Applied

Mathematical
Sciences
EDITORS
FrttzJohn J.E. Marsden l..awntnce Sirovich
Coutant Inslitute d Department d DlvisIond
MathemtJtJc8J ScIences Mathematics ApplIed Mathematics
New YOft( Univ8rslty University of California Brown University
New York, NY 10012 Bertteley, CA 94720 Providence, RI 02912

ADVISORS
H. cabanne8 University of ParIs-VI J. Kaler Stanford University
M. GhI New YOft( University G.B. Whitham California Inst. of Technology
J.K. Hale Brown University

EDITORIAL STATEMENT
The mathematization of all sciences, the fading of traditional scientific bounda-
ries, the impact of computer technology, the growing importance of mathematical-
computer modelling and the necessity of scientific planning all create the need both
in education and research for books that are introductory to and abreast of these
developments.
The purpose of this series is to provide such books, suitable for the user of
mathematics, the mathematician interested in applications, and the student scientist.
In particular, this series will provide an outlet for material less formally presented and
more anticipatory of needs than finished texts or monographs, yet of immediate in-
terest because of the novelty of its treatment of an application or of mathematics
being applied or lying close to applications.
The aim of the series is, through rapid publication in an attractive but inexpen-
sive format, to make material of current interest widely accessible. This implies the
absence of excessive generality and abstraction, and unrealistic idealization, but with
quality of exposition as a goal.
Many of the books will originate out of and will stimulate the development of
new undergraduate and graduate courses in the applications of mathematics. Some
of the books will present introductions to new areas of research, new applications
and act as signposts for new directions in the mathematical sciences. This series will
often serve as an intermediate stage of the publication of material which, through
exposure here, will be further developed and refined. These will appear in conven-
tional format and in hard cover.

MANUSCRIPTS
The Editors welcome all inquiries regarding the submission of manuscripts for
the series. Final preparation of all manuscripts will take place in the editorial offices
of the series in the Division of Applied Mathematics, Brown UniverSity, Providence,
Rhode Island.

SPRINGER SCIENCE+BUSINESS MEDiA, LLC

Applied Mathematical Sciences I Volume 25
Applied Mathematical Sciences

1. John: Partial Differential Equations, 4th ed.

2. Sirovich: Techniques of Asymptotic Analysis.
3. Hale: Theory of Functional Differential Equations, 2nd ed.
4. Percus: Combinatorial Methods.
5. von Mises/Friedrichs: Fluid Dynamics.
6. Freiberger/Grenander: A Short Course in Computational Probability and Statistics.
7. Pipkin: Lectures on Viscoelasticity Theory.
8. Giacaglia: Perturbation Methods in Non-Linear Systems.
9. Friedrichs: Spectral Theory of Operators in Hilbert Space.
10. Stroud: Numerical Quadrature and Solution of Ordinary Differential Equations.
11. Wolovich: Linear Multivariable Systems.
12. Berkovitz: Optimal Control Theory.
13. Bluman/Cole: Similarity Methods for Differential Equations.
14. Yoshizawa: Stability Theory and the Existence of Periodic Solutions and Almost
Periodic Solutions.
15. Braun: Differential Equations and Their Applications, 3rd ed.
16. Lefschetz: Applications of Algebraic Topology.
17. Collatz/Wetterling: Optimization Problems.
18. Grenander: Pattern SynthesiS: Lectures in Pattern Theory, Vol I.
19. Marsden/McCracken: The Hopf Bifurcation and its Applications.
20. Driver: Ordinary and Delay Differential Equations.
21. Courant/Friedrichs: Supersonic Flow and Shock Waves.
22. Rouche/Habets/Laloy: Stability Theory by Liapunov's Direct Method.
23. Lamperti: Stochastic Processes: A Survey of the Mathematical Theory.
24. Grenander: Pattern Analysis: Lectures in Pattern Theory, Vol. II.
25. Davies: Integral Transforms and Their Applications, 2nd ed.
26. Kushner/Clark: Stochastic Approximation Methods for Constrained and
Unconstrained Systems.
27. de Boor: A Practical Guide to Splines.
28. Keilson: Markov Chain Models-Rarity and Exponentiality.
29. de Veubeke: A Course in Elasticity.
30. Sniatycki: Geometric Quantization and Quantum Mechanics.
31. Reid: Sturmian Theory for Ordinary Differential Equations.
32. Meis/Markowitz: Numerical Solution of Partial Differential Equations.
33. Grenander: Regular Structures: Lectures in Pattern Theory, Vol. III.
34. Kevorkian/Cole: Perturbation Methods in Applied Mathematics.
35. Carr: Applications of Centre Manifold Theory.

(continued on inside back cover)

B. Davies

Integral Transforms
and their Applications
Second Edition

With 50 Illustrations

Springer Science+Business Media, LLC

B. Davies
The Australian National Vniversity
Department of Mathematies
Post Oftiee Box 4
Canberra, A.C.T. 2600
Australia

AMS Classifieation: 44-01, 44AIO, 44A15, 44A20, 44A30

Library of Congress Cataloging in Publieation Data

Davies, B. (Brian)
Integral transforrns and their applieations.
(Applied mathematieal seienees; v. 25)
Bibliography: p.
Includes index.
1. Integral transforrns. I. Title. II. Series:
Applied mathematieal seienees (Springer-Verlag
New York Ine.); v. 25.
QA 1. A647 voI. 25 1984 [QA432] 510 s [515.7'23] 84-14015

© 1978, 1985 Springer Science+Business Media New York

Originally published by Springer-Verlag New York Inc. in 1985

AII rights reserved. No part of this book may be translated or reprodueed in any form
without written permis sion from Springer-Verlag, 175 Fifth Avenue, New York, New
York 10010, V.S.A.

987 654 3 2 1

ISBN 978-0-387-96080-7 ISBN 978-1-4899-2691-3 (eBook)

DOI 10.1007/978-1-4899-2691-3
 Preface to the Second Edition

In preparing this second edition I have restricted

myself to making small corrections and changes to the first
edition. Two chapters have had extensive changes made. First,
the material of Sections 14.1 and 14.2 has been rewritten to
make explicit reference to the book of Bleistein and Handelsman,
which appeared after the original Chapter 14 had been written.
Second, Chapter 21, on numerical methods, has been rewritten to
take account of comparative work which was done by the author
and Brian Martin, and published as a review paper. The material
for all of these chapters was in fact, prepared for a transla-
tion of the book.
Considerable thought has been given to a much more com-
prehensive revision and expansion of the book. In particular,
there have been spectacular advances in the solution of some
non-linear problems using isospectra1 methods, which may be re-
garded as a generalization of the Fourier transform. However,
the subject is a large one, and even a modest introduction would
have added substantially to the book. Moreover, the recent
book by Dodd et al. is at a similar level to the present volume.
Similarly, I have refrained from expanding the chapter on num-
erical methods into a complete new part of the book, since a
specialized monograph on numerical methods is in preparation
in collaboration with a colleague.

B. Davies
Canberra, 1984

v
 Preface to the First Edition

This book is intended to serve as introductory and

reference material for the application of integral transforms

to a range of common mathematical problems. It has its im-

mediate origin in lecture notes prepared for senior level
courses at the Australian National University, although I owe
a great deal to my colleague Barry Ninham, a matter to which
I refer below. In preparing the notes for publication as a
book, I have added a considerable amount of material addi-
tional to the lecture notes, Hith the intention of making the
book more useful, particularly to the graduate student in-
volved in the solution of mathematical problems in the physi-
cal, chemical, engineering and related sciences.

Any book is necessarily a statement of the author's

viewpoint, and involves a number of compromises. My prime
consideration has been to produce a work whose scope is
selective rather than encyclopedic; consequently there are

many facets of the subject which have been omitted--in not a

few cases after a preliminary draft was written--because I

vii
believe that their inclusion would make the book too long.
Some of the omitted material is outlined in various problems

and should be useful in indicating possible approaches to

certain problems. I have laid great stress on the use of
complex variable techniques, an area of mathematics often
unfashionable, but frequently of great power. I have been
particularly severe in ex~ising formal proofs, even though
there is a considerable amount of "pure mathematics" associated
with the understanding and use of generalized functions,

another area of enormous utility in mathematics. Thus, for

the formal aspects of the theory of integral transforms I

must refer the reader to one of the many excellent books
addressed to this area; I have chosen an approach which is
more common in published research work in applications. I

can only hope that the course which I have steered will be
of great interest and help to students and research workers
who wish to use integral transforms.
It was my privilege as a student to attend lectures
on mathematical physics by Professor Barry W. Ninham, now at
this university. For several years it was his intention to
publish a comprehensive volume on mathematical techniques in
physics, and he prepared draft material on several important
topics to this end. In 1972 we agreed to work on this pro-
ject jointly, and continued to do so until 1975. During

that period it became apparent that the size, and therefore

cost, of such a large volume would be inappropriate to the

current situation, and we decided to each publish a smaller

book in our particular area of interest. I must record my

gratitude to him for agreeing that one of his special

interests--the use of the Mellin transform in asymptotics--
viii
should be included in the present book. In addition there
are numerous other debts which I owe to him for guidance

and criticism.
References to sources of material have been made in

two ways, since this is now a fairly old subject area. First,
there is a selected bibliography of books, an~ I have
referred, in various places, to those books which have been
of particular assistance to me in preparing lectures or in

pursuing research. Second, where a section is based directly

on an original paper, the reference is given as a footnote.
Apart from this, I have not burdened the reader with tedious
lists of papers, especially as there are some comprehensive
indexing and citation systems now available.
A great deal of the final preparation was done while
I was a visitor at the Unilever Research Laboratories (UK)
and at Liverpool University in 1975, and I must thank those
establishments for their hospitality, and the Australian
National University for the provision of study leave. Most
of the typing and retyping of the manuscript has been done
by Betty Hawkins of this department while the figures were
prepared by Mrs. L. Wittig of the photographic services de-
partment, ANU. Timothy Lewis, of Applied Mathematics at
Brown University, has proofread the manuscript and suggested
a number of useful changes. To these people I express my
gratitude and also to Professor Lawrence Sirovich for his
encouragement and helpful suggestions. This book is dedicated
to my respected friend and colleague, Barry Ninham.

Brian Davies

Canberra, Australia
1977
ix
 Table of Contents
Page
PART I: THE LAPLACE TRANSFORM 1
1. Definition and Elementary Properties 1

1.1. The Laplace Transform 1

1.2. Important Properties 3

1.3. Asymptotic Properties: Watson's lemma 8

2. The Inversion Theorem 15

2.1. The Riemann-Lebesgue Lemma 15

2.2. Dirichlet Integrals 17

2.3. The Inversion Integral 19

2.4. Inversion of Rational Functions 20

2.5. Taylor Series Expansion 23

3. Ordinary Differential Equations 26

3.1. First and Second Order Differential

Equations 26
3.2. Higher Order Differential Equations 29

3.3. Simultaneous Differential Equations 33

3.4. Equations With Polynomial Coefficients 42
4. Partial Differential Equations 47
4.1. The Diffusion Equation 47

4.2. Wave Propagation 50

5. Integral Equations 59

5.1. Convolution Equations of Volterra Type 59

5.2. Convolution Equations Over an Infinite Range 64
5.3. The Percus-Yevick Equation for Hard Rods 67
6. The Inversion Integral 76
6.1. Inversion of Meromorphic Functions 76
6.2. Inversions Involving a Branch Point 79

6.3. Watson's Lemma for Loop Integrals 82

xi
 Page
6.4. Asymptotic Forms for Large t 84
6.5. Heaviside Series Expansion 86
PART II: THE FOURIER TRANSFORM
7. Definitions and Elementary Properties 89
7.1. The Exponential, Sine and Cosine Transforms 89

7.2. Important Properties 95

7.3. Spectral Analysis 97

7.4. Kramers-Kronig Relations 101

8. Application to Partial Differential Equations 110

8.1. Potential Problems 110

8.2. Water Waves: Basic Equations 114

8.3. Water Waves Generated by an Initial

Surface Displacement 117
8.4. Waves Due to a Periodic Disturbance:
Radiation Condition 120
9. Generalized Functions 130
9.1. The Delta Function 130
9.2. Test Functions and Generalized Functions 131

9.3. Elementary Properties 136

9.4. Analytic Functionals 143

9.5. Fourier Transforms of Generalized Functions 145

10. Green's Functions 155

10.1. One Dimensional Green's Functions 155

10.2. Green's Functions as Generalized Functions 159

10.3. Poisson's Equation in Two Dimensions 162

10.4. Helmholtz's Equation in Two Dimensions 168

11. Fourier Transforms in Two or More Variables 178
11.1. Basic Notation and Results 178

11.2. Diffraction of Scalar Waves 182

11.3. Retarded Potentials of Electromagnetism 185

xii
 Page
PART III: OTHER IMPORTANT TRANSFORMS

12. Mellin Transforms 195

12.1. Definitions 195
12.2. Simple Examples 196
12.3. Elementary Properties 201
12.4. Potential Problems in Wedge Shaped Regions 203
12.5. Transforms Involving Polar Coordinates 204
12.6. Hermite Functions 207
13. Mellin Transforms in Summation 214
13.1. Mellin Summation Formula 214
13.2. A Problem of Ramanujan 216
13.3. Asymptotic Behavior of Power Series 219
14. Integrals Involving a Parameter 225
14.1. Preliminary Example 225
14.2. A General Class of Problems 225
14.3. Ascending Expansions for Fourier Integrals 227
14.4. Multidimensional Integrals 230
15. Hankel Transforms 237
15.1. The Hankel Transform Pair 237
15.2. Elementary Properties 240
15.3. Some Examples 242
15.4. Boundary Value Problems 243
15.5. Weber's Integral 245
16. Dual Integral Equations 254
16.1. The Electrified Disc 254
16.2. Dual Integral Equations of Titchmarsh
Type 255
16.3. Erdelyi-Kober Operators 258

xiii
 Page
17. Integral Transforms Generated by Green's
Functions 267
17.1. The Basic Formula 267
17.2. Finite Intervals 269
17.3. Some Singular Problems 272
17.4. Kontorovich-Lebedev Transform 276
17.5. Boundary Value Problem in a Wedge 278
17.6. Diffraction of a Pulse by a Two-Dimen-
sional Half-Plane 280
PART IV: SPECIAL TECHNIQUES
18. The Wiener-Hopf Technique 288
18.1. The Sommerfeld Diffraction Problem 289
18.2. The Wiener-Hopf Procedure: Half-plane
Problems 299
18.3. Integral and Integro-Differential
Equations 301
19. Methods Based on Cauchy Integrals 313
19.1. Wiener-Hopf Decomposition by Contour
Integration 313
19.2. Cauchy Integrals 315
19.3. The Discontinuity Theorem 320
19.4. The Riemann Hilbert Problem 321
19.5. Simple Applications 323
19.6. Problems in Linear Transport Theory 324
19.7. The Albedo Problem 329
19.8. A Diffraction Problem 332
20. Laplace's Method for Ordinary Differential
Equations 342

20.1. Integral Transform Solutions 342

20.2. Hermite Polynomials 344

20.3. Hermite Functions 347

xiv
 Page
20.4. Bessel Functions: Integral Representations 351
20.5. Bessel Functions of the First Kind 353
20.6. Functions of the Second and Third Kinds 356
20.7. Poisson and Related Representations 362
20.8. Modified Bessel Functions 364
21. Numerical Inversion of Laplace Transforms 374
21.1. Gaussian Quadrature Formulae for the Laplace
Inversion Integral 374
21.2. Use of Laguerre Polynomials 379
21.3. Approximation of FCp) by Chebyshev
Polynomials for Real p 382
21.4. Representation by Fourier Series 385
21.5. Pade Approximation 386
21.6. Rational Approximation of FCp) 389
APPENDICES
A: The Factorial Function 396
B: Riemann's Zeta Function 399
C: The Exponential Integral 402
BIBLIOGRAPHY 403
INDEX 406

xv
 Part I: The Laplace Transform

1
§l. DEFINITION AND ELEMENTARY PROPERTIES

1.1. The Laplace Transform

Let f(t) be an arbitrary function defined on the
interval 0 < t < 00; then

F(p) = f: e- pt f(t) dt (1)

is the Laplace transform of f(t), provided that the integral

exists. We shall confine our attention to functions f(t)
which are absolutely integrable on any interval 0 < t ~ a,
and for which F(a) exists for some real a. It may readily
be shown that for such a function F(p) is an analytic func-
tion of p for Re(p) > a, as follows. First note that the
functions
Io
epCp,T) = T e-pt f(t) dt (2)

are analytic in p, and then that epCp,T) converges uni-

formly to FCp) in any bounded region of the p plane sat-
isfying ReCp) > a, as T + 00. It follows from a standard

1
 2 PART I: THE LAPLACE TRANSFORM

2
theorem on uniform convergence that F(p) is analytic in
the half-plane Re(p) > a.
As simple examples of Laplace transforms, we have
(i) Heaviside unit step function
I, t > 0
h(t) = {
(3)
0, t < 0

H(p) ( e- pt dt

lip, Re(p) > 0 , (4)

(ii)
iwt,
f (t) e w real (5)

F(p) J: e- pt e iwt d t
1 Re(p) > 0, (6)
p-iw ,
(iii)
St
f (t) tY e , a real > -1 (7)

F(p) = Y! Re (p) > Re (S). (8)

(p- S) Y+l

An important feature of these examples, and indeed of

many of the Laplace transforms which occur in applications,
is that the analytic function defined by (1) in the half-plane
Re(p) > a can be analytically continued into the remainder
of the plane once the singularity structure has been eluci-
dated. Thus the functions defined by (4) and (6) exhibit
only a simple pole; in the case of (8) there is a branch

point at p = S except for the special case that Y is an

integer, when we get a pole.
sl. Definition and elementary properties 3

1.2. Important Properties

There are a number of simple properties which are of
recurring importance in the application of the Laplace trans-
form to specific problems. In order to simplify somewhat
the statement of these results, we introduce the notation

~[f] = pep) = J~ e- pt f(t) dt (9)

o
which emphasizes the operator nature of the transform.

Linearity: If we consider the linear combination

(10)

where the ak are arbitrary constants, then

n
~[f] = I a k ~[fk] . (11)
k=l

One immediate consequence of this is that if f depends on a

variable x which is independent of t, we have

~[df/ax] a ..st[f] / ax, (12)

~u: f d~ = f: ~[f] dx· (13)

These results follow by trivial manipulation of the integrals

in the half-space Re(p) > a in which all the integrals con-
verge absolutely and uniformly (in x). But then they must
also hold over the entire region of the complex p plane to
which the transforms may be analytically continued.

Derivatives and Integrals: If we apply integration by parts

to (1), we obtain
4 PART I: THE LAPLACE TRANSFORM

Y[f' (t)] pY[f] - f(O+), (14)

f(O+) = lim f(t).

t+O
The distinction which we have made between the value of f(O)
and the limit of f(t) as t + 0 is of importance in prob-
lems where there are discontinuities at t = O. In many
problems, initial values of functions are specified with the
implied meaning that they are limiting values for small t,
and the distinction becomes unimportant and may be neglected.
By repeating the procedure of integration by parts, we can
derive the general result

Y[f(n) (t)] = pn Y[f] _ f pn-k f(k-l) (0+). (15)

k=l
A similar result holds for differentiation of the Laplace
transform with respect to p. By differentiating under the
integral sign, we obtain

dn
-n:
dp
F(p) (16 )

Suppose now that we define get) by

get) = ft
f(T) dT. (17)
o
Then by interchanging orders of integration, we get

G(p) =
r e- pt dt
I: f (T) dT

r
0
00
-pt
f(T) dT e dt
0 IT (18)

P
1
f: e- PT f(T) dT
1
F(p),
P

where the real part of p must be sufficiently large to

§l. Definition and elementary properties 5

ensure that all of the integrals converge. A complementary

result can be obtained by considering

Y[t- l f(t)] = Jo e-pt !ill dt

OO

(19)

= rp
F(q) dq,

which is valid provided the integrals exist. Both of these

procedures may be iterated to give more general results,
which we will not list here.

Translations: Let T > 0 and suppose that f(t) o for

t < 0, then

Y[f(t-T)]

fOO e-p(t'+T) f(t') dt' (20)

o
e -PT ,Sf[f (t)] .

This result applies to translations to the right; in parti-

cular the inverse Laplace transform of exp(-pT)F(p), where
F(p) = .sf[f(t)], will give f(t-T) for t > T and zero for
t < T. For translations to the left we have

Y[f(t+T)] = f OOo e- P t f(t+T)dt

foo e-p(t'-T) f(t') dt' (21)

J:
T

e PT .sf[f(t)] - eP(T-t')f(t') dt'.

The finite integral cannot be neglected unless f(t) 0

 6 PART I: THE LAPLACE TRANSFORM

for t < T, as it accounts for the part of the function which

has been 'lost' by translation to negative t values where
the Laplace transform does not operate.

Convolutions: The convolution of two functions fl(t) and

f 2 (t) is defined by

get) = J: fl(T) f 2 (t-T) dT. (22)

Now we take the Laplace transform of get), and by chang-

ing the order of integration and writing t' = t-T, we ob-
tain
G(p) fo e -pt
CO
dt
ft
o
fl (T) f2 (t-T) dT

(23)
pt
f CO e-pT f (T) dT Jco e- ' f (t')dt'
o 1 0 2

= Fl (p) F2 (p).

Thus the transform of a convolution is simply the product of

the individual transforms--a result which is of considerable
importance. Obviously this result can be iterated to obtain
a connection between an n-fold convolution of n functions
and the product of the transforms of these functions.

Simple Applications:
(i)
1 - iwt
.5f[sin wt] 2i .5f[e ]
(24)
§1. Definition and elementary properties 7

(i i)

..st'[ co s wt]
(2 S)

(iii)
.,St[t sin wt] ~p .st'[sin wt]

2pw (26)

(iv)
-at w
..st'[e sin wt] (27)
(a+p)2+w2
[by replacing p by a+p in (24)].

r
(v)

..st'[t -1 sin wt] .st'[sin wt] dq

p

JP
OO w dq
q2+w 2
(28)

arc tan (w/p)

(vi) Let

then
Si (z) r0
sin t dt
t '
(29)

..st'[Si (z)] p -1 ..st'[t-1 sin t]

P -1 arc cot (p). (30)

Less trivial applications of the properties of the

transform, particularly in the solution of differential equa-
tions and integral equations of convolution form, are the
subject matter of Sections 3-S.
8 PART I: THE LAPLACE TRANSFORM

1.3. Asymptotic properties: Watson's Lemma

Consider equation (1) for large p. By inspection,
it seems reasonable to assume that the only significant re-
gion of integration is 0 < t < lip, so that we could write
as an approximation

F(p) '" f(O) f

OO

0 e
-pt
dt
(31)
= f(O)/p, p » 1.

Such information, linking properties of functions and their

transforms directly, may be very useful in application. How-
ever, the example given in equations (7) and (8), where

F (P) - ~
y+l'
- p » i3 (32)
P
shows that we need a sharper result than (31).

Definitions: If two functions f(x) and g(x) satisfy the

relation
lim [f(x)/g(x)] = 1, (33)
x-+-xO
then we say that they are asymptotically equal as x -+- x O'
and write
f(x) ~ g(x), x -+- xO' (34 )

In the event that x is a complex variable, we may need to

add some restriction about the way in which x approaches

x O' for example

-z 1, z 00,
1 + e ~ -to

larg(z)1 < 'IT /2 . (35)

If now (33) is replaced by the condition that

§l. Definition and elementary properties 9

lim [f(x)/g(x)] 0, (36)

x .... x O

then we write
f(x) = o(g(x)), (37)

and if If(x)/g(x)1 is bounded as x approaches x o' then

we write
f(x) = &(g(x)). (38)

In this book we shall frequently use the notations (34) and

(38); the small 0 notation (37) will not occur often.

Asymptotic Expansion: An expansion of the form

00

f(x) ~ L gv(x), (39)

v=l
is called an asymptotic expansion if

( 40)

The meaning of such an expansion is that

n
f(x) = v~l gv(x) + &(gn+l(x)) (41)

so that a finite number of terms of the series gives an ap-

proximation to the function f(x) of "order" gn+l (x) when
x approaches x . Viewed as an infinite series (39) may be
o
convergent or it may be divergent.

Watson's Lemma: We will now state and prove an important

result, of which (31) is a special case, linking the asymp-
totic expansion of a function f(t) about t = 0 with the
asymptotic expansion of pep) as p .... 00. Suppose that f(t)
has the asymptotic expansion
10 PART I: THE LAPLACE TRANSFORM

fCt) 'U t-+-O ,

( 42)

then FCp) has the corresponding asymptotic expansion

co
F (p) 'U r
v=l
Ipl-+-co,

-n/2 < arg(p) < n/2. ( 43)

Note the effect of the restriction larg(p)1 < n/2 is to en-

sure that Re(p) becomes infinite as Ipl does in this sec-
tor. To derive the stated result, we introduce the function

(44)

in terms of which F(p) is given by

F(p) (4S)

To compute bounds on we choose positive numbers

tn and Kn so that
Re (;I. )
< K
n
t 'J , (46 )

Also, we know that there must be some real value a for

which the integral defining F(p) converges, and we use this

constant to define the functions

n
<Pn(t) = f: e
-CIS
(47)

The importance of the choice of a is that the functions

~
~n
will be bounded, and we write An for the maximum value

of I cf>n (t) I , t
n -
< t < co. Using these definitions, we can
§1. Definition and elementary properties 11

break up the integral defining 3lf[f (t)) into two parts,

n
and calculate the following bounds:
Re (A ): K
< n n
Re(X n )+l
[Re(p)]
(48)
-A -1
O'(p n )
and

If: e- pt fn(t) dtl I(p-a) J: e-(p-a)t $n(t)

n n ( 49)
~Ip-al
< exp {-(Re(p)-a)t }.
Re (p)-a n

This latter integral tends to zero exponentially as p tends

to infinity in the given sector, consequently (48) shows
that
- A -1
Y[fn (t)) = O'(p n ) (50)

and Watson's lemma (43) is proved. The result must be used

with caution. It gives information about the behavior of
F(p) for large p which is consequent upon the behavior of
f(t) for small t. The question of a converse implication
is discussed in Sections 2 and 6.

Problems
Deduce the following general relationships. 3

1. If f(t+T) = f(t). t > 0 where T > 0 is a constant.

F(p) = Cl-e -T p) -1 IT0 e- P t f(t) dt.

1 d
2. If f(t) = t dt get).

FCp) = 100

p
q G(q) dq.
12 PART I: THE LAPLACE TRANSFORM

If f(t)
t u
-1
g(u) du,

r
3.
0
F (p) p-l G(q) dq.

4. If f(t) r t
P
u- l g (u) du,

J~
F(p) -1
p G(q) dq.

5. Joo t- l f(t) dt = foo F(p) dp

o 0
Find the Laplace transforms of the following functions. 4
6. sinh (at)
7. cosh (at)
8. cos (at) cosh (bt)
9. t- l sinh (at)

10. sin (at l / 2 )

12. tV cos (at), V > -1

Using Problem 5, evaluate

14.
I: t- l sin(wt) dt

15.
I: t
-1
{e
- at
-e
- bt
} dt, a > 0, b > O.

By taking the Laplace transform with respect to t, evaluate

the integrals

16.
I: x sin (xt) dx
1+x 2

17. f: r
l
t21 dx.
exp _x 2 -Z
x J
~1. Definition and elementary properties 13

18. Let
{ t p-1 , t > 0, Re (p) > 0
f(t)
o , t < 0,

get)
= L: :r q-1 t >

t < 0,
0, ROe (q) > 0

h(t) =
r
'J
(E - 1) ! (9. - 1) !
(p+q-1)! t
p+q-1
, t > 0

o.
L 0 t <

Show that the Laplace transform of the convolution of f(t)

with get) is equal to ~[h(t)]. Hence derive the
formulaS

-1 q-1 (p-1)!(q-1)!
J1 xP (I-x) dx
o (p+q-l)!

19. Show that

(-1) v-1(v/2_1)!
(Xl

):
v=l 2p(v/2+1)

20. Show that

-1/2 -1/2
~[t (t+2) ] 'V

(-l)v 1 2 .3 2 .5 2 .... (2v-1)2

L v
v=O v! (8p)
14 PART I: THE LAPLACE TRANSFORM

Footnotes
1. The results given in this section may be found in many
places. We mention in particular DITKIN & PRUDNIKOV
(1965), DOETSCH (1971), and WIDDER (1944).
2. AHLFORS (1966), Ch. 5.
3. Many more general relationships may be found in ERDELYI,
et al. (1954), Ch. 4.
4. Extensive tables of Laplace transforms are available; for
instance, ERDELYI, et. al. (1954).
5. Anticipating the result that the Laplace transform has a
unique inverse.
§2. The inversion theorem 15

§2. THE INVERSION THEOREM

2.1. The Riemann-Lebesgue Lemma

As necessary preliminaries to a statement and proof
of the inversion theorem, which together with its elementary
properties makes the Laplace transform a powerful tool in
applications, we must first take note of some results from
1
classical analysis. Suppose that f(x) is a function con-
tinuous on the closed interval a < x < b (and hence uni-
formly continuous). We will investigate the asymptotic
properties of the integral
b
I(w) = fa f(x) e
iwx
dx (1)

for large real w. By some trivial changes of variable we

can write

fa +7T / w f(x) iwx b iwx

I (w) =
a
e dx +
fa +7T / w f (x) e dx

fa +7T / w
a
f(x) e
iwx
dx - t- a
7T / W
f(x+7T/W) eiwxdx (2)

and
I (w) t-
a
7T / W
f(x) e
iwx
dx +
b
Jb - 7T / W f(x) e iwx dx, (3)

and thus
1 fa+7T /w il'lx 1 iwx
I (w)
2 a
f(x) e dx +
2 J:- f(x) e dx

t-
7T / W
(4)
l 7T / W iwx
+ [f (x) - f (X+7T /w)] e dx.
2 a
It is easily seen, by a mean value theorem for integrals,
that the first two integrals in (4) are functions of asymp-
-1
totic order w furthermore, since f(x) is uniformly
continuous, we can make the integrand in the third integral
16 PART I: THE LAPLACE TRANSFORM

arbitrarily small by choosing w sufficiently large. Thus

we have proved that

lim
w+oo
Jba f(x) e
iwx
dx = 0, (5)

which is known as the Riemann-Lebesgue lemma.

Infinite Interval: The extension of (5) to the case where

one limit or both may be infinite will also be needed. For
example, if f(x) is a function defined on the interval
o< x < 00, for which

r o
I f(x) 1 dx (6)

converges, then we can write

r o
f(x) e iwx dx Ja
o
f(x) + E,

r
(7)

IE I ~ I f (x) I dx,
a
and because of the absolute convergence (6), it is possible
to make I EI arbitrarily small by a suitable choice of a.
Using (5) on the finite integral, we have its extension to
the infinite integral, i.e.,

iwx
lim J OO
f(x) e dx o. (8)
w+oo 0

Dirichlet Conditions: We say that a function f(x) satis-

fies Dirichlet's conditions in the interval a < x < b if
it has at most a finite number of maxima, minima, and points
of discontinuity in the interval, and takes only a finite
jump at any discontinuity. The importance of the Dirichlet
conditions to the theorems which we need is that they enable
the interval a < x < b to be divided into subintervals, in
~2. The inversion theorem 17

each of which the function is both uniformly continuous and

monotonic. This latter property allows us to use the second
mean value theorem for integrals, which states that if f(x)
is a monotonic function and g(x) a continuous function on
the interval a < x < b, then there is a point c in the

r
interval such that

a
f(x) g(x) dx = f(a) ra
g(x) dx + feb) t
c
g(x) dx. (9)

Returning now to equation (1), if f(x) satisfies Dirichlet's

conditions, then we can take the interval a < x ~ b to be

one of the subintervals in which it is monotonic and continu-

ous; then the integral is equal to

f(a) fC
a
e iwx dx + feb)
r c
e
iwx
dx (10)

6'(w - 1) , w + 00

For an arbitrary interval, we must add up a finite number of

such results, and so equation (5) is replaced by the much
stronger condition

fb.
a
f(x) e 1WX dx = 6'(w
-1
), W + 00. (11)

Note however, that we may not set a = -00 or b = 00 in this

result without imposing restrictions on f(x) in addition to
the convergence of equation (6).

2.2. Dirichlet Integral s

In addition to integrals of the form (1), we must con-

sider what are known as Dirichlet integrals, viz.,

Jba f(x) sin(wx) dx

x
(12)
18 PART I: THE LAPLACE TRANSFORM

in the limit that w tends to infinity. Suppose now that

a = 0, b > 0, and f(x) satisfies Dirichlet's conditions on
o < x < b. Choose c so that f(x) is continuous and mono-
tonic on the interval 0 < x < c; then an application of the
Riemann-Lebesgue lemma shows that

lim Ib f(x) Sin!w X ) dx

w+oo 0
= lim
w+oo
I C
0
f(x) sin!Wx) d~ (13)

and in addition, we can use the second mean value theorem to

I:
write
f(O+) sin~Wx) dx

[f(x)_f(O+)]sin(w x ) dx
x (14 )
f(O+) Io sin(wx)
C
x
dx

+ [f(c)-f(O+)] JC sin(wx) dx,

h x

where 0 < h < c. It is a standard result that the integral

sin x
dx (15 )
x

is convergent, and has the value n/2. Consequently, the

I: Sin~wx~
expression
I dx (16)

is bounded by some constant M, so that we can write

lim
w+oo
fco f(x) sin(Wx) dx
x
1l f(O+) +
2
£

(17)
1£1 < If(c) - f(O+)IM

We can make £ arbitrarily small by choosing c so that

§2. The inversion theorem 19

If(c)-f(O+) 1M is sufficiently small, and this does not af-

fect the restriction previously placed on c. Consequently

lim Ib f(x) sin(wx) dx = ~2 f(O+). (18)

w+oo 0 x

By a similar argument, it can be shown that

lim
w+oo
Io -b
f(x) sin(wx) dx
x
~f(O-). (19)

Finally we note that these results are unchanged if b is

set to infinity, since the added integral tends to zero in
the limit w + 00 by the Riemann-Lebesgue lemma.

2.3. The Inversion Integral

Let f(x) be a function with the Laplace transform
F(p) for which the defining integral

F(p) = I: f(x) e- PX dx (20)

converges in the half plane Re(p) > c. Consider the inte-

gral
1
2ni Iy-iR
Y+iR
e PX F(p) dp, Y > c. (21)

We substitute for F(p) the integral (20), and interchange

the orders of integration (an operation which is valid be-
cause (20) is uniformly convergent with respect to p when
T > c), to transform (21) to the formula

IR(x) I: fey) dy Z;i r+iR eP ( x-y )

y-iR
dp

1
"if I: fey) e y(x-y) sin R(x-r) dy
x-y (22)

1 e- Yu sin Ru duo
"if ex f(x+u) u
20 PART I: THE LAPLACE TRANSFORM

If we break the integral into two, from -x to zero, and

zero to infinity, and allow R to become infinite, there are
three possibilities; namely,
0 ,x < 0

IR(x) + { if(O+), x = 0
t[f(x-O) + f(x+O)], x > O. (23)

This result is generally known as the inversion theorem for

Laplace transforms, and is expressed by the reciprocal pair
of equations

FCp) r0
f (x) e- Px dx, Re(p) > c, (24)

1
fCx) = 21T i r+~oo F(p) e Px dp, Y > c, (25)
Y-l 00

1
where f(x) is taken as Z[f(x-O) + f(x+O)] at a point of
discontinuity.

2.4. Inversion of Rational Functions.

In many situations it is necessary to calculate the
inverse Laplace transform of a rational function

F(p) = ~
B (p) ,
(26)

where
A(p)

B (p) = (27)

and n > m. The need for such inversions arises particularly

in the solution of equations with constant coefficients (Sec-
tion 3), and in techniques of rational approximation (Sec-

tion 22). We commence with the integral

§2. The inversion theorem 21

1
hi JC A(p) e Px dp,
B(p)
(28)

Im(p)

Re(p)

C=L+ r

Figure 2.1

where the contour is shown in Figure 1. We are only concerned

with the case x > 0, and since n > m, we can obtain a bound
for the integral around the semi-circle by writing

ia
p = y - Re , -Tl/2 ~ a ~ Tl/2 (29)

and taking R sufficiently large so that IF(p)1 < K/R.

Then

I Jr &12l
B(p)
e Px dPI < K JTl/2 e- Rx cos ada
-Tl/2

O'(R- l ), R .... co. (30)

Thus the integral (28) is the inversion integral in the

limit R .... co. Since the only singularities of the integrand

are poles produced by the zeros of the denominator, we find

22 PART I: THE LAPLACE TRANSFORM

that the original function f(x) is the sum of the residues

at these poles.
The evaluation of these residues can be calculated by
decomposing F(p) into partial fractions and then inverting

each term. In particular, if the roots aI' a 2 , a 3 , ... , an

of B(p) are distinct then we can write immediately
n A(a i )
F (p) L
i=l B' (a.) (p-a.)
1 1
(31)
n A(a i ) ai x
f(x) = L e
i=l B' (a.)
1

obtaining an original function which is a sum of exponentials.

A similar, but more complicated analysis can be made
if B(p) has multiple zeros. For simplicity, we concentrate

here on the contribution from one such root, as the extension

to the general case is trivial in principle but tedious in
practice. If a is a root of multiplicity m, then the
partial fraction expansion of F(p) will contain the terms

Y2
+ + ... + (32)
(p_ a) 2

and the corresponding contribution to the original function

f(x)
Y2
-
1: x + -
Y3
2: x 2 + ••• +
(m-l):
Ym
x J
m- II
e
ax
. (33)

An Example: If
f (p) = 1 (34)
l_p4

then a straightforward partial fraction decomposition leads

to
-1 + 1 i + i
F(p) (35 )
'47(p=--'l") 4 (p+ 1) 4 (p - i) 4 (p+ i)

and
~2. The inversion theorem 23

f(x) "2I sin x - I2 sinh x. (36)

2.5. Taylor Series Expansion

For small values of x, it may be more appropriate to
have an expansion for f(x) as a Taylor series, instead of
a cumbersome expression involving a complete knowledge of
the roots of the polynomial B(p). Returning to the integral
(28), we assume that R is so large that all of the poles
lie inside the contour, and then deform the contour to be a
circle, centered at the origin, with all of the zeros of
B(p) still on the inside. On this circle we may expand
F(p) in a convergent power series in inverse powers of p;

term by term evaluation of the integral will then give the

Taylor series 2 for f(x). An example will demonstrate the
technique. Suppose that

F(p) (37)
then
FCp) (38)

and term by term inversion on a contour which is a circle of

radius more than unity gives

(_l)n x2n+1
f(x) = l: (39)
n=O (2n+l) !
sin x.

For rational functions, then, there is a converse to Watson's

lemma: an expansion in inverse powers of p which is both

asymptotic and convergent for large p implies an expansion

of the original function in powers of x, again both asymp-
totic and convergent. In the example chosen here (37-39)
24 PART I: THE LAPLACE TRANSFORM

all of the series are elementary and can be written out in

full; however, knowledge of the first few terms of the one

expansion is sufficient to construct the first few terms of
the other in more difficult problems.

Problems

1. Show that if f(x) has derivatives up to fen) (x), and

if fen) (x) is absolutely integrable, then

. wnJ
l~m
b
f(x) e iwx dx = 0
W·HO a
where a and/or b may be infinite.

Find the inverse Laplace transforms of the following func-

tions using the inversion integral.

\p+]J
2.
(p+ a) 2

\p+]J
3.
(p+a) (p+b)

4. Show that if F(p) has the expansion

an
F(p) = l: n+l
n=O p
which is convergent for Ipl > R, then the inverse func-
tion has the power series expansion
00

f(t) l:
n=O n!
§2. The inversion theorem 25

Footnotes

1. For a thorough treatment of the material in Sections 2.1-

2.3, see, for example, APOSTOL (1957), Ch. 15.

2. Often known as the Heaviside series expansion. See

Section 6.5 for the general case.
26 PART I: THE LAPLACE TRANSFORM

§3. ORDINARY DIFFERENTIAL EQUATIONS I

3.1. First and Second Order Differential Equations

Linear differential equations with constant coeffi-
cients are an important area of application of the Laplace
transform. As a prelude to the discussion of such problems
we discuss first two particularly simple examples, since the
connection with the classical methods of solution is readily
apparent in these cases.

First Order Equations: Consider the initial value problem

y' (t) + by(t) f (t) , t > 0,

(1)
Y (0)

which can be solved by using the integrating factor exp(bt)

to give
-bt
yet) = Yo e (2)

Now we take the Laplace transform of (1) ; after applying

(1.15) we have

[p yep) - yO] + b Y (p) = F (p) , (3)

which is an algebraic equation. yep) is found immediately,

viz. ,
yep) G(p) [y + F(p)],
o (4)
G (p) [p + br l ,

and this is obviously equivalent to the classical solution

(2) since G(p)F(p) is the transform of a convolution, and
[p + b]-l is the transform of exp(-bt). The advantage of
the Laplace transform over the classical method is not
§3. Ordinary differential equations 27

apparent from this simple example, however, it is interesting

to see how (4) gives a different emphasis from (2). In
particular, the function G(p), which contains information
about the analytical behavior of the solution, plays a
prominent role, while the initial value, which is no more
important than the function f(t), enters on an equal foot-
ing with that function and is incorporated from the outset.

Second Order Equations: Now we consider the second order

initial value problem

y" (t) + by' (t) + cy(t) f (t) , t > 0 ,

yeO) Yo (5)

y' (0) vo

This equation arises in many elementary applications which

2
may be found in standard texts. If we take the Laplace

transform, and again use (1.15), we obtain

pyO - vol .... b[p'r(p) - yO] + cY(p) F(p). (6)

The equation for yep) is algebraic, and can be solved

immediately to give

yep) G(p) [(p+b)yo + Vo + F(p)],

2 -1 (7)
G(p) = [p +bp+c] .

Once again, inversion gives the solution as a term depending

on the initial conditions plus a convolution integral. In

some cases it is more convenient to invert the function

G(p)F(p) directly, rather than write it as a convolution
and evaluate the latter; nevertheless, the general form of
28 PART I: THE LAPLACE TRANSFORM

the solution is important for understanding the role of G(p),

An analysis of (7) depends on factoring the quadratic ex-
pression p2+ bp +c ; two different cases emerge:

2
Unegual Roots: If P + bp + c = (p - 0. 1 ) (p - 0. 2 ) with 0. 1 f

0. 2 '
then we can write

G(p) = 1
0. 1 -0.2 G~o.l - P~o.2J, (8)

yep)
_1_
0. 1 -0. 2 t
rco. l +b)yO+vO
P-\ -
(o. 2+b)Yo+v;l
p-o. 2 J + G(p) FCp),

Inversion of the various terms then gives

yet) [(o.l+b)yo + vO]gl(t) + [(o. 2+b)yo + v O]g2(t)

r
=

+ g(t-T) f(T) dT,

o

(9)

2
Egual Roots: If p2 + bp + c (p-o.) , then we have

1
G(p)
(p _a.) 2 '
(10)
Yo (o.+b)yo + vo
Y (p) = + ----"--,2,.--=- + G (p ) F (p) ,
p-o. (p-o.)
and inversion gives
§3. Ordinary differential equations 29

y (t) Yo g (t) + [ (a+ b)y 0 + v 0] h(t)

+
t
f
h(t-T) f(T) dT,
0

g (t) eat ,
(11)
at
h(t) t e

General Features: Two important features are implicit in

these results. First, information on the behavior of f(t)
for large t is contained in the function G(p), since its
poles determine both the functions which determine the de-
pendence on initial conditions, and the influence function
which is used in the convolution integral. If both poles of
G(p) have negative real part, then the influence of the
initial conditions dies out for increasing time. Also in
this case the influence of the function f at time T on y
at time t (t > T) diminishes as t becomes large. The
second feature of note is that yet) is continuous even if
f(t) has discontinuities. This property follows from the
fact that yet) depends on f(t) via an integral with a
finite integrand, so that it is automatically continuous.

3.2. Higher Order Differential Equations

The analysis of the nth order differential equation

a yCn) (t) + a y(n-l) (t) + ... + aoy(t) = f(t), t > 0 (12)

n n-l

proceeds in a similar fashion. The Laplace transform of (12)

gives the algebraic equation

n k k-l k-t-l (t)

l a [p yep) - l p y (0)] + aOY(p) F(p) (13)
k=l k t=O

which can be reversed to give

30 PART I: THE LAPLACE TRANSFORl~

yep) G(p) [H(p) + F(p)],

G (p) LL akP~-l (14)

n-l n
H(p) = I y(p)(O) ) a k pk-~-l
~=O k=t+l

where we have defined a polynomial H(p) which contains all

the information about initial conditions. A formal solution

to (12) can now be found by inverting the functions G(p)

and G(p)H(p), namely

t
yet) y-l[GH] + fo g(t-T) f(T) dT,
(15 )
get) =y-l[G],

which exactly parallels the solutions given in (2), (9), ~nd

(11) for first and second order equations.

Stability: The inversions involved in (15) are of rational

functions, and have been considered in Section 2.4. A most
important question is that of the stability of the solution,
that is, whether the function yet) increases without bound
for large time without a corresponding increase in the driv-
ing function f(t) to cause this behavior. This asymptotic

behavior depends solely on the position of the poles of

G(p) in the complex plane, for if we turn off the driving

force at some time T > 0, then we can write (15) as

T
yet) = y-l[GH] + fa g(t-T) f (T) dT. (16 )

We know from Section 2.4 that the inversions of G(p) and

G(p)H(p) have the general form

~3. Ordinary differential equations 31

Ct.t
.,st'-l[G] = .. t i e J ,
~ g 1J
i ,j
(17)
Ct.t
.,st'- 1 [GH] .. t i e J
~ h 1J
i,j
where the poles of G(p) are at p = Ct. , and on substitution
J
into (16) it is readily seen that the large time behav:ior of
the contribution of the pole at Ctk is determined as follows:

Re (Ct k ) < 0, exponentially damped;

Re (Ct k ) > 0, exponentially growing;
Re (Ct k ) 0, bounded i f the root is simple,
otherwise unbounded.
For stability, we want the solution to remain bounded, hence

all the poles must be in the left-hand half-plane, except

possibly for simple poles on the imaginary axis.

Transfer Function: Our analysis of (12) shows that G(p)

plays a central role in determining yet). In many physical
applications, the function f(t) represents an input to a
system, and the corresponding response is measured by yet).
The relation
yep) = G(p) F(p) (18)

can be represented diagrammatically as

F
G .,v

Figure 3.1

where the box labelled G represents a linear system de-

32 PART I: THE LAPLACE TRANSFORM

scribed mathematically by G. G is known as the transfer

function in this context, and this is a most important con-

cept in the analysis of linear systems. Suppose now that
the function yet) is used as the forcing function in an-
other system. Then we have, for the second system,

Gl(p) yep)
(19)
Gl (p) G(p) F(p),

and this may be represented by the diagram

--~F~~~L- ____
G____ ~~-----~.Y------~~___ G_l____ ~~
Figure 3.2

In the analysis of more complicated systems, this method of

representation leads to a very clear formulation of the over-
all problem. For details the reader should consult special-
ized books on the subject. 3

An ExamEle: If

y(4) (t) + 4y (t) = sin t,

(20)
y(3) (0) y(2) (0) = /1) (0) yeO) 0,
then we get
1
yep) G(p)
p2+l '
(21)
1
G(p)
p4+4 .

Now we can factor yep) as

§3. Ordinary differential equations 33

yep) -i + i 3-i 3+i

lO(p-l) IO(P+l) 80(p-l-l) 80(p-l+l)
( 22)
+ 3+i + 3-i
80(p+l-l) 80(p+l+i)
from which we obtain yet) as

yet) = i sin t - io e t (3 cos t + sin t)

(23)
I -t
+ 40 e (3 cos t - sin t)

Note the fact that yet) grows exponentially for large t,

a fact which is evident from the factorization

(p2 + 2i) (p2 _ 2i)

(24)
(p-l+i) (p+l-i) (p-I-i) (p+l+i).

3.3. Simultaneous Differential Equations

As we have shown, the Laplace transform is an effec-

tive method for dealing with the solution of a single dif-
ferential equation with constant coefficients. However, the
full power and elegance of the method only become apparent
when it is applied to a system of simultaneous differential equa-

tions. Moreover, it is possible to gain an insight into prob-

lems with a comparatively small amount of calculation, es-
pecially as compared to the classical methods. We consider
first two examples to illustrate what can happen.

Exam~: Consider the mechanical system shown schematically

in Figure 3. Two springs of negligible mass and spring con-
stant k support two masses of equal mass m in the manner
shown. The bottom mass is attached to a linear damping de-

vice which applies a resistance proportional to velocity.

34 PART I: THE LAPLACE TRANSFORM

If we denote the (vertical)

displacements of the upper
and lower masses by YI(t)

and YZ(t), respectively,

taking a downward displace-

ment as positive, we can k

write the equations of

motion in the form m

my I + kYI - k(YZ-YI) 0,
k
(Z 5)
my z + cY z - k(YZ-YI) O.

After taking the Laplace trans-

m
form and rearranging, the equa-
tions become c
Figure 3.3
(pz+Zwz)Y (p) - WZYz(p) = PyI(O) + Yi(O),
Z I (Z6)
-w YI(p) + (pZ+yp+wZ)Yz(p) = (P+Y)Yz(O) + Yi(O),

where wZ = kim and y = c/m. Solutions for YI(p) and

Yz(p) are readily found by elimination, viz.,

YI (p) G(p)HI (p),

Y Z (p) G(p)H Z (p) ,
G(p) [(pz+wz) (pZ+yp+wZ)_w4]-I,

[pZ+yp+wZ] [PYI (0) + Yi (0)] (27)

+ wZ[(p+ )y Z(0) + YZCO)] ,

WZ[PYl(O) + Yi(O)]
+ [pz+zwz] [(p+y)y Z (0) + Y (0)] . z
Now the analysis of this expression may proceed analogously
to that of the single fourth order equation; for simplicity
§3. Ordinary differential equations 35

we assume here that the damping is,small, so that we can

write

2 2 -1 2 2 -1
G(p) (p +rlP+D l ) (p +r 2P+D 2 ) ,

D 2 .!. w2 (3+ / 5) '" 2.62 w2 ,

1 '" 2
1 2 2
2 '" 2" w (3 -15) '" 0.38 w ,
D 2 (28)

r l '" .!. y (1-1/15) '" 0.27 Y,

2

r 2 '" i y (1+1/15) '" 0.72 y.

Each of the quadratic factors has a pair of complex conjugate

roots with negative real part, hence the system is stable.
Also, the polynomials Hl(p) and
H2 (p) are of lower degree
than l/GCp), so we can find partial fraction decompositions
for YlCp) and Y2 Cp). As a simple example, we consider the
solution for YlCt) when the initial conditions are ylCO) =
YiCO) = Y2(0) = O. Then we have

where we have again used the conditions r l « Dl , r 2 « D2

to find the poles of G(p). Solutions for arbitrary initial
conditions can be found by similar algebraic manipulations.

Example 2: We consider the electrical circuit shown

36 PART I: THE LAPLACE TRANSFORM

E (t)
L

Figure 3.4

in Figure 4. Here we want to determine the voltage Ef

across RZ from a knowledge of the input voltage E(t) and
initial conditions. Ef is equal to i Z RZ' and must also
equal the voltage across L, which is L(ii - ii), since
il - iZ is the current through L. A further equation comes
from the fact that E is the sum of Ef and i l Rl . Put-
ting these facts down, we have
E i l Rl + Ef ,
L(ii - ii) (30)

i Z RZ•

Now we introduce the notation Yl

iI' YZ = = i Z' and elimi-
nate Ef from the problem in favor of i Z. Then (30) be-
comes the simultaneous pair of equations

RlYl + RZY Z = E,
(31)
LYi - LyZ - RZYZ = O.

On taking the Laplace transform, and writing E(p) .5t'[E] ,

we have
RlYl(P) + RZYZ(p) = E(p),
(3Z)
pL Yl(p) - (pL+R Z) YZ(p) = L[Yl(O) - YZ(O)].
§3. Ordinary differential equations 37

These equations can be solved for Yl and YZ to give

Y1 (p) G(p) [HI (p) + (pL-. RZ) E(p) ] ,

YZ(p) G(p) [HZ(p) +pLE(p)],

-1
G(p) [pL(Rl+R Z) + RlR Z] , (33)

HI (p) RZL[YI (0) - Yz(O)] ,

This time we have a solution in which new features appear,

which becomes evident if we try to write Yl or YZ as the
inversion of the initial value term plus a convolution involv-
ing E(t). For Yl ' this convolution would be
t
fo gl(t-T) E(T) dT,
(34)
gl(t) = ~-l[(pL + RZ) G(p)],

but the inversion integral for gl(t) does not exist in the
classical sense. The most straightforward way out of this
problem is to rearrange (33) by writing (pL+RZ)G(p) and
pLG(p) as quotients plus remainders, where each remainder
is a fraction with the denominator of higher order in p
than the numerator. Explicitly,

Yl(p) = i(;a + G(p)

1 Z
.(35)

YZ (p)

Inserting the functions G, HI' and HZ' we can now write

the formal solutions
38 PART I: THE LAPLACE TRANSFORM

E (t) + -at
e
Rl+R Z R +R
1 Z
R Z t
+ z J e-a(t-T) E(T) dT,
(Rl+RZ)ZL 0

~ _ Rl[Yl(O) - YZ(O)] e- at
YZ (t) (36)
Rl+RZ

This solution differs from previous solutions which we

have discussed in two important respects. First, it will not
necessarily satisfy the initial conditions which we try to
impose. In fact, i f we put t = 0 in' (36), we have the
relation RlYl(O) + RZYZ (0) = E(O), which is (3la). So
there is no real contradiction; the basic equations imply
that the possible initial values of YI , YZ' and E are re-
lated, and the solution is consistent with this restriction.
The other interesting new feature is the appearance of E(t)
as a component of the solution. This means that if E(t)
has a discontinuity at some time, then yl(t) and Yz(t)
will also be discontinuous at that time since the other term's
in these functions cannot be discontinuous if E(t) is fi-
nite. But this raises another problem, since we have assumed
that and are differentiable in writing down (31).
It is a remarkable fact, which we will consider in more de-
tail in Section 9, that the Laplace transform still handles

the situation correctly.

Alternative Formulation: We consider the last example again,

this time formulating the problem so that we do not differ-
§3. Ordinary differential equations 39

entiate a discontinuous function. If we define variables

E (t).
(37)

We can now eliminate u1 algebraically to get the first or-

der differential equation

R2
uZ(t) + au 2 (t) = E(t). (38)
(R 1 +R 2)L
However, the output voltage Ef = R2 i Z is given by
RR
E(t) - ~ u (t) (39)
Rl +R2 2
and this will exhibit the phenomenon of being discontinuous
wherever E(t) is discontinuous.

Normal and Anomalous Systems: We now consider the system of

differential equations
n n
l a .. y! (t) + l b .. y. (t) = fl.. (t), i = 1,2, ... ,n, (40)
j=l l.J J j=l l.J J
where the and b·l.J. are constants. Any set of differ-
a ..
l.J
ential equations with constant coefficients can be reduced to
this form without making assumptions of differentiability be-
yond those implicit in the original set. For example,
(25) can be written

my' + 2kY1 - ky = 0,
3 2
cy' + mY4 - kY1 + kY2 0,
2 (41)
y' - Y3 0,
1
yZ - Y4 0,
40 PART I: THE LAPLACE TRANSFORM

by introducing two extra variables and Now we take

the Laplace transform of (40), and get

L
j
(a .. p
1J
+ b .. ) Y.(p)
1J J
( 42)
H·=La .. y.(O).
1 j 1J J

The equations can be solved for Yi(p):

Y i (p) L G .. (p) [F. (p) + H.], (43 )

j 1J J J

where the functions G· . (p) are the elements of the inverse

1J
of the matrix C with elements a .. P + b ... Using Cramer's
1J 1J
rule, we can express Gij (p) as a ratio of determinants,

G.. (p)
1J
= (-1) i + j I CJ..1 (p) I / I C I , (44)

where is obtained from C by deleting row and

column i. In practice, Cramer's rule is unlikely to prove
a viable method of constructing the functions Gij ; our in-
terest here, however, is simply to discover their analytic
structure. In particular, we are interested to know the or-
der of the polynomials IC(p) I and Ic .. (p) I. Since they
J1
are sums of products of linear factors in p, the maximum or-
der is n for Ici and n-l for Icjil. In fact the co-
efficient of pn in ICI is obviously IAI, where A is
the matrix formed by a ij . This leads to the following dis-
tinction: if IAI t- 0, the system is said to be normal; if
IAI =0 it is said to be anomalous. For a normal system,
the functions G.. (p) are therefore rational functions
1J
which can decomposed as partial fractions, exactly as G(p)
was decomposed in Section 2.4. Consequently, the inverse
transforms g .. (t) may be defined, and the formal solution
1J
§3. Ordinary differential equations 41

to (40) is

-j=l)
r H.
n
g .. (t)
1)
+ r
n
Jt
gi·(t-T)
j=l 0 )
f.(T) dT.
)
( 45)

From this we see that the solutions are continuous for finite
inputs f. (t), and that no restrictions are placed on the
)

possible initial values by the solution.

Anomalous systems are different, as we have seen in
Example 2 above. Some of the functions Gij(p) cannot be
inverted as ordinary functions, consequently the solutions
may be discontinuous for discontinuous inputs. Moreover, the
solution will place certain restrictions on the initial

values. The reason for this is not hard to find; if IAI = 0

then there are non-trivial solutions of the homogeneous equa-
tion

i=l
f Cl.
1
a ..
1)
= 0, j = 1,2, ... ,no (46)

If we multiply (40) by these coefficients and sum, the deriv-

ative terms vanish and we obtain

r r
n n
Cli b i )· y).(t) = Cl.
11'
f. (t) (47)
i,j=l i=l
which is a linear relationship between the unknown functions
Yi(t) and the inputs fi(t). One possibility is to use
this relationship to eliminate one unknown from (40), which
will give a new system of n-1 equations. If necessary,
this can be repeated until eventually we obtain a normal
set.
42 PART I: THE LAPLACE TRANSFORM

3.4. Equations with Polynomial Coefficients

The Laplace transform can sometimes be used to obtain
solutions of ordinary differential equations with non-constant
coefficients, as we now show in connection with Bessel func-
tions. Bessel's equation for functions of order v is

J~ (x) + x1 J ~ (x) +
rII - xv2]2 Jv(x) = 0. ( 48)

Near the origin, the two linearly independent solutions of

+v
this equation have the asymptotic form x- , except when
v = 0, in which case 4 the second solution behaves like

£n x. We will consider only the functions of the first kind,

defined by
v
J (x) ~ (x/2) x + 0, ( 49)
v V!'

and for this purpose we make the substitution Jv(x)

-v
x fv(x) , leading to the new differential equation

x f" (x) - ( 2v-I) f' (x) + x f (x) = 0. (50)

V V V

On taking the Laplace transform, this second order equation

for fv(x) is converted into a first order differential equa-
tion for Fv(p), namely

( 51)

The point of the substitution of fv(x) for Jv(x) is that

this equation does not depend on fv(O) or f~(O). The
general solution of (51) is

Av
(52)

The constants A can be fixed by appealing to the relation-

v
§3. Ordinary differential equations 43

ship between the asymptotic forms of fv(x) for small x, and

Fv (p) for large p [equations (1. 4Z) and (1.43)], giving

ZV(V_ 1:.)'
2 •
(53)

Some discussion of the inversion of this Laplace transform

may be found in Section 6.

More General Applications: Direct use of the Laplace trans-

form on equations such as (4S) depends on first finding the
asymptotic form of the solution near the origin, and then
using this information in such a way that unknown quantities
such as f(O) are eliminated from ~he transformed equation.
Since these steps are only intermediate, the final result
being a particular integral representation of the desired
solution, it is better to write down the solution as an inte-
gral from the outset. This approach is generally known as
Laplace's method, and is the subject of Section ZO.

Problems

In Problems 1-11, solve the given differential equations. If

initial conditions are not stated, take them as arbitrary.

1. y' + y = 1, y (0) = Z.
Z
Z. y" + w y = cos (vt) , v " w.
3. y" + y = sin (Wt).
4. y" + 4y' + Sy = 1, y (0) = y'(O) = O.
5. y'" + y 1, y (0) y' (0) y" (0) O.
6. y'" + Y t, Y(0) y' (0) y" (0) O.
44 PART I: THE LAPLACE TRANSFORM

7. y' - z -t 2 ,
y - z' 2t e -t
8. x' y,
y' z,
z' x.
9. y" + 2z 0,

Y - 2z' O.
10. x" + ay' bx 0,
y" - ax' - by O.
11. x'n + a(x n + xn - l ) 0, n > 1,
x'0 + ax O = 0,
x (0) 0, n > I,
n
Xo (0) 1.

12. A constant voltage EO is applied from time t = 0 to

a circuit consisting of a resistor R and capacitor C
in series. The charge on C is initially zero. Find
an expression for the current i(t). Is it possible to
specify i(O) arbitrarily?

13. A constant voltage EO is applied from time t = 0 to

a circuit consisting of a resistor R, a capacitor C,
and an inductance L in series. Find an expression for

the charge on C for t > O.

14. For the same circuit as in Problem 13, find the charge
on C if the applied voltage is EO sin (wt).

IS. Express the solution to Problem 13 as a convolution if

the applied voltage is an arbitrary function E(t).
§3. Ordinary differential equations 45

16. Two flywheels of moment of inertia II and 12 are

coupled by an elastic shaft of stiffness k. The first
flywheel is coupled to a motor which applies a constant
torque from t = 0, the second drives a load which ap-

plies a torque proportional to the angular velocity.

Describe the motion of the system. (Neglect the moment
of inertia of the shaft.)

17. We wish to find the motion of a particle projected from

a fixed point on the earth's surface. Let the origin
of the coordinates be 0, the x-axis point east, the
y-axis north, and the z-axis opposite to the accelera-
tion due to gravity. Then the equations of motion are

x" 2w(y' sin A - z' cos A),

y" -2w x' sin A,
z" g + 2w x' cos A,

where A is the latitude of 0, and w the angular

velocity of the earth. Find the solution for arbitrary
initial values of x', y', and z'.

18. A motor produces a torque proportional to a control

voltage fed into a regulator. It is coupled to a fly-
wheel of moment of intertia I which drives a load
exerting a constant retarding torque N. If the regula-

tor control voltage is proportional to w-w O' where w

is the angular speed and wo a constant, find an ex-

pression for wet). Assume that w(O) = O.

46 PART I: THE LAPLACE TRANSFORM

19. Suppose that in Problem 16 the torque supplied by the

motor is proportional to a control voltage fed into a
regulator. Let the voltage be proportional to
where w is the angular velocity of the second f1y-
wheel. Discuss the stability of the system.
n
CHint: If the polynomial anP + a n-1 P
n-1 +
...
+ aO

has real coefficients, then the ratio -a fa is

n-1 n
equal to the sum of the real parts of the roots) .

Footnotes

1. A thorough treatment of the material in this section may

be found in DOETSCH (1971), Ch. 3.

2. A very large number of applications may be found in

THOHPSON (1957), Ch. 3.

3. See, for example, KAPLAN (1962) and VAN DER POL &BREMMER
(1955), Ch. 8.

4. Logarithmic terms appear in the second solution whenever

v is an integer. See Section 20.6.
§4. Partial differential equations 47

§4. PARTIAL DIFFERENTIAL EQUATIONS

4.1. The Diffusion Equation

As an example to show how the Laplace transform may
be applied to the solution of partial differential equations,
we consider the diffusion of heat in an isotropic solid body.
For temperature gradients which are not too large, it is a
good approximation to assume that the heat flow is propor-
tional to the temperature gradient, so that the temperature
u(E,t) must satisfy the partial differential equation

K-
au (1)
at
where K is a constant given by K oc/k, and 0 is the
density, c the specific heat, and k the thermal conductivity
of the material. (We have assumed that 0, c, and k are
all constants.) A method of solution of (1) for particular
problems is illustrated by the following two examples. l

Semi-infinite Region: Suppose that the body fills the region

x ~ 0, and that u is a function only of x and t. We

want to find u(x,t) if the temperature is initially equal

to TO' and the plane x = 0 is held at the temperature
Tl · For this problem the diffusion equation (1) reduces to
the form

au (2)
at
By analogy with Section 3, where we showed that the Laplace
transform reduces the problem of solving ordinary differen-
tial equations with constant coefficients to an algebraic
problem, we might expect that a similar method applied to a
48 PART I: THE LAPLACE TRANSFORM

partial differential equation would afford some simplifica-

tion. We introduce the Laplace transform with respect to
time
U(x,p) = Joo e- pt u(x,t) dt (3)
o
and take the transform of (2), to get

I 2
pU(x,p) - TO = - -d U(x,p). (4)
a dx 2
2
Thus the problem will be solved if we solve the ordinary dif-
ferential equation (4) subject to appropriate boundary con-
ditions. The boundary conditions on u(x,t) are:
(i) u(O,t) = TI and (ii) u(x,t) remains finite as x goes
to infinity. The boundary conditions on U(x,p) are obtained
by taking the transforms of these, so we have

U(O,p) (5)

U (x, p) finite, x -+- 00,

and the unique solution of (4) subject to these restrictions

is
(6)

To complete the solution we need the inverse Laplace trans-

form of (6), which we obtain in Section 6.2. The result is

u(x,t) = TO + (TI-T O) erfc (a x / 2 /t) , (7)

where the function erfc (x), known as the complementary

error function, is given by the integral

erfc (x) =
2
I1r rx
e- u 2 duo (8)
§4. Partial differential equations 49

Infinite Slab: We consider the case when the body fills the
region 0 < x < ~ and is initially at temperature TO' One
face (x = 0) is maintained at this temperature, while the
second face is supplied with heat from time t = 0 at a con-
stant rate H. We want to find the heat flow through the
first face as a function of time. Our partial differential
equation is again (2), and the Laplace transform is again
(4). The general solution of (4) is
TO
U(x,p) -- + A sinh (Bx/p) + B cosh (Bx/p) , (9)
p

where the constants A and B will be determined by the

boundary conditions. Since u(O,t) = 0 we immediately have
B = 0. To determine the other constant, and also the heat
flow through the face x = 0, we must consider the function
q(x,t) = -K au/ax, which represents the heat flow at an ar-
bitrary point. From (9), we can write for the transform of
q

Q(x,p) = C cosh (Bx/p) , (10)

where C -KBA/p. The boundary condition at x = ~ is

q(~,t) -H or Q(~,p) = -H/p, which determines C. We re-
quire the heat flow through the face x = 0; the Laplace
transform of this quantity is

Q(O,p) = H (11)
p cosh (B~/p)
This function may be inverted using the methods of Sections
6.2 or 6.1. The results are, respectively,
00
(_l)r-l erfc 1
q(O,t) = 2H L {B~(r-2)/lt},
r=l (12)
00 r
4 1 2 n2t/B2~2}].
H[l+ TI L (-1) exp {-(r--)
2
r=l (2r-l)
50 PART I: THE LAPLACE TRANSFORM

The first expansion is useful for computation for small t,

while the second is a good expansion for large t. In par-
ticular, we see that q(O,t) + H for large t, as it must.

4.2. Wave Propagation

The simplest continuous vibrational system is a uni-
form flexible string of mass P per unit length, stretched
to a tension T. If the string executes small transverse
vibrations in a plane, then the displacement u(x,t) must
satisfy the partial differential equation

aZu aZu f (x, t)

a2 + P , (13)
dt Z ax Z
Z
where a T/p and f (x, t) is the external force per unit
length. In addition, u (x, t) must satisfy boundary condi-
tions, which depend on the manner in which the string is sup-
ported. We consider two simple problems involving (13) to
further illustrate how the Laplace transform may be applied
to such equations.

Semi-infinite String: We consider (13) with f =0 for the

region x ~ 0, with u and au/at initially zero. For
t ~ 0, the end x = 0 is subjected to the time varying dis-
placement ¢(t). The Laplace transform of (13), together
with the boundary condition at x = 0, give the equations

(14 )
U (0, p) = <P (p ),

and the solution which is bounded for x + 00 is simply

U(x,p) = <p(p) e -px/a . (15)

§4. Partial differential equations 51

The corresponding displacement is easily found using the

translation properties of the transform, viz.,

u(x,t)
{:(t - x/a), x < at
(16 )
x > at.

Hence the displacement which is imposed on the end propagates

down the string at velocity a.

Finite String: Suppose now we fix the string at x = t,

while still subjecting the end x = 0 to an arbitrary time
varying displacement. Then we must solve (14) subject to
the additional boundary condition U(~,p) = O. The solution

is easily found to he

sinh[p(~-x)/al
U(x,p) = <I>(p) (17)
- sinh[p~/al

This function may be inverted to find u(x,t) in a manner

which describes the physical picture very well: By replacing
the hyperbolic sine functions by exponentials and expanding
the denominator, we get

-px/a p(x-Z~)/a -p(x+U)/a

U(x,p) = <I>(p)[e -e +e

-p (x- H) / a
-e + ••• l,
(18 )
u (x, t) cp(t-x/a) ¢(t+(x-2~)/a) + CP(t-(x+Z~)/a)

- ¢(t+(x-4t)/a) + ... ,
where we assume in writing the expression for u (x, t) that
¢(t) = 0 for t < O. This result represents the propaga-
tion of ¢ (t) at velocity a while t < t/a; however, the

wave reaches x = t at this time and a second term,

-CP(t+(x-Zt)/a), begins to contribute. This represents the
52 PART I: THE LAPLACE TRANSFORM

reflection of ~ at the boundary, with the same amplitude

but opposite sign. Further reflections occur at t = Z'X-/a,
3t/a, etc., as the disturbance travels up and down the string.

Infinite Transmission Line: Coaxial cables and other e1ec-

trica1 transmission lines are systems which are amenable to
simple analysis via the Laplace transform. We consider here
a rather simple example: a line which has no resistive
losses. The line is described by two parameters, the induc-
tance L and capacitance C per unit length. Consider a
small length ox of the line (Figure 1).

L 6x

i(x. t) i (x+6x. t)

e(x+8x,t)
e(x,f) C6x

o~--------~----------~0 T
Figure 4.1

By equating the currents and voltages at x and x + ox

and taking the limit ox + 0, we find that they must satisfy
the simultaneous partial differential equations
ae(x,t) L di(x,t)
ax at
(19)
ai(x,t) _ C ae(x,t)
ax at
Suppose now that we connect a voltage source ~(t)

at x = 0, commencing at t = O. For initial conditions we

take e(x,O) = i(x,O) = O. Taking the Laplace transform of
§4. Partial differential equations 53

(19), we get

oE(x,p) - pLI(x,p),
ax (20)
aI(x,p) - pCE(x,p),
ax
which must be solved subject to E(O,p) ~(p). The solution
which is bounded as x + 00 is

E (x,p) ~(p) e- px / v ,
(21)
v2 l/LC,

so that the disturbance ~(x,t) propagates at velocity v

exactly as for an infinite string.

Termination of a Finite Line: Suppose now that the line is of

finite length i, terminated by a resistance R (Figure 2).

<1»( t) Transmission line R

Figure 4.2

Again we apply a voltage ~(t) at x = 0, and look at the

way the signal propates from the source to the load R.
Equation (20) must be solved subject to the boundary condi-

tion E(i,p) = RI(i,p), which is Ohm's law for the load.

The solution now becomes

R cosh [p(i-x)/v] + Lv sinh [p(i-x)/v]

E(x,p) = ~(p) (22)
R cosh [pi/v] + Lv sinh [pi/v]

Now if R = 0, we recover (17); the physical interpre-

S4 PART I: THE LAPLACE TRANSFORM

tation is the same as before, the electric signal being re-

flected back and forth along the line. The solution for
R = 00 is similar; we will not pause to consider the details.
From a practical point of view, the purpose of a transmission
line is to transfer energy from the source to the load. We
therefore ask if R can be chosen so as to eliminate any

reflection at x = t, and it is evident from the form of (22)

that the choice R = Lv = I(L/C) is the only one which
achieves this aim, since we then have

E(x,p) = ~(p) e -px/v , (23)

which is equivalent to (21). It is of interest to calcu-

late in this case the ratio of voltage to current at x = O.

Using (23) in (20) we have E(O,p)/I(O,p) = R, but only if
R = I(L/C). Thus for this particular choice, the system,

transmission line plus load, appears to the voltage source

to be the load without an intervening line. The input sig-
nal is transmitted at velocity v without change of form or
loss of energy, and delivered to the load R without reflec-
tion. The quantity R = I(L/C) is known as the impedance
of the line, and the line is usually referred to as a Heavi-
side distortionless line.

Problems

1. The plane boundary (x = 0) of a semi-infinite body is

maintained at temperature f(t) from t = O. The body
is initially at a uniform temperature TO. Find an ex-
pression for the subsequent temperature u(x,t) at each
point in the body.
§4. Partial differential equations 55

2. Obtain explicit solutions to Problem I for the following

special cases:
o < t < to'
(i) f(t)
to < t,

(ii) f(t) = At.

3. Solve Problem I if, instead of maintaining the boundary

x = 0 at a fixed temperature, we supply heat at a rate
Q(t) per unit area. Give an explicit solution if Q(t)
is a constant. Find also an expression for the tempera-
ture at x = o.

4. The plane boundary (x = 0) of a semi-infinite body

radiates heat according to Newton's law (Q proportional
to ~T, where ~T is the temperature difference between
the body and the medium in contact with it). If the body
is initially at uniform temperature TI and the other
medium is at TO' find an expression for the temperature
distribution of the body at subsequent times.

5. Two semi-infinite bodies, initially at uniform tempera-

tures TI and T2 respectively, are brought into thermal
contact at t = O. Describe the subsequent equalization
of temperature.

6. A slab of thickness ~ is initially at temperature TO.

From t 0, one face (x = 0) is held at temperature
TI , the other (x =~) at T2 . Find the subsequent tem-
perature distribution. Give forms useful both for small
t and large t.
56 PART I: THE LAPLACE TRANSFORM

7. Solve Problem 6 if the face x = 0 radiates according

to Newton's law, while the face x = t is held at tem-
pera ture T1 .

8. Solve Problem 7 if the face x = t is supplied with

heat at a constant rate Q.

9. A solid sphere of radius a is initially at uniform tem-

perature TO' From t o the surface is kept at tem-

perature Tl . Find an expression for the temperature

distribution at subsequent times.

10. An infinite solid has in it a circular cavity of radius

a. It is initially at temperature TO; from t = 0 the
surface of the cavity is held at temperature Tl . Find
the subsequent temperature distribution.

11. Solve Problem 9 if the sphere is surrounded by a medium

of temperature Tl into which it radiates according to

Newton's law.

12. A stretched string, fixed at x = 0 and x = t, is

plucked at its mid-point and released (from rest) at

t = O. Find an infinite series solution for the subse-
quent motion of the string.

13. Show that

a > 0
1 fY+1.· 00
e
aA dA
2 1T l· y- 1oo a < 0

where Y > O. Hence obtain from the Laplace transform

solution of Problem 12 a direct picture of the displace-
ment of the string in the time interval 0 ~ t ~ 2 t /c.
§4. Partial differential equations 57

14. A capacitor CO' initially charged to potential EO' is

connected at t = 0 to a semi-infinite transmission
line with inductance L and capacitance C per unit
length. Find the distribution of current at subsequent
times.

15. A finite line of length ~ and parameters Land C

is terminated by a resistance R. If the end x = 0 is
connected to a constant potential EO from t 0, show
that the potential across the load at subsequent times
is given by

{ 0, 0 < t < t/v

e(t,t)
EO [1 - (~ : ~)n], (2n-l)t/v < t < (2n+l)t/v

z = IL7C ,
v = l/ILC".
16. A finite line of length ~ and parameters Land C,
is connected at x = 0 to a potential source E(t) in
series with a resistance R. The end x = ~ is open.
Find the potential at x = ~ for t > O. Is there any
value of R for which transmission takes place without
distortion?

17. A line of inductance L and capacitance C per unit

length also has resistance R and leakage conductance
G (G = the inverse of the leakage resistance) per unit
length. If the line is infinite, and if a potential
~(t) is applied at x =0 from t = 0, find an expres-
sion for the voltage distribution for t > O. In par-

ticular, study the case RC = LG.

58 PART I: THE LAPLACE TRANSFORM

18. Using residues, find a series solution for problem 15.

19. A line with RC = LG, of length 9" is short-circuited

at x = 9,. From t = 0 a potential ¢(t) = A sin (wt)

is applied at x = O. Find the voltage distribution in
a form which explicitly shows the role of reflections
at the ends.

20. A constant potential EO is applied from t o to the

end of a semi-infinite cable (a line with L G = 0).

Find the voltage at subsequent times.

21. Solve Problem 20 if the cable is finite and terminated

at x = 9, by a resistance RO.

Footnotes

1. Many more examples may be found in CARSLAW & JAEGER

(1941), Chs. 5-10.
§5. Integral equations 59

§5. INTEGRAL EQUATIONS

5.1. Convolution Equations of Volterra Type

Integral equations in which the unknown function ap-
pears in a convolution occur in some important situations.

The equation
b
g(x) = f(x) + \ Ja k(x-y) g(y) dy, (1)

where f(x) and k(x) are given functions and \ a given

constant, is an example of a Fredholm integral equation of
the second kind. (An equation of the first kind is one in
which the unknown function g does not appear outside the
integral.) If the upper limit of integration b is replaced
by the variahle x, then (1) is said to be of Volterra,
rather than Fredholm, type. By the change of variables
x' = x-a, y' = y-a, (1) may then be written
x'
g(x') = f(x') + \ Ia k(x'-y') g(y') dy'. (2)

General Method of Attack: Applying the Laplace transform to

(2) leads immediately to the algebraic equation

G(p) = F(p) + \ K(p) G(p) (3)

with the immediate consequence

G(p) F(p) (4)

1-\ K(p)

and inversion yields the solution. Another problem of fre-

quent interest in connection with (2) is the determination of
the resolvent kernel, that is, the function ret) such that

g(x) = f(x) + JX r(x-y) fey) dy. (5)

a
60 PART I: THE LAPLACE TRANSFORM

Now (4) may be written

G(p) = F(p) + 1 K(p) F(p), (6)

1-1 K(p)
so that ret) is the inverse Laplace transform of 1K/[1-1 K].
We illustrate these few comments with some examples.

Example 1: Consider the equation

s = Iso e s-t get) dt. (7)

The Laplace transformation yields

1 1
.. = -p-l G(p) '
~ (8)
P
hence E.:.l _ 1 1
G(p)
p2 - P - p2 '
(9)
get) l-t.

Example 2: Consider the equation

g(x) = 1 - IX (x-y) g(y) dy. (10)

o
Then
G(p) = ! - ~ G(p), (11)
p p

which gives for the solution

G(p)
(12)
g (x) = cos x.

Example 3: Let r be the resolvent of the equation

g (s) f(s) + 1 Iso e s-t get) dt. (13)

Then, denoting the Laplace transform of r by n(p), we can

easily obtain
§s. Integral equations 61

_A_
n (p)
p-A-l
-(A+l)s . (14)
r (s) A e

Equation (5) reads in this case

g(s) ,. f(s) + A ro
e(A+l) (s-t) f(t) dt. (15)

Example 4: A less trivial example is furnished by the equa-

tion
g(t),. 1
(1+t)2
+ a ft0 g(s) ds
(1+t-s)2 '
(16)

which occurs in the solution of a semi-empirical model for

1
subsidence caused by mining operations. We introduce the
-2
notation k(t) ,. (l+t) ; then for Re(p) > 0 we have

K(p) Jo e-pt
OO dt
(l+t) 2

1 - P fooo e -pt dt
(l+t)
(17)

JOO
e
-u
u
du
'
p

where the last step follows from the substitution u,.

p(l+t). The last integral is the exponential integral (see
Appendix C), so
(18)

which defines K in the entire complex plane cut along the

negative real axis. Laplace transformation of (16) now
gives
G(p) ,. K(p) + a K(p) G(p). (19)

Evaluation of get) will involve numerical approximations to

. 2
the inversion 1ntegral; however, we can deduce some impor-
62 PART I: THE LAPLACE TRANSFORM

tant information, particularly about the behavior of get)

for large t, from simple analytic information. Consider the
inversion integral for g (t) ,

g (t) 21Ti
1 r+ ico e pt K(p)
Y-ico I-a K(p)
dp, (20)

where Y > 0 is chosen so that the contour is to the right

of all zeros of the denominator I-a K(p). Now we know from
Appendix C that the exponential integral has a logarithmic
branch point at the origin. Furthermore, it may be shown
(Problem 6) that the imaginary part of K(p) is not zero
except for real positive p, so that I-a K(p) can only have
zeros for p > O. Again, for real p we have

r
K(O) 1,
(21)
K' (p) = - t e -pt k(t) dt < 0,
o
so that if a < 1, I-a K(p) has no zeros; if a = 1 the
origin is a zero; and if a > 1 there is one simple zero for
some real positive PO' We assign the analysis of the spec-
ial case a = 1 to the problems. For a < 1, we can deform
the contour in (20) to encircle the negative real axis.
Using the superscripts ± to denote the values of a func-
tion at p = -~ ± iE, ~ > 0, E + 0, we obtain in this

case
get) l(t)
(22)

and using the properties

-
[K (- 0 ]*
(23)
~5. Integral equations 63

r
we have
E; e-(l+t)E; dE;
I (t) = > O. (24)
o 11 - a K+ ( - E;) I 2

This integral can easily be bounded, since 11 - a K+I must

have a minimum value, and replacing the denominator by this
value we find that there is some constant A for which

A
I (t) < 2 . (25)
(1 +t)

Hence the solutions to (16) tend to ~ero for a < 1. If

a > 1, a similar analysis can be made, except that when we de-
form the contour we must pick up the residue at the simple
pole p = PO. (See Figure 1.) This gives

1m (p)

isolated pole
Re(p)

branch cut

Figure 5.1
Pot
Po e
get) = + I(t), a > 1 (26)
a (PO - a-I)
showing that the solution is exponentially growing in this
case.
64 PART I: THE LAPLACE TRANSPORM

5.2. Convolution Equations over an Infinite Range. 3

We consider the integral equation 4

f (x) A foo k(lx-yl) g(y) dy, x > 0

o (27)
-as
k(s) = e a > O.

Taking the Laplace transform of both sides and splitting

the integral over y into two, over y ~ x and y ~ x, we

ro r
obtain
pep) = A[G(p)K(p) + e- Px dx k(y-x)g(y) dy]. (28)
x

The double integral can be rearranged as

f: g(y) dy U~OO k(y-x)e- px dx - f~oo k(y-x)e- px dX]

G(p) K(-p) - foo g(y) dy foo k(x+y)e Px dx (29)

o 0
-1
= G(p) K(p) + (p - a) G(a) ,

providing all the integrals converge. S Substituting this

result into (28) and solving for G(p), we obtain

G(p) = A-lp(p) - (p-a)-lG(a)

K(p) + K(-p)
(30)
A(p+a)G(a) - (p2_ a 2)p(p)
2aA
The appearance of p2 p (p) shows that the solution for
g(x) will involve ~I(X); hence we rearrange (30) as

2 2
2aA G(p) = - [p pep) - pf(O) - fl (0)] + a pep)
(31)

- p [f (0) - A G(a)] - [f 1 (0) - aA G(a)] .

§S. Integral equations 6S

This expression has a convergent inverse only when the terms

of the form Ap + B are identically zero, i.e., only if

f (0) x G(a),
(32)
f' (0) aX G(a).

These restrictions may be obtained directly from the integral

equation, so it is not surprising that they occur as neces-
sary conditions for the convergence of the inversion inte-
gra1. Subject to these restrictions, it follows from (31)
that the integral equation has the explicit solution
2
g(x) = a f(x) - f"(x) (33)
2aX
although it must be noted that the solut~ons to this differ-
entia1 equation are not necessarily related by an integral
equation of the type (27); the initial conditions (32) are
also needed.

General Considerations: We consider an integral equation of

the form
a g(x) = bf(x) + X Joo k(lx-yl) g(y) dy, x > 0 (34)
o
where either a or b may be chosen to be zero. Laplace
transformation, followed by rearrangement of the double inte-
gral, yie1ds 6

a G(p) = b F(p) + X[G(P)K(P) + G(p)K(-p)

-f: g (y) dy f: k (X'y) ,px dj.( 35)

Now we suppose that the kernel function is a linear combina-

tion of exponential functions with polynomial coefficients,
namely
66 PART I: THE LAPLACE TRANSFORM

n -a..s
k (s) I
i=l
CP.(s) e
1
1

(36)
cp.1 (s) = ~i
j=O
Substituting a representative term from (36) into the double
integral occuring in (35) yields

fo g(y)
00

dy
foo
0
a iJ. (x+y)J e
. -a.. (x+y) + px
1 dx

a· . f j!
1J k=O k!(j-k)!
00

fo g(y)yJ-ke
• -a..y 00 k (p-a.i)x
1 fo x e dx

t
G(j -k) (a..)
a .. j! 1 (37)
1J k=O (j - k) ! (p_a..)k+l
1

that is, we obtain a partial fraction expansion whose coef-

ficients are constants, which are related to the Laplace
transform G(p) at the points p = a. i . To solve (35), we
temporarily regard these constants as arbitrary. Denoting
the double integral in (35) by N(p), we obtain the explicit
solution for G(p)
A N(p) - b F(p)
G(p) (38)
A[K(p) + K(-p)]-a
In general the inversion integral for this G(p) will not
converge, and this will restrict the constants, which were
temporarily assumed to be arbitrary, to certain fixed values.

Consequently (34) will lead not only to a functional rela-

tionship between f(x) and g(x), but may also give a set
of subsidiary conditions, as we saw in the example discussed
above.

A Further Example: We solve the equation

§5. An integral equation 67

g(x) f(x) + A I~ Ix-yl e-alx-yl g(y) dy,

o
(39)
x > 0, a > O.

The Laplace transform gives

( 40)

Solving for G(p), we obtain

G(p) = rF(p)
L
Yl
+ -- +
p-a
Y2
(p-a) 2
J
[1 + 'I'(p)],

'I' (p) ( 41)

(p2_a 2 )2 _ 2A(p2+a 2 ) .
Inversion of (41) is possible for any values of and Y2'
hence we have

g(x) = f(x) + (Y l +Y2x) e ax

-J:
( 42)

[fey) + (Yl+Y2x)eaYlw(x-y) dy.

It is not difficult to show that the constants Yl and Y2

are indeed arbitrary in this case (see Problem 18).

5.3. The Percus-Yevick Equation for Hard Rods

One of the central problems of statistical mechanics

is the determination of the pair distribution function. 7
Several approximate integral equations have been proposed to
determine this function; the most successful appears to be
the Percus-Yevick equation. We consider here the one-dimen-
sional form which can be written
68 PART I: THE LAPLACE TRANSFORM

~(x) Q2 + n foo ~(x') f(x')~(x-x') e(x-x') dx',

-00

( 43)
Q2 = 1 - n f~oo ~(x') f(x') dx',

where e(x) = exp [-SV(x)], f(x) = e(x)-l, S = l/kT, n is

the density, Vex) the interaction energy between a pair of
particles, k Boltzmann's constant, and T the temperature.
For hard rods of length a, we put Vex) 00, Ixl < a;
Vex) = 0, Ixl > a, so that e(x) and f(x) are step func-
tions.
Equation (43) determines a function ~(x) which is
related to the pair distribution function g(x) by

g(x) = ~(x) e(x). (44 )

In the ensuing treatment, we shall also employ the function

hex) = Hx) f(x), (45)

known as the direct correlation function. 8 Substituting

these definitions into (43) and taking the Laplace transform,
9
we have
2 roo 1
G(p) - H(p) = ~ + n
P
.sflf
~_CXl
g(x') hex-x') dx'J. (46)

The integral in (46) is subjected to manipulations similar

to, but more complicated than, those which we employed on
(34). We split it up into three regions: (i) x' < 0,
(ii) 0 < x' < x, (iii) x' > x, and deal with each in turn.
(i)

f: e- PX dx I~CXl g (x') hex-x') dx

I: e- Px dx r
x
h(x") g(x-x") dx" (47)

0,
§5. Integral equations 69

where we have used the variable change x" x-x', and the
properties
g (s) 0, Is I < a,
( 48)
h (s) 0, Is I > a.

(ii) If 0 < x' < x, we have a convolution, giving the con-

tribution
G(p) H(p). ( 49)

(iii) By interchanging orders cr integration,

JOOo
e- Px dx foo g(x') hex-x') dx
x

J: g(x') dx' [J:~ e- Px hex-x') dx (50)

-f: oo e -px h(x-x') dj.

In dealing with both of these integrals we need to note that

g(x) and hex) are even functions. In the first integral,
the substitution x" = x-x' yields the contribution
G(p) H(-p); in the second a change of sign of both variables
gives (47) with p replaced by -po Hence (46) becomes
2
G(p) - H(p) = ~ + n G(p) [H(p) + H(-p)]. (51)
P

The problem with this result is the occurrence of H(-p); we

now show how this can be circumvented. IO Solving for G(p),
we obtain
Q2p-1 + H(p)
G (p) = ---'-~----";~-- (52)
I - n H(p) - n H(-p)
The function H(p) is an entire function of p, since
hex) =0 for Ixl > a, and we assume it to be finite for
Ixl < a. Also G(p) is regular for Re(p) > 0 (except at
70 PART I: THE LAPLACE TRANSFORM

p = 0), so that the denominator cannot have zeros for

Re(p) ~ O. But the denominator is an even function of p,
hence the function G(p) is entire except for a simple pole
at p = O. Consequently, the function

(53)

is an entire function; it is not difficult to show (problem

21) that it is also bounded as Ipl + 00 It follows that the
function is a constant; evaluating it at p = 0 we have

2 2 -1
P G(p) [Q P - H(-p)] (54 )

We use this result to eliminate H(-p) from (51), obtaining

G(p) - H(p) = Q2[~ - ~ + n G(p) [H(p) + Q2p-1]. (55)

The function hex) can be obtained by simple considerations,

using (48). First note that the inversion of (55) gives

g(x) - hex) = Q2 [l-nx] + n fXo g(x') [hex-x') + Q2 ]dx', (56)

a considerable simplification on the original equation (43).

For Ixl < a, the convolution is zero because g(x') = 0
for Ix'i < a, hence

hex)
{-Q 2 (1 - nx), Ixl < a
(57)
0, Ixl > a.

The constant Q2 can be evaluated by inserting this result

into the definition of Q2 (43b); this gives Q2 =
-2
(l-na) . Eq. (56) is now a convolution equation of Volterra
type, and is amenable to analysis using the methods of Sec-
tion 5.1. Details are left as a problem.
§S. Integral equations 71

Problems

l. Show that Abel's integral equation

( CP(T) dT = f(t), o < Re(a) < 1

o (t_T)a

has the solution

2.
<PC t)

Solve the integral equation

sinCa'll)
'II
r.tiQl
Lt l - a - r
0
f' (T)
(t_T)l-a dT Jl
sin 5 = J: JO(s-t) get) dt.

3. By introducing the change of variables 5 = xl/2, u =

yl/2, show how the solutions of the equations

f:
2 2
f(s) = k(s -u ) g(u) du

f:
and
<P (x) = k(x-y) 1jJ(y) dy

are related.

4. Solve the integral equation

S. Find the resolvent for the equation

g(s) = f(s) + A fo (s-t)a

5
get) dt.

6. Define
<p(x) = lim 1m [K(x+i£)]
£+0

where K(p} is given by (17). Show that

0 , x > 0
<p (x) {
1Txe x , x < O.
72 PART 1: THE LAPLACE TRANSFORM

Since it may also be shown that

t~i+oo 1m [K(p)] : 0,

deduce the fact that 1m [K(p)] > 0 for all p satisfy-

ing Im(p) > O.

7. Analyze (16) when a: 1 to determine the behavior

for large t.

8. Consider the integral equation ll

f(x) : IX k(x-t) get) dt.

o
Under what condition (onthe kernel function k) may the
solution be written
X
g(x) : J k(x-t) Pn(d/dt) f(t) dt,
o
where k is the same kernel function, Pn is a poly-
nomial of degree n, and f(O) : f'(O) f(n-l)(O):O.

Solve the following integral equations in the form given in

Problem 8:

9. f(x) :
r 0
e-(x-t) get) dt

10. f(x)
r 0
[sinh(x-t) - sin(x-t)] get) dt

11. f(x)
J: 2
sin (x-t) get) dt

12. f(x)
r 0
e
x-t
erf(/x-t) get) dt

13. f(x) J: JO(/x-t) get) dt

§5. Integral equations 73

Solve the integral equations:

14. f (s) = 2 ts
t g (t) dt
ft2:2
t -s

fs
00

15. f(s) g' (t) dt

s
v't-"S
. x
1 -a/4x
16. g (x)
Ix
e + ~fo &lrl
IX::Y
dy

17. Obtain (Z7) from (3Z) and (33).

18. Prove that the constants and YZ in (4Z) are arbi-

trary. [Hint: examine the equation
x
¢(x) f ~Cx-y)¢(y) dy.]
o

I:
19. Solve
g(x) sinClx-yl) g(y) dy.

r
20. Solve
g(x) = 1 + Ix-yl cos(lx-yl) g(y) dy.
o
21. Show that if the pair distribution function g(x) is
bounded as x + 00, then the function defined in (53)
is bounded as Ipl + 00.

22. Show that (57) implies

QZ = (l-na)-Z
where QZ is defined in (43).

23. Investigate the Volterra integral equation (56) for

g (x).

24. In three dimensions, the Percus-Yevick equation may be

written
74 PART I: THE LAPLACE TRANSFORM

+ n I Hr') f(r') Hr-r') e(r-r') dr'·

Assume a spherically symmetrical solution, and introduce

variables IE' I, Ir-r'l, and the angle 8 between r'
and r-r' into the integrals. By considering the func-
tion x ~(x), in the case of hard spheres rVer) = 00

for r ~ a, VCr) = 0 for r > a] find the direct cor-

relation function explicitly and derive a convolution
integral equation for the pair distribution function
g(x) = ~(x) e(x).

Footnotes

1. See J. H. Giese, SIAM Review (1963), ~, 1.

2. Some numerical values for the case a = -2 computed by

Pade approximation may be found in L. Fox and E. J.
Goodwin, Phil. Trans. Roy. Soc. Lond. (1953), A245,
501.

3. N. Mullineux and J. R. Reed, Q. Appl. Math. (1967), ~,

327.

4. Equations of this type may be solved by the Weiner-

Hopf technique (see Section 18). However, we are inter-
ested here in a class of problems which can be solved by
more elementary methods.

5. We must first take Re(p) < a, and then use analytic con-
tinuation on the final result to extend it to Re(p) > a.
§5. Integral equations 75

6. As with (29), a process of analytic continuation may

be involved.

7. This is the probability of finding two particles at the

stated positions. For an infinite uniform system it is
a function only of the relative positions of the two.

8. This identification is only valid in the Percus-Yevick

approximation.

9. M. S. Wertheim, J. Math. Phys. (1964), ~, 643. The more

general case where Vex) r 0 for a ~ Ixl ~ ~ is also
analyzed using Laplace transforms.

10. The ensuing procedure is a simple example of the type

of argument which is used in the Wiener-Hopf technique
(Section 18).

11. Problems 8-13 and some related material may be found in

D. O. Reudink, SIAM Review (1967), ~, 4.

76 PART I: THE LAPLACE TRANSFORM

§6. THE INVERSION INTEGRAL

6.1. Inversion of Meromorphic Functions

Analytic information about the inversion integral is
usually obtained by "closing the contour", as in Section 2.4
for rational functions. More generally, we may consider a
Laplace transform F(p) which is meromorphic, that is, a
funct~on having no singularities other than poles, and in-
vestigate the integral
~
~7Tl
f
L+r
F(p) e pt dp, (1)

where the contour is shown in Figure 1.

1m (p)

Re (p)
poe

Figure 6.1

We suppose (as is usually the case) that it is possible to

choose a sequence of values ~ of R so that ~ + 00 as
n + 00, while on the corresponding contours r the inte-
n
grand satisfies the inequality

k > 0 (2)
§6. The inversion integral 77

for some T. (The reason for choosing a discrete sequence of

contours rather than allowing R to vary continuously is to
avoid having poles lying on the contours.) It can readily
be shown from (2) that

lim __1__ J pep) e pt dp = 0, t > T, (3)

n-+-co 2Tfi r
n
so that as we take the limit n -+- co of (1) we recover the
inversion integral. The only singularities enclosed by
these contours are poles, hence the inversion integral is
given by the sum of the residues at these poles.

Heaviside Expansion Theorem: Suppose that the poles of Pcp),

at p = ak' are all simple; then the function H(p) = l/P(p)
has simple zeros at p = a k , and the residues of Pcp) are
given by l/H ' (a k ). More generally, if there is a conveni-
ent factorization
Pcp) = G(p)/H(p), (4)

where G(p) is an entire function and H(p) has only

simple zeros, the inversion integral is given by the series

f(t) =
I G (ak)
(5)
k=l H' (a k )
This result, first formulated by Heaviside in relation to
his operational calculus, is known as the Heaviside expan-
sion theorem.

Examples:
1 , c > 0
(i) Pcp) (6)
P cosh (yip)

This transform was derived in Section 4.1 in connection

with the solution of a partial differential equation. Notice
78 PART I: THE LAPLACE TRANSFORM

that the Taylor series for the cosh function has only even
powers of its argument, so that F(p) does not have a branch
point at the origin. For negative t, we can close the con-
tour in the right-hand half-plane (Figure 2), and readily
show that the integral along r goes to zero for large R.

1m (p)

Re(p)

Figure 6.2

Since the function is analytic in the right-hand half-plane,

this gives
f(t) = 0, t < 0. (7)

This feature is a general one in the inversion of Laplace

transforms as is shown by the inversion theorem [equation

(2.23)].

For t > 0, we may close the contour in the left-hand

half-plane, and the Heaviside expansion theorem gives

f(t) (8)
§6. The inversion integral 79

(ii)
F(p) cosh (tS/p) c > 0 (9)
P cosh (yip)

Again the function does not have a branch point, and the
Heaviside expansion theorem gives
(_l)k cos[tS(k-l)n/y] _(k_l)2n2t/y2
00

f(t) = 1 + ~ L 2 e 2 (10)
n k=l (k- ~)

6.2. Inversions Involving a Branch Point

If the Laplace transform has a branch point, possibly
in addition to singularities in the form of poles, then it is
appropriate to consider the integral (1) with a new contour
as shown in Figure 3. For convenience, we have assumed that
the branch point is at the origin, and that there is only
one. Extension of the following techniques to more general

1m (p)

Re(p)

Figure 6.3

situations is not difficult (in principle). Assuming that we

80 PART I: THE LAPLACE TRANSFORM

can again make the contribution from r vanish by taking R

sufficiently large, we have

f(t) = E residues at poles

+ _1_
21Ti
JB F (p) e pt dp. (11)

Special Case: The treatment of the loop integral depends on

the behavior of the integrand near the branch point. If
F(p) 'V pY, with Re(y) > -1, then we can "shrink" the con-
tour onto the branch cut. Writing p = u exp(± i1T) accord-
ing as Im(p) is positive or negative, this leads to

J F(p) e pt dp = (12)
B

In some cases it may be possible to evaluate the integral

explicitly; in other cases an asymptotic series for large
t follows immediately by the use of Watson's lemma. As an
example, consider the function

1 -yip
F(p) = - e (13)
Ip
Substituting into (12) leads to

f(t) = ! foo u- l / 2 e- ut cos(YUl/ 2 ) du, t > o. (14 )

1T a

The integral can be reduced to a more standard form by the

2
substitution ut = s , giving for f(t) the expression
2
f(t) = (1Tt)-1/2 e- y /4~ (15)

We leave it to the reader to verify that the application of

Watson's lemma to (14) yields the Taylor series of (15) in
2
ascending powers of (y /4t), which is the asymptotic series
for large t.
[6. The inversion integral 81

More General Case: The Laplace transform

F (p ) = 1:. e - Y Ip , (16)
p

which occurred in Section 4.1, cannot be treated by substitu-

tion into (lZ) because it diverges too strongly at p = o.
We consider the loop integral more carefully, breaking it up

into three parts (Figure 4). For the integral around the

Figure 6.4

small circle, we write p = £ exp (is), and readily show that

1
fB F(p) e pt dp = 1 + 6'(£). (17)
Z
For the other two contributions, we can use (lZ) with the
lower limit replaced by u = £; subsequently setting £ to
zero gives

f(t) = 1 - _1 Ioo e- ut sin(yul/ Z) du . (18)

'IT 0 U

Comparing this result with (14), we see immediately that

8Z PART I: THE LAPLACE TRANSFORM

e- ut sin ( yU liZ) -du

u

(19)

Z ~ -1/2 fYIZlt Z
e _u du
o
= - erf (y/Z/t),

and hence the complete inverse of (16) is

f(t) erfc (y/Z/t), t > O. (ZO)

6.3. Watson's Lemma for Loop Integrals

The above examples involving a branch cut were reduced
quite readily to real integrals to which Watson's lemma could

be applied. We consider here an extension of Watson's lemma

to loop integrals which has the advantages of being direct in
application to the inversion integral, and of working for a
wider range of integrals than may be treated by (lZ).
Specifically, we will show that if F(p) has the asymptotic
expansion
'" Av
F(p) L
'" v=l avp , p .... 0,
(Zl)
-~ < arg(p) < ~

where Re(A l ) < Re(A Z) < Re(A 3 )< ••• , and Re(A v ) increases
without bound as v .... "', then the loop integral

f(t)
1
Z~i
fO_:_ F(p) e P t dp (Z Z)

has the asymptotic expansion

§6. The inversion integral 83

a\l
f(t) 'V
L A +i
v= 1 v
(-Av-l)! t

- 7f/2 < arg(t) < 7f/2 . (23)

In (22) the notation -~, 0+ means that the contour begins

and ends at p = -~, and circles the origin once in the posi-
tive direction, as in Figure 4.
The proof is quite simple. Define a set of functions

(24)

and substitute into (23) to get

n
1 (+ Av pt
f(t) = L av
hi
p e dp
v=l -~

+ -
1
-
27fi
t+
-~
F (p) e pt dp.
n
(25)

If Re(t) > 0, i t is permissible to make the substitution

u = pt in the first integral while using the same contour
for u as for p; the Hankel integral representation of
liz! (Appendix A) then gives

1
!iT
JO+
_~
A pt
p e dp
1
(26)

To deal with the remainder term, choose n sufficiently large

so that Re(A n+ 1) > -1; then we can shrink the contour onto
the branch cut--which we could not do with (22) because there
is no restriction on Re(A). Equation (12) followed by
1
Watson's lemma for real integrals then yields the estimate

-
1 f O+ F (p) e pt dp = tJ(t
-A
n+ 1
-1
). (27)
27fi -~
n

This completes the proof.

84 PART I: THE LAPLACE TRANSFORM

6.4. Asymptotic Forms for Large t

The information gleaned above may be applied to many
inverse Laplace transforms (and, as we shall see in Part II,
to Fourier transforms also) to recover asymptotic informa-
tion for large values of the time. If the singularities of
F(p) all take the form of isolated poles and/or isolated
branch points, then by a suitable deformation of the inver-
sion contour we may reduce the integral to a sum of resi-
dues at the poles plus a sum of loop integrals around the
branch points. These latter may usually be estimated for
large t by Watson's lemma for loop integrals. For example,
if there is a branch point at p = a, and if F(p) has the
asymptotic expansion

F (p) 'V (28)

then the substitution p' = p-a reduces the loop integral to

0+
I f_oo F(p'+a)e(p'+a)t dp'
271 i

'V e at (29)

Formulas appropriate for asymptotic expansions invo~ving

logarithmic functions may also be derived--see Problem 9.

For large t, the contribution from each pole and each branch
point is dominated by the exponential factor, and the asymp-
totic form of the complete inversion integral will be gov-
erned by the singularity whose position p = a has the most
positive real part.

Examples: (i) Consider the function

 §6. The inversion integral 85

1
F(p) = a > 0 (30)
Ip (p+a)
which has branch points at p o and p = - a. Sine e a > 0,

we need consider only the origin for large t, so that

F(p) '" _1_

!pi lrl - L +
2a 8a 2
~+
... l, (31)

with the corresponding asymptotic expansion

f(t) '" -l-llr

hat
+ -
1
- +
4at
9
32a 2 t 2 +"'J'
'1 (32)

(ii) The Bessel function JO(t) has the Laplace transform

[(3.53)]
F(p) = _1_. (33)
42+1
There are two branch points, both on the imaginary

axis, and consequently of equal importance for large t.

The necessary asymptotic expansions of F(p) are

[1 J
~L
e- i7r / 4 E.:i. 3 (p-i) 2
4i 32 +... , p ... i
12(p-i)
F(p) '"
e i7f / 4
12(p+i)
p+i
[1 + '2f1
3(p+i)2 +
32
J
. .. , p .... - i,
(34)

from which it follows that

J O(t) '" I27rt cos (t-7f/4) [1 - 9/128t 2 + ... ]

(35)
+ 1217ft sin (t-7f/4) [1/8t - ... ] , t ... co •

Because there are no other singularities in this case, there

are no neglected terms which are exponentially small.l
86 PART I: THE LAPLACE TRANSFORM

6.5. Heaviside Series Expansion

For small values of the time, it is often possible to

extend the technique of Section 2.5 to derive an expansion
in ascending powers of t. Sometimes this expansion will be
a convergent Taylor series, but more often it will be an asymp-

totic expansion. We deal with the latter, since it includes

the former as a special case. Suppose then that the Laplace
transform F(p) has an asymptotic expansion 2
00 -A
F(p) ~ rap v (36)
v=l v

then for any n we can define the function Fn(p) in the

usual way by

(37)

and deform the contour into the right-hand half-plane so that

(38)

Some elementary considerations, the details of which we omit,

then lead to the Heaviside series expansion, namely,
A -1
av t v
r
00

f(t) ~ (39)
v=l (Av-l)!

An Example: We consider again the Bessel function JO(t).

Expanding (33) in descending powers of p gives

lIT"
00

F(p) ~ l ( 40)
k=l (-k-i)! k! p(2k+l)
with the corresponding Heaviside series expansion
00
(-l)k ft/2)2k
J 0 (t) = r (41)
k=O k! k!
Since the expansion of F (p) is a convergent series for
§6. The inversion integral 87

Ipi > 1, the series (41) is also convergent.

Problems

Find the inverse Laplace transforms of the following functions

using the inversion integral.

1
1.
p/p+1

2. 1
a+1p
-ap
3. 1-e
p
, a > 0

e-ap_e- bp
4.
p
o< a < b

- ap - bp
5. e -e 0 < a < b
2
P

6. R.n
[~J
p+a

7. R.n
[~]
p2+a

8. Show that if FCp) has the asymptotic expansion

()() X
FCp) ~ L a v p vR.n p, p + 00,
v=l

-7r ~ arg(p) -< 7r ,

where ) in- Re O.
v
creases without bound, then the loop integral (22) has
the asymptotic expansion
()()
av 1jJ( - Xv) - R.n t
f (t) ~
I
v=l c-x v -1)! X +1
t v
88 PART I: THE LAPLACE TRANSFORM

where

1/1 (a.+ 1)
d
da. [R.n a.!].

10. Invert
2 2
F(p) = P R.n (1 - a /p ).

11. Find power series for the functions whose inverses were
found in Problems 1, 2, 6, and 7.

12. Find an asymptotic expansion for the inverse of

F(p)
122
Ip +a.

Footnotes

1. For a discussion of the possible importance of exponen-

tia11y small terms, see OLVER (1974), pp. 76-78.

2. If the expansion is convergent, then so is the inverse

(39). See CARSLAW & JAEGER (1941), pp. 271-273.
 Part II: The Fourier Transform

§7. DEFINITIONS AND ELEMENTARY PROPERTIES

7.1. The Exponential, Sine, and Cosine Transforms

Let f(t) be an arbitrary function; then the (expo-
nential) Fourier transform of f(t) is the function defined
by the integral

F(w) (1)

for those values of w for which the integral exists. We

shall usually refer to (1) as the Fourier transform, omitting
any reference to the term exponential. The Fourier trans-
form is related to the Laplace transform; indeed, on denot-
ing by f±(p) the following Laplace transforms:

f: e- pt f(±t) dt, Re (p) > CI.±' (2)

we have
F (w) (3)

89
90 PART II: THE FOURIER TRANSFORM

Furthermore, we see that (1) will converge for values of w

in the strip a+ < Im(w) < -a_, corresponding to the regions
of convergence of (2).

Inversion: Consider the inversion integrals

_1_ r+ ioo ept

271"i . f± (±p) dp. (4)
Y-l°O
If Y> a+, the first integral gives f(t) for t > 0
and zero for t < O. Similarly, on making the substitution
p ~ -p, we see that if -y > -a , the second integral gives
f(-t) for t < 0 and zero for t > o. Adding these two
results we have

1 JY+l.· 00 pt
f(t) = "ZiTI e [f+ (p) + f (-p)] dp
Y-l°O
(5)
1
271"

where the last step follows from the substitution p + -iw,

and the use of (3). Hence we have the reciprocal transform
pair
F(w) JOO _00
eiwt
f(t) dt, a < Im(w) < S
(6)
1 fl.·Y+OO e-iwt
f(t) = 271" F(w) dw, a < Y < S.
lY-OO
Sine and Cosine Transforms: Consider the functions defined
by the integrals

F
s
(w) =2 JOO sin(wt) f(t) dt,
o

f:
(7)

Fc(W) = 2 cos(wt) f(t) dt,

known as the Fourier sine and Fourier cosine transforms.

They may be related to the Laplace transforms (2) by
§7. Definitions and elementary properties 91

F (w)
s
-
[f+(-iW) + f+(iW)] ,
(8)

so that the integrals (7) converge (if at all) in the strip

a+ < Im(w) < -a_, which includes the real axis. Employing
the Laplace inversion theorem with y = 0, we may readily
show that the inversions of (7) are

f(t) = !.1T foo0 sin(wt) F (w) dw,

S
(9)
f(t) 1 foo cos(wt) F (w) dw .
1T 0 C

Examples:

(i) f(t) e- altl , Re(a) > 0

(10)

Here the region of convergence of the integral defining F(w)

is IIm(w)1 < Re(a). Inserting F(w) into the inversion
integral (5) with y = 0, we can easily evaluate the latter
by residues. If t > 0, we must close the contour in the
lower half-plane; if t < 0, we must close in the upper half-
plane. The effect of switching from one pole to the other
1
gives the result
e- at , t > 0
f(t) { +at
e , t < 0
(11)

e -altl .

(ii) f(t) Re(a) > 0 (12)

92 PART II: THE FOURIER TRANSFORM

Few)
2 2
e-w /40. (13)

('lr/a.) 1/2

In (13) we have written u = t - iw/2a., and changed the con-

tour from -00 < t < 00 to -00 < u < 00. In this example, the
integral defining F(w) converges for all w, and the
Fourier transform is an entire function of w as a conse-

quence. Since F(w) has the same functional form as f(t),

the inversion integral is evaluated by a trivial modification
of (13).

1, -1 < t < 1
(iii) f(t) = { (14 )
0, It I > 1

r1 iwt
F(w) e dt
Ll (15 )
2 sin w
w

To invert (15), we write the inversion integral as

I
1
~i
JC e -iw w(t -1) dw _ ---:-
1
~1
fC e-iw(t+l)
w
dw (16)

where the contour is shown in Figure 1. If t < -1, we

can close the contour in the upper half plane for both in-
tegrals, giving I = O. If -1 < t < 1, we close in the

lower half-plane for the second integral, obtaining I = 1.

Finally, if t > 1, we close in the lower half-plane for both
integrals, and the residues cancel. Hence I = f(t).
§7. Definitions and elementary properties 93

Im(w)

c
__________________~~----------------------~ Re(w)
'pole

Figure 7.1

(iv) (17)

For this function, the integral (1) will diverge unless w

is real, so there is no strip a < Im(w) < B for conver-
gence. For real w we can evaluate (1) by writing2

F (w) = lim
r
-00

2 . 2 -1/2 2 . 2 -1/2 (18)

lim {[a - (W-1E:) ] + [a - (W+1E)] },
E-+-O
where we have used the Laplace transform of JO(at) in
(3). On taking the limit E -+- 0, we have

Iwl < a
(19)
Iwl > a

where the result for Iwl > a depends on a consideration of

the phases of the two terms in (18). The inversion integral
may be transformed into Bessel's integral (20.50) by the sub-

stitution w =a sin a, viz.

94 PART II: THE FOURIER TRANSFORM

-iwt
JO(at) =!11 ( _ a e dw
2 2 I a _w

= ~ Ja cos(wt) dw ( 20)

r/
11 0 / a 2-w 2

=~ 2
cos (at sin e) de.
11 0
ft- l / 2 t 0
to :
>
(v) f (t) (21)
t < 0

F(w)
J: t
-1/2
e
iwt
dt
(22)
e ill/4 ITT Im(w) > 0
rw
The inversion integral is (see Figure 2 for details)

fet) = e
-ill/4
2 ITT
Ic --iwt
e - dw.
IW
(23)

1m (w)

-c ...
Re (w)

r--..... branch cut

of W -1/2

Figure 7.2
97. Definitions and elementary properties 95

If t < 0, we can close the contour in the upper half-

plane, giving zero. If t > 0, we can close C in the lower
half-plane (giving Cl ), and 'shrink' the contour about the
= iw, this gives
r
branch cut. With ~

e- i1T / 4 e-~t le h / 4 _ e- 3i1T / 4 l d~

2 In 0 L~ ~ j
1 Joo ~-1/2 e-~t d~ (24)
I1T 0

t -1/2 .

Hence we recover f(t) for all t.

7.2. Important Properties

We will derive a number of simple but important pro-
perties of the exponential Fourier transform; the correspond-
ing properties for the sine and cosine transforms, which are
also simple, are given in the problems.

Derivatives: Suppose that get) = f' (t); then

G(w) = foo f'(t) e iwt dt

_00

iw Joo f(t) e iwt dt (25)

-00

-iw F(w),

provided 3 f(t) + 0 as It I + 00. Similarly if ~(t) =

t f(t), then differentiation of the integral which defines
F(w) yields
~(w) = -i ~w F(w), (26)

provided the order of integration and differentiation may be

changed. Equations (25) and (26) represent a duality between
96 PART II: THE FOURIER TRANSFORM

operations on a function and the corresponding operation on

its Fourier transform; symbolically we may express this by
the correspondence

d
-++ -iw,
dt
d (27)
-++
dw it.

Translations: Simi1ar1y,there is a duality between transla-

tions of a function and multiplication by an exponential fac-
tor. Denoting by jF[f] the Fourier transform of f(t), we
have
jF[f(t- T)] r _00 e
iwt f(t-T)
dt

iWT
e (00 e iwu feu) du (28)

iWT
e F(w),

~[eiat f(t)] = f~oo ei(a+w)t f(t) dt

(29)
F(a+w).

Convolutions: A convolution integral of the type given in

(1.22) has a particularly simple Laplace transform. The
corresponding result for the Fourier transform stems from re-
placing the integration limits by too; that is, we consider

the Fourier transform of the function defined by the convo1u-

tion integral
get) = fOO k(t-T) f(T) dT. (30)
-00
Assuming that the necessary changes of orders of integra-
tion are valid, the application of (28) gives
§7. Definitions and elementary properties 97

G(w) r co Y[ k (t - T)] f (T) dT

K(w) fco e iwT f(T) dT (31)

-00

= K (w) F (w).

There is a similar result, which again reveals a duality be-

tween operations on functions and their Fourier transforms,
for the Fourier transform of the product of two functions.
By replacing one of the functions by its inverse Fourier
transform and using (29) we obtain

Y[f(t) get)] = ;71 Joo

_00
F(w') dw' Y[e
-iw't
g (t)]

(32)
~7I f~oo F(w') G(w-w') dw'.

Parseval Relations: One immediate and important consequence

of (32) is obtained by putting w = O. The resulting
equation, which involves the function G(-w), may be made
more symmetrical by replacing get) by its complex conjugate
function g*(t), and noting that

y[g*(t)] G*(-w). (33)

Hence on writing Y[f(t) g*(t)]w=o as an integral we have

Coo f (t) g*(t) dt = -1271 r _

00
F(w) G*(w) dw, (34)

which is Parseval's relation.

7.3. Spectral Analysis

Suppose that f(t) represents the value of some
physical quantity at time t. Then if the Fourier transform
98 PART II: THE FOURIER TRANSFORM

exists for real w, the representation

1
f(t) = 2'IT
foo
_00 F(w) e
-iwt
dw (35)

is a way of expressing f(t) as a linear combination of

simple harmonic functions cos(wt) ± i sin(wt). This means
that the frequency content of the signal f(t) is spread
over a continuous range of frequencies w, the amplitude of

a given frequency being proportional to F(w). If If(t)1 2

is a measure of the intensity of the quantity f(t) at time
t, then we may regard the function IF(W)1 2 as a measure of
the intensity at the frequency w. Parseva1's relation a1-
lows us to given these two statements a consistent quantita-
2
tive meaning: if If(t)1 ot is the power content of f(t)
in the time interval from t to t + ot, then we may interpret
2
IF(w) I ow/2'IT as the power content in the frequency range w
to w + OW, for then the relation

fOO 2
If(t)1 dt = 2'IT
1 foo IF(w)1 2 dw (36)
_00 _00

gives an unambiguous meaning to the concept of the total

power content of the quantity f(t).

Illustrative Example: Consider the following simple mechani-

cal problem. A mass m is suspended by a spring with force
constant k, subject to a linear damping force proportional
to its velocity and driven by an external force f(t). The
equation for the displacement of the particle from equi1ib-
rium is
m x"(t) + y x' (t) + k x(t) f(t). (37)

For simplicity we put m= 1 and k = 1, and consider the

case of light damping, y «1. First we make the driving
§7. Definitions and elementary properties 99

force a periodic function, f(t) = sin(wt), and look for

steady-state solutions x(t) = B sin(wt +8), where Band
8 are functions of w but not t. Direct substitution into
(37) gives us two relations for these quantities, namely

B[(l-W Z) cos 8 - wy sin 8] 1,

(38)
B[(l-w Z) sin 8 + wy cos 8] 0,

from which we obtain

1
(39)

The rate at which energy is dissipated by friction is

ylx'(t)I Z, hence the energy dissipated per cycle is

JoZW / w ylx'(t)1 Z dt=wYlwBI.

w Z
( 40)

A graph of this quantity is shown in Figure 3.

Figure 7.3

Notice that for small y the peaks at w ~ ±l are

high and narrow.
100 PART II: THE FOURIER TRANSFORM

Now we apply the force

f
0 t < 0
f (t) sin t, 0 < t < 2'1T (41)

l 0 t > 2'1T

which is one cycle of a sine wave at the resonant frequency.

The solution of (37) for Y« 1 and this force is approxi-
mately given by

= {}(Sin t - t cos t), o < t < 2'1T

x (t) (42)
-'1Te-Yt/ 2cos t, t > 2'1T.

Let us calculate the total energy dissipated by friction as

the result of this 'one-cycle' signal. There may seem to be
two methods, viz.:
(i) Use the solution (42) to calculate the integral of
ylx' (t)1 2 . Explicitly, this gives

( 43)

(ii) Use (40) with w = 1, and multiply by the period 2'1T

during which the force is applied. This gives '1T/Y as the
energy, a result which disagrees completely with (43).
The resolution of this problem is quite easy if we
avply the concept of spectral analysis to the force f(t).
Writing
f(t) = ~'1T foo F(w) e- iwt dw (44 )
-00
and applying the steady-state result (38) to each harmonic
component separately, we obtain for x(t)

1 [00 -i[wt+8(w)]
x(t) = --
2'1T _00 F(w) B(w) e dw. (45)

To compute the total energy, we apply Parseval's relation to

the integral of Ylx' (t)1 2 , giving

§7. Definitions and elementary properties 101

2 2
E ~1T f oo
_00 IF(w) I IwB(w) I dw
2 (46)
1T
"2

which agrees with (43). This illustrates the fact that the
energy is spread out over a wide range of frequencies.

7.4. Kramers-Kronig Relations

Consider a linear physical system with input x(t)
and output (response) yet). We suppose that the law of
cause and effect holds, that is, that the output yet) de-
pends only on values of the input x(t') for times t' < t.
Then the most general linear relation which we may write is a
convolution integral involving an influence function
k(t-t'); viz.

yet) = f~oo k(t-T) X(T) dT, (47)

where k(s) = 0 for s < 0 because of causality. We sup-

pose, further, that the system is unconditionally stable, so
that if x(t) = 0 for t > to' then yet) + 0 as t + 00.

This means that the Fourier transform K(w) has no poles in

the half-plane Im(w) > O. Finally, suppose that a real in-
put results in a real output; then we may show that the real
and imaginary parts of K(w) are, respectively, even and
odd functions of the real variable w.

Now consider the contour integral

fC K(W) W-rl
dw (48)

where the contour is shown in Figure 4.

102 PART II: THE FOURIER TRANSFORM

1m (w)

Re (w)
Figure 7.4

We know that K(w) has no poles inside the contour, hence

the integral has the value zero. Evaluating one-half the
residue at w = n therefore gives
ex>
K(w) dw
iTT K(n) + PV = 0, ( 49)
Lex> w-n
provided KCW) ... 0 as Iwl ... ex> in the upper half plane.
Equating real and imaginary parts we have [with K(w) =

r
Kr (w) + iKi (w)]
Ki (w) dw
KrCn) - -1 PV
TT -ex> w-n

r
(50)
Kr(w) dw
Ki(n) = !TT PV _ex>
w-n
which are the Kramers-Kronig relations. Thus the require-
ment of causality leads to a connection between the real and
imaginary parts of K(w) for a very general class of linear
systems.
Another important relation of this type is obtained
by considering the integral (see Figure 5)
§7. Definitions and elementary properties 103

Im(w)

Figure 7.5

(51)

Exploiting the fact that Kr(w) is an even function of

(real) w, this gives

K(i~) = 2
'If
Joo0 (52)

an important relation with many physical applications.

Problems

Prove the following general properties of the Fourier trans-

form.
1. ~[f*(t)] = F*(-w)

2. If f (t) is an even function, F(W) = Fc(W).

3. If f(t) is an odd func ti on , F(w) -iFs(w) .

4. ff[f(t/a+b)] = a e iabw F(aw)

5. ff[e ibt f(at)] ! F(w+b)

a a
104 PART II. THE FOURIER TRANSFORM

1 [F (w+b) + F (w-b))
6. ~ [cos (bt) feat)]
s 2a s a s a

1 [F (w-b)
7. ~s [sin (bt) feat))
za c a
Fc(W:b)]

1 [F (w+b) + F (W- b) ]
8. ~ [cos(bt) feat)]
c Za c a c a

1 [F (w+b) _ F (w-b)]
9. ~
c
[sin(bt) feat)]
za s a s a

10. §' [f(n)(t)] _2f(n-1) (0) + wF [fen-I) (t))

c s

II. Y[f(n)(t)] -wF [f(n-1) (t)]

s c

12. ~ [f" (t)] Zw f(O) - wZF (w)

s

13. g [f" (t)] - Zf' (0) _ w2 F c (w)

c

14. In foo F (w)G (w) cos (wt) dw

2 0 c c

= ! foo g(u)[f(t+u) + f(lt-ul)] du

Z 0

15. --Zl fOO F (w) G (w) cos (wt) dw

nos s

= ! foo g(u) [f(t+u) + f(lt-ul)] du

Z 0

16. ~2 fOO
F (w) Gc(w) sin(wt) dw
nos

1 foo feu) [g(lt-ul) - g(t+u)) du

"2 0

= -1 fOO g(u) [f(t+u) - f(lt-ul)) du

f:
2 0

17. f(t) get) dt

§7. Definitions and elementary properties lOS

18. For a function f(x) with Fourier transform F(p), we

define the quantities

<xn> = J~oo x n lf(x)1 2dx,

(~x)
2 = <x 2> - <x> 2 ,

(~p)
2 = <p 2> - <p> 2

Show that 4

(~x) (~p)
- 2
> 1:.
for any function f(x).
[Consider the inequality

roo 1{xf(x) - <x>f(x)} + cdf' (x) + i<p>f(x)} 12 > 0

where ~ is an arbitrary real number.]

Verify the following list of Fourier transforms:

iat
,
{:
19. P < t < q
f (t)
, t > q or t < P
eip (a+w) _ eiq(a+w)
F (w)
w

20. { lit, t > 1

f (t)
o , t < 1

2
21. f(t) cos (at )
F (w) (~/a)1/2cos[(w2/4a)-(~/4)]
106 PART II: THE FOURIER TRANSFORM

22. f(t) sin(at 2)

F(w) (w/a)1/2 sin[(w 2/4a)+(w/4)]

23. f(t) It/-a, 0 < Re(a) < 1

F(w) 2(-a)! sin(1fa)

/w/ 1 - a
e- a / t /
24. f (t)
/t/1/2

[ fa2::2 al 1 / 2
+
a 2+w 2 J
F (w)

25. t < a

t > a

F(w) 2\1+1/2 liT \I! \1+1/2 -\1-1/2 (

a w J\I+1/2 awl

26. f(t) sinh (at) -1f < a < 1f

sinh (1ft)

F(w) sin (a)

cosh(w) + cos (a)

27. f(t) cosh (at) -1f < a < 1f

cosh (1ft)

F(w) 2 cos (a/2) cosh(w/2)

cosh(w) + cos (a)

28. f(t)
sin[bia 2 + t 2]

1a2 + t 2

F(w)
/w/ < b
§7. Definitions and elementary properties 107

29. f (x) erf (ax)

_w 2/4a 2
F (w) e
w
-t
30. f(t) e

2w
F (w)
s 1+w2
2
-t
31. f(t) = e

2
F (w) = liTe- w /4
s

32. f(t) = sin (t)

t

F (w) = tn 11+wl
s 1-w

33. o < t < a

t > a

34. f(t) = cosh (at)

sinh (7ft)

sinh(w)
Fs(W)
cosh(w) + cos(a)

sinh(at)
35. f (t) coshC'lft)

F (w) 2 sin(a/2) sinh(w/2)

s
cosh (w) + cos (a)

Use the Parseva1 relations to evaluate the following integrals.

dt 7f
36.
2ab(a+b)
108 PART II: THE FOURIER TRANSFORM

'" 2
37. t dt 'IT
fo (a 2+t 2) (b 2+t 2) 2(a+b)

38.
r 0
sin(at) sin(bt)dt
t2
{ oa/2,
'lTb/2,
a < b
a > b

Poisson Summation Formula

39. Let f(x) be an integrable function, and define

f+ (x) f (x) h(x),

f_(x) f(x)h(-x),
F+ (w) Y[f+ (x)],
F (w) Y[f (x)].
Then show that

'"
~f(O) + l f+ (an)
n=l

!f(O) +
2
-Y
n=-l
f_(an) =! Y F (2'ITm)
a m=-oo - a '
by representing f±(an) by its inverse Fourier trans-
form and evaluating the integral (after performing
the summation first) by residues. Adding these gives
00 00

l f(an) = !a l F(2'ITm)
a ,
n=-'" m=-'"
known as the Poisson summation formula.

40. Show that

1!. coth (:!!"o)

a a

'" 2 2
41. l e -n q
n=-'"
[7. Definitions and elementary properties 109

42.
l
n=-oo
JO(na) ir ~ 0 < a < 21T
a'

~
a
+ 4 ~
k=l
1 , 2m1T<a<2(m+l)1T

00

43. l J O(n1T) cos(n1Ta) 2 -1 < a < 1

n=-oo

*l
00

44. l
n=-oo ~2 1T2 n 2 m=-oo
I"a +

Footnotes

1. This method leaves the value of f(O) undetermined, a

matter of no practical consequence since inverse trans-
forms are unique only to within a null function.

2. Alternatively, we could appeal to Bessel's integral

(20.50) immediately to obtain the result (19).

3. We will see in Section 9 that (25) and (26) are

valid for generalized functions with no additional as-
sumptions.

4. In quantum mechanics, this is the uncertainty principle.

110 PART II: THE FOURIER TRANSFORM

s8. APPLICATION TO PARTIAL DIFFERENTIAL EQUATIONS

The use of the Fourier transform to obtain a form of

solution to a partial differential equation (together with
associated boundary conditions) is a very general technique.
For simple problems, the integral representation obtained as
the solution will be amenable to exact analysis; more often
the method converts the original problem to the technical
matter of evaluating a difficult integral. Numerical methods
may be necessary in general, although asymptotic and other
useful information can often be obtained directly by appro-
priate methods. We illustrate some of the more simple prob-
lems in this section, leaving applications involving mixed
boundary values, Green's functions,and transforms in several
variables until later.

8.1. Potential Problems

Problems in electrostatics and steady-state heat con-
duction involve the solution of Laplace's equation

(1)

subject to prescribed boundary conditions on the function

¢. We consider three examples.

Example 1:

2
V ¢(x,y) = 0, _00 < x < +00, y > 0
(2)
¢(x,O) = ~(x)

If we take the Fourier transform of (2) with respect to

the variable x, we obtain the ordinary differential equation
§8. Apnlication to partial differential equations 111

d2
~(w,y) - w2 ~(w,y) 0,
dy2 (3)
~ (w, 0) 'I' (w) ,

with the solutions

(4 )

To choose the functions A(w) and B(w)--constants with re-

spect to y--we need a further condition. This is because
we have not specified the behavior of the solution for large
y. Assuming that the solution is bounded for large y we
obtain l
'I'(w) e- wy , Re(w) > 0
~(w,y) = { (5)
'I' ~) e+ wy , Re(w) < O.
The Fourier transform ~(w,y) is a product, so we introduce
a function K(w,y) by ~(w,y) = K(w,y)'I'(w). From (7.11)
we see that K(w,y) is the transform of an elementary func-
tion, and on using the convolution theorem we obtain as the
solution to (2) the general formula

¢(x,y) = I.
".
I"" -00
(6 )

Example 2: Consider the electrostatic field produced by the

arrangement shown in Figure 1, where a finite section of an
infinite electrically conducting cylinder of radius a is
held at potential V while the remainder is grounded. Using
cylindrical polar coordinates, and the fact that the poten-

tial is axially symmetric, we obtain the equations

1 a (r~) + a2~ 0,
r aT ar az2
(V, Izl < ~
cjJ(a,z)
1. 0, Izl > L
(7)
ll2 PART II: THE FOURIER TRANSFORM

q,=O q,=v t/>=O

Z =--l Z=.f
Z

Figure 8.1
Now we take the Fourier transform with respect to z, so that
equations (7) become

d 2 <Jl(r,w) 1 d 2
+
r dr <Jl(r,w)
- w <Jl(r,w) 0,
dr 2
(8)
2V sin (w,Q,)
<Jl (a ,w) =
w

The differential equation can be solved in terms of modified

Bessel functions. The solution which is finite at r = 0
and satisfies the boundary condition at r = a is

2V sin (w£) IOCwr)

<Jl(r,w) = ~~w~~~ IOCwa)' (9)

The expression for the potential follows immediately from the

inversion integral, which may be evaluated over real values

of w. Noting that <Jl(r,w) is an even function of w, we

can write the solution as a real integral, namely

Joo cos(wz) sin(w£) IO(wr)

°
2V
Hr , z) = I ( ) dw. (10)
'IT W 0 wa

This solution could also have been obtained by applying the

Fourier cosine transform to the problem for z > 0, since the
§8. Application to partial differential equations 113

potential is obviously an even function of z, which implies

that d~/dZ = 0 at z = o.

Example 3: As a final example of a potential problem, we

find the electrostatic field inside an infinite conducting
cylinder of radius a due to a point charge q on its
axis. Z If we take the origin of our (cylindrical) coordi-
nates at the point charge, the potential has the form

~(r,z) = q + u(r,z), (11)

IrZ+zZ

where the first term is the field due to a point charge in

the absence of the boundary, and u(r,z) is a harmonic func-
tion chosen to make ~ = 0 on the surface of the cylinder.
Taking the Fourier transform with respect to z, we obtain
from Laplace's equation (VZu = 0) and the boundary condi-
tion the equations

1 d (r dU(r,w)) _ w2U(r w) 0,
17 dr dr '

U(a,w) e iwz dz
(12)
Iz2+a 2
= - 2q KO(wa).

The solution of (12) which is finite at r = 0 is

(13)

from which we have

¢(r, z) = -q>--- -~
'If
fo KO (wa)
oo 10 (wr)
cos (wz) dw. (14 )
/r 2+z Z 10(wa)
114 PART II: THE FOURIER TRANSFORM

8.2. Water Waves: Basic Equations 3

Water waves are perhaps the most easily observed os-
cillatory phenomenon in nature. We will discuss the applica-
tion of integral transforms to the mathematical analysis of
water waves in several places; in the present section we out-
line the basic equations to be solved and discuss some simple
problems which are amenable to analysis via the Fourier
transform.

Equations of Motion: We will briefly sketch those equations

of hydrodynamics which are appropriate to the theory of
water waves. 4 The basic assumption is that water may be re-
garded as an incompressible, inviscid fluid. We denote its
density by p, and its velocity at a point r and time t
by v(r,t). A small element of the fluid is acted upon by
two distinct forces: the pressure which acts across the
boundary of the element, and external forces such as gravity
(generally called body forces). Denoting the pressure by p
and the body forces by F, Newton's second law gives the
equation of motion
Dy
F - Vp, (15)
P Dt

where the operator DIDt is the time rate of change for an

observer moving with the element, at velocity v. In a
stationary coordinate system, assuming that the body forces
are simply gravitational, (15) becomes

p
av
~ + p(v·V)v + Vp = Pg, (16)
at - - - -
where & is a constant vector, the acceleration due to grav-
ity. One consequence of (16) is that if V x v = 0 at one
§8. Application to partial differential equations 115

instant of time, then it is true for all time. We will make

this assumption (irrotational flow). Then we can introduce
a velocity potential ~ from which v is obtained by

v = ~~. (17)

There is a fundamental conservation law of hydrodynam-

ics which expresses the fact that fluid is neither created
nor destroyed during its motion. For a fluid of arbitrary
density p(!,t), it is

~i + ~. (p~) = o. (18)

In the present case, p is a constant, so (18) becomes

V·v = o. With (17), this implies

(19)

so that the velocity potential is a harmonic function.

Another equation can be obtained from (16) and (17).
Writing v = ~~ in (16) and switching the order of differ-
entiation in the first two terms, we have

v !t
- at
+ l2 -V(v 2) + VE - g =
p-
o. (20)

With the z-axis as the upward vertical direction, g =- g~z,

we can integrate (20) along any path in space (and within

the fluid) to obtain Bernoulli's equation

a'"'I' 1
+ - V
2 + n
~ + gz A(t) , (21)
'IT 2 p

where A(t) is an arbitrary constant of the integration.

Boundary Conditions: We assume that the water has a bound-

ary surface S with the property that any particle on the
116 PART II: THE FOURIER TRANSFORM

surface remains on the surface. At a fixed boundary, such as

the bottom of the sea, the normal velocity must be zero, so
that the velocity potential has to satisfy the condition

o. (22)

At a free surface conditions are more complicated, since we

will want to specify the pressure and let this determine where
the surface is. If we assume that the equation of the sur-
face is z = n(x,y,t) then by differentiation with respect
to t we have

~ + an o on z = n. (23)
az at

A second boundary condition follows from Bernoulli's equa-

tion. Assuming the pressure is equal to a prescribed func-
tion on S, we have

a<p + !. IV<P12 + PO + gn = A on z = n. (24)

at 2 p

These boundary conditions are both nonlinear, and in addi-

tion must be used to determine n. Hence the general equa-
tions which govern the theory of water waves, even in this
simple model, are extremely complicated and intractable to
present analytic methods.

Small Amplitude Waves: We shall suppose that the elevation

of the free surface n and the pressure p are small per-

turbations from equilibrium values n = 0 and p = PO' and

that the velocity of the flow is small. Then we can linear-

ize (23) and (24) by neglecting products of perturbation

quantities and applying the resulting conditions at z = 0

rather than z = n. Thus we have

§8. Application to partial differential equations 117

o on Z = 0, (25)

= 0 on Z = 0, (26)

where op = p - PO. We can eliminate n completely to

get a boundary condition on ~,namely

d2~ + g ~ + l ~ (op) = 0, (27)

dt 2 dZ P at

after which the free surface elevation n may be obtained

from (25). Apart from the simplified form of these boundary
conditions, we have the additional fact that the region in
which ~ must be determined and the boundary at which (25)
and (26) must apply are fixed by the equilibrium solution.

8.3. Water Waves Generated by an Initial Surface Displacement

We consider first waves on water of infinite depth

which are generated by an initial displacement of the sur-
face elevation, the pressure at the surface being a given
constant. In this section we confine our analysis to waves
in two dimensions, that is, solutions of the relevant equa-
tions which are independent of y. For such functions, the
Fourier transform (with respect to x) of (19) is

d2 2
---2 ~(w,z,t) - w ~(w,z,t) = 0, (28)
dz
and the solution for real w which is bounded as z + -00 is

~(w,z,t) = A(w,t) e
Iwlz (29)

Now the boundary condition (27) with p o yields for

A(w,t) the differential equation
118 PART II: THE FOURIER TRANSFORM

(30)

with the solution

A(w,t) B(w) sin (t~) + C(w) cos(t~). (31)

Suppose that the water is initially at rest, with surface

elevation given by n(x,O) = nO(x). The condition
¢(x,z,O) = 0 gives C(w) = O. Denoting the transform of
n by H, (25) gives

g HO(w) + /j~ B(w) = O. (32)

Thus W(W,z,t) and H(W,t) are uniquely determined. Ex-

plicitly,
W(w,z,t) Ho (W)[fwrr /2 sin(tl"fWlg)e lwlz
(33)
H(w, t) = HO(w) cos (t/rz;-rg)

The expressions for the velocity potential ¢ and surface

elevation n follow immediately. Tke integrals for these
expressions are intractable even for simple forms of HO(w);
nevertheless they yield useful information, either through
numerical studies or asymptotic analysis.

An Asymptotic Form: We examine in some detail the asymptotic

form of the solution if we write

nO(x) =
{' o ,
2a' Ixl
Ixl
<

>
a
a
(34)

Then the expression for ¢(x,O,t) which we obtain from in-

verting (33a) is
§8. Application to partial differential equations 119

<p (x ,0, t) -l:SfOO sin(aw){sin(wx + t 1Wg)

21T 0 aw
(3S)
- sin(wx - t 1Wg)} dw
rw
An important feature of this last expression is that it is
unchanged i f we replace x by Ixl. Physically this must be
so since the initial displacement no(x) was symmetrical
about x = 0, and hence the disturbance must propagate in
both directions symmetrically. We now make the substitutions

-1(;i1"i'T"
¥WIAI -
+ .!.. I I = a. ,
2/1xl
(36)
i/~ = a,
after which the expression for <p becomes

<P(x,O,t) - n l
_ 1 (..L1l/2{ foo sin[(a/lxl)(a.-a)2]
2 sin(a. 2 -a 2)da.
llxlJ a (a/lxl)(a.-a)

foo sin[(a/lxl) (a.+a)2] 22 1

-------=---sin(a. -a ) da.!>-. (37)
-a (a/lxl)(a.+a)2 J
We now make some approximations. First, we note that the
major contribution to each integral comes from the regions
1a.1 - lal, so that if a/lxl « 1, we can writeS

<f>(x,O,t) '" - ~ (&)1/2 f: sin(a. 2 -a 2) da.,

(38)
Ixl » a.

The corresponding approximation to n(x,t) is readily

found to be
n(x,t) _ !. a<f>(x,O ,t)
g at
-t gl/2 fa 2 2
Ixl »
'" -3/2 cos(a. -a ) da., a. (39)
1Tlxl 0
120 PART II: THE FOURIER TRANSFORM

It is of interest to investigate these integrals for large

s. The two integrals are the real and imaginary parts of
-i1T/4
2
e-is2+i1T/4JS: e- s ds
( 40)

For large S, the error function tends to unity, so that we

have the approximation

n(x,t) ~ -t g 1/2 cos [

21TlxI 3 / 2
~ 2 -
41xl
i~ ' gt 2 » Ixl » a. (41)

It is interesting to discuss the character of the motion fur-

nished by this solution. The crests of the waves are given
2 1
by the condition gt /41xl = (2n+ 4)1T, hence they move at an
increasing velocity as time progresses. Another feature is
that the distance between two successive crests increases
with time. Hence the waves furthest away from the initial
disturbance move more rapidly and become longer as the pat-
tern is drawn out. SimultaneouslY,new waves of shorter wave
length continually appear and also propagate outwards. These
conclusions remain unchanged if we consider the three dimen-

sional case;6 furthermore they may easily be observed by

throwing a small stone into a calm pond.

8.4. Waves Due to a Periodic Disturbance: Radiation

Condition
In this section we consider water waves generated by
a pressure fluctuation which is periodic in time. We commence
§8. Application to partial differential equations 121

by investigating the solution of the hydrodynamic equations

when the water is initially at rest with surface elevation
n = 0, but subject to the pressure fluctuation

op = 1jI(x) cos (nt), t::: 0, n> o. (42)

For simplicity of the ensuing algebra, we replace the func-

tion cos (nt) by exp(-int) and afterwards take the real
part of the solution. The Fourier transform ~(w,z,t) is
again given by (29), but now (30) is replaced by

a2A(w,t)
--"-:~':'" + g I w I A(W, t) in 'l'(W) e
-int (43 )
p ,
at 2
which is obtained from the boundary condition (27). The
solution is

A(w,t) in'l'(w) e- int + B(w) e itlfWTg

p(lwlg - n 2 )
(44)
+ C(w) e- itlfWTg

The initial condition ~(w,z,O) = 0 gives

B(w) + C(w) = -in'l'(w) (45)

p(lwlg - n 2 )
and from (25), together with the initial condition nO(x) = 0

we obtain the further relation

B(w) - C(w) i~ 'l'(w) (46)

p(lwlg - n 2 )
It is now an easy matter to determine ~(w,z,t); viz

e it/lwlg
2 (n + If(;j''fg)
(47)
-i til wi g ]
+ ..;;.e_ _.........,==-
2 (n-/rwrg)

Using the Fourier transform of (26) and noting that

122 PART II: THE FOURIER TRANSFORM

a~/az = Iwl~, we obtain also the transform of the surface

elevation as
Iwi '1'( w)
H(w,t)
( 48)
+ i( n ) sin(t/lwlg)].
/Iwlg
Steady-State and Transient Solutions: The inversion integral
for n(x,t) is over real values of w, and it is evident
from (48) that there is no singularity at n 2 = Iwlg, al-
though each of the terms taken individually will lead to a
singularity there. We will use this fact to split n(x,t)
into two terms: one term (coming from the factor exp(-int)
in the square brackets) is the steady periodic response to
the periodic perturbation; the other, coming from the remain-
ing two terms, dies away as t increases. To do this, we
first write the inversion integral in the form

n(x,t) =
1
2'11
foo0 [H(w,t)e
-iwx
+ H(-w,t)e
iwx
] dw ( 49)

and then deform the contour to the one shown in Figure 2 which
avoids the point w = n2 /g. This enables us to consider the
various terms in (49) separately, and to write n(x,t) as
the sum of two functions n(s)(x,t) and nCt)(x,t), where

n
(s)
(x, t) e
-int I w ['I'(w)e
-iwx
+ 'I'(-w)e
iwx
] dw,
z:;;-P c n2 -wg

n(t)(x,t) 1
fc r
IWI'I'(w)
{ e
- iwx+it/Wg
_ e
-iwx-it,/Wg
2'11P 2;g L n + rwg
iwx+it/wg
+ 'I'(-w) { ~e ________ dw . (50)
n + Iwg
§8. Application to partial differential equations 123

1m (w)

c
Re (w)
w= Q~g

Figure 8.2

(t)
Now we may show that for any fixed x, Tl is of order
-1
t provided that ~(w) falls off sufficiently fast as
w+ 00. To do this, we proceed as follows. First observe
that the terms having the factor n + ~ as denominator may
-1
be evaluated as real integrals, and are of order t for
large t by Problem 2.1. Also, by the same consideration,
the contribution from the other two terms which come from
integrating along portions of the real axis are of order
-1
t . If we denote the integral around the small semi-circle
by I, the argument of w - n 2/g on the semi-circle by $,
and the minimum value of In - IWgI by tJ., then we have the

t
bound
sin $ d$
II I < A
-'IT
e ttJ.
(51)
= tr(t -1)
(t)
so that is indeed a transient term, and the steady
Tl

state is given by Tl (s) •

124 PART II: THE FOURIER TRANSFORM

Radiation Conditions: Suppose that 1/J (x) = 0 for I x I > a;

then ~(w) is an analytic function which grows no faster
than exp (alwl) as Iwl + 00. We may use this to estimate

n(s)Cx,t) for large x by deforming the integration contour

to the imaginary w axis. If x > a, we deform the contour
to the positive imaginary axis for the term involving
exp(iwx), and to the negative real axis for the term involv-
ing exp(-iwx). In the former case we must also pick up the
2
residue at w = rl !g. Hence

n(s) (x,t)

(52)

-1
For large x, the integral is of order x leaving the
°
secon d term as t h e maJor contrl°butlon.
° 7 This describes

traveling waves moving away from the disturbance at velocity

g!rl. Similar considerations show that for x large and nega-
tive, we again have outgoing traveling waves.
Our separation of n(x,t) into transient and steady-
state solutions has been quite tedious,and since we usually
only want the steady state component, we now ask how this
could be obtained directly. We therefore go back to the
original equations, and assume that the pressure perturba-
tion (42) is periodic for all t, -00 < t < 00. Consistent
with this we must also assume that all other functions have
the time dependence exp(-irlt), which is equivalent to put-
ting B(w) = C(w) = 0 in (44). This leads to functions
~Cw,z,t) and H(w,t) which have poles on the real axis at
2
w = ± rl !g; consequently,the inversion contour must avoid
§8. Application to partial differential equations 125

these poles. All that our analysis of the transient terms

has achieved has been to indicate the appropriate choice of
contour. We could have done this more easily, however, by
one of two methods.
Ci) We could analyze the expression for n(s)cx,t) obtained
for the various contour choices, and choose the contour
which gives outgoing waves.
Cii) We could replace n by n + ie, so that the driving
force is increasing exponentially. By taking the limit
e + 0, we then recover the correct result. This procedure is
known to physicists as "turning on the perturbation adiabati-
cally" since the effect of the exponential growth is that
there is no driving force for t + _00.

Either of these procedures is simpler than the above

analysis. The condition that the steady-state solution has
outgoing waves only is known as a radiation condition.

Problems
1. Find the stationary temperature distribution u(x,y) of
a semi-infinite body y > 0, i f the boundary is held at
the temperature

2.
u (x, 0)
f' 0,

Find the stationary temperature distribution

Ixl < a
Ixl < a.

u(x,y) of
a quadrant x ~ 0, y ~ 0, if the face y = 0 is held at
zero temperature while the other, x = 0, is thermally in-
sulated for y ~ b, while heat flows into the strip
o~ y < b at a constant density q. Find the distri-
bution of heat flow through the face y = o.
126 PART II: THE FOURIER TRANSFORM

3. If a function u(x,y) satisfies Laplace's equation in

the quadrant x ~ 0, y > 0, and if it also satisfies the
boundary conditions

U x (0, y) f (y) , Y > a

u (x, 0) a x > a

r
then show that
u (x, 0) 2 t f(t) dt.
y 'IT
a x 2+t 2
4. Show that the solution ~(x,y) of Laplace's equation in
an infinite strip _00 < x < 00, a~ y < a, subject to the
boundary conditions
~ (x, 0) f(x) ,
Hx,a) g (x),

is
f
f(t)dt
00

¢ (x, y)
za Sln (~)
1 .
a
~f
l-oo cosh'IT(x-t)/a - cos'ITy/a

g(t)dt 1
cosh'IT(x-t)/a + cos'ITy/aJ .
[Use Problem 7.26.]

5. Derive the solution to Problem 4 when the boundary condi-

tions are
(i) ¢Cx, 0) f(x) , ¢ (x,a) g (x),
Y
(ii) ¢ ex, 0) f(x), ¢ (x, a) g (x),
Y
(iii) ~ (x, 0) f(x), ¢y(x,a) g (x).
y

6. Solve Problem 4 in the special case

( Va' Ixl ~ a
(i) f(x) = g(x)
La, Ixl < a

L
( Va' Ixl ~ b
(ii) f(x) g (x)
a , Ixl > b
§8. Application to partial differential equations 127

7. Investigate the solutions of Laplace's equation in the

semi-infinite strip 0 ~ x ~~, 0 ~ y ~ a, using the
Fourier sine or cosine transform as appropriate to the
boundary condition at x = O.

8. Show that the potential due to a point charge q placed

on the axis of an infinite conducting cylinder of radius
a is

q 2 I~ KO (w a )
cp(r,z) - ~ 0 IO(wa) IO(wr)cos(wz) dw.
1r2 +z2
9. The end of a semi-infinite cylinder o< r ~ a, 0 ~ z ~ ~

is held at constant temperature TO' while the cylindri-

cal surface is held at zero temperature. Show that the
steady temperature distribution is given by

2 foo IO(wr) sin(wz) dw

u(r,z) = TO [ 1 - ~ 0 IO(wa) w

where IO(x) is a modified Bessel function.

10. By taking the Fourier sine transform in x, solve the

one-dimensional diffusion equation

on the line x > 0, subject to the boundary conditions

u (0, t) f(t) , t > 0,

u (x, 0) g (x), x > o.

11. Solve Problem 9 if the first boundary condition is re-

placed by
Ux (0 , t) = f (t) , t > 0,

using the cosine transform.

128 PART II: THE FOURIER TRANSFORM

12. Show that, if in considering (51) we took the con-

tour around the pole on the other side, the conclusion

(t)
regarding n would be invalid.

13. Consider waves on water of finite depth h generated

by an initial displacement nO(x) of the surface eleva-
tion. In particular, investigate the asymptotic form
of the solution if

Ixl < a
Ix I > a.

14. Show that the radiation condition of Section 9.4 applies

to the generation of water waves on water of finite
depth by a periodic pressure fluctuation.

15. Consider (two-dimensional) waves on a stream of uni-

form depth h, whose unperturbed motion is a uniform
velocity U in the positive x direction. Then the
velocity potential may be written as Ux + ~(x,z,t), and
the free surface conditions, after linearizing, are

n a~ a~
~ + gn + at + U ax 0,

an + U an _ a~ 0
at ax az .
Show that if p(x,t) = p(x) cos(nt), t > 0 with the
motion undisturbed initially, then the following be-
haviour is predicted: 8
2
(i) If U > gh, the disturbance dies out both upstream
and downstream of the region where p(x) is non-

zero.
[8. Application to partial differential equations 129

(ii) If U2 < gh, the disturbance dies out upstream,

but at any downstream point there is, after suf-
ficient time has elapsed, a steady periodic dis-
turbance.

Footnotes

1. Note that ~ is not a meromorphic function even if ~

is.

2. This problem anticipates some of the discussions of

Section 11.

3. The standard reference on water waves is STOKER (1957).

4. A lucid exposition may be found in CURLE &DAVIES (1968),

Ch. 21.

5. This follows because in this case

sin(a/lxIHa ± S) 2
'" 1
(al Ixl) (a±S)2

when (a± S) '" o.

6. STOKER (1957), Ch. 4.

7. The result follows from Watson's lemma.

8. This problem is considered by K. K. Puri, J. Eng. Math.

(1970), i, 283.
 130 PART II: THE FOURIER TRANSFORM

§9. GENERALIZED FUNCTIONS

The subject of generalized functions is an enormous
one, and we refer the reader to one of the excellent modern
books l for a full account of the theory. We will sketch in
this section some of the more elementary aspects of the
theory, because the use of generalized functions adds con-
siderably to the power of the Fourier transform as a tool.

9.1. The Delta Function

Generalized functions have their origin in Dirac's
delta function, denoted o(x-x O)' which is typically defined
in books on quantum mechanics by: "o(x-x O) is zero every-
where except at x = x o' where it is infinite; moreover it
has the property that

for any function which is sufficiently well behaved." Now

it is evident that this definition is inconsistent, since
if o(x-x O) is a function in the ordinary sense, then
fo(x-xO)f(x)dx = 0 regardless of whether or not o(x-xO)
is infinite at Thus if we wish to use (1), we
must generalize the concept of a function so as to give the
required formula (1) a precise meaning.
In the applications of mathematics to physical prob-

lems, functions are used to represent variables. For example,

E(t) might represent a voltage at time t. Now it is im-

possible to observe the instantaneous value of a voltage;

we can only measure the effect of the voltage acting dur-
ing a finite time interval. To consider a concrete example,
suppose that the measuring process is linear, so that the
§9. Generalized functions 131

measured value B(t) is

E(t) = f: k(t ' ) E(t-t ' ) dt'. (2)

In this situation, E(t) can not be measured for any choice

of function k, and so a theory which dealt with values of
E(t) directly would be sufficient. The essential difference
between using E(t) and E(t) is seen more clearly by set-
ting t = 0 in (2), for then we see that E(O) is a
function of the function k(t ' ). Functions which act on func-
tions rather than numbers are usually called functiona1s;
-
E(t) is a functional, assigning a value to each pair of
functions k(t ' ) and E(t).
Other examples of functiona1s are the Fourier and
Laplace transforms. The Fourier transform, for instance,
assigns the value F(w) to the function pair f(x) and
exp(iwx). The essential difference between these examples
and the generalized functions which we are about to define
is that we may evaluate the functiona1s by classical methods
involving the use of functions. We have already seen that
there is no such interpretation possible for (1); rather
we may define the delta functional by

(3)

instead of (1). This we will now proceed to do, using a

more convenient notation than (3).

9.2. Test Functions and Generalized Functions

We begin by defining the range of our generalized
functions; that is, we define the functions, called "test
132 PART II: THE FOURIER TRANSFORM

functions", on which the functionals act. We choose 2 the

set of all complex-valued functions of a real variable having
the properties:
(i) Each function ¢(x) has derivatives of every order for
all x,
(ii) Each function ¢(x) is zero outside some finite inter-
val a < x < b. This interval is arbitrary, depending
on the particular test function.

An example of a test function is

2
e-l/(l-x ), Ixl < I
¢ (x) = { (4 )
o , Ixl > 1.

The conditions imposed on the test functions are very re-

strictive,3 so it is reassuring to note that for any con-
tinuous function f(x) which is absolutely integrable there
are test functions which are arbitrarily close, i.e., for
any £ > 0 we may find a test function ¢(x) such that

I f (x) - ¢ (x) I < E, - 00 < x < 00. (5)

Such a function may be constructed as follows:

Choose a so that If(x) I < £ for Ixl > a, and

then let the functions ¢a(x) be constructed from (4) by

¢ (x) = p(x/a) (6)

a foo ¢(x/a) dx
00

Then it is not difficult to show that the functions ~a(x)

defined by
a
~a(x) = J ¢a(x-x') f(x') dx' (7)
-a
are test functions, and that we may choose a O so that for
§9. Generalized functions 133

all a > aO' Wa(x) satisfies (5).

Properties of Test Functions: Some of the simplest and most

useful properties of our test functions are as follows:

(i) They form a linear space. In particular this means

that a finite linear combination of test functions is again
a test function.

(ii) If ~(x) is a test function and f(x) an infinitely

differentiable function, then f(x) ~(x) is again a test
function.

(iii) The Fourier transforms of test functions have a par-

ticularly simple form. Suppose that ~(x) = 0 for Ixl > a;
then
~(w) = fa-a ~(x) e
iwx
dx. (8)

Now this integral may be differentiated with respect to w,

so that it is an entire function. 4 Moreover, if we write
w = cr + iT, then

1~(w)1 < J~a 1~(x)1 e- TX dx

(9)
< A e alTI

so that ~(w) is an entire analytic function whose growth

for large Iwl is bounded by an exponential function. Con-
versely, given any function ~(w) with these properties it
is easy to show that it is the Fourier transform of a test
function.

Linear Functionals: A complex-valued function f of a real

variable x may be defined as a rule which assigns a

134 PART II: THE FOURIER TRANSFORM

complex number [the value of f(x)] to each real x. The

key to the theory of generalized functions is that this con-

cept be relinquished in favor of a less restrictive one so

that (1) can be given a precise meaning. This is afforded by

the concept of a linear functional, that is, a rule, de-
noted <f, >, which associates with every test function ~

some complex number <f,~>, such that

(10)

where ~ and ware test functions and a and Bare

arbitrary constants. The important thing to note is that the
linear functional assigns a value to each test function, not
to each value of x.
An important class of linear functionals is the fol-
lowing: if f(x) is any function which is integrable, then

we define <f,~> by

J~oo f(x) ~(x) dx. (11)

Generalized Functions: Since the concept of continuity is of

prime importance in the theory of ordinary functions, we de-
fine a similar concept for linear functionals. We will say

that the sequence ~n of test functions converges to

zero if there is some interval Ixl ~ a outside which all
the n
ex)
~ vanish, and inside which they converge to zero
uniformly. Further, we will say that a linear functional
<f, > is continuous if the sequence of numbers <f'~n>

tends to zero whenever the sequence of test functions ~n

converges to zero. Finally, we define generalized functions

as the set of all continuous linear functionals acting on a
set of test functions. In particular (11) defines a gener-
§9. Generalized functions 135

alized function for any integrable function f(x), since it

is easy to show that it is continuous. s Functionals of this
type, llhich correspond to ordinary integrable functions,

are said to be regular generalized functions. 6 All others,

(e.g., the delta function) are said to be singular.
We will often denote the generalized function <f, >
simply by f. Thus we refer to the delta function as

o(x-xO)' although it is not a function of x, and what is

actually defined is the operation on test functions

(12 )

It may readily be seen that this defines a continuous linear

functional, i.e., a generalized function.

Addition and Multiplication: Suppose that f and g are

generalized functions and ~(x) is infinitely differenti-
able (but not necessarily a test function). Then we define
the generalized functions f+g and ~f by

<f+g,¢> = <f,¢> + <g,¢>, (13 )

<~f,¢> = <f,~¢>. (14 )

Equation (14) is possible because ~¢ is a test function.

As an example we define the generalized function x o(x):

<xo(x),¢(x» <0 (x) ,x¢ (x) >

O'¢(O) (15 )

0,

which we write symbolically as

x 0 (x) = 0, (16)

since the zero generalized function maps every test function

136 PART II: THE FOURIER TRANSFORM

to zero. Equation (16) shows that the equation x f(x) = 0

has as one possible solution f(x) = A o(x), where A is an

arbitrary constant. This is quite different from the situa-

tion with ordinary functions, for which the solution would be

f(x) = 0 for x r 0, f(O) = A, giving the zero generalized
function when this is substituted in (11) regardless of the
value of A.

Generalized Functions over Finite Intervals: The test func-

tions considered above are defined in the infinite interval
-00 < x < 00. It is sometimes useful to use a finite interval
a < x < b. In this case the additional restrictions
¢ (n) ( a ) -_ ¢ (n) (b) -_ 0, n 0,1,2, ... .
are applled. This en-
sures that the generalized functions have the same proper-
ties as for an infinite interval.

9.3. Elementary Properties

The manipulations of functions which are commonly use-
ful in applied mathematics, such as differentiation, in-
finite summation, and changes of order of limiting processes,

are accompanied by various restrictions on their validity.

Nevertheless these conditions are often ignored in applica-

tions; for example, the delta function arose from the chang-
ing of orders of integration in quantum mechanics. We
shall show in this section that the natural extensions of
everyday concepts from functions to generalized functions
leads to the lifting of many restrictions, and this is why
generalized functions find so many applications.

Transformation of Variables: Given an integrable function

§9. Generalized functions 137

f(x), we may define a regular generalized function by (11).

Suppose, however, that we wish to use the function f(g(x))
in (11), where g(x) is an infinitely differentiable mono-
tone function satisfying g(x) + too as x + too. We may
relate the two generalized functions by

<f(g(x)), $(x» = J~oo f(g(x)) $(x) dx

roo [f(g)$(x(g))/g' (x(g))] dg

(17)
= <f(g) ,1fJ(g) > ,

w(g) = Hx(g))/g'(x(g)).

Because of the properties of g(x), x(g) exists and is in-

finitely differentiable, and thus W(g) = $(x(g))/g' (x(g))
is a test function of g, so that the last line gives the
meaning of f(g(x)) in terms of f(x). We may use this
relation to define the meaning of f(g(x)) when f(x) is a
singular generalized function; if f(x) symbolically repre-
sents such a generalized function, then f(g(x)) is defined
by (17) also. In this case, the intermediate calculations
have no significance as the steps in an argument, rather they
lead to a final result which is true by definition. As an
example, o(g(x)) is the functional

<o(g(x)),Hx» = Ha)/g'(a), (18)

where a is the unique solution of g(a) = O.

Differentiation: We again commence with a regular general-

ized function (11) where the function f(x) has a first
derivative which is also differentiable. Then we may define
138 PART II: THE FOURIER TRANSFORM

the generalized function f' (x) by (11), and relate it to

the original generalized function by integrating by parts,
viz.
<f' ,cp> J~oo f' (x) CP(x) dx

-J~oo f(x) cpr (x) dx (19)

= - <f,cp'>.

Now we define the derivative of an arbitrary generalized func-

tion by <f' ,Cp> <f,cp'>, and also denote this definition
by
Joo
_00 f' (x) ¢Cx) dx =
_Joo
_00 f(x) cpr (x) dx. (20)

If f(x) is not a function satisfying the conditions for

validity of equation (19), then (20) is a symbolic way of
writing the definition. It emphasizes the important principle
that we define the properties of generalized functions so
that desirable manipulations are always valid.

Example I: Consider the Heaviside step function, hex), which

has the value unity for x > 0 and zero for x < O. In the
ordinary sense it may not be differentiated, but as a gen-
eralized function this causes no trouble. From (20) we have

J~oo h' (x) CP(x) dx = -f~oo hex) cpr (x) dx

(21)
¢CO) ,
which is the defining relation for the delta function, hence

h' (x) = 0 (x) • (22)

Example 2: Using (22) we may differentiate--in the

§9. Generalized functions 139

generalized sense--a function whose only problem is a finite

number of finite discontinuities. We have represented such a

f~

----------~----r---~r_----~~------------~x

Figure 9.1

function schematically in Figure 1. In precise terms we as-

sume that we may write
n
f(x) = f 1 (x) + L Ak h(x-xk)' (23)
k=l
where f 1 (x) is differentiable in the ordinary sense. Then
we have
n
f' ex) = f1' (x) + L Ak 0 (x-X k ) (24)
k=l
for the generalized function.

Convergence Properties: A sequence of ordinary functions

fv would be said to converge to the function f if for
each x the numbers fv(x) converged to the number f(x).
Similarly, we shall say that the generalized functions fv
140 PART II: THE FOURIER TRANSFORM

converge to the generalized function f if for each test

function ~(x) the numbers <fv'~> converge to the number
<f,~>. This definition is the same for integer values of v
or continuous values of v. We give two examples:
(i) Let
-ax
e hex), a > O. (25)

Then for any test function ~

lim
a+O
foo _00
f (x)
a
~(x) dx I: ¢Cx) dx; (26)

hence we write

lim hex). (27)

a+O

(ii) Let

V, Ixl < 1/2v

fv (x) = { (28)
0, Ixl > 1/2v.

Then we have, for any test function ~,

f f1/2V
00

lim fv(x) ~(x) dx = lim v ~ (x) dx

v+oo _00 v.... OO -1/2v
(29)
¢CO) ,

which may be written as

lim f (x) = o(x). (30)

v+oo v

This example shows that a sequence of regular generalized

functions may converge to a singular generalized function.

Regular and Singular Generalized Functions: We will state

7
here two important facts without proof. First, if a
§9. Generalized functions 141

sequence of generalized functions converges, then it con-

verges to a generalized function. Second, every generalized
function is the limit of a sequence of regular generalized
functions. As an example of the first property, we mention
the generalized function x -1 This has no meaning as a re-
gular generalized function, since the integral

(31)

diverges. However, the functions f e; (x) = l/x for Ixl > e;

and 0 for Ixl < e; define regular generalized functions,
and the limit as e; + 0 is also defined for all test func-
tions. Hence we may define the singular generalized function
-1
x by

r
_00
~ (x) dx
x
lim
e;+O
f
Ixl>e;
~(x) dx.
x
(32)

This is well known as the principal value of the integral.

Differentiation of Sequences: For any convergent sequence of

generalized functions and any test function ~,we can
write
lim <f' ,~> lim<fv'~'>
v
<f,~'> (33)

<f' ,~>.

Hence we can reverse the order of taking a limit and dif-

ferentiating. In particular, if fv is the partial sum of

an infinite series whose sum is f, then (33) shows that we

may differentiate term by term.

As an example, consider the series sin x + 21 sin 2x

+ !. sin 3x + which converges uniformly to a function
3
142 PART II: THE FOURIER TRANSFORM

f(x) equal to (n-x)/2 for 0 < x < 2n and periodic with

period 2n for other real x. For any test function, we
can multiply this series by ~(x) and integrate term by
term, so that we can write

f(x) sin x + 1 S1° n 2x + 1 sin 3x + ... (34 )

2 3

in the sense of generalized functions. Differentiating term

by term we obtain the further relation

...
00

-1 + cos x + cos 2x + cos 3x + n L o (x- 2nn). (35)

2 n=-oo
If we apply this to a test function Hx) with Fourier
transform <!> (w) , then on writing the cosine as a sum of com-
plex exponentials we obtain
00

<!>(n) H2nn), (36)

n=-OO n=_OO
which is known as t h e P01sson summat10n f ormula. 8
0 0

Consider the functions ge(x) = ~n Ixl h(lxl - e),

which converge to the integrable function ~n Ixl as e + O.
Now g~(x) = h(lxl - e)/x + ~n e [o(x-e) - o(x+e)], and on
taking the limit e + 0 we readily show that

lim <g' (x) ,Hx» = d:., Hx». (37)

e+O e x
Thus d(~n Ixl)/dx = l/x. Furthermore, we can apply this re-

suIt to ~n (x+iy), y > 0, where the branch of the logarithm

is defined by

~n (x+iy) = l ~n (x2+yZ) + i arctan(y/x). (38)

Z

As Y + 0, the function ~n (x+iy) converges to a function

f(x) defined by
§9. Generalized functions 143

f(x) : ~n Ixl + i n he-x). (39)

Differentiating and using (37), we obtain the useful result9

lim 1 1
i n o(x), (40)
y+O x+iy x
-1
where the generalized function x is the principal value
integral when applied to a test function.

9.4. Analytic Functionals

We have already shown that the Fourier transform of a
test function is an entire analytic function of w which
grows at most exponentially for large Iwl. Let ~(w) be
such a function; then it is the Fourier transform of some
test function ~(x). Hence for real w we can write

1~(w)1 : IJ~oo ~(x) e iwx dxl

< foo 1~(x)1 dx : A (41)

_00

so that ~(W) is bounded as Iwl + 00 Moreover, the nth

derivative of ~(x) is also a test function, so its Fourier
transform is bounded for real w. Applying (7.27), this
means that
n : 0,1,2, ... (42)

i.e., the functions ~(w) falloff faster than any finite

power of w. Also, by using (7.27) we see that ~(w)

is infinitely differentiable. Thus we may use this set of

functions lO to set up generalized functions exactly as for
the original test functions, and all the properties we have
proved above will again apply.ll In particular, regular
144 PART II: THE FOURIER TRANSFORM

generalized functions corresponding to (11) may be construc-

ted; for integrable functions F(w) we write

<F,W> = I~oo F(w) W(w) dw. ( 43)

A particularly useful class of generalized functions

over the test functions W(w) makes use of the fact that the
latter are entire analytic functions. Then we define an
analytic functional as

<G,W> = fr G(w) W(w) dw, (44)

where r is a given contour whose specification is an inte-

gral part of the definition of G.

Examples
-1
Ci) Consider the function G(w) = (w-w o) ,where is
real. We may use it to define two different analytic func-
tionals, namely

( 45)

where a > O. From the property of residues, we see that

(46)

which is written, in the notation of generalized functions,

as
(47)

(ii) Motivated by the last example, we consider the func-

tional
§9. Generalized functions 145

1
27Ti
I q,(w) dw ( 48)
C w-a

where C encircles the point w= a in a positive direc-

tion. From residue theory, the value of the integral is
~(a), hence (48) is the generalized function o(w-a). Ana1y-
tic functiona1s are thus seen to encompass a wider class of
generalized functions than regular functiona1s, at least in
some respects.
2
(iii) The function exp(w ) may be used to construct the
analytic functional

w2
<F,q,> e q, (w) dw. ( 49)

9.5. Fourier Transforms of Generalized Functions

The concept of the Fourier transform of a generalized

function is a very powerful one and has many practical ap-
p1ications. To motivate the definition, we first consider
those regular generalized functions which are constructed
from ordinary functions having Fourier transforms for real
w. Then Parseva1's relation (7.34) may be applied in either
of two ways, viz.

f~oo f(x) ¢(x) dx

(50)

The integrals on the right hand side are particular types of

analytic functionals, so we introduce the notation

(51)

Now (50) is a correspondence between generalized functions

146 PART II: THE FOURIER TRANSFORM

over the test functions ¢(x) and generalized functions over

their Fourier transforms ~(w). In the appropriate notation,
we have

<F,~(w» 27f <f,¢(-x» (52)

and

<F,~(-w» = 27f<f,¢(x». (53)

Either of these relations will define a generalized function

F corresponding to the generalized function f, even if f
is not regular. Therefore we use (52) and (53) to define
the Fourier transform of any generalized function. Thus by
definition every generalized function f has a Fourier
transform in this sense. We consider some simple examples.
(i)
(54)

Here (52) gives <F,~(w» = 27f¢(-x O)' Using the Fourier in-
version theorem to represent ¢(-x O), we obtain the analytic
.
f unctlona 112

(ii)
<F,~(w»
r_00
e
+iwxO
~(w) dw. (55)

- iwOx
f(x) e (56)

Here (52) gives

00 - iwOx
27f ¢(-x) dx
<F,~(w»
Loo e
(57)
27f ~ (w O)

which is equivalent to F = 27f o(w-w O)' From (48) above,

we may also write F as an analytic functional, even though
it is a singular generalized function.
§9. Generalized functions 147

Elementary Properties: Because of the restrictions on the

test functions, we may apply (7.27) repeatedly to obtain

.9"[cp(n)(x)] = (_iw)n ~(w)

(58)
~[(ix)n CP(x)] = ~(n)(w)

for arbitrary n. We use these results to show that they

also apply to arbitrary generalized functions. Considering
first the derivatives of the generalized function f(x), (58)
yields
2n <f(n) (x),CP(x» 2n (_l)n <f(x),cp(n) (x»
(_l)n <F(w),(iw)n ~(-w» (59)

«-iw)nF(w),~(-w»

and comparison with (53) shows that (-iw)n F(w) is the

Fourier transform of f(n)(x). A similar argument shows that
F(n)(w) is the Fourier transform of (ix)n f(x).
Another important property of Fourier transforms fol-
lows directly from their definition. Suppose the sequence of
generalized functions fv converges; then the sequence
of Fourier transforms Fv also converges. This is frequently
useful in finding the Fourier transform of a singular gen-
eralized function as we shall show.

Examples:
(i)
f(x) <5 I (x)

(60)
<F,~>
(co (- iw) Hw) dw
(E)
f(x) = x
(61)
F(w) -iw <5 I (w)
 148 PART II: THE FOURIER TRANSFORM

(iii) -iw x
f (x) e 0 hex), (62)

where hex) is the unit step function. The simplest approach

is to define the regular generalized functions f£(x)
exp(-£x) f(x), £ > 0, and then take the limit £ + O. In
this way we obtain

<F,il?> = lim
£+0 r _00 W
iil? (w)
- wo + i£
dw

(63)
i <l?(w) dw
Ie w - wo

where the contour e is parallel to the real axis, and

passes above the singularity at w =

Problems

1. Show that
b
¢Cx) = f ~a(x-y) dy,
a
with ~a defined by (4) , is a test function, and that

~(x) 1, a + a < x < b - <x.

2. Show that if f is infinitely differentiable, there is

a test function ~ with the property

~ (t) = f(t), a < t < b.

3. Show that if two regular generalized functions are equal

«f,~> = <g,~> for all ~), then at any point where

f(x) and g(x) are continuous functions,

f(x) g(x).
§9. Generalized functions 149

4. Show that

d 1
(i) ac;: o (o.x) = - 2" 0 (x)
a.
d 2
(ii) ac;: 0' (o.x) - 0' (x)
0. 3
(iii) f(x) o(x) = f (0) o (x)

(iv) f(x) 0' (x) - f' (0) o (x) + f(O)o' (x)

(v) o(e o.x - S) = 1 o(x - ~nS

o.S

r
a.

5. lim cos (wx) dw = 710 (x)

a+oo 0

a.
6. lim w sin(wx) 0
w+oo

v _v 2 x 2
7. lim - e o (x)
v+oo I7T
a. -o.lxl
8. lim "2 e o (x)
0.+
00

9. Consider the Fourier series

inx
00

1:
n=-oo
ane
where
a < A nk , 0, k > O.
n n "
t>how that i f P ~ k + 2, then

inx
1: a e = a O + f(P)(x)
n=-oo n
where
00 a
_n_ inX
f(x) e
n~o
n=-oo
(in)P
150 PART II: THE FOURIER TRANSFORM

10. Evaluate
00

L m > o.
n=-oo

11. xn 0 (m) (x)

r 0, m n <

= ~ (_l)n m! o(m-n)(x),
L (m-n)! m > n

o(m) (ax+b) 1 Ii (m) (x + b/a)

12.
ami al

13. lim Ii (x+l1) - Ii (x) Ii' (x)

11+0 11

14. Consider the regular functional defined by

A
x cjl(x) dx, Re (A) > -1.

Show that for Re(A) > -n-l, analytic continuation in

A yields
n-l
<x A,¢> = Jl x A[cjl(x)-cjl(O)-x¢' (0)- x ¢ (n - 1) (0) ] dx
+ 0 (n -1) !

+ I OO

1
A
x ¢(x) dx +
n
I
¢(k-l)(O)
k=l (k-l)! (A+k)
15. Show that

A f 0,-1,-2, . . . .

16. The generalized function x A is defined by

A iTIA fO A
<x_,¢> = e -00 (-x) ¢(x)dx, Re(A) > -1,

and by analytic continuation for Re(A) < -1. Show

that the generalized functions (x ± iO)A are related

to x A by
±
§9. Generalized functions 151

lim (x ± ie:) A
e:+0
A ±i7TA A
x+ + e x

and that they are entire functions of A.

-1
17. The generalized function x was defined in Section
9.3 as the principal value integral, and we showed that

(x ± iO)-l = x- 1 ; i7To(x).

We could define x- n by differentiating this expres-

sion, viz.
(x ± iO)-n = x- n ± (_l)n i7T o(n-1) (x)
(n-1)!
-n
Derive an explicit formula for x and in particular
show that
_1
(i) <x -n ,CP> 1 < x , cp (n -1) >,
(n -1) :

(ii) <x -2 ,CP>

J: cjJ(x) +CP (-x) - 2CP (0) dx.
x2

18. Show that the general solution of

m
x f(x) 0

is

where ak are arbitrary constants.

[First show that, if CP(x) is an arbitrary test func-
tion and w(x) is a test function satisfying W(O) = 1
and W' (0) = W" (0) = W(m-1) (0) = 0, then

m-1 xkcp(k) (0)

x(x) cjJ(x) - w(x) L
k=O k:
152 PART II: THE FOURIER TRANSFORM

is also a test function, for which <f,X> = 0.]

19. By using the Fourier transform, show that the general

solution of
f(n)(x) = 0
is
n-1 k
f(x) = L akx,
k=O
where are arbitrary constants.

20. Show that

a < Im(w) < a .

21. ?[h(x)] = 'IT o (w) + 1.

w

22. 9Te ax ] = 2'lT o (w- ia)

23. ?[sin ax] = -i'lT [0 (w+a)-o (w-a)]

+" ±i'lTA/2 A! (w ± iO)-A-1

24. ff[X~] = -1e
2'lTe±i'lTA/2
25. Y[(x ± iO)A] = w-A-1
(-A-1)! +
26. ff[x- 1 ] i'lT sgn (w)

27. ff[x - 2] -'lTlwl

im'lT m-1
28. ..'f [x -m] w sgn (w)
(m-1) !
29. " "b ut10n
Prove that every u 1 tra d 1str1 " 11 h as t h e conver-

gent Taylor series

an F(n)
L
00
F(w+a) Cw)
n=O n!
for all a.
§9. Generalized functions 153

30. Using the Taylor series, show that

Footnotes

1. ZEMANIAN (1965), GELFAND &SHILOV (1964).

2. This set of test functions is usually referred to as D.

Another important possibility is the set S of infinitely

differentiable functions which fall to zero faster than

any power of l/x as Ixl + 00. The reason for our
choice of D will become apparent later.

3. In particular, every trial function is identically zero

for Ixl > a for some a; any function of a complex
variable with this property must have essential singu-
larities at points on the real axis.

4. It may readily be shown that the Fourier transforms of

functions in S (see footnote 2) are again in S.
This is why we choose to use Dj the Fourier transforms
of functions in D are all entire functions satisfying
(9). This set of functions is usually denoted Z.

5. By some standard theorems of classical analysis, we may

write the series of inequalities

a
< max l¢n(x)lf If(x)ldx,
-a<x<a -a
and if the sequence converges uniformly to zero,
154 PART II: THE FOURIER TRANSFORM

so must the right-hand side of this inequality.

6. Thus the functional E(t) which we discussed in Section

9.1 is regular.

7. See GELFAND & SHILOV (1964), p. 15.

8. See problem (7.39).

9. The proof is easy because (37) and (38) are both regular,
whereas to prove (40) directly is more difficult.

10. It is interesting to note that D and Z (see footnote

4) are both subspaces of S, although D and Z have
no functions in common except zero.

11. Generalized functions defined over Z (see footnote 4)

are sometimes called ultradistributions, in distinction
from generalized functions over D, which are called
distributions.

12. If we ignore the fact that exp (iwt) is not a test

function and write

iwx
e o(x-x ) dx =
o
we lose essential information regarding the contour of
the analytic functional.
§10. Green's functions 155

§10. GREEN'S FUNCTIONS

One approach to the solution of non-homogeneous bound-

ary value problems is by means of the construction of func-
tions known as Green's functions. Historically, the concept
originated with work on potential theory published by Green
in 1828. Green's work has provided the germs of a much wider
formulation for solving a variety of eigenvalue, boundary
value, and inhomogeneous problems, particularly since the ad-
vent of generalized functions. We shall not attempt a sys-
tematic treatment in this book; rather we will discuss prob-
lems and methods where integral transform techniques are use-
ful. l In particular, we will discuss in this section problems
where the Fourier transform in one variable is applicable.

10.1. One-dimensional Green's Functions

We consider the linear second-order equation

L[y] p (x) y" (x) + q (x) y' (x) + r (x) y (x)

(1)
fex)

where the coefficients p(x), q(x), rex) and the function

f(x) are given. If we impose on the solution y(x) the
boundary conditions

aly(a) + a 2y' (a) o

(2)
bly(b) + b 2y' (b) o (a < b)

then we have a two-point boundary value problem of a type

quite common in applications. Any two solutions of this prob-
lem differ by a solution of the homogeneous problem

L[y] = 0 (3)
156 PART II: THE FOURIER TRANSFORM

together with the boundary conditions. If this latter prob-

lem has no solution other than y(x) = 0, then it follows
that the original problem can have only one solution, which
we will now construct.
Let uL(x) be the (unique) solution of the homogeneous
problem (3) satisfying the boundary condition at the left-
hand boundary x = a. Similarly, let uR(x) be the solution
satisfying the boundary condition at the right-hand boundary.
The functions uL(x) and uR(x) cannot be linearly dependent,
since there are no solutions of (3) satisfying both boundary
conditions. We may therefore use these functions to con-
struct a solution of the original boundary value problem by
the method of variation of parameters. We write

(4)

where we must determine suitable functions vL(x) and vR(x).

We have a considerable amount of freedom in choosing these
functions; following the usual procedure we require that

vL(x) uL(x) + vk(x) uR(x) = O. (5)

Substitution of (4) and (5) into (1) yields

vL(x) -[uR(x)/p(x) W(x)] f(x),

vR(x) [uL(x)/p(x) W(x)] f(x), (6)

W(x) uL(x) uk(x) - uR(x) uL(x).

We must also satisfy the boundary conditions; from (4)

and (5) we can write these as vL(b) = 0 and vR(a) = O.
Hence we can integrate the expressions given in (6) to obtain
§10. Green's functions 157

vL(X)
t uR(~)
f(~) d~,

r
x p (~) w(~)
(7)
uL(q
vR(x) = f(~) d~.
a p(O W(O
Equations (4) and (7) may now be combined to read

y(x) Iba g(x,O f(~) d~,

UL (x) UR (0 (8)
x < ~
p(q W(q
g (x,~)
uR(x) uL(O
x > C
p (0 W(~)

The function g(x,~) thus defined is known as the Green's

function for the given boundary value problem.

Adjoint Problem: Suppose u(x) and vex) are two functions,

twice differentiable but otherwise arbitrary. Using integra-
tion by parts, we can derive the identity

fba vex) L[u(x)]dx =

fb
a
v[pu" + qu' + r] dx

I: u[(pv)" - (qv) , + rv] dx (9)

+ [p(vu'-uv') + uv(q_p,)]b
a

This process has introduced a new differential operator,

which we denote by Lt; it is defined by

Lt [v] = (pv)" - (qv) , - rv. (10)

We also define adjoint boundary conditions as follows: vex)

satisfies the adjoint boundary condition to (2) if
158 PART II: THE FOURIER TRANSFORM

Jba vL[u] dx =
Jb
a
u Lt [v] dx (11)

for every u(x) satisfying (Z). From (Z) and (9), we see
this is equivalent to the boundary conditions

(lZ)
[blP (b) + bZp' (b) - bzq (b)] v(b) + bZp (b) v' (b) = O.

Finally, we say that the boundary value problem Lt[v] = f

subject to the condition (lZ) is the adjoint problem to
(1) and (Z).

We will denote the Green's function for the adjoint prob-

t
lem by g (x,q. It may be constructed by the same procedure
as was followed for g(x,~). We now proceed to do this.
First (see Problem 1) note that the left-hand and right-hand
solutions to the homogeneous adjoint problem are related to
uL(x) and u R(x) by
uL(x)
uI(x)
p(x) W(x)
(13)
t u R(x)
u R(x)
p (x) W(x)
Also, the Wronskians are related by p (x) wt (x) l/p(x)W(x).
Hence, from (8) we have

ut (x) u~ (0
r P (~) wt (~)
x < ~

1
t
g (X,O
4. (x) ut (~)
x > ~ (14 )
p (0 wt (~)

g(~,x).

Thus the two Green functions are related quite simply by

[10. Green's functions 159

Self-adjoint Systems: If the boundary value problem is

identical with its adjoint, it is said to be self-adjoint.
For boundary conditions of the type given in (2), a self ad-
joint boundary value problem is one for which q = p', so that
we may write
L[u] = (pu')' + ru. (15)

The Green's function for such a problem has the important pro-
perty of symmetry: g(x,~) = g(~,x). Furthermore, since we
may easily show that p(x) W(x) is a constant, (8) simpli-
fies to

uL(x) uR(~), x < ~

uR(x) uL(S) , x > ~

(16)

/:; = 1
p(~) W(~)

for a self-adjoint problem. These results are often of prac-

tical importance as a step in the process of constructing
Green's functions for partial differential equations.

10.2. Green's Functions as Generalized Functions

We restrict our comments in this section to Green's
functions for problems over an infinite range, that is, the
boundary conditions (2) are taken in the limit a + -~,

b +~. Obviously, g(x,~) defines a generalized function

which depends on a parameter x, viz.

(17)

Let us consider the differential equation for g(x,~). We

160 PART II: THE FOURIER TRANSFORM

operate on the function y(x) <g,~> with L; then by (8)

we have
Hx), (18)

which may be written as

L[g(x,S)] = o(x-S). (19)

We have derived (19) from the original definitions for L

and g; however, we may consider the reverse process. In-
stead of beginning with g(x,~) as a given function, we ask
for solutions of (19) subject to the boundary conditions

lim o
x-+-oo (20)
lim
x++oo

It may be shown that if p(x) f 0 for -00 < x < 00, then the
only solutions of the homogeneous problem L[g] = 0, where
g is a generalized function, are the solutions in the ordi-
nary sense. 2 (That is, they are the regular functionals con-
structed from the usual solutions.) Since we have assumed

that there are no such solutions, the Green's function, if it

exists, will be unique. Using (8), we ~herefore write

uL(x) uR(S)
g(x,O ----- h(~-x)
p (0 W(S)
( 21)
uR(x) uL(~)
+ h(x-~)
peo W(O

We may apply the operator L to this generalized function

directly using the results of Section 9.3, and show that it
satisfies (19).

Fourier Transform: This viewpoint is particularly useful if

§lO. Green's functions 161

the operator L is sufficiently simple for the Fourier trans-

form to be of use. As an example, we consider the equation

(22)

leaving aside the question of boundary conditions temporarily.

On taking the Fourier transform, we have

2 2
(w +a ) G(w,~) = -e iw~ . (23)

In addition, there is a restriction on the contour of inver-

sion, which is brought about by the fact that the Fourier
transform of o(x-~) is not the function exp(iw~) but an
analytic functional whose contour begins at w= -~ and ends
at w = +~. This information is insufficient to specify
the Fourier inversion integral for G(w,~), since there are
poles at w = ±ia. The choice of contour with respect to
these poles is equivalent to a choice of boundary conditions,
and is in no way restricted or determined by (22) and
(23), which have in fact yielded all their available informa-

tion. We illustrate by constructing three examples from

(23) .

(i) If we write

(24)

then by the application of standard techniques for contour

integrals we obtain

-1 -alx-~I
g(x,~) = 2a e . (25)

(ii) Choosing a different contour which is consistent with

the restrictions on the end points, we may write
162 PART II: THE FOURIER TRANSFORM

1 f~Y+oo eiw (!;-x)

g (x,S) dw, Y > a, (26)
21T lY- OO

and we obtain a different function, viz.

r 0, x < !;
Li
g (x,S) = ~ (27)
sinh[a(x-!;)), x > !;.

(iii) We may choose the contour of Figure 1, which gives

1 alx-!;I
g (x,S) 2a e • (28)

1m (101)

Re(w)

Figure 10.1

10.3. Poisson's Equation in Two Dimensions

We consider first the equation

2
'iJ g(r,r') OCr-r')
(29)
o(x-x') o(y-y')

where r = (x,y) is unrestricted. The problem does not

§10. Green's functions 163

have a unique solution, although any two solutions differ by

a harmonic function, i.e., a solution of the homogeneous equa-
tion. It can be shown quite readily that a solution is
g(r,r') = (1/2n) tn 1:-:' I. We shall show how to construct
this function from the defining equation via the Fourier
transform. On taking the transform with respect to x we
have
(~ - w2 ) G(w,y,x' ,y') = e iwx ' <5 (y-y'). (30)
ay2
Now we know that for any w this equation has more than one
solution; the corresponding Green's functions differ only by
harmonic functions. In order to obtain for G a form which
has an inverse Fourier transform in the usual sense, we choose
the solution of (30) as

.1L
iwx'
-e e -wly-y'l , Re(w) > 0
2w (31)
G(w,y,x',y') iwx'
e l.llly-y'l
e , Re(w) < O.
2w

It is important to note that this function is not analytic

across the imaginary axis and is also singular at w = O.
For the inversion contour we may use any curve which begins
at and ends at +00. [This requirement comes from taking
the Fourier transform of <5(x-x').] We choose a contour
which crosses the imaginary axis only once, at w = ia r O.
The resulting inversion integral is

1 Jia e-iw(x-x')+wly-y'
z;{ -00 2w
1
dw

- Jia
oo e-iw(x-x')-wly-y'l
2w
dw} (32)

- !n{E1 (a(x-x')+ialy-y' I) + E 1 (a(x-x')- ia ly-y' I)}.

164 PART II: THE FOURIER TRANSFORM

It is shown in Appendix C that the exponential integral

can be written in the form El(z) = -tn(z) + ~(z), where
~(z) is an entire function. Hence (32) may be written as

geE,!') = ~1T[R.n a + R.nlr-r'l + Re{H-a(x-x')-ia(y-y')}] (33)

The first and third terms are harmonic functions (depending

on the choice of contour through a), and we discard them.
This leaves us with (1/21T) tnIE-E'I, and because of its
fundamental significance in the solution of problems in re-
gions with boundaries, we call it the elementary solution of
(29). Denoting this elementary solution by e(:,!') we
have

~1T tnl r-r' I· (34 )

The significance of this function is that if we define the

function u(r) by

(35)

then it will satisfy Poisson's equation

This relationship holds for a wide class of functions and

generalized function~ feE); in particular it holds for any
function fer) which is absolutely integrable.

Poisson's Equation in a Bounded Region: Suppose we consider

a region R of the plane whose boundary we denote S. We
can solve the equations
§10. Green's functions 165

V2 g(E'E') 0(:-:'),
g(:,:') 0, r on 51' (36)

n'yg(r,r') 0, r on 52'

where 51 and 52 together constitute the whole boundary 5.

We will show that this function can be used to construct a
solution to the boundary value problem

2
V u(!) feE),
u(r) $1 (r), r on 51' (37)
~·~u(!) $2(E), r on 52'

and so it is called the Green's function for this boundary

value problem. To construct the solution, we first define the
region RE to be the region enclosed by the boundaries, with
a circle of radius E centered at r = r' deleted. R is
the region in which we desire the solution to (37).
2
In RE we have V geE,!') = 0, since the function g(!,E') -
e(r,r') is harmonic in RE , and e(:,:') is harmonic in

RE. Now we multiply (37) by g, and integrate over RE us-

ing Green's theorem to change an area integral to a line
integral around the boundary. Explicitly, using polar
coordinates for the small circle of radius E, this gives

fR g(r,r') fer) dr
E

IR E
[g(E-E') V 2U(E) - U(E) v2g(!'E')] dE
(38)
I5 [geE,!') $2(E) - $l(E) n'V g(E,E')] d~
+ ~w J:w [~n E ~~ + ~] dB.

On taking the limit E + ° and using the boundary conditions

for g(E,E'), this gives the representation
166 PART II: THE FOURIER TRANSFORM

u(r' ) IR g(::,::') f(::) d::

+ fSl ~l(r) n'~ g(r,r') d~ (39)

- f ~2(r) g(r,r') d~.

S2
Symmetry of the Green's Function: By a similar trick we may
show a most important result, namely g(r,r') = g(r' ,r). To
2 2
do so, we use the fact that V geE,!') = 0 and V geE,!"~) =0
in the region obtained by deleting circles of radius € with
centers at r' and r" from R. Application of Green's
theorem then gives

o I [g(~:.!') V2g(!,!") - g(!,!") V2g(!,E')] d:

Is [geE,!') !!.~ g(!,!") - geE,!"~) n'V geE,!')] d~ (40)

+ € I27T fg (!,r") g(r,r'fi

- - :"'1 de.
2n 0 L € € J
The products which occur in the integral over S are
zero because of the boundary conditions, and on letting
€ +0, the last integral yields g (r' ,r") = g (r" ,r')·

An Example: We will find the Green's function determined by

the equations

e(!-!,'), -00 < x < 00, 0 < y < ~,

(41)

g (!,!') = 0 r on the boundaries.

Taking the Fourier transform in x, we have

2 iwx'
d - w2 )
(0
y
G(w,y,x' ,y') = e e(y-y'), (42 )
§10. Green's functions 167

which is the same equation as (30). This time, however,

there are two boundary conditions, and on using (8) we
have
_eiwx' sinh(wy<) sinh[w(t-y»]_ ( 43)
G(w,y,x' ,y')
w sinh wt
< and y', arid y>
where y is the smaller of y is the 1ar-
ger of y and y'. The Fourier inversion of G gives a
unique result, provided we require g(E,E') + 0 as IE-E'I
+ O. If this is the case, the Fourier transform must exist
on the real axis, hence we choose the real axis for the in-
version integral. This point is not irrelevant, since G
has poles on the imaginary axis at wt = 2TIin, n = ±1,±2,
±3, ... ' Choosing a contour which does not pass between the
poles at ±2TIi will lead to a Green's function which grows ex-
ponentia11y as Ir-r'l + 00 We conclude, then, that the
solution of (41) is

g(x,y,x' ,y')
(44 )
e-iw(x-x') sinh(wy<) sinh[w(t-y»]d
= - ~TI [00_00 w.
w sinh(wt)

Relation to Images: In applications, it may be useful to

leave g as an integral, since it cannot be evaluated in
finite closed form. An infinite expansion may, however, be
obtained, as we now show. We first write (44) as twice
the integral over the range 0 < W < 00, and then apply the
expansion

sinh(wy<) sinh[w(t-y»] 1 Y e-(2n+1)wt

sinh(wt) 2 n=O ( 45)
168 PART II: THE FOURIER TRANSFORM

If we replace the lower limit of integration by E, we can ex-

press each term in the integral as an exponential integral.
Using the result El(-En) = tn(En) - y + O(En), we can then
take the limit E + 0 to get

g(x,y,x',y') = 2l1f Re I
n=-oo
tn rY - Y '+2nt +iCX-X')]
LY+Y'+2nt +i(x-x')
(46)

which is the form of solution which results from applying

the method of images to the problem.

Normal Derivative: It is of interest to calculate the normal

derivative, ~.~g, on the boundary. For y = 0, we obtain
from (44)
-iw(x-x')
-g (x 0 x' y') +
1
'Z'if
J00
e sinh[w(t-y')] dw. (47)
y " , _00 sinh(wt)
If y' + 0, the integral tends to 8(x-x'); if y' = t, it is
identically zero. Both of these properties are necessary for
the validity of (39) for the present system.

10.4. Helmholtz's Equation in Two Dimensions

It is frequently of interest to consider solutions of
the inhomogeneous wave equation

( 48)

which have the time dependence exp(-int). This converts

the wave equation to Helmholtz's equation

(49)

which is generally solved subject to boundary conditions.

We will consider Green's functions for this latter equation,
commencing with the elementary solution in two dimensions,
§10. Green's functions 169

that is, the solution of

(50)

with no boundary conditions required on a finite boundary.

The most straightforward method of solving (50) is to look
2 2
for solutions of (V +k ) e(E,E') 0 which have a singu-
larity at r = !'; these turn out to be the Hankel functions
of argument klr-r' I. The arbitrary constants may then be
chosen by applying Green's theorem to the integral

(51)

where ~ is a test function, exactly as was done in connec-

tion with (37-40) above. We leave this approach to the
problems, choosing rather to attack (50) directly. On
taking the Fourier transform with respect to x, we obtain

a2 2 2
[-2 - w + k] E(w,y,x',y') = e
iwx'
o(y-y'). (52)
ay
Suppose that we choose the integration contour for inversion
so that Re(,r~2_k2) ~ 0 on it; then the solution which is
bounded for large Iy-y' I is
iwx' _!w2_ k 2Iy-y'l
E(W,y,x',y') -e e (53)

and substitution into the Fourier inversion theorem gives ex-

p1icit forms for the elementary solution. Figure 2 shows
the only two contours consistent with the condition
Re(Iw-2- k 2) > 0; they lead to two different elementary solu-

tions. We will evaluate the inversion integral for contour

C1 . First, we introduce the polar coordinates r, e by
170 PART II: THE FOURIER TRANSFORM

1m (w)

branch cut

Figure 10.2.

x - x' p cos e
(54)
Iy-y'l p sin e

and then carry out the change of variables

W : -k cos (~+it),

- k cos ~ cosh t + ik sin ~ sinh t, (55)

_00 < t < 00

The path described by this new variable is a hyperbola, pas-

sing between two branch points at w: ±k, and is shown in
Figure 3.
In terms of t, we may write

-iw(x-x') - ~2_k2 Iy-y'l : ikp cos(~-e+it),

(56)

__d""w~_: _ dt,
Iw-Z"7
where we have chosen the branch of ~2_k2 which makes it
§10. Green's functions 171

\
\
\
'\
\
III =-k \

/
/
/
/
/
/
/

Figure 10.3

agree with the choice appropriate for C • Now we may de-

1
form the contour Cl to the hyperbola provided the contri-
butions from the arcs Ll L2 are of order R- l (see Figure
3). On Ll , this imposes the requirement ~ ~ e; on L2,
~ ~ e. The only consistent choice is ~ = e, whereupon the
Fourier inversion integral of (53) becomes

~ J(~ e ikp cosh t dt

4'1T _00
(57)
- i H (1) (k )
4 0 p •

The identification of the integral as a Hankel function fol-

lows from (20.66). Using the results of Section 20 we
may write down asymptotic forms for e(:,:') for small and
large Ir-r'l; they are
172 PART II: THE FOURIER TRANSFORM

1
e(!,!') - In ~n p, kp « 1

(58)
-3in/4
_ .::.e_-,-_
4
[~Jl/2
nkp
kp » 1.

The logarithmic singularity at p = 0 is characteristic of

problems involving the Laplacian operator in two dimensions;
the asymptotic form for large p shows that we have out-
going waves as the boundary condition at infinity. It is
easy to show that the inversion of (53) using the contour
C2 leads to

e(r,r')
(59)
- ikp
e , kp » 1.

This corresponds to incoming waves at infinity. Thus we have

two distinct elementary solutions; the decision on whether to
use either one, or a linear combination of the two as the
solution of (50) depends on the physical content of the prob-
lem at hand.

Helmholtz's Equation in a Bounded Region: The developments

of the previous section may now be replaced. If we define a
Green's function by

(V 2 + k 2 ) g(~,~') o(r-r'),

g(~>E') 0, (60)

n·Vg(r,r') 0,

it leads to the same representation (39) for the solution to

the inhomogeneous problem
§l0. Green's functions 173

(V 2 + k 2) uC:O -
fer) ,
u(r) tPl (~) , r on Sl' (61)

n'V u(!) tP2 (!) ,

-r on S2'

Furthermore, we can again show that

(62)

Specific examples are left to the reader (Problems 5-8).

Problems

1. Show that the functions defined in (13) do satisfy

the stated differential equation and the adjoint boundary
condition, and that the Wronskian is given by

p (x) W[u tL , u~ ] = ---:--::1,:-:-:_ _-:-

p(x)W[uL,U R] .
2. Show that, for a self-adjoint second-order differential
equation,
~x[P(X) W(x)] 0.

3. Find elementary solutions of

in two dimensions by using Hankel functions directly.

4. Show that the Green's function for Poisson's equation in a

three-dimensional half-space -00 < x < 00, -00 < y < 00,

z > 0, subject to g = 0 on z = 0 is

where
174 PART II: THE FOURIER TRANSFORM

R / ' (x-x') 2 + (y-y') 2 + (z-z') 2

R' = / (X-x,)2 + (y_y,)2 + (z+z,)2

S. Show that the Green's function for Helmholtz's equation

in the three-dimensional half-space z > 0, satisfying
og/on = 0 on z = 0 and having the form of outgoing
waves for large R, is

with Rand R' defined in Problem 4.

6. A metal disc of radius a is set into an infinite metal

wall, separated by a thin insulator. If the potential
of the disc oscillates at frequency n, show that the
potential far from the disc has the approximate form
J 1 (ka sin a)
e
ikR-int
ka sin a

;z---z
ck, tan a Ix~+y~

7. Show that the Green's function for Helmholtz's equation in

the strip < x < 00, 0 ~ y < a, satisfying g = 0 when

y = 0 or a, is
1 Ioo sinh(sy<)sinh s(a-y»
e
-iwlx-x'i
dw
g(x,x',y,y')
2n 0 s sinh (s~)

where
2 >
s2 w y< min(y,y'), y max(y ,y').
910. Green's functions 175

Using (45) express the solution as an infinite series

of Hankel functions.

8. Consider the boundary value problem 5

HO,y) 0, y < 0,

4>x(O,y) fey), y > O.

Show that, in polar coordinates R,6 (x R sin 6,

y = R cos 6) the function

x
1/2 t _ cos 6
v(R,cosex)
Icos ex + cos 6
d(cos ex)

is the solution, provided that v satisfies

1 av
0,
P ap
P R sin 6,
1',; R cos 6,
and Iz f( -1',;), < 0
av {
= 1T/i7I
1',;

lim p
p+O ap 0 1',; > O.

Using the Green's function for Helmholtz's equation in three

dimensions, deduce that
1 Ico 1/2 exp (ikl p2+ R2+2PR cos 6)
v(R,cos 6) = - - P f(p) dP.
'11"12 0
/ p2 + R2 +2p R cos 6

9. Consider the boundary value problem 6

f(r)} r
on S,
gC:O
176 PART II: THE FOURIER TRANSFORM

with k 1 2 f k22. Show that, if we define ~l and ~2

by

then the solution may be written

with the functions ~l , 2 determined by

2
Cv 2 + k ,2 H l,2 = 0,
l

WI , 2 Cf) = k 22 ,1 f(f) + g (r) , f on S.

Hence represent the solution in terms of the Green's func-

tion for Helmholtz's equation with satisfies g = ° on S.
10. Consider the boundary value problem

CV2+k2)2~ 0,
~
-
(r) fCr) "I
g(~) J
r on S.
V2~ C:)

By applying the limiting procedure kl + k2 to Problem

9, show that

WeE) = f fCE') ~n G(E,E') dS

- ~k Is {k 2 f(E') + geE')} a2
akan
GCE.E') dS'

11. Find an integral representation for the solution of

4
V cp(x,y,z) 0, -00 < x < 00, -00 < y < 00, z > 0,

subject to
§10. Green's functions 177

~(x,y,O) = f(x,y),
[V2~lz=0 = g(x,y).

Footnotes
1. Excellent accounts are given in STAKGOLD (1968) and MORSE
& FESHBACH (1953), Ch. 7.

2. See GELFAND & SHILOV (1964), pp. 39ff.

3. The difference between any two solutions of V2g

o(r-r') satisfies v2~ = 0; therefore we may write

4. The normal derivative of (47) is a Fourier transform

which is given in Problem 7.26.

5. This problem is adapted from a paper by W. E. Williams,

Q. J. Mech. Appl. Math., (1973), ~, 397, where some more
general results may be found.

6. Problems 9-11 are based on results given by G. S. Argawal,

A. J. Devaney and D. N. Pattenayak, J. Math. Phys. (1973),
!i, 906.
178 PART II: THE FOURIER TRANSFORM

§11. FOURIER TRANSFORMS IN TWO OR MORE VARIABLES

ILL Basic Notation and Results

The theory of Fourier transforms of a single variable
may be extended to functions of several variables. Thus, if
f(x,y) is a function of two variables, the function F(~,Tl)

defined by

F(~,Tl) = I~oo dx J~oo dy ei(~x + ny) f (x ,y) (1)

is the two-dimensional Fourier transform of f(x,y), and,

provided that the inversion formula (7.6) may be applied
twice, we have

An important point to note about this formula is that it in-

volves functions of more than one complex variable. The
theory of such functions is exceedingly complicated, and there
are no well developed techniques of the same generality and
power as for functions of one complex variable. Usually it
is necessary to treat each variable in turn, temporarily re-
garding the others as constant. Some of the subtlet'ies which
emerge will become evident in this and later sections, through
concrete examples.
An elegant notation may be used if the variables are
components of a vector; thus for a function fer) in n
dimensions we write

F(k) J f(::) ik' r n

e - - d r
(3)
1 -ils'r dn~.
f(O
(h)n J
r F(~)e
§ll. Fourier transforms in two or more variables 179

Elementary Properties: Formal manipulations, which we leave

to the reader as an exercise, lead to the following parallels
to the properties derived in Section 7.2.

1
(i) Derivatives:
~[~f(!) ] - i~ F (~) (4)

~[rf(r) ] - i!5YF (!5) (5)

or
~r ++ -ik
(6)
~k ++ +ir

(ii) Translations:

ik·a
~[f(!-~)] = e - - F(!5) (7)

(8)

(iii) Convolutions: If

n
her) = J g(r-r') f(:')d :', (9)

then
H(k) G (k) F (k). (10)
Also
g (r)] = 1 G(k') dnk'.
~[f(r)
(27T) n J F(k-k') (11)

(iv) Parseval relation:

J
n
f (r) g*(r)d r =
1
(27T)n
I F(!5) G*(!5)d n !5. (12)

Illustrative Example: As a simple application, we will re-

derive the results of Section 10.4 using a two variable

transform. The algebraic manipulations involved are trivial,
but the analysis of the inversion integral already exhibits
 180 PART II: THE FOURIER TRANSFORM

some interesting and illuminating subtleties. We want to

solve the equation Z

(17 Z + k 2) e(r) Ii (r) . (13)

Taking the two dimensional Fourier transform, we have

1, (l4 )
where

E(g) = J e io' -r~ e(~) d

Z
~. (15)

If we solve (14) for E(g) and substitute into (3b),

we obtain
1 Z
(2";)2 d g. (16 )

This solution is not unique, since the Fourier transform of

3
a delta function does not specify the inversion contour.
We denote the components of q by ~ and n; it is our in-
tention first to evaluate the n integral for each value of
~ which is needed, and then to integrate over ~. Great
care is needed at this point, since the n integral depends
critically on the position of the n contour relative to
each value of ~. Hence, we choose the ~ contour first;
our choice is indicated in Figure 1. On this contour,

o < arg(~Z_kZ) < TI, so we may define the function s(~,k) by

;-Z-Z-
s (k ,;) = I~ -k ,
(17)
o < arg(s) < TI/Z.

Turning to the n integral, we need to evaluate

I(~,y) = _I OO e-iny dn
(18)
-00 (n- is) (n+is)

We integrate along the real axis, since the poles lie off it
§ll. Fourier transforms in two or more variables 181

1m (~)

c
-k
Re (~)
k

Figure 11.1

by virtue of (17). By residues, we have

1(;, y)
1Te -slyl (19)
s

and on using this in (16), we obtain

-inx _1n2 _k 2 Iyl
e(r) = - !1T fe e
dn, (20)

a result already obtained in Section 10.4.

Use of Radiation Condition: The difficulty in the above

treatment arises because k is a real quantity, so that we
must choose an inversion contour which goes off the real axis.
One method of treating such problems, discussed in detail in

Section 9.4, is the application of radiation conditions.

Thus we replace k by k + ie:; then both components of q

may be confined to real values.

Changing to polar coordinates, with r defining the

polar direction, (16) becomes

182 PART II: THE FOURIER TRANSFORM

qdq J27f (21)

o
The a-integral is one form of Bessel's integral for JO(qr),
hence
1
e(r) = 2~"
J'0" JO(qr)
qdq k2 _ q2 (22)

This integral defines ~n analytic function of k for 0 <

arg (k) < 7f, and in particular, if k = ia, Problem 20.23

yields
1
- 27f KO(ar)
(23)
1
27f KO(-ikr).

Having evaluated the integrals, we must set £ = O. Equa-

tion (20.79) then shows that

e(r) = - ! H (1) (kr) (24 )

4 0

in agreement with (10.57).

11.2. Diffraction of Scalar Waves

The mathematical solution of diffraction problems is
generally very difficult, and explicit exact formulas are
known for only a small number of relatively simple cases.
Fortunately, a large number of problems of interest in op-
tics (and other fields) may be usefully approximated by a
method which is due to Kirchhoff. We discuss here the dif-
fraction of a scalar wave, satisfying the equation

(25)

by an aperture in a plane screen at z = O. The basic idea

is to express the solution at an arbitrary point in terms of
§ll. Fourier transforms in two or more variables 183

the values in the aperture; these aperture values are then

set equal to the strength of the incident wave, calculated
in the absence of the screen. (If we can calculate the aper-
ture function in the presence of the screen, the problem is
exactly solved.) In the following we denote the aperture
plane by A, and the remainder of the z = ° plane by B.
Also, corresponding to any three-dimensional position vector
r having components x,y, and z, we introduce the two-
dimensional vector s having components x and y.
We consider monochromatic waves, i.e. we set

Re(rI) > 0, Im(rI) > 0, (Z6)

where we have taken Im(rI) > ° as a radiation condition.

After determining the necessary solutions, we may set
Im(rI) = 0. We introduce the two-dimensional Fourier
transforms
iq.s Z
¢ (E)
Je - - <Hg,z)d g, (Z 7)

(~1T)Z f
e-ig.~
1>(g,~) ¢(~,z)d~. (Z8)

Substitution into (Z5) yields the following ordinary dif-

ferential equation for 1>(g,z):

d Z",
'!'

dz Z
-
Z Z
(q -k )1>
°, k rile, (Z9)

with the general solution

/2-Z 12:2
<Hg,z) F(g)e- z/q -k + G(g)ez/q~-k~ . (30)
184 PART II: THE FOURIER TRANSFORM

From (26) we have -n < arg(q2_k2) < 0 for all real q,

and thus we may choose the square root by -n/2 < arg(~2_k2)
< O. Hence the second term in (30) is an incoming wave, con-
trary to our physical requirement that the diffracted wave

be outgoing, so that G(g) = O. Thus F(g) = ~(g,O), which

may be obtained from ~(s,O) by putting z = 0 in (28).
Equation (27) may be written as
. r?2
= J d2 9 lq·S
e - - e
-z/q--k-

x
l(1 --
-
(2n)2
f d 2 s' e - iq·
-
__s'

_l_J d 2 s' ¢(s' 0)

4 2 - -'
xn td
J d2 iq'(s-s')
__ k2}
- z/q2_ __
__________
(31)
~e~- ~-~-

Qz g /c2
q -k
2
~n J d2~, ¢(~' ,0) h [e~kRj,
z2 + /s-s,/2.

The evaluation of the 9 integral, which leads to the last

step, is dealt with in Problem 15. This formula is particu-
larly appropriate if ¢ satisfies the boundary condition
¢ = 0 on B, for then the integral in (31) extends only over
the aperture A. Thus, in this case we have expressed ¢ in
terms of its value in A.

Fraunhofer and Fresnel Diffraction: In practice, these for-

mulas have their most important applications in an asymptotic

limit which we now derive. Suppose that the aperture is

small compared to z; then if we choose the origin to lie in

A, and we have /~'/« z in (31), we may write

§ll. Fourier transforms in two or more variables 185

R
.
~' ~' - (32)
'" r - r's' +
2r
This replacement is made in the exponential function; else-

where we simply write R '" r. The case where it is necessary

to retain the quadratic terms in (32) is called "Fresnel dif-

fraction," the simpler case, R '" r - f'~" is called

"Fraunhofer diffraction." For Fraunhofer diffraction, (31)
becomes

ikr
ze -
-
27rr2
f d 2s' e-ik~.?'
_- ~'I'(~' ,0)
(33)
27rz
-;Z e ikr F(k~/r).

Thus the angular distribution of the wave diffracted by an

aperture in this limit is determined principally by the

Fourier transform of the illumination of the aperture. If the

aperture is narrow compared with the wavelength, the diffrac-
tion pattern will be broad; converselY,a wide aperture pro-
duces little diffraction.

11.3. Retarded Potentials of Electromagnetism

In the classical theory of electromagnetism, the elec-
tric and magnetic fields E and H satisfy Maxwell's equa-
tions, which in c.g.s. units are

V'E 47rp,
y. x .§ 1m
c at
(34 )
V'H 0,

v x H
186 PART II: THE FOURIER TRANSFORM

where p and j are the sources, i.e., the charges and cur-
rents which generate the fields. It is usual to express E
and H in terms of a vector potential A and a scalar poten-
tial ~ via the relations

H v x ~,
(35 )
1 a~
at --y~,
E
c

together with the subsidiary condition (called Coulomb gauge)

V'A + 1:.c ~
at
= o. (36 )

After some simple algebra, it is readily shown that A and

are determined from

lV'
and by
~
- p

-~
1
2' 1
_ A
at Z -
4'TT
c
j ,

[ vZ _ ,,'atZ
cZ
] ~
- 4'TTp.
(37)

The problem is to express the potential in terms of the

sources by an explicit formula. Such a formula involves the
Green's function. We consider here the equation for ~,and

introduce the four-dimensional Fourier transform

@(~,w) = f ~(r,t) e
i(k'r+wt) 3
- - d r dt, (38)

with a similar definition for p(~,w). Equation (37b) now

reduces to simple algebra, with the solution

4'TT
p(~,w). (39)

From the product form of (39), we deduce that the Fourier

2 Z 2
transform of the Green's function is 4'TT/(k -w Ic ). This is
only defined once we apply boundary conditions. We take the
§ll. Fourier transforms in two or more variables 187

radiation condition Im(w) > 0, so that

iwr/c
1 e ( 40)
Z;Z r

The w is not defined in the usual sense; but since we are

dealing with generalized functions we use the result
~[<5 (t-t')] = exp (iwt') to get
<5(t-t' - 1.:-.:'l/c)
G(!-!', t-t') = (41)
I!-!' I
and
pC:' ,t-I.:-.:' lie)
<pc!, t) J I r-r' I
d3 ,
! '

A(r,t) 1
c
-j(r',t-lr-r'l/c)
- - ~ d 3;r.
, (42)
I r-r' I
This is called a retarded solution because the source at
(!' ,t') only influences the potential at r at a later time t

related by t' = t - I!-!' I Ie. Advanced potentials are ob-

tained by taking Im(w) < O.

Lienard-Wiechert Potentials: For a point charge q at posi-

tion !O(t) with velocity vet) = ~o(t), we have

( 43)

j (r, t)

<PCr,t) f <5 r'(r'- -r- -rO)

I_ 1
3
d r'
-,
(44 )
_0

1 ( <5(!'-!0) ~(t-I!'-!ol/c) d3r'.

( 45)
c J 1:'-':0 1
The integrals must be evaluated carefully, since
188 PART II: THE FOURIER TRANSFORM

_ __0 I/c)
__I'D = r(t-Ir'-r is a function of r'. The simplest pro-
cedure is to introduce a new variable s = E'-Eo; then it is
readily shown that the Jacobian of this transformation is
1 + R'v(t')/R, where R :0-: = and t' = t-R/C. Now the
delta function is simply o(~), so we have

<PC!' , t ) = -=-_e=-c~--,~-::­
cR + R'v(t')
C46)
ev(t')
~(!"t) =
cR+R'v(t')
These are the Lienard-Wiechert potentials.

Problems
Verify the following Fourier transform pairs.

-ar 2 2 + y2
1. f(x,y) e Ir, r x

F(/;;,T)) 2Trl ~
+a, k2 /;;2 + T) 2

2 2
2. f(x,y) xe-ar/r 2 , r x2 + y

F(I;,T)) = 2Trl; [1 k/A2+a 2 ]/k 2 , k2 1;2 + T) 2

-a 2r 2
3. feE) e r = (x ,y, z)
I 2 2
F(k) (71 3 2/a3)e-k 14a

-ar
4. fer) e Ir, r = (x,y,z)

471
F(k)
k 2+a 2
(1, r < a
S. fer)
lo, r > a

471
F(k) {sinCka) - ka cos(ka)}
k3
§11. Fourier transforms in two or more variables 189

6. Show that the vector operators (in three dimensions)

transform as
~~(:) ~ - i~ ~(~),

y.~(!) ~ - i~'Y(~),

'Vxu(r) ~ - ikxU(k).

7. Apply a two-variable Fourier transform to the problem

u(r,e,O) fer),
to show that

u(r,e,t) = -1 J
oo 2
pf(p)e- P /4Kt IO(pr/2Kt) dp.
4'ITKt 0

8. At a plane boundary z = a between two dielectrics,

the electrostatic potential satisfies the boundary condi-
tions

[e:a~/az] z=a- = [e:a~/az] z=a+

Using a two variable Fourier transform in x and y, find

the Green's function for Laplace's equation in an infinite
region which consists of a plane slab of material of
thickness ~ and dielectric constant e: in vacuum (di-
electric constant 1).

9. Solve the initial value problem

2
'V u(x,y,t)
1 a2u(x,y,t)
c2 at 2
u(x,y,O) f(x,y) ,
u t (x,y, 0) 0,

using a two-variable Fourier transform.

190 PART II: THE FOURIER TRANSFORM

10. Solve the boundary value problem

V4 u(x,y) f (x, y), _00

< x -< 00
, y > 0,
ux (x, 0) 0,
uy(x,O) 0,

by using a Fourier transform in x and a Fourier cosine

transform in y.

11. Solve the heat diffusion equation

au
at
2
KV U + <P

where <P(E,t) is the rate of production of heat,

subject to the initial condition

In particular, show that

2
u(!,t) If- 3/ 2
f f(!+2~ 1Kt) e- s ds
2
t
+ If- 3/ 2
f0 dt' J <p(!+2~ 1Kt'", t' )e
-s
ds.

12. The solution of the Dirichlet problem

(V2_A2)u(~,z) 0,

u (~, 0) g (~),

(xl x 2 '···, x n _l ),
may be written
u(~,z) f g(~') L(~-!' ,z) d!'.
Show that 4
L(~,z) = 2(A/2lf) n/2 Kn / 2 (AP),
2 2
P + z •
§11. Fourier transforms in two or more variables 191

13. Initially, the half-space x > 0 is at a constant tem-

perature. From time t = 0, the plane boundary x = 0

is held at the temperature fey). By applying the La-

place transform in x and the Fourier transform in y,
derive the formula

u(x,y, t) x Joo f(s) e-p2/4Kt d

7T
_00 P2 s,

x 2 + (y_s)2 ,

for the subsequent temperature distribution.

14. By using the Laplace transform in x and the Fourier

transform in y, show that the solution of the wave equa-

tion
2
'V u

in the region x > 0, -00 ~ y ~ 00, subject to the ini-

tial conditions
u(x,y,O) 0

ut(X,y,O) 0
and the boundary condition
u(O,y,t) = fey)
is
t < x/c
u(x,y,t)
Jyy+y'
_y ,
f(s)
p2
ds
-::/c;':;c2~t::;2F_=p:;;2 '
t > x/c

15. Evaluate the g integral in (31) using the methods

of Section 10.4.
192 PART II: THE FOURIER TRANSFORM

16. Show that the Green's function defined by the equations S

z>O, z'>O,

G(~-:>::r) 0, s = (s x , s y ,0),
~
z > 0,
is given by

GCE-E' ,y) e
-ik'(r-r')
~ ~ ~ d\.
1 + ik'v

6
Double Laplace Transforms: If f(x,y) is defined in the
quadrant x > 0, Y ~ 0, we define the double Laplace trans-
form F(p,q) of f(x,y) by

F(p,q) = J""o J"" f(x,y)e- px - qy dx dy = ~[f(x,y)].

0

Prove the following general properties of double transforms

[under suitable restrictions on f(x,y)].

17. ~[f(x+y)] peg) - F(I1)

p-q
where
F(p) Y[f(x)].

18. ~[f(x-y)]
peE) + P(q) f even
p+q
F(2) - F(g) f odd.
p+q

19. Yz [au/ax] pU(p,q) - UO(q)

where 7

UO(q) Y[u(O,y) ; y -> q] .

2
20. ~ [a 2u/ax 2 ] p U(p,q) - p IT (q) - IT (q)
1
0
where

ITI (q) Y[u (0 ,y) ; y -> q].

x
§11. Fourier transforms in two or more variables 193

21. Solve the partial differential equationS

au x > 0, y > 0
ax
subject to
u(x,O) = a(x)
using the double Laplace transform.
[Hint: U(p,q) must be analytic for Re(p) > a, Re(q)
> a, for some fixed a,a. This imposes a restriction on
the possible value of u(O,y), and thus determines the
solution uniquely.]

22. Solve the heat conduction problem

au a 2u
at = K , x > 0,
ax 2
u (x, 0) 0, JC > 0,

u(o,t) TO' t > 0,

using the double Laplace transform.

23. Consider the wave equation

a 2u 1 a 2u
a;z , x > 0, t > 0,
c 2 at 2
u (x, 0) f (x),
u t (x, 0) g (x),

u(O,t) o.
Show how the solution, which may be constructed by
D'Alembert's method, can be recovered using the double
9
Laplace transform.
194 PART II: THE FOURIER TRANSFORM

Footnotes
1. These results apply either to functions having the neces-
sary behavior at infinity to allow integration by parts,
or to generalized functions with no restrictions.

2. aCE) = a(x)a(y). The theory of generalized functions may

be extended quite simply to several variables, but we do
not need to concern ourselves with the details here.

3. See Section 9.5.

4. This result is given in I. N. Sneddon, J. Eng. Math.

(1974), ~, 177, together with a discussion of the connec-
tion with the half-space Dirichlet problem for Laplace's
equation.

5. This is an example of the collisionless linear transport

equation. See section 19.6 for an example of the use of
this Green's function in the solution of the linear trans-
port equation with collisions.

6. See DITKIN &PRUDNIKOV (1970) for more information on

double Laplace transforms.

7. We use the notation ~[f(x,y); y + p] so as to indicate

which variable is transformed. Thus ~[f(x,y); y + p]
is a function of x and p.

8. See J. C. Jaeger, Bull. Am. Math. Soc. (1940), 46, 687.

9. The application of the double Laplace transform to a more

general second-order partial differential equation in the
quadrant x ~ 0, y ~ 0 is discussed in K. Evans and E. A.
Jackson, J. Math. Phys. (1971), g, 2012.
 Part III: Other Important Transforms

§12. MELLIN TRANSFORMS

12.1. Definitions
In this and the next two sections we study the Mellin
transform, which, while closely related to the Fourier trans-
form, has its own peculiar uses. In particular, it turns
out to be a most convenient tool for deriving expansions, a1-
though it has many other applications. We recall first that
the Fourier transform pair can be written in the form

A(w) roo a(t)e iwt dt, a. < Im(w) < B , (1)

and

a (t ) = 12n I y+oo
i
.
1Y-00
A("')e-
~
.
1wt d w, a. < y < B. (2)

The Mellin transform and its inverse follow if we introduce

the variable changes

P iw,
x = et , (3)

f(x) a(.tn x),

195
196 PART III: OTHER IMPORTANT TRANSFORMS

so that (1) and (2) become

F(p) f: x p - 1 f(x) dx, a < Re(p) < S, (4 )

f(x) = 2~i
1 JC+ 1.· '" x- p F(p) dp. (5)
C-1'"

Equation (4) is the Mellin transform, and (5) is the Mellin

inversion formula. The transform normally exists only in the
strip a < Re(p) < ~, and the inversion contour must lie in
this strip.

12.2. Simple Examples

We now study three simple examples which illustrate
the most important and peculiarly useful features of the
Mellin transform.

f (x)
-ax
(i) e , a > 0, (6)

F (p) I'o" e -ax x p-l d x

(p-1) ! Re (p) > o. (7)

aP
By the inversion formula we thus have the integral representa-
tion

1
f(x) = 2~i J
C + 1.·'"
(p-1)! (ax)-Pdp, C = Re (p) > O. (8)
C-1'"

From the asymptotic behavior of (p-1)! for large p, we

readily conclude that the contour of the inversion integral

can be closed in the left-hand half-plane for any value of

x, leading to the expansion

-ax '" ~ r
e =
r=O
L r! (ax) (9)
§12. Mellin transforms 197

corresponding to the poles and residues of the integrand.

(ii) f (x) = (1 + Bx)-y, y > 0, larg 131 F TI (10)

r p-l dy
x p - ldx
(
00

F (p) B- P (ll)
fo (l+Bx)Y (l+y)Y

The substitution y = z/(l-z) reduces the integral to the

standard form

_p (p-l)!(y-p-l)!
13 , (12)
(Y -1) !

where for the integral to converge, we must have

o < Re(p) < y. (13)

The inversion formula then gives us

(y-l)! f(x) = 1
y-:-
TIl
fc-i
C+ ioo
oo
(p-l)! (y-p-l)! li3x)-Pdp, (14 )

where the contour separates the two sets of poles as indicated

in Figure 1.

Im(p)

poles of (p-O! poles of (y-p-1)!

Re (p)
------~--~~--~--~~----~--~--_M~--~--_M~
-3 -2 -1 c y Y+1 Y+2 y+3

Figure 12.1
198 PART III: OTHER IMPORTANT TRANSFORMS

In order to close the contour so as to utilize the

poles and residues of the integrand, we must first consider
the asymptotic form of the integrand for large Ipl. From
Appendix A, we see that

... 00
, (15)

and thus we can close in the left-hand half-plane if IBxl < 1

and in the right-hand half-plane if IBxl > 1. This leads
immediately to both ascending and descending expansions, a
common feature of the Mellin transform inversion.

Ascending Expansion: If we close the contour to the left the

poles are those of (p-l)! Evaluating the residues at these
poles we have
r
L ~
1 00

f(x) = (y-l) ! r! (y+r-l)! (Bx) ,

r=O (16 )
= 1 - y(Bx) +
y(y+l) (Bx) 2
2! - ...
This last expression is just the binomial expansion of f(x).

Descending Expansion: The poles of (y-p-l)! are at

p = r+y, r = 0,1,2, ... , with residues (_l)r+l/r! Therefore,
closing the contour to the right we have the expansion
00
( _l)r+l
(y-l)! f(x) = (-1) L (r+y-l)! (Bx) -r- y , (17)
r!
r=O
where the additional factor (-1) arises since we are clos-
ing the contour in the negative (clockwise) direction.
Written out explicitly the expansion is

f(x) = (BX)-y~
L
_ 1- + y(y+l) _ X(y+l) (y+2) + ..
Bx 2! (Bx)2 3! (Bx)2
J, (18)
§12. Mellin transforms 199

which is the binomial expansion valid for large values of

I exl·
(iii) The exponential integral is

El (x) = r
x
-u duo
_e_
u
(19)

Denoting by

r
n(p) the Mellin transform of E1 (x), we have

n(p)
r0
xp - 1 dx
1
e
-wx
-w-- dw

f OO dw foo x p - l e- wx dx (20)
1 w 0
(p-l) ! Re(p) > O.
p
Thus,
1 fC+l.· OO
El(x) = (p-l)! x- p ~ c > O. (21)
2~i C-l OO P ,

Closing the contour to the left, which is permissible because

of the asymptotic form of n (p), we recover a sum of residues
at p = 0,-1,-2, ... The pole at p = 0 is a double pole
with residue -in x y, where y is Euler's constant (equal
to the value of d in(a!)/da at a = 0) ; the other poles
are simple poles. Our ascending expansion appears as
00

- in x - y - L (22)
k=l
In this case, as with the exponential function [example (i)],
we do not recover a descending expansion, because there are
no singularities in the right-hand half-plane. The real
reason is deeper than this; the exponential function has an
essential singularity at infinity, and hence no expansion in

powers of l/x. It is easy to see, either from (19) or (20),

that El(x) ~ exp(-x) for large x; consequently we consider
200 PART III: OTHER IMPORTANT TRANSFORMS

the function

f(x) = eX EI(x). (23)

Taking the Mellin transform we obtain

fm
= J0 x p - l dx I e w
m -x(w-I)
F(p) dw

=Jm dw fm e-x(w-l) xp - l dx

f:
I w 0 (24)

(p-l)! w- l (w-l)-P dw

= (p-l)! (-p)! (p-l)!, 0 -< Re(p) < 1.

There are now poles in both half-planes, but we cannot close

the contour to the right and throwaway the integral around
the large semi-circle, because F(p) grows exponentially as
p ->- +m.

Ascending Expansion: Closing the contour to the left, we

must evaluate the residues at the double poles of [(p_I)!]2.
This can be done by writing
2 x- p
F(p) x- p = _1[-;:-__ (25)
sin2(1[p) (-p) !
leading to the expansion

-x ~ xk
EI(X) =e L k.' [W(k) - tn x],
k=O
(26)
w(a+l) = dd a tn(a!).

Descending Expansion: The inversion integral gives

EI(x)
x
e- .
= ---2
1[1
f C + im
. F(p) x -p dp, O<C<1. (27)
C-1 m

The contour may not be closed to the right; however we can

shift it a finite distance, since F(p) goes to zero
§lZ. Mellin transforms 201

exponentially as p -> c ± ioo. Thus we can write

n k+l c+n+ioo
El (x) = e- x { L
k=l
(-1)
xk+l
k: + 1
TnT fc+n-l°O
. F(P)x-PdPJ. (Z8)

We leave i t to the reader to verify that the remainder term

is of order x- n - Z as x -> 00 , showing that we have recovered
the well-known asymptotic expansion

(Z9)

lZ.3. Elementary Properties

Mellin transforms have a number of important elemen-
tary properties, which we now investigate. We use the nota-
tion
oo 1
~[f;p] = J0 f(x) x P - dx (30)

where this simplifies the appearance of the results.

Derivatives:

L[f' ;p] f oo
f' (x) x P -
1
dx
o
~(X) xp-~: (p-l) I: f(x) xp - Z dx. (31)

We assume that F(p) exists for a < Im(p) < s; consequently

we must have

0, Re(p) > a,
(3 Z)
0, Re(p) < S,

and thus

1[f' ;p] -(p-l) F(p-l), a < Im(p-l) < S. (33)

202 PART III: OTHER IMPORTANT TRANSFORMS

Powers:
1
J
00
~[x~f(x);p] = x~f(x) x P - dx = F(p+~). (34)
o
Laplacian in Plane-Polar Coordinates: In two dimensions, the
2
Laplace operator V is

(35 )

If we take the Mellin transform F(p,~) of f with respect

to the radial variable, we obtain the simple relation

(36)

by the application of (33) and (34). Thus problems involving

this operator may be simplified by use of the Mellin trans-

form.

Convolutions: If

hex) = fOO y~f(xy) g(y) dy, (37)

o
then
H(p) fOO x p - l dx foo y~ f(xy) g(y) dy
o 0

fOOo y~g(y)dy fOO x P - 1 f(xy) dx

r
0

Jr y~-Pg (y)dy t p - l f(t) dt

oo
(38)
o 0

F(p) G(~-p+l).

Similarly, if

k(x) J: y~ f(x/y) g(y) dy, (39)

then
K(p) F (p) G(~+p+l). (40)
§lZ. Mellin transforms Z03

A further relation which is sometimes useful concerns

.L[fg ;p]; it is

.L[fg;p] fOO f(x) g(x) x p - l dx

o

fo g(x)
OO
xP -
1
dx
1
Zrri
fc-iioooo F(s)x -s ds
C+
(41)

1 fC+ 1.' 00
Zrri F(s) G(p-s) ds.
C-1 00

lZ.4. Potential Problems in Wedge-Shaped Regions

Consider the boundary-value problem

0, O<r<oo, -a<a~a, ( 4Z)

I, o < r < a
u(r ,±a) = ( (43)
0, r > a,
which determines the solution of a potential problem in an
infinite wedge of angle Za. If we assume that u(r,a) is
-8
bounded as r + 0, and that as r + 00, u (r, a) 'V r for some
8 > 0, then the Mellin transform of u with respect to r
exists. Before applying (36) to the partial differential
Z
equation, we multiply by r so as to obtain an equation for
U(p,a) rather than U(p-Z,a). Then (4Z) and (43) transform
to

["d- ZZ +
zl
p °1 U (p , a) 0,
Lda J
p (44)
U(p,a) = U(p,-a) alp,
wihh the solution

U(p,a) = aP cos pa (45)

p cos pa
204 PART III: OTHER IMPORTANT TRANSFORMS

We shall consider the inversion of Mellin transforms of this

type in the next section; anticipating these results we ob-

tain the solution of our potential problem as

r1 - - arctan
1 ~(ar)S
L cos sel,
J 0 < r < a

u(r ,e)
= JI ~ arc an
't ~(ar/~::r::J
2 ' r > a
(46 )

L 7T r 2S _a S

where S = 7T/20..
1
12.5. Transforms Involving Polar Coordinates
In problems involving polar coordinates, one is con-
fronted with transforms of the type F(p)sin(pe) and
F(p)cos(pe), as in (45). Suppose that F(p) is the Mellin
transform of a real function fer); then proceeding formally
we have 2

ie roo i e -1
~[f(re ); r + pl f(re ) rP dr
J 0

e- ipe J f(s) s
P -1
ds (47)

e- ipe F(p),

provided the s integral is equivalent to a Mellin transform.

This leads to the useful formulas

ie
F (p) cos (pe) ~[Re f(re ); r + pl,
(48)
ie
F(p) sin (pe) = ~[-Im f(re ); r + pl.

Sector of Validity: In order to carry out the variable

change s = r exp (ie) in (47) we must assume that fer) is
the value of an analytic function fez) defined in some
sector -a. < arg(z) < a., with r = Izl. Replacing the upper
§lZ. Mellin transforms ZOs

limit for r in (47) by R, we can write

t o
f(re i9 ) r P - l dr = e- iP9 {t 0
f(s) sp-l ds
( 49)

+ Ie f(s) sp-l dS},

Im(s)

c
Re(s)
o R
Figure lZ.Z

(see Figure Z). We need the second integral to become zero

as R + 00. A sufficient condition is

zP fez) + 0 as Izl + 00, larg(z) I < a. (SO)

We have already assumed that ~[f(r)] exists in some strip

PI < Re(p) < PZ' thus

r P fer) + 0 as r + 00, Re(p) < PZ. (51)

The usual situation is that (SO) is also valid when

Re(p) < Pz provided that a is suitably chosen; consequently
(SO) is a restrictior. on a rather than fez). For example,
if fer) = exp (-r), then fez) = exp (-z) and a = TI/Z.
Z06 PART III: OTHER IMPORTANT TRANSFORMS

Applicat.ions:

n
Ci) 1IR.n(1+r)] -1 < Re (p) < O. (5Z)
p sin pn
Hence

1 Z n cos pe
1IZ R.n(l+Zr cos e + r )] (53)
p sin p'!T
and
1Iarctan ( r sin e )] -n sin pe
l+r cos e p sin p'!T
- n/Z < e < '!T/Z,
-1 < Re(p) < O.

n
(ii) 1I; - arctan r]
Zp cos (p'!T/Z)
(54)
o < Re (p) < 1.

To check the sector of validity of (SO) in this case we write

arctan z = ~ + ~i [~n(z-i) - ~n(z+i)] (55)

where the branches of the logarithm are chosen so that

o< arctan r < n/Z. It follows immediately that a = n!Z;
using (55) we find that

'!T - arc t an [ zr cosZ eJ o< r < 1

l-r
Re I; - arctan z] ( 56)

arctan , r > 1

and this function has as its Mellin transform the function

cos(pe)/p cos (p'!T/Z). Equation (46) may be recovered by
applying the results of Problems 10 and 11 in succession.
§12. Mellin transforms 207

12.6. Hermite Functions

The Mellin transform may sometimes be of use in solving
ordinary differential equations with polynomial coefficients,
using a technique which was first employed by Barnes in his
investigations of the hypergeometric function. We choose
here to discuss the Hermite equation

H"(x) - xH' (x) + vH (x) =

v v v
o. (57)

On taking the Mellin transform, we get

(p-1) (p-2) S(p-2) + (v+p) S(p) 0, (58)

S(p) = .L[H ;p].

v

It is apparent that the Mellin transform does not give us

S(p) directly, but rather a difference equation for S(p).

Solution of the Difference Equation: We first reduce the dif-

ference equation for S(p) to standard form (in which the
arguments differ by an integer) by writing p = -2s and
S(p) = T(s). Thus we have to solve

T (s+l)
(s- jv)
'1' (s) (59)
2 (s+i) (s+l) •
A particular solution is

-s 1 1
T(s) = K 2 (s-r-1)! (-s-Z)! (-s-l)!, (60)

but it is not unique, since we may multiply T(s) by any

function which is a solution of the difference equation

YCs+1)
Y (s)
1. (61)

At this point, we appeal to the fact that S(p) is a Mellin

Z08 PART III: OTHER IMPORTANT TRANSFORMS

transform, defined only in some strip a < Re(p) < a. There-

fore, equation (58) is valid only in the overlap of the two
strips
a < Re(p) < a
(6Z)
a < Re(p-Z) < a
and there is no such overlap unless a > a+Z. Thus, yes)
cannot have poles, since they would give rise to a row of
poles in S(p) separated by exactly two units. Also, yes)
cannot grow faster than lsi as Im(s) + 00 in the inversion
strip; otherwise the inversion integral would diverge. There-
fore, by (61), Yes) is a bounded entire function, and thus
equal to a constant. Hence (60) is the only acceptable solu-
tion, and then only if Re(v) < -Z, so we have

HV (x) = K
Z'1Tl
r+ ioo (s-ZV-1)
c-ioo
1
! (-s-2)!(-s-1)!"2
1 [x')' ds, (63)

Re (v) < - 2.

By convention the coefficient of XV in Hv(x) is unity,

and since the pole of (S-~V-l)! at s = lv has residue
2
l, we have
Zv/Z
K (64)

1
Complete Descending Expansion: The poles of (s-ZV-l)! lie
1
at s = (zv-r) , r = 0,1, ... , with residues ( -1 ) r / r ! Thus

2v/ Z
H (x)
v

( l)r (--v+r--)!
1 1 1
00

x I ----
r! 2 2
(--v+r-l)!
2
(65)
r=O
xv- v(v-l) xV-Z + v(v-l)(v-Z)(v-3) x v - 4 + ••••
2 2! z2
§12. Mellin transforms 209

The restriction on v may now be lifted by analytic continua-

tion. If v is a positive integer or zero, we have a poly-
nomial of degree v.

Complete Ascending Expansion: Closing the contour to the

right leads to the expansion

2v/2
(-.!.v-!.)! (-!.2v-l)!
2 2

x
r=O
00

I
r_l)r(
7:'- t 11
(r-iv-i)! (- r+i)!
1
(66)

1
z
+ (r- V-l)!(-r+'2)!
1
( ]r}
x2
'2 .

If v is zero or a positive integer, we must first calculate

the ratios of the factorials outside and inside the summation
before using this formula, which then gives us a polynomial.
In other cases, the expansion is an infinite series.

Problems
Prove the following general properties of the Mellin trans-
form.
1. 1[f (ax) ;p] a- P F(p)

2. 1[f(x a );p] a -1 F(p/a)

-1 -1
3. 1[x f(x ) ;p] F(l-p)

d
4. 1[tn x f(x);p] dp F Cp)

S. n n n
1[(xd/dx) f(x);p] = (-1) p FCp)

6. 1n: f (U)dU;~ = - ; F (p+l)

210 PART III: OTHER IMPORTANT TRANSFORMS

7. I
.A<-
fJoox
L.: l
f(U)du;pJ = i; F(p+l)
Verify the following Mellin transforms.

-a (E-l)! (a-E-l) !
8. ...I[ (l+x) ;p] o< Re(p) < Re(a)
(a-I)!

-1 '11'
9. ..L[ (l+x) ;p]
sin 'I1'p

a -1 '11'
10. .L[ (l+x) ;p]
a sin('I1'P/a)

11. ..L[e
iSx
;p] (E- 1 2! ei'ITp/2 o< Re(p) < 1
sP

12. ..L[cos Sx;p] (E-l) ! cos ('I1'E/2) o < Re(p) < 1

sP

(P-l) ! sin ('I1'P/2)

13. ..L[sin Sx ;p] -1 < Re(p) < 1
sP

2P-l(iv+~-I)!
14. L[J (x) ;p] -v < Re (p) < v+2
v c}v-}P) !

[Hint: use the Poisson integral representation for

Jv(x).]

-ax 2 !. -p/2(1 -I)'

15. ..L[e ;p] 2 a F'

a-I (p -1) ! (a -1) !

16. ..L[ (I-x) h(l-x) ;p]
(p+a-l)!

17. L[e-xR,nx;p] (p-l)! W(p-l)

18.
r l+x cos e l '11' cos pe
1G+2X cos e+x2 ;~ sin P'l1'

19.
..Lr x sin e l '11' sin pe
Ll+2X cos e+x2;~ sin P'l1'
H2. Mellin transforms 211

R(p-l): (-p-2/3): (~P-l/3):

20.
4 (-2/3): (iP+7 /6):

21. ~ [J oOO
u
2/3
(u+x)
-1/3 _ 2
e u du;p
~
=
(p-1):(-p-2/3):(~P-l/3):
2(-2/3):

22. The complementary error function is

erfc(x) = ~ foo e- u2 du;

Iii' x
1 1
show ~[erfc (x) ;p] = (Zp-7) :/plii'.

23. The cosine integral is defined by

Ci(x) _fCO ~ duo

x u '

show ~[Ci(x) ;p] = -(p-l)!

p cos(~p/2) •

24. Find the steady state temperature distribution inside a

wedge 0 ~ r < 00, 0 ~ e ~ a, if the boundary e = 0 is
held at temperature zero, while the other boundary is
maintained at
r < a
u(r ,a)
r > a.

25. The boundary e = 0 of an infinite wedge o< r < 00 ,

o < e ~ a, is held at zero temperature. Through the
other boundary, the concentrated heat flow

q(r) = Qo(r-a)
is maintained. Show that the steady state temperature

distribution is
212 PART III: OTHER IMPORTANT TRANSFORMS

u(r,a) =
Q
2w K cosh[wtn(r/a)/2~]

x tn
rcosh[wtn(r/a)/2~] + sin(wa/2~)l
ILcosh[wtn(r/a)/2~] - sin(wa/2~)
J'
26. A thin charged wire with charge q per unit length is
placed along the line r = ro' a = aO' inside a wedge
shaped region 0 < r < 00, 0 ~ a ~ ~, whose boundaries
are held at zero potential. Show that the electrostatic
potential may be written as

i oo
sin p (~-a 0) sin pa

cj>(r ,a)
(
¥ f_ioo p sin p~
[:or dp, a < a
0

l~ J~:oo sin pa O sin p(~-a)

p sin p~
r dp,
[roJp a > a o.

27. In the preceding problem, show that if ~ = 2w, rO a,

and a O = w, then
1+21TTa sin(a/2)+(r/a)
Hr,a) = q tn
l-2!T7a sin(a/2)+(r/a)
Calculate the charge density induced on the boundary
a = 0, r > O.

28. If in Problem 24 the boundaries are at a ±~, with the

boundary conditions

u(r,-~) f (r) ,

u(r,~) g (r) ,

then show that

a {I ua-lf(u) du
S:" 0u 2a-2rau a sin aa
OO

u(r,a) = cos(aa) + r2a

where a w/2~.
§12. Mellin transforms 213

29. Solve the Laguerre equation

xy" (x) + (l;+l-x) y' (x) + ny(x) =0

by the Mellin transform. In particular, derive the

Laguerre polynomials corresponding to the choices

11 m-R.-1,
I; R. (R.+1),
where R. and m are non-negative integers.

30. Show that if the integral transforms

F(x) f: k(xt) f(t) dt

f(t) f~ R.(tx) F(x) dx

o
are reciprocal to each other, then

L(p) K(l-p) = 1.

31. By considering the ~e11in transform of Jv(kx) verify

the Hankel transform pair (15.1,2).

Footnotes

1. See W. J. Harrington, SIAM Review (1967), ~, 542.

2. We write ~[f(r,6); r ~ p] to indicate that r is the

variable being integrated out to give a function of p.
214 PART III: OTHER IMPORTANT TRANSFORMS

§13. MELLIN TRANSFORMS IN SUMMATION

13.1. Mellin Summation Formula l

Suppose we wish to evaluate the sum

S = L f (n). (1 )
n=l
If the function fen), regarded as a function of a continuous
variable, has the Mellin transform F(p), then we may write

fen) (2)

and consequently

S = 2711 i r+~oo F (p) L n -p dp

C-1 00 n=l
(3)
1 r+~oo F (p) <;;(p) dp,
271 i C-1°O

where dp) is the Riemann zeta function, whose more impor-

tant properties are discussed in Appendix B.

An Example: We consider the sum

cos Sn
S = L o < S < 271. (4)
n=l n2

From Problem 12.12, we have

fOO cos(Sn) n P- l dn = (p-l)! cos(7Ip/2),

o sP
o < Re(p) < 1, (5)

and thus the Mellin transform of cos(sn)/n 2 is

F(p) = _ (p-3)! cos (7Ip/2). (6)

Sp-2

The sum now becomes

§13. Mellin transforms in summation 215

s 1 rc +ioo Cp-3)! cos ('ITP/2) 1;(p) dp,

2'ITi J c-ioo 8P- 2
(7)
2 < Re (p) < 3,

where the interchange of order of integration and summation

is made possible because the sum converges uniformly in p
on the inversion contour.

Using the Riemann relation (Bll) the integral may be

cast in the more convenient form

s -~ 2
4 'IT
fc+ioo
c-i 1 oo
(2'IT)P
~
r;Cl-p)
(p-l) (p-2)
dp. (8)

We have three simple poles at p = 2, 1, and 0, from which

we obtain the simple result

s = _ ~ [(2'Tf)2 1;(-1) -
2 2 L8 ~'IT 1;(0) - iJ (9)
2 2
'IT _ 1LB. + .fL
"6 2 4

A Further Example: We consider the finite series

S (x) I
m=l
(1_xm2/3)1/2
m2/ 3
(10)

-3/2
M largest integer < x

For small x the series is slowly convergent; we will ob-

tain an asymptotic formula which is rapidly convergent.
First we write
2/3 2/3
00

S (x ) = x f(xm) , (11)

where
_ J(1_t 2/ 3 )1/2/ t 2/3, t < 1
f (t)
-l 0 , t > 1,
(12)

and, using the relation

216 PART III: OTHER IMPORTANT TRANSFORMS

!.a F(E. - 1)
a (13)
we obtain
3 [(3p/2) -2]! (1/2)!
F (p) (14 )
"2 [(3p/2)-(1/2)]!
and

1 IC + ioo3 [(3p/2)-2]!(1/2)! -p-2/3

(15)
hi c-ioo '2" [(3p/2)-(1/2)]! x l,;(p),

c > 2/3.

The factor [(3p/2)-2]! has poles at p = 2/3, 0, -2/3, ... ,

-(n-l)2/3, with residues (2/3)(-1)n/n! and 1;(P) has a
pole at p = 1 with residue 1. Hence

3Tf 1 1 2
Sex) " - + 1;(2/3) + iX - g1;(-2/3)x + ••• , (16 )
4IX
where the remainder term, which is a difficult integral, even-
tually increases with increasing n. Nevertheless, (16) is
a useful asymptotic series for small x.

2
13.2. A Problem of Ramanujan
Consider the function

~ -xe n -eiil
f(x) = n~-oo e J' (17)
00
- e

This infinite sum converges for all positive x, but it can-

not be summed by the Mellin summation formula. It obviously

satisfies the functional equation

f(ex) = fee) + f(x) (18 )

which is the same functional equation satisfied by the func-

tion in x. Is our function then identically equal to
in x (up to a constant factor)? We can show that f(x) f

in x as follows.
§13. Mellin transforms in summation 217

The relation

r0
e- y yp-l dy = (p-l)! , Re (p) > 0 (19)

allows us to write, by the Mellin inversion formula,

-y _ 1 fC+i~ -p
e - 2Tri . (p-l)! Y dp, c > O. (20)
C-l~

Hence, substituting this integral representation into (17),

we can write

r+~~
~

1
f(x) = L 2Tri (p-l) ! e- np (x-P-l)dp
n=-oo C-l~

r+~~
~

1
L
2Tri (p-l) ! e-np(x-p-l)dp (21)
n=O C-l~

_1_ r+~~
~

+ L 2Tri (p-l) ! emp(x-p-l)dp

m=l C-l~

The first sum 51 converges uniformly with respect to p, so

that we can change the order of integration and summation
and carry out the sum to write

1 fC+i~ (x-P-l) dp. (22)

51 = ---2. (p-l)!
Trl c-i~ l-e-P
This interchange is not permissible in the second sum
but we can overcome the problem by observing that the inte-
grand has no pole at p D, the pole due to (p-l)! being
cancelled by the zero of (x-P-l). Thus we can translate
the contour to the left to write

fc'-i~
~ c'+i~
5 = L _1_ (p-l)! emp(x-p-l)dp, -1 < c' < O. (23)
2 m=l 2Tri

On the new contour our sum converges, so bringing the sum-

mation inside we have
218 PART III: OTHER IMPORTANT TRANSFORMS

S2 -
1
21fi
r'
c'-ioo
-p
+ioo (p-l) ! (x -1) dp
(l-e- P)
(24)

and
f(x) S + Sz
1
(x-P-l)
Z;i Ie (p-l)!
(l-e- P )
dp (25)

where the closed contour of integration is indicated in

Figure 1. The poles of the integrand are a single pole at

poles of _'_=-
1-e-P

/
poles of(p-1)!

Figure 13.1

p = 0, and single poles due to zeros of the function l-e- P

at p = ±Z1fin along the imaginary axis. The residue of the
integrand at p = 0 is

-J/,n x (26 )

Addition of the residues at the poles gives us

I
00

f (x) -J/,n x + (hin-l)! (x - 21fin_ l ) (27)

n=-oo
§13. Mellin transforms in summation 219

Our function f(x) then does not coincide with the function
-~n x, except at isolated points given by

n = 1,2,3, ... (28)

The function in fact wobbles about the function -~n x. A

similar problem led Ramanujan to his fallacious proof of
the Prime Number Theorem--he forgot the wobbles!

13.3. Asymptotic Behavior of Power Series 3

A problem which sometimes occurs in that of finding
the asymptotic (large z) behavior of a function defined as a
power series in z. We have already seen that an integral
representation may be a useful starting point for asymptotic
analysis; the Mellin transform is the most direct method for
a power series because it involves ascending and descending
expansions simultaneously.
Consider the integral representation

fez) = 2!i r+~oo FCp) z-P dp, c > O. (29)

C-1°O
n
This will equal the power series E a z provided that the
n
contour may be closed to the left, and that F(p) has simple
poles at 0,-1,-2, ... , with residues at p = -no A

possible choice of FCp) which satisfies this latter condi-

tion is
F(p) = n(-l)P a cosec (np) (30)
-p

where 0 < c < 1, and we have assumed that the coefficients

are expressed by a suitable formula.

4
Stirling's Series: It is shown in many places that for
lal < 1 we can write
220 PART III: OTHER IMPORTANT TRANSFORMS

'"
Jl.n(a!) = -ya + }; (_l)n an r;;(n)/n. (31)
n=2
Using (30), we have

Jl.n(a!) = -ya + 1
2~i
fC +i '" ~--~~-
~as r;;(s)
ds, 1 < c < 2. (32)
c-i'" s sin(~s)

By the principle of analytic continuation we can replace the

restriction lal < 1 by larg(a) I < ~; subsequently moving
the contour to the left yields the asymptotic expansion

'" (-1) n r;;(-n)

}; (33)
n=l n an
which is known as Stirling's series.

Incomplete Factorial Function: As a second example, consider

the functions

(a,z)! = J: t a e- t dt, Re(a) > -1

'" (_l)n zn+a+1

}; (34 )
n=O n! (a+1+n)
and
[a,z] ! f"'z t a e -t dt

a! - (a,z)! (35)

For Re(z) -+- "', la, z)! obviously behaves as exp (- z) , so i t

is expedient to deal also with the function f (z)
exp (z) (a, z) ! which is defined by the power series

fez) =z
a+1 '"}; zk '"}; (- z)n
k!
k=O n=O n! (a+1+n)

za+1 '"}; zm
m=O
ry ( -1) JI.
LJI.=O (m-Jl.)!JI.! (a+1+JI.)J
l (36)

a+1 '"}; a! zm
z
m=O (a+1+m) !
§13. Mellin transforms in summation 221

Using (29) and (30) we tentatively write

z
(a,z)! = a! za+1 e- fC+~oo Cp-1)!(-p)! (-z)-Pdp,
2ni C-l°O (a+1-p)! (37)
O<c<1.

The question of the convergence of (37) can be settled by re-

placing p by c+i~ and using the asymptotic forms for
large ~ for the factorial functions; this gives for the ab-
solute value of the integrand the behavior
exp [-(n~/2) + ~ arg(-z)] and thus (37) will converge in the
sector
larg(-z) I < n/2. (38)

To obtain an integral representation in the sector

larg (z)1 < n/2, we must work with the power series for (a,z)!
directly, to yield

(a, z) ! za+1 fC+~oo (p-1)! z-p dp,

hi C-l°O a+1-p (39)
larg (z)1 < n/2.

Moving the path of integration in (37) and (39) to the right,

we obtain the asymptotic information
a -z
(a,z)! =a! +O(z e ), larg(z)1 <n/2,

a+1 -z 00
1
-a! z e L (40)
k=O (a-1)! zk+1 '
larg(-z) I < n/2.

A complete descending expansion of [a,z]! may be obtained

by a direct application of the Mellin transform; this leads
to the complete expansion
222 PART III: OTHER IMPORTANT TRANSFORMS

co
(k- a-I)!
(a.,z)! I (_ z) k
(-a-I) ! k=O
(41)
I arg (z) I < TT/2.

Further examples of these techniques may be found in the

problems.

Problems
co
sin an
1. Evaluate I
n=l
n

2. Show that
co
(_l)n+l n -s (1- 2l - S g (s),
I
n=l
co
-s -s
I
n=O
(2n+l) = (1-2 )1;(s).

co J 1 [C2n+l)y]
3. Evalute I
n=O 2n+l

4. Derive, by the method of Mellin transforms, the Jacobi

theta function transformation
2 2
m Ia.
co

L l. + (1fla.)1/2 ~' e
-1f

n=l 2 m=O

5
5. Show that
co
TT2 _ 1 x
[. R.n(l-e -nx )
~
6x ZR.n(x/2TT) + 24
n=l
00
2
+ I R.n(1-4TT nix).
n=l
§13. Mellin transforms in summation 223

6. Show that S

L ~n[1_e-(2n-l)x]
n=l
I
00

x
- 24 - n=l
7. By writing

(CL+~)-
1 fl u CL+~-l du,
o
show that

y (_1)R. (a.-I) !
~=O (m-R.)! R.! (CL+~) (m+CL) !

8. Let
00

f(x) L
n=O n! (2n+l)
and

x
g(x) e £(x).
Show that
I1r xn
00

g (x) =Z- I
n=O (n+j) !
Hence deduce the integral representations

r 1
2~i
fy + ioo (p-l)!x-Pdp
l-2p ,
larg(x)1 < ~/2

lZ!i I::;:
f (x)

(p-l)!C-p)!(_x)-PdP larg(-x)1 < ~/2

( -P+ 1) ., '
Z
and the asymptotic series

J~ x- 1/2 + o(x-le- x ), larg(x)1 < ~/2

l
£(x)
e- x -1/2)' 00 -1/2 (
- In +. + O(x ),larg(-x)1 < ~/2.
21Ti" n=O (-x)n 1
224 PART III: OTHER IMPORTANT TRANSFORMS

Footnotes

1. This section is based on G. G. MacFarlane, Phil. Mag.

(vii) (1949), 40, 188. MacFarlane considers the more gen-
eral problem of evaluating sums of the form
00

I f[ (n+a) v].
n=O
2. This treatment is due to B. w. NinLam.

3. DINGLE (1973), Ch. 2.

4. For example, OLVER (1974), p. 64.

5. This is a transformation in the theory of elliptic modu-

lar functions.
§14. Integrals involving a parameter 225

§14. INTEGRALS INVOLVING A PARAMETER

14.1. Preliminary Example l

Consider the function g(y) defined by
2
f""o e -yk
e 1r/k - 1
k dk (1)

r
The Mellin transform is simply
2

f:
2lrl/2 e- yk
G(p) yp-l d y y 3/2 k dk
0 e lr/k - 1

I: (p+ l)
2 ., k- - dk
2p 2
2lr l / 2 (2)
e lr/k - 1
1
21r - 2p-Z (p + l) , (2p)! r.;(2p+l).
2 .

On using the inversion integral, and closing the contour to

the left, we obtain the convergent expansion
1
1/2 (l)n 1 zn+l
2g(y) 1 - (lrY) + L
co

n=O
Ii:- (2 n)! r.;(n+2)y (3 )

which may be used to compute the value of g up to Y = 10

quite easily. In the right hand half plane, G(p) has no poles,
and the Mellin transform does not yield a descending expansion
in this case.

14.2. A General Class of Problems 2

The integral (1) is a convolution, of a type which has
a particularly simple Mellin transform. A very general class
of integrals whose asymptotics may be inve~tigated using the
Mellin transform is

h(A) = I: fey) g(AY) dy (4 )

It was shown in Eq. (12.38) that the Mellin transform of (4) is

226 PART III: OTHER IMPORTANT TRANSFORMS

H(p) = F(l-p) G(p) ( 5)

The range of p for which this is initially defined - which

fixes the placement of the inversion contour - must be deter-
mined in any particular case.
As an example, Watson's lemma is easily derived. In
this case, the integral of interest is

h(A) = f: fey) e- AY dy ( 6)

and the Mellin transform is

H(p) = (p-l)! F(l-p) (7)

There are two restrictions on p. (i) Re(p) > 0, in order

that the Mellin transform of the exponential should be defined,
and (ii) Re(l-p) must be in the range of convergence for the
Mellin transform of f(x).
Convergence of the integral defining F(l-p) depends
on the asymptotic behavior of f(x) for both small and large
values of s. In order to separate the two problems, it is con-
venient to define two new functions fl and f2 as

{ f (x) , x < 1
flex)
o , x > 1
(8)

{ 0 , x < 1
f 2 (x)
f(x) , x > 1

with a corresponding splitting of h(A) and H(p). We now

assume that fl and f2 have the asymptotic forms

- ~ a xAn x + 0
L n '
n (9)
-1 < Al < A2 < •••
§14. Integrals involving a parameter 227

(10)
III < 112 < •••

More general forms involving logarithms are readily included

at the expense of more algebra. Note that the restriction on
Al is necessary in order to make the original integral con-
verge. With these asymptotic forms, it is readily shown (an
analogous case is analyzed in'§14.3) that the singularities of
Fl(l-p) are simple poles at p = An + 1 with residues -an.
The singularities of F 2 (1-p) are simple poles at p = 1 - lln·
Descending expansions of h(A) are obtained by moving the in-
version controus for hl(A) and h2 (A) to the right. For
Hl(p), the contour must originally be chosen so that 0 <

Re(p) < 1 + An and for H2 (p) so that Re(p) > 0 and

Re(p) > 1 - Ill. When we move the contours, we pick up a resi-
due at each pole of Fl(l-p); there are no contributions from
F 2 (1-p). In this way Watson's lemma is recovered.

14.3. Ascending Expansions for Fourier Integrals 3

Consider the integral

f(x) = f: A(k)e ikx dk (11)

which is a typical Fourier integral representation for the func-

tion f(x). To obtain an expansion for small x, we might try
to replace the exponential term by its Taylor series and inte-
grate term by term, which gives

f(x) r
n=O
(ix)n
n! I: A(k) kn dk. ( 12)

This will be a useful asymptotic series if at least the first

228 PART III: OTHER IMPORTANT TRANSFORMS

few terms converge, but in most cases it is not particularly

valuable.
Suppose now that A(k) has the asymptotic form

o< v < 1. (13)

The analysis of the cases v = 0 and v = 1 requires some

modifications to the following argument, and is left to the
problems. In the present case, we note that if A(k) satis-
fies some rather general conditions, then f(x) + 0 as
x + ~ (see Section 2.1). Furthermore, it is easily shown
that the asymptotic form (13) implies that f(x) = O(x v - l )
as x + O. For v < Re(p) < 1, we consider the Mellin trans-

form of f(x), viz.

F(p) = J ~o xp-l dx I~0 A(k) e ikx dk

(p-l)! e iwp / 2 I: A(k) k- P dk, (14)

1 - v < Re(p) < 1.

In order to obtain an ascending expansion for f(x), we need

to know the analytic structure of F(p) for Re(p) < I-v. To
this end, we define the functions

(15)

Now for Re(p) > I-v we can write

J: A(k)k- P dk

I~ A(k)k-Pdk + J: Rn(k)k- P dk + t~O atJ: k- P - v - t dk

n at
r
nn(p) + t=O p+v+ t - 1 ' (16)
§14. Integrals involving a parameter 229

where nn(p) is the sum of the first two integrals, which is

an analytic function for -v-n < Re(p) < 1. Thus (16) is an
analytic continuation into this larger region. Moreover we
note that if we restrict p to the strip -v-n < Re(p) <
I-v-n, then
n at
I
t=O p+v+t-l '
(17)

so that one analytic continuation of F(p) into this strip

is obtained by replacing A(k) by
Rn (k) in (14) .
Returning to F(p), we see that there are poles at
p = -n, n = 0,1,2, ... , due to the factor (p-l) ! . The resi-
dues in the Mellin inversion are

~
n.
JW0 R (k)k n dk.
n
( 18)

There are also poles from the integral (16) at p I-v-n,

n = 0,1,2, ... , with corresponding residues

(-v-n)! e iw (1-v-n)/2 an x v +n - l

w v-I iw(1-v)/2 an(ix)n

sinewv) x e (v+n-l)! (19)

Consequently, moving the contour of the inversion integral to

the left gives the asymptotic series

J: A(k) e ikx dk

J:
W
(ix)m
I
m=O iii!
m
Rm(k) k dk (20)

w v-I iw(1-v) /2
W am(ix)m
+
Sln wv x e I
m=O (m+v-l)!

It is instructive to compare this with (12) . In particular,

i f the first N coefficients an are zero, then (20) and
(12) coincide to the first N terms.
230 PART III: OTHER IMPORTANT TRANSFORMS

14.4. Multidimensional Integrals 4

Many research problems degenerate at some point or an-
other to the evaluation of rather complicated multidimensional
integrals containing a parameter. Normally one is interested
in only the leading terms in an asymptotic expansion of such
an integral; we give below some typical examples.

Example 1. Partition Function of an Electron Gas: An impor-

tant correction to the equation of state of a classical elec-
tron gas is given by the integral

f(A) = f: {e-Sq(x) - 1 + Sq(x) - ~Isq(x) 12}x 2 dx, ( 21)

where
q(x) 2 e -Kx
e: x
(22)

We require an asymptotic expansion for small values of the

parameter A = KSi 2 . Consider the integral representation of
the exponential function (12.8); we can translate the contour
there to the left by writing

e-
Y
-1 + Y -
Y2
2 _ _1_ fC+ioo (p-l)! y-Pdp, -3 < c < -2. (23)
hi .
C-1 OO

Now we replace y by Sq(x)

giving

f(A) = Jooo x 2 dx 2;1 JC+~oo (p-l)! (Se: 2 e-:X)-P dp. (24)

C-1°O

A change of variable to t = X/Si 2 and the substitution

KSe: 2 = A give

(25)
-3 < c < -2.
§14. Integrals involving a parameter 231

For the contour above the orders of integration may be inter-

changed, and the t integration performed at once. Thus

( 26)

so that

2;1 JC+1.· 00

C-1 OO
(p-l)! (p+2)! A- 3 - p (_p)-3- Pdp. (27)

The integrand has double poles at p = -3,-4,-5, ... ; evalua-

ting the residues at these poles, we obtain the expansion
n-3
[tn
00

n~3 ~~C~-3)! A + tn n - (n~3) - 1/J(n)

- 1/J(n-3)]. (28)

The first term of this expansion

Mtn 3 + 2y - (11/6) + tn A], (29)

where y is Euler's constant, was first obtained by Abel by

a very much more cumbersome calculation. Note that the
Mellin transform method yields the complete expansion of the
integral f(A) almost trivially.

Example 2. Correlation Energy of a Degenerate Electron Gas:

Gellmann and Brueckner first gave the leading corrections to
the correlation energy of a degenerate electron gas in the
form

J J
OOo du J~ dq ~l (-~) n [R(u)]n (~)n-2, (30)
q n=2 q
where
Ru = 1 - u arctan (l/u) (31)

and x is a parameter. We require the leading terms for

both large and small x. In the original paper the integral
Z3Z PART III: OTHER IMPORTANT TRANSFORMS

J was evaluated numerically. Writing y = (x/qZ)R(u), we see

that the sum over n is just the formal expansion of

- (~)-Z[tn (l+y) - y). ( 3Z)

q

We can give this function an integral representation by ob-

serving that

p
Ltn(1+y) 1
00
- -1 Joo ~1
p dy
POP 0 lJ.+YJ

-1 Joo (z-l)Pdz - -1 II (l-n)Pn -P-1 dn

P 1 P 0
lr , -1 < Re (p) < 0, ( 33)
p sin '7fP
so we have

1
tn(1+y) = Zlri r+~oo lr yP dp, o< Re(p) < 1. (34)
C-;lOO P SIn lrp
Hence, translating the contour to the right, the sum which
occurs in (30) may be written as
n
L
00

n=Z
(-~) [R(u))n (xq z)n-Z
1 IC+ I.' lr p-z
[R(U))P(~)
00

IiI C-IOO P sin lrp dp, ( 35)

1 < c < Z.

After substitution of this expression into (30), the q in-

tegration can be carried out immediately. The q integral is

I~ ~ (:z)P-Z = J: q 3-Zp dq = 4-Zp

1
, (36)

and our original integral J reduces to

1 ioo
J = I i I r+
'If
~ x P - Z dp,
c-ioo Zp sin 'lfP p-
I < c < Z, ( 37)
§14. Integrals involving a parameter 233

I:
where
~(p) = [R(u)]P duo (38)

A new feature here is the appearance in the integrand of a

function ~(p) which cannot be evaluated explicitly in terms
of elementary functions. Ho~ever, in order to find an ex-
pansion for J it is sufficient to know only the poles and
residues of ~(p). We can find these as follows.
Consider the integral (38). As u + 0, R(u) - 1 -
(uw/2) + 1, so the integrand is well behaved at the lower limit
for all values of p. On the other hand, for large values of
u, we have

arctan (l/u) (l/u) - (l/3u 3 ) + ••• , (39)

hence
( 40)

Now choose L sufficiently large so that (40) is a good ap-

proximation to R(u) for u > L, and split the integral (38)
as follows:

~(p) = J: [R(u)]P du + J: [R(u)]P duo (41)

The first integral is well-behaved (has no poles) as a func-

tion of p for all p. The second has a pole, since when
we use (40) we have

f
OOL [R(u)]P du = fooL' u-3p2P du + (an integral convergent at p

+ (remainder).

Thus as P + 21 from above, ~(p) has a single pole at

I
P =2 with residue 1/213. The leading terms in the des-
234 PART III: OTHER IMPORTANT TRANSFORMS

cending expansion are thus

J (42)

The ascending expansion is obtained from (37) by

translating the contour to the right. The result is

1 1
J = - 2 [~(2) tnx - 2~(2) + ~'(2)] + O(x tn x). (43)

Problems
1. Obtain expansions corresponding to (7) for s not equal
to an integer.

Obtain ascending the descending expansions for the following

integrals.

2. erfc(x) [See Problem (12.22).]

3. Ci(x) [See Problem (12.23).]

u 2/ 3 Cl_u 2)1/2
4.
f~ (u+x)1/3
du [See Problem (12.20).]

I:
2/3 _u 2
S. u e du [See Problem (12.21).J
(u+x)1/3

6.
I: u 2 tn(l-e -/u 4+a) du

1
7. If u(y) = fo t 3 (1_t 4)1/2 e- ty sin(ty) dt,

show that for large y

8. By taking the Mellin transform with respect to x in

the Poisson integral representation of Jv(x), deduce
the asymptotic expansion
§14. Integrals involving a parameter Z35

- l'Zl1Tx"J
l / Z{cOS[X-(V1T/Z)-(1T/4)]
00 r (4v Z-1) (4v Z-3 Z) ... (4v Z- (4r-l) ZJ
)
L (-1)
r=l CZr)! Z6r x Zr

+sin[x-(v1T/Z)-(1T/4)] [ L
00

(_l)r
Z Z Z Z
(4v -1)(4v -3 ) ... (4v -(4r-3) )JJ .
Z""'

r=l (zr-l)! Z6r-3 x Zr - l

9. For integer s, show that

n r
s(s-r+l) _ x~{~nlxl-~(s)+~(O)}
r~s x r! s.
r=O
n+l
+ 0 (x ).

10. If
Z
f(a) = J1 dx Joo dy l-exp[-ay X(l-X)],
o 0 l+yZ
show directly that for small a

1T 3/ Z 1T
f' (a) '"
l6al/Z - 12 '
3/Z l/Z 1Ta
rCa) '" 1T 8 a -TI

Use the Mellin transform to find complete ascending and

descending expansions.

11. If (13) is replaced by

n -~ 1
A(k) L a~k + O(k- n - ),
~=1

then show that the ascending expansion for f(x) is

f(x) L (iX~n {foo R (k)k n dk

n n. 0 n

+ an+l[-~nlxl+tny +(i1Tsgn(x)/Z) + ~ llJ

r~l ~ ,
236 PART III: OTHER IMPORTANT TRANSFORMS

where
n
Rn(k) = A(k) - L a r k-r_a k- n - l h(k-l),
r=l n+l

and h is the Heaviside step function.

Footnotes

1. B. Davies & R. G. Storer, Phys. Rev. (1968), 171, 150.

2. A comprehensive analysis of the use of Mellin transforms
to investigate integrals of the form (4) may be found in
Bleistein and Handelsman (1975).
3. These results were obtained by H. C. Levey and J. J.
Mahony, Q. Appl. Math. (1967), ~, 101, by a direct analy-
sis. It is interesting to compare the two methods of
derivation.
4. Based on material written by B. W. Ninham.
§ls. Hankel transforms 237

§ls. HANKEL TRANSFORMS

15.1. The Hankel Transform Pair

Bessel functions have frequently occurred in our in-

vestigations of the Laplace and Fourier transforms; indeed,

we could rewrite most of the formulas we have derived in

terms of Bessel functions of order ±~, since
liZ
(Zx/n) Kl/z(x) = exp(-x), with similar relations for
sin(x) and cos(x). Furthermore, we noted in Problem 12.31
that the integral transform

(1)

has for its inverse the reciprocal formula

These formulas constitue the Hankel transform pair. Proof

of the validity of these results for various classes of

functions f(x) (such as functions satisfying Dirichlet

conditions) may be found in various places;l we will be con-

tent here to reproduce a rather elegant treatment due to

MacRobert which is sufficient to cover many situations oc-
curing in practice.

MacRobert's Proof: We consider the integral

1o J v (kt)
00
k dk fb J (kx) f(x) x dx,
a v
0 < a < b, (3)

where we assume that f(x) is analytic in some region of

the complex plane containing the line a < x < b. Now we
split up one of the Bessel functions into Hankel functions,
and deform the x contour onto the contours Cl and C2
238 PART III: OTHER IMPORTANT TRANSFORMS

shown in Figure 1. Thus (3) becomes

Im(x)

c,
d b Re(x)

Figure 15.1

'21 r0 Jv(kt) k dk J H(l) (kx) f(x) x dx

C v
1 (4)

+ t fooo J v (kt)
k dk J H(2) (kx) f(x) x dx.
C v
2
We may reverse the order of integration because the Hankel
functions falloff exponentially on the contours Cl and
C2 as k + 00; using Lommel's integra1 2 this yields

1 r
(t/x)v-l f(x) dx (5)
= hi Jc {x\ - x!t}

_{£Ct), a < t < b

o , x < a or x > b

where C is the loop

C2-C l . From this the transform pair
(1) and (2) follow, provided that either f(x) or F (k)
v
. . 3
has the necessary ana 1 yt1C propert1es. This is usually the
§lS. Hankel transforms 239

case for practical applications, although (1) and (2) are

valid for a wider class of functions. 4

Connection with the Fourier Transform: Consider the two-

5
variable Fourier transform pair

2
f (r) e ik'r
1
F (k) 27f J - d r,
(6)
-ik·r
f
1
fer) F(~) e - - d2~.
27f

Suppose that we introduce polar coordinates

r ->- (r,e),
(7)
is->- (k,¢),

and expand both f(r,e) and F(k,¢) in Fourier series:

00
ine
f(r,e) L
n=-oo
fn(r) e
(8)
in¢ ,
F (k, ¢) L
n=-oo
Fn (k) e
where
fne r ) 1
171 f: 7f fer,e) e- ine de ,
(9)
1
Fn (k) 27f o F (k, ¢) e- in ¢ d¢.
t7f

On substituting (6a) into (9b) and using (8a) to represent

f(r,e), we obtain

Fn(k) = (2~)2 I:
7f
e- in ¢ d¢I: r dr f: 7f
de e
ikr cos (e- ¢)

ime
x e
m=-oo
(10)
1-
27f fo r
OO
dr

dr J n (kr) f n (r).
240 PART III: OTHER IMPORTANT TRANSFORMS

Similarly we may derive the relation

f (r) = foo k dk J (kr) F (kr). (11)

nOn n

Equations (10) and (11) are the Hankel transform pair with
v = n.

15.2. Elementary Properties

Because of their increased generality over the Laplace
and Fourier transforms, Hankel transforms do not have as many
elementary properties as do the former. We will recount here
those elementary properties which correspond to Sections 1.2,
1. 3, and 7. 2.

Derivatives: Suppose that Fv(k) is the Hankel transform of

order v of the function f(x); then the Hankel transform of
the function g(x) = fl (x) is

Gv (k)
f: f t ex) J ekx) x dx
v
00

[xf(x) Jv(kx)]o (12)

-f: f(x) £X [x J v (kx)] dx

We assume that the behavior of f(x) at 0 and 00 makes

the bracket zero, and use (20.46) and (20.47) to write

Hence we have

(14 )

Formulas for transforms of higher derivatives may be ob-

tained by repeated application of this result.
§lS. Hankel transforms 241

Bessel's E9.uation: Let f ex) be an arbitrary function, and

consider the transform of the combination

d2 1 d v 2
g ex) = f (x) +
x dx
fex) - 2" f(x). (15)
dx 2 x
Integrating by parts, assuming at each stage that the con-
tributions from x = 0 and x = 00 are zero, we have

fo Ldx
OOI~ x df _ v
dx x
2 f (x)l J" (kx) dx
j v

(16)
- k 2 roo f(x) Jvekx ) x dx
J0

- k 2 F v (k) .

Thus Hankel transforms may lead to significant simplifica-

tion in problems involving Bessel's equation.

Parseval's Theorem: There is no simple addition formula for

Bessel functions such as exist for the exponential and tri-
gonometric functions, thus the Hankel transform ~es not
satisfy any simple convolution relation. However, a simple
relation of Parseval type can be derived as follows. Let
Fv(k) and Gv(k) be Hankel transforms of order v; then

fo Fv (k) Gv (k) k dk
OO

fo Fv (k) k dk g(x) J v (kx) x dx

oo foo
0

f: g(x) x dx J: Fv(k) Jv(kx) k dk (17)

Joo f(x) g(x) x dx.

o
 242 PART III: OTHER IMPORTANT TRANSFORMS

The similarity with (7.32) and (7.34) is obvious.

15.3. Some Examples

Let

v 2 2]1
f(x) = x Ca -x) hCa-x), ]1 > -1; (18)

then, on expanding Jv(kx) by (20.45), we have

F (k)
v
(19)
00
C_l)m(k/2)v+2m fao 2v+2m+l( 2 2)]1
l
m=O (v+m) !m!
x a -x dx.

The latter integral is a Beta function (Appendix A); expres-

sing it in terms of factorials we get

(-l)m]1! (k/2) v+ 2m a 2]1+2v+2m+2

FvCk) = l
m=O 2(]1+v+m+l)!m!
(20)

Using the reciprocal Hankel transform, and replacing ]1+v+l

by ]1, we obtain another useful integral, namely

Jo x l-]1+V
oo
J (ax) J (bx) dx
]1 v

where the restriction ]1 > v is needed to make the integral

converge. A further result is obtained by setting ]1 = 0
in (20) and using the Parseval relation (17), so that
roo Jv+l(ax) Jv+l(bx)

r
J dx
o x
(ab)-v-l k2v+l h(k-a) h(k-b) dk
0 (22)

r 2C~+1) (~}V+l, a < b

= <
l2 C ~+l) (%] v+l , a > b.
§lS. Hankel transforms 243

Many other similar and related results may be obtained; some

of them are stated in the problems.

lS.4. Boundary-value Problems

The Hankel transform can be used to solve numerous
boundary-value problems in a relatively straightforward way,
using various properties of Bessel functions. We solve two
illustrative problems here; others are found in the problem
section.

Heat Conduction: Suppose that heat enters a semi-infinite

body of thermal conductivity K through a disc of radius a,
at a constant rate Q. The remainder of the surface at
z = 0 is insulated. We will find the steady-state tempera-
ture distribution of the body, u, which satisfies Laplace's
equation with appropriate boundary conditions. Using cylin-

drical polar coordinates, we can write for u(r,z) the

equations

(23)
J Q/a 2 , r < a
l
-KU =
z 0 , r > a.

Taking the Hankel transform of order 0 with respect to r,

the equations become

2
Uzz(k,z) - k U(k,z) 0,

(24)

The solution of these equations which remains finite as

z -+ 00 is
244 PART III: OTHER IMPORTANT TRANSFORMS

J 1 (ka) -kz
U(k,z) = Q... e (25)
Ka k

leading to a temperature distribution given by the integral

representation

u(r, z) (26 )

An Electrostatic Problem: We will find the electrostatic

potential generated in the space between two grounded plates

at z = ±a by a point charge q at r = 0, z = O. The

potential ~ satisfies Laplace's equation except at the
origin where it has the singular behavior ~(r,z) ~ q/~.
Writing ~(r,z) = q/~+ ~(r,z), we are faced with the

equations

~rr(r.z) + !.r ~ r (r,z) + ~ zz (r,z) 0,

(27)
~(r,±a) + q = O.
1r2+a 2
The Hankel transform of order zero turns these into the
simpler equations

0,
(28)
~(k,±a) = -qe -ka Ik,

and the solutions follow immediately, viz.

-ka
~(k ) = _ cosh(kz e (29)
,z q cos a k

q Joo cosh(kz) -ka (30)

Hr,z) Ir2+z2 - q 0 cosh(ka) e JO(kr) dk.
915. Hankel transforms 245

15.5. Weber's Integral

For some applications, a generalization of the Hankel
transform using Weber's integral may be useful. We sketch a
few salient points here, relegating most of the details to
the problems. We commence by considering the cylinder func-
tions

Z Ckr)
v
= J v (kr) Y (ka) - Y (kr) J (ka),
v v v (31)

6
chosen so that Zv(ka) = O. It can be shown that if Zv(kr)
rather than Jv(kr) is used in (1), we obtain the transform
pair

FV(k) [00 f(x) Zv(kx) x dx, (32)

a
Zv(kx)
f(x) = foo Fv(k) k dk. (33)
0 J~(ka) + Y~(ka)

A Simple Application: We consider an infinite slab of uni-

form solid material of thickness 2~, through which there is
a circular hole of radius a. If the plane faces are held at
temperature zero, while the circular surface is heated to
the temperature TO' then the steady-state temperature
u(r,z) will satisfy the equations

u(r,±~) 0, (34)

u(a,z) = TO.

Taking the transform (32) of (34) with v = 0, and using the

result of Problem 18, we obtain the ordinary differential
equation
246 PART III: OTHER IMPORTANT TRANSFORMS

2 2
Uzz(k,z) - klJ(k,z) =;:rTO' (35)

The solution, chosen to satisfy the boundary conditions

U (k , H) = 0, is

U(k, z) _ 2TO f1 _ cosh k~ (36)

rrk2 ,_ cosh k~ .

Use of the inversion integral (33) now yields an integral

representation of the solution.

Connection with Fourier Series: This problem may also be

solved by expanding in a Fourier cosine series in the z
variable, and we are led to seek the connection between the
two solutions. Expressing the Bessel functions in (33) as
Hankel functions, we can write u(r,z) as
TO 00 fHa1) (kr) H~2) (kr)]
u(r,z) - rri fo LH£l)(ka) - Hb 2 ) (ka)]
(37)
x [1 cosh(kz)l dk
- cosh(hl] k
The functions Ha 1 ,2)(Z) have no zeros for Re(z) > 0, hence
we may deform the contour of integration as follows:

(i) The term multiplied by Ha1)(kr)/H~1) (ka) is integrated

along the contour C1 of Figure 2, which is chosen to coin-
cide with the imaginary axis except for indentations around

the poles of U(k,z) at ~k = irr(n+t), n = 0,1,2, ...

After this change of contour, we introduce the new variable
s = -ik, so that the contour in s is C2 .

(ii) The term mUltiplied by Ha 2 ) (kr)/Ha 2 ) (ka) is inte-

grated along C3 , and subsequently we write s = ik to
bring the s contour to C •
4
§15. Hankel transforms 247

5;"1'12 .t

3i7l'/21

~
brlzl.
C.

C2
- i11'''2.e

Cll
-3i7l'/2.{'

- 5i11'IU

Figure 15.2

It is evident that the contours C2 and C4 coin-

1
cide, except for the indentations at ~
n = 'TT (n+ Z) /Q" where
they pass on either side. Hence (37) becomes

u(r,z) TO f KO(~r) I _ COS(~Z)I

'TTi C KO(~a) 1 cos(~t)
2

+ TO
'TTi
JC KO(~r)
KO(~a)
11 _COS(~Z)I
cos(~t)
it
~
(38)
4

= - 2T 0 I
n=O

which is the Fourier series solution.

Z48 PART III: OTHER IMPORTANT TRANSFORMS

Problems

Prove the following general properties of the Hankel trans-

form of order v. (Here we denote the integral (1) by

Z.

v-I d I-v
3. Hv[x ax{x f(x)};k] = -kFv_l(k)

4.

5.

P 1 1
Z CZV+ F _ 1:.)'
Z •
8.

9. Sonine's First Integral: Show that

xv+l JTf/Z ) . )1+1 Zv+l

J (x) = -- J (x sin e SIn e cos e de.
)1+V+1 ZVv! 0 )1

[Use (Zl).]

10. Show that

§lS. Hankel transforms 249

11. Sonine's Second Integral: Use the result of Problem

20.24 to show that

12. Show that

v-n-l n-v -n
= 2 (v-n-l)! k a In(ka).

(Set u2 = t 2 + a 2 in Sonine's second integral and

let x + 0.)

13. Let ~ be a solution of Laplace's equation in the half-

space z > O. The boundary z = 0 is held at the po-
tential ~ = fer). Show that the potential elsewhere is

given by the expression

<j>(r,z) = r
0
JO(kr)e
-kz
k dk
00

fo JO(ks)f(s) 5 ds.

Examine the special case fer) = v h(a-r) and show that

Hr,z) = Va r0 e -kz J l (ka)Jo(kr) dk.

14. The initial temperature distribution of an infinite

uniform region is u(r,O) = f(lrl). Show that

u(r,t) = ~
2t
f0 e-K(r +s
oo 2 2 /
) 4t I O(Krs/2t) f(s) s ds.

(Use Problem 10.)

250 PART III: OTHER IMPORTANT TRANSFORMS

15. The vibration of a thin elastic plate is described by

the equation

where c is the ratio of the rigidity of the plate

(against bending) and its mass per unit area. Show that
the motion of an infinite plate, starting from the
axially symmetric initial conditions

fer),
0,

subsequently is described by the expression

w(r,t)

Derive the alternative formula

w(r,t) = 2~t f: J o Crs/2ct)sin[Cr 2+s 2 )/4ct] f(s) s ds.

[Extend Problem 10 to verify the relation

Io JOCkr)
OO
JOCks) cos Cctk 2 ) kdk

16. Two point charges +q and -q are placed in vacuum on

either side of a slab of material of dielectric con-
stant E. The geometry is shown in Figure 3. Find an
expression for the electrostatic potential in each of
the three regions.
§15. Hankel transforms 251

-q +q

x=-1, X=-a X=a

x

Figure 15.3

(Hint: in each vacuum region, separate off the singu-

lar part of the solution.)

17. A direct current J enters a semi-infinite region

z > a of conductivity a through an electrode of

radius a making contact at the surface z = a. Show

that the current density j is given by

j -l7 u

where
fco
u(r,z) =
J
7Tao a e -kz dk
J l (ka) Ja(kr) k

Show that as a ->- a

J
u(r,z) '"
27TO/r +z
1z---2'

18. If
1
f(x) g"(x) + - g' (x)
x
252 PART III: OTHER IMPORTANT TRANSFORMS

show that the Weber transforms (32) of f and g are

related by

2 2
F (k) -k G (k) - - g(a).
v v 'IT

19. By considering the integral

f OO
Zv (kx) x dx
ICb Zv(px)
F (p) pdp
o J 2 (pa) + y2(pa) v
v v
and using Lommel's integral,2 construct a proof of the
inversion theorem for Weber's transform.

20. Construct an integral transform using the functions

and show how the transforms of functions f (x) and

g (x) , where
2
1
f (x) = gil (x) + -
x g'(x)
- V
g (x)
x2
and g(x) is arbitrary, are related.

21. A cylindrical hole of radius a is drilled in an in-

finite body, and the walls of the hole are maintained
at temperature Tl from time t = O. If the body is
initially at temperature TO' show that the temperature
distribution is given by
2

J:
2 (T 1 - T0) Zo(kr)(l-e-Kk It) dk
u (r , t) = TO
J~(ka) + Y6(ka) k
+
'IT

where K is the thermal conductivity and Zv is de-

fined in (31).
§15. Hankel transforms 253

Footnotes

1. For example, SNEDDON (1972).

2. Lomme1's integral is

for any pair of cylinder functions Uv and Vv [WATSON

(1958), p. 134] • It may be used to obtain results such
as
= _ ;i (A/]..l) v ,
eA 2_]..l2)
J: J v (Ax)H(l)
v
(AX)X dx Im(]..l) > O.

3. If Fv(k) is analytic in a region of the complex plane

containing a < k ~ b, then we replace (3) by
b
J: Jv(kx) x dx Ja Jv(px) Fv(p) p dp.

4. In particular the case b -+- co is easy to handle. Also,

i f the interval o< x < co can be split up into a
finite number of subintervals in each of which the con-
dition of MacRobert's proof applies, then the proof is
easily generalized. This covers most functions which
arise in applications.

5. We have chosen the constants 2~ in a more symmetrical

way than in (11.1) and (11.2).

6. Another transform is obtained from the choice Z~(ka) O.

254 PART III: OTHER IMPORTANT TRANSFORMS

§l6. DUAL INTEGRAL EQUATIONS l

16.1. The Electrified Disc

To motivate this section, we first solve a classical
problem of electrostatics. We wish to find the electrostatic
potential ~ created by an isolated thin conducting disc of
radius a, whose potential is V. Noting the symmetry of the
problem about the axis of the disc and introducing cylindri-
cal polar coordinates r, 6, and z, we reduce the problem
to that of satisfying the equations

~rr + r1 ~r + ~zz = 0 (1)

and
Hr,O) = V, r < a,
(2)
~z(r,O+) = ~z(r,O-), r > a.

Applying the Hankel transform of order zero, we easily find

from (1) that
~(k,z) = A(k)e -kl z I , (3)

and that the boundary conditions (2) reduce to the "dual

integral equations"

I: ACk) JoCkr)kdk V, r < a,

I:
(4)

kACk) JOCkr) kdk = 0, r > a.

If we differentiate C4a) with respect to r, we obtain an

alternative pair of equations, namely

~ (r,O)
r
(5)
0, r < a,
§16. Dual integral equations 255

(6)
0, r > a.

From (15.21) we see that the function

A(k) C(ka) - 3/2 J l / 2 (ka) (7)

satisfies both of these equations; furthermore, with this

form for A(k), (15.21) gives

<P (r 0) = - ~/f her-a) (8)

r' a 1f /
r1r2- a 2
and thus

<p(r,O) -f: <pt(t,O)dt

J ~2;1f 'in-'(a/r), r > a (9)

l ~2I!f, r < a.
2
Finally, this implies C = Va /ZTIT, so the solution is

'" . (k ) - k Iz I
<p(r,z) = ~V J0 Sln ~ e JO(kr)dk . (10)

16.2. Dual Integral Equations of Titchmarsh Type

Equations of the type

f'o" k-2a A(k) J (kx) kdk

]l
f (x), x < a,

(11)

J'"o k-2B A(k) J v (kx) kdk = g(x), x > a,

where f(x) and g(x) are only known over part of the
range 0 < x < '" and A(k) is sought, occur in certain
mixed boundary value problems of which the electrified disc
256 PART III: OTHER IMPORTANT TRANSFORMS

is a simple example. A convenient formalism for the solution

of these equations can be developed using a modified Hankel

transform defined by

oo -a
S f = (2/k)af x f(x)J 2\)+a (kx)x dx. (12)
\) , a 0

It is readily verified that the inversion formula for this

transform is given by

S-l = S (13)
\),a \)+a,-a

The dual integral equations (11) can be written

Sl A(k) = (2/x)2 a f (x), x < a,

7).l-a,2a
(14 )

Sl A(k) = (2/x)2 Bg (x), x > a.

'2\)-B,2B
Suppose now that we can find two operators Ll and L2 with
the following properties:

(i) S Y, 8 '
(15 )

LZ Sl = Sy 8;
'2\)-B,2B '
(ii) Llf(x) only involves values of f(x) for x ~ a;

(iii) L2g(x) only involves values of g(x) for x > a.

Then (14) will become

rLl (2/x) 2a f(x),

1
x < a
S A(k) (16 )
y,8 = L2 (2/x) 2Bg (x) , x > a

and A(k) can be found by applying the inverse operator

S to the right-hand side, which is a known function.

y+8,-o
§16. Dual integral equations 257

Choice of Operators: Using the inversion operators on (14),

we find that Ll and L2 must satisfy

Ll S Sl ,
Y,8 -lJ+a. -2a.
2 ' (17)
LZ S Sl
y,o ZV+I3,-ZI3
We will deal here with Ll , leaving the corresponding calcu-
lations for LZ to the reader. Written as a double inte-

gral, Llf is

(2/x)o foo k-oJ 2 (kx)kdk

o y+8
(18)

and, if we interchange the order of integration, this appears

r
as
Llf = w(x,u) feu) du, (19)
o
w(x,u) = Z8-2a.x-o ul+Za. x f: k2a.-oJ2Y+8(kX)JlJ(kU)kdk. (ZO)

This equation represents the first of three conditions which

we want the operators Ll and LZ to satisfy. The second
condition requires that w(x,u) = 0 when u > x; reference
to (15.Zl) shows that this is easily satisfied by choosing

Y = tlJ-a., for which

L f = 2x Za.-28- lJ IX ul+Za.+lJ(xZ_u2)O-Za.-l feu) duo (Zl)

1 (8-2a.-l)! 0
In a similar way, the choice °= l(v-lJ)
1
+ a.-13 gives
lJ-Za. 1 1
LZg = 2x Joo l+ZI3-v 2Z a.-I3+ZV-ZV-l
( 1v 1 1)' u (u -x ) g(u)du.
l -ZV+a.-I3- . x (ZZ)

With these expressions in (16), the problem is formally solved.

258 PART III: OTHER IMPORTANT TRANSFORMS

Restrictions on Parameters: If feu) and g(u) tend to

finite (non-zero) limits as u + 0 and u + 00 respectively,
then we need
-11-2 < 2a < 11 (23)

for both integrals (21) and (22) to converge at these limits.

This causes no difficulty, since we can choose a new a by
redefining A(k) by

k - 2a A (k) = k - 2a 1 A' (k) . (24 )

In order for the integrals to converge at u = x, we also

need the restrictions

1
Z(v-11) ± a-B > O. (25 )

However, by extending the definitions of the operators suit-

ably, we can lift this restriction. We turn to this task in
the next section.

16.3. Erdelyi-Kober Operators

The operators Ll and L2 are usually known as the
Erdelyi-Kober operators of fractional integration, and in the
conventional notation are defined as
-2a-2n
I f
2x fX u 2n +l (2
x -u 2)a-l feu) du, (26)
n,a (a-I)! 0

K f
n,a
=
2x 2n
(a-I)!
These definitions are restricted by
r x
u
-2a-2n+l (u 2_x 2 )a-l feu)

Re(n)
duo

> -
1
Z' Re (a)
(27)

> O.

Properties: We will investigate here the operators I

n,a
only, relegating the derivations of the corresponding proper-
ties of K to the problems. First we note that
n,a
§16. Dual integral equations 259

28
(28)
x I
n+0."a f (x).

Secondly, consider

(29)
2u-2n-2a-28 fU 2n+2a+l 2 2 8-1
x t (u -t) f(t) dt.
(8-l)! 0
Interchanging the order of integration and evaluating the
inner (u) integral by the variable change
2 2 2 2
s = [l-(t lu )]/[l-(t Ix )], which transforms it to a Beta
function, we find that

I n,a I n+ a,.,of(x) I f (x) . (30)

n,a+8

A similar treatment shows that

I I f(x) = I f(x). (31)

n+a ,8 n,a n,a+8

Connection with Differentiation: We introduce the differen-

tial operator
1 d
D (32)
x 2x dx

Then, using integration by parts, we see

I x - 2n Dx x 2n f(x)
n,a

2x- 2a - 2n
(a-I)! I: (2
x -u 2)a-l
1 d 2n
"2 du [u f (u)] du (33)

x- 2 In,a- l f(x),

and recursive use of this formula yields

260 PART III: OTHER IMPORTANT TRANSFORMS

(34)

Re(a-m) > O.
Similarly

x- 2n - 2a D x 2n +2a I f(x)
x n,a

2x-2n-2a-l d
ax IX u
2n+l 2 2 a-I
(x -u ) feu) du (35)
(a-I)! 0

and in this case, repeated application yields

2m-2n-2a m 2n+2a
x Dx x I (36 )
n,a

Re(a-m) > O.

Analytic Continuation: We will now lift the restriction

Re(a.) > 0 from the definition of I and formulas involv-
n,a
ing I First we use (34) to define I when
n,a n,a
Re(a) < 0 by choosing an integer m such that
Re(a+m) > 0 and writing

I
n,a
f(x) = x -2n-2a Dxm x 2n+2a+2m I n,a+m f(x). (37)

It is trivial to show that with this definition equations

(28), (34) and (36) hold without the restriction on Re(a).

Moreover,
I (38)
n, Of (x) = f (x) .

Now on setting B -a in (30), we find that if Re(a) < 0

then
§16. Dual integral equations 261

I I
n,a n+a,-a
-2n-2a m 2n+2a+2m
x Dx x I
n,a+m I n+a,-a
(39)
x- 2n - 2a Dm x2n+2a+2m I
x n+a,m

= I n+a,O'

Thus, from (38), another possible definition of I for

n,a
Re(a) < 0 is
-1
I
n,a
= I ( 40)
n+a, -a

Now let Re(a) > 0 and Re(B) > 0; then on taking the in-
verse of (30) we see

-1 -1 -1
I
n+a,B
I
n,a
= I
n,a+B'
( 41)

which, using (40) and making the substitutions n+a+B + n,

-B + a, and -a + B, becomes (30) again, except that now
Re(a) < 0 and Re(B) < O. By similar arguments, (30) and
(31) may be extended to all values of a and B.

Connection with Modified Hankel Transform: Consider

I S f (x)
n+a,B n,a

2x -2n-2a-2B fXo d 2 2 1 2 B 1
u u n+ a+ (2
x -u ) -
(B -1) !

( 42)

l+a -2n-2a-2B fooo I-a

2 x t f(t) dt
(B-1)!

(X d 2n+a+l 2 2 B-1
x u u (x -u ) J2 (ut).
)0 n+a
262 PART III: OTHER IMPORTANT TRANSFORMS

The inner (u) integral can be evaluated by using (15,19) and

(15,20), to give

I n+a,B Sn,a = Sn,a+B' ( 43)

By similar methods, we can also derive the identities

S I S (44)
n+a,B n,a n,a+B'

K
n,a Sn+a,B S (45)
n,a+B'

S K S (46 )
n,a n+a,B n,a+B'

Sn+a,B Sn,a I n,a+B' (47)

S S K ( 48)
n,a n+a,B n,a+B'

Finally, we note that in the notation of the Erdelyi-Kober

operators, the solution of the dual integral equations (11)
is given by

Sl A(k) h (x) , ( 49)

Z]..1-a,A-]..1+2a
where
1 1
A -]..1+-v-a+B (50)

.1,,"
2 2
and - 2a
x II f x < a
2]..1,A-]..1
hex) (51)

x -2B Kl
l'2S -]..1-a+B V-A
2 '
g, x > a,
§16. Dual integral equations 263

Problems
1. Show that the solution of the dual integral equations

fO A(k)
CIO

Jl(kx) dk = -1, x < 1

ICIO kA(k) Jl(kx) dk 0, x > 1

o
is
A(k) = cos(k) - 1
k

2. Show that the solution of the dual integral equations

I: A(k) sin(kx) dk f(x), x < 1

I: kA(k) sin(kx) dk 0 x > 1

is given by

A(k) = ~
7T
II
0
J
0
(kx){~ IX
ux 0
y fey) d Y } dx.
I 2 2
vX -y

3. Show that the solution of the dual integral equations

Io k A(k) JO(kx)
CIO
dk = f(x), x < 1

o x > 1

II r
is
A(k)
_2
sin (kx) dx Y fey) dy
7T 0 o
Ix7
4. Consider the problem of finding a function u(x,y) which

is harmonic in the half-plane y ~ 0, and satisfies the

mixed boundary conditions
264 PART III: OTHER IMPORTANT TRANSFORMS

uy(x,O) = vex), Ix I < 1,

u (x, 0) = 0 Ix I > 1.

Obtain a pair of dual integral equations by writing

v (x) v+(x) + v- (x),

v+ (x) 1
z[v (x) + v(-x)],

v- (x) i[v (x) - v(-x)] ,

with similar definitions of u+ and u , and using the

+
Fourier cosine transform on u and Fourier sine trans-
form on u. Show that, if v = 0, then

u(x,y)
2
1T f: A(k) cos (kx)e
-ky
dk,

where
A(k) =
_JI
0
t J 0 (kt) dt r !tV
o
v~s) ds

Find the remainder of the solution when v "f o.

S. Show that the solution of Problem 4 with v = -1 is

u(x,y) = r 0
J 1 (k) cos(kx)
e-ky dk
k

6. Verify the relations

2S
K x f(x)
n,ct

K K = K
n,ct n+ct,S n,ct+S,

x 2n+ 2n nxn x- 2n Kn,ct+n f( x ) = Kn+n ,ct f( x.

)

7. Consider the dual integral equations

f: G(k) A(k) Jv(kx) dk = f(x), X < 1,

f: k A(k) Jv(kx) dk = 0 x > 1,

§16. Dual integral equations 265

where G(k) is a given function with the asymptotic

form

By defining the functions

and
g (x) = r0
k A(k) Jv(kx) dk, x > 0

H(x)
xv-~+l
II YI-V (y 2-x 2) B-1 g(y) dy,
2B- l (B-1)! x
show that H(x) is determined by the Fredholm integral
equation

H(x) + x II K(x,y) H(y) dy

o
B v+l
2 (B-1)! sin(l-B)w B-v d IX y fey)
-----------------x
w
--d
x 0 (x2.y2)B
dy,
where
K(x,y) = Jl [k 2B G(k) - k] Jv_B(kx) Jv_B(ky) dk.
o
8. Consider a condenser made of two equal, coaxial, parallel,
circular metal discs of unit radius and separation t,
which are charged to potentials Vo and -VO' Show that
the potential ~(r,z) can be represented as (see
Figure 1)
[vof:{e-k'-e-kC'-l)) ACk) JOCkr) dk, z > t

Hr,z) = lvo/:{e-k'-ekC'-l)) ACk) JOCkr) dk, o< z < t

11VoJo{ekZ-ek(Z-t)}
l oA(k) J (k )
r dk , z < 0

provided that the function A(k) satisfies the dual

integral equations
266 PART III: OTHER IMPORTANT TRANSFORMS

Figure 16.1

J"" (1 - e- k R. ) A(k) JO(kr) dk = 1, r < 1,

o

J"" k A(k) JO(kr) dk = 0, r > 1.

o
Show that an application of Problem 7 leads to

2k2
A(k) = -;- t0 get) cos (kt) dt

where g (r) satisfies the Fredholm equation

g (r) _ ~ (1 g(t)dt = 1.
2
)-1 R. +(r-t)2

r
Show also that the capacity of the condenser is

C = -1 g(r) dr.
7T 0

Footnotes
1. The most comprehensive reference on mixed boundary-value
problems and dual integral equations is SNEDDON (1966).
§17. Integral transforms generated by Green's functions Z67

§17. INTEGRAL TRANSFORMS GENERATED BY GREEN'S FUNCTIONS

17.1. The Basic FOlmula

In this section we will investigate (in a purely for-
mal manner) some properties of the self-adjoint differential
operator [see (10.15)]

L[u] = [p(x) u' (x)]' + q(x) u(x), (1)

where p(x) and q(x) are given functions on the interval

a < x ~ b, and the functions u(x) under consideration all
satisfy homogeneous boundary conditions of the type [see
(lO.Z)]

alu(a) + aZu' (a) 0,

(Z)
blu(b) + bZu' (b) O.

We shall not take special note of the cases where a and/or

b are infinite, although this is crucial to a rigorous
analysis.

We first recall the following results from Section

10.1. Let ~A and WA be solutions of the equation

L[u] = Ar(x)u(x), (3)

where A is a constant and rex) a given function such that

~A and W
A satisfy the boundary conditions

al~A(a) + aZ~~(a) 0,
(4)
blwA(b) + bZw~(b) O.

Then the Green's function for the operator L-Ar which sat-
isfies the complete boundary conditions (Z) is
268 PART III: OTHER IMPORTANT TRANSFORMS

< >
cf>A (x )1/\ (x )
g(x,x' ;A) = (5)
II (A)

lI(A) = p(x) W[cf>A'W A]'

< >
where x is the smaller of x and x', and x the
larger. As we showed in Section 10, lI(A) is independent of
x, although in the present problem it is a function of A.
The Green's function is undefined when lI(A) = 0, that is,
when the functions cf>A and WA are linearly dependent,
making each one a solution of the eigenvalue problem given by
(2) and (3) together. Thus there is a close connection
between Green's functions and eigenfunctions; we refer the
reader to one of the many excellent texts for relevant de-
l
tails.

Consider the partial differential equation

ir(x) acf>(x,t) = L[cf>(x,t)] (6)

at
together with the initial conditions

cf>(x,O) = f(x) (7)

and the boundary conditions (2). Taking the Laplace trans-

form with respect to t, we obtain

[L - isr(x)] q,(x,s) = -ir(x)f(x) (8 )

where s is the transform variable. In terms of the

Green's function (5), the Laplace transform q,(x,s) is
b
q,(x,s) = -i fa g(x,x' ;-is) f(x')r(x')dx'. (9)

If we apply the inverse transform to q,(x,s), we recover a

§17. Integral transforms generated by Green's functions 269

function which is zero for t < 0 and equal to ~(x,t) for

t > 0; hence on setting t = 0 in the inversion integral we
recover the average. Explicitly,

f(x) 1
'IT
JC+l.·~ ds fba dx' r(x')g(x,x';-is)f(x'). (10)
C-l~

The contour in s must pass to the right of all the singu-

larities of g(x,x';-is) in the s-plane; these are at the
points s = iA corresponding to eigenvalues A of the opera-
2
tor L-Ar. Using the standard result that the eigenvalues
of a self-adjoint operator of the present type are real num-
bers, we find that c may be any positive number. In our
subsequent use of (10), we shall replace s by iA and
write

f(x) = 7fT
1 fl.' c+~ dA Jba dx' r(x')g(x,x' ;A)f(x'), c > o. (11)
lC-~

Similarly, by considering the equation ir~t = -L[~], we

obtain

f(x) = -
1'lTi fl.·C+~ Jba dx'r(x')g(x,x' ;A)f(x'), c < o. (12)
lC-~

We have derived these formulas without regard to a rigorous

justification of the steps involved. Such a justification
can be provided 3 when appropriate conditions are applied to
the functions p(x), q(x), and rex), although we will not

attempt this here. Alternatively, the formulas may be used

to generate useful particular results whose validity must be

checked by some other method.

17.2. Finite Intervals

If a and b are finite numbers, and p(x), p'(x),
q(x), and rex) are all continuous on the closed interval
270 PART III: OTHER IMPORTANT TRANSFORMS

a < x ~ b, and if in addition p(x) > ° and rex) > ° for

a < x ~ b, then the eigenvalue problem defined by (2) and (3)
4
is a regular Sturm-Liouville problem. It is shown in many
texts on mathematical physics that the eigenvalues \n and

eigenfunctions ~n(x) for such a problem have the following

properties:

(i) The eigenvalues are real and denumerable.

(ii) There is no point of accumulation of the eigenvalues;
i.e., there are only a finite number of eigenvalues in
any finite interval.
(iii) There is only one eigenfunction (to within an arbit-
rary multiplicative constant) for each eigenvalue.
(iv) Different eigenfunctions are orthogonal in the sense
that
b
Ja ~ n (x)~ m(x)r(x)dx = 0, n f m. (13)

Fourier Sine Series: The simplest regular problem corresponds

to p(x) = -1, q(x) = 0, rex) = 1, a = 0, and b =~. It is
trivial to show that the eigenvalues and eigenfunctions are

2
(mf/~) ,
(14 )
= sin(mTx/~), n = 1,2,3, . . . .

The Green's function (5) is also easy to construct; it is

sin(kx<)sin[k(~-x»]
g(x,x';\) = - (15 )
k sin U
Despi te the appearance of ff in these formulas, the Green's
function does not have a branch point at \ = 0, but merely
a series of simple poles at \ = (nTI/~)2. Adding (11) and

(12), we find that

§17. Integral transforms generated by Green's functions 271

f (x) = -
1 f dA
f
dx'
sin(kx<)sin[k(~-x»]
f (x' ) , (16)
2wi C k sin k~

where the contour is shown in Figure 1.

Imo.)

c
Re(A)

poles at A= (nn/1)2

Figure 17.1

Interchanging the orders of integration, we can

readily evaluate the A integral as a sum of residues. This
gives the familiar formula
2
f ~o
00

f(x) ="f L sin(nTIX/~) sin (nwx'/~)f(x')dx' . (17)

n=l
Eigenfunction Expansion: Now consider the general case. The
zeros of ~(A), which determine the poles of g(x,x' ;A), are
all discrete and simple; hence we can proceed by adding (11)
and (12) and evaluating residues as for the Fourier series.
To find the residues, we need the value of d~(A)/dA at
A = A i.e., we need to calculate
n'

(18)

To get d~A/dA and d~A/dA, we differentiate (3) with res-

pect to A:
272 PART III: OTHER IMPORTANT TRANSFORMS

dCPA
[L - Ar(x)]-= r(x)cp,(x), (19)
dA /I.

and solve using the Green's function. After some straight-

forward algebra, we get

(20)

When we put A = An' CPA and ~A become the same function,

apart from a normalizing factor, and in evaluating the resi-
due this factor cancels, so that the residues are

CPn (x)CPn (x') (21)

b
J r (x)CP 2 (x) dx '
a n
where we have written cP (x) for CPA (x). The complete ex-
n n
pansion of a function f(x) is

f (x) (22)

Note that f(x) satisfies the boundary conditions (2), as is

evident from its definition, and so there is no conflict
caused by the fact that every term in (22) also satisfies
these boundary conditions.

17.3. Some Singular Problems

It is evident from the foregoing that we must consider
singular problems if we are to obtain integral transforms.
Before generating any new transforms, we will show how the
Fourier, Mellin, and Hankel transforms are related to
Green's functions.

Fourier Transform: Let p(x) = -1, rex) = 1, q(x) = 0,

a = _00, and b = +00. We require as our boundary conditions

§17. Integral transforms generated by Green's functions 273

that ~\(x) and ~\(x) remain finite as x + +00, respec-

tively, remembering that Im(\) > 0 in (11). The functions
we need are

-ikx
¢\ (x) e

~\(x) = e ikx
(23)
t. (\) 2ik,
k2 \ , o < arg (k) < Tf/2.

and so (11) reads

f(x)
1
2Tf
t c + oo
lC- OO
d\ r _00
e
- ik(x<-x»
k
)
f (x') dx', c > O. (24)

Similarly, we may evaluate the functions relevant to (12) to

r
get
f(x) 1
IiI fiC +
oo
d\
_00
c > 0,
ic- oo (25)

-Tf/2 < arg(k) < O.

When \ is on the negative real axis, these two expressions

differ only in sign. Adding them and letting c + 0, we get

f(x) =
1
2Tf foo0 I'r
d\ foo
_00 cos[k(x-x')] f(x') dx'. (26)

If we replace 1\ by k and break up the cosine into com-

plex exponentials we obtain the exponential Fourier trans-
form.

An Alternative Formula: In each of our examples above, the

eigenvalues of the operator have had a lower bound and the
Green's function g(x,x' ;\) has been analytic across the
real axis to the left of the lowest eigenvalue. Whenever
this is the case, we can add (11) and (12), to obtain the
formula
274 PART III: OTHER IMPORTANT TRANSFORMS

b
f(x) 2;i Ie dA fa g(x,x' ;A)f(x')r(x')dx', (27)

1m ().)

c
Re ().)

minimum
eigenvalue

Figure 17.2

where the contour e is shown in Figure 2. In the regular

case, the eigenvalues are discrete, and the value of the
integral is given by a sum of residues. With singular prob-
lems, however, the Green's function generally has a branch
cut along the real axis, and other appropriate methods of
evaluating the integral must be found.

Mellin Transform: If we set p(x) x, q(x) = 0 and

-1
r (x) = x in the interval o< x < 00, then we again have a
singular problem. The equation for </l.A and 1/1 A is

(xu')' = AU/X (28)

with the solutions On the contour e we choose

by 0 < arg If" < 7T; then Re (ilr) < O. Thus
§17. Integral transforms generated by Green's functions 275

and the integral reduces to

-1 Joo d\ fIX>
[g(x,~;\+ic) - g(x,~;\-ic)]f(O
d~
-"F
2'Jfi 0 0 <,

r r
(30)
1
d\ [ -ilX" ilX" ilA -iv'rJ
x ~ +x ~
d~
f(~) ~
4'Jf o IX 0
2 2
Now write q =\ in the first term and ( -q) \ in the
second; this gives

(31)

which is the Mellin transform formula with p = iq.

Hankel Transform: Proceeding as above, we consider the

2
interval 0 < x < IX> with p(x) = -x, q(x) = v Ix and
1
= x, where Re(v) > - Z· The Green's function, finite
r (x)
as x -+- 0 or IX> , is easily constructed as

g(x,x' ;A)
(32)
o< arg(k) < 'Jf/2.

Using the relations

-e - i'Jfv H~2) (z), (33)

-'Jf < arg(z) < 0,

we readily reduce the integral around the branch cut

o< A< IX> to the Hankel transform formula
276 PART III: OTHER IMPORTANT TRANSFORMS

f(x) Jo kdk
OO foo J (kx)J (k~)f(~) ~d~. (34 )
0 v v

17.4. Kontorovich - Lebedev Transform

The Hankel transform is the first example we considered
for which q(x) ~ O. In fact, the Hankel transform, with v
1
replaced by v + 2' may be obtained from the choice p (x) = 1,
2 2
q(x) = v Ix , rex) = 1 on the interval 0 < x < 00. Setting
v = 0 gives the Fourier sine transform as a special case,
thus we can regard the Hankel transform as a generalization
of the Fourier sine transform to nonzero q(x). In a simi-
lar way, we can generalize the Mellin transform by the choices

p (x) -x,
2
q (x) k x,
(35)
rex) l/x,
0 < x < 00.

The functions ~A(X) and WA(x) must satisfy the differen-

tial equation

~"A +
1
:x ~~ - rk 2 - ~l2 ~A = 0, (36)
l x J
whose solutions are modified Bessel functions of order FA.
If we choose a = ~ by Re(a) > 0 for A on the contour
C, then the Green's function which goes to zero as x + 0 or
x + 00 is
(37)

To evaluate the integral (27) we need the value of

g(x,x' ;A+iE) - g(x,x' ;A-iE) for real positive A and
E + O. Using the results of Problem 20.17 we can write
917. Integral transforms generated by Green's functions 277

(with v ill

I -lV
. (kx) - 1.1V (kx) 2i
= ~ . h ( nv ) Kiv (k)
sln x,
(38)

and hence

g(x,x' ;A+iO) - g(x,x' ;A-iO)

(39)
2i sinh(nv) K.1 v (kx) K1.,,(kx').
n v

Using this result in (27), and subsequently changing the in-

tegration variable to v, we obtain a form of the Kontorovich-
Lebedev transform, viz.

f(x) = -
2 Joo vsinh(nv) dv Joo K. (kx)K. (kx')f(x')+-.
d ' ( 40)
n2 0 0 1V 1V X

An Alternative Formula: Equation (40) is the original trans-

form given by Kontorovich and Lebedev; however, there is an
alternative formula which demonstrates the close connection
with the Mellin transform. To obtain it, we note that (38)
demonstrates that the Ia functions may be written in terms
of K, so that the distinction between x< and x> could
a
be dropped, provided we maintain convergence of the inte-
grals. We therefore set x< = x and x> = x' in (37), and
substitute into (27). Using a as a new variable, this
gives

f (x)
1
ni
J i~
-1 00
a da I
a
(kx) f 0 Ka (kx')f(x')+.
OO d '
x
( 41)

Regions of Convergence: Consider the Kontorovich-Lebedev

transform of a function f(x), defined as

F(k,v) = f o Kv (kx)f(x)
OO d
~.
x
( 42)
278 PART III: OTHER IMPORTANT TRANSFORMS

If we assume that f(x) has the asymptotic form

fex) - )
Jx B
r:J.'
x ... 0
(43)
lX ,

then, using the asymptotic forms of the modified Bessel func-

tions
x ... 0,

x ... 0, (44 )

-x x .... co,
e

we see immediately that if f(x) is a "reasonable" function

(for example, if it satisfies Dirichlet's conditions) then
the inte.gra1 (42) converges in the region Re(v) < B for all
k > O. Thus the inversion integral is defined whenever
B > 0; it can be shown that the transform pair may be ex-
tended to functions for which B < 0 by moving the inversion
contour.

Relation to Mellin Transform: If r:J. < B, then in the strip

r:J. < Re(v) < B we may let k ... 0 in (42) to obtain

F(O,v) = 1 1[f(x);v]. ( 45)

2vv!
Furthermore, in this limit the inversion integral (41) is the
Mellin inversion.

17.5. Boundary-value Problems in a Wedge

To illustrate a simple use of the Kontorovich-Lebedev
transform, we consider the problem of determining a func-
tion u(r,a,z) which is harmonic in the wedge
§17. Integral transforms generated by Green's functions 279

o ::. r < 00,

o < e < a., (46)
o < z < .Q.,

and which satisfies the boundary condition

u(r,a.,z) f (r, z) (47)

on one boundary and is zero on all the other boundaries.

First, we introduce a Fourier series in z, using the func-
tions sin(nnz/.Q.), which satisfy the required boundary condi-
tions at z = 0 and z =.Q.. Thus we write

u(r,e,z) E un (r,e) sin(nnz/.Q.),

n=l
( 48)

f(r,z) = l fn(r) sin(nnz/.Q.),

[n
n=l
and the coefficients un (r, e) are determined by

fL
l ar2
+ 1..L
r ar
+ -
I
r2 ae 2
a2
--- ~. un (T • ') = 0,

tinCr,a) a, ( 49)

un(r,a.) fn(r).

We denote the Kontorovich-Lebedev transforms of un(r,e) and

fn(r) with respect to r by Un(k,v,e) and Fn(k,v) res-
pectively; on multiplying the differential equation for un

by rKv(nnr/.Q.) and using integration by parts twice 5 we

reduce (49) to

J d\ 2}U (nn/.Q.,v,e)
1 de
+ v
n
0,

U (nn/.Q.,v,O) 0, (SO)
n
Un(nn/.Q.,v,a.)
280 PART III: OTHER IMPORTANT TRANSFORMS

These equations are readily solved to yield the expression

sinh(v6)
UnCnTI/t,v,8) = sinh(va) Fn(nTI/t,v), (51)

from which an explicit integral representation of the solution

can be constructed.

17.6. Diffraction of a Pulse by a Two-Dimensional Half-Plane 6

As a more difficult example of the use of the
Kontorovich-Lebedev transform, we will construct an explicit
representation of the Green's function for the two-dimensional
scalar wave equation,

(52)

subject to the boundary condition g = 0 on the positive

x-axis and the initial conditions g = ag/at = 0 at t = o.
This Green's function represents the wave pattern generated
by a pulse at : = :0' t = 0, including the effects of dif-
fraction by a semi-infinite barrier. We have solved the free-
space problem in Section 10; we must now find the effect of
the barrier on this solution. Introduce polar coordinates,
and also the variables

2 2
R [r + r 0

(53)
2 2 1/2
RO = [x + r 0 - 2 x r 0 cos 8 0 ] .

All of these quantities are depicted in Figure 3.

§17. Integral transforms generated by Green's functions 281

( r,9)
R

(X,O)

barrier
Figure 17.3

Laplace Transform: We introduce a function ~ by g = gO+~'

where is the free-space solution. 7 Then the equations

which determine ~ are
[1{2 _ L~] ~(r,e,ro,eo,t) = 0,
c2 at 2

{ ~:/2.~2~R;;::' (54)

~(r,e,ro,eo'O) 0,
~t(r,e,rO,eO'O) o.

We denote the Laplace transform of ~ with respect to t

by ~; then the Laplace transform of these equations is

(55)

Kontorovich-Lebedev Transform: It is convenient to intro-

duce another new function
282 PART III: OTHER IMPORTANT TRANSFORMS

(56)

2
which goes to zero as r -+- O. After multiplying by r , the
equation for W is

( r2 ~
~l
+ d
r "r + ::2 - p~~2 lw(r,a,ro,ao'p)
dr 2 a (57)
_ p2r2 ±o (prO/c)
~ 2'IT
and the boundary conditions are 8

(58)
at a = 0 and a = 2'IT.

These equations are now transformed by the Kontorovich-Lebedev

transform with respect to r. The transform of w is

(59)

and (57) and (58) transform to

{
~ + v2 } = _ vKO(prO/c) (60)
da 2 4 sin('lTv/2) ,
Kv(prO/c)cos v('IT-a O) KO(prO/c)
(61)
2v sin ('lTv) 4v sin('lTv/2)

at a = 0 and a = 2'IT.
The solution to these last two equations is

(62)
KO(pro/c)
4vsin('lTv/2)
The Solution: The function W is obtained by evaluating the
inversion integral
§17. Integral transforms generated by Green's functions 283

(63)

and subsequently we must invert the Laplace transform to get

~ and g. The technical details of this inversion, which are
given in Turner's paper, are quite complicated; we merely

quote the result here, which is

C.64 )

(65)

t > Ric

(66 )

L Pn
00

lr 2+r 2-c 2t
0
j 1 1
sin(n+Z)6sin(n+Z)60,
n=O 2rrO

Ir-rol < ct < r+rO (67)

0, otherwise

2 2 1/2
where R1 [r +r O-2rr O cos(6+6 0)] is the distance to the
image of r O,6 0 in the plane y = O. For a further discus-
sion of this solution, the reader should consult Turner's
original paper.
284 PART III: OTHER IMPORTANT TRANSFORMS

Problems
1. By setting p(x) = -1, q(x) = 0, and rex) = lover the
interval a ~ x < 00, with the boundary conditions g = 0
at x = a, g finite as x ~ 00, derive the transform pair

F(k) I {sin(kx) cos

= oo
a
(ka) - cos (kx) sin (ka)} f(x) dx,

f(x) = J: {sin(kx) cos (ka) - cos (kx) sin (ka)} F(k) dk.

Consider the limit a ~ o.

2. Repeat Problem 1 with g =0 replaced by dg/dx o at

x = a.

3. Consider the Green's function obtained over the interval

o< x < 00 by setting p(x) = -1, g(x) = 0, rex) = I, and

ax
~x + h g = 0, x = a.

Show that if h < 0 the resulting integral transform is

F(k) J: ~(k,x) f(x) dx,

f(x) 2h 2 Joo m(k,x) F(k) dk,

'IT 0 h 2 + k2
where

Hk,x) sin (kx) - (k/h) cos (kx) ,

and that if h > 0, there is an extra contribution from

a pole at A _h 2 , giving
2 00

f(x) = ~ J ~(k~x) F(k) dk + 2Ahe- hx ,

'IT 0 h + k2

A= J: e- hx f(x) dx,
§17. Integral transforms generated by Green's functions Z85

with the other quantities defined as before.

4. Recover the Weber transform (Section 15.5) by using

Green's functions.

5. By considering the Hermite equation (see Section ZO), re-

cover the eigenfunction expansion
Hn(x) Joo -x
I
00
f(x) = e H (x) f(x) dx.
n=O zn n! liT -00 n
6. A quadrant-shaped slab 0 ~ x < 00, 0 ~ y < 00, 0 < z ~ ~,

has the face x = 0 held at temperature TO' while the

other faces are held at temperature zero. Show that the
temperature distribution is
8T O 00 sin[(Zn+l)nz/~]
u(r,6,z) = I
n Z n=O Zn+l
cosh(nv/Z) sinh(6v)
Kiv[(Zn+l)nr/~] dv.
sinh (nv/Z)
Using the integral representation

K. (x) = 1 Joo cos(x sinh t) cos(nv) dt,

lV cosh(nv/Z) 0

reduce the result to. the simpler form

sinh[(2n+l)nz/~]
u = I
n=O 2n+1
00 cos[(Zn+l)nr sinh(t/~)]
xJ o - ----
cosh Zt + cos Z6
dt.

7. A point charge q is placed near the edge of a conduc-

tor of rectangular shape held at zero potential (see
Figure 4). Find expressions for the potential and the
density of charge induced on the boundary.
286 PART III: OTHER IMPORTANT TRANSFORMS

q -- a

Figure 17.4

8. A line source of current, J = J O exp(iwt), is placed

parallel to the edge of a thin conducting sheet

o~ x < ~, -~ < y < ~, in the plane of the sheet at a
distance a from the edge. Show that the density of
current induced in the sheet is

j J0 [~) 1/2 e - ik(x+a)

21T x x+a

k w/c.

9. Plane waves whose propagation is governed by the

Helmholtz equation are incident on a screen in the form
of a half-plane r ~ 0, 9 = n, on which the boundary con-
dition is a~/an O. The incident wave is
§17. Integral transforms generated by Green's functions 287

Show that the total field ~ is given by

~ 1 e- ikr COS6[1 + e in / 4 erf{12kr sin(6/2)}]

2

+ ~e-ikr cos(6-2a)[1 _ e in / 4erf{!2kr sin(6/2-a)}].

Footnotes
1. For example, STAKGOLD (1968).

2. STAKGOLD (1968), Ch. 4.

3. TITCHMARSH (1953), Ch. 6.

4. If anyone of these conditions is not satisfied, we have

a singular problem.

5. These manipulations involve assumptions about the solu-

tion which can only be verified a posteriori. Alterna-
tively, we could work with a suitable set of generalized
functions from the outset.

6. R. D. Turner, Q. Appl. Math. (1956), .!i, 63.

7. MORSE & FESHBACH (1953), p. 842.

8. The evaluation of the right-hand side (58) is discussed

at length in Turner's paper.
 Part IV: Special Techniques

1
§18. THE WIENER-HOPF TECHNIQUE
The solution of boundary value problems using integral
transforms is comparatively easy for certain simple regions.
There are many important problems, however, where the bound-
ary data is of such a form that although an integral trans-
form may be sensibly taken, it does not lead directly to an
explicit solution. A typical problem involves a semi-
infinite boundary, and may arise in such fields as electro-
magnetic theory, hydrodynamics, elasticity, and others. The
Wiener-Hopf technique, which gives the solution to many
problems of this kind, was first developed systematically by
Wiener and Hopf in 1931, although the germ of the idea is con-
tained in earlier work by Car1eman. While it is most often
used in conjunction with the Fourier transform, it is a
significant and natural tool for use with the Laplace and
Mellin transforms also. As usual, we develop the method in
relation to some illustrative problems.

288
§18. The Wiener-Hopf Technique 289

18.1. The Sommerfeld Diffraction Problem

In this section we will study a problem involving the

reflection and diffraction of waves in two space dimensions,

commonly known as the Sommerfeld diffraction problem. 2 We

commence with the wave equation

a2<p 2 2
c V <p (1)
at7
in the unbounded region -00 < x < 00, -00 < y < 00. We will not
investigate the initial value pr6blem, but rather look for

particular steady-state solutions with the time dependence

exp(-i~t). Then the wave equation becomes the Helmholtz

equation in two dimensions, namely

2 2
(V +k ) <p (x,y) 0, k = ~/c. (2)

We impose three conditions on (2) to complete the specifica-

tion of the problem:

(i) We suppose that the motion is caused by a steady inci-

dent plane wave

-ik(x cos e + y sin e)

e (3)

which represents plane waves proceeding in a direction

making angle e with the positive x-axis.

(ii) We assume that the positive x-axis is a barrier to the

waves. Specifically, we impose the boundary condition

a<play = 0 for y 0, x > O. If we introduce as the new

unknown function ~ = <p - <p.Inc' this amounts to the boundary
condition

i k sin e e -ik(x cos e + y sin e) x > O. (4)

290 PART IV: SPECIAL TECHNIQUES

Because of this ~ may be discontinuous across the positive

x-axis. However, we must have continuity for negative x,

giving the further boundary condition

~(x,o+) - ~(x,O-) 0, x < o. (5)

(iii) In choosing the inversion contour, we must ensure that

the resulting solution is the steady-state component of the

(more complicated) initial-value problem which we ought to

have solved. We saw in Section 8.4 that one way to do this
is to replace n by n + ie, where e > 0 if n > 0; in the
present case this amounts to replacing k by k + i£. After
the problem is solved we allow £ to become zero.

Preliminary Considerations: The Wiener-Hopf technique relies

on the use of Liouville's theorem,3 and hence on having some
information about the analytic properties of the Fourier
transforms involved. It is obvious that in using an integral
transform to solve any problem we are making some assumptions
about the unknown function. In the present case, we will
need information about the analytic properties of the trans-
form of ~, and this comes from physical considerations.
Referring to Figure 1, there are three regions in which we
expect ~ to behave quite differently, which we have
labeled I, II, and III. In region I, ~ should consist of

the reflection of the incident plane wave plus an outgoing

diffracted wave coming from the edge of the barrier. In
region II, we expect ~ to be only a diffracted wave.
Region III is in the 'shadow' of the barrier, and here the

complete solution ¢ must be only a diffracted wave. Hence

~ = ¢ - ¢inc consists of a diffracted component and the
§18. The Wiener-Hopf technique 291

waves

__________________
II
-;~~----------------~x

Figure 18.1

negative of ~inc'

We are particularly concerned with the behavior of

these functions on the x-axis, since this is where the boun-

dary conditions are applied. The incident waves have ampli-

tude exp(Ex cos e), and we can estimate the amplitude of the
diffracted wave by the following argument: the diffraction
is caused by the edge of the barrier, and the strength of
this term at a distance r from the origin must be propor-
tional to the strength of the incident wave at the origin at

time ric previously. However, ~inc is increasing in time

as exp(5t), where 5 = EC, so the diffracted wave must de-

crease as exp(-Er) for large r. This gives us the esti-

mates
cos e
Je EX x > 0
1jJ(x,O) - (6)
'le EX x < 0
From (6) we expect that the Fourier transform will converge
292 PART IV: SPECIAL TECHNIQUES

in the strip E cos e < 1m (w) < E, so we confine the inver-

sion contour to this region.

Basic Procedure: We take the Fourier transform with respect

to x of the equation for

r
~,obtaining

d_ 2 + 2 2l
k - w"' ~(w,y) = 0, (7)
Ldy2 J
with the independent solutions

~(w,y) = exp(± y ;G2 - k 2 ) few). (8)

The execution of the method requires that we consider only

transforms which are analytic in a strip containing the in-
version contour. From phy'sica1 considerations, moreover, we
must choose from the solutions (8) a function which is bounded
as Iyl + 00, and this requires that we have knowledge of the
sign of Re (~2 - k 2) on the contour. It is easy to show
that if we choose the branch of (8) so that ~ = ik for
w = 0, and cut the w-p1ane as indicated in Figure 2, then

Im(w) branch
cut

w = k+i£
Inversion contour
----------------~--4---~-------------------Re(w)

w=-k-i£
L=(k+;£) cos 0
branch
cut

Figure 18.2
§18. The Wiener-Ropf technique 293

Re (;G2 - k 2) ~ 0 in the strip -E < 1m(w) < E, and suitable

solutions of (7) are given by

'I' (w,y) 'I'(w,O±) exp (-Iyl ,lwZ_kZ), (9)

where 'I'(w,O±) are still to be determined.

We must now incorporate the boundary conditions at
y = 0, recognizing the fact that they are different for posi-
tive and negative x, which is the cause of the complication.
We therefore split the Fourier transforms of the boundary
values into two parts, defining the four functions

A_(w) = 2
1 fO {w(x,O+) - w(x,O-)} e
iw x
dx
_00

0,

A+ (W)
i f: {w (x, 0+) - W(x,O-)} e
iwx
dx,

B (w)
t wy(x,O) e
iwx
dx, (10)

I:
_00

B+ (w) wy(x,O) e iwx dx

-k sin 6 1m (w) > E cos 6.

w - k cos6
The whole difficulty with the problem is that while we
know the functions A_(w) and B+(w), we do not know A+(w)
and B_(w), so that there is not enough explicit information
to write down 'I'(w,O±) immediately from the boundary condi-
tions. We do know that wyand hence 'I'y is continuous at
y 0, and on differentiating (9) and setting y = 0- and
y 0+, this gives the relation

Z 'I' (w,O) 'I'y(w,O+) + 'I'y(w,O-)

y
(11)

- ;GZ_k2 [ 'I'(w,O+) - 'I'(w,O-)],

294 PART IV: SPECIAL TECHNIQUES

or in terms of the functions defined above

(12)

Further progress can only be made by appealing to the ana1y-

tic properties of the functions A+(w) and B_(w). It fol-

lows from (6) that A+(w) is analytic in the region

Im(w) > E cos 8, and B_(w) is analytic in the region

Im(w) < E. The factor ~ has branch cuts in both of
these regions, so we write it as the product ~ 1W+k,
which separates the two branch points. Using this factoriza-
tion, we can rearrange (12) as

_1_ _ [B (w) + B (w)] (13)

Iw-k +

There are three combinations here. B_(w)/IW7K, which is a

new unknown function, is analytic in the region Im(w) < E,

and -/w+k A+(w), which is also unknown, is analytic in the

region Im(w) > E cos 8. We therefore examine the third
function, B+(w)/Iw-k, for which we have an explicit formula,
and write it as the sum of two functions, each analytic in
one or other of the two regions mentioned. By trivial

algebraic manipulations, we can write

lIM
( 1 1
Ik(cos
(14 )

We have here denoted the first term, B+(w)/;k(cos 8-1), by

F+(w); it is obviously analytic in the region Im(w) > E cos8

since the denominator is independent of w. For the second

term, we have removed the only singularity in B+(w), a pole

§18. The Wiener-Hopf technique 295

at w = k cos 8, by arranging for the terms in braces to have

a simple zero there. Consequently F_(w) is analytic in the

region Im(w) < E. Using this decomposition we can again

rearrange (13) to define a new function E(w) by

B_ (w)
E (w) - - + F (w)
IW-1<
(15 )

The point of this is that E(w) is an entire function, since

it is defined in two overlapping half-planes by functions

which are analytic in those half-planes and which coincide in

the strip of overlap, E cos 8 < Im(w) < E. Therefore each

function is the analytic continuation of the other and E(w)

is entire. Under rather weak assumptions (see Problem 1 for
further details) we can show that B_(w) and A+(w) tend to
zero for large w in the respective regions Im(w) < E and
Im(w) > E cos 8, so that the entire function E(w) is
bounded and tends to zero for large w. Hence by Liouville's
theorem, 3 we conclude that E(w) = O. Equation (15) now
gives explicit formulas for the unknown functions A+(w) and
B_(w), and by working backwards through the definitions we ob-
tain for ~(w,y) the explicit form
-i sgn(y) l2l<: cos(8/2) exp(-Iyl /w2_ k2)
~(w,y) = (16 )
(w-k cos 8) ~

The Solution: We insert (16) into the inverse Fourier trans-

form and allow E to become zero, whi~h also involves mov-

ing the contour off the real axis. The solution can then
be written as the integral
296 PART IV: SPECIAL TECHNIQUES

sgn(y) .fk[2cos(8/2) r e- iwx - Iyl ~

1/!(x,y) =
7Ii
J dw, (17)
Cl (w-k cos 8) ,!i;j':"l(

where k is real and positive and the contour Cl is shown

in Figure 3.

1m (w)

w =k
Re(w)

w=-k w=k cose

c,

Figure 18.3

We will demonstrate that the integral (17) does in-

deed describe a solution having the general properties
which we discussed in connection with Figure 1. For this
purpose, we introduce the variables r and X by4

x = r cos X,
(18)
Iyl = r sin X,
so that the regions of Figure 1 correspond to

I: 0 < X < 71-8, Y > 0,

>
II: 71-8 < X < 71, Y < 0, (19)

III : 0 < X < 71-8, Y < O.

Next we introduce the change of variable

§18. The Wiener-Hopf technique 297

W : -k cos(X+it), _00 < t < 00 (20)

Elimination of t shows that the contour described by (20)

is a hyperbola whose major axis is the real axis in the w-
plane. The vertex is at the point w: -k cos X, and the
asymptotes make an angle ~-X with the real axis (see Figure

4). It can be shown that the integrals along the arcs rl

and r2 tend to zero as their radius R + 00, hence we can

deform the contour Cl in (17) to this new contour, provided

we pick up the residue at the pole at w: k cos e if the new
contour is on the opposite side of the pole from the original
contour. Temporarily denoting this new integral by J, we
have the following results for the solution ~:

( i) In region I, k cos X > k cos(~-e) : -k cos e, so that

the two contours enclose the pole. Thus

ik(x cos e - y sin e) J

~ : ~inc + e + . (21)

Here the second term, which is the residue at the pole, is a

reflected plane wave as expected.

(ii) In region II, the contours are on the same side of the
pole, and we have

~inc + J. (22)

(iii) In region III, we again have a contribution from the

pole, but because of the different sign of y, it exactly
cancels lnc ,and we obtain
~.

~ : J. (23)

The integral J can be written by straighttorward

298 PART IV: SPECIAL TECHNIQUES

substitution of (20) into (17) as

J = } sgn(y) sin(8/2) J00

-00
e
ikr cosh t . . /
sln[(x+1t)2] dt. (24)
cos 8 + cos(X+it)

It is possible to perform further manipulations on this inte-

gral which reduce it to the Fresnel integral, but we will not

do that here. What we will note is that for large r the

major contribution comes from the region t ~ 0, since the

Re (w)

c,

Figure 18.4

exponential function varies rapidly as t increases. As a

first approximation then, provided that cos 8 + cos X is
not too small, we will approximate the factors sin[(x+it)/2]
and cos(X+it) by sin(x/2) and cos X respectively.
Comparison with (20.66) shows that we then have a Hankel
function, and since we have already assumed that r is
large, we replace this Hankel function by its asymptotic
form (Problem 20.11) to obtain the approximation
§18. The Wiener-Hopf technique 299

J '" e
i7T/4
sgn(y) rSin (6/2)
l cos 6 +
sin(x/ 2 ) 1L~.J
cos X J vrkrJ
1/2e ikr . (25)

Thus J represents an outgoing diffracted wave whose amp1i-

tude is proportional to the factor
sin(6/2) sin(x/2)/[cos 6 + cos xl. Analysis of J using the
method of steepest descents confirms this conclusion, and
also shows how it behaves asymptotically when
cos 6 + cos X '" 0, where it is evident that (25) is invalid.
(See also Problem 3.)

18.2. The Wiener-Hopf Procedure: Half-Plane Problems.

The typical problem which may be solved by the Wiener-
Hopf technique involves the solution of equations which only
give explicit information over a semi-infinite range of a
variable. In a mixed boundary-value problem, for instance,
we may know the boundary value of one combination of the un-
known functions for x ~ 0, and of a different combination
for x < O. After taking the Fourier transform, and finding
the general relationship between the partially specified
but still unknown functions, we are then faced with the
following problem: find unknown functions ~+(w) and ~_(w)

satisfying
A(w) ~+(w) + B(w) ~_(w) + C(w) = 0, (26)

where this equation holds in a strip a < Im(w) < S, ~+(w)

is analytic in the half-plane Im(w) > a, and ~_(w) is

analytic in the half-plane Im(w) < S. The functions A(w),

B(w) and C(w) are analytic in the strip. The fundamental
step is to find a 'factorization' of A(w)/B(w), i.e., to
find functions K+(w) analytic in Im(w) > a and K_(w)
300 PART IV: SPECIAL TECHNIQUES

analytic in 1m (w) < 13 such that

A(w) K+ (w)
(27)
B (w) K_ (w)

Then we can rewrite (26) as

K+ (w) <1>+ (w) + K_ (w) IJI_ (w) + K_ (w) C(w) IB (w) = o. (28)

For some problems, this decomposition can be found by inspec-

tion; a contour integral method which we give in Section 19

may be useful with other problems, and more techniques may
be found in the 1iterature. S Assuming that the factorization
(28) has been performed, we must now effect the further decom-
position
(29)

where again F+(w) is analytic for Im(w) > a and F_(w)

is analytic for Im(w) < 13. Taking (28) and (29) together,
we can now define an entire function E(w) by

(30)
-K_ (w) IJI_ (w) - F_ (w).

At first sight it may seem that we have merely defined a

function which is analytic in the strip a < Im(w) < 13, but
in fact each of the two ways of defining E(w) makes it

analytic in a semi-infinite region, and since the two regions

overlap we can use the principle of analytic continuation to

show that E(W) is entire.

Now suppose that we can show that as Iwl + 00
§18. The Wiener-Hopf technique 301

iK+(w) ~+(w) + F+(w)i t!5'(iwi r ), Im(w) > ex,

(31)
iK_(w) ~_(w) + F (w)i t!5'( i wis), Im (w) < s;

then from Liouville's theorem we can conclude that E(w) is

a polynomial of degree no higher than the largest integer
smaller than both rand s. This may reduce the problem
to that of determining the coefficients of a polynomial, or
it may be that the solution to the problem is not unique, in
which case the coefficients play the role of arbitrary con-
stants.

18.3. Integral and Integro-differential Equations.

The original work of Wiener and Hopf was in conjunc-
tion with the integral equation

Hx) }J: E(x-t) Ht) dt,

(32)
E (u) = -E1C-iui)'

where El (z) is the exponential integral. This equation oc-

curs in the study of radiative processes in astrophysics, and
is known as Milne's equation. More generally we may con-
sider the problem of solving the equation

AHx) +
J: k(x-t) Ht) dt = f(x), x > O. (33)

The first move is to extend the range of the equation to all

x, which can be achieved by writing

A~(X) + f: k(x-t) ~(t) dt = f(x) + ~(x), (34 )

where
302 PART IV: SPECIAL TECHNIQUES

</> (t) 0, t < 0,

f (x) 0, x < 0, (35)
1/J (x) 0, x > O.

The Fourier transform of (34) is

H+ (w) + K(w) if>+(w) = F+(w) + 'I'_(w) , (36 )

where we have added subscripts to the transforms to indicate

the regions of the complex w-plane in which they should be
analytic. This equation is a special case of (26), and should
therefore be amenable to the Wiener-Hopf technique. The more
general case of an integro-differential equation obtained
from (33) by replacing A by a linear differential operator
can be analyzed in a similar manner (see below for an example).

Example 1: To illustrate, we put k(x) = exp(-alxl), a > 0,

and consider the homogeneous problem

"" e-alx-yIHy)
dy = Hx), x > O. (37)
fa
Equation (36) now becomes

[ .' }+ (.)
2a
a2 +
= if>+Cw) + 'I'_Cw), (38)

which can be immediately factored as

( 2a - a
2 2 I
l - w t+(W) = (w- ia) '1'_ (w)
w + ia
(39)
E (w).

If </>(x) and 1/J(x)

are bounded as Ixl + "", then if>+(w)

and 'I'_(w) are of order Iwl -l for large Iwl in the

upper and lower half-planes respectively, and (39) defines a
§l8. The Wiener-Hopf technique 303

bounded entire function E(w). By Liouville's theorem,

E(w) = A, where A is an undetermined constant. Thus
~
( w + ia
Al-2-a;;"--a"';2~--w-2 J'
( 40)
A
'¥ (w)
w - ia
Note that the inversion contour must pass above the poles of
~+(w) and below the poles of '¥_(w) so as to satisfy (35).
Inversion now yields

~(x) A[cos(bx) + (a/b) sin(bx)),

( 41)
b

It is instructive to reflect on the fact that there is only

one arbitrary constant in this solution, while the integral
operator in (37) is the Green's function for the second-order
differential operator (d 2 /dx 2 - a 2 ). Acting on (37) with
this operator we get the differential equation

2
~" + (2a-a H = 0, C42)

which has two independent solutions. However, the integral

equation also contains the boundary condition

~'(O) = a foo e- ay ~Cy) dy

o
(43)
= a ~ (0) ,

and when this is included with (42) it leads to the solution

(41) again.

Example 2: We consider again an equation solved in Section

5.2, namely
304 PART IV: SPECIAL TECHNIQUES

A r o
e-alx-yl¢ey)dy f(x). (44)

Here we must factor the equation

[ ,22:'w 2 ].+CW) , F.Cw) + '_Cw), ( 45)

Proceeding as for (38) we obtain

1 1
l
( 2 2

j
w + a + ia
<P+(w) = y+(w) (47)
'2aA 2aA
and
2
a f(x) f" (x)
¢ex) = - + B[ao(x) + 0' (x)], (48 )
2aA

where B = iA/2A is an arbitrary constant. This solution

involves generalized functions in two ways: explicitly in
the combination ao(x) + 0' (x), and implicitly through the
appearance of f"(x), the second derivative of a function
which may be discontinuous at x = O. As was observed in
Section 5.2, (44) implies the boundary condition

a f(O+) - f' (0+) = 0, (49)

so that if we use the notation

x < 0
f" (x) = { 0 (50)
r f" (x) , x > 0

we can write [see (9.24)]

f"(x) = f(O)[ao(x) + o'(x)] + f"(x). (51)

r

Hence we can replace f" (x) by f" (x) in (48) by adjusting

r
the value of B. In particular, the choice B = -f(O) is
the only one for which the solution is an ordinary function
rather than a generalized function. We leave it to the
reader to show by direct substitution that the constant B
§l8. The Wiener-Hopf technique 305

is indeed arbitrary if we allow the solution to be a gen-

eralized function.

Example 3: We continue to use the same integral operator to

illustrate the variety of phenomena which it may contain, and
consider the integro-differential equation

I/>"(x) + ! Joo e-a!x-y!lj>(y) dy = 0, x > O. (52)

£. 0

Proceeding with the Wiener-Hopf method, we obtain from this

a 2 <Xl+ (w)
_w 2 <Xl (w) - 1/>' (0) + iwlj>(O) + '¥ (w), (53)
+ w2 + a2

and factorization yields

w + ia
(w- ia) '¥ _ (w) (54)
E (w).

We may now examine, in retrospect, the conditions necessary

for the validity of our procedure. From the fact that
I/>(x) = 0 for x < 0, we obtain for w(x) the simple expres-
sion ~(x) = (constant) exp(ax), so that its traRsform con-
verges if Im(w) < a. This behavior is reflected in (54),
from which we see that ,¥_(w) has a pole at w ia. We
need an overlapping strip to ensure that E(w) is an entire
function, so the pole in '¥_(w) forces us to assume that
I/> (x) grows at a rate less than exp(ax) for large positive
x. Applying this restriction we conclude that E(w) is an
2
entire function which is of order at most w in the upper
half-plane, and bounded in the lower half-plane. Consequently

E(w) = A = constant, and we obtain

306 PART IV: SPECIAL TECHNIQUES

A(w+ia) + (w 2+a 2 )[Ij>'(0) - iwIj>(O)]

<I>+(w) = (55)
a 2 _ a 2w2 _ w4

The poles of <I>+(w) occur at the zeros of the denominator,

namely 2w2 = _a 2 ± 14a 2+a 4 . Three of these lie in the
region Im(w) ~ 0, and the other in the region Im(w) > a.
This latter pole clearly violates our original conditions on
<I>+(w). The way out of this difficulty is to choose A so
that the numerator of (55) is zero at the awkward point,
making <I>+(w) analytic there. Thus A is not an arbitrary
constant, but is determined by our assumption regarding the
rate of growth of Ij>(x). Inversion of <I>+(w) gives for
Ij>(x) a linear combination of three exponential functions de-
pending on two arbitrary constants, namely 1j>(0) and Ij>' (0).

Boundary Conditions: To investigate the significance of

these findings, we use the fact that the integral operator
in (52) is a Green's function to convert the problem into a
differential equation. Acting on the original equation with
the operator d 2 /dx 2 - a 2 yields the fourth-order equation

(56)

whose solution is
4
Hx) = L (57)
j =1
4 2 2
where the r·J are roots of r - a r If we im-

pose the condition that Ij>(x) grow more slowly than exp(ax) ,
one of the exponential functions is disallowed, and we re-
cover the solution found above, except that it appears to
depend on three arbitrary constants. In fact there is a
boundary condition implicit in the original integro-
§18. The Wiener-Hopf technique 307

differential equation, namely

cp(3) (0) = a CP"(O), (58)

and this reduces the number of independent constants to two.

Problems

1. Show that if the function W(x,O) of (6) has the be-

havior
w(x,O) - xll, x -+ 0,

where II > -1/2, then the entire function E(w) of (15)

is identically zero. Investigate the solution obtained
for the Sommerfeld diffraction problem under the weaker
assumption that II = -1/2.

2. By using a suitable free-space Green's function for the

Helmholtz equation in a half-plane, show that the solu-
tion of the Sommerfeld diffraction problem may be written

1
as
e
-ik(x cos 8+y sin 8) + e -ik(x cos 8-y sin 8)

~i J~oo H~l)(kR) h(~) d~,

l
- y > 0

Ii fO H(l) (kR)
2' _00 0
h(~) d~, y 5- 0,

R2= (x_~)2+y2,
where the unknown function h(~) is determined by the
integral equation
o
i f _00
H~l) (klx-~I) h(~) d~ = 2 e
-ikx cos 8

Solve these equations using the Wiener-Hopf technique.

308 PART IV: SPECIAL TECHNIQUES

3. Derive an asymptotic expansion for the function defined

by (24) by writing u = cosh t and deforming the con-
tour so as to employ Watson's lemma for loop integrals.

4. Show that the solution of the mixed boundary-value prob-

lem
0, _00 < x < 00, y > 0,

Hx,O) 0, x > 0,

g(x), x < 0,
is given by

~n J ~(w,O) e- iwx - lyl/w -k dw,

2 2
Hx,y)
C
where
3in/4 0
f fo
00

~(w,O) = e e iwu du ~-1/2 eik~ g(u-~) d~.

,!n(w-k) _00

5. If the boundary conditions in Problem 4 'are replaced by

cp(x,O) f(x) , x > 0,

~y(x,O) o x < 0,

then show that

6. Solve the mixed boundary-value problem

(V 2 k 2) ~ (x,y) 0, _00 < x < 00 , -00 < y < ,

00 •

iflx
<py (x, 0) e , x > 0,

~(x,y) ... 0, x2 + y2 ... 00.

§18. The Wiener-Hopf technique 309

7. Investigate the Sommerfeld diffraction problem when the

boundary condition on the screen is replaced by

~(x,O±) = ± i6 ~y(x,O±), x ~ 0,

and show that this leads to the Wiener-Hopf problem

'!'~(w,O) -(1 + i6y) ['!'~(w,O+) - '!'~(w,O-)]

+ 2i6yk sin 9/(w-k sin 9)

where ~ = ~ - ~inc' '1" (w,y) is the Fourier transform of

~y(x,y), and

8. Solve the integro-differential equation

~It(x) + a. 2Hx) + foo e -ix-yi ~(y) dy = 0, x > 0,

o
subject to
HO) 1,

~(x) + 0, x + 00.

9. Find the Green's function for the equations

00

+ fo e- ix - yi G(y,x') dy 6(x-x'), x > 0,

G(O,x') 0,

G(x,x') + 0, x + 00,

and show that it is related to the solution of Problem

8 by6
310 PART IV: SPECIAL TECHNIQUES

X,
G(x,x') -f0 ¢(s) ¢(s+x-x ') ds

foo ¢(s+x) ¢(s+x') ds + GO(x-x'),

o
where GO is the Green's function for the infinite
problem, i.e.,

{3 2/3X 2 + a 2 } GOCx-x')

+ foo
_ 00
e- Ix - yl G Cy-x') dy
0
= o(x-x'), _00 < x < 00
'

GoCx-x') ... 0, Ixl'" 00

10. Derive the factorization

7[1/2 C-iz) !/C-iZ-~)!,

K_ (z) = K+ (-z).

11. If ¢Cx,y) is determined hy

<y2¢ - ¢x = 0,
¢(x,y) ... 0, x 2 + y2 ... 00
,
-ax
¢(x,O) = e x > 0,
then show that
;r:;:a- e- i 7[/4 -I yl Iw 2 - iw
<I>(w,y) e
(a-iw) IW-T
while the inversion contour lies in the strip
o < 1m (w) < 1.
§18. The Wiener-Hopf technique 311

12. Consider the infinite strip -b ~ y ~ +b, _00 < x < 00,

along which a wave ~i = exp(ikx) is incident from

x = _00. The total wave field ~,which consists of the

incident traveling wave and waves diffracted by a semi-

infinite strip at y = 0, x ~ 0, satisfies the equations

~y(x,±b) = 0, _00 < x < 00,

~(x,O) = 0, ° < x < 00.

Find explicit expressions for ~(x,y).

13. Solve the previous problem with the boundary condition

on the strip replaced by ~y(x,o) = 0, ° < x < 00

Footnotes

1. The Wiener-Hopf technique is mentioned in a number of

books; for a comprehensive review of the method see
NOBLE (1958).

2. This problem may also be solved using the Kontorovich-

Lebedev transform: see Sections 17.4-5.

3. One form of Liouville's theorem is as follows: if E (z)

is an entire function, and i f

E(z) _ ~(zs), Iz I .... 00,

then E (z) is a polynomial of degree n, where n is an

integer less than or equal to Re(s). See AHLFORS (1966),

pp. 122-123.
312 PART IV: SPECIAL TECHNIQUES

4. See Section 10.4 for another example of this transforma-

tion, which is also discussed at some length in NOBLE
(1958), p. 31ff.

5. See NOBLE (1958), p. 93ff. for a list and some references.

In addition to problems in one complex variable, Kraut
has considered mixed boundary value problems which may be
resolved using a Wiener-Hopf type of decomposition in two
complex variables. See E. A. Kraut, Proc. Amer. Math.
Soc. (1969), ~, 24, and further references given there.

6. This relationship holds for a wide class of kernels, of

which exp(-Ix-yl) is the simplest. See G. A. Baraff,
J. Math. Phys. (1970), g, 1938.
§19. Methods based on Cauchy integrals 313

§l9. METHODS BASE]) ON CAUCHY INTEGRALS

19.1. Wiener-Hopf Decomposition by Contour Integration

The major difficulty in using the Wiener-Hopf tech-
nique is the problem of constructing a suitable factoriza-
tion. We consider here a method based on contour integra-
tion which leads by natural extensions to the use of Cauchy
integrals in the solution of mixed boundary-value problems.
Suppose then that the function ~(z) is analytic in the
strip a < Im(z) < 8, and that we wish to find functions
~+(z) and ~_(z) analytic in the half planes Im(z) > a
and Im(z) < 8 respectively, such that

~ + (z) + ~ _ (z), a < 1m (z ) < 8. (1)

First choose z to lie inside the contour shown in Figure 1;

then Cauchy's integral formula gives

~(z) = ~
27f1
I !iU
I;;-Z
dl;;. (2)

4~

f'i6
ip
- C,

C3 ,~ C4
Re (~)
-L L

l/ iY
~ i Q -
... C2

Figure 19.1
314 PART IV: SPECIAL TECHNIQUES

We consider only functions for which the integrals along C3

and C4 become zero as L ~ 00 Taking this limit, we obtain
the result

<I> Cz) 1
2ni
fiO +OO
<I> (1;)
---- d~. (3)
io- oo ~-z

The first integral defines a function which is analytic for

ImCz) > a [since for any such z we can always choose y
by a < y<Re(z)], and the second integral defines a function
which is analytic for ImCz) < 8, so that (3) gives the de-
sired decomposition.
Equation (3) gives an additive splitting; however,
the same trick may suffice to give a multiplicative split-
ting, that is, a factorization of the type
K+ (z)
<I>(z) = -- (4 )
K_ (z)

If K+(z)rO for Im(z»a and K_(z)rO for Im(z)<8, then

we can reduce this problem to the previous one by either of
two methods. The first is to take logarithms, writing

9.n <I> (z) = 9.n K+ (z) - 9.n K_ (z). (5)

Alternatively, we may differentiate to write

(z) K~ (z)
<1>'
~~
----- - (6 )
K+ Cz)

Both approaches depend on K+(z) and K_(z) not having

.
zeros, and both have been used extensively 1n th '
e l1terature. 1

Milne's Equation: On taking the Fourier transform of equa-

tion (18.32) we are led to the problem of factoring

[arc~an w _ 1] •. (w) • ,_ (w). (7)

§19. Hethods based on Cauchy integrals 315

-1
The function (w arctan w)-l has branch points at w = ±i,
so the strip in which we must operate is -1 < Im(w) < 1.
It is not difficult to show that in this strip the only zero
of this function is a double root at w = O. We must remove
this zero in order to apply a contour integral method; there-
fore we consider the problem of finding the factorization

rw 2+l1 r arctan w _ 11 = G+(w)

l w2 H w J G_ (W) •
(8 )

This is suitable for use in the contour integral method since

the logarithm of the function tends to zero as Iwl + 00 in
the strip -1 < Im(w) < 1, and the double zero has been re-
moved. Assuming that tn G+(w) and tn G_(w) have been
found by contour integration, (7) now admits the factoriza-
tion

(W~iJ G+(w) ~+(w) = (W~ij G_(w) 'P_(w), (9)

and the functions ~+(w) and 'P_(w) may be found by the

usual arguments. For details and numerical computations the
reader is referred to the literature. 2

19.2. Cauchy Integrals 3

Equation (3) leads us to consider functions of the
type
F (z) = 2;i fC ~~~) dr;, (10)

which are usually known as Cauchy integrals. We restrict

ourselves to contours C which are piecewise smooth, and
4
functions fer;) which satisfy the Holder condition wherever
the contour is smooth. Restrictions on fer;) near a corner
or an end of C will be specified below. It is a standard
result that, with the restrictions applied to (10), the
316 PART IV: SPECIAL TECHNIQUES

integral defines a function Fez) which is analytic in the

entire complex plane excluding C. Of particular interest is
the value which Fez) approaches as z approaches an arbit-
rary point of C. Since the definition of C includes a
specification of the direction in which the contour is to be
traversed, we may define the meaning of the left-hand and
right-hand side of the contour. It is conventional to refer
to the left-hand side as positive and the right-hand side as
negative, and to define the functions

F+e~) = lim F(z), (11)

- z.... ~
where the point ~ on C is approached from the side which
is indicated by the suffix. We also define the principal
value by

F (r,;)
P = lim
E .... O
1
2 .
TIl
I
C'
feu) du
u-'"
~
(12)

where C' is that part of C satisfying lu-~I ~ E.

Plemelj Formulas: We now derive some very important results

in the case that f(~) is an analytic function. These re-
sults are in fact true for functions f(~) which satisfy a
Holder condition, although the proofs are more intricate in
that case. 5 Consider first a point ~ on C which is not
an end point or a corner; then, by some elementary considera-
tions from the theory of contour integrals, we have
§19. Methods based on Cauchy integrals 317

1 f (u) du
F+(r;) 21T i
Ic'+c 1 u-1; ,

F _ (1;) 1
hi
Jc'+c feu) du
u-1; ,
2 (13 )
-1 feu) du
f (r;)
1Ti Ic 1 u-1;

1 :ti!!l d u,
7f1 Ic 2 u-1;
where the various contours are shown in Figure 2. On taking
the limit e: ... 0, we recover the Plemelj formulas

Figure 19.2

1 (14)
F- (1;) = Fp (1;) - -2 f(1;).

Near a corner of included angle a (see Figure 3), we have

318 PART IV: SPECIAL TECHNIQUES

1 feu) du
f (I;;)
ai JC u-I;;
1
(15 )

1
(21T-a)i
JC ~ du,
u-I;;
2
and hence the P1eme1j formulas are changed to

(16)

In either case, subtraction of one result from the other

leads to the important common result that

f (I;;) , (17)

so that the Cauchy integral gives a construction for an ana1y-

tic function having a prescribed discontinuity across C.

Figure 19.3
§19. Methods based on Cauchy integrals 319

Behavior Near End Points: For simplicity we will consider a

smooth arc C with end points a and b. We assume that
f(s) satisfies a Holder condition at every interior point of
C, and examine the behavior of F(z) near z = a. If f(s)
satisfies the Holder condition at z = a, then we can write
(10) as

1 Ib f(s)-f(a)+f{a) ds
F (z)
2Tf ia s-z
(18 )
1 z-b + Q(z),
f (a) tn (z-
hi a)
where
Q(z) =
1
2Tf i
ta
f(S)-f(a) d s .
s-z
(19)

Since the integral defining Q(z) converges when we set

z = a, the singularity in F(z) near z = a is of the form

:.f..(tl
F(z) ~ 2Tfi tn (z-a). (20)

Similarly, if f(s) satisfies the Holder condition at

s = b, F(z) has a singularity of the form

(21)

near that end point.

More generally, we may consider functions f(s) which
are not finite at the end points, but for which there are
constants Ya and Yb satisfying the inequalities

o < Re(y a ) < 1,

(22)
o < Re(Yb) < 1,

so that the functions (s-a)Y a f(s) and (s-b)Yb f(s)

satisfy the Holder condition at the end points a and b

respectively. Then by a subtraction procedure similar to
320 PART IV: SPECIAL TECHNIQUES

that employed in equations 6 (18) and (19), we may show that

the singularities of F(z) near the end points take the
form
i'rry
F (z) 'V _ _ e~_a--J;<p..::.(_a:...)___y_' (23)
2i sin (7rYa) (z-a) a

y
<I> (a) lim (I;-a) a f(l;)
1;"" a

near z = a, and

F(z) 'V -
y , (24)
2i sin (7rYb) (z-b) b

<I> (b)

near z = b.
19.3. The Discontinuity Theorem
Cauchy integrals represent analytic functions which
are discontinuous across an arc (or collection of arcs) C.
The discontinuity theorem states that if C is a contour
and f(l;) a function defined on C, both satisfying the
restrictions introduced in Section 19.2, then the only func-
tions ~(z) satisfying the three conditions

(i) ~(z) analytic in the entire complex plane excluding

C,
(ii) ~+(I;) - ~_(I;) = f(I;), (25)

(iii) for any end point or corner,

1~(z)1 < Alz-cI S , S > -1 (26)

are of the form

§19. Methods based on Cauchy integrals 321

<I> (z) 1
2'1fi
JC f(l:;)
I:;-z
dl:; + E(z), (27)

where E(z) is an entire function. The proof consists of

using the properties of Cauchy integrals near end points (re-
membering that a corner is the adjunction of two end points)
to show that the function <J>(z) - F(z) has no singularities,
and is therefore entire. Details may be found in the liter-
ature; Muskhelishvili's book is a particularly comprehensive
reference.

19.4. The Riemann-Hilbert Problem

The discontinuity theorem allows us to solve problems
concerning analytic functions which have their properties
specified on a contour C. The Riemann-Hilbert problem is
to find a function <I>(z) , analytic in the entire plane ex-
cluding C, which satisfies the condition

(28)

on C. To solve this problem, we first set f(s) to zero,

and find functions K+(z) and K_(z), with the property

(29)

Leaving aside the construction of such a pair of functions

until later, we see that (29) reduces the Riemann-Hilbert

problem to
<I> _ (I:;)
I:; on C. (30)
L(I:;)

The discontinuity theorem gives a solution immediately, viz.

322 PART IV: SPECIAL TECHNIQUES

<I>

K(z)
(z) 1
2TIi JC (I; - fz)(I;)K+ (I;) d I; + E (z ) . (31)

Construction of K+ and K: Taking the logarithm of equa-

tion (29) we obtain

(32)

and the discontinuity theorem will again provide a solution,

provided we can handle any singularities near end points in

such a way that the function f(I;)/K+(I;) satisfies the

necessary restrictions for the application of the discon-

tinuity theorem to equation (30). For applications there

are two important cases which we now consider.

(i) C is not a closed contour, and does not cross itself.

Then we write

tn L(z) = 2;1' J tn gel;) d~, (33)

C ~-z

and using the properties of Cauchy integrals near an end

point, we find that near the beginning (z=a) of C the
function L(z) behaves like
Ya
L (z) '" (z-a) ,
(34 )
tn g (a)
Ya = - 2TI i

Now choose an integer ka so that 0 ~ Re(Y a + k a ) < 1, and

k
multiply L(z) by the factor (z - a) a. The singularity
of this new function at z = a is sufficiently weak to allow
its use in equation (31). The other end point is treated
similarly, so that the factorization (29) is given by
§19. Methods based on Cauchy integrals 323

ka kb
K(z) = (z-a) (z-b) L(z). (35 )

Because the constants ka and kb are integers, K(z) sat-

isfies equation (32) whenever L(z) does.

(ii) C is a closed contour which does not cross itself.

Then the function tn g(s) will increase by 2nin in one
circuit of the contour, so we introduce a new function goes)
by
(36)

where a is an interior point of the region bounded by C.

It is easy to check that the desired decomposition is given

in terms of goes) by

(37)

tn K (z) n tn(z-a) +
1
2ni
fC tn s-zgoes) ds .

19.5. Simple Applications

(i) If C is a smooth arc joining the points z = a and
z = b, with the Riemann-Hilbert problem given as

(38 )

then we first evaluate the integral

1 rb in
tn[L(z)] = 2ni J s-z ds , (39)
a
giving
( 40)

The behavior of L(z) at z =b is inappropriate for the

324 PART IV: SPECIAL TECHNIQUES

next step; dividing by the factor (z-b) yields the cor-

rect factorization

K(z) = (z-a) -1/2 (z-b) -1/ 2 . (41)

Thus the solution of (38) is

li.U
K(z) - 1 Jb f(1;) dr,; + E(z). ( 42)
2ni a K+(r,;) (r,;-z)
For example, if a = -I, b = 1, and f(r,;) 1, then

~(z) = 1 z + ~ (43)
2" - 2/z2-l 1z2_ l
(ii) The Wiener-Hopf problem. The Riemann-Hilbert problem
given in equation (18.26) is

A(w) ~+(w) + B(w) ~ (w) + C(w) D,

(44 )
(l < Im(w) < 8.

Now we no longer need a strip of overlap; indeed on choosing

a constant such that (l < Y < 8, we can take the contour
y
7
C as the line Im(z) = y. Application of the method of
Cauchy integrals leads immediately to the formulas of Section
19.1.

19.6. Problems in Linear Transport Theory8

Fourier transform methods have been extensively ap-
. 9
plied to the linear transport equatIon

[1 + ;t + ~.~) $(!",~,t) ( 45)

=c I
Here $, the phase-space density, is to be determined in a

region V with the boundary condition on the surface S

§19. Methods based on Cauchy integrals 325

given by
y inwards. (46 )

The scattering kernel cr and source function q are, like

l/Is' presumed to be given. We shall consider in this section
half-plane problems for which V is the region z > 0, and
show that whenever the function cr has the separable form

cr(y,y') = fey) g(y'), (47)

the Fourier transform method may be used in conjunction with

Cauchy integrals to give a general method of attack for the
analytic solution of the problem. We confine our attention
to time independent solutions and drop the reference to t

in what follows.
The co11ision1ess Green's function (see Problem 11.16
for the various properties used here)

GCr-r' ,v) = __
1 ___ f e -ik'Cr-r')
- - - d 3k (48)
- - - (2TI)3 1 + i~'~
can be used to rewrite the basic equations as an integral
10
equation, namely

fz>O G(!'-!,~) [c pC!' ,~) + q(!' ,~)] d 3r'

( 49)
+ Vz f z=O G(rs-r,v) l/Is(:s'~) d 2r ,
- - - _s
where
p(:,~) = f crC~,~') l/I(:,~') d 3v', (SO)

and the subscript s refers to the restriction to the bound-

ary z = O. We now take the three-dimensional Fourier trans-
form so as to make use of the convolution property. In so
doing, we shall have to add to the left-hand side of the
326 PART IV: SPECIAL TECHNIQUES

equation an unknown function which is given by the right-

hand side for z < O. Without any fear of confusion we de-
note this function by w(r,Y), introducing the conventions

(51)

After these manipulations, we obtain the integral equation

c P+(~,~) + Q+(~,~)

1 - ik·v
(52)

+
Vz J ik' r 2
e - -ws(r,Y) d r.
1 - ik'v z=O
Solution for Separable Kernels: When cr has the form (47),
we can introduce the integrated density

(53)

with Fourier transform p(~), and by multiplying by g and

integrating over v immediately get

A (1)) p+ (15) + P_ (15) = B(15) , (54)

where
A (15) = 1 - c
f f(~) g(~)
d 3y (55)
1 - ik'v

is the dispersion function, and

BCk) = J
g (v)
-
[<Q (k,v) + Vz - -~ (r,v)d 2~
J oe ik·r r d 3v
- 1 - ik'v + - - z= s - - - -
(56)

represents the known contributions from the sources q and

the boundary function ws'

§19. Methods based on Cauchy integrals 327

For convenience we denote the z-component (normal

component) of ~ by k; dependence on the transverse com-
ponents kx' ky will not be explicitly mentioned. It is
evident from the above formulas that the real k-axis will
separate regions of differing analytic behavior, and that
equation (54) is a standard Riemann-Hilbert problem in the
complex variable k. For suitably behaved functions BCk),
the solution of (54) will be given by the Cauchy integral
B (k') !I. _ (k' )
dk' , (57)
k' - k

where the functions 1I.+(k) and II._Ck) factor the disper-

sion function lI.(k) according to
11.+ (k)
!I. (k) = !I. (k) , k real. (58)

One-Speed Isotropic Scattering,~: In the particularly

simple case that all the particles travel at the same speed
and that collisions result in a random directional change, we
can replace the velocity variable y (which is three-
dimensional) by a direction variable, ~, which is a unit
vector. The dispersion function becomes simply
d~
!I. (k) = I - ~ J
4'lT I - ik'rl
(59)
I + ic ~n I + i~ I,
Zlk2+K2 I - i;k4'i(21
where k2
y'
The analytic structure of this func-
tion for c < I is indicated in Figure 4 ,. there are two
branch points at k +"
_1 h + K2 , and two zeros 11 at

k = ±iK O' where KO < ~.

328 PART IV: SPECIAL TECHNIQUES

branch cut

i~

-i Ko

-i~

c
branch cut

Figure 19.4

It is clear that the function

(60)

is analytic in the entire complex plane with the two branch

cuts of Figure 4, and that it can be additive1y decomposed
according to

(61)

Now we deform the contours to C± of Figure 4, so that the

factorization of the dispersion function can be written as
k ± iKO
A±(k) X+ (k) ,
k ± i/1+K2 -

X±(k) exp !t 27Ti1 fc_ ~~~~) dkll.

J
(62)
1 +
§19. Methods based on Cauchy integrals 329

This can now be substituted into equation (57) to solve

particular problems.

19.7. The Albedo Problem

The Albedo problem is the problem of obtaining the
(neutron) phase-space density in a source-free half-space
z ~ 0, if a parallel beam is incident on the surface z = O.
We consider a beam which is uniform in the x and y
variables,12 which means that the solutions will be indepen-
dent of these variables, and we can set kx and ky to zero.
Then we have

n inward, (63)

where go is the direction of the incident beam, with

nOz> O. From this, using equations (56) and (57), and
writing n ~, nOz ~O' it follows that
z

B(k) i (64)
k + (i/~O)
and
iA_(-i/~O)
p+(k) (65)
[k+(i/~O)]A+(k)

Substitution of this result into the inverse Fourier trans-

form gives the result that p(z) = 0 for z < 0, while for
z ~ 0, we can deform the contour into the lower half-plane,
where the integrand has a simple pole and a branch cut, to

obtain
330 PART IV: SPECIAL TECHNIQUES

p (z)
-1 foo dk
21Tl _00
[k+ (i/]10) ]A+ (k)
[(1/]10)+K O] (l-K O) X_ (-i/]10) -K Z
e 0 ( 66)
[(1/]10)-K O] [1+(1/]10)] X+(-iK O)

___1__ [(1/]10)+K o]X o(-i/]10)J e-ikz(k+i) dk.

21Ti [(1/]10)+1] C [k+ (i/]10)] (k+iKO)X+ (k)

Connection with Singular Eigenfunction Expansion: We have

chosen to write the Fourier inversion of p(z) in some de-
tail so that we can show, in a reasonably simple situation,
the connection with the solution by singular eigenfunctions.
In the notation of Case and Zweifel, the expression for
p (z) is
-z/v
p (z) =- e 0
cvO X(v O)
(67)
2 (v O-]10) Y (1l0) {_~_ _l _ _ _ } -x/v
A(v) e
cv(vO-v) X±(v) A+(v)

where we have added a circumflex to their functions A (v)

and XCv) to distinguish them from the functions appearing

above. The eigenfunctions in (67) are

<Pv (]1) ~v P V~]1 + A(V) O(V-]1), (68)

where the P refers to principal value integration. The

meaning of that expansion is most easily expressed in terms
of Cauchy integrals using
l
A

1 A+(v) A(v)
fo A(v) <Pv(]1) dv = dVJ+
(69)
A_ (v) A(v) l
dVJ _ '
§19. Methods based on Cauchy integrals 331

where ~ is the complex variable to which the subscripts ±

apply. With the help of these formulas, we can write p(z)

as
2y(~0)
p (z)
cvO X(v O)
(70)
+ !VO-~O) Y(~O)J 2 e- z / v
A dv,
2ni C cv(vo-v)(v-~O) XCv)
where the contour is shown in Figure 5.

1m (v)

c
Re(v)

branch cut

Figure 19.5

Under the variable change y -ilk, it is easily shown that

we have the correspondences
A

A(v) A(k),
A

X (v) (71)

so that the integral in equation (70) becomes

332 PART IV: SPECIAL TECHNIQUES

-ikz
1 2kO e
dk (72)
2rri Ic
cll (k+iK 0) [k+ (i/ll O)] X+ (k)
0
with the contour of Figure 4. Thus the correspondence is es-
tablished, although it may be seen, as pointed out by Case
and Hazeltine,S that the Fourier transform method provides
a more direct approach and avoids the difficulties inherent
in the interpretation of singular eigenfunction methods as
explicit formulas suitable for direct computation.

Angular Density of Emergent Particles: Using (65) in (52),

we obtain the result
i I) (g-gO)
1/J (k ,Q)
k+ (i/llO)
(73)
c A_(-i/ll O)
- 4rr
II A+ (k) [k+ (i/llO)] [k+ (i/ll)]
This function is analytic in the upper half k-plane except
for a pole at k = -i/ll when II < O. Hence, for z = 0,
the inverse Fourier transform involves only a simple residue
calculation, and the result of this calculation is an expli-
cit expression for the angular density of particles which
escape from the surface. This expression is

1/J(x,y,O,Q) (74)

19.8. A Diffraction Problem l3

Consider the diffraction of scalar waves in an
infinite two-dimensional region by a strip of finite width.
The differential equation is (V 2+K2)¢ = 0, and there are
boundary conditions on the strip, which we take to be y 0,
Ixl ~ a, plus the additional condition that the solution con-
§19. Methods based on Cauchy integrals 333

sists of incoming plane waves, ¢i' and outgoing diffracted

waves. The details of the problem will depend on the parti-

cular boundary condition applied on the strip; here we take

¢ = O. Denoting by G(r,r') the free-space two-dimensional
Green's function and applying Green's theorem to a region
whose boundaries are a large circle centered at the origin,

and the two sides of the strip, we readily deduce the for-
mula

¢Cx,y) - ¢i(x,y)
1
-4
(a G(x,y,x',O) p(x') dx', (75)
'IT J-a
where
(76)

We recall 14 that the Green's function may be written

as
-ik(y-y') - lx-x' I ~
G(x,y,x' ,y') e dk, (77)
,{~
where we take as the branch cuts the contours Cl and C2
of Figure 6. Setting y = 0, Ixl < a in (75), we obtain

the integral equation

-4'IT ¢i(x,O) = fa K(x-x') p(X') dx',

-a
(78)
K(x-x')
r -co

The Function K: By deforming contours around the branch

cuts and then shrinking them onto these cuts, we may deduce
the representations
334 PART IV: SPECIAL TECHNIQUES

1m (k)

-K Re (k)

C1

Figure 19.6

-ik(x-x')
K(x-x' ) e 6(k) dk,

Ll (k) (79)

where, for convergence, C1 is the contour if x < x' and

Cz is the contour if x > x'. From this representation of
K(x-x'), the following important property emerges:

fa-a K(x-x') ok' ,

e- 1 x dx'

-ikx-ia(k-k') -ik'x'
e - e dk (80)
-i(k'-k)

e -ik'x - e ia(k-k')
+ J Ll(k) dk.
Cz i(k'-k)
§19. Methods based on Cauchy integrals 335

We look for a solution of equation (78) having the form

-ik.x -ikx
J
p (x) L AoJ e + J <1>1 (k) e dk
j Cl
(81)

+ ICz <l>z(k) e- ikx dk,

which can be substituted into the integral equation to con-

vert it to

I
-ik;X -ikx+ia(k-ko)]
= ~ Aj { tJ. (k) [ e "' 0 - e J dk
J Cl 1(k j -k)

+ I Cz
tJ.(k) ~e
-ikX-ia(k-ko)

i(kj-k)
J - e
-ikoXj
J dk
}

+ {J Cl <l>l(k') dk' + IzC

<l>z(k') dk'}

·{leI
re-ik'x _ e-ikx+ia(k-k'~
tJ.(k) t i (k' - k)
dk

+ ICz tJ. (k) le-ik'x-ia(k-k') - e-ik'Xl dk}.

[ i(k'-k) J (8Z)

Reduction Using Cauchy Integrals: None of the integrals in

this expression is divergent, since the expressions in square
braces are finite when their denominators are zero. What we
do here is regard all of the integrals as principal value
integrals, which allows us to separate the various terms
which are presently grouped together to effect these can-
cellations. In doing this we need the following Plemelj
formula:

Ic z
tJ. (k')
i (k' - k)
dk' - I tJ. (k')
1(k'-k) dk
,
Cl

=~ Zn n(k), k not on CI or Cz
(83)
l 0 k on CI or CZ'
336 PART IV: SPECIAL TECHNIQUES

With the aid of this formula, the original integral equation

is reduced to
-ik.x
l: Z1T A. n (k J. ) e J
j J
-iak·
,ik(a-x) {- l A. e
J
J

4> (k')e- iak ' - iak'

+
Ic
1 dk' +
Ic
4> Z (k' ) e
dk'~ dk
1
i (k' - k)
Z i (k' - k) j

-I 11 (k) e
-ik(a+x) .1 - L
Aj e
iak·
J

Cz l i(k-k.)
J

e
iak'
dk' +
4>2 (k' ) e
iak'
I (84 )
dk'J dk,
i(k'-k) i(k'-k)

For the present problem, with incoming plane waves, 4>. (x;O)
1

is simply one exponential term, and a solution of this last

equation is achieved by simultaneously solving the three
equations

(85)

- iak' -iak l -iak'

4>l(k')e A e 4> Z(k' ) e
dk' dk' , (86)
fC l i(k'-k) i(J<-k l ) fez i(k'-k)

k on Cl '
iak' iak l iak'
<pz(k')e A e 4>1 (k') e
dk' - dk' , (87)
Ic z i(k'-k) iCk-k l ) Ic 1 i(k'-k)

k on CZ'

Reduction to Coupled Equations of Fredholm Type: lS No ex-

plicit solution of these equations is known; however, we may
§19. Methods based on Cauchy integrals 337

easily reduce them to a much simpler form not involving any

singular integrals. In the following we assume normal inci-
dence, so that kl = 0, and the algebra is somewhat simpli-
fied. First introduce the functions

-l/Z
Xl (k) (k+K) ,
(88)
XZ(k) = (k_K)-l/Z,

which have the properties

Xl + (k) + Xl _ (k) 0 on Cl'

(89)
XZ+ (k) + XZ_(k) = 0 on CZ '
and the Cauchy integrals

I
4>1 (k)
4>l(z) = ~dk,
Z!i C
1 (90)

4>Z(z) = Z!i IC2

4>z(k) dk
-y:z-- .

Then the Plemelj formulas enable us to express the functions

tP1Ck) and tPZCk) in terms of these Cauchy integrals as

(91)
on CZ '

The equation for tPl(k) now appears as

where
-ik'a
2iK 4>z (k') e
l/I (k) - k- dk' , (93)
k' -k
and the solution of this Riemann-Hilbert problem is
338 PART IV: SPECIAL TECHNIQUES

1 ( Xl + (k) - Xl _ (k)
<PI Cz) Xl (z) = Z'ITi J -';;"'--:-k---z-~-1/J (k) dk. (94)
Cl
Introducing the expression for 1/J(k) into (94), the
first term of 1/J(k) leads to the integral

k (k- z)

(95)

-ZiK
= -z- [Xl (z) - Xl (0)],

which therefore gives to <Pl(z) the contribution

-ZiK (
-z 11 - Xl (0) ')
Xl (z) f' (96 )

and to ~l(k) the contribution

iK {Xl (0) Xl (0) } ZiK Xl(O)

Z- Xl_(z) - Xl+(z) -z-~) (97)

The second term is dealt with using partial fractions in a

similar manner, so that the singular integral equations (86)

and (87) for the unknown functions ~l(k) and ~Z(k) are
reduced to the ordinary integral equations
-ik'a
e
dk' ,

(98)
ik'a
e
i'ITXl_(k) ~Z(k) - ZiK X (0)
k Z
+ JC ~l(k')XZ(k')
dk' ,
k' -k
1
k on CZ'
We refer the reader to the literature, and particularly to
the work of Wolfe,16 for the solution of these equations by
iteration and the relation of these solutions to other
§19. Methods based on Cauchy integrals 339

methods of treating this particular problem of diffraction.

Problems

1. Use Cauchy integrals to derive the factorization of

(18.14).

2. Investigate the various solutions of (18.52) which may be

obtained by choosing to take the Fourier transform along
different contours, subsequently using this contour in
the relevant Cauchy integrals.

3. Find an additive decomposition of the function

1
<P (z)

by deforming the contours of the Cauchy integrals to turn

them into loop integrals.

4. Find an additive decomposition of the function

<P (z) = 1z2+k 2 •

5. Use Cauchy integrals to derive the decompositions needed

in the solution of Problem 18.6.

Footnotes

1. See NOBLE (1958), p. 93ff. for details.

2. C. Mark, Phys. Rev. (1947), 'lJ.., 558; G. Placzek, Phys.

Rev. (1947), '!.l, S56.

3. MUSKHELISHVILI (1953).
340 PART IV: SPECIAL TECHNIQUES

4. The function f(s) satisfies a Holder condition on a

contour C if there exist real positive constants A
and V such that

for any two points sl and Sz on C.

5. MUSKHELISHVILI (1953), Ch. Z.

6. MUSKHELISHVILI (1953), Ch. 4.

7. We could in fact use a Fourier inversion contour which

is not a straight line parallel to the real axis, thus
achieving a generalization of the Wiener-Hopf technique
by using the P1eme1j formula.

8. Based on work by K. M. Case and R. D. Hazeltine, J. Math.

Phys. (1971), 11., 1970.

9. See CASE & ZWEIFEL (1967) for the derivation and inter-
pretation of this equation.

10. In particular, we have used the property that

11. Proved in CASE & ZWEIFEL (1967), p. 62 ff.

12. See ref. 8 for another example.

13. Based on a paper by K. M. Case, Rev. Mod. Phys. (1964),

~, 669.

14. See Section 10.4.

§19. Methods based on Cauchy integrals 341

15. The techniques used for the solution of these singular

integral equations are quite standard; see MUSKHELISHVILI
(1953).

16. P. Wolfe, SIAM J. Appl. Math. (1972), ~, 118.

342 PART IV: SPECIAL TECHNIQUES

§20. LAPLACE'S METHOD FOR ORDINARY DIFFERENTIAL EQUATIONS

20.1. Integral Transform Solutions

Transform methods are useful in finding solutions of

ordinary differential equations far more complicated than
those considered in Section 3. In fact, we have already seen
in Section 3.4 that an explicit formula for the Bessel func-
tion JO(x), defined as the solution of an ordinary differen-
tial equation with variable coefficients, can be found with
the Laplace transform. One advantage of the technique
developed in this section over the simpler method for solu-
tion in terms of a power series expansion is that the trans-
form method gives the solution required directly as an inte-
gral representation. In this compact form various properties
of and relations between different solutions of an equation
become quite clear, convenient asymptotic expansions can be
obtained directly, and numerical computation may be facili-
tated. For applications, the analytic properties, asymptotic
expansions, and ease of computation of a function are of
primary interest.

Laplace's Method: We consider the differential equation

whose coefficients are linear polynomials. Laplace's method

consists of assuming that the solutions have the integral
representation

y(x) = JC S(p) e Px dp, (2)

where the contour is to be chosen as part of the process of

§20. Laplace's method 343

solution. Equation (2) therefore represents a straightfor-

ward extension of the Laplace transform method, for which
the contour is a vertical line in the p-plane. If we in-
sert (2) into (1), and integrate by parts to remove the
terms in x, we get

o= x G(p)] S(p) e PX dp
Ie [F(p) +

px r d l
Ie e ~(p) S(p) Up G(p) S (p)J dp (3)

+ [G(p) S(p) e px]2l'

where
k
FCp) ~ akP ,
k=O
n k (4)
G(p) I
k=O
bkP .

The second term in (3) is evaluated at the end points of the

contour, which are often at infinity. If we choose the con-

tour so as to make this contribution vanish, then (2) will

represent a solution to (1) provided the function S(p) sat-
isfies the differential equation

d
F(p) S(p) - dp G(p) S(p) = o. (5)

There is an immediate formal solution to (5), namely

S(p) -- _ A exp
G(p) [I E.ill
G(p) dp ] , (6)

where A is an arbitrary constant. Using (6), it is possible

to determine suitable contours to complete the solution.

The simplification which allows us to write the expli-

cit formula (6) arises from the fact that (5) is a first-

order differential equation. This, in turn, is due to our

requirement that the coefficients of the original equation
344 PART IV: SPECIAL TECHNIQUES

. 1
be only linear polynomlals. If they were quadratic, the

process of integrating by parts in (3) would lead to a second-

order equation for S(p). There is no general method for

solving for S(p) in this case, although if a solution can

be found the method will still succeed.

20.2. Hermite Polynomials

As an illustration of Laplace's method, we consider
Hermite's differential equation

w" (x) - 2x w' (x) + 2vw(x) 0, (7)

where v is an arbitrary constant. We obtain solutions to

this equation by using equations (2)-(6), which give

1
S(p) 2p exp

(8)

and this can be substituted into (2) to obtain an integral

representation. In order to conform with established conven-
tion, we first introduce a new variable s = p/2, so that
2
w(x) =A fC e-s +2sx
ds, (9)

where A is an arbitrary constant, and the path of integra-

r
tion must be chosen so that

e-s2+2Sxl2
0, (10)
L sv+l J1
where [~]2 refers to the increment in ~ as we traverse
1
the contour C.
§20. Laplace's method 345

Choice of Path: In general a variety of paths will satisfy

the condition that the integrated term vanish, and each path
leads to a different solution. The particular path of in-
terest is usually dictated by the physical or boundary condi-
tions we impose. Possible paths which satisfy (10) are

(i) -00 < s < +00 (avoiding s = 0),

(ii) 0 < s < 00 (Re(v) < 0),

(iii) A contour which encircles the origin (van
integer),
(iv) The path shown in Figure 1.

1m (5)

Re (5)

_v_1
branch cut of 5

Figure 20.1

We will prove in Section 20.3 that Hermite functions

2
have the behavior w(x) '" A exp(x ) either for x-+-oo or
x -+- -00, or both, except when v = 0,1,2, ... [This conclu-

sion may also be reached using Watson's lemma for the paths
(i), (ii) and (iv).J
346 PART IV: SPECIAL TECHNIQUES

Turning first to the bounded functions, for which v

must be an integer, we take C to be path (iii) , and write
e-s2+2sx
wn(x) = A
Ic n+l ds, n = 0, 1, 2, .... (11)
s
The solutions (11), which will turn out to be polynomials,
n
are normalized by the convention that the coefficient of x

is 2n; this gives

-s2+2sx
'
Hn(X) = n2~i JC e sn+l ds, (12)

where we use the symbol Hn as the (standard) notation for a

Hermite polynomial.

Derivative Form: Another expression for Hn(x) follows if

we use the change of variables u = s-x in (12) and apply

the standard formula for evaluating the residue at a pole of

order (n+l). This gives

(13)

Explicit Formula: From (12), we have

2
e -s +2sx} ,
n
H (x) = n!{coefficient of s in (14 )
n

giving the explicit formula

[nf 2 ] (-l)kn ! (2x)n-2k,

Hn(x) = (15)
k=O k!(n-2k)!

where [n/2] is the largest integer .:. n/2.

Generating Function: It follows immediately from (14) that

the function exp(2xt-t 2 ) generates the Hermite polynomials

in the sense that

§20. Laplace's method 347

(16)

This formula plays an important role in the theory of Hermite

. 2
polynomials.

20.3. Hermite Functions

We return to the problem of finding solutions of
Hermite's equation for arbitrary v. Consider the integral

JC e-p2+2PX
pV+l
dp, (17)

which is the Laplace integral solution to (7), with p re-

placed by 2p. We take the contour shown in Figure 1, so
that for v = 0,1,2, ... (17) gives Hn(x). When v = -1,
-2, ... , (17) is indeterminate. However, for Re(v) < 0 we
can obtain an alternative representation by 'shrinking' the
contour around the branch cut; with the change of variables
p =t exp(± in), this gives
-t 2-2tx
_-v~!~s~i~n~(_n_v~) foo
~e~~___ dt
n 0 tV+l
(18)

-(---V~~-l-)-! J: ~ -t 2-2tx
t V+l dt, Re(v) < O.

There are two important immediate consequences. The first is

that Hv(x) is an entire function of both x and v; the
second is that Hv(O) can be evaluated. For, on putting

x = 0 in (18) , we have
(-I-I)!
Hv(O) = (19)
2(-v-l)!
which holds for all v by analytic continuation.
348 PART IV: SPECIAL TECHNIQUES

Recurrence Relations: If we multiply (17) by 2x and inte-

grate by parts, we find that
r -s2 1
2xH (x) = -v!
v 21Tl Ic e
+2sx d
as L- e
sV+1 JdS

-v!
-s2+2sx r 1
(20)

t
e
21Ti Ic sv+1 2S - ~Jd
s s ,.

and after using (17) again we see

(21)

This recurrence relation (called a three-term relation be-

cause it connects Hermite polynomials of three different in-
dices) can be used to tabulate the Hermite polynomials step
by step, beginning with HO(x) = 1.
Another recurrence relation may be obtained by dif-
ferentiating (17) with respect to x. This gives

V!
---
21Ti
IC 2s e - s + sx ds
sv+1
2
2

(22)
= 2v Hv - 1 (x) •

The manipulations involved here are only valid for

Re(v) > -1, but the results are not restricted, because we
can employ the principle of analytic continuation to extend
them to all v.

Independent Solutions: Hv(x) is one solution of Hermite's

equation, and it is trivial to show that Hv(-x) is also
a solution. The crucial question is whether this second
solution is linearly independent of the first, and this de-
pends on the Wronskian. For any two solutions of (7), we
3
have
§20. Laplace's method 349

2
W[u,w] = Ke x (23)

In the present case, we find the constant by putting x = 0

and using (22), (19), and the properties of the factorial
function. This gives
V+l;- 2
W [H (x), H (-x)] = 2 7T eX (24)
v v (-V-I)!

Consequently Hv(x) and Hv(-x) are linearly independent

solutions provided v r 0,1,2, ..• In these exceptional

cases, we can show that

Other solutions can be found by noting that the func-

tions exp (z2) H_ v _ l (±iz) also satisfy Hermite's equation.

The Wronskians can again be obtained by direct calculation

for x = O. This gives

2
W [Hv(x), eX H_v_l(±ix)] (25)

so that these new functions provide solutions independent of

Hv(x) for arbitrary v. We return to this fact below, but

first we wish to use these new solutions to obtain an inte-
gral representation similar to (18) for Re (v) > -1. We
commence by noting that any three solutions of a second-order

differential equation must be linearly dependent; hence there

are constants A, B, and C such that
2
AHv(x) + eX [BH_ v _ l (ix) + CH_ v _l (-ix)] = 0. (26)

By evaluating (26) and its derivative with respect to X

at x = 0, we obtain the relation
350 PART IV: SPECIAL TECHNIQUES

Substitution of (18) into this result yields the integral

representation
v+l x 2
Joo V \l1T
H,,(x) = 2 e t cos(2xt- Z-)dt, Re(v) > -1. (28)
v I7T 0

Formulas (18) and (28) are useful in deriving asymptotic

forms.

Asymptotic Forms: We commence with the integral (17), and

replace the branch cut by the straight line t = -p exp(ia),
where a is chosen by the following considerations: (i) the
integral must be unchanged and convergent, hence
-1T/4 < a < 1T/4, and (ii) we want the factor exp(+2xp) to go
to zero, which imposes the restriction Re(xp) < O. Denoting
arg(x) by 8, this implies -1T/2 < a + 8 < 1T/2 and thus
2
-31T/4 < B < 31T/4. Now we replace exp(-t) by its Taylor
series and apply Watson's lemma for loop integrals, which
gives
(_l)k (2k-v-l)!
L
00
Hv(x) ~ (2x)\I -31T/4 < arg(x) < 31T/4.
k=O k! (-\I-I)! (2x) 2k'
(29)
We also need a formula to cover the region excluded from (29).
Using (27) rearranged in the form

i1Tv -i1T i liT x 2 i1Tv/2 (-i1T/2)

H,,(x) = e Hv(x e ) + -v--- e e H 1 xe (30)
v 2 v! -v-

and applying (29) to each term, we get

H () (2x) \I L (_l)k (2k-v-l)!

00

v x ~ k=O k! (-v-I)! (2x)2k

11T_ e inv e x 2 L (2k+v)!

(31)
(-v-I)! x v+l k=O k!v! (2x)2k '
1T/4 < arg(x) < 51T/4
§20. Laplace's Method 351

5'IT 'IT
with a similar formula for 4 < arg(x) < - 4'
The essen-
tial difference between (31) and (29) is in the term exp(x 2);
unless v = 0,1,2, ... this causes the corresponding func-
tions to be unbounded for iarg(x)i > 3'IT/4, in support of the
assertion made in Section 20.2.

20.4. Bessel Functions: 4 Integral Representations

The Hermite equation investigated above is typical of
the simplest class of second-order equations whose general
solutions are nontrivial. Slightly more difficult is the
Bessel equation, which we consider here. Bessel functions
satisfy the equation
1 2
f" + - f' + (1 - ~) f = O. (32)
z z2

Again, the solution is most easily effected if we remove the

singularity at z = 0; this may be achieved by the substitu-
tion
f (z) (33)
to obtain the equation

zu" + (2v+l)u' + zu 0, (34)

which is amenable to solution by our method. Equation (34)

is closely related to the confluent hypergeometric equation.
However, it is worthwhile to discuss Bessel's equation in its
own right, as there are many important properties and methods
which are specific to Bessel functions.
The direct application of Laplace's method to (34) is
straightforward. In the notation of Section 20.1 we have
352 PART IV: SPECIAL TECHNIQUES

F(p) (2v + 1) p,
2
G (p) P + 1, (35)
S(p) A (p2 + l)v- 1/2

For large Izl, equation (34) may be approximated by

u" + u = 0, the solutions of which are periodic. It is there-
fore conventional to make the substitution p = iw in the
Laplace integral, so that the method yields

u (z) A
(w
2 l)v- 1/2 e iwz dw,
=
21T Ic - (36)

where the contour C and arbitrary constant A must be

chosen. A variety of contours are suitable for defining
various types of Bessel functions. These will be discussed
as the occasion arises.

Another Representation: The integrand in (36) has two branch

points at w = ±l, a feature which is characteristic of the
confluent hypergeometric equation, and which causes some
practical difficulties. An alternative approach, applicable
only to Bessel functions, commences by replacing (33) by

v 2
fez) = z g(z). (37)

After writing ~ = z2 and substituting into Bessel's equa-

tion (1), this yields the new differential equationS for

g (0
4~ gil + 4(v + l)g' + g = o. (38)

Application of Laplace's method to this new equation gives

§20. Laplace's method 353

S (p)
1
exp U
4 (v+l)p + 1
d~
4p2 4p2
(39)
1 v-I e -1/4p ,
;rP
1
v-I pl;-
gel;)
2~i Ic P e 4p dp. ( 40)

More symmetrical formulas may be obtained by a change of

variables. If we substitute back into (37) after replacing
p by p/2z, we obtain the representation
1 ( l)
fez) = 2rri
A J p
v-I
e
7z P - p
dp, ( 41)
C

where the constant A and contour C have yet to be chosen.

Since the parameter v occurs in Bessel's equation only as
v 2 , its sign is undetermined, and another integral representa-
tion is given by
1 . 1
A -v-l Zz(p - -)
f (z) J e Pdp. ( 42)
zrri C p

20.5. Bessel Functions of the First Kind

By choosing A = 1 and the contour shown in Figure 1,
we obtain functions generally known as Bessel functions of
the first kind, and denoted Jv(z), i. e. ,
1 1
Zz(p - -)
e p dp, Re (z) > O. ( 43)

The restriction Re(z) > 0 is necessary to make the integral

converge.

Analytic Continuation: It is a simple matter to perform an

analytic continuation of (43) to all z, and to elucidate

the behavior of Jv(z) about z = 0 at the same time. If

we temporarily restrict z to be real and positive, then the

354 PART IV: SPECIAL TECHNIQUES

change of variables u = pz/2 yields

2
u_ z
J (z) = (z/2)'I) J u -v-l e 4u du, (44)
'I) 2lTi C

where the contour is unchanged since z is real. But the

integral in (44) defines an entire function of z since it
is single-valued and absolutely convergent for all z.
Hence it is a valid representation for all z, and we see
that Jv(z) has a branch point at the origin, but that it
has no other singularities. To complete the definition of
J'I)(z) we must introduce a branch cut; the usual convention
is to make this the negative real axis, specifying the branch
by the restriction -IT < arg(z) < IT.

Series Expansion: A series expansion for J'I)(z) can be ob-

2
tained from (44) by replacing exp (-z /4u) by its Taylor
. 6 ~
serIes and integrating term by term. Using (AB) we thus
obtain
(z/2)'I)
J
'I) (z) 2lTi

00 ( _ 1) k (z / 2) '1)+ 2k
( 45)
k~O ('I)+k)! k!

Recurrence Relations: Recurrence relations can easily be ob-

tained from (43) by temporarily assuming Re(z) > 0 and then
using analytic continuation to remove this restriction. If
we differentiate under the integral sign, we have

J~ (z) e
tz (p- !)
p dp
(46 )

Alternatively, we can integrate by parts to get

§20. Laplace's method 355

-1
27Ti

1 ( !.)
1
27Ti
JC Z(l
Z
+
1
p2)
p-V eZZ p- p
dp (47)

From these two relations a number of others can be deduced

(see Problem 2).

Bessel's Integral: We modify the contour of Figure 1 to that

shown in Figure 2 and let the straight line sections tend
toward the negative real axis. The integral then splits up
into two terms:

1m (p)

Re(p)

Figure 20.2

(i) Contribution from the circular path. If we write

p = exp(i6), we have
356 PART IV: SPECIAL TECHNIQUES

I fIT -ive+izsine I fIT

ZIT e de = rr cos (ve - z sin e) de, (48 )
- IT 0

which comes from writing the complex exponential in terms of

sine and cosine functions, and using the fact that the former
is an odd function and the latter an even function.
(ii) Contribution from straight paths. On the upper path
we put u = exp(s - iIT), on the lower path u = exp(s + iIT).
This yields the contribution

I
ZITi
fO e
-vs+iITV I
expf - 2"z (e s - e -s ) l ds
00
L J
+ ZITi
I
r e
-vs-iITV I
expf - 2"z (e s - e - s ) l ds (49)
0 L J
sin(vIT)
IT f: exp (- z sinh s - vs)ds.

Adding (48) and (49) we have

fIT
J (z) = rrI cos (z sin e - ve) de
v 0 ( SO)
sin(ITv) f"" exp (-z .
sInh s - vs)ds, Re(z) > O.
IT 0

For integer v, the second integral gives no contribution.

The first integral is known as Bessel's integral; the complete
formula (SO) which is valid for all v is a generalization
of Bessel's integral.

ZO.6. Functions of the Second and Third Kinds

For arbitrary v the functions Jv(z) and J -v (z)

both satisfy Bessel's equation. To see if they are indepen-

dent we evaluate the Wronskian. This is most easily done by

first noting (see Problem 3) that for any pair of solutions

of (3Z) the Wronskian must have the form

§20. Laplace's method 357

(51)

so we need only evaluate the constant A. Using (46), we

have

A tZJv (z) ~ -v-l (z) - J -v+l (Z~

(52)

- tZJ_v(z) ~V-l (z) - J V+l (Z~ ,

and on inserting the power series (45) and considering the

limit z + 0,

A 1 1
v!(-v-l)! (-v)!(v-l)!
(53)
-2sin(nv)
n

Hence Jv(z) and J_v(z) are an independent pair of solu-

tions provided v is not an integer. If v is an integer,
then it may be shown (Problem 4) that J
-n
(z)
Other solutions of Bessel's equation may be found
from (42) by using a contour which has one end at the origin.
Because of the essential singularity there, it is necessary
to choose the contour so that it approaches the origin in
the sector !arg(pz)! < n/2, so that the integrand will be-
come zero along this path. This essential singularity may
be removed by the substitution p = exp(t), leading us to
consider the functions

Z (z) = A
v C
f
e Z sinh t-vt dt , (54)

for suitable contours C. By repeating the arguments which

led to (46) and (47), it can readily be shown that they
 358 PART IV: SPECIAL TECHNIQUES

satisfy the recurrence relations?

2Z~ (z) = ZV_l (z) - ZV+l (z) ,

(55)
;V Zv(z) = ZV_l(z) + Zv+l(z),

Choice of Path: The integrand in (54) is an entire function

of t, consequently closed contours yield the trivial func-

tion Zv = O. Furthermore, the integrand has no zeros, so
both ends of the contour must approach infinity in such a
way that Re(z sinh t) + If we restrict our attention
to Re(z) > 0, then these considerations impose the follow-
ing restrictions:

(i) Re(t) + 00, Im(t) + (2n+l)rr,

(i i) Re (t) + - 00, 1m (t) + 2n rr .

We also note that if two contours are related by the dis-

placement t + t + 2rrni, then the functions which they de-
fine are identical except for the constant multiplier
exp(-2rr inv). We consider the four contours shown in
Figure 3--any other suitable choices will give linear com-
binations of the functions which we obtain from these.

The Functions J±v(z): By a simple change of notation it

can readily be shown that (54) reduces to the generalization

of Bessel's integral (50) if A = 1/2rri. Hence

J (z) =
1 f e
z sinh t-vt
dt, Re(z) > O. (56)
v C
---2
.
rrl
3
Next we consider the contour C4 . The substitution
t = irr - s maps the contour C4 into C3 and changes the
§ZO. Laplace's method 359

1m (t) 1m (t)

iTT

Re (t) Re (t)

-iTT

1m (t) 1m (t)
2 iTT

iTT

Re (t) Re (t)

-iTT

Figure ZO.3

integrand to exp [z sinh s + VS - inv]. Hence, on choosing

A = -exp(inv)/Zni, we obtain

J
-v
(z) =
_e inv
Zni
f
e Z sinh t-vt dt, Re (z) > O. (57)
C4
Thus we recover functions of the first kind from the con-

tours C3 and C4 .

Hankel Functions: Functions of the third kind, named in

honor of Hankel, are obtained from the contours Cl and CZ'
Explicitly, they are defined by
360 PART IV: SPECIAL TECHNIQUES

1. f e zsinht-vt dt,
1fl C
1 (58)
H(2)(z) = - ~
v 1fl
fC ezsinht-vt dt, Re(z) > O.
2
It follows immediately from (56) and (58) that

J v (z) = ~.rH(l) (z) + H(2) (z)l. (59)

2L v v J
Furthermore, i f we translate the contour C2 by the substi-
tution t + t + 21fi, we obtain from (57) and (58)

J (z) (60)
-v

These relations have been proved under the restriction

Re(z) > O. However, analytic continuations of Hankel func-
tions may readily be effected; if we introduce the negative
real axis as the branch cut, (59) and (60) are valid in the
entire cut plane.

Wronskian: The Wronskian of the two Hankel functions is con-

venient1y evaluated by using (59) and (60). This gives

J J] sin(1fv) W[H(l) H(2)] (61)

W[ v' -v = 2i v' v '

and after substituting the value of

W[H~l) (z), H~2) (z)] 4i (62)

1fz

Thus the two Hankel functions are linearly independent for

a11 v.

Weber Functions: Functions of the second kind, named after

Weber, are defined by
§20. Laplace's method 361

Yv(z) = l...-f
21L v
H(l) (z) - H(2) (z)
v
l.J (63)

The origin is again a branch point, and it is the usual con-

vention to use the negative real axis as a branch cut. The
pair of functions J v ' Yv are linearly independent for all
v, as may be seen by evaluating the Wronskian

(64 )

Other properties of Weber functions are given in the problems.

Another Integral Representation: We conclude hy deriving an-

other integral representation for the Hankel functions which
will prove useful later on. We commence by making the sub-
stitution t + S ± i~ in (58) so that

e-i~v/2
H(l) (z)
v ~i Ir 1
e
izcoshs-vs
ds,
(65)
i~v/2
H(2) (z)
v
e
~i Ir 2 e -izcoshs-vs d s,
Re(z) > 0,

where the contours rl and r 2 are shown in Figure 4. If

1m (Z)

Re (Z)

Figure 20.4
362 PART IV: SPECIAL TECHNIQUES

we further restrict z by H(1) and

Im(z) > 0 for
v '
Im(z) < 0 for H~2), we can deform these contours in the
s-p1ane to the real axis. Hence

(1)
Hv (z) =
e-i~V/2
1Tl
foo e
izcoshs-vs
ds, Im(z) > 0, (66)
-00

( 2) (z) ei1Tv/2 foo e -izcoshs-vs d s

= - .::::.....--;--
1 T i ' Im(z) < 0,
H (67)
v _00

where we have removed the restriction Re(z) > 0 since these

formulas achieve analytic continuations of (58).

20.7. Poisson and Related Representations

In Section 20.4 we derived two distinct integral
representations of Bessel functions, although we have only
considered one of them (42) so far. We now examine the in-
tegral (36) with the contours shown in Figure 5. For the

1m (w)

C2 C,
branc.h cuts
Figure 20.5

contour Cl we temporarily assume that z = is, with ~

real and positive, and consider the function defined by
1
fC e
-w~
(w -1)
2 v-Z
dw. (68)
l
We specify the branch of by requiring arg (w 2 -1)
§20. Laplace's method 363

to become zero as we move to large Iwl along the contour.

Now we use the result (see Problem 1.17) that

(69)

which is valid provided we require larg(w)1 < rr/4, which is

consistent with w lying on the contour Cl . We therefore
use this integral to represent exp(-ws) in (68) and then
interchange the orders of integration. This leads to the
problem of evaluating
1
-w2t2 2 v-Z
J we (w -1) dw, (70)
C1
which can be done by the substitution w2 -1 u exp(-irr),
giving 1 1
1 -irr(v+I)
ze e- t
2 fO +
u
V-I
e
_ut 2
duo (71)
-""
With the help of (A8) we can easily evaluate this new
-2v-l 1
integral as [2rrit /(-v-I)!]' and on using these results
in (68) we have
1
v--
e-ws(w2_1) 2 dw
JCl
2/rr e-irrVJ"" e- t2 -(s/2t)2 t- 2V - 1 dt
(72)
(-V-i)! 0

lIT e- irrv / 2 (z/2)-Vf""_"" eiZcoshs-vs ds,

(-V-i)!

where the last step follows from the substitution

t 2 = (z/2) exp(-S-~irr). Comparing this with (66) we see that
we have recovered a Hankel function, and since (68) and (72)
both define analytic functions for Im(z) > 0, we may lift
the restriction z = is to write
 364 PART IV: SPECIAL TECHNIQUES

Hv
( 1)
(z) =
l )!
(-V--
Z
. 3/Z
Cz/Z)v f 'e lWZ Z
(w -1)
l
V--
Z
dw, Im(z) > 0. (73)
11T Cl

A similar argument leads to

(Z) (_v_!-)!(z/Z)v iwz Z v-i
H (z) = - / J e (w -1) dw, Im(z) < 0, (74)
v i1T3 Z C
Z
where the phase of (wZ_l)v-l/Z is again chosen so that it be-
comes zero as we move along the contour after circling the
branch point.

ZO.8. Modified Bessel Functions

The differential equation

1
f" +
z f'- 0, (75)

which is closely related to Bessel's equation, occurs fre-

quently in applications. In principle, it may be solved by
the substitution z = ±it, which turns it into Bessel's equa-
tion, but because of its frequent occurrence, special term-
inology has grown up for it. This has been brought about
particularly by the fact that Bessel functions are defined in
the cut plane because the origin is a branch point.

Modified Bessel Functions of the First Kind: The functions

±i7Tv/Z
e Jv (~iz) (76)

are solutions of (75) which are finite as z + 0 if

Re(v) ~ 0, and identical if Re(z) > ° (Problem 13). From

them we may construct a solution of (75) in the region

-1T < arg(z) < 1T by

§20. Laplace's method 365

(77)

The function Iv(z) thus defined is usually referred to as

a modified Bessel function of the first kind. For real v
and z, it takes on real values.

Macdonald's Function: For a second independent sOlution of

(75), it is often convenient to have a function which be-
comes small for large real z. Such a function cannot be ob-
tained by replacing Jv by Yv in (77). However, it may
be verified that the functions

'ITi i'ITV/2 H(l) (.lZ,

)
2 e v
(78)
'ITi -i'ITv/2 H(2) ( . )
Z- e v -lZ,

which are both solutions of (75), are equal for Re(z) > 0
(Problem 13). Hence, it is conventional to define as the
second solution to (75) the function Kv(z), known as
Macdonald's function, by

f Z- e i'ITv/2 H(l)
'ITi v (l'Z), -'IT < arg(z) < 'IT/2

l-
(79)

;i e- i 'ITv/2 H~2) (-iz), -'IT/2 < arg(z) < 'IT.

Recurrence Relations: Recurrence relations corresponding to

(55) are easy to derive from the basic definitions of Iv(z)
and Kv(z) and (55). They are
366 PART IV: SPECIAL TECHNIQUES

21' (z) I V_l (z) + IV+l(z),

v

- 2K~ (z) K (z) + KV+l (z),

v-I
2v I (80)
z v (z) Iv_l(z) IV+l (z) ,

- -2v K (z) Kv _l (z) - KV+l (z).

z v

Problems

1. Show that

-t 2+2itx n
e t dt, n = 1,2,3, . . . .

2. Prove the recurrence relations

d v v
[z J (z)] z J,,_l (z),
dz v
d -v -v
az [z J
v
(z)] -z J V + l (z) ,

J~ (z)

J v' (z) = -J
v+
l(z) + ~z J v (z).

3. Show that the Wronskian of any two solutions of Bessel's

equation is of the form A/z, where A is a constant.
(Consider the differential equation satisfied by the
Wronskian.)

4. Show that

J (x) (-l)n J (x).

-n n
§20. Laplace's method 367

5. Show that

H(l) (z) eirrvH(l) (z) ,

-v v
B(2) (z) e -irrv H(2) (z).
-v v

6. Prove the relations

Jv(z) cos(rrv) - J_v(z)
Yv (z) =

t'\ (Z)j
sin (rrv)

Y (z)
n
= 10.
rr Clv v=n
-~
rr
[Cl,T _v(z)1
Clv Jv=n'

Y (z) = (_l)n Yn (z).

-n

7. Verify the expansion

_ 10. n~l (n-k-l)!

Y (z) [.-
n rr k=O k!
I 00 (_I)k(~z)n+2k
+ rr l
k=O k! (n+k)!
where the first sum is set equal to zero if n 0, and
w(z) = d R,n[(z-l)!]/dz.

8. Show that

Yn(z) ft.
·V
_(n-l)!
rr (1o.z)-n
2' z + 0, n > I.

9. Prove the following properties governing the behavior of

Bessel functions on the branch cut Im(z) = 0, Re(z) < 0:

J (-x-iO) 2i sin(rrv) Jv(x),

v

2i[cos(rrv) J v (x) + J -v (x)],

368 PART IV: SPECIAL TECHNIQUES

H(l) (-x+iO) - He l ) (-x-iO)

v v

H(Z) (-x+iO) H(Z) (-x-iO)

v v

where x > O.

10. Using Watson's lemma in conjunction with the Poisson

integral representation, show that
(-l)k(V+k-t) !
k! (v-k-i)! (Ziz)k'

z + 00, -n/2 < arg(z) < 3n/Z,

1
(v+k- Z)!
k! (v-k-l)! (2iz)k'
Z
z + 00, -3n/Z < arg(z) < n/2.

11. Verify the following relations between Bessel functions

and circular functions:

H~l) (z) '" (~zr/z e i[z-(nv/Z)-(n/4)] ,

H~Z)(z) '" [;zr/ z e- i [z-(nv/Z)-(n/4)],

J v (z) '" (nZzJ 1/ Z cos[z-(nv/Z)-(n/4)],

Yv(z) " ' " [:z] l/Z sin[z-(nv/Z)-(n/4)].

lZ. Verify the following formulas for Bessel functions of

order integer plus one-half:

Jl/Z(z) '" (;zf/Z sin z,

Yl / 2 (z)
z Jl/Z
- (- cos z
nz '
§20. Laplace's method 369

(1) iz
H1/2 (z) -i (;zr/2 e

H(2) (z)
1/2
i (;z] 1/2 e -iz

J 1 (z) ( _l)n [~r/2 zn+1/2 (~Jn sin z

n+- zdz z
2

y 1 (z) (-1) n+1 [2f/2

:rr zn+1/2[~r ~
n+- zdz z
2

13. Prove that for real x > 0

i'ITV/2
e J (-ix)
v
= e - i1TV/2 J v (ix),

e i1TV/2 H(l)
v (0lX ) -_ -e -i1TV/2 H(2)
v ( -lX.
0 )

Prove the following integral representations.

1
1 2 v-2"
14. J (w -1) e
-UlZ
dUl ,
-1
1
Re (v) > - 2

15. K (z)
v
1T 1/2 z v
2V (V-i) ! r0
e-z cosh t
51nh2v t dt,
o

1
Re(z) > 0, Re(v) > - 2

2v (v-t)!

xV 1T1/2 r cos (xt)

0 (1+t 2) v+1/2
dt, x > 0,
1
Re(v) > - 2'

17. Show that

sin (1TV)
370 PART IV: SPECIAL TECHNIQUES

18. Verify that, for x > 0

I (-x+iO) - I (-x- iO)

v v

-i 1f [ I (x) + I (x)],
-v v

2 cos(1fv) Kv (x).

19. Show that

(;zJ
"
l /2
sinh z,

[;zf/2 e
-z

In+1/2 (z) [~r/2 zn+1/2[~]n

zdz
sinh z
z

-z
e
z

Integrals Involving Bessel Functions: There is an enormous

amount of literature on the evaluation of integrals involving
Bessel functions. WATSON (1958) is a primary reference on
methods; extensive tables are also available. The two most
important techniques are: (i) use an integral representation
for one of the factors and interchange the order of integra-
tion, and (ii) expand one of the factors in a power series
and integrate term by term. Verify the following:

J:
1 1
2il (~ + IV - 2")!
20. xil J v(x)dx 1)
(- ill +
2"
Iv
2 2

Re(il) < -~, Re(il+v) > -1.

[20. Laplace's method 371

[/aV - a] v
21. f
ooD e
-ax
J v (bx)dx
bV~

a > 0, b > 0, Re(v) > -1.

[Use (43).]

22.

2 2
(Expand exp (-a x ) and use Problem 20.)

23. K (ab) ,
V-ll

a > 0, b > 0, -1 < Re(v) < 2 Re(ll) + i

[Use the representation

1 Joo 2 2
(x 2 +a 2 ) -ll-l = -, e -(x +a )t t ll dt.]
ll· 0
24. Using the convolution formula for Laplace transforms,
obtain the result

Airy Functions:

25. Show that two independent solutions of Airy's equation

utI - zu = 0

are
372 PART IV: SPECIAL TECHNIQUES

1/2
z
Ai(z)
3

Bi(z) = ("3z) 1/2

where ~ = 2z 3/2/ 3 . These solutions are known as Airy

functions of the first and second kinds, respectively.

26. By the application of Laplace's method, show that two

solutions of Airy's equation are

f
1
=
1
11 f: cos (} t 3+xt) dt,

f: ~
- -t
13 3 +xt +
f2 =
1
7f
e Sin(~ t 3+xt)]dt, x > O.

27. Prove that the solutions flex) and f 2 (x) of Problem

25 are the Airy functions Ai(x) and Bi(x),

respectively.

28. Show that as Iz I + 00

e-I; 27f 27f

Ai (z) 3 < arg(z)
<
'" 27f1/2 z 1/4' 3

e-I; 7f
Bi (z) '"
1T
3
< arg(z) <
3
,
7f1/2 z 1 / 4 '
2 3/2
where ~ = "3 z .

29. Show that as x + 00

cos[l;- (7f/4)]
Ai(-x) '"
7f1/4 xl/4

sin [I; - (7f / 4) ]

Bi(-x) '" - 7f172 x 1 / 4 x > 0,

where I; ~ x 3/ 2 .
§20. Laplace's method 373

30. Show that

and hence derive the asymptotic series

e- s (_l)k (3k-t)!
2
00

AiCz) ~
2nzl/4 k=O 3 2k (2k)! sk
where s i z3/2.

Footnotes

1. More complicated equations can sometimes be reduced to

this form by suitable transformation.

2. For details beyond those given in this section see, for

instance, ABRAMOWITZ &STEGUN (1965), Ch. 22, and
LEBEDEV (1965), pp. 60ff.

3. Since on differentiating Wand using (7) we have

W' = 2xW, whose solution is (23).

4. The classic and monumental reference on Bessel functions

is WATSON (1958).

5. Bessel's equation is a special case of the confluent

hypergeometric equation; one of its distinguishing fea-
tures is that under this transformation it remains an
equation of the same form.

6. This is permissible even though the function has an es-

sential singularity at u = O.

7. Functions satisfying (55) are known as cylinder functions.

They satisfy Bessel's equation as a consequence of (55).
374 PART IV: SPECIAL TECHNIQUES

§2l. NUMERICAL INVERSION OF LAPLACE TRANSFORMS I

There are many problems whose solution may be found in

terms of a Laplace or Fourier transform, which is then too
complicated for inversion using the techniques of complex
analysis. In this section we discuss some of the methods
which have been developed - and in some cases are still being
developed - for the numerical evaluation of the Laplace inver-
sion integral. We make no explicit reference to inverse
Fourier transforms, although they may obviously be treated
by similar methods, because of the close relationship between
the two transforms.

21.1. Gaussian Quadrature Formulae for the Laplace Inversion

Integral
If F(p) is a Laplace transform having the form

(1)

where ~(p-l) is a polynomial of degree N, then we may

easily find the inverse function f(t) numerically using n
point Gaussian quadrature formulae which are exact whenever
2n-l > N. Many such rules have been derived for real inte-
grals of the type

t a
w(x) f(x) dx (2)

depending on the choice of weight function w(x) and limits

a and b. We give here the derivation of a similar rule for
the evaluation of the Laplace inversion integra1 2 , under the
assumption that, for some s > 1, the function pSF(p) may
-1
be approximated by polynomials in P Setting u = pt,
and writing
§2l. Numerical inversion of Laplace transforms 375

F(p) '" p-s G(p) ( 3)

our aim is to approximate the inversion integral

f(t) 1 r
C+ ioo
2Wl Jc-ioo
F(p) e pt dp

I
(4)
t s -l Ct + 1.· OO u-s
2wi e u G(u/t) du
Ct-lco
as

f(t) ( 5)

in such a manner that the approximation is exact whenever

-1
G(p) is a polynomial in P of degree less than or equal to
2N-1.
Now it is well known 3 that a Gaussian quadrature for-
mula for (2) may be constructed once we have a set of poly-
nomials PN(x) of degree N, each of which satisfies the
orthogonality relations

Ib
a
k
w(x) x PN(x) dx '" 0, k '" 0,1,2, ... , N-l ( 6)

In the present case, we need to find polynomials PN,s(u- l )

with the orthogonality properties

2iI
1 IY+ 1.· OO eU
Y-l oo
u
-s-k
PN,s(u- ) du '"
1
° ( 7)
k '" 0,1,2, ... , N-l

The problem is most easily solved using the convolution prop-

erty of Laplace transforms, since the inverse of polynomials
in p-l must be polynomials in t. Therefore we define a
set of functions ~N,s(t) by
376 PART IV: SPECIAL TECHNIQUES

(8)

and after noting that 5([t s +k - 2 ] = (s+k_2)!/ps+k-l, we see

that equation (7) is equivalent to

J: (1_t)s+k-2 ~N,S(t)dt = 0, k = 0,1,2, ... ,N-l (9)

The substitution t + (2t-l) converts this to a standard for-

mula expressing the orthogonality of the Jacobi polynomials;
using the notation of hypergeometric functions, we have,

~N,s(t) = 2Fl(-N,N+s-l;1;t) (10)

The polynomials PN,s(p-l) are readily found by taking the

Laplace transform, they are

(11)

Therefore, by the standard techniques for the derivation of

Gaussian quadrature rUles 3 , the points uk in (5) are the
-1
zeros of the polynomials PN,s(p ), and the weights Ak are
given by the formula

(_l)N-l (N-l)! [2N+S-2 ]2 (12)

NCN+s-Z)! -1
uk PN-l,s (uk)

Tables of uk and Ak for various values of sand N are

available. 4

Restrictions on application: Suppose that the function G(p)

has the asymptotic expansion

Ip I + 00 (13)

then the Heaviside series expansion gives for the inverse

§2l. Numerical inversion of Laplace transforms 377

function f(t) the asymptotic expansion

00
t s +n - l
f(t) "v ~ an (s+n-l)! (14)
n=O

Denoting now by fN(t) the approximation to f(t) obtained

from the N point Gaussian rule, we may write

N
f (t) = t s - l I Ak G(uk/t)
N
k=l

Y
00

"v t s-l ~ an [ Ak(t/uk)n]

n=O k=l
00
(15)
I a r t n+s - l
n=O n n,N

r n,N

where the interchange of order of summation is valid because

the series are only asymptotic. Now r n,N is the approxi-
mate value of the inverse of the function u- s - n , hence by
the properties of the quadrature rule

1
r n,N (s+n-l)! ' n < 2N - 1 (16)

and we see that the method recovers a sequence of functions

whose asymptotic series match that of f(t) more and more
closely. Conversely, if G(p) has the asymptotic expansion
00 -A
G(p) "v ~ a p n
n=O n
(17)

where some of the An are not integers, then the Gaussian

formula will not reproduce the coefficients of the corres-
ponding terms in the asymptotic expansion for any value of N.
378 PART IV: SPECIAL TECHNIQUES

Hence the method must be restricted to functions for which

(13) holds, and we have

(18)

for such functions.

Convergence in N: 5 Now assume that G(p) is analytic at

infinity, in which case (13) is a convergent series for
ipi > PO' Now it is well known that Gaussian quadrature is
equivalent to replacing the integrand by a Lagrangian poly-
nomial interpolation formula, and then integrating this for-
mula analytically. For the present case, the Lagrangian
formula is

N
~
k=l (l/u - l/u k ) [PN, s (u)] u=u -1
k

1 [P N sew) - PN,s(u- l )]
ZiT
G(l/tw)dw (19)
fW-l/u)PN ,s ew)

where the representation as a contour integral follows from

the use of standard residue theory, and the contours CN are
circles centered on the origin, and large enough to contain
all the zeros of PN,s (u -1 ). Since G(p) is analytic at
infinity we may write

GCu/t) 2d
1 fC G(l/tw) dw
w-l/u (~O)
N

and on substituting this into (19) we derive an expression

( 21)
t s - l fct+i~ u -s-l -1 G(l/tW)dw}du
= ~ . e u PN s(u ) w-l/u
ct-l~ ,
§2l. Numerical inversion of Laplace transforms 379

The orders of integration may be interchanged, which gives

the result in the more useful form

(_l)n n! t s - l
2n(2n+s-I)!
I C
(22)
N

From this formula, it is not difficult to show that

lim EN(t) 0, 0 < t < 2/PO (23)

N+co

Thus the approximations fN(t) are shown to converge to f(t)

as N increases for a range of values of t which depends on
the analytic behaviour of G(p). In fact, a function with
the assumed properties of G(p) would give an inverse such
that t-s+lf(t) is analytic for all t, and it seems plaus-
ible that the functions fN(t) should converge to f(t)
for all t. No proof of such a property seems to be known.

21.2. Use of Laguerre Polynomials

It is a fundamental property of Laplace transforms that
they are analytic functions in a half-plane Re(p) > PO.
Polynomial approximation, on the other hand, is generally
useful in a finite disc, not a half-plane. Thus, to use
polynomial approximation of F(p) as a means of numerical
inversion, it is useful to first carry out the bilinear trans-
formation

z=~ (24)
p+D

which maps the half-plane Re(p) > -(a+b)/2 into the interior
of the unit disc \z\ < 1. F(p) is then written as a func-
tion of z, and suitable polynomials are chosen to enable
convenient and accurate numerical evaluation of the neces-
sary approKimations. In this section we discuss the use of
380 PART IV: SPECIAL TECHNIQUES

generalized Laguerre polynomials; in the next the use of

Chebyshev polynomials. In either case, the choice of suit-
able parameters in (24) is obviously a compromise: on the
one hand it is essential that none of the singularities of
F(p) be mapped too near to the disc Izl < 1, since singu-
larities generally upset numerical processes; on the other
hand the singularities should not be mapped too far away,
since they contain essential information.
Laguerre polynomials are a natural choice 6 since the
Laplace transform of the function

-t/2
$~ (t) = t Cl e L~(t) (25)

is simply
(Cl+n) ! (p -z
l)n
( 26)
Cl! (p + i)n+Cl+I
1 1
which may be simplified by writing z = (p - Z)/(p + I)· To
apply this property in a practical way, we first write the
approximation

f(t) (27)

It is assumed that the exponent Cl is known from the asymp-

totic form of F(p) for large p. T simply fixes the scal-
ing of the time variable for the particular problem. The
Laplace transform of (27) is

1 n
N (p - c - IT)
F(p) (28)
n~o an (p-c + }T)n+Cl+1

Introducing the bilinear transformation

p-c-i T
z = 1
(29)
p - c + I T
§21. Numerical inversion of Laplace transforms 381

reduces (28) to
N
(pT - cT + ~)a+l F(p) ~ L (30)
n=O

Also, from (29) we see that the disc Izl < 1 is the image
of the half plane Re(p) > c, which offers some help in choos-
ing an appropriate value for c.
To calculate the values of f(t), we must calculate the
coefficients an and the Laguerre polynomials Ln(t). The
second task is easily carried out by using the recursion
relations

1 + a - t (31)

(2n+a-l-t)L~_1(t) - (n-l+a)L~_2(t)

which are stable. The coefficients an may be found by using

trignometric interpolation on the unit circle, which relies
on orthogonality properties of the trignometric functions over
discrete sets of points. Setting

z = e i9 (32)

which is equivalent to

i 9
p c + TI cot 2" (33)

Eq. (30) reads

h(9) = (~ 2"
9r+ l F( c + i cot
+ i cot "2
TI ;) (34)
N
~
L
n=O
an(cos n9 + i sin n9) .

This gives the following formulae for the coefficients an

382 PART IV: SPECIAL TECHNIQUES

1
N
aO L Re h (e.)
N+l j=O J

1
N
an L cos(ne.)Re(h(e.))
2CN+l) j=O J J
( 35)

e. (2j+l) 'If
J N+l "2

The choice of parameters c and T involves some experimenta-

tion in any particular application. However, a useful start-
ing point is given by Weeks. 7 It is
t
T
max
-N-
(36)
1
c = PO +
t max

Here t is the maximum value of t at which f(t) is to

max
be computed, and Po gives the half plane in which F(p) is
analytic.

12.3. Approximation of F(p) by Chebyshev polynomials 8 for

Real p
Laguerre polynomials are capable of giving extremely
accurate results in a wide range of cases, but the method suf-
fers from the disadvantage that it is necessary to compute
F(p) along the line Re(p) = c. An alternative method, which
involves only real values of p, makes use of the expansion

-s ~I -1
F(p) = P L an Tn (l-b p ) (37)
n=O

in place of (27) . Here Tn is the Chebyshev polynomial of

degree n, b is a free constant to be chosen later, and the
prime on the summation symbol means that the first term must
be multiplied by one half. If this series is truncated after

N terms then we have a least square approximation of F(p)

§Zl. Numerical inversion of Laplace transforms 383

over the interval ~b 2 p < 00 with weight function

p-3/2 (1_b/Zp)-1/2 (38)

Inverting the series term by term, we obtain

f(t) (39)

where ¢n(x) is a polynomial of degree k. The first three

of these polynomials are readily found to be

¢o(x) 1

¢l(x) 1 - -Zx
s ( 40)

1 - 8x + 8 2
¢Z(x) s s(s+l) x

For n > 3, the direct evaluation of the ¢n(x) is a numeri-

cally unstable procedure owing to cancellation of large a1ter-
nating terms; Piessens has shown that they obey the recurrence
re1ations 9

4>n (A+Bx)4>n_1 + (C+Dx)4>n_Z + E4>n_3

A Zn -
(n-1)(Zn-3)(s+n-Z)
(n-Z) (s+n-1)

B -4 (41)
n+s-l
C 1 - A - E

D
4(n-1)
(n-2) (s+n-l)
(n-1) (s-n+2)
E = - (n - 2) (s +n -1)

and that these relations are numerically stable and thus

suitable for automatic computation.
The coefficients cn in the expansion may be expressed
384 PART IV: SPECIAL TECHNIQUES

as definite integrals in the usual manner for orthogonal poly-

nomial expansions. However, the Chebyshev polynomials enjoy
some remarkable orthogonality properties over finite sets of
points, one of which results in the following property: 10 if
f(x) is approximated by the Chebyshev expansion

( 42)

then the coefficients are obtained by the formula

2 N
'IT
I~
L
k=l
f(I;.) T (1;.),
J n J
n = 0,1, ... ,N-1 ( 43)

where the are the zeros of TN(x). Applied to the pre-

sent problem, in which we have truncated the fundamental ex-
pansion (37), it gives the formula

-N2 Nil
k=O
~(cos (2k+1)
N '!
w) cos ((2k+1) ~)
N 2
(44 )

where

~(u) = (~)S
1-u
F(~).
1-u
( 45)

It only remains to choose the parameter b. Clearly

we need b sufficiently large so that F(p) is analytic for
1
Re(p) > Zb, apart from that fact the choice is arbitrary. The
polynomials ~n(x) have n real positive zeros, so that
they are initially oscillatory, and then increase like xn
once the last zero has been passed. The argument of the poly-
nomials in equation (40) is }bt, thus we will not want }bt
too large or the series will contain serious cancellations.
On the other hand, we see from the formula for the coeffici-
ents that a large value of b will make them rapidly
§2l. Numerical inversion of Laplace transforms 385

converge to zero for large n. In fact, if we take the first

neglected term in the truncated series (38) as an indication
of the error, we will need to choose b large enough to make
1
c N small, and small enough to avoid having ~N(2bt) too
large. Clearly there is a need for some experimentation in
any particular calculation.

21.4. Representation by Fourier Series ll

An alternative and more explicit form of the inversion
integral is given by either of the formulae

f(t) 2e ct
~
f=0 dw Re[F(c+iw)]cos(wt) (46)

f(t) = _2ew
ct f=0 dw Im[F(c+iw)]sin(wt) (47)

where it is assumed that c > PO. If the trapezoidal rule

is applied to (46), then we have the approximation

2e ct ~I
f(t) = ---- L Re[F(c+iwk/T)]cos(~kt/T) (48)
T k=O
where T is a scaling parameter and the prime on the summa-
tion means that the k o term has weight one-half. There
is no theoretical reason for using the real part of F(p) as
in (48), and another formula may be obtained by applying the
trapezoidal rule to the average of (46) and (47). This gives

e ct ~I
f(t) = -r- L {Re[F(c+i~k/T)]cos (wkt/T)
k=O
(49)
- Im[F(c+i~k/T)]sin (~kt/T)}.

The error involved in these formulae is essentially bounded

by exp (PO-c)T , becoming worse as t approaches T/2 for
(48) and T for (49). This is a most unexpected but welcome
386 PART IV: SPECIAL TECHNIQUES

result, since for large values of T it becomes small much

faster than the usual T- 2 rule 12 . The reason for this be-
haviour is that both methods give Fourier representations of
functions which are closely related to f(t).
The sums in (48) and (49) must be truncated in any
practical calculation, so that Eq. (48) or (49) is used to
calculate the Nth partial sum fN(t). Two procedures may
now be used, either separately or together, to improve the
computational efficiency of the method. First, a convergence
speeding technique may be used on the partial sums. For
examp 1 e, t h e epsl-I on a 1 gorlt
- h m13 -
VIZ

N N+l + (N+l
£s-l £s - ~Ns)-l (50)
£s+l ~

N N
may be used, where £-1 = 0 and £0 is the Nth partial sum.
In many applications of the Fourier series method, particu-
larly to discontinuous functions f(t), it is necessary to
calculate to rather high values of N, and in this case the
-
F ast F ourler trans f orm tec h nlque
- 14 may b I oye
e emp d. This is,
of course, only a computational variation, but nevertheless
it may be the essential step in rendering the method viable.

21.5. Pade Approximation

As a prelude to the discussion in Section 21.6, we
give here a brief treatment of approximations using rational
functions. Suppose that f(x) has the representation

p+q k 1
f(x) = L ck x + &(x p +q + ) (51)
k=O
then we call the rational function
§21. Numerical inversion of Laplace transforms 387

f (x)
Fp,q (x)
QxT
P
fp ex) L Ilk xk (52)
k=O

gq(x) r
k=O
Sk x
k

So 1

a Pade approximation to Fp,q has the same re-

f(x) if
presentation (51) up to, but not including, the remainder
term. This is equivalent to the requirement that

gq ( x ) f (x) - f p ( x ) = x + + 1 h p,q (x)

p q
(53)

Clearly, the functions Fp,q may be arranged as a matrix,

and this is usually known as the Pade table for f(x) . Expli-
cit formulae may be given for the elements of the Pade table
for certain special functions f(x) ; we refer the reader to
specialized literature for details. 15
In general, the coefficients of the polynomials may
be obtained by equating powers of x up to x p +q in (53),
which gives
k
L c J" Sk - J"'
j =0
k 0,1, ... , p

(54)
k
o L
j=O
c"
J
Sk .,
-J
k = p+l, p+2, ... , p+q.

This is a set of p+q+l inhomogeneous equations; for practi-

cal purposes they are not very convenient, and we now show
how they may be solved recursively.16 The method uses the
fact that the first row and the first column of the Pade
table are easy to compute. For the first row, we have
388 PART IV: SPECIAL TECHNIQUES

E
L ck
xk
k=O
(55)
gq(x) = I

and for the first column, equations (54) become

Co
(56)
k
- L c J' Sk-J" k I, ... , q
j=l

which may be solved recursively. Suppose that we already

have p,q-l and F
F
p- l
,q' We denote the corresponding co-
efficients by a~, S~(Fp,q_l) and a k , Sk(Fp_l,q). If we
write out the corresponding equations to Eq. (53) for these
two approximations, we have

q-l
f(x) I
k=O
+ k
ak x
k=O
+ k
L Sk x = 6'(x p +q )
(57)

f(x)
p-l
k=O
L a*k x
k
r
k=O
k
S*k x = 6'(x p +q )

and on subtracting the second relation from the first

p-l + k
f(X)[k~O (a k - apx + a; x p ]
(58)
- [qi l (Sk+ - S*)x k - S~ x q ] = 6'(x p +q ) .
k=O k

Now, So = I and aO Co for any element of the Pade table,

so we may divide out a factor x in (58) . This gives
p-l k q-l
f(x) L ak x L b k x k = 6'(x p +q - l )
k=O k=O

+
a~+l' k 0, I, ... , p- 2
ak r~+l
ap k p - I (59)
§ 21. Numerical inversion of Laplace transforms 389

+
{ 13 k+l - *
13 k +l , k 0,1 , ... , q-2
bk ( 59)
-13*q k q-l

Eq. (59) is almost identical to the defining relation (53) for

Fp-l,q-l' If 13+1 r 13 1*, then we may divide by the factor bO
to obtain

(60)

for the coefficients of the element Fp-l,q-l' If b O = 0,

a l+ = * and (59) gives for F
then we have aI' p-l,q-l a reduc-
ible rational function, a circumstance which needs special
consideration. 17 Normally, however bO r 0, so we have a
recursive relation which enables a triangular portion Fp,q'
p+q ~ r, to be built up from a knowledge of the functions

Fk , 0' F0 , k' k = 0,1, ... , r.

21.6. Rational Approximation of F(p)

Writing, as usual,

F(p) p-s G(p), s > 1 ( 61)

where s is chosen so that G(p) + 1 as p + 00, we may ap-

proximate G(p) by rational functions, using a Pad~ table,

and then invert these rational functions using standard
18
methods. [See Section 2]. For relatively low order approxi-
mations, it will be possible to determine the position of the
pole, and to calculate the various residues. More often, how-
ever, particularly in conjunction with high order approxima-
tions, it is necessary to find some method of inversion which
"
d oes no t re 1 y on f lrst compu t"lng t h e zeros 0 f po 1ynomla
"1 s. 19
390 PART IV: SPECIAL TECHNIQUES

We discuss such a method here, assuming that s = I so that

we want to invert

**
pep) (62)
B P
where
m
A(p) L a. p i
1
i=O ( 63)
n
B (p) =
i=O
L n > m.

If we assume that the inversion is carried out by decomposing

Pcp) into partial fractions and then inverting each term,
and that the roots aI' a 2 , •.. , an of B(p) are distinct,
then we may write immediately

n A(a i )
Pcp) L
i=l (p-a) B'(a i )
(64)
n A(a i ) a.t
f(t) L e 1
i=l B ' (a i )
with similar but more complicated results if one or more
roots is repeated. The problem is to replace these formulae
by an algorithm which does not employ the a ..
1

Special Case: Suppose that A(p) = I and that all the roots
of B(p) are distinct. Since Eq. (64) for f(t) is a linear
combination of a finite number of exponentials, we may ex-
pand each exponential as a Taylor series and reverse the order
of summation. This gives

n a.kt k
L
00

f(t) ILl
i=l B'(a i ) k=O -ry--
(65)

where
§2l. Numerical inversion of Laplace transforms 391

(66)

Now it is easily shown by using a contour integral representa-

tion for the coefficient uk that uk = 0 for k 2 n-2, and
-1
that u n-l = bn . Now for k ~ n, we rearrange the defini-
tion of uk using the fact that B(u i ) = 0, to give

bnuk
I
i=l
Ui
k-n
B'Cu i ) bn Ui
n

n a.. k - n n-l
- L ~1r--........... L b J. 0. 1) (67)
i=l B'(u i ) j=O

n-l
- j=oJ
Lb. uk
-n+J
..

Hence we may calculate the coefficients uk which we need

2
in (65) recursively, starting with un -bn_l/b n , u n +l =
b~_l/b~ - bn_2/b~, etc.
As a simple example, consider the function

F(p) 1 (68)
-2-
P +1
then we have b2 1, b 1 = 0, b O = 1, and hence

k > 2 . ( 69)

Equdtion (65) now gives

(_1)1, t H + l
()()

f(t) '= L (2.t+l)! sin t. (70)

1,=0

General case, Distinct Roots: We now relax the condition

A(p) = 1, whilst still assuming that the roots of B(p) are
distinct. Then we must replace equations (65) and (66) by
392 PART IV: SPECIAL TECHNIQUES

co vk t k
f(t) L
k=O ~
(71)

where
n A(Ct.) Ct. k
vk L 1
BI(Ct.)
1

i=l 1

n Ct. k m
L a j Ct j
1
BI(Ct i ) L (72)
i=l j=O
m
L a J. u k + J·•
j=O
Hence we may again calculate the coefficients of the Taylor
series (71) without calculating the roots of B(p), only we
need a second recursion relation (72) to determine the co-
efficients from the coefficients As a simple
example, we consider

F(p) = ~ . (73)
p +1

Then we have aO a l = 1, and hence

(74)

For f(t), we obtain

(-1) t t 2t + l
co co

f(t) L L (2t+1)!
t=O t=O
(75)
cos t + sin t.

Repeated Roots: If some of the roots of B(p) are repeated,

we may still write the series representation (71) for f(t).
Now, however, the derivations given for the recursion rela-
tions (67) and (72) break down, since some of the B'(Ct i )
are zero. Nevertheless it may be shown that (71) still holds
with the coefficients and determined from these
§2l. Numerical inversion of Laplace transforms 393

recursion relations. In a sense, this result is not surpris-

ing, since any method for determining f(t) without a know-
ledge of the roots Qi ought not to break down if the roots
happen to have a particular property. As a simple example,
consider

1
F(p) = 2 ( 76)
(p-l)
for which it is easily shown that uk k, k > 0 and hence

f(t) ~ ktk
k=O i('"!
( 77)
t et .
More difficult (and realistic) examples have been given by
Longman. 20

Footnotes
1. A more comprehensive survey and evaluation may be found
in B. Davies and B. Martin, J. Compo Phys. (1979), ~, 1.

2. Based on H. E. Salzer, Math. Tables and other aids to

computation, (1955), ~, 164; Journal of Maths. and Phys.
(1958), ~, 89.

3. For example, see STROUD (1974).

4. H. E. Salzer, Journal of Math. and Phys. (1961), ~, 72:

STROUD &SECREST (1966).

5. This argument is given in LUKE (1969), vo~. II.

6. Laguerre polynomials were suggested by F. Tricomi, R. C.

Acad. Nat. dei Lincei 21 (1935), 232 and D. V. Widder,
Duke Math. J., ! (1935), 126. Their practical use was
developed by W. T. Weeks, J. ACM. 13 (1966), 419 and
394 PART IV: SPECIAL TECHNIQUES

R. Piessens and M. Branders, Proc. IEEE, 118 (1971), 1517.

7. W. T. Weeks, J. ACM, !i (1966), 419.

8. Based on R. Piessens, J. Inst. Maths. Applics. (1972),

lQ, 185. In the original paper, Piessens writes

FCp)

where the p(a,a) are Jacobi polynomials. We consider

n
only the special case a = a = -1/2, which forms the main
body of Piessens' papers.

9. We have corrected Piessens formulae for the coefficients

to remove some errors.

10. RIVLIN (1974), p. 47.

11. H. Dubner and J. Abate, J. ACM. ~ (1968), 115.

M. Silverberg, Electron Lett. ~ (1970), 105.

12. F. Durbin, Comput. J., !2 (1974), 371.

13. J. R. MacDonald, J. Appl. Phys. ~ (1964); 3034.

14. J. W. Cooley and J. W. Tukey, Math. Comp., 19 (1965),

297.

15. A very thorough treatment may be found in LUKE (1969),

vol. II.

16. I. M. Longman, Int. J. Compo Math. B, (1971), l, 53.

17. Obviously such a circumstance would cause peculiar diffi-

culties.

18. Some other possibilities for the use of Pade approxima-

tion are discussed in LUKE (1969), vol. II.
§21. Numerical inversion of Laplace transforms 395

19. See I. M. Longman &M. Sharir, Geophys. J. Roy. Astr.

Soc., (1971),~, 299.

20. I. M. Longman, J. Compo Phys. (1972), .!.Q., 224.

396 APPENDIX A

Appendix A: The Factorial Function

Definition: The factorial function is defined by

Jo xa-x
00

a! = e dx, Re(a) > -1, (1)

since it has the value n! when a = n. Analytic continua-

tion can be effected by splitting the integral, i.e., by
writing

a' =f~ x a e - x dx +
foo
1

=to L
n=O n!
dx + rI(a) (2)

L
n=O n! (a+n+l)
+ rI(a)

where rI(a) is an entire function. We see that the function

has simple poles at a = -(n+l), with residues (-l)n/n!,
n = 0, 1, 2, ...

Functional Relationships: The factorial function satisfies

a number of important functional relationships. The most
important three are

a! = a (a-I)! (3 )

a! (-a-I)! -1T/sin(1Ta) (4 )

, ( l)
a.a+ Z ·
, (5)

Derivations of all these results can be obtained by manipula-

tion of the appropriate integrals, and may be found in many

places.
Appendix A 397

Hankel's Integral Representation: Consider the loop integral

fez) = fO
-00
+
t
z t
edt. (6)

where the contour is as in (6.22). For Re(z) > -1, we can

shrink the contour to encircle the branch cut, giving

2i sin (nz) z! (7)

By analytic continuation, the restriction Re(z) > -1 can

be removed, and, after using equation (4) together with the
change of variables z + -(z+l), we have Hankel's integral
representation

1
(8)
~

This result shows that liz! is an entire function, so that

z! has no zeros.

Asymptotic Form: An application of the method of steepest

descents to (8) gives the important result

1
tn(a!) ~ (a+ 2) tn a - a + 21 tn(2n), (9)

a + 00, !arg(a)! < n.

The complete asymptotic expansion, of which we have written

down the first three terms, is derived in (13.33) using the
Mellin transform.

Beta Function: Related to the factorial function, and often

occuring in applications, is the function

398 APPENDIX A

B (p ,q) (10)

Re(p) > 0, Re(q) > O.

Another common form, related by a variable change, is

B(p,q) = f
CO
x
p-l
dx (11)
o (l+x)P+q
In terms of the factorial function, the Beta function has the
value
B(p,q) = (p-l)! (q-l)! (12)
(p+q-l) !
a result whose derivation may be found in many. places. (See
also Problem 1.18.)
Appendix B 399

Appendix B: Rieman's Zeta Function

Definition: The zeta function may be defined by
co
1,; (s) = t -s
l n Re(s) > 1. (1)
n=l '
An integral representation is obtained quite easily by obser-
ving that

fo x s-l
co
e -nx dx = n -s (s-l)!, Re(s) > o. (2)

After summing both sides and reversing the order of summation

and integration, this gives

Re(s) > 1. (3)

Analytic Continuation: Consider the loop integral

~
1 f O+ -zS-l
-dz . (4)
.:;1Tl -z 1
-co e -
For Re(s) 1 we can shrink the contour about the branch
cut to get

- !.
1T sin[1T(s-l)]
(5)

= - !.
1T sin[1T (s-l)] (s-l)! 1,; (s).

Using (A4) this gives

1,;(s) = ~
,
f 0+ s-l
_z_ _ dz. (6 )
21Tl
-co e -z - 1
This shows that 1,;(s) is analytic except possibly where
(-s)! has poles.
400 APPENDIX B

Integer s: When s is an integer, the integrand is single-

valued, and we may replace the contour by a small circle
about the origin and use residue theory. It follows im-
mediately that the integral is zero for s = 2, 3, 4, ... ,
cancelling out the poles of (-s)! there. For s < 1, we
recall the Taylor series
2n
Bn z
!. cot ~ = 1 -
2 2 L
n=l (2n)!
(7)

where Bn are the Bernoulli numbers. Using this in the

present case to construct a Laurent expansion of the inte-
grand of (6), we find that when s = l-m, m = 0, 1, 2, ... ,
we have

z - (m+ 1)
~1 z
- -2 + (8)

Excluding for the moment the case m 0, we can evaluate (6)

immediately to give

1;(0) - 1/2,
1; (- 2n) 0, n = 1, 2, 3, ... , (9)

1;(1-2n) (_l)nB /2n, n = 1, 2, 3, ...

n

When s = 1 (m = 0), the value of the integral in (6) is

- 1, and since (- s) ! has a simple pole with residue -1 at

s = 1, we conclude that 1; (s) has a simple pole with resi-

due +1 at s = 1. Apart from this, it is analytic for all

s. Some other special values are

2
1; (2) 'IT /6,
1; (4) 'lT4/90, (10)

1;' (0) - 1:. R,n 2'IT.

2
Appendix B 401

Riemann's Functional Relationship: By further arguments in-

volving the integral (6), which we shall not reproduce here,
it can be shown that

'ITs 1;(l-s) = Zl-s(s_l)! cos ('ITs/Z)1;(s). (11)

Asymptotic Forms: It is important to know the asymptotic be-

havior of 1; (s) for large s. For Re (s) > 1, (3) can be
rearranged as follows:
oo -x s-l
1; (s) 1 J e x {l _ e- x + e- x } dx
(s-l)! 0 l-e- x

1 +
1
(s-l)!
Joo
0 e
- 2x s-l
x
+ --- -
x
l-e-
1
~} dx

Zl-s
1 + s=-r-

+
1
(s -1) !
Jo
OO { _1
_ _
l-e x
l}
x
e - Zx x s-l dx. (lZ)

This last formula is an analytic continuation to Re(s) > O.

It is easy to bound the integral using the fact that the ex-
pression in the brackets is less than unity, and this gives
the asymptotic information

1;(s) ~ 1, s ~ 00, Re(s) > O. (13)

The picture is completed by the use of Riemann's functional

relationship, giving

-s
1; (l-s) ~ (Z'IT) cos ('ITs) (s-l)! ,

(14 )
s ~ 00, Re(s) > O.
402 APPENDIX C

Appendix C: The Exponential Integral

The exponential integral is defined by
'" -t
El(z) = f _e_ dt. (1)
z t

It is a multi-valued function, since its value along a closed

contour encircling the origin increases by 2ni for each
counter-clockwise circuit, due to the simple pole in the inte-
grand. It is conventional to take the negative real axis as
a branch cut, restricting the contour appropriately. If we
split up the integral as follows:

(2)
_Jz ~
-t
dt _ II dt
o t z t

then the sum of the first two integrals is a constant; on

substituting t = l/u in the first integral and t = u in
the second, we find that this constant is

-y = - Jlo l-e-uu _e- l / u du

'
(3)

which is one expression for Euler's constant, y 0.5772157 ....

Thus we have

E (z) = -,Q,n (z) - y -

l O t
z e-t-l
- - dt J (4)

where ,Q,n(-z) is the principal branch. Since the remaining

integral is an entire function, we see the behavior at the

branch point explicitly; also by using a Taylor series expan-

sion and integrating we obtain the series representation

El(z) = -,Q,n z - y - I (5)

n=l

Asymptotic forms for large z are derived in Section 14.2.

 Bibliography
Abramowitz, M. & Stegun, I. A., 1965, "Handbook of Mathemati-
cal Functions", National Bureau of Standards, Applied
Mathematics Series 55.
Ahlfors, L. V., 1966, "Complex Analysis", McGraw-Hill.
Apostol, T. M., 1957, "Mathematical Analysis", Addison Wesley.
Bellman, R. E., Kalaba, R. E., & Lockett, J. A., 1966, "Numeri-
cal Inversion of the Laplace Transform", American
Elsevier.
Berg, L., 1967, "Introduction to the Operational Calculus",
North Holland.
Bleistein, N. &Handelsman, R. A., 1975, "Asymptotic Expansions
of Integrals", Holt, Rinehart and Winston.
Carrier, G. F., Krook, M., & Pearson, C. E., 1966, "Functions
of a Complex Variab10", McGraw Hill.
Carslaw, H. S. & Jaeger, J. C., 1941, "Operational Methods in
Applied Mathematics", Oxford University Press.
Case, K. M. & Zweifel, P. F., 1967, "Linear Transport Theory",
Addison Wesley.
Dingle, R. B., 1973, "Asymptotic Expansions: their derivation
and interpretation", Academic Press.
Ditkin, V. A. & Prudnikov, A. P., 1962, "Operational Calculus
in Two Variables and its Applications", Pergamon Press.
_ _ _ , 1965, "Integral Transforms and Operational Calculus",
Pergamon Press.
Dodd, R. K., Eilbeck, J. C., Gibbon, J. D., &Morris, H. C.,
1982, "Solitons and Nonlinear Wave Equations", Academic
Press.
Doetsch, G., 1971, "Guide to the application of the Laplace
and Z transforms", Van Nostrand.
Erdelyi, A., 1962, "Operational Calculus and Generalized Func-
tions", Holt, Rinehart and Winston.
Erdelyi, A., Magnus, W., Oberhettinger, F. &Tricomi, F. G.,
1954, "Tables of Integral Transforms", (2 volumes)
McGraw Hill.
Friedmann, B., 1956, "Principles and Techniques of Applied
Mathematics", Wiley.
Gelfand, 1. M. & Shilov, G. E., 1964, "Generalized Functions",
vol. 1, Academic Press.
404 BIBLIOGRAPHY

Goldberg, R. R., 1961, "Fourier Transforms" (Cambridge Tract

No. 52), Cambridge University Press.
Kaplan, W., 1962, "Operational Methods for Linear Systems",
Addison Wesley.
Krabbe, G., 1970, "Operational Calculus", Springer-Verlag.
Krylov, V. I. & Skoblya, N. S., 1977, "A Handbook of Methods
of Approximate Fourier Transformation and Inversion of
the Laplace Transformation", MIR Publishers, Moscow.
Lebedev, N. N., 1965, "Special Functions and their Applications",
Prentice Hall.
Luke, Y. L., 1969, "The special functions and their approxima-
tions" (2 volumes), Academic Press.
Miles, J. W., 1961, "Integral Transforms in Applied Mathe-
matics'" Cambridge University Press.
Morse, P. M., & Feshbach, H., 1953, "Methods of Theoretical
Physics", McGraw Hill.
Murnaghan, F. D., 1962, "The Laplace Transform", Spartan Books,
Washington.
Muskhelishvili, N. I., 1953, "Singular Integral Equations",
Noordhoff.
Noble, B., 1958, "Methods based on the Wiener Hopf technique
for the solution of partial differential equations",
Pergamon Press.
Oberhettinger, F., 1971, "Tables of Bessel Transforms",
Springer-Verlag.
&Badd, L., 1973, "Tables of Laplace Transforms",
Springer-Verlag.
, 1974, "Tables of Mellin Transforms", Springer-Verlag.
Olver, F. W. J., 1974, "Asymptotics and Special Functions",
Academic Press.
Papoulis, A., 1963, "The Fourier Integral and its Applications",
McGraw Hill.
Rivlin, T. J., 1974, "The Chebyshev Polynomials", Wiley-
Interscience.
Smith, M. G., 1966, "Laplace Transform Theory", Princeton.
Sneddon, 1. N., 1966, "Mixed Boundary Value Problems in Poten-
tial Theory", North Holland.
, 1972, "The Use of Integral Transforms", McGraw Hill.
Bib lio graphy 405

Stakgold, I., 1968, "Boundary value problems of Mathematical

Physics" (2 volumes), McMillan, New York.
Stoker, J. J., 1957, "Water Waves", Interscience Publishers.
Stroud, A. H., 1974, "Numerical Quadrature and Solution of
Ordinary Differential Equations", Springer-Verlag.
&Secrest, D., 1966, "Gaussian Quadrature Formulas",
Prentice Hall.
Szego, G., 1959, "Orthogonal Polynomials", American Mathemati-
cal Society Colloquium Publications, vol. 23.
Thompson, W. T., 1950, "Laplace Transformation", Longmans,
Green & Co.
Titchmarsh, E. C., 1948, "An Introduction to the Theory of
Fourier Integrals", 2nd ed., Oxford University Press.
, 1953, "Eigenfunction Expansions Associated with Second
--- Order Differential Equations", Oxford: Clarendon Press.
Tranter, C. J., 1956, "Integral Transforms in Mathematical
Physics", Methuen.
Van der Pol, B., &Bremmer, H., 1955, "Operational Calculus
Based on the Two Sided Laplace Transform", Cambridge
University Press.
Watson, G. N., 1958, "A Treatise on the Theory of Bessel
Functions", 2nd ed., Cambridge University Press.
Whittaker, E. T. & Watson, G. N., 1963, "A Course of Modern
Analysis", Cambridge University Press.
Widder, D. V., 1944, "The Laplace Transform", Princeton.
, 1971, "An Introduction to Transform Theory", Academic
Press.
Wolf, K. B., 1979, "Integral Transforms in Science and Engin-
eering", Plenum Press.
Zemanian, A. H., 1965, "Distribution Theory and Transform
Analysis", McGraw Hill.
, 1968, "Generalized Integral Transformations", Inter-
science.
406 INDEX

Index
Abel, 231 Cauchy integrals, 313-341
Abel's integral equation, 71 Causality, 101, 102
Adjoint problem, 157-159 Chebyshev polynomials, 382-
385
Advanced potential, 187
CO~~lementary error function,
Airy functions, 371-373
Albedo problem, 329-332 Continuity of linear func-
t ionals, 134
Analytic functionals, 143-145
Convergence, of test func-
Anomalous system, 40 tions, 134
Asymptotic expansion, 9 of generalized functions, 139
Asymptotically equal, 8 Convolution equations, 59-67,
70
Barnes, 207
Convolutions, 6, 96, 202
Bernoulli's equation, 115
Cosine transform; see Fourier
Bessel functions, 85, 86, 93, transform
351-371
Coulomb gauge condition, 186
of the first kind, 42, 353-
356 Cramer's rule, 40
integrals involving, 370, Cylinder functions, 373
371
D'Alembert's method, 193
integral representations of,
351-353, 362-364 Delta function, 130, 135
of the second and third Diffraction problems, 182-185,
kinds, 356-362 280-283, 289-299, 307, 309,
332-339
Bessel's equation, 42, 241,
351 Diffusion equation, 47-50, 127,
l89~ 199; see also Heat
Bessel's integral, 93, 94, conauctlon
182, 355, 356
Dirac's delta function; see
Beta function, 397, 398 Delta function

Branch point, Laplace trans- Direct correlation function,

form Inversions involving a, 68
79-82
Dirichlet conditions, 16
Carleman, 288
Dirichlet integrals, 17-19
Case and Zweifel, 330
Discontinuity theorem, 320,
321
Index 407

Distributions, 154 Fourier transform, cont.

Double Laplace transform, relation to Green's func-
192, 193 tions, 277, 278
Dual integral equations, 254- relation to Hankel trans-
266 form, 239
Eigenfunction expansion, 271, relation to Laplace trans-
272, 330 form, 89
Electrical circuit problems, Sine and Cosine transforms,
35,44, 52-54, 57, 58 90, 91
Electron gas, 229, 231 of test functions, 133
Electrostatic problems, 111- in two or more variables,
113, 127, 189, 212, 244, 178-194
250, 251, 254, 255, 265,
266, 285, 286 Fractional integration, 258
Equations of motion, 114 Fraunhofer and Fresnel dif-
fraction, 185
Erdelyi-Kober operators,
258-262 Functionals, 131
Euler's constant, 199, 231, linear, 133, 134
402
analytic, 143-145
Exponential integral, 402
Gaussian quadrature, 374-
Factorial function, 396-398 379
Fermi-Dirac integral, 236 Gellman and Breuckner, 231
Fourier integrals, ascending Generalized functions, 130-154
expansions for, 227-229
addition and multiplication
Fourier series, 246, 247, 270, of, 135
271, 385-386
convergence of sequences of,
Fourier transform 139-143
applications to partial dif- definition, 134
ferential equations, 110-
129 differentiation of, 137-139
definition, 89 Fourier transforms of, 145
of generalized functions, on finite interval, 136
145-148
properties of, 136-143
inverse; see Inverse Fourier
transform properties of, 136-143
properties of, 95-97, 103, regular, 135
104
sequences of, 141-143
singular, 135
408 INDEX

Green's functions, 155-177 Helmholtz's equation, cont.

adjoint, 158, 159 elementary solution, 168-172

as generalized functions, Green's function, 172-174

159-162
Hermite equation, 207, 285
for Helmholtz's equation,
172-174 functions, 207-209, 347-351
integral transforms genera- polynomials, 344-346, 366
ted by, 267- 287
Holder condition, 340
one-dimensional, 155-157
Hydrodynamic equations, 114
for Poisson's equation, 165-
168, 173 Images, 167, 168
symmetry of, 159 Impedance, 54
Hankel functions, 171, 172, Incomplete factorial func-
359-362, 367-369 tion, 220-222

Hankel transform) application Influence function, 29

to boundary value problems,
243, 244 Integral equations, 59-75,
301-307
definition, 237
classification of, 59
inverse, 237
dual, 254-266
properties of, 240-242, 248
Integral transforms generated
relation to Fourier trans- by Green's functions, 267-
form, 239, 240 287
relation to Green's func- Integrals, Fourier, 227-229
tions, 275, 276
involving a parameter, 225-
Heat conduction problems A 47- 236
50, 54-56, 125, 127, 1~0,
191, 193, 211, 212, 243, multidimensional, 230-234
245, 249, 252, 285
Integro-differential equa-
Heaviside tions, 301-307

distortionless line, 54 Inverse Fourier transform, 90

expansion theorem, 77 Sine and Cosine transforms,
91
series expansion, 25, 86, 87
Inverse Laplace transform, 19,
step function, 2, 138 20

Helmholtz's equation, 168-173, asymptotic forms of, 84, 85

286, 287, 289, 307
of meromorphic functions,
76-79
Index 409

Inverse Laplace transform, cont. Lienard-Wiechert potentials,

187, 188
numerical evaluation of, 374-
395 Linear functionals, 133, 134
of rational functions, 20-23 Linear transport theory, 194,
324-329
Taylor series of, 23, 24
Liouville's theorem, 311
Inversion integral for Laplace
transforms, 19, 20, 76-88 Lommel's integral, 238, 253
Ihv-ersion theorem for Laplace Longman, 391
transforms, 15-25
Macdonald's function, 365
Jacobi polynomials, 376
MacRobert, 237
Kirchoff, 182
Maxwell's equations, 185
Kontorovich-Lebedev transform,
276-283 Mechanical problems, 33, 45,
46, 98, 250
relation to Mellin trans-
form, 278 Mellin summation formula, 214
Kramers-Kronig relations, Mellin transform, 195-236
101-103
definition, 196
Laguerre polynomials, 213,
379-382 inverse, 196
Laplace transform properties of, 201-203,
209, 210
asymptotic properties, 8-11
relation to Fourier trans-
definition, 1 form, 195
double, 192, 193 relation to Green's func-
tions, 274, 275
inverse; see Inverse Laplace
transform in summation, 214-223
inversion integral, 19, 20, Meromorphic functions, 76
76-88
Method of images, 167, 168
inversion theorem, 15-25
Milne's equation, 301, 314
properties of, 3-7
Modified Bessel functions,
relation to Fourier trans- 112, 127, 364-366, 369, 370
form, 89
Modified Hankel transform,
Laplace's equation, 110, 126 261
127, 244, 249
Multidimensional integrals,
Laplace's method, 43, 342-373 229-234
Newton's law, 55
410 INDEX

Normal system, 40 Principal value integral, 141

Numerical inversion of Laplace Radiation condition, 124, 125,
transforms, 61, 374-395 128, 181, 182
Ordinary differential equa- Ramanujan, 216-219
tions, Green's functions for,
155-162 Rational approximation, 386,
388-391
Laplace transform methods
for, 26-46 Rational functions, inverse
Laplace transforms of, 20-23
Laplace's method for, 342-
373 Regular generalized functions,
135
Stability of solutions of,
30, 31 Resolvent kernel, 59
Pade approximation, 74, 386- Retarded potential, 185-188
387
Riemann-Hilbert problem,
Pair distribution function, 67 321-324

Parseval relation, 97, 241 Riemann-Lebesgue lemma, 15-17

Partial differential equa- Riemann's zeta function, 399-
tions, Laplace transform 401
methods for, 47-58
Second mean value theorem,
Fourier transform methods 17
for, 110-129
Self-adjoint, 159
Partial fractions, 22
Sine transform; see Fourier
Percus-Yevick equation, 67-71, transform
73, 74
Singular generalized func-
Plemelj formulas, 316-318 tlons, 135

Poisson integral representa- Sommerfeld Diffraction prob-

tion, 362-364 lem, 289-299, 307, 309
Poisson summation formula, Sonine's integrals, 248, 249
108, 142
Spectral analysis, 97-101
Poisson's equation, 162-168,
173 Stirling's series, 219, 220

Potential problems, 110-113, Stretched string, 50-52, 56

125-127, 203, 204, 249,
278-280; Sturm-Liouville problem, 270
see also Heat conduction
and Electrostatic problems Symmetry of Green's function,
159
Power series, asymptotic be-
havior of, 219-222 Taylor series of inverse
Laplace transform, 23, 24
Prime number theorem, 219
Index 411

Test functions, 131-133

Titchmarsh ty~e, dual integral
equations oI, 255
Transfer function, 32
Transmission line, 52-54, 57,
58
Two point boundary value prob-
lem, 155
Ultradistribution, 152, 154
Uncertainty principle, 109
Variation of parameters, 156
Watson's lemma, 9-11
for loop integrals, 82, 83
Wave equation, 50, 168, 189,
191, 193, 280, 289
Wave propagation, 50-54
Weber functions, 360, 361
Weber transform, 245, 252
Weber's integral, 245-247
Wiener-Hopf Technique, 74,
288-312, 313, 324
Zeta function; see Riemann's
zeta function
Applied Mathematical Sciences
cant. from page ii

36. Bengtsson/GhiIlKallem: Dynamic Meterology: Data Assimilation Methods.

37. Saperstone: Semidynamical Systems in Infinite Dimensional Spaces.
38. Lichtenberg/Lieberman: Regular and Stochastic Motion.
39. Piccinini/StampacchialVidossich: Ordinary Differential Equations in Rn.
40. Naylor/Sell: Linear Operator Theory in Engineering and Science.
41. Sparrow: The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors.
42. Guckenheimer/Holmes: Nonlinear Oscillations, Dynamical Systems and
Bifurcations of Vector Fields.
43. OckendonlTayler: Inviscid Fluid Flows.
44. Pazy: Semigroups of Linear Operators and Applications to Partial Differential Equations.
45. Glashoff/Gustafson: Linear Optimization and Approximation: An Introduction to the
Theoretical Analysis and Numerical Treatment of Semi-Infinite Programs.
46. Wilcox: Scattering Theory for Diffraction Gratings.
47. Hale et al.: An Introduction to Infinite Dimensional Dynamical Systems - Geometric
Theory.
48. Murray: Asymptotic Analysis.
49. Ladyzhenskaya: The Boundary-Value Problems of Mathematical Physics.
50. Wilcox: Sound Propagation in Stratified Fluids.
51. Golubitsky/Schaeffer: Bifurcation and Groups in Bifurcation Theory, Vol. I.
52. Chipot: Variational Inequalities and Flow in Porous Media.
53. Majda: Compressible Fluid Flow and Systems of Conservation Laws in Several Space
Variables.
54. Wasow: Linear Turning Point Theory.
55. Yosida: Operational Calculus: A Theory of Hyperfunctions.
56. Chang/Howes: Nonlinear Singular Perturbation Phenomena: Theory and
Applications.