet'senko, V. Ya., Integral Equations - A Re- ference Text, Noordhoff International Publishing, Leyden, 1975. 
Index Abel's integral equation, 1, 3, 37 Abelian group, 80 Abstract Hilbert space, 80, 109 Adjoint equation, 99, 110 operator, 89, 99, 110 Adjugate matrix, 116 Associative law, 80, 86 Basis, 70, 73, 83 Bernoulli, 14 Bessel function, 37 Bessel's inequality, 77, 83 Best mean square approximation, 77 Bifurcation point, 11 Bilinear formula, 139 Biorthogonal series, 111 Bacher's example, 109 Bounded linear operator, 89 Bounds on characteristic values, 146 Cauchy inequality, 71 Cauchy principal value, 56, 63 Cauchy-Riemann equations, 57, 63 Cauchy sequence, 73, 79, 82 Cauchy-type integral, 56 Characteristic functions, 101 values, 7, 21, 101, 110, 136 bounds on, 146 vectors, 110 Commutative law, 80, 86 Compact operator, 92 Completely continuous operators, 92, 111, 125 Completeness relation, 78, 83 Complete linear vector space, 80 orthonormal system, 82 space, 74, 76, 79, 82 system, 76, 83 Continuous functions and kernels, 44, 102, 126 Convergence of Neumann series, 2, 44 strong, 72, 78, 82 weak, 92 Convolution theorem for Fourier transform, 25 for Laplace transform, 24 for Mellin transform, 25 Cramer's rule, 116 Critical speeds, 21 Degenerate kernels, 7, 114 approximation by, 120 Delta function, 18 Dense, 82 Difference kernel, 5 Differential equations linear, 14 first order, 14 second order, 15, 17 Dimension of Hilbert space, 82 Dini's theorem, 150 Dirac, 18 Dirichlet's problem, 2, 55 Dissection, 122 Distance, 72 Distance function, 81 Distributive law, 80, 86 Domain, 89 Du Bois-Reymond, 1 Euclidean space 3-dimensional, 69 n-dimensional, 70 Expansion theorems, 139 for iterated kernels, 143 Field, 80 Finite Hilbert transform, 65, 67 Finite rank, 114 Flexible string, 16, 153 Foppl's integral equation, 67 Fourier transforms, 1, 25, 26, 29, 35, 40 coefficients, 77, 91 
Fox's integral equation, 39 Fredholm alternatives, 124 determinant, 129 equations, 2 of first kind, 3, 34 of second kind, 3, 26, 43, 98, 102, 109, 144 formulae, 126 minor, 130 theorems, 121, 124, 125 Full orthonormal system, 139 Function space, 74, 79 L 2 , 79 Fundamental set, 82 Gamma function, 38 Generalized functions, 19 Gram-Schmidt orthogonalization, 76, 82 Green's function, 18, 54 in two dimensions, 54 in three dimensions, 55 Hadamard's inequality, 130 Hermitian adjoint, 88 Hermitian kernels, 5, 88, 136, 147 Higher dimensions, 53 Hilbert formula, 143 integral operator, 64 kernel, 61 positive, 147 non-negative definite, 147 singular integral equation, 65 space, 69, 79 dimension of, 82 of sequences, 71 transform, 63, 64 Hilbert-Schmidt kernel, 136 series, 141 theorem, 141 Homogeneous equation, 3 Huygens tautochrone problem, 2 Impulse function, 19 Index, 10, 116 Influence function, 20 Inner product, 69, 71, 74, 80 Integral equation Abel's 1, 3, 37 Index 159 convolution type, 24 Foppl's, 67 Fox's, 39 Fredholm first kind, 3, 34 second kind, 3, 26, 43, 98, 102, 109, 144 Homogeneous, 3 Lalesco-Picard, 28 singular, 9, 53 Stieltjes, 34 Volterra first kind, 3, 4, 36 second kind, 4, 31, 45, 105 Integral transform, 24 Fourier, 1, 25, 26, 29, 35, 40 Hilbert, 63, 64 Laplace, 1, 24, 31, 34, 36, 38 Mellin, 25, 39, 40 Integral operators, 85 completely continuous, 92 continuous, 90, 92 norm of, 87 Iterated kernels, 46, 143 Inverse element, 80 Inverse point, 59 Kernel, 3 continuous, 102 convolution type, 5 degenerate, 7, 114, 120 difference, 5, 24 Hermitian, 5, 88, 136, 147 Hilbert, 61 Hilbert-Schmidt, 136 iterated, 46, 143 L 2 ,9 matrix, 91 Pincherle-Goursat, 114 polar kernel, 53 positive, 147 R2,9 regular value of, 99, 110 resolvent, 6, 47, 145 self-adjoint, 88 separable, 6 skew-Hermitian, 154 skew symmetric, 119, 154 singular, 4, 9, 53 square integrable, 8 symmetric, 4 Weyl, 11 Kneser, 137 
160 Index ' 2 space, 74 L 2 function, 8, 79 kernel, 9 Lalesco-Picard equations, 28 Laplace equation, 2, 55 uansfonn, 1, 24, 31, 34, 36, 38 Lebesgue, 8 Limit element, 82 Linear integral equation, 3 Linear manifold, 89 Linear operators, 85 bounded, 89 completely continuous, 92 continuous, 90, 92 in Hilbert space, 85 integral, 85 Linear vector space, 80 Liouville, 2, 43 Liouville-Neumann series, 43 Matrix representation, 91 Mean square convergence, 77 limit, 78 Mellin uansform, 25, 39, 40 Mercer's theorem, 148 Metric, 80 Minkowski's inequality, 75 n-dimensional space, 70 Neumann, 2, 43 Neumann series, 43, 102 convergence of, 2, 44, 102 Non-linear equations, 11 Norm, 69, 70, 71, 74, 87, 89 Normal operator, 96 Null operator, 87 Operator, adjoint, 89, 99, 110 bounded, 89 compact, 92 completely continuous, 92, 111, 125 continuous, 90, 92 Hermitian, 89 identity, 86 integral, 85 norm of, 87 normal, 96 null, 87 resolvent, 98, 110 self-adjoint, 89 Orthogonalization, Gram-Schmidt, 76 Orthonormal system, 70, 75, 82 Parseval's formula, 78 generalized, 78 Pincherle-Goursat kernel, 114 Poisson, 2 Poisson's formula, for half plane, 60 for unit disc, 59 Polar kernel, 53 Positive Hermitian kernel, 147 R 2 function, 8 kernel, 9 Range, 89 Rank, 10, 114, 116 finite, 10, 114 infinite, 10 Rayleigh-Ritz variational method, 152 Regular value, 99, 110 Reimann, 8 Riesz-Fischer theorem, 79, 140 Ritz, 152 Resolvent equation, 98, 110 kernel, 6, 47, 145 Joperator, 98, 110 \ set, 99 -......: Scalar product, 69 Schwartz's inequality, 72, 74, 81 Self adjoint kernel, 88 Shaft, 21 Singular kernels, 4, 9, 53 Singular integral equations, 9, 53 of Hilbert type, 65 Singular point, 12 Space complete, 76, 79, 82 Euclidean, 69 Hilbert, 69, 79 Spectrum, 10, 101, 136 Square integrable functions, 8 kernels, 8 Stieltjes integral equation, 34 String, 16, 20, 153 Strong convergence, 72, 78, 82 Successive approximations, method of, 2,43 Symmetric kernel, 4 
Index 161 Tautochrone, 2, 37 Trace, 133, 144 Transforms Fourier, 1, 25, 26, 29, 35, 40 Hilbert, 63, 64 Laplace, 1, 24, 31, 34, 36, 38 Mellin, 25, 39, 40 Trial function, 153 Triangle inequality, 70, 72, 75, 81 Uniqueness theorem, 99, 105 Variational principles, 150 Volterra, 2 Volterra equation, 2 of first kind, 3, 4, 36 of second kind, 4, 31, 45, 105 Weak convergence, 92 Weyl kernel, 11 Wronskian, 20