Вы находитесь на странице: 1из 515


Second Edition

California Institufe of Technofogy

o ,
o •
o •
+ ~"/

Addison-Wesley Publishing Company, Inc.

The Advl1nced Book Program
Redwood City, Ca[jfornia· Menla Park, Califomia· Rcading, Massachu~etts
New York • Amsterdam' Doo Mills, Ootaria· Sydney
Bonn· Madrid· Singapore • Tokyo
San luan· WOkingham, United Kingdom

For the last ninetecn ycars a course in mathematical

methods in physics has becn taught, by the authors and by
others, at the California Institute af Technology. It is
intended primarily rar lirst-year physics graduate studcnts;
in reccnt year~ mos! senior physics undergraduatcs, as wc!l
as gradualc students fram other departments, have beco
taking it. This book has evolvcd from lhe notes thal
have becn used in this cour~c.
We make [ittle cffort to tcach physics in this text; lhe
emphasis is on prcscnting mathematical techniqucs that
have provcd to be useful in analyzing prohlcrns in
physics. We assume Ihat the studcnt has beco cxposed
to the standard undergraduate physics curriculum:
mcchanics, e1ectricity and magnctism, introductory
quantum mechanics, etc., and that we are free to se!ect
examples from such arcas.
In short, this is a book about malhematics,for physicists.
80th motivation and standards are drawn fram physics;
that i5, the choice of subjects is dictated by their useful-
ness in physics, and the levei of rigor is intcndcd to reflect
current practice in theoretical physics.
It is assumed that the student has hccome acquainted
with tlle following mathcmatical subjects:
1. Simultaneous linear equations and determinants;
2. Vector analysis, including dilferential operations in
curvilinear coordinates;
3. Elementary differential equations;
4. Complex variables, through Cauchy's theorem.
VI Preface

However, these should not be considered as strict prerequisites. On the

one hand, it will often proflt the student to have had more background, and
on the other hand, it should not be too difficult for a student lacking familiarity
with one of the above sllbjects to remedy the defect by extra work andfor out-
side reading. In fact, the sllbject of dilferential equations is discussed in the
first chapter, partly to begin on familiar ground and partly in arder to treat
some topics not normally covered in an e1ementary course in the subject.
A considerable variation gencrally exists in the amount of preparation
dilfcrent studcnts have in the thcory of functions of a complex variablc. For
this reason, we usually give a rapid review of the subject before studying
contour integration (Chapter 3). Material for such a review is presentcd in
the Appendix. AIso, there are some excellent and reasonably brief mathe-
matical books on the ~ubjcct for the studcnt unfamiliar with this material.
A considcrable nllmbcr ofproblems are inc1uded at the end of each chapter
in this book. These form an important part ofthe course for which this book
is designed. The cmphasis throughout is on understanding by means of
A few remarks may be made about some more ar less unconventional
aspccts ofthe book. In the first place, the material which is presentcd does
no! necessarily fiow in a smooth, logical pattern. Occasionally a new subject
is introduccd without the student's having becn carefully prepared for the
blow. This occur~ not nccessari\y through the irrationality or irascibility of
the authors, but becausc that is the way physics is. Students in theoretical
physics often need considerable persuasion before they will plunge into the
middle of an unfamiliar subject; thi8 course is intended to give practice and
confidence in dealing with problems for which the student's preparation i5
A related point is that there is considerable deliberate nonuniformity in
the depth of presentation. Some subjects are skimmed, while very dctailcd
applications are worked ou! in other areas. Iflhe course is to give practice in
doing physics, the student must be given a chance to gain confidence in his
ability to do detailed calculations. On the other hand, thls is a text, not a
refercnce work, and the material i5 intended to be covcred fully in a year.
It 18 thcrefore not possible to go into everything as deeply as one might like.
The plan of the book remains unaltered in this Second Edition; wc havc
rewritten a number of sections, though only one (Section 14-6) is ncw. We
have tried to make dcarcr certain material which has proved particularly
difficult for students, such as the discussion of the Riemann P-symbol in
Section 7-·3 and the Wiener-Hopf integral equation in Scction 11 5.
A number of new problems have been added at the ends 01' chapters, and
several new detailed examples are included, such as the triangular drum-
head of Scction 8-3 and the recursive calculation of Bessel functlons in
Section 13-3.
Preface VII

Several acknowledgements are in order. The course from whieh this text
evolved. was originaUy based on lectures by Professor R. P. Feynman at
ComeU University. Mueh of Chapter 16 grew out of fruitful eonversations
with Dr. Sidney Coleman. The authors are grateful to Mrs. Julie Cureio for
rapid, accurate. and remarkably neat typing through several revisions.

Pasadena, California
May 1969
~w4 J..:k"mJ:;i..l.t.. J";".",,,, J, ,rJ;w,"1«LJNJ'tl..,
}e OO...md.et1l> ><JL.f.i.. '"""01'''';:'
ll"+ P(")f+Q(~)~~Ol
li, --~~(~) '1 0~) ~,;:- ~ .a .W/;A «1t4<
"."+'Pr,,)u,'-tQel)U,~O '1--- " I
< < , ; ,i\C~,,~.).M,-~,~, I1VMl)VJ~-,arwJ
{ ~'~-+1'C~)~;-tQ(~)~,~O L _

~e. ~. e~) 1ll ~ ,elo) /4tm;/; "'/"'" JtIY'iJ,

: ~z(,,):;~.~;)-}e.·~)<91 M~~';:)-:~}
w.al)a W~.,,"b ' __ -- -,
I~-t-1'C1-J~=o ->;> r~(i)=H*o),,~[-s:.'P~tJJ:t_~J
í ;~"o~o(~)~~ -r '" ]"'? [Ú~i +o 1
I\ ToJ (To
~z )
6(;' )

'0" Mu) J Do"" ~. (>l -tlft

~ZO\)= I,e}\) 5 ~~(u) ~ = ""MO uo 1-& , ,h,
= ~ (f )' (t) ["__ 5'0" 'P(iJ,jt ]
"I'J-,'Cu) ,lu
~, (lO M

I CDe" f" I'ti"

o 4' J )t-~ t~
. . . , ArnÍL'l'Gt "'N' ,

IFltJ: -c. ~,Ct)H,~z(J(;--\ ~fiQ f"''''e. D ,wafj~ arMr nl

""ih I."rl'''-
~ ili WMi <wi!M &k"ri/i..e,.
y "",,,,clt m . W<,*,~ J;P"",*',J;,
[ ~'+1'[tl~'+Q('l~=OJ·
'PQ"a...e '. WM.1u f"!"J:, "''t>If'k "f'&t
.k<ii. .
1' C>ll-\,C'l/C<->O.) ]. {I'C~)-~ hC't-i.l" J(~'*'"-~.J,,,,).
{ QCltl- 4C,,)/(~_~.J2 4(~)-fr.4"("-1,.Jn

[u1lo.J'~II;-PC~)(~-~.Jt+F*) ~rl'tm/l. rw~)

'" LrtJlU.f"'/""- ... ~. di. r-... . . __ ]
u~r'->l.J.\ >~ C,(,,_,.)n_.i:c"(l!.-~.)·+n (c.+o).
[ o:r
-_.- ---
n-o - -
_ _ " _ o _,. _ _ _ _ - - _____ - --

'" ~~CM<Í,~ .~ .."~"Me '" )otfVl" "'_~.. o. uu.4ei

eo..rc,(emhl, .k4 rlI- •• ). ,
c,[/"[h-j)+P.M4.J~c,[h2+r/>.-nô+l.J-o. c.fo "'>

[m) ~_A'+r~_-A)~~1._ 0_ \ EcM.t:;o._...,)icl~.

c.ofi,,'fYit.L -&; r~_ •• )A+n(O_V2,3._ .. )·
CO [C MO)' +( />.-A l(Mn)+ 4. J+.{; Co_IcLhCh+n- ~ )+'lk ] = ]

1 ~
-cnHA+nl++-rcn-~ [ h(J+n-l)+4d~o.
11. A" A2 ~x~ _4ilÀ Ji~,,~,J.. ~ i"~ '" .
F(A.J~ A~+(~.-Oh,"+l. ~o ] { ô,H,,=-h+1 lJ
{ 2
HA,) EA. + (j>.- ~ JA,' 1.~ o
(A, ...,.).
ô,." = 1. . C,""I"'"iiLkr .
'J+t. Lh(A.+n-il+h J~o
a. f'C,;,+ ao-L

6, f(A,'O)+ t.- b._..J Pk ql]=O (I,,+n-t)+

F(.II,+n) +0 (11=4,2, '3 J - - ) . L1,-A,z +n C!; F('ó~+"J+f(.4d=o Cn'" 0, ~j2,3J_ • .)_

1I,(.)~(.-1, ).1':2:. .,[.-l.)' ~." -Bvr.1NY'

,. O"
(j gL h ( A.L
A_-:(. 1'O,~, 2, "
) .J'U,.. i. +-.
[~-~. 1 '..LI. n [.-~.)
0·0. '

ri) =
""-r"""''''''' .
L<L«i o"d, , .krw. i,(;I,)~~.rl!J. ú""'" "",Iik m Cm=!,2, 3,. )
0Iif/J :meLê h.-6,~m, ~I\cl F(.1,+"')~ F{A,j=O "! mdvJa.
MriA~ .u puitri pro<Lce hnJ.1MÀ ~tiC>. !/z) .
Ü'dM,,~ cJ"i &I-a J"", r4 í", reM akJ d.ijcern:h .'<f<'''r{iBvr
J1;(j, "''''."''" '" )I~ l""Ii~.
~ -/.~- ~~: 'm_ ~--J ry [~~~:~ ~~,_~--P~+~--J (m = o, _, 2, ~ ~ ... ).
'l-'rirr/lC( f<I4J·Âd-í.e .ik : ___
! IL
}"Ci) (~::~;-.;,-:(~) =(~-~,)', >: ~"(*- '.l'
____ _
______ • _ _ _ _ _ • _________ ._ •• _ _ _ _ _. _ .

Mcd1 (Ih • jl.oWtie '"'..t.1í~ 11 "(~,)=ao"'o.


h dw.~ o.Jv.ll e lf ~~fta ""'_J~"ifliÚ'c" '

1,(.)~~.(~lGf~.) ~ff[- 51'(.)., Jh J
._- ---- - - - --- -- ----

(V(·)"jji{ .. ),"f[-j"P(t)'.Jü~f~~;(.)~xp[-U(;: J~Jd.~

~ 5(.-i,f~·-"·ó(.)d.~ SC.-t,fm-'3CitlJt. "",d" ."
I, ~(")=s..~o(.-t,)
i n
• ~.-~C •• )-; I'

í V(')~ Sc,,- •• r"-' QC<l'" =

i, -',m-l,C.- •• )+Cli,-",f
<J TYl
•. ) . 11

I - . --_ .. - . ···· ..· ---·-1


~/·)=Jm~.(.)t(.-.,J+c.- •. ) ·ZbnC'-~.l
.. --- ---------

ri'" o





1-1 Solution in Closed Form 1
1-2 Power-Series Solutions 13
1-3 Miscellaneous Approximate Methods 22
1-4 The WKB Method 27
References 37
Problems 38


2-1 Convergence 44
2-2 Familiar Sedes 48
2-3 Transformation of Series 50
References 55
ProbJems 56


3-1 Elementary Methods 58
3-2 Use af Symme,ry Arguments 62
3-3 Contour Integration 65
3-4 Tahulated Integrais 75
3-5 Approximate Expansions 80
3-6 Saddle-Point Methods 82
Reference5 90
Problems 91
x Contents


4-1 Fourier Series 96

4-2 Fourier Transforms 101
4-3 Laplace Transforms 107
4-4 Other Transform Pairs 109
4-5 Applications of Integral Transforms 110
References 119
Problems 120



5-1 Conformai Transformations 123

5-2 Dispersion Relations 129
References 135
Problems 135


6-1 Linear Vector Spaces 139

6-2 Linear Operators 141
6-3 Matrices 143
6-4 Coordinate Transformations 146
6-5 Eigenvalue Problems ISO
6-6 Diagonalization of Matrices 158
6-7 Spaces of Infinite Dimensionality 160
References 163
Problems 163


7-1 Legendre Functions 167

7-2 Bessel Functions 178
7-3 Hypergeometric Functions 187
7-4 Confluent Hypergeometric Functions 194
7-5 Mathieu Functions 198
7-6 ElIiptic Functions 204
References 210
Problems 210



8~ I Examples 217
8-2 General Discussion 219
8-3 Separation af Variables 226
8-4 Integral Transform Methods 239
8-5 Wiener-Hopf Method 245
References 252
Problems 253


9-1 Simple Examples af Eigenvalue Problems 260
9-2 General Discussion 263
9-3 Solutions af Boundary-Value Problems as
Eigenfunction Expansions 266
9-4 Inhomogeneous Problems. Oreen's Functions 267
9-5 Green's Functions in Electrodynamics 277
References 282
Problems 282


10-1 Conventional Nondegenerate Theory 286
10-2 A Rearranged Sedes 292
10-3 Degenerate Perturbation Theory 293
References 296
Problems 296


11-1 Clasification 299

11-2 Degenerate Kernels 300
11-3 Neumann and Fredholm Series 302
11-4 Schmidt-Hilbert Theory 305
1l~5 Misçellaneous Devices 311
11-6 Integral Equations in Dispersion Theory 316
References 318
Problems 318
XIl Contents


12-1 Euler-Lagrange Equation 322
12-2 Generalizations of the Basic Problem 327
12-3 Connections between Eigenvalue Problems and
the Calculus of Variations 333
References 341
Problems 341


13-1 Interpolation 345

13-2 Numericallntegration 349
13-3 Numerical Solution of Differential Equations 353
13-4 Roots of Equations 358
13-5 Summing Series 363
References 368
Problems 368


14-1 Introduction 372

14-2 Fundamental Probability Laws 373
14-3 Combinations and Permutations 375
14-4 The Binomial, Poisson, and Gaussian
Distributions 377
14-5 General Properties of Distributions 380
14-6 Multivariate Gaussian Distributions 384
14-7 Fitting ofExperimental Data 387
References 395
Problems 396



15-1 Cartesian Tensors in Three-Space 403

15-2 Curves in Three-Space; Frenet Formulas 408
15-3 General Tensor Analysis 416
References 421
Problems 421
Contents Xl11



16-1 Introduction; Definitions 424

16-2 Subgroups and Classes 426
16-3 Group Representations 430
16-4 Characters 433
16-5 Physical Applications 440
16-6 Infinite Groups 449
16-7 Irreducible Representations of SU(2), SU(3),
and 0(3) 457
Reference5 467
Problems 468


A-I Functions of a Complex Variable. Mapping 471
A-2 Analytic Functions 477
References 483
Problems 483




We begin this chapter with a brief review af some af
the methods for obtaining solutions af an ordinary
differential equation in closed formo Solutions in the
form af power series are discussed in Section 1-2, and
some methods for obtãining approximate solutions are
treated in Sections 1-3 and 1-4.
The use af integral transforms in solving differential
equations is disCllSsed Jater, in Chapter 4. Applications
af Green's funchan and eigenfunction methods are
treated in Chapter 9, and numerical methods are
described in Chapter 13.


The arder and degree af a differential equation refer to the derivative af
highest arder after the equation has been rationaJized. Thus, the equation

d'y3 + X ~ + xZy = O
is of third order and second degree, since when it is rationalized it contains
the term (d 3 yfdx 3 )'l.
2 Ordinary Differential Equations

We first recall some methods whieh apply partieularly to first-order

equations. If the equation ean be written in the form
A(x) dx + B(y) dy _ O (I-I)
we say the equation IS separable; the solution 15 found immediately by


l_x 2 -
dy, dx
F-y'+ JI_x i -
sin- 1 y+sin- 1 x=C
or, taking the sine of both sides,
xJI- y2+y../I_X2 =sinC= C'

More general1y, it may be possible to integrate immediately an equation

of the form
A(x, y) dx + B(x, y) dy - O fI-J)
If the left side of (1-3) is the differential du of some funetion u(x, y), then
we ean integrate and obtain the solution
Sueh an equation is said to be exact. A necessary and sufficient conditioD
that Eq. (1-3) be exact is
àA àB
oy = ax

(x + y)dx + x dy= O (1-5)


The solution is
l-I Solution in Closed Form 3

Sometimes we can find a function À(x, y), such that

.I(A dx + Bdy)
is an exact differential, although A dx + B dymay not have been. We
caIl such a function À an integrating factor. One can show that such factors
always exist (for a first·order equation), but there is no general way offinding
Consider the general linear first-order equation

~~ + f(x)y ~ .(x) (1-6)

Let us try to find an integrating factor À(x). That is,

.I(x)[dy +1\x)y dx] ~ .I(x)g(x) dx
is to be exact. The right side is immediately integrable, and our criterion
(1·4) that the left side be exact is

dA(x) ~ A(x)f(x)
This equation is separable, and it~ solution

A(x) ~ exp (f f(x) dx] (1-7)

is the integrating factor we were looking for.


xy' + (1 + x)y = eX
y'+ (
1+ X) e"
x y=-X

The integrating factor 1S exp{J[(l + x)jx]dx} = xe x

Now oU! equation is exact; integrating both sides gives

xe Xy = elx dx = !elx + C
é C_li:
2x x

4 Ordinary Differential Equations

One ean often simplify a differential equation by making a judicious

ehange of variable. For example, the equation
y' ~f(ax + by + c) (1-9)
becomes separable ir one introduces the new dependent variable,
v = ax + by+ c
As another example, the so-called Bernoulli equation
y' + f(x)y ~ g(x)y" (1-10)
hecomes linear if one sets v = yi-". (This substitution becomes .. obvious"
if the equation is first divided by y".)
A functionf(x, y, ... ) in any number of variables is said to be homogeneous
of degree r in these variables if
f(ax, ay, ... ) ~ a'f(x, y, ...)
A first-order differential equation
A(x, y) dx + B(x, y) dy ~ O (1-11)
is said to be homogeneous if A and B are homogeneous functions of the same
degree. The substitution y = vx makes the homogeneous equation (1-11)


y dx + (2JXY - x) dy ~ O (1-12)
y=vx dy=vdx+xdv
'x dx + (2xj~ - x)(v dx + x d,) ~ O
2,'" dx + (2j~ - l)x dv ~ O
This equation is clearly separable and its solution is trivial.

Note that this approaeh is related to dimensional arguments familiar

from physics. A homogeneous funetion is simply a dimensionally con-
sistent funetion. if x, y, ... are ali assigned the same dimension (for example,
length). The variable v = y/x is then a .. dimensionless" variable.
This suggests a generalization of the idea of homogeneity. Suppose
that the equation
is dimensionally eonsistent when the dimensionality of y is some power m of
the dimensionality of x. That is, suppose
A(ax, a"y) ~ a'A(x, y)
B(ax, a"'y) = á- m +1 B(x, y)
l-I Solution in Closed Form 5

Such equations are said to be isobaric. The substitution y = vx'" reduces

the equation to a separable one.


xy2(3y dx + x dy) - (2y dx - x dy) = O (1-14)

Let us test to see if this is isobaric. Give x a .. weight" 1 and y a weight
m. The first term has weight 3m + 2, and the second has weight m + 1.
Therefore, the equation is isobaric with weight m = -!.
This suggests introducing the "dimensionless" variable v = y~. To
avoid fractional powers, we instead let
Equation (1-14) reduces to
(3v - 2)y dv + 5v(l - v) dy = O
which is separable.

An equation of the form

(ax + by + c) dx + (ex + Iy + g) dy = O (1-15)
where a, ... , 9 are constants, may be made homogeneous by a substitution
x=X+C( y~ y+p
where C( and pare suitably chosen constants [provided aI #: be; if aI = be,
Eq. (1-15) is even more trivial).
An equation of the form
y-xy'~f(y') (1-16)
is known as a Clairaut equation. To solve it, differentiate both sides with
respect to x. The result is
y'U'(y') + xl ~ O
We thus have two possibilities. Ir we set yN = O, Y = ax + b, and substitu-
tion back into the original equation (1-16)givesb = I(a). Thus y = ax + f(a)
is the general solution. However, we also have the possibility
I'(y') +x ~ O
Eliminating y' between tbis equation and the original differential equation
(1-16), we obtain a solution with no arbitraI}' constants. Such a solution
is known as a singular solution.
6 Ordinary Differential Equations


y = xy' + (y')2 (1-17)

This is a Clairaut equation with general solution
y=cx+c 2
However, we must also consider the possibility 2y' + x = 0, which gives
x' + 4y ~ O
This singular solution 1S an envelope of the family of curves given by the
general solution, as shown in Figure l-I. The dotted parabola is the singular
solution, and the straight lines tangent to the parabola are the general solution.

There are various other types of singular solutions, but we shall nol discuss
them here. See Cohen (C4) or Ince (12) for more complete discussions
and further references.
Next we review some methods which are useful for higher-order differential
equations. An important type is the linear equation with constant coefficients:
Iff(x) = 0, the equation is said to be homogeneous; otherwise it is inhomoge-
neous. Note that, if a linear equation is homogeneous, the sum of two
solutions is also a solution, whereas this is not true if the equation i8 in-
The general solution of an inhomogeneous equation is the sum of the
general solution of the corresponding homogeneous equation (the so-called
complementary function) and any solution of the inhomogeneous equation

Figure 1-1 Solutions or the differential equation (1.17) and their
I-I Solution in Closed Form 7

(the so-called particular integral). This is in fact true for any linear differen-
tial equation, whether or not the coefficients are constants.
Solutions of the homogeneous equation [(1-18) with f(x) = O] generally
have the form

Substitution into the homl geneous equation gives

lfthe n roots are mi' m2 •... ' m~ the complementary function is

(C j are arbitrary constants)

Suppose two roots are the same, m 1 = m2 • Then we have only n - 1

solutions, and we need another. Imagine a limiting procedure in which
m2 approaches mi' Then
e"''''' - e""'"

is a solution, and, as m2 approaches m l , this solution becomes

= xe""'"

Thís is our additional solution. If three roots are equal. ml = m2 = m3'

then the three soIutions are

and so on.
Arguments involving similar limiting procedures are frequently useful; see,
for example, the discussion on pp. 17 and 18.
A particular integral is generally harder to findo If f(x) has only a finite
number of linearly independent derivatives, that is, is a linear combination
of terms of the form x", eU, sin kx, cos kx, or, more generally,
X'r cos rlX X'e"'~ sin rlX

then the method of undetermined coefficients is quite straightforward. Take

for y(x) a linear combination of f(x) and its independent derivatives and
determine the coefficients by requiring that y(x) obey the differential equation.


y" + 3y' + 2y = e" (1-19)

8 Ordinary Differential Equations

Complementary function:
m2 +3m+2=O
m= -1,-2
Y = Cle-X + cze- 2;r;
Particular integral: Try y = Ali". Substitution into the differential equation
(1-19) gives

Thus, the general solution is

y "'" ié + cle-x + c2e- 2x
If [(x), ar a term in [(x), is also part of the complementary function,
the particular integral may contain this term and its detivatives multiplied
by some power of x. To see how this works, solve the above example
(1-19) with the right-hand side, eX, replaced by e-"'.
There are several formal devices for obtaining particular integraIs. If D
means d(dx, then we can write our equation 0-18) as
(D - m,)(D - m,)" ·(D - m,)y ~f(x) (1-20)
A formal sotution of (1-20) is
y ~ =-""",,'-"''-;-;;:-"'""'7
(D - m,)'" (D - m,)
or, expanding by partial fraction techniques,
f(x) f(x)
y = AI
D ml
+ ... + An -,f-""-,-

What does [(x)(D - m) mean? It is the solution of (D - m)y = f(x),

which is a first-order linear equation whose solution is trivial [see (1-6)1.
Alternatively, we can just peel off the factors in (1-20) one at a time.
That is,
(D - m,)(D - m,) ... (D - m,)y ~ ~"'--­ (1-22)
D - ml

We evaluate the right side, divide by D - m 2 , evaluate again. and so on.

Next, we consider the very important method known as variation of
parameters for obtaining a particular integral. This method has the useful
feature of applying equally well to linear equations with nonconstant co-
efficients. Before giving a general discussion of the method and applying
it to an example, we shall digress briefly on the subject of osculating parameters.

l-I Solution in Closed Form 9

Suppose we are given two Iinearly independent functions Yl(X) and Yz(x).
By means af these we can define the two-parameter family af functions
Now consider some arbitrary function y(x). Can we represent it by ao
appropriate choice af Cl and C 2 in (1-23)1 Clearly, the answer in general
is no. Let us try the more madest approach af approximating y(x) in the
neighborhood of some fixed paint x = Xo by a curve af t1te family (1-23).
Since there are two. parameters at our disposal, a natural choice i8 to fit
the value y(xo} and slope y'(xo} exactly. That is, cI and c2 are determined
from the two simultaneous equations
y(xo} = C1Yl(XO) + Clh(xo)
y'(xo} = CIY~(XO) + C2 Y2(XO}
The c1 and c 2 obtained in this way vary from paint to paint (that i8, as Xo
varies) along the curve y(x). They are ca11ed osculating parameters because
the curve they determine fits the curve y(x) as c10sely as possible at the
point in question. 1
One can, of course, generalize to an arbitrary number N of functions
YI and parameters Cj • One chooses the c, to reproduce the function y(x)
and its first N - 1 derivatives at the point xo.
We now return to the problem of solving linear differential equations.
For simplicity, we shall restriet ourselves to second-order equations. Con-
sider the inhomogeneous equation
p(x)y· + q(x)y' + ,(x)y ~ ,(x) (1-25)
and suppose we know the complementary function to be
C1Yl(X) + c2 Y2(X)
Let us seek a solution of(1-25) ofthe form
Y = U1(X)Yl(X) + U2(X)Y2(X) (1-26)
where the u/(x) are functions to be determined. In order to substitute
(1-26) ioto (1-25), we must evaluate y' and yW. From (1-26),
Before going on to calculate y", we observe that it would be coovenient to
impose the condition that the sum of the last two terms in (1-27) vanish,
that is,

1 Latin osculari = to kiss.

, ,
, Ji'+p()/,)~:+Q(l)~,~ O
i . Jj (~4) ~z)=
I ~. ,Z ,
,'* O•
, 1"+Hn<+Q(')~<~Q
I" ' ,Z
, ~
2 " "I
"(»,_" (~/R(tJ~,(O Jt+u (1'(' k(O,.(+j Jt
ti" (~J MO J~ j tJ(O
10 Ordinary Difl'erential Equations

This prevents second derivatives of the UI from appearing, for now


and difl'erentiating gives

Note that the condition (1-28) nol anly simplifies the subsequent algebra
but also ensures that Ui and Ul are just the osculating parameters discussed
above; compare (1-26) and (1-29) with (1-24).
The rest af the procedure is straightforward. If (1-26), (1-29), and (1-30)
are substituted into the original difl'erential equation (1-25), and we use
the fact that Y1 and Yz are solutions af the homogeneous equation, we

p(x)(u~y~ + u; Yí) = s (1-31)

(1-28) and (1-31) are now simultaneous linear equations for the uj, whose
80lution i5 straightforward. The student i8 urged to complete the 80lution;
see Prablem 1-26.


Cansider the ditrerential equation

x:Zy"-2y=x (1-32)
The complementary function, that is, tbe general solution of
x 2 y" - 2y =O (1-33)
is most easily found by noting that y = x'" is a natural trial solution.:Z Sub-
stitution into (1-33) shows that m = 2 ar - I, so that the complementary
function is
, c,
CI X +-

Therefore, we are led to try to find a solution of(I-32) oftbe form


1 In facl, }' = x" is an obvious lrial solulion for any linear differenlial equalion of lhe
l-I Solutioo in Closed Form 1i

We ditferentiate, obtaioiog
1 + X 2 u,
y '=
2XU - 2:
1 u2 1,
1 + - U2

aod impose the cooditioo

2, 1,
XU 1 +-U 2 = O (1-35)



1 ,

Substituting (1-34) and (1-37) back ioto the ditferential equation (1-32),
we obtaio

Solviog this together with (1-35) gives

, 1 , X
u1 = - U2 = --
3x' 3
1 x'
U1=--+C, u2 = --+C,
3x 6

Then the general solution of (1-32) is

X 2 C2
Y = - -
+ CIX +-

Changes of variable sometimes help. We shaU discuss one general trans-

formation which is particularly useful. Consider the second-order linear
y. + f(x)y' + g(x)y ~ O (1-38)

The substitution
will give another linear equation in v(x) which may be easier to solve for
12 Ordinary Differential Equations

certain choices of p(x). The resulting equation is

V" + (2 ~ + I)V' + (Pll + I:' + gP)v = O (1-40)

If we know one solution of the original equation (1·38), we can lake p to

be that solution and thereby eliminate the term in v from (1-40). This
is very useful because we can then find the general solution by two straight·
forward integrations.
Alternatively, we may choose


and eliminate the fust derivative term of (1-40). This procedure is helpful
as an aid to recognizing equations, and it is especially useful in connection
with approximate methods of solution.


Bessel's equation is
x2y~ + xy' + (x 2 _
n2)y = O (1-42)
The substitution y = u(x)p(x) with p(x) = X- t/2 gives

uI! + (1- n2x~!)u =0 (1-43)

This equation, with no first derivative term, is very convenient for finding
the approximate behavior of Bessel functions at large x, for example, by the
WKBmethod (Section 1-4). The details areleftas an exercise (Problem 1-41).

In conclusion, we shall briefiy mention some other devices for solving

differential equations. For details, refer to a text on differential equations.
If the dependent variable y is absent, let y' = p be the new dependent
variable. This lowers the order of the equation by one.
If the independent variable x is absent, let y be the new independent
variable and y' = p be the new dependent variable. This also lowers the
order of the equation by one.
If the equation is homogeneous in y, let v = log y be a new dependent
variable. The resulting equation will not contain v, and the substitution
v' = p will then reduce the order of the equation by one.
If the equation is isobaric, when x is given the weight 1 and y the weight m,
the change in dependent variable y = vX". followed by the change in indepen·
dent variable u = 10g x, gives an equation in which the new independent
variable u is absent.

1-2 Power-Series Solutions 13

Always watch for the possibility that ao equation is exact. Also consider
the possibility offlnding ao integrating factor. For example, the commonly
occurring equation y" = f(y) can be integrated immediately ir both sides
are multiplied by y',


Before discussing sedes solutions in general, consider a simple (although
noolinear) example:



Then (1-44) becames

Equating coefficients af equal powers af x gives

C1 = -!c~

c3 =i-tcO Cl
c4 -- -lZC
"+"1 12 CO, etc.

Suppose, for example, we want the solution with y = 0, y' = 1 at x = O.



This method af solution is a very useful aoe, but we have beeo careless
about justifying the method, establishing convergence af the series, etc. We
now briefly outline the general theory of series solutions of linear differential
equations. 3
Consider the equation

"This theory is discussed more fully, for example, in Copson (C8) Chapter X, or
Jeffreys and Jeffreys (J4) Chapter 16.
14 Ordinary Differential Equations

If fo(x), ... ,J,.-l(X) are regular (in the sense of complex variable theory;
see the Appendix, Section A-2) at a point x = x o , Xo is said to be an ordinary
point of the differential equation. Near an ordinary point, the general
solution of a differential equation can be written as a Taylor series whose
radius of convergence is the distance to the nearest singularity of the differen-
tial equation; by a singularity, of course, we mean a point which is not
The Taylor series is an ordinary power series

Y = L c",(x -
Xo)'" (1-47)

whose coefficients c'" may be convenient1y found by substitution in the

dilferential equation, as in the example above.
If Xo is not an ordinary point, but (x - XO)fn-l(X), (x - xo)2fn_2(X), ... ,
(x -xo)nfo(x) are regular at Xo, Xo is said to be a regular singular point of
the differential equation. Near a regular singular point, we can always
find at least one solution of the form

y = (x - XoY L
c",(x - Xo)m with Co ::F O (1-48)

where the exponent s of the leading term will not necessarily be an integer.
The series again converges in any circ1e which includes no singularities
except for xo.
The algebraic work involved in substituting the series (1-47) or (1-48)
in a differential equation is simplified if Xo = O. Thus it is usually convenient
to fust make a translation to Xo as origin, that is, to rewrite the equation in
terms of a new independent variable, z = x - Xo.
If a point is neither an ordinary point nor a regular singularity, it is an
irregular singular point.


We first consider an example of expansion about an ordinary point:

Legendre' s differential equation is

(I - X2)yN - 2xy' + n(n + l)y = O (1-49)

The points x = + 1 are regular singular points. We shaU expand about

the ordinary point x = O. Try

y = L c",x'" = Co + c1x + C2 X 2 + ".

1-2 Power-Series SoIutions 15

Substitution into the differential equation (1-49) gives

n(n + 1)
c2 = - 2 Co

c3 = 6 C1

n(n + 1)[6 - n(n + I)]

c4 =- 24 Co

,s-- [2 - n(n + 1)][12 - n(n + I)] ,

120 1

The general recursion relatian is

C;+2 i(i + 1) - n(n + 1) (i + n + 1)0 - n)
--~ (I-50)
cI O+l)(i+2) (; + I)(H 2)
Note that this is a two-term recursion relation, that is, it relates two co-
efficients only. Their índices differ by two. These facts could have been
noticed at the beginning by inspection af the ditferential equation.
The general solution af Qur differential equation (1-49) is therefore
2 4
y = co [1 - n(n + 1) x + n(n +
l)(n - 2)(n + 3) x
- + ... ]

+ c1 [x-<n -l)(n +2);:

+ (n - I)(n x' - + ...]

+ 2)(n - 3)(n + 4) Si (1-51)

For physical applications, it is necessary to consider the behavior af the

infinite series in (l-51) near the singular points x = + 1. Ir we let i_ 00 in
the recursion relation (1.50), we see that

Thus, if we go farther and farther out in either series, we find it more and
more resembling a geometric series with x 2 being the ratio of successive terms.
The sum of such a series clearly approaches infinity as x 2 approaches 1. 4

4 This discussion is incomplele. Infinite series are discussed in more delail in Chapter 2,
and lhe series of (I-51) are studied as a special case on p. 47. The series in
facl "almosl" converge as x 2 approaches 1; Iheir sums go lo infinity logarilhrnically
[y '" In(l - x 2 »).
16 Ordinary Differential Equations

In cases of physical interest, however, we often require solutions y(x)

which are finite for -1 :::;; x:::;; + I. There are two ways we can arrange this:
(1) Let C1 = O, and choose n to be one of the integers ... , -5, -3, -1,
0,2,4,.... Then the first series in (l-51) terminates, and the second is
(2) Let Co = O, and choose n to he one of the integers ... , - 6, -4, - 2,
I, 3, 5,.... Then the second series in (l-51) terminates, and the first is
We see that if we wish a solution of the Legendre differential equation
(1-49) which is finite on the interval - 1 :::;; x :::;; + I, n must be an integer. The
resulting solution y(x) is a polynomial which, when normalized by the
condition y(x) = I, is known as a Legendre po/ynomial Pn(X). 5

As an example of the prablems that can arise at a regular singular point,

consider Bessel's equation
x 2 y" + xy' + (x 2 - m 2 )y = O (1-52)
This has a regular singular point at x = O. Thus, a solution of the form

is guaranteed to exist, and we can see fram the differential equat;on that
a two-term recursion relation will be obtained. If we substitute into the
differential equation, the coefficient of x' is
CO(S2 - m2) =O
This is called the indiciai equation. Its roots are s = + m.
Next consider the coefficient of x'+ 1. It is

Thus C 1 = O except in the single case m =tl/2, s = _ = -1/2, and in

that case we can set C 1 = O, since the terms thereby omitted are equivalent
to those which make up the other solution, with s = _ = + 1/2.
We therefore confine ourselves to even values of n in the sum, writing
y = x±m(co + C2X2 + C4X4 + ...)
The recursion retation is easily found to be
Cn + 2 -1 -1
~= (s + n + 2)2 m2 - (~n-+C-=2)"(72,-'+-n-+'""""'2) (1-54)

~ Note that Pn(x) = P _I _.(x).

y = co r [ l - 4ú ; 1) + 4 x 8(s :41)(S + 2) - + ... ] (1-55)

Suitably normalized, this senes is called a Bessel function; we shall discuss

Bessel functions in more detail in Chapter 7.
If m is not an integer, we have two independent solutions to our equation,
namely, (I-55) with s = +m. If m is an integer (which we may assume
positive ), we can only choose s = +m; for s = -m, al1 denominators
in (I-55) beyond a certain term vanish. If we multiply through by ($ + m)
before setting s = -m, to canceI the offending facton, we get a multiple
Df the solution with s = + m, as may be readily verified.c~)
This is a situation reminiscent of our difliculty on p. 7 in obtaining indepen-
dent solutions Df a differential equation with constant coeflicients. The
resolution of the difliculty is quite analogous. We begin by letting y(x, s)
denote the series

y(x, s) ~ x' ~::C"x'

with Co = I, c1 = O, and ali other coefficients given by (1-54), but without

making use ofthe indiciaI equation (l-53). That is, we set
y(x, s) ~ x' [1 - ( 2
)( 2 )

+ (s + 2 + mXs + 2 -

m)(s + 4 + m)(s + 4 - m) -
+ ...
If we abbreviate by L the Bessel differential operator,
d' d
L = Xl_+1
x_+(x1 _ m 2 )
dx dx
then the series Ly(x, s) will contain only a term in x", because the recursion
relation makes the coeflicients of ali higher powers vanish. We obtain
[compare (I-53)J
Ly(x, s) ~ (s - m)(s + m)x'
We see (again) that s must equal ±m in order that y(x, s) be a solution
Df Bessel's equation Ly = O. However, if m is a positive integer, we have
remarked that
[(s + m)y(x, s)). __ m
18 Ordinary Differential Equatiotis

is a constant multiple of y(x, m), and we must find a second solution. To

do so, consider the result obtained by substituting (s + m)y(x, s)"into Bessel's
equation. We get

L[(s + m)y(x, s)] = (s + m)2(s - m)x'

The derivative of the right side with respect to s vanishes at s = - m.



is a solution of Bessel's equation. and is in fact the second solution we were

looking for.


Let llS find a second solution of Bessel's equation for m = 2.

~ x' ]
y(x, s) = x' [ 1- s(s + 4) + s(s + 2)(s + 4)(s + 6) - + ...
,[ s +2 2 x (l-58)
(s + 2)y(x,s) = X (s+2)- s(s+4)x + s(s + 4)(s+ 6)

x' ]
- '(H 4)(H 4)(H 6)(H 8) + - ...
Remembering that (djds)x' = x' In x and
~ (UV .. ') = uv ... (U' + v' + ... _ w' _ ... )
dx W'" W··· U v w

we obtain

-a [(, + 2)y(x, ,)) ~ (, + 2)y(x, ,) In x

+x"l- [ s(sH2(1 1- 1),x
+ 4) s + 2 s s + 4

+ s(s + 4)(s + 6)
(1 1 1).
- '$ - s + 4 - s + 6 x - + ...
1-2 Power-Series Solutioos 19

Setting s = -2, and observing that [(s + 2)y(x, s)]~," -2 =- /6 Y(X, 2),
we obtain

- --y(x
16 •
2)lnx+x- 2 1+-+-+···
4 64
(X'x') (1·59)

This solution and y(x, 2) are two independent solutions of Bessel's equation
with m =;: 2.

Let us now consider the differential equation

dx' + (E -x')~ - O (1·60)

This is the Schrõdinger equation for a one-dimensional quantum-mechanieal

harmonic oscillator, in appropriate units. If ODe tries a direct power series
solution of (1-60) about x = O, one obtains a three-tenn recursion relatioD.
These are a little inconvenient, 6 so that it is advantageous to look for a
transformation of variables which willlead to a simpler equation.
A trick that often helps in such a situation is to .. faetor out" the behavior
near some singularity or singularities. Where are the singularities of this
equation 1 There are none in the tinite z plane, but we must digress for a
moment on singularities at infinity.
Consider the differential equation
y' + p(x)y' + Q(x)y - O (1-61)

X=O is an ordinary paint if P and Q are regular there

Recall that x = O is a regular singular paiot if xP and x 2 Q are regular
( there
Let z = l/x. The equation becomes ,


p(z) = P(x) q(z) = Q(x)
Then x = 00 is an ordinary point or a singularity of Eq. (1-61) depending
on whether z = O is an ordinary paint or a singularity of Eq. (t -62). That

1& Three-term recursion relations can often be trcatcd by continued fraction methods

(see Scction 7-!$).

J~ JWimf • -~ k
'1'(,,)<1::. h ~-Q
{ 'Z
• Q(>o)-
,?, 9~"
- n
20 Ordinary Differential Equations

is, x = 00 is an ordinary point if 2x - x 2 p(x) and x 4 Q(x) are regular there,

and x = 00 is a regular singular point if xP{x) and x 2 Q(x) are regular there. (I!
This concludes our digression.
Using these criteria, we see that our differential equation (l ~60) is highly
singular at x = 00. For large x, the equation is approximately

~2~ _ x 2", =O (1-63)

and the solutions are roughly


in that, if we substitute either function into the differential equation'(l~63).

the tenns which are dominant at infinity cancelo
In quantum mechanics, physically acceptable solutions must not become
infinite ato lxi_ 00; therefore. try

This change of variable does nol ensure the desired behavior at infinity,
of course, and- we shall still have to select solutions y(x) which give this
behavior. In fact, we may expect in general y(x) _ e'" so as to give the
divergent solution, 1/1_ e hZ !2.
The differential equation (1~60) becomes
y" - 2xy" + (E -1)y ~ O (1-66)

[If we write E - I = 2n, (l~66) is known as the Hermite differential equation.]

We can obtain the general solution of this equation in the form of a power
series which will converge everywhere, and the recursion relation for t,he
coefficients will contain only two terms. The recursion relation is
em+-: (2m + 1) - E

em (m + 1)(m + 2)

and the solution is

y ~
c, 1 + (1 - E) - + (1 - E)(5 - E) - + ... X']

+ c, [X +(3 - X'
E)- + (3- E)(7 -E)- + ...
3! 5I
X'] (1-67)

If E = 2n + I where n is an integer, -ODe series terminates after the term

in ~ (the even or odd series, deperu:ling on n). The resulting polynomial,

1-2 Power-Series Solutions 21

suitably normalized, is called a Hermite po/ynomiaJ1 af arder n, H,,(x).

The other series ean be eliminated by setting its coefficient, Co ar C1 , equal
to zero, and the resulting solution (1-65) will go to zero at infinity.
Ir either series af (1-67) does not terminate, its behavior at large x is deter-
mined by the terms far out where the recursion relatian is approximately


'. m
Thus either series behaves for large x like eX2 , and I/J -+ e"zll as expected.
A solution l/t, whicb remains bounded as x -+ + 00, is therefore aaly possible
ir E = 2n + 1 with integral n; that is.
l/t = "',,(x) = H,,(x)e- xZ / 1
E=E,,=2n+ 1 (1-68)
This is another example ofbow boundary conditionsmay imposerestrictions
00 the acceptable values for a constant appearing in a differential equation.
The acceptable solutions I/I~ are called eigenfunctions of the differentiaI
operator _(d 2 fdx 2 ) + x 2 belonging to the eigenvalues E" (see Chapter 9).
As another example of series solutions, we consider briefty the associated
Legendre equation
(1 - x )y" - 2xy' + [n(n + 1) - m 2]Y
=O (1-69)

Note that when m = O, this reduces to Legendre's equation. The origin

is an ordinary point, and x = ± 1 are regular singularities.
A straightforward attempt at a power series 'solution about x = O again
yields a three-term recursion relation. Let us try to factor out the behavior
at x = ± 1. Let x = + I + z and write the approximate form of the differen-
tiaI equation valid for Izl 4: L Near either singularity it is
zy" + y' - 4z Y = O

This has solutions y = zhl2 and z-,"12, the former being well behaved if,
as we may assume without 10ss of generality, m ~ O.
We are therefore led to make the change of variable y = v(1 - X2)'"f2 in
(1-69), thus factoring out the behavior at both singularities simultaneously.
The equation becomes
(I - x')v' - 2(m + I)xv' + [n(n + I) - m(m + I)]v - O (1-70)

'For a precise definition of HK(X) sce Eq. (7-103).

~-p { -~
- ml2

M 1:<.
*J v_ p 1r -l

O-lõt.)Il"-2-l!AI+n(O+~)1l =0 7-
(A-~·).'M+')-Z(m+A) .. "("'+'\)+[n(n+~)-m(m+4)J.("')~O " v~ .,,)
22 Ordinary Differential Equations

with a straigbtforward seties solution about x = O. The recursion relation

for the coefficients is

C. H
(r + m)(r + m + 1) -
n(n + 1)

- (r +m- n)(r + m + n + 1)
(r + I)(r + 2)

Again we obtain even and odd series solutions for v(x), both of wruch behave
Iike (I - x 2)-'" near x = ± 1 if they do ~ot terminate. 8 A bounded soIution
exists onIy if n and m are such that one series terminates after some term x.
From (1-71), the condition is

(n - m) = r = an integer;;.: O (1-72)

Usually in physical appIications n and m are both integers. It may be

vermed that y is then simply a constant times

(I - x')'''(~)·
P (x)


which is called an associated Legendre function. These functions will be

ruscussed in more detail in Chapter 7.

One can often get a qualitative picture of the solutions of a differential
equation by means of a graprucal approach.
For example, consider the first-order equadon


We may draw little lines in the xy-plane. each indicating the slope of the
solution passing through that peint, as iIIustrated in Figure 1-2. The
approximate shape of the solutions is obvious.
As a second example, consider the second-order, nonlinear equation
yJr=X_ y 2 (1-75)

a Scc Problem 1-32.

1~3 Miscellaneous Approximate Methods 23

I ,
- -
, - - --
- - -- " I I I , x

- - I
, I

Figure 1-2 Slopes af the SOIUtiODS of (1-74)

If we draw the parabola x = y2, as in Figure 1-3, we see that yH is positive

inside the parabola, and negative outside af it. If we limit ourselves to
solutions starting near the origin and going toward the right, we can identify
several types which are sketched in Figure 1-3:

I. Solutions oscillating about the upper branch af the parabola.

2. A criticaI nonoscillating solution approaching the upper branch of the
parabola fram above.
3. Solutions going to y = - 00.
4. A criticai nonoscillating solution approaching the lower branch af the
parabola fram above.

The power-series solution (1-45) we found eaclier [withy(O) = 0, y'(O) = 1]

is af type I. Ao approximation to solution 2 is given by expression (1-80).

Figure 1-3 Several types af solution for the equation y" = x _ yZ

24 Ordinary Differential Equations

V,- ------;?---


V,- - ~---"

Figure 1-4 A potential V(x) for the Schrodinger equation

As a third example, consider the one-dimensional Schrodinger equation

for a particle of mass m in a potential V(x):
d 2 !/t 2m
dx' ~ - h' [E - V(x)J~ (1-76)

If E > V(x), then !/t JIJJ < O, and ljJ curves toward the x-axis; that is, !/t has an

oscillatory or "sinusoidal" character. If E < V(x), then ljJ"jIJJ > O, and !/t
curves away from the x-axis; that is, !/t has an "exponential" character. If
we impose the boundary condition that !/t remain finite everywhere, then a
solution with unbounded exponential behavior is unacceptable.
Suppose V(x) looks as shown in Figure 1-4.
If E > V3 , ali solutions are oscillatory everywhere and ali are acceptable,
that is, none blow up at infinity.
If V2 < E < V3 ' .. most" solutions blow up at x + 00. However,

if we start just right. we can find a solution that faUs off exponentially as
x + 00. This defines the acceptable solution uniquely (except in amplitude)

and, in particular, fixes the phase in the left-hand region where the solution is
oscillatory. Such a solution is illustrated quaIitativeJy in Figure l-S.

V, ,~ ------
-,y'- - - - -
, I

, x,
" , ,, ~~~x
v, ""
Figure 1-5 A physica1ly acceptable solution of the Schrõdinger
equation for V2 < E < V3

1-3 Miscellaneous Approximate Methods 25

Ir Vi < E < V2 , things are reaIly difficult. 1{1 behaves "exponentially"

at both ends. If we adjust it so that it does not blow up at the left, it is
almost sure to do so at the right. Only for certain values Df E ean satis-
factory solutions be faund. These are the eigenvalues; compare (1-68).
Ir E < Vl' no satisfactory solutions existo

Another method, which ean often provide useful approximate solutions.

consists af experimenting with the equation, that is, dropping small terms,
iterating, and 80 foeth.



Let's look for a solution near x ---+ 00. Assume y > O. Thus, we try
-~O-J.yR;a (= constant)
Now put this guess back into (1-77)


Iterate once more.

dxR;exp [ (a- 2a
-2x ,O''')]

Y R; a -
2a -
1 (
1 ) 44x
X + 40 e- + ... etc. (1-78)

If th:! procedure seems to be converging, it is probably giving us a solution.

Sometimes, though, it does not converge.


26 Ordinary Differential Equations

(See the graphical discussion on pp. 22-23.) Try y:::.:: x. Substitutiog J

into the right side of (l~79) gives y" ~ O, which in tum implies y = ax + b,
with a and b constant. This does not look anything like our first try. The
mistake, of course, was substituting our first guess ioto the "Iarge" term y2,
rather than iota the "small" term y". Doing it the other way gives
y2 = x_ y":::.:: x- (J xy = x{1 + !X- SI2 )

Iterating once more gives

y = J"-; + !x- 2
- lV'é X - 912 + ... (1-80)
This is the solution (2) enumerated 00 p. 23.

Altematively, we may try a differeot method. Write

Substituting into the differential equation (1.79) gives
_!X- 3/ 2 + I}" = -21}J-; - I}2 (1-82)
We neglect the last term, but the equation is still complicated. We
may try further neglecting either the first or second termo If we assume
1""1 ~ 12"j~l, we find

and we obtain our previous solution (1·80). We should, of course, check

the consistency of the solutioo and the approximation made in finding it;
namely, that
(ix·')" ~ 2(tx·')Jx
This is satisfied for large x.
3 2
We can obtain another solution by neglecting the X- / term of Eq. (1-82)
instead of I}". That is, if we assume
1""Il> 1-'x·'''1
we obtain the differentiaJ equatioo

"'+2Jx"~0 (1-83)
An approximate solution ofthis equation may be found by the WKB method,
which we discuss in the next section. The result is

I} :::.:: ~
sin (4 J 2: x SI4 + ô)
1-4 The WKB Method 27

which gives an oscillatory solution of type 1 listed on p. 23. Again, we

should go back and verify the consistency ofthe assumption
1'(1 p Itx- 312 1
Why can we not neglect the first terro 00 the right of(I-82)1


The WKB method provides approximate solutions of differential equations
of the form
dx 2 + f(x)y = O (1-85)

provided f(x) satisfies certain restrictions discussed below, which may be

summarized in the phrase "f(x) is slowJy varying." RecaJl that aoy linear
homogeneous second-order equation may be put in this form by the trans-
formation (1-41). The one-dimensional Schrõdinger eguation is of this
form and the method was developed for guantum-rnechanical applications
by WentzeJ, by Kramers,9 and by Brillouin, whence the name. The method
had been given previously by Jeffreys.lC
The solutions of Eg. (1-85) with f(x) constant suggest the substitution

The differential equation becomes

( 1-87)

If we assume f/J" small, a first approximation is

~'~ +J! (1-88)

The condition of validity (that f/JH be "small ") is

W'I '" ~ :;[ "" Ifl (1-89)

From (1-86) and (1-88) we see that 1/J7 is roughly lj(2n) times Ofie "wave-
length" or one "exponentiallength" of the solution y. Thus the condition
of validity of our approximation is simply the intuitively reasonable one that
the change inf(x) in one waveJength should be small compared to Ifl.

9 Sec, for example, Kramers (K5).

10 See Jeffreys and Jeffreys (J4).
28 Ordinary Differential Equations

A second approximation is easily found by iteration. From (1-88),

"" ~ +,1"1'1'
Substituting this estimate for the small term 1J" into (1-87), we obtain

F- •, f'

,,(x) ~ + f J f(x) dx + ~ Inf

The two choices of sigo give two (approximate) solutioos which may be com-
bined to give the general solution

y(x) ~ (f(:))''' (" exp [; JT(x) dx] + '- «p [-; J f(x) f dX]) (1-90)

where c+ and c_ are arbitrary constants.

We havc thus found an approximation to the general solution of the
original equation (1-85) in any region where the condition of validity (1-89)
holds. The method fails if f(x) changes toa rapidly or if f(x) passes through
zero. The latter is a senous difficulty since wc oftco wish to join ao

'.".". J(x)
' .•..• I , - ....
'" "".
;."1":. \
Yl(X).,,'" I ••••• "
..... ' I \ ...-
'. , x .-
... \ :'
.'. \
\ .,'
\ "
Y 2 (x)" ........... ',

'" '-

Figure 1--6 Graph of f (x) and two exact solutions of the equation
(1-85); one of them. YI(X). being the special solution
which decreases to zero on the left
1-4 The WKB Method 29

oscillatory solution in a region where f(x) > O to an "exponential" one in a

region whcref(x) < O [see, for example, the discu~~ion of Eg. (1-76)].
We shall investigate this problem in some detailll in order to derive the
so-called connection formulas relating the constants c + and c _ of the WKB
solutions on either side of a point where f(x) = O.
Suppose f(x) pas~es through zero at Xo and is positive on lhe right as
shown in Figure 1-6. Suppose further that f(x) sati~fies the condition of
validity (1-89) in regions both to the left and right of Xo so that any spedfic
solution y(x) may be approximated in these regions by
x ~ Xo ,f(x) < O:

y(x) ~ 4/ a
v - i(x)
ex p [+ JXo,,/-=" f(;j dx]

+ b "P [- r"/ _I(;) dX] (1-91)

~/-f(x) 'x
x ~ xo,f(x) > O:

y(x) ~ ~/;(x) cX P [ + i (,,//ex) c/x]

'\ j(x)

where the symbols vr and ~r mean positivc real roots throughout. If

(x) is real, a solution \Vhich is real in the lert region will also be real on
the right. Thi~" reality condition" states that ir a and b are real, d = c*.
Qur problem is to "connect" lhe approximations on either side of Xo
so that they rder to the same exact solution, that is, to find c and d ir we
know a and b and vice versa. To make this connection, we need to u~e an
approximate solution which is valid along the entire lenglh of some path
connecting the regions of x on either side of xo, where the WKB approxima-
tÍons are valido One procedure,12 followed by Kramers and by Jelrreys,
is to use a solution valid on the real axis through Xo (see Problcm 7-36 and also
lhe discussion at the end of Section 4-5),
A second procedure, used by Zwann and by Kemble,13 is to avoid the
real axis near Xo and use a path encirc1ing Xo in lhe complex plane, along

11 lf we were using fine print in this book, thL' material from here to Eq. (1-122)
would be 50 displayed. The eonnection formulas (1-113) and (1-122) are, however,
important and should be understood even by lhose who wish to ,kip lhe inl~rvemng
" See Schiff (S2) Scction 34, for a discussion of this approach, and numerous reference~.
13 See Kemble (KI) Seetion 21.
30 Ordinary Differential Equations

which the WKB approximations themselves remain valido We shall follow

this second procedure because it provides not only the connection formulas,
but also the means to estimate errors in the WKB approximations. Also,
the techniques employed are more generally instructive.
The question of errors is important because we wish to use the approxi-
mate solutions over a large range of x covering many "wavelengths" or
"exponentiallengths." Thus, one might worry that errors would gradually
accumulate and that the approximate solution would get badly "out of
phase," for example, in the region where it is oscillatory.
Define W KB functions associated with Eq. (1-85) as

W±(x) = [f(x)r 1/4 ex p [ + i (J f(x) dX] (1-93)

We must consider these to be functions of the complex variable x, and

we should draw suitable branch cuts in the x plane to avoid ambiguities
from the roots af f(x) (see the Appendix, Section A-I). These functions
(1-93) satisfy (exactly) a differential equatian which we can find by differ-

W', ~ ( 1') I
+ij!(x)-.! w,

I J"
w±+ [ f(x)+----5 (1'
- )'] W+=O (1-94)
4! 16! -
If we define


then W±(x) are exact solutions af

W'; + [J(x) + g(x)]W, ~ O (1-96)
and approximate solutions of
y" + f(x)y ~ O (1-97)
provided Ig(x)l.:g If(x)l. Equation (1-97) is regular at xo, whereas Eq.
(1-96) has a singularity there. Correspondingly, y(x) is regular at Xo but
the W ± are singular there.
We now define functions IX±(X) by
y(x) = IX + (x)W+ (x) + 1X_(X)W_(x) ( 1-98)
y'(x) = 1X+(X)W'+(x) + 1X_(X)W'_(x) (1-99)
where y(x) is a solution of (1-97). The IX± are just osculating parameters
as discussed on p. 9 in connection with the variation of parameters method.
1-4 The WKB Method 31

The WKB approximation corresponds simply to setting 0:+ and 0:_ cqual to
Solving (1-98) and (1-99) for o:±(x),
yW'-- - y'W_
(X_ =- W+W'_ _ W'+ W_

where the denominators are the Wronskian of W+ and W _. The Wronskian

is a constant, as may easily be shown from the dilferential equation (1-96).
From the explicit forms (1-93), it is
W+ W'_ - W~ W_ = -2i
so that

a+ ="2'( yW ,_ - y 'W_
) (1-100)

Differentiating, and using the differential equations (1-96) and (1-97) to

eliminate yfl and W± , we obtain
da± j i
dx = +"2 (yW'~ - y"W1') = +"2 g(x)yW+ (1-101)

or, remembering (1-93) and (1-98),

d;; = + ~ [f~;~;1/2 {(X± + (X1' ex p [ +2i (vi dX]}

f(x) (1-102)

One or the other of these expressions (1-101) or (1-102) may be used to

estimate the error which may accumulate in the WKB approximation
over a long range of x.


The WK.B functions associated with the equation

y" + xy = O (1-103)
for x ~ O are

W±(x) = x- t / 4 exp ( + j (X112 dx) = X- 1/ 4 exp (+ i-ix 3i2 ) (l-104)

These satisfy exactIy the equation (1-96):

W' + (x - -/fiX-2)W = O
That is,
) =-lóX
32 Ordinary Differential Equations

We may write the solution of (l-103) in the form (1-98)

y(x) = IX +(x)W+ (x) + 1X_(X)W_(x)
At large x, a general solution of (l-103) is accurately described by the
WKB approximation:
as X""" 00 (1-105)
so that IX+ ...... (Aj2)él, cc ...... (A/2)e- id for x ...... 00. We wish to investigate
the errar in this WKB solution when it is carried in to small values of x,
this errar being_measured by the deviation of IX+(X) and CL(X) from the
constant vahies above. If the êt±(x) do not change much, we have appraxi-
mately, fram (1-102),

d;: ~ +~ (_ 156x-2)X-I!2[~ eW +; e'Fld exp ( +2j~X3!2)]

and AIX±, the changes in 01± fram Xl to infinity, are given by

óa:± ~ ~ ( ' dlX± dx

Af2 A:x, dx


The complicated second term in the bracket is less important than the
first (why?), and we see that the relative errar IAIX± I/IIX±I is only of the arder
of 10 to 20 percent even for a value of X as low as Xl = I, and at larger Xl
the errar becomes very small. In particular, no significant errar in phase
accumulates even after an arbitrarily large number of" wavelengths."

We now return to the problem of estimating the variation in the functions

IX±(X) in the general case. Integrating Eq. (1-102) along a path r in the
complex plane from Xl to x 2 ,

AIX±=f dlX±dx=+~f j(X) IX±dx+~f j(X) __ IXt=(W')dX (1-107)

r dx 2 r f(x) 2 r f(x) W±
We shall consider only paths such that

f• J g(x)
ds <>i 1 (1-108)

where ds = Idxl is the element of Iength along r. Kemble calls such a

path a good palh.
Along a good path, the first term on the right side of (1-107) does not
contribute a significant change Llcx:±-. The second term can be significant
only if IW./W± I ~ 1. That is, along a good path 0:+ can change only in
a region where "its" function W+ is small in magnitude compared to W_.
1-4 The WKB Method 33

Even there, cx+ cannot change if cx_ is zero or nearly zero. Similar statements
hold for cx_ if we interchange + and - everywhere.
In the regions where IW_/W+I pior IW+/W_I p I, the coefficient of
the small function may vary markedly, a behavior known as Stokes'
phenomenon. 14 (This variation has no appreciable effect on lhe function
y(x) in this region, so that there is no inconsistency with the fact lhat the
WKB approximation, with constant cx+ and cx_, is supposed to be valid!)
The above results will now be used to derive the WKB connection formulas.
We recaJl the definitions (1-93) and draw a branch line from Xo to + 00
along the real axis. We take roots [j(x)r/ 4 and U(x)]112 to be the positive
real roots on the top side of this branch cut. Then, for x < Xo,

W±(x) = e- i ("/4)[ - f(x)]- JI4 ex p [ + ("J= /(x) dX] x<x o (1-109)

Referring to 0-91) and (1-92), we have

cx+ ..... ae i ("14) (l_ -+ be i (,,14l
(l+ ..... C cx_ ..... d
Inordertolocatethe"Stokes'regioos",wherelW+/W_1 p 10rIW_/W+1 p 1,
we notei S that near Xo ,f(x) ::::: K(x - xo), and the boundaries ofthese regioos
will be qualitatively similar to those for f(x) = K(x - xo). These boundaries
are where J~oJ f(x) dx is real since then IW+ I = IW_I· Forf(x) = K(x - xo),
ifwe put x - Xo = re i9 ,

f..• Jf(x) dx = K 1/2t(X - x o)3/2 = iKI12r3/2e,(3812l (1-1 li)

This is real in directioos e = o, -i1t', ~1t', and it is easy to see from (1-93) which
of W+ or W_ dominates in each region between these boundaries.
A map of the Stokes' regions is given in Figure 1-7, and a good path r
is 5hown connecting the regioos x ~ Xo and x p Xo 00 the real axis, where the
WKB solutions are to be connected.
It is now easy to find the connection formula for the special solutioo
Yl(X) which has a decreasing exponential behavior to the left of xo, as shown
in Figure 1-6. This solution has a = O, in the notation of (1-91), ar, from
(1-110), cx+(x) = O at the start of path r. Thus, ( L (x) remains constant
along r even though IW+/W_I p 1 in region 2. cx+ remains essential\y zero.
In region 1, IW_/W+I ~ 1 50 that CL remaios constaot although cx+ may

14 For a rather amusing description of Stokes' diseovery of this phenomenon, as well as

one of the earliest diseussions of the conneetion problem, see R. E. Langer (L4).
l ' The reader will note a tacit assumption: wc assume that a radius for lhe path r ean
be ehosen which is small enough for lhe assumplionf(x):::::: K(x - xol to be valido and at
the same time large enough Ihat r is a "good" path.
34 Ordinary Differential Equations

region 1

region 2
branch line

region 3

Figure 1-7 Complex x-plane showing the Stokes' regions around the
point Xo wbere f(x) vanishes

(aod does) change. Since cc is constant ali along r, we find from (l-llO)
The reality coodition gives c = dI/< if b is real.
The conoectioo formula for the solution Yl(X) with decreasing exponential
behavior for x < Xo is thus (normalizing b = I)

~ ex1p [- rJ I(x) dx]_ 2 ~ ccs[!"

J I(x) dx -~]
'/(x), ' I(x)" 4

Reversal ofthe arrow is delicate. We koow ooly that the growiog exponen-
tial term 00 the left will be absent for some phase near -1t/4 00 the rigbt.
Next, consider a more general solutioo Yl(X) such that a oF O. Since the
basic equation is linear, the coefficients in expressions (1-91) and (1-92) must
be Iinearly related:

d= Ca+ Db

1-4 The WKB Method 35

We have found in (1-112) that

B = e- i (fI{4) and D = e i (fI/4) (1-115)

The reality condition d = c·, ir a and b are real, says

C =A* and (1-116)
the second af which we already know.
Another general relation may be obtained from the fact that the Wronskian
YIY~ - Y2 y't, af two independent solutions af our equation yH + f(x)y = O,
is a constant. (To see this, just differentiate the Wronskian.) Write

Remembering that y' = C(+ W~ + cc W~ and the Wronskian

ODe finds

Evaluating this constaot at x <a; Xo [from (1-110)],

where OI and b1 are the coefficients a and b in the WKB approximation

(1-91) fOf Yl(X), and a2 and b 2 are the coefficients fOf yix). Similarly,
on the other side, where x ~ xo,

el (ft/2) = i = AD - BC= AB'" - A"'B = 2i Im AB* (1-119)

But B = e- I(If/4 J, 50 Ae i (lI/4) = R + ij2 where R is an undetermined real

constant. Collecting our results,
A = Re-I(~/4) + !e l (lI/4) B = e- I ("/4)
C = Rel(~/4) + !e- i (lI/4) D = e i (lI/4)

and because R is undetermined, we still cannot write c and d in terms of

a and b. However, this is as far as we can proceed; R is indeterminate.,
Another way of expressing this indeterminacy is to observe that the asympto-
tic form of a solution like Y2(X) in Figure 1-6 for x < Xo does not fix yix)
for x > Xo . What about the reverse problem of writing a and b in terms of

36 Ordinary Differential Equations

c and d1 Equations (1·114), with the constaots A, B, C, D given by (1·120),

are easily inverted with the result
a = e- I (1I/4)c.+ e i (1I/4)d
b = [1e 'C1I/ 4) _ Re-i(~/4)]c + [!e-I(~/4) _ Rei(~/4)]d

Thus we can find a but not b, which might be expected since cc(x) ftuctuates
wildly in regioo 2 of Figure 1-7 except in the special case ct:+ = O. We might
also expect this result from the fact that, when a #= O, the value of b is im·
material since the decreasing exponential is negligible compared to the
increasing one. Omitting b for this reason, we can write a second formula
connecting the two expressions (1-91) and (1-92) for y(x). It is conventional
to choose a real solution by setting c = fé/! and d = !e-Jo/I. Then

provided cp is not toa near - n/4. The arrow cannot be reversed, as we

have seen, since the phase cp cannot be determined if we only know how
much large exponential is present.
It is very easy to remember the connection formulas (1-1l3) and (1-122)
by drawing qualitative pictures like Figure 1-6. Consider first Yl(X) with
the decreasing exponential to the left of Xo. To the right of x o , Yl(X) looks
like a cosine wave for which the phase is between -n/2 and O when x = xo.
We remember the phase is -n/4. Furthermore, the amplitude ofthe cosioe
wave 00 the right "Iooks" larger than that of the exponential on the left.
We remember it is twice as large. These comments reproduce the result
(1-113). A similar mnemonic for remembering (1-122) results from con-
sidering a solution Y2(X) which is 90° out of phase with Yl(X) in the right-h~nd
region. By this same procedure, it is a straightforward matter to write
down connection formulas similar to (1-113) and (1-122) for the case where
the slope off(x) is negative at X o instead of positive.
We shall conclude by giving an example of the use of the WKB method
in quantum mechanics. Consider the Schrõdinger equation for a particle
in a potentia! well

with V(x) as shown in Figure 1-8.

2m . (Positive for a < x < b
[(x) ~ f,' [E - V(x)] IS negative for x < a, x > b
References 37


a b x

Figure 1-8 A typical one-dimensional potential well

If ifi(x) is to be bound.ed for x < a, then in a < x < b our connection formula
(1-113) teUs us

~(x) '" (E AV),,, co, U:J~7 (E - V) dx - i) (1-124)

where A is an arbitrary constant. If ifi(x) is to be bounded for x > b. then in

a < x < b similarly

ifi(x):::=: (E _ BV)I/4 cos (f'" J2m

r;r (E - 4
V) dx - ") (1-125)

where B is an arbitrary constant. These two expressions must be the same,

whichgives the condition
f• .j2m(E - V) dx ~ (n + !)n* (1-126)

This result is very similar to the Bohr-Sommerfeld quantization condition

ofpre-1925 quantum mechanics.

There are many good books 00 the subject of ordinary differential equa-
tions. Two small but remarkably complete ooes are Burkill (B9) aod
Ince (11). The latter is a condensed version of the treatise (12) by the
same author. Another treatise, excellent and well worth looking tbrough,
even if it is slightly old-fashioned, is Forsyth (F5).
Modero texts on this subject tend to place more emphasis on numerical
solutions, integral transform methods, and special techniques for oonlinear
equations, than do the older texts mentioned above. Severa) good examples
are Birkhoff and Rota (B3); Golomb and Shanks (G5); and Rabenstein (RI).
The WKB metbod and its applications to quantum mecbanics are discussed
in any good book 00 quantum theory, such as Schiff (S2) Section 34;
38 Ordinary Differential Equations

Merzbacher (M4) Chapter 7; or Landau and Lifshitz (L2) Chapter VIL

Note that the last-named book refers to the WKB method as the "quasi-
classical case."

Find the general solutions of Problems l-I to 1-20:
1-1 x 2y' + y2 = xyy'
, xJl+y2
1-2 Y ~
yJ 1 + x

y' = ('x-'+'--y)'"
1-4 y' + ycosx = t sin 2x
1-5 (1 - x 2)y' - xy = xy2
1-6 2x J y' 1 + 1 + 4x 2Y
= J
1-7 y" + y'2 + 1 = O
1-8 y" = ri'
1-9 x(l - x)y" + 4y' + 2y = O
1-10 (1 - X)y2 dx - x 3 dy = O
1-11 + y + X 4y4(?"'" = O
1-12 (I + x 2 )y' + y = tan~l x
1-13 x 2 y'2 _ 2(xy - 4)y' + y2 = O (general solution and singular
1-14 yy" - y'2 _ 6xy2 = O
1-15 x 4 yy" + x 4 y'2 + 3x 3 J'Y' - 1 = O
1-16 x 2 y" - 2y = x
1-17 y" -2y" - y' + 2y = sin x
1-18 yio + 2y" + y = cos x
1-19 y" + 3y' + 2y = exp [e>]
1-20 a 2 y"2 = (I + y'2)3
1-21 The differential equation obeyed by the charge q on a capacitor C
connected in sedes with a resistance R to a voltage

= Vo(~re-tlf

Find q(t) ir q(0) ~ O.

Problems 39

]·22 In the activation af ao indium foil by a constant slow neutron flux,

the number N af radioactive atoms obeys lhe equation
dt = AN. - J.N

where N. is the constaot number after "saturation." Find N(t)

ir N(O) ~ O.
1-23 Find the general solution af

A(x)y"'(x) + A'(x)y"(x) + Y«X) ~ O

A x)
where A(x) is a known function and y(x) is unknown.
1-24 Find the general solution af
xy" + 2y' + n2 xy = sin wx
Hint: Eliminate the first derivative termo
1-25 Note that y = x would be a solution af
(1 - x)y" + xy' - y = (1 _ X)l

ir the right side were zero. Use this fact to abtain the general soJution
af the equation as given.
1-26 Consider the differential equation
y' + p(x)y' + q(x)y ~ O
00 the interval a ~ x ~ b. Suppose we know tWQ solutions, Yl(X)
and Yz(x), such that
Yt(a) = O
y,(b)" O
Give the solution of the equation
y' + p(x)y' + q(xly ~ f(x)
which obeys the conditions y(a) = y(b) = O, in the form
y(x) ~ i"•
G(x, x')/(x') dx'

where G(x, x'), the so-called Green's function, involves only the
solutions Yl and Y2 and assumes different functional forms for x' <:; x
and x' > x.
IIlustrate by solving
y" + k 2 y = /(x)
y(a) ~ y(b) ~ O
40 Ordinary Differential Equations

1-27 Find the general solution of the differential equation

,,3 3
xy +- Y= 1+ x
in real form (no i's in answer).
1-28 Consider the equation

+ ~ dy + (K + ~ _ l(l +2 l»)y =
d y O O~x~oo
dx 2 X dx X x
where I = nonnegative integer
Find all values of the constant K which can give a solution which is
finite on the entire range of x (including 00). An equation like this
arises in solving the Schrõdinger equation for the hydrogen atom.
Hint: Let y = v/x, then "factor out" the behavior at infinity.
1-29 For what values ofthe constant K does the differential equation

" (1 K)
y-4+~y= O (O<x<oo)

have a nontrivial solution vanishing at x = O and x = oo?

1-30 Find the values of the constant k for which the equation
xy" - 2xy' + (k - 3x)y = O
has a solution bounded on the range O ~ x < 00.
1-31 A solution of the differential equation
xy" + 2y' + (E - x)y = O
is desired such that y(O) = I, y(oo) = O. For what values of E is this
1-32 By considering Eq. (1-71) for r-+ 00, verify that v(x) indeed behaves
like (I - x 2 )-m as x -+ + 1.
1-33 Bessel's equation for m = O is
x 2 y" + xy' + x 2 y =O
We have found one solution

Show that a second solution exists of the form

Jo(x) In x + Ax2 + Bx 4 + Cx6 + ...
and find the first three coefficients A, R, C.

Problems 41

1-34 Consider the differential equation

xy" + (2 - x)y' - 2y = O
Give two solutions, ORe regular at the origin and having the value 1
there, the other af lhe form
- + A(x) In x + B(x)
where A(x) and B(x) are regular at the origino Three terms af each
senes will suffice.
1-35 Find ao approximate expression (say three terros) fOf large x for the
solution af
" 1
Y = (x2 + y2)2
which approaches the x axis as x -+ + 00.
1-36 Consider the differential equation
yH _xy+y3=O
for large positive x.
(a) Find ao oscillating solution with two arbitrary constants.
(b) Find a (nontrivial) particular nonoscillating solution.
1-37 Find ao (approximate) oscillating solution af
y" = (y _ X)2 _ é(Y-X)

Vou may let (X denote the real number such lhat

(IX ~ -0.57)
1-38 Consider the difTerential equation
dy = e/x
(a) Suppose y(1) = O. Give a series expansion for y(x) which is
valid for x near l. Neglect terms of order (x _ 1)4.
(b) Suppose y(xo) = + CI) (xo > O). Give an approximate expression
for y(x) which is useful for x slightly less than xo.
1-39 Use the WKB roethod to find approximate negative values of the
constant E for which the equatiofl
-+ [E - V(x)Jy ~ O
has a solutiofl which is finite for ali x between x =- 00 and x = + 00,
inclusive. V(x) is the function shown below.
42 Ordinary Differential Equations


-a x

1~ Find a good approximation, for x large and positive, to the solution

of the equation

y" _ ~ y' + (~ + XIIZ)y = O

x 4xz
Hint: Remove first derivative termo
1-41 Obtain an approximate formula for the Bessel function J",(x) by
the WKB method and give the limiting form of this expression for
large x(x ~ m). Do not worry about getting the constant in front
correct. Vou may assume m ~ 1-.
1-42 Consider the differential equation y" + xy = O.
1 2
(a) If y ,.., :Ti4 cos - x 31Z as x ---+ + oo,y ,.., ? as x -+ - co.
x 3

(b) Ify'" (_~)lf4exp[ _~(_X)3/2] asx---+ -00,

y'" ? as x ---+ + co.

(c) If y"" (_~)1/4 exp [ + ~ (_X)3I Z] as x -+ -co,

y'" ?asx-+ +CO.
Note: The answer to one of (a), (b), (c) is not defined. Be sure to
indicate which one is not defined, as well as giving correct answers
for the other two.
1-43 Consider the solution ofthe differential equation
_+x 2 y=O
which has zero value and unit slope at x = 10.
(a) Give (approximately) the Iocation of the next zero of y(x) greater
than 10.
(b) Give (approximately) the value of y(x) at its first maximum for
x> 10.
Problems 43

1-44 Consider a solution Yl (x) of the differential equation

y"+x 2 y=O
such that Yl(X) has a zero at x = 5. Give approximately the location
of the 25th zero beyond the one at x = 5, and estimate the error in your


In this chapter we recall some tests for the convergence
af series and present a number af methods for obtaining
the sum af a series in cIosed formo Numerical and
approximate methods are not discussed here, but in
Chapter 13.

An infinite series

IOn=a1 +02 + a 3 + ...

n= 1

is said to converge to lhe sum S provided the sequence af partial sums has
the limit S; that is, provided

If the series 1:::..\ a

n converges to the sum S, we will simply write this as an

The series L:'= 1Qn is said to converge absolutely ir the related series :2::'='1 Ia. I
converges. Absolute convergence implies convergence, but not vice versa;
for example, the series
2-1 Convergence 45

converges (to the sum In 2), but it does not converge absolutely, because

does not converge (that is, it diverges).
It should be emphasized that the numbers an may be complex numbers.
We must, of course, give lhe symbol lanl of the preceding paragraph its
usual meaning when an is complex:

That is, la,,1 denotes the absolute value (or modulus) of a".
The simplest means for determining the convergence of an infiníte sedes
is to compare it with a series which is known to converge or diverge. For
example, the geometric sedes

1+x+x 2 +x 3 + ... (2-1)
1 x

converges for Ixl < 1 and diverges for Ixl > I. This leads to the ratio test:
If the ratio 10,,+ lia" I of successive terms in the infinite series

has as a limit a number less than one as n . . . . 00, the series converges, and, in
fact, converges absolutely. if the limit is greater than one, the series diverges.
Ifthe limit is one (or ifthere is no Iimit), we must investigate further.
A second criterion is comparison with an infinite integral. The series

J(I) + J(2) + J(3) + "" "

converges or diverges with the infinite integral

f, f(x) dx

provided f(x) i8 monotonically decreasing. For example, consider the

series for the Riemann zeta function
<.01\ 'jI' ~Cl :~.If .6 >~
1 1 1
2' 3' 4'
)~')i'i"e ~,., <",. --: os A .
The ratio of successive terms is

0,,+1=(" )'~(l+~)-'-l-:+ ... (2-3)

0" n+l n ".... 00 n
46 Infinite Series

The ratio approaches one as n -+ 00. Thus, comparison with-the geometric

series fails. However,


so that the criterion for convergence of the zeta-function series is s> 1.

This enables us to sharpen our ratio test; if

all +1 -+ 1 _ :: (2-5)
". o
with s greater than I, the series converges (absolutely).


Çonsider the hypergeometric series

ab x a(a + 1)b(b + 1) x 2
F(a, b; c; x) = 1 + --,
c 1.
+ c(c + I) -2'
. +... (2-6)

The ratio of successive terms is

a ll + I
(a + n)(b + n) x
ali (c + n)(n + 1)

(1+ ~)(I + ~) x
(I +~)(I +~)
= (1 + a + b: c-I + .. .)x
Thus the series converges if Ixl < I, ar, when Ixl = I, if 1

a+b-c<O (2-7)

We can sharpen our ratio test even more by considering a very slowly
converging series such as
00 1 1 1
.~, o(ln n)' ~ 2(ln 2)' + 3(ln 3)' + ... (2-8)

1 If we allow a, b, c to be comp{ex numbers, condition (2-7) becomes


2-1 Convergence 47

By comparing with the integral

dx -1 1
f x(ln xy - s 1 (In X)s-l

we see that the series (2-8) converges provided s > 1. The fatio af successive
terms in (2-8) is
°,,+1 n [ Inn ]'
~ = n + 1 In (n + 1)

~ (1 - ~n + ... ) [1 + nln1n + ...]-'

1 s
n n In n
+". (2-10)

Thus a series for which 10"+lla,,1 is of the form (2-10) converges ir $ > I and
diverges ir s ~ 1.

Consider the series solution (I-51) af Legendre's equation

XZ x4
1 - n(n + 1) - + n(n + l)(n - 2)(n + 3)- - + ...
21 4!
The ratio af successive terms is
a, (n - 2i + 4)(n + 2i - 3) 2
0,_1 (2i - 3) (2i - 2)

For large i, this fatio is

- [1 - ~ + o( 1,)]x'
l i
Thus, ir Xl = I, the series diverges.

There are, af course, other convergence criteria; fOf example, ir the sigos
af the 0" alternate and a.. approaches zero monotonically, then the series
In 0" converges (but not necessarily absolutely).The reader interested in
a more rigorous and more complete treatment af convergence, absolute
convergence, and so forth, should consult any of the standard references,
such as those listed at the end of this chapter.
48 Infinite Sedes


The reader should be familiar with at least the following simple series:
x2 '" n! x,.
(I +x)"_1 + nx+ n(n -1)-+
000 - L
,.=0 (n cx)!-cx!


Using the Euler relation e l '" = cos x + i sin x, we deduce from (2~12)

x3 x5
sinx=x--+-- +'" (2-13)
3! 5!
x4 x2
cosx=I--+-- +'" (2-14)
2! 4!
Term~by-term integration of the series for (1 + X)-1 and (l + X 2)-1 yields
x2 x3 x4
In(1 + x) = x - - +- - - + - ' .. (2-15)
2 3 4


The above series are simple enough to be easily remembered, Other

elementary functions such as tan x, ctn x, sec x, and so on, have sedes
expansions which are probably less familiar, For some, the coefficients
are expressed in terms of the so-called Bernoulli numbers or the Euler numbers
(which may be defined, in fact, by these coefficients), See Problem 2-12
for sec x,
We shall introduce the Bernoulli numbers by an expansion of the function
{e'" - 1)-1, Multiplying by x to eliminate the singularity at x = 0, we can


x - (c o + c i x + C
x 2 + ",) ( x + -x 'x'
+ -
2! 3! + '"

and, dividing by x,

I-(c o +cx+c x 2
+",) ( 1+-+-+'"
.x x' )
1 2 2! 3!
, ,n
(,n]! 'B,,,* .
2-2 Familiar Series 49

Let c" = Bn/n!. Then

B x B x'
1= ( B +_'_+:2 + ... )( 1+-+-+'"
x x' ) (2-18)
o 11 21 21 31

Equating powers of x in (2 18), w

1 =.Bo
Bo B,
2! 11
o _-Bo .!L Bi
- 31 +,2-111 + 1121
Except for the first one, Eqs. (2-19) may be written symbolically in the
concise form (B + 1)" - B" = O. with the understanding that B' really means
B.. The tirst few af these Bernoulli numbers are

Do =1 D2 = i- B. = - 'iõ B6 =n B8 =- 1
3 0 •••

B 1 = -} B3=B5=B,=""=0 (2-20)
~fl+(=O (r.=.(,2,3, ... ).
Notations .vary, 80 lhat care must be exercised when looking up formulas.
For example, the notations af Dwight (OS) and Pierce and Foster (Pl)
differ fcom ours.
As an example of the usefulness af Bernoulli numbers· as coefficients
in power series af familiar functions, consider
.eb+e- 1X
ctnX=l Ix -Ix (2-21)
e - e
Let ix = yf2. .Then

~;fl + 2·)
\ e' - I

y' 2
e1'- l'

~2' I B.y·
Y "cwn n!
~ )'
"lo t -In, = ' - H_.~ (" )'n
y<. -{=o [:.:(O;! '-' :<11 ",'l, .

50 Intinite Series

Replacing y by 2ix gives

ctn x = ~
L (-1)'" ",B,,.,(:;:2x,,-)' (2-22)
x .. evu n!
The relation tan x = ctn x - 2 etl'l. 2x enables us to deduce the power series
for tan x.


Various devices may be used to reduce a given unfamiliar series to a
known oDe.
Differentiation and integration are often useful.


f(x) = 1 +2x+ 3>' +4x' + ... (2-23)

Integrating term by term,

= x + x 2 + x3 + ... = x
!(x) dx
1 x
Then, differentiating,

f(x) = :X(l ~J =(1 1 X)2

[We might have recognized this immediately from the binomial senes


2 3
1 XX X
!(x) = - + - + - + - + ... (2-24)
1. 2 2 .3 3.4 4.5

Differentiating twice,
(X"!)" = 1 + x + x 2 + Xl + -.. =
1 x
From this, two integrations give
1 1- x
f(x) =- + ,In (1 - x)
x x
2-3 Transformation Df Series 51

The constants of integration which arise in this procedure must be evaluated

by knowing the series at certain values of x; for example, it is obvious from
(2-24) thatf(O) ~ t.

Complex variables sometimes provide useful transformations.


f(O) = 1 + acos O + a 2 cos 20 + .. .

= Re (1 + aei9 + a 2 e 216 + ...) (2-25)
This is just a simple geometric sedes. Therefore,

f(O)=Re 1= l-acasO
l_ae l9
I 2acosO+a 2

The trick of dilferentiation or integration may be employed even if the

series does not contain a variable.


S~-+-+-+··· (2-26)
2! 3! 4!
x2 2Xl 3X4
f(x) ~ - +- +- + ...
2! 3! 4!
Tho" S ~f(I).
x3 X4
f '(x) = X + x + - + - + ... = x~
2! 3!
f(x) = f,xe" dx = xe" - ~ +1
s ~ f(l) ~ 1


m(m - 1) m(m -l)(m - 2) m!

S~I+m+ 2'• + 3'• +···~LR n.'(m _ n)1 (2-27)

""' m! " S ~
f(x)=~n!(m_n)!x f(l)
52 Infinite Series

J(x) - (1 + x)"
Therefore S = 2m•
We shall now consider the deceptively simple series [compare (2·2)]
S- L - 2 - 1 + -4 + 9- + -16 + ... -
n=l n
(2) (2-28)

Previous examples suggest defining

x2 x3 x4
J(x) - X + 4 + "9 + 16 + ... S - J(I) (2-29)

X x2 1
I'(x) - 1 + - +- + ... - - -In (1 - x)
2 3 x

J(x) - - I,
x In (1 - x) dx
and S __ f,'In (1 -x x) dx (2-30)

Unfortunately, this integral is just as difticult as the series was.

The series (2·28) can be expressed in terms of a Bernoulli number. To
find this relation, we begin by writing the Fourier series for cos kx. (We
shall discuss Fourier series in more detai! in Chapter 4.)
Ao ~ .
cos kx = - +~ (A~ cos nx + Bftsmnx) (2-31)
2 n'" 1

cos kx is an even function, so there are no sine terms (B, - O). The
coefticients An are given by
2 I~ 2k sin kn
An =- cos nx cos kx dx = ( - l t (k' ') (2-32)
Jto Jt -n
_ 2k sin kJt
coskx- Jt
cos x(_1_
cos 2x _ cos 3x ...)
2k2-k2_1+e_4 k2_9+ (2-33)

Setting x = lT in (2·33) gives

kJtctnkn=1+2k e 1-1 + k 2
4+ 2
9+'" ) (2-34)

This is the partial fraction representation of the cotangent. For those who
know the relevant mathematics, this partial fraction represenlation could
2-3 Transformation of Sedes 53

have been written down immediately by using the Mittag-Lerter theorem 2

from complex variable theory.
[n any case, ~y expanding (2-34) as a power series in k, we obtain

k1tctn k1r = l- 2k2(1 + ;2 + ;2 + .. -)

-2k,( 1+-+-+···
) _ ...
2 3

- 1- 2 L (2n)k"
,- ,
CH-±~.±2, ---) (2-35)

where {(211) rs defined in (2-2). Comparing (2-35) with our power series
(2-22) for the cotangent in terms of Bernoulli numbers, we find
_ (_l)N+ 1 B 2N (21t)2N
( n) - 2(2n)! (2-36)



1 1 B 16n4 n- 4
16 81
... -(4)- - ' 48 -90
- (2-38)

We shall conclude this section by discussing a transformation which is
useful both for analytical and numerical summing of series. Let


be a series whose sum we know, and suppose we want to sum the series

f(z)- Lc.b,z·
• (2-40)
'Scc Copson (C8) Section 6.8, or Whittaker and Watson (W5) Section 7.4. The
tbeorem, rougbJy speakina:, says that a function /(:) which is analytic everywhere in the
flnite z-plane exccpt for simple peles at z = alo Ql, ••• with residues b h b 2 , ••• , and which
(except ncar the peles) remains finite as Izl approaches infinity, may be written in the form

f(')-f(O)+'E b.( I
• z a.

(2-34) folloWi immediately ifwe take for I(z) the function TrZ ctn TrZ.

~ ~ ooA
..::..1í~ 1-,1.:L="+2~2.""'\'"
""t .... =-:;:--,.'2:>''''''-
~ .... 4.L- 2. 2. 2. -{ .
",~L . . ... 'I L - . 2. n-
... O-A 2; -n li n"'~ -
54 lnfinite Series

We eliminate the b n from (2-40):

!(z)=CObO+Clb1Z+ClblZ2+C3b3Z3+ ...
= cog(z) + (c 1 - cO)b 1z + (c 1 - co)b 1 z 2 + (C3 - co)b 3 z 3 + ...
= cog(z) + (c 1 - co)zg'(z) + (c 2 - 2c, + cO)b1z 1
+ (C3 - 3c 1 + 2co)b 3 z 3 + ...
= cog(z) + (c 1 - co)zg'(z) + (c 1 - 2c 1 + co) 21
Z3 g '''(Z)
+(C3 - 3c1 + 3c. - co) 3! + '" (2-41)

Tbe successive coefficients are just the leading differences in a difference

table of the coefficients cn :

Cl - Co

c1 Cl - 2Cl + Co
Cz - C1 C3 - 3Cl + 3c1 - Co
C2 C3 - 2'c] + Cl

. ..
• ••

Tbe most common example ofthis transformation is Euler's transformation.

g(z) = =1- z + Zl _ Z3 + - , .. (2-42)

bn =(-I)n

zg'(z) = (1 + Z)l
z2g"(Z) Zl
2! - (1 + Z)3

ReCerences 55

Thus, if/(z) = I:'-o (-lrcnzn, the Euler transformation teUs us that

fez) = 1 ~ z[co - (c} - co)C : z)

+(c 2 -2c +cO)(I: zf - +"J
I (2-43)

For example, consider the series

fez) = 1 - 1-Z + !Z2 - !Z3 + - ... (2-44)
[We already know fez) = (l/z) In (I + z), af course.] This is a rather
poorly converging series; in fact, ir diverges for Izl > 1. From the difference

-i -t
, t
, TI
, '"

we see that

!(,) ~ 1 [1 + 21+z ( z )' + ... j

1( z ) + ~31+z (2-45)
This new series converges much better for z near 1, for example. In fact,
it converges for ali positive z.
This is just one example of the tricks that are available to turn one series
into another, more useful, one. The important practical problem of how
to find the numerical value of a series will be treated in Chapter 13.

Infinite series are discussed quite thoroughly in numerous books on
complex variable theory. See, for example, Whittaker and Watson (WS)
Chapter 11 and IH, ar Apostol (AS) Chapters 12 and 13.
Several fairly elementary references are Hyslop (HI3); Green (G7);
and Stanaitis (SI1). The monograph by Hirschman (HIO) is also very read-
able, but it is at a somewhat more advanced leveI than the three preceding
56 Infinite Series

Two standard treatises, with somewhat of a nineteenth-century flavor,

are those by Brornwich (B8) and Knopp (K3).
Davis (D2) presents in a systematic manner rnany devices for summing
series. Many elementary series, inc1uding those leading to Bernoulli numbers
and Euler numbers, are given near the beginning of Dwight (08).

Find the sums of the sedes 2-1 through 2-7:
1 1 1 1 1
2-1 1+-----+-+--- - + + ...
4 16 64 256 1024
1 1 1 1
2-2 10+2·4+0+4'6+'"

1 1 1 1
2-3 1 - 5. 32 - 7.33 + 11 . 3 5 + 13.3 6 - - + + ...
1 2 3
2-4 01 + I! + 2! + ---

1 1 1
2-5 1+-+-+-+'--
9 25 49
1 1 -1
2-6 1 + 92 + 252 + 49 2 + . --

I 1 1
2-7 1--+----+-··-
2 2 2
4 9 16
2-8 Evaluate in closed form the sum
f(O) = sin O +! sin 20 +t sin 30 +t sin 40 + ---
(you may assume O < O < n for definiteness).

Do the foJlowing sedes converge or not?

,., ,.,., "!'!'!
;;-~OÓc-" + 2 2 + + . -.
9-7-25·1! 11-9-49'21 13-11·81·3!
(1'3)' (1'3'5)' (1'3'5'7)'
2-10 i-,.--é" d'-;-:""-;;é"o",
1.1'1 2 + 4-2-(1-2)2 + 16'3-0-2·3)2
(1 . 3· 5 . 7 . 9)'
+ + ...
64· 4 '(1' 2· 3 '4)'
Problems 57

2-11 Evaluate the series

'" (~lt+ln2x2~-1
f(x) ~ Jo (20 - I)!

4x 3 9x 5
3! 5!
in closed form, by comparing with
x3 x 5
sinx=x--+-- + ...
3! 5!
2-12 Consider the so-called Euler numbers, defined by

~ E2~ 2
secz= L,(-lt-z n
n~O (2n)!
(a) Show that

(E + l)k + (E - V = O (k even)
Give E2' E 4 , E6' Es.
(b) Using the partia! fracHoo expansion of the secaot


(1 )

1 1
(2) 1-32n+l+52n+l-+'"

In the first three sections af this chapter we discuss var-
ious techniques for the analytic evaluation af definhe
integrais. These techniques include differentiation ar
integration with respect to a parameter, the exploitation
af symmetries, and evaluation by contour integration.
Tabulated integrais such as the gamma function, beta
function, exponentiaI integral, elliptic integrais, and 50
forth, are described in Section 3--4. Approximate expan-
sioos, especial1y asymptotic expansions, are treated in
Section 3-5, and the saddle-point method is discussed
in Section 3-6.


We first review some useful elementary techniques for doing integrais.
The simplest device is changing variables. For example, the standard trick
for evaluating
, ,<
J, e-" dt ~ ,-.In (3-1)

by using polar coordinates is presumably familiar. From this result, by

setting t = uJawe find
f, -..' du -- J
< 1 "
e ~ - (3-2)
2 a
3-1 Elementary Methods 59

ar by setting t = ul , we may deduce

(0 e-li' u du = !J~ (3-3)
How about Só e-t' dt? This is not so easy. Let


Make the change of variable

t~ =u dt = ! U(l/~)-l du


This does not look any easier, but one defines the gamma function r(z) by


so that


We shall say more about the gamma function In Section 3--4; note that
(3-1) give, rm J'.
Another useful technique is to introduce complex variables.


I = f,• e- cos Ax dx

= Re f,• e-aXé'X dx



I Note Ihat our method seems to require Ihat À be real, but the results are in fact corceet

for À anywhere in lhe slrip 11m Àj.0Ç.4, How did Ihis happen 1
..p,,,/f-flj.fl~ 1m À>-IÃ
60 Evaluation of Integrais

This method gives us another integral at the same time, from the imaginary


The method of integration by parts is very useful. and is presumably

familiar by now.
Another useful trick is differentiation or integration with respect to a


f• e-ti;; cos Ax dx
1= x (3-9)

I e-ucosÀxdx= 2
o a +A
d a 2 - À.2
l=-d-/(a)=(, .,),
a o + Ao
We have assumed that the order of differentiation and integration can
be reversed. For necessary and sufficient conditions see mathematics
books, such as Whittaker and Watson (WS) Chapter lV, or Apostai (AS)
Chapter 9. In physical applications it will nearly always work.


f• - dx

'" e-lU: sin x
1(0:) =
I• x
dx so that 1 = 1(0)

I" -1
e- u sinxdx=2 1
do: o o: +
da ,
1(0:) =- fa 2
= C - tan- o:

But l( ct) = O. Therefore C = tr/2.

1(0:) = "__
tan - I o: and (3-11)
3-1 Elementary Methods 61

Sometimes one can combine several derivatives of an integral to form a

differential equation.



/"(a) + 1(0:) = f e-~X dx

= -

o "
This equation is easily solved by the variation of parameters method dis-
cussed in Section l-L The result is

fI> COS t
~ sin t .
l(a) = -cos ct -,- dt + sm ct -,- dI

But [(ct) and ali its derivatives vanish at a = 00. Thus

l(ct) =
sm a f~ -
cos-t di - COS ct f" sin
- -t dI
00 t "'" t
The cosine-integral and sine-integral functions are defined by

· = f"'costd . f"sint.
C IX -- I SlX= --dI (3-13)
00 I o I

Thus [(Ix) = sin ct Ci a + cos ct(tr/2 - Si a). The functions Si o: and Ci (X are
tabulated in Jahnke et aI. (13) and Abramowitz and Stegun (AI).

Finally, the useful integraIs JO' e-ax~x" dx (n = 0,1,2, ...) can be obtained
from the first two,

foOO e - •• ' dX -- -1
J" -
a f

e-ax2x dx =-

by repeated differentiation.
We observe in passing that if a parameter in an integral also appears in
the limit(s), the differentiation with respect to that parameter proceeds
according to the following rule:
62 Evaluation of Integrais


The evaluation of some integraIs may be greatly simplified by exploiting
the symmetries present in the problem. We shall illustrate the principies
involved by means of some integrais over solid angle in three dimensions.


Consider the integral

dO 2.. +1 1
II(k) = f 1 +k'i = fo dtP LI d(cosO) 1 +k.i (3-16)

where i is the unit radius vector and (O, tP) are conventional spherical polar

i", = sin Ocos tP iy=sinOsintjJ

Since the orientation of our coordinate system is arbitrary, we may choose
the z axis along k and obtain (we assume k < I)

Il(k)=f2~dtPf+l d(cosO) = 21t ln(1

o _1 1 + k cos O k
1 k

The integrais
I.(k) ~ f (I + k _f)' (3-18)

may be obtained from /1 by differentiation (replace the I in the denominator

by (l, differentiate with respect to CJ:, and set (l equal to I again). For example,
da 4,
12(k)= (1+k'i)2=1 e (3-19)

Another example is
II(k,a)= f 1 +k'f' (3-20)

This integral is complicated by the fact that two directions are given by the
two vectors a and k. and we cannot choose our polar axis along both of
them. However, the direction a is trivial, in that it may be "factored out."
For, consider
J(k)~ 1 +k-f (3-21)
3-2 Use of Symmetry Arguments 63

Il{k, a) =a ' J
Now J{k) is a vector and must point in the direction k, since no other
direction is specified in the definition (3-21) of J. Therefore,
To evaluate the scalar A, we dot k into both sides of (3-22) and obtain

A ~.!.k
- J(k)~.!.f
k -, dll
k k 1+ k ' ;

~ :' f dll( 1- 1+IH)

= 4n: (1 _ -.!.. In 1 + k)
k2 2klk
Thus our original int~gral (3-20) is


What about the integral

12 (k, a) = f (I + k ' ? tY
We could obtain 12 (k, a) from ft(k. a) by replacing the I by IX and differentiat-
ing as before, but a simpler method is to differentiate ll(k) with respect to k.
On the one hand,
fà dll f d!l1
~ ak I +k"~ - (I +k'i)'

On the other hand, using (3-17),

aI ,(k) ~ ~ aI ,(k) ~ 2nk [......:2:.." _ ~ In

ôk k õk k 2 "1 k 2 k I
(1+ k)]

Comparing (3-24) and (3-25) gives

d!l1 2nk ( - 2 11 I + k)
f (1 + k' ;)2 = k} 1 _ e +k n 1 k
and therefore

64 Evaluation of lntegrals

Other examples of the use of symmetry arguments are the evaluation

of integrais such as

$2(a, b, c, d) = fdOf'a f'b f·c f'd (3-27)

To evaluate $), we observe that it is a scalar which is linear in both a and b.
The only passibility is that $1 = Aa . b, where A is a number. To find A,
let a and b both equal 2. Then

= A = f dO(f' 2)2 = J dO cos 2 0='3

Therefore $1(a, b) = (4nj3)a' b.

The evaluation of $2 proceeds similarly. Since $2 is a scalar, linear in
the four vectors a, b, c, d, as well as being invariant under any interchange
af these vectors, it must have the form
$2 = B(a'b c'd + a'cb'd + a'd b'c)
where B is a number. Bis found by setting a = b = c = d = 2, so that

$2(2. Z, 2, i) = 3B = f dO(f • i)4 = f dO cos 4 O = ~n

Therefore B = 4n/15, and
!fI 2 (a, b, c, d) = 15 (a' b d + C' a' c b' d + a' d b· c)
As a final example of the sort of device one can use to simplify integrais,
we mention the identity

If' du
ab = o [au + b(l _ U)]2

which Feynman has used to simplify lhe evaluation of integrais aflSlng

in quantum field theory.2 As ao application of (3~28), we evaluate the

-"kl-f dO (3-29)
( , ) - (l+k-i)(l+I-i)

2 See the Appendix of R. P. Feynman CF3).

3-3 Contour Integration 65

Use of (3-28) converts (3-29) to

f' f
~(k, ~ ~ o du (I + t, [ku + 1(1
u)lJ' (3-30)

The solid-angle integral in (3-30) is just Il[ku + l(i - u)], as given by

(3-19). Thu,
I/t(k, I) = 411: o 1 _ [ku
+ 1(1 U)]2

This is an elementary integral, although rather tedious; the answer has the
interesting form



B~ )(1 - k')(i' -I')

The proof of (3-31) is left as an exercise (Problem 3-37).


One ofthe most powerful means for evaluating definhe integrais is provided
by the theorem of residues from the theory of functions of a complex vari-
able. We shall iIlustrate this method of contour integration by a number of
examples in this section. Before reading this material, the student who does
not know the theory of functions of a complex variable reasonably well
should review (or learn) certain parts of this theory. These parts are pre-
sented in the Appendix of this book to serve as an aid in the review (or as a
guide to the study).
The theorem of residues [Appendix, Eq. (A-15)] teUs us that if a function
fez) is regular in the region bounded by a c10sed path C, except for a finite
number of poles and isolated essential singularities in the interior of C,
then the integral of fez) along the contour C is

tf(z) dz = 2n:i I residues

where L residues means the sum of the residues at ali the poles and essential
singularities inside C.
The residues at poles and isolated essential singularities may be found
as follows.

66 Evaluation of IntegraIs

If/(z) has a simple pote (pole or order onerat z = zo. the residue is
ao' - [(z - zo)f(z)],... (3-32)
If fez) is written in the farm fez) = q(z}fp{z), where q{z) is regular and p(z)
has a simple zero at zo. the residue of fez) at Zo may be computed from


Ir Zo is a pole af arder n, the residue is

a_I = (n _1 I)! ((d)"-'

dz [(z - I
zo)"/(z)] ~~~o (3-34)

If Zo is an isolated essential singularity, the residue is found from the

Laurent expansion (Appendix, Section A-2, item 7).
We illustrate the method af contaur integration by some examples.



Consider § dz/(l + Z2) a100g the contour af Figure 3-1. Alang the rcal
axis the integral is U. Along the large semicircle in the upper-half plane
we get zero, since

z = Rei' dz - iRe" dO

f,--=d,,-z'l ~ .!. fe- 19

de -+ oas R -+ 00
1+ z R



Figure 3-1 Contour for Ih. integral (3-35)
3-3 Contour Integration 67

The residue of l/O + Z2) = l/(z + i)(z - i) at z = i is 1/(2i). Thus

21 = 21ti(;j) = 1& 1-- •2

Note that an important part of the problem may be choosing the .. return
path" 50 that the contribution from it is simple (preferably zero).


Consider a resistance R and inductance L connected in series with a

voltage V(t) (Figure 3-2). Suppose Y(t) is a voltage impulse, that is, a
very high pulse lasting for a very short time. As we shall see in Chapter 4,
we can write to a good approximation

V(t) A
= -2n f o
e''''' dw


Figure 3-2 Sedes R-L circuit

where A is the area under the curve V(t).
The current due to a voltage e'''' is e, .. t/(R + jwL). Thu5 the current due
to our voltage pulse is
A f"" i
e "" dw
1(.) - 2. -o R + 'roL (3-36)

Let us evaluate this integral.

II t < O. the integrand is exponentially small for Im w -+ - 00, so that we
may complete the contour by a large semicircle in the lower-half w-plane.
along which the integral vanishes. 3 The contour encloses no singularities,
so that l(t) = o.
If t > O, we must complete the contour by a large semicircle in the upper-
half plane. Then
I(t) = 21&i _
(A) e-""IL. = _A
21& L
e- Rt / L

3 A rigorous justification or this procedure is provided by lordan's lemma; see Copson

(C8) p. 137 for example.
68 Evaluation of:,{ntegrals


• dx
1=tl+x3 (3-37)

The integrand is not even, so we cannot extend it to - 00, Consider the

j 1 + Z3
The integrand is many-va:lued; we may cut the plane as shown in Figure 3-3,
and define In z real (= In x) just above the cut. Then In z = In x + 27r;
below the cut, and integrating along the indicated contour,
1.. In z dz _ 2 'I
jl+z 3 --

On the other hand, using the method of residues,

i1n z dz = _ 41t 2 ;.J3

j 1 + Z3 9
Thud ~(2.J3)/9.
When integrating around a branch point, as in this example, it is necessary
to show that the integral on a vanishingly small circle around the branch
paint is zero, In this example, this part goes like r In r, which approaches
zero as r -+ O,



~ '"

i. 5 ii/l
-{!. =-e

Figure 3-3 Contour for the in.tegral (3-37)

3-3 Contour Integration 69

A more straightforward method for evaluating the integral (3-37) consists

of evaluating the integral
1, dz
J=11+z 3
along a c10sed contour consisting of (i) the real axis from O to + 00, (ii) one-
third of a large circle at Izl = 00, and (Ui) a return to the origin along the line
arg z = 211./3. Evaluating J along these pieces, we find
J = (I - é~i/3)I

On the other hand, the integrand of J has a simple pole at z = e~i/3; the
residue theorem gives

2n:i e- 2 "i/3 n 2n
I~- = =-
3 1 e2ni/3
3 sm '"
... 3j3-
as before.


" de
I-foa+bcosO a>b>O (3-38)

The integrand is even, so

21 ~ f" -:-,d-::e~
o a+bcos8
If we integrate along the unit circ1e, as in Figure 3-4,
dz = ie i8 dO

efO+e- iO
1( z+-1) (3-39)
2 2 ,
21 = fc a + (bI2)[,
+ (llz)]
2 f. dz
= i bz + 2az + b
70 Evaluation of Integrais



Figure 3-4 Contour for the integral (3-38)

The integrand has two poles, at the roots ofthe denominator. The product
of the roots is h/h = 1; one is outside and one inside the unit circ1e. They
are at

aJa'- -
,= - -b +- b2

21 = i2 27ri ( residueatx 1 = - ') + b2 - 1

I _ 2n( 1 ) _ ~n~
- 2bx 1 +2a - Ja 2 _b 2


Consider the integral
i';; d,
j 1+ Z2

along the contour of Figure 3-5. We choose JZ positive on top of the cut.

But, using residues,

Therefore 1= n/J 2.

3-3 Contour Integration 71

Figure 3-5 Contour for the integraI (3-40)


(O < a < 1) (3-41)

We c1early wish to consider

é" dz
f ". + 1
along some contour. One choice is shown in Figure 3-6, involving the
famíliar large semicircle. Note that we must give a a small imaginary part in


Figure 3--6 Possible contour for the integral (341)

72 Evaluation of Integrais

order to he certain that the integral along the semicircle vanishes. Then

! = 2n; L residues

There is an infinite number of poles.

At z = ;n, residue is _ei~a

At z =3in, residue is _e 3i:n:a, etc.



Now we can let Im a ....... O.

Alternatively, we could use the contour shown In Figure 3-7. Along
the real axis, we get I. Along Im z = 2ni, we get

(I - e2 :n:ia)/ = 2ni (residue at z = ni)

1= n as before

* z-plane

Figure 3-7 Another contour for the integral (3-41)
3-3 Contour Integration 73

R~ 00

Figure 3-8 Contour for the integral (3-42)



We consider
I dz
r ~1 _ z2(1 + Z2)
along the contour of Figure 3-8. On the top side of the cut we get I, and
we get another I from the bottom side. Therefore,

21 ~ hi [2i~ + 2i~1

Consider the integral
I~l*)dz (3-43)
j SIO 1tZ

around the contour of Figure 3-9, wheref(z) has several ísolated singularities
(indicated by crosses in the figure) and goes to zero at least as fast as Izr 1
as Izl-+ 00.
00 the one hand, the integral I may be evaluated by summing the residues
at the zeros of sin 1tZ, indicated by dots in Figure 3-9. The result is

74 Evaluation of Integrais


x x


Figure 3-9 Original contour for the integral (3-43)

On the other hand, the integral along the contour of Figure 3-9 is clearly
the same as that along the contour of Figure 3-10 since the integral around
the circ1e at Iz) = 00 vanishes. The singularities enc10sed by the contour of
Figure 3-10 are now the singularities of f(z); if the locations and residues
are denoted by Zk and R~, respectively, then

I = - 21tl,'" R,
"" ~,=­
t sm 1tz.

Comparison with (3-44) gives the summation formula

00 R
L (-)'i(n)~ -nL ' ' (3-45)
n~-", k Sln1tz.

x x circle at
Izl= 00

Figure 3-10 Deformed contour for the integral (3-43)

3-4 Tabulated Integrais 75

This device, which converts an infinite sum into a contour integral which
is subsequently deformed, is known as a Sommerfeld-Watson transformation."
As aD example, consider the series
sin x 2 sin 2x 3 sin 3x
8(x)- 1
- a + 1 - a1 + 4 + a l + 9 - +".

This is just the Ieft side of equation (3-45), if we set

f(z) = _~zsinxz
2a 1 + Zl
(Note that we must require Ixl < n in order that the large circle in Figure
3-10 make no contribution to the integra!.) Now the locations and residues
of the poles of fez) are
Z1 = ia Zl = -ia
R1 =
. h
ax Rl =~sinhax
Therefore, we obtain from formula (3-45)

[ i' h i'h]
- 4 810 ax -4'10 ax
. h ax
7t SlO
S(x) = -7t i8inha7t + - i sinh an =2sinhan

As we remarked above, this result ia valid only if Ixl < n. For x outside
this range, we simply observe that S(x), from its definition, is periodic in
x with period 2n.


We fiest mention (again) the gamma function, defined for Re z > O
by the integral '


• Sce Sommerfeld (S9) Appendill to Chapter VI, or G. N. Watson <W1).

76 Evaluation of Integrais

Integration by parts gives

r(z) ~ (z - I)r(z - I) (Rez> I) (3-47)
Thus. from r(l) = 1. we obtain
r(2) ~ I. r(3) ~ 2. r(4) ~ 6•...• r(nl - (n - I) I (3-48) o ! =" r(o=~,
Tbe gamma function may be analytically continued into !he entire complex
plane by means of the recursion relation (3-47). except for simple poles at
A related integral of some interest is encountered if we consider


Let x = uI.
r(r)r(s) = Ja) e-'V+ I - 1 du (dt ,r-lei _ 1)*-1
o o
- r(, +a)B(,. s)
where B(r, s) is the beta funclion

B(,. ,) ~ r(,)r(s) ~
r(r+s) , o
f'..- '(I - xy-' dx (3-49)

This integral representation for B(r, 9) is obviously valid only for Re r > O,
Re 3 > O; the fust equality in (3-49) defines B(r, $) for ali r, s in the complex
An interesting special case of this relation is
r(z)r(1 - z) ~ r(I)B(z. I - z)

Letx ~ '/(1 +I). The.

a) ~-1 dt
r(z)r(l - z) ~ o I I +•
This integral can be done by using the contour shown in Figure 3_11. 5 The
result is
3-4 Tabulated IntegraIs 77

t -plane

Figure 3-11 Contour for the integral J; [t=-l/(l + t)] dt

Our derivation has only been valid for O < Re z < I, but both sides of (3-50)
are analytic functions of z (except at z = O, + 1, ±2, ... ), so we can extend
the result into the entire plane.
Another integral which has been given a name is the so·called exponential
' e' dt
Ei(x) = f -
_. t

This definition is conventional1y supplemented by cutting the x·plane along

the positive real axis. Thus Ei(x) is well defined for negative real x, but for
positive real x we must distinguish between Ei(x + ie) and Ei(x - ie).
Two functions related to Ei(x) are the sine and cosine integrais:
S IX = f"'sintdt
o t

. f"'COS tdt
CIX = (3-53)
• t
The error function is defined as

erf X
= J- J'
< o
e-I' dt (3-54)

The associated trigonometric integrals are known as Fremel integra/5 6 :

C(x) ("') dt
= fo,cos 2 S(x) = Ia,sin «t')
2 dt (3-55)

6 Conven1ions differ; our defini1ions (3-55) agree with Magnus el ai. (MI) but oot with
Erdelyi tI ai. (E5).
78 Evaluation of Integrais

Another class of integral whieh is tabulated in many plaees is the class of

elliptic integraIs. These are integrais of the form

dx A(x) + B(x)jS(Xj
f C(x) + D(x)JS(x)

where A, D, C, D are polynomials and S is a polynomial of the third or fourth

degree (not a perfeet square, of course!).
We shall not go into the general theory ofsueh integraIs, butjust make a few
remarks, which the student can verify. 7 In the first place,

A::--:-+.::B",J",S ~ E + .!... (3-57)

C + D-.fi -.fi
where E and F are rationaI (ratios of poIynomials). By deeomposing F in
partial fraetions, we see that the only nonelementary integrais we need are
H. f
~ (x _ ,)"Js (l-58)

But we can evaluate ali the Jn from J o , J I , and J 2 , and alI the Rn from Ri> J o ,
Ji> and J 2 • Thus we only need

f dx
fO J,~
f xJ-dx J,~
f x' dx/O H,~
f dx
JS (3-59)
vS S vS (x - ,)
There are various standard forms for S; we shall consider only Legendre's,
namely, S = (1 - x 2 )(I - k 2 x 2 ). Then
J ,-
- f J(l x 2)(1 ex 2 )
The integral
• dx
- ,J(I x 2 )(1 k 2x 2)

is ealled the Legendre elliptic integral of the first kind. Generally, one sets
x = sin (jJ, and defines
• d~
F(~. k) ~ f, J7.I~k~':;si::'n,F~f (3-61)

J 1 is ao "elementary" integral, if we set x 2 = u.

7 See, for eumple, Abrarnowitz and Stegun (AI) Chapter 17 for a more complete and
useful discussion.
3-4 Tabulated Integrais 79

Instead of J 2' one takes for a standard integral the Legendre elliptic
integral 01 the second kind,

E~ f Jl - k x
2 2

o Jl- x
Again one general1y sets x = sin ~, and tabulates
E(~, k) ~ f•Jl -
k' sin' ~ d~ (3-63)

The integral H I is hard. One defines the Legendre el/iptic integral 01 the
third kind,
• dx
n (~, n, k) = fo (1 + nx2)J(l _ x2)(1 _ k2x·i) (3-64)

• d~

= fo (1 + n sin ~)J1 - 2
k 2 sin 2 ~ (3-65)

When ~ = x/2, we have the complete elliptic integrals,

K(k)=FG,k) = (12 Jl-~~sin2~ (3-66)

E(k) = EG, k) = (f2J1 - k 2 sin 2 ~ d~ (3-67)

n (n, k) = n G, n, k) = f:'

(1 + nsin2 ~~1 (3-68)

The formulas for reducing S to a standard form are given in various places,
for example, Magnus et aI. (MI) Section 10-1 or Abramowitz and Stegun
(AI) Chapter 17. Elliptic integrais occur, among other places, in
I. The motion of the simple pendulum
2. Finding inductances of coils
3. Finding lengths of conic sections
4. Calculating the solid angles of circles seen obliquely
A large number of integrais are related to tabulated functions; these
integral representations are sometimes very useful. For example,
I cos tx dt 11:

fo J 1 - t
'~-2J,(x) (3-69)

I cos (n + t)~ n:
7;;";;~~'" d~ ~ - - P.(cos O)
J2 (3-70)
80 Evaluation of Integrais

Jo(x) is a suitably normalized solution of Bessel's equation, v.ith m = O,

and is known as a Bessel function; Pn is a Legendre polynomial. We
shall study such integral representations of the .. special functions" in
greater detail in Chapter 7.


One can often obtain a useful expression for an integral by expanding the
integrand in some sort of series.


erf x =
e-I' dt

2 ( x' x' x' )

= JTc x~3+5·2!~7·3!+~··· (3-71)

This series converges for ali x, but is only useful for small x, (x;$ 1).

Integrations by parts are often useful.


Suppose we want erf x for large x. As x --+ 00, erf x --+ 1. Let us compute
the dilference fram 1.
2 •
1 ~ erf x = r= J e-r' dt
v" ,
We perform a sequence of integrations by parts.

"" e-r'dt=___
e-X' f"" e-I' dt
fX 2x" 2t
e-X' f"" 3 e-r'
=h~ 4x 3 + x 4"7 dt
3-5 Approximate Expansions 81

The result after n integrations by parts is (n > 1):

erfx = l_..2.-e-Xl[~ __1_+~_1· 3·5 + _ ...

..r;r 2x 3
22X 4 723 X 5 2 x

1 _1 1 . 3 . 5 ... (2n-3)]
+ (- )" 2"x'" ,

+ ( _l)n
1 . 3 . 5 ... (2n - 1) 2
2n r:
f <ll e- 12
---z,;- dt (3-72)
v 1t x t
The terms in the brackets do not form the beginning of a convergent
infinite series. The series does not converge for any x, since the individual
terms eventually increase as n increases. Nevertheless, this expression witb
a finite number ofterms is very useful for large x.

The expression (3·72) is exact if we include the "remainder," that is, the
1ast term, containing the integral. 1hls remainder alternates in sign as n
increases, which means that the error afier n terms in the series is smaller
in magnitude than the next termo Thus the accuracy of the approximate
expression in the brackets is highest if we stop one term before the smallest.
The series in brackets in (3·72) is an example of an asymptotic series if
continued indefinitely.8 The precise definition of an asymptotic series is
the following:
S(z) = Co + C-1 + -Cz + ... (3-73)
z z'
is an asymptotic series expansion off(z) [writtenf(z)"" S(z), where ,.., reads
"is asymptotically equal to"] provided, for any n, the error involved in
terminating the series with the term cnz- n goes to zero faster than z-n as Izl
goes to 00 for some range of arg z, that is,
Um z'[f(z) - S,(z)] ~ O (3-74)
for arg z in the given interval; S~(z) means Co + c1 /z + ... + c,.}%".
[A convergent series approaches fez) as n -. 00 for given z, whereas ao
asymptotic series approachesf(z) as z-+ 00 for given n.]
From the definition, it is easy to show that asymptotic series may be
added, multiplied, and integrated to obtain asymptotic series for the sum,
product, and integral of the corresponding functions. Also, the asymptotic
expansion of a gíven functionf(z) is unique, but the reverse is not true. An
asymptotic series does not specify a function f(z) uniquely [see Whittaker
and Watson (W5) Chapter VIII or Jefl'reys and Jeffreys (J4) Chapter 17].

• The idea is probahly due to Poincaré; see reference (P2) Chapter VIU.
82 EvaIuation of Integrais

We have not proved that our series in (3-72) is an asymptotic series,

but that is not hard; we leave it to the interested reader.
Another example:

Ei(-x) =
f -· e'
-dt= f' e- r _dt (x> O)
-00 t "" t
-'" ",-r
=-~-f ~dt
x "" t

Continuing in this way, we obtain

,-' [ 1 21 3! (-1tn!]
-Ej(-x)=~ 1-~+X2-X3+-···+ x"

" f x e-r
+(-1)(n+1)! 00t"+2dt (3-75)

We can use this result in two ways:

1. As an asymptotic senes for Ei( - x):

,-' ( 1 2! 3! )
-Ei(-x),..,~ 1-~+ x2 - x 3 + - ... (3-76)

2. As an exact expression for computing certain integraIs, assuming we

have available a table of Ei( -x). From (3-75)

·_'--;;',dt~ (-1)" Ei(-x)+ (-1)" _,-_.
t" (n 1)! (n-l)! x

x [1-~ +
2! _ + ... +(_1)"(. -_2)1]
x2 X~2


Other important methods for approximating integrais are known as
saddle-point methods, for reasons which will 5000 be c1ear. The most
importaot of these is the method 0/ steepest descent. We shall iIlustrate
the method by looking for an approximation to r(x + 1) for x large. positive,
and real.

r(x + 1) = f,• t"'e- r dI (3-78)

3-6 Sadd1e-Point Methods 83

For large x, the integrand looks as shown in Figure 3-12. To find the
maximum, we have

O=!!..dI (e-It"') = e-'[ - t'" + Xt"'-l]

We now approximate the integrand in a particular way, by writing it in
an exponential form e'(I) and using a Taylor's series approximation for
f(t) near its maximum. (A Taylor's expansion of.the integrand itself, for
example, would not be useful if only a few terms were kept.) The integrand
af (3·78) is
e!<t) = t"'e-r = e"'lnt-t
so that
!(I) _ x In I - I

f'(t)-~-1 j'=Ofort=x

Then, expanding about t = x,

r(x + 1) ~ fo""explx In x - x - ~ (I _X)l] dI

~ e"'h""-'" f~""expl- ~ (I - X)l] dI

(If x> I,
we make a very small error by extending the integral to - 00.)
Doing the t integral gives

Figure 3-12 Graph of t"'e-' for large x

84 Evaluation of IntegraIs

This is the first term of Stirling' s formula, which is the asymptotic expansion
ofn! = r(n + 1):


We shall return to this asymptotic series at the end of this chapter.

The method of steepest descents is applicable, in general, to integrais

of the form
I(IX) = fce"/<') dz (IX large and positive) (3-80)

where C is a path in the complex plane such that the ends of the path do
not contribute significantly to the inlegral. [This method usually gives
the first term in an asymptotic expansian of [(IX), valid for large IX.]
If fez) = u + iv, we would expect most af the contribution to I(IX) to come
from parts of the contour where u is largesL The idea of the method of
steepest descents is to deform the contour C so that the regioo of large u
is compressed into as short a space as possible.
To see how to deform the contour it is necessary to examine the behavior
of the functions u and v in more ar less detail for the example at hand.
However, some general features apply to any regular functionf(z). Neither
u nor v can have a maximum or minimum except at a singularity, because
V2 u = O and V2 v = 0. 9 For example, if
ôx 2 < O then

so that a "fiat spot" of the surface u(x, y), where

ôu=ôu=O (3-81)
ôx ôy
must be a "saddle point," where the surface looks like a saddle or a mountain
pass, as shown in Figure 3-13.
By the Cauchy-Riemann equations (A-8), we see that (3-81) implies
ôv/êy = O and av/ax = O, so that F(z) = O. Thus a saddle point of the
function u(x, y) is also a saddle point of v(x, y) as well as a point where
1'(,) ~ o.
Near the saddle point zo,

• These follow immediately from the Cauchy-Riemann equations, equations (A-8) of

the Appendix.
Figure 3-13 Topography of the swface u= Ref(z) near the saddle point Zo. for a typica1 functionf(z). Theheavy
solid curves follow the centers of the rldges and vaUeys from the saddle point. and the dashed curves
foUow leveI contours. u = u(xo, yo) = constant. The curve AA is the path of steepest descent
86 Evaluation of Integrais

Figure 3-14 Alternadve possibiUdes for steepest descent contoun

Letf"(zo) =pe 19, z - Zo = se'., Then

u ~ u(xo. Yo) + i PS 2 cos (9 + 2t/J)
v ~ v(xo. Yo) + ; pS2 sin (6 + 2tjJ)
We see that on the surface u(x, y) paths af steepest desceot Crom the sad-
dle paiot into the val1eys start out in the directions where cos (8 + 2.p) -
-1. In these directions, v"'" v(xo, Yo) is constant. As we 80 further from
Zo. the steepest path follows the direction - grad u, which is normal to
the contours af equal u. Tbus the path fallaws a curve v(x. y) = constaot =
v(xo, Yo), 80 lhat the faetor ellll' in lhe original integrand af (3-80) will
nol produce destructive oscillations~~) We therefore defarm our original
contour C 80 as to go over the pass 00 tbis path af steepest descent. Using
the approximation (3-82) in lhe integral (3-80) givcs

1(1X)::::: ~e«I(~a)e'.
--.j-.p (3-83)

where tP has one of the values - 9/2 ± 'n/2, depending on which direction
we traveI over the passo (tP is just the inclination of the path at the saddle
paint.) For example, ir (J = 1[/2 we have the two alternatives or Figure 3-14.
The first is much more likely to be correot, but to be sure we should examine
the .. mountain range" we are passing through. Ir it looks anytbing Iike
that shown in part (a) or Figure 3-15, then tP = 1[/4 is indeed correct. On
the other hand. ir the range is like tbat shown in part (b) or Figure 3-15,
the second alternative. tP = - 31[/4, is the right one.
As an example or the method or steepest descents with a more general
contour, we shall work out the asymptotic approximation to r(z + 1) for
3-6 Saddle-Point Methods 87

complex z, without making any simplifications based on our previous work

with r(x + 1).

r(z + 1) = t• e- t +zlnt dt (3-84)



f(t) = (In t - ~)ei~

r(t) = G- De lP
to = z

I"(t) = - ,"

feto) = (In z - l)e iP

(a) (b)

Figure 3-15 Alternative "mountain ranges" for the two cases of

Figure 3-14
88 Evaluation of Integrais

Thusp= 1/a.2~9=1t_p. Whatis4>? Use ofthe condition cos (9 + 24»=

-1 gives
or --n
In our previous example, with z real, p = O and 4> = O, so the choice
4> = PI2 would clearly seem most reasonable. Since it is sometimes necessary
to examine the topography ofthe surface u = Ref(t) in more detail, however,
we show the ridges and vall~ys for the present example, with P= n/4, in
Figure 3-16. This figure confirms the choice 4> = P/2.
The steepest descent approximation for r(z + 1) may now be written
down from (3-80):
r(z + 1) "" ~edU-Zei'!~
r(z + 1) "" ...[2nzZ+I!2e- Z (3-86)
The original integ,.al representation for r(z) is only valid for Re z > o.

~ ti = constant /
I _
ridge /
co.".,o' I / -
~_~/;~~~":':" constant

v = constant li = constant\ \

Figure 3-16 Tapography af the surface u= Ref(t) = Re (In t- tlz)e'P

for the case z = 3ei ("14) • It is c1ear that the original path
af integradoo along the real t-axis should be deformed
50 as to go over the saddle point, to = z, 00 the path
af steepest descents (valleys) in a direction specified by
lhe angle 4> - P/2
3-6 Saddle-Point Methods 89

However, the result (3·86) is aetually valid for alI Izl- 00 provided we stay
away fram the negative real axis (see Whittaker and Watson (W5) Seetion
The result (3·86) is just the first term of an asymptotie series for the gamma
funetion, whieh we shall now findo First write

Let us now set


The constants A, B, ... may be found from the recursion relation for the
gamma funetion:
r(z + 1) = zr(z)
For, from (3·87),

r(z + 1) = J2n(z + l)Z+1/2e~(Z+1)[1 + A + B + ... ]

z + 1 (z + 1)2

= J2~exp[(z +D log(z + I)-(z + 1)]

x [ 1+ A + B + ... ]
z +1 (z + 1)2

~ J2n exp [(, +~)

logz - z + (_I
- _I
12 z 3
+ 2...
40z 4
- + ...)]
.x [ 1+ A + B + ... ]
z+l (z+I)2

A B-A C-2B+A ]
[ + + ...
z Z2 Z3
90 EvaIuation of Integrais

On the other band, (3 87) gives immediately


zr(z) = J2tt zZ+ lf2 e- Z( 1 + : + ~ + ...)

Equating corresponding terms in the two series we find
A ~ 1/12, B ~ 1/288, ...
This result agrees with the formula (3 79) previously quoted for z = n.

We have not proved the validity of this heuristic procedure, but it gives the
right answer.
One last procedure for approximately evaluating integrais should be
mentioned. This is the so-called method of stationary phase; it is concerned
with integraIs of the form


where IX is large and positive, andf(z) is real aIong the contour C. Unless
f'(z) = O, the contributions to I from the neighborhood of z will largely
cancel because of the rapidly oscillating character of el~f(z). Thus we look
for points along the contour where f'(~) = 0, and use the result

This method is clearly very closely related to the method of steepest descent.
Note that if the integrands for the two methods are written in the same form,
é'"f(z>, the paths of integration are along a curve Imf(z) = constant for the
method of steepest descent, and along a curve Re f(z) = constant for the
method of stationary phase.

Convenient collections of useful integraIs are to be found in Dwight (D8)
and Pierce and Foster (PI). Grobner and Hofreiter (G8) and Gradshteyn
and Ryzhik (G6) give more extensive Iists. The possibility of finding a
given integral in a table of integral transforms, such as tbose of Erdelyi
et ai. (E6), should not be overlooked.
Many useful tables and graphs of functions defined in this and later chapters
may be found in Jabnke et ai. (B), and Abramowitz and Stegun (Al).
Additional properties of these functions are collected in Magnus et ai. (Ml)
and Erdelyi et aJo (E5). In particular. the reader who is- faced with the
problem of actual1y evaluating some elliptic integral will find convenient
transformations and reduction formulas in Magnus et aI. (MI) Chapter X;
Problems 91

Abramowitz and Stegun (AI) Section 17; lahnke et aI. (J3) Chapter V;
MUne-Thomson(M7) pp. 26-38; and Hancock (H6). Also, the lovely mono-
graph by Artin (A6) on the gamma function should not be overlooked.
Evaluation of integrais by contour integration is treated in many books;
among them are Copson (C8) Chapter VI; Whittaker and Watson (W5)
Chapter lU; and Morse and Feshhach (M9) Chapter 4.
Asymptotic expansions are treated by Jeffreys and Jeffreys (J4) Chapter
17; Whittaker and Watson (W5) Chapter VIII; and 8mith (85) Chapter 8.
Two smaller books, devoted specifically to asymptotic expansions and related
questions, are Erdelyi (E3) and de Bruijn (D4). Saddle-point methods are
discussed in the last-named reference, as well as by Jeffreys and Jeffreys
(J4) Chapter 17; Smith (S5) Chapter 8; and Morse and Feshbach (M9)
Chapter 4.

Evaluate the integrais in Problems 3-1 to 3-9:
3-1 folO e-li' - e- b1 dy
o y

3-2 Ia" sin bx dx (applya convergence factor; do integral; remove tbe

convergence factor)
olO cos ax dx
o 1 + x2

(the symbol dJx stands for dx dy dz, ar in general a

volume element in three dimensions)

i dx 1 + x
f o x
In .,--'--::
1 x
Hint: Expand the integrand in a power series.
" dx
fo cash x
Hint: Expand the integrand in a series wbicb is useful near x = 00.
92 Evaluation of Integrais

Do the integrais 3-10 to 3-24 by contour integration:


3-11 (put poles slightly above real axis)

<Xl X2 dx
f -<Xl (a 2 + X 2 )2
f (a 2 + r 2 )3
+1 dx
3-14 { (a>b>O)
-1 Ji-x 2 (a+bx)
<Xl X dx
f o 1 + XS
h sin2 8 d9
f o a+bcos8
(a> Ibl)

<Xl sinh ax d
3-17 { . x
-<Xl smh n:x

1+ x


00 dx
3-20 (a> O, b > O)
fo (a + bx 2 f

"" dx
o (a 2 + X 2 )3
'" sin x dx
f x(a + x
o 2 2

h d9
t (a + b cos 8)2
(a> b > O)

Problems 93

'" In x dx
. 3-24
fo (x + 1)2

3-25 Evaluate Io e-"'z Ci (ax) dx

3-:16 Evaluate J,• e- u

erC x dx

3-27 Evaluate (a) rm~fi" (d) B('. -tI

(b) rm (e) rm r( -tI
(e) B(l. 3)

3-28 Evaluate {' e- U [ -Ei(-x)] dx

J..2' Evaluate I-,(4 - Xl )116 dx in terms ofbeta and gamma functions.

• 3-30 Consider the integral F(z) = Se dt( - ty-1e- r where the t-plane
is cut along the positive real axis, (_W- 1 is detined to equal
exp [(z - 1) In (- t)] with In (-I) real on the negative reaI t-axis,
and the path af integration C comes in from t = + 00 below lhe cut,
goes around the origin and retums to t = + 00 above the cut. This
integral defines tbe gamma function r(z) throughout the complex
plane [unlike the detinition (3-46)]; more precisely, F(z) = (something)
x r(z). Evaluate ,(something). (~tN.ttit.)
3-31 (a) Consider the contour integral §f(z) ctn 1I:Z dz around a suitable
large contour, and obtain tbereby a formula for the sum

(b) EvaIuate

3-32 The absorption mean free path À for neutrons in a certaio material
is measured by ao absorptioo experiment as follows: A thin foi!
detector in the sbape of a disk of radius b is irradiated at a distance
a from a point source of oeutroos, and its aetivity Ao is measured.
Ao absorber of tbickness Tis tben placcd betweeo the source and foil
as sbown. aod tbe activity A produced in tbe same length or time is
messured. Find À if a = b = 12 em, T - 1 em, and A/Ao = 0.25.

94 Evaluation of Integrais

1. Source emits neutrons isotropicaJly.
2. Neglect seattering of neutrons.




3. In traveling a distance d through absorber, neutron intensity is

reduced by a faetor e-di). .
4. Activity produced by neutron in foi! is proportionaI to distance
traveled by neutron in foil.

3-33 Evaluate, r(z)~ dz around the contour Izl = t once in a positive

3-34 Obtain two expansions of Si x, one usefuI for small x and one usefuI
for large x.
3-35 Evaluate l(x) = Só' dt e"'t-e' approximately for Iarge positive x.
3-36 Energy in a star is produced by nuclear reactions. The number of
collisions with' eM kinetie energy in the interval from E to E + dE is
per unit time, where k is Boltzmann's constant, Tis the temperature,
and N is a constant. The probability that a collision with eM
kinetic energy E will result in a nuclear reaction is

where M and OI: are constants. Find an approximate expression for

Prohlems 95

the total number of nuclear reactions per unit time, assuming


3-37 Verify the result (3-31).

3-38 Find an asymptotic approximation for large x (real and positive)
to the function

f~(x) = Ice-i" .Int+intdt

where C is the contour shown below.



In this chapter we discuss Fourier series, Fourier trans-
forms, and Laplace transforms. Some other transform
pairs are Iisted in Section 4-4, but not discussed. The
use ofintegral transforms in soIving problems is briefiy
illustrated by a few examples in Section 4---5. Further
applications af integral transforms to the solution af
partial differential equations and integral equations will
be made later, in Chapters 8 and 11.


We begin by considering a function 1(0) defined for O ~ 8 < 2n. We
seek an expansion af the form
A 00

f(O) = 2° + "~l (Ali cos nO + BII sin nO) (4-1)

The coefficients may be found by multiplying both sides af (4-1) by cosnO

(or sin nO) and integrating from O to 21'1::
Ali = -
" o
1(0) cos nO dO
[ ,.
Bit = . . f i(O) sin nO dO
" o

C1rJ mi3 . M ti B J .f) = (li
t.;{n ml1, mn fl "6 ~4)- == Ti drnn. l o,
,, «'l OYJm'lt. /4Ío 0'9- J4}= ('l"'if ltil(rn~' UY:lrtfl JlI =- O == 1,'2, ~""._).
•• o
4--1 Fourier Series 97


-2~ -~ 9

Figure 4-1 An even function of period 2n

The series (4-1), known as the Fourier series of f(O), converges at alI points to
-!-[f(O+) + f(e-)] provided f(e) is of bounded variation! in O ~ e < 2n.
We have not worried about rigor in the above discussion, and refer the
interested student to a mathematical reference book such as WhiUaker and
Watson (WS) Chapter IX or ApostaI (AS) Chapter 15.
A Fourier series is periodic so that it repeats f(O) in 2n ~ e < 4n,
-2n:S:;; < O, etc. We need not have started with the interval from O to 2n;
any interval oflength 2n would do. Often the interval from -n to n is more
One should be able to recognize types of functions whose Fourier series
obviously have certain terms missing.
I. Even functions: f( -e) = f(O), ar f(2n - O) = f(O), etc. Only cosine
terms occur, that is, B" = O. An example is shown in Figure 4--1.
2. Odd functions: f( -O) = -f(O). Only sine terros occur, that lS,
A" = O. An example is shown in Figure 4--2.

-2~ ~ 9

Figure 4-2 An odd functioD af period 2n

1 For a function f(x) defined on some interval, choose a set af n points X, fram that
interval and evaluate V (the "variation") = If(x,) - f(X2)1 + If(X2) - f(x 3 )1 +."
+ If(x__ I) - f(xJI. If there is a constant B. such that V < B for ali choices af x. as
11 -+ <XI, we say thatf(x) is of bounded variatwII on thatfntervaL
98 Integral Transforms


Figure 4-3 An even function symmetrlc about 71:/2

3. Even functions symmetric about n/2 :f«7I:/2) + 8) = f«7I:/2) - O).

Only cosine terms with even n occur, that is, B~ = O, A2~+1 = O. Figure 4-3
shows an example.
Obviously there are many such possibilities. We shall now consider
a few examples of Fourier series.


1(0) ~ (+1

This is an odd funtion, so there are no cosine terms. It is symmetric

about n/2, which means that there are no even sine terms. For odd n,

B" = -
1 f'"1(0) sin nO d8
= -
1(0) sin n8 dO = -
71:0 no nn
sin30 + sin50 + ... )
1(8)=- sinO+ (4-4)
" 3 5
This series exhibits the nonuniformity of the convergence of a Fourier
series near a discontinuity. Successive approximations are illustrated in
Figure 4-4. Note the overshoot, which is called Gibbs' phenomenon. In


\w --- +---f---l- ---
O 6
...., -I

Figure 4-4
Partial sums of tbe Fourier series (4-4), witb increasing
number of terms
4-1 Fourier Sedes 99

the limit af infinitely many terms, tbe overshoot remains finite-approxi-

mately 0.18 (see Problem 4-2).
Ir we set (} = n/2 in tbe Fourier senes (4-4), we obtain
11: 1 1 1
4 3 5 7

a senes known as Gregory's series. lt may aIso be derived from the series
x3 x!
tan- x 1
=x - - + - - + ...
3 5


f(8) - cos k8 (-x<8<x)

The Fourier sedes for cos k8 was already mentioned in connection with
Bernoulli numbers (Section 2-3).
Sincef(8) is even, anly cosiDe tenns are present.

A" = -2.f", cos k8 cos n8 de



We can, af course, generalize Fourier senes to represent functions wbich

are periodic with some period L other thao 2n. Let
Then ao interval af length 211: in the variable (} becomes ao intervaI af lcngth
L in the variable x. Qur formulas become

f(x) = iA + 1I~1
• (
. 2nnX)
+ B. Stn L
2 f
A" = L o!(x)cosLdx
L 2nnx

k-,> ~[fCe+o)+-f(ho)J
Ct-I--'"'oQ )

Ali our remarks about evenness, oddness, and 80 aR may be carried aver to
this new interval.
One must leato to be clear about what the fundamental interval L is for
aoy particular problem. Suppose a functionf(x) ia given to us in the interval
O < x < a. We may expandf(x) in a Fourier series af period a:

BII SIO =::=

A, • ( 21tnx . 2nnX)
!(x) = -2 +n=1
~ A~ cos
+ a

Clearly we need both the cosines and sines for an arbitrary!(x). In other
words, sines ar cosines (with period a) alone are incomplete.
00 the other hand, we can expand in sines a/one by defining /(x) on
-Q < X < O by f( -x) = -!(x). Now take 2a to be lhe period af our
Fourier series, 80 that
f(x) = L B, , i n

- (4-9)
11"" 1 a
Thus, 80 to speak, we have thrown away the cosines but doubled the number
af sine terms, and we still have a complete set of fUDetioos. We could, of
course, also expand !(x) in a series containing only cosines of period la
by. defining !( - x) = +!(x) for O < x < a.
It is often convenient to write Fourier series in complex form:

f(8) = L a,e'"
l i " -00

ar, more genera11y,

li" -00

To evaluate the coefficients a" in (4-11), multiply both sides by e- 2d..u;IL and
integrate from O to L (ar any other interval of length L). Then we observe
f, e-2..J..."ILe2..JIIXILdx = Lb".,. (4-12)

where ó",,, is called the Kronecker delta and is defined by

ifm # n
ifm =n

4-2 Fourier Transforms 101

Be sure to natice the change in sign in the exponential:

f(x) = L a n e2ninx/L (4-14)

The quantity (l/L) J5lf(xW dx, namely, the average absolute square
af I(x), is often af interest in physical applications. Expanding f(x) In a
complex Fourier series (4-11), we obtain 2


This result (4-15) shows that each Fourier component af f(x) makes its
separate contribution to the integral J~ If(xW dx independently af the
other Fourier components. There are no interference terms af the form
a: am • [Compare the comments made below in connection with the example


We begin with the complex Fourier series af a functionf(x):

f(x) = L a n ehinxlL a" 1

= ~
fL'2 .
j(x)e-1n,nxIL dx
• L -L/2

We wish to consider the case L --l> 00. Then the sum may be converted into
an integral as follows: define
-~y and
Then, since n increases in steps of unity in the sum,

1 Notice the general rule:

L (anythlng)m S... = (anything)"

provided the range of summation over m includes tbe value n; otberwise, of course, the
sum 15 zero.
102 Integral Transforms

where F(y) = FII' Our formulas thus become

f(x) - -1 f" g(y)e'" dy

21& - ""
g(y) - L/(x
v '" dx (4-16)

g(y) is called the Fourier transforrn of f(x), or vice versa. The position of

- -
the 2n: is quite arbitrary; oeten one defines things more symmetrically,

f(x) =J L
""y(y)e'''' dy g(y) -
Jf 1 "
-2n: _ <Xl f(x)e-'" dx (4-17)

We may combine the expressions (4-16) [or (4-17)] to obtain

f(x) = -1 f"
2n: _""
dy e'''' f" f(x')e-
dx ' l

= f dx'!(x')-f e'("-"')Ydy
" 1 "
-<Xl 21t -""
The fact that (4-18) holds for any functionf(x) teUs us something remarkable
about the integral

considered as a function of x'. It vanishes everywhere but at x' = x, and _

its integral with respect to x' over any interval including x is unity. That
is, we may think of this function as having an infinitely high, infinitely
narrow peak at x· = x.
It is conventional to define the (Dirac) deltafunction 3 b(x) by

..f-. ~(x) - O

~(x) dx - 1

a•. b > O

These equations imply

f f(y) ~(x - y) dy -f(x) (4-20)

for any functionf(x), provided the range of integration includes the point x.
Comparison of(4-18) and (4-20) now shows that

27t - ""
b(x) = -1 f" e'''' dy (4-21)

which is an integral representation of the b-function.

3 See Dirac (D6) Section IS.

4-2 Fourier Transforms 103

We may use the delta fundiao to evaluate the important integral

Lrolf(x)l' dx
in terms ofthe Fourier transform (4-16) af/(x), as follows:

f ""

_ DO
If(xW dx = f dx -2n1 f_ g*(y)e~IX1 dy -2n1 J- '" g(y')é"" dy'

- co


~ - f dy g*(y) f dy'g(y') - f
1 00 1 '"
21t _'" 2n -'"

~ -2n1 f • -00
dy g'(y) f•
dy'g(y') b(y' - y)

1 "
= 2nf_oo dy g*(y) g(y) (4-22)

This is called Parseval's theorem. It is useful in understanding the physical

interpretation af the transform function g(y) when the physical significance
afj(x) is known [see the discussion following (4-26)].
Suppose f(x) is ao even function. Then
ro ,
g(y) = f, !(x)e- dx f .f(x)e-

= f,•f(x)(é"Y + e-i",) dx
2 f" f(x) cos xy dx (4-23)
Now note that g(y) is even. Therefore
f(x) = - f g(y) cos xy dy

rr ,

[(x) and g(y), which now need only be defined for positive x and y, are called
Fourier cosine transforms of each other.
By considering the Fourier transform af an odd function, we similarly
obtain the relations hetween Fourier sine transforms
=- f J, f(x) sin xy dx
1 •
f(x) g(y) sin xy dy g(y) = 2 (4-25)
n o
We may symmetrize by putting ./i/rr
before each integral in (4-23), (4-24),
and (4-25) ifwe wish.
Extensive tables of Fourier transforms are given by Erdelyi et ai. (E6).
104 Integral Transforms


O (t< O)
f(t) ~ (4-26)
(e- r/T sin W o t (t > O)

This function is shown in Figure 4-5. I(t) might represent the displacement
of a damped harmonic oscillator, or the electric field in a radiated wave,
or the current in an antenna, for example.
The Fourier transform of I(t) is

g(w) ~ f-.• f(ty'" dt

1 1 1
w+w o - -
w-w o - -

We may interpret the physical meaning of g(w) with the help of Parseval's
theorem (4-22). For example, if/(t) is a radiated electric field, the radiated
power is proportional to If(tW and the total energy radiated is proportional
to J~ If(tW dI. This is equal to

by Parseval's theorem. Then Ig(w)1 2 may be interpreted as the energy

radiated per unit frequency interval (times some constant).


- /

Figure 4-S A damped sine wave

4-2 Fourier Transforms lO5

Ig(ro) I'



Figure 4-6 Energy spectrum for the damped oscillation of Figure 4-5

Let us assume that T is rather large (w o T p 1). Then our "frequency

spectrum" g(w) is sharply peaked near w = +wo. For example, near
w = wo,
1 1
g(w) '" -
2 ,
w - Wo - T

Ig(w)1 '" ~ J (w -
W O)2 + T1

When w = W o + I/T, the "amplitude" g(w) is down by a factor J-lji,

and the radiated energy Ig(wW is down by a factor 1/2. In other words,
the width í at half (power) maximum is given by r;:::: 2/T. The energy
spectrum Ig(wW is sketched in Figure 4-6.
This result is a typical uncertainty principie and is very closely related to
the Heisenberg uncertainty principie in quantum mechanics. The length
of time (T) during which something oscillates is inversely proportional to
the width r, which is a measure of the " uncertainty" in frequency.

Fourier transforms may easily .be generalized to more than one dimensiono
For example, in thrce-dimensional space we have the transform pair
q,(k) = f d xf(x)e-
3 I
f(x) = f (2,)

Frorn these we rnay deduce, as before, the integral reprcsentation

c5(x) = f d 3k
e'''''' (4-28)
106 Integral Transforms

where the three-dimensional delta function is defined by

ó(.) ~ O .,;0

f d x ó(x)
= I
provided origin is ins;de region of
integration (4-29)

fd'xf(.) ó(x - y) ~ f(y)

Such transform pairs are of interest in quantum mechanics. If f(x)
is the wave function of a particle, the Fourier transform <jI(k) is the so-caIled
"wave function in momentum space." If(xW and 1<jI(k)1 2 are the probability
distributions for the position and momentum, respectively.



This is a wave function which gives a Gaussian 4 probability distribution,

If(x)1 2 , centered at r = 0, and normalized so that Jd 3 x If(xW = I. The
Fourier transform is

Introduce polar coordinates, with the z axis along k. LeI cos e= ct. Then

cP(k) = 2rrN J"

1"2 dr f "' dct e-,2/a'e-ih~

~ -4rr N fOC> r dI" e-r '",a sin kr

k O

__ N f'" "
rdre- r I~ 2 e'., r
k 2i -OC>

= ~1T Ne-k'a2/4 f'" r dr exp [_ ~2 (r _ ika ),]
Ik -OC> a 2

2rr N , -k'a'/4
00 (
y +-
ika- )
d y' -y'la'

.. Funçlions of lhe form Ae-s~2 (A, B constant) are often cal1ed Guussian funclions
because of Iheir occurrence in lhe least-squares method of data anaJysis, which originated
with Gauss. See also lhe foolnole on p. 378.
4-3 Laplace Transforms 107

Finally, remembering N = (2/na 2)31\ we obtain

$(k) = (2na2)314e-k2d~f4 (4-31)
Note that the Fourier transform of a Gaussian is another Gaussian. The
narrower the distribution in x (that is, the smaller lhe value of a), the broader
is the distribution in k. The widths 6x and Ai< are roughly inverse to eaeh

The preseription p = hk from qllantum meehanics (p = momentllm) converts

his into another qllantllm mechanical llncertainty relation

L<.xL<.p '" h
The probability distribution 1$(kW found above is a Gaussian centered
at k = O. We might expect that a similar distribution centered at an arbitrary
ko could be obtained, starting with the same probability distribution in
position If(r)1 2 • This may be done by simply mllltiplying f(x) by a factor
e/ ko ''', as the reader may easily verify.


We are often interested in funetions whose Fourier transforms do not
exist. For example, the simple funetionsf(x) = A = constant, andf(x) = x 2 ,
have Fourier transform integrais whieh do not converge. For many fune-
tions, the trouble at x -+ + 00 may be fixed by multiplication with a faetor
e-c" if c is real and larger than some minimum value (J;. This faetor may c. >0< '''C
make the behavior at x -+ - co worse, but we are often interested in a funetion
only for positive x. Thus we can take care of the behavior for negative x
bya second faetor, the (Heaviside) step funetion:

[Note that (dfdt)H(t) = (j(t), as can be shown by integrating the (j functiou.]
The functionf(x)e-C"H(x) now has the Fourier transform:

The inverse transform is

f(x)e-C"H(x) = -
f •
27t _ '"
g(y)é X
)' dy
108 Integral Transforms

It is conventional to introduce a new variable,

s = c + iy (4-33)

and to define F(s) = g(y). The above two integrais become



where the path of integration eis upward along the straight line, Re s = c =
constant (Figure 4-7).
F(s), as given by (4-34), is called the Laplace transjorm of j(x). The
integral exists only in the "right-half" s-plane, Re s > a, where a is
the minimum value for c mentioned above. In this region, F(s) is ana-
lytic; F(s) may usually be defined in the left-half plane by analytic continua-
The second integral (4-35) is called the Laplace inversion integral. Note·
that for x> O it gives f(x) {or more precisely Hj(x +) + j{x- )]} but for
x < O it automatically gives zero; when x < O the contour C may be closed
by the addition of a large semicircle on the right, where F(s) is analytic.
We shall hereafter omit the H(x) in (4-35), and write simply


with the understanding that ali functions j(x) which are to be Laplace
transformed vanish for negative arguments.


Figure 4-7 Contour for the Laplace inversion integral (4-35)

4-4 Other Transform Pairs 109

f(x) ~ 1
F(s) = S"'!ex)e-U dx
= f""o e-": dx = ~,
where the integral exists for Re s > O (that is, IX = O). Note that F(s) has a
singularity (a sim pie pole in this example) 00 the limiting line, Re s = O.
We may verify the inversion formula:
1 e'X
f(x) = - .
f c+l",

F(s)e'X ds = -.
f - d, 0+100

c-ioo S

where c > O.
If x> 0, we complete the contour by a large semicircle to the left, and
f(x) ~ 1

Ir x < O, we complete the contoue to the right, and

f(x) ~ O
A short table af Laplace transforms appears at the end af this chapter.
More complete tables may be found in Erdelyi e/ai. (E6); Magnus et aI.
(M1); and elsewhere.


Fourier-Bessel transforms:
(or Hankel transform)
g(k) ~ So• f(x)J.(kx) x dx (4-38)

f(x) ~ 1,• g(k)J.(kx) k dk (4-39)

Mellin transform: cfJ(z) = Ia t f(/) dt Z


Hilbert transform: g (y) ~ -1 P S" f(x) dx (4-42)

1t _"" X Y

f(x) ~ ~ p
r -00
g(y) dy

[P denotes that the (Cauchy) principal value of the integral is to be taken;

see Appendix, Section A-2, number 8.]
ilO Integrai Transforms


We shall first summarize some basic properties of Fourier and Laplace
i. Both transforms are linear, that is, thetransform of af(x) + fJg(x) equals
a times the transform of f(x) plus fJ times the transform of g(x).
2. Derivatives:

e-sxf'(x) dx = e-SX!(x)I:f + s J'oo e-SXf(x) dx
We shall write Z[J(x), s] or Z[J(x)] for the Laplace transform of f(x),
and similarly 3"[j(x)] for the Fourier transformo Then our reiation above
may be written
2'[l'(x)] ~ ,2'[f(x)] - j(O) (4-44)
2'[/"(x)] ~ ,'2'[f(x)]- ,j(O) - 1'(0) (4-45)
and so forth (note that O reaIly means 0+, the limit as zero is approached
from the positive side). For Fourier transforms the integrated parts vanish,
F[f'(x)] ~iyF[f(x)] etc. (4-46)
3. IntegraIs:

fo dlf,

~ -f

dtf(r) e-sI



For Fourier transforms, things are not quite so sim pie. Suppose
g(x) = Sf(x) dx is an indefinite integral off(x). Then, from (4-46),

F[f(x)] ~ iyF[g(x)]
4-5 Applications of Integral Transforms 111

However, we cannot imreediately conc1ude that

"'[g(x)] ~ "'[f(x)]
Why not? Consider the equation xf(x) = g(x). Can we conclude that
f(x) ~ g(x)
No, not at x = O. The general result is
f(x) ~ g(x) + C .(x)
where C is an arbitrary constant. Thus

"'[J f(x) dx 1~ "'[~X)] + C .(y) (4-48)

This arbitrariness is also obvious from the fact that Jf(x) dx is uncertain
to within an arbitrary additive constant C, and
"'[C] ~ 2xC .(y)
4. Translafion:

f f(x + a) e-

9'(f(x + a)] =

= f_""f(x)
e-iy(x-a) dx

"'[f(x + a)] ~ "·''''[f(x)] (4-49)
For Laplace transforres, we must be a little more careful. Consider the
cases a > O, a < O separately.
For a > O,f(x + a) is shown in Figure 4-8b.

f(x) f(x + a) f(x - a)

(a) (b) (c)

Figure 4-8 j(x) translated to the left, f(x + a), and to the right,
f(x - a)
112 Integral Transforms

Since the Laplace transform ignores f(x) for x < O (in fact, assumes it is
zero), we must chop off some of our function, and

so that

a> O (4-50)

On the other hand,J(x - a) is shown in Figure 4-8c, and

.T[J(x - a)] = t 00

f(x - a) e- u dx

= f-.f(x) e-;(""+~) dx

2'[/(x - a)] ~ e-"2'[I(x)] a>O (4-51)
5. Mu/tiplication by an exponential:
The following two formulas are easily verified:
F[e7(x); y] ~ F[I(x); y + i"] (4-52)
2'[,"[(x); ,] ~ 2'[/(x);' -"] (4-53)
6. Multiplication by a power Df x:
g(y) ~ J 00

f(x) e-o" dx



F[xf(x)] ~ i <i. ,"U(x)] (4-54)

An analogous result is easily shown to hold for Laplace transforms:
2'[xf(x)] ~ - d, 2'[J(x)] (4-55)

4--5 Applications af Integral Transforms 113

7. Convo/ution theorems:
Let !J(x), f2(X) be two arbitrary functions. We define their convofution
(faltung in German) to he

g(x) ~ LJ'(y)!,(x - y) dy (4-56)

What is the Fourier transforrn af such a convolution? A straightforward

change af variable shows that
§[g(x)] = §[fl(x)]F[f2(X)] x constaot (4-57)
The value af the constaot in (4-57) depends 00 our convention for Fourier
transforms, that is, whether we use (4-16) or (4-17) and, ir (4-16), which af
f(x) and g(y) is considered the original functioo and which the Fourier
transform. The student is urged to evaluate the constaot in (4-57) for at least
ooe convention. In any case, the importaot result is that, to within some
constaot, the Fourier transform of a convolution is the product ofthe Fourier
transforms of the "factors" of the convolution.
An analogous result holds for Laplace transforms; if

g(x) ~ fo• dI !,(I)f,(x - I) (4-58)

-"'[g(x)] ~ -"'[f,(x)]-"'[f,(x)]
An interesting converse relation holds for Laplace transforms. Suppose
where the Laplace integrais for 91(S) and 9is) exist for Res > ai and Re s >
<1. 2 , respectively. Then the Laplace transform ofthe productflf2 is given hy


where the path of integration is along the line Re z = c, with (Re s - a J ) >
c> ai' This result may be obtained by substituting the Laplace inversion
integral for !tez) in the integral Z[fd2].
The corresponding relation for Fourier transforms is s


~ Again, the faclor lj21T depends on one's convention for Fourier transforms; see remarks
after Eq. (4-57).
114 Integral Transforms

We now consider several examples ofthe application ofintegral transforms.

The general procedure is to "transform" the problem, ar restate it in terms
ofthe transform function, and hope that in this form it is easier to solve.


Find the current in the circuit of Figure 4-9 if the switch is closed at time
t = O, and the initial charge on the condenser is Qo.

dI Q
RI + L- + - = Eo (4-61)
dt C

Strict1y speaking, we should write Eo H(I), where H(t) is the step function
(4-32), but in dealing with Laplace transforms we assume everything vanishes
for t < O anyway.
Since dQfdt = /, Q(t) = Qo + J~ I(r') dt', and (4-61) may be written

RI + L dI + ~
dt C
[Q, + f'l(t') dt'] ~ E,

Now take the Laplace transform ofboth sides. Let 2[/(1)] = i(s):

Ri(s) + L[si(s) - 1(0)] [Q,

+ -1 - + -i(')] ~-
C s s s

Solving for i(s) gives (by simple algebra!)

i(s) = L C 1
(s + a)l + b 1

b= J~C- :~~ [and 1(0) ~ O]

Figure 4-9 Serles RLC circuit
4-5 Applications oflntegral Transforms 115


,ze . b1
"[ -" Slnt=( b
)' ,
s+a +b


E _ Qo
I(t)= o C e-a<sinbt
L b


Consider the coupled pendulums shown In Figure 4-10. Assume the


Figure 4-10 Coupled pendulums

initial conditions XI = X2 = O, Xl = V, Xl = O at t = O. Newton's equations


Let .P[x/t)] = Fls). Then the Laplace transforms of our two differential
equations (4-63) are
116 Integral Transforms

We must now solve these simultaneous algebraic equations for FI and Fl'
We find

~- 1 +_----,1-,
2 9
s +-
+ 2 -mk S2 +-9I
J--- J 9
I m


J g
Similarly, we can work out x 2 (t). Note that g/I +'2(k/m) and J J U/'
are the (angular) frequencies of the two normal modes.

We shalI conclude this chapter by carrying out an alterna tive derivation

of the WKB connection formula (l-Il3) given in Chapter I, relating the
exponentialIy small solution on one side of a turning point to an oscillatory
solution on the other side. Consider the differential equation
dx 2 + xy = O (4-64)

This equation has a solution y(x) which looks something like the sketch in
Figure 4-11. Let

gero) = f-.• y(x)e- iW

>: dx

Then the Fourier transform of our differential equation (4-64) is

_ro 2g(ro) + i dU = O
which gives, inverting the transform,
g(ro) = Ae- i (W'/3l (A = arbitrary const.)

y(x) = A f~oJ ~: exp [i(rox _ ~J)] (4-65)

4-5 Applications af Integral Transforms 117

ri Y(x)


v v

Figure 4-11 Sketch of the solution of (4-64), which has a decreasing

exponential character for x < O

We shall drop the irrelevant constant Aj27t.

Unfortunately, this integral is nontrivial. Partly for this ceason it has a
name, being known as an Airy integral. For the present application af
finding the WKB connection formula, we anly need the asymptotic forms
for large Ixl, which may be found by lhe saddle-point method af Section 3-6.
The two cases x -. + (J) must be considered separately.
I. Let x---+ +00. The saddle points oecur on the real axis, and we must
go over both af them. Using lhe notation introduced in Section 3-6.

I'(w) ~ i( 1 - :') r(w)~


() = +-
~~ +-4


y(x) ~ I J2n:..;X exp [+j-iX 3/2] exp [+ in]

± 2x 4

y(x) - ~t cos (iX3/2 - ~) (4-66)

118 Integral Transforms

2. Let x---+ -00. Now the large positive parameter is -x, and

f(w) = -i(W _~~) j'(w) = -1(1 _:2) 1"(0) = 2:W

1'(0)0)=0::;'0)0= +i~
The saddle points now occur on the imaginary aXÍs.
2' +2
p'" ~ -'x ( + iFx) ~ j---,';=~
- x
8 ~ (~)

ridge ridge
v = constant = O v = constant = O


" w1 - +i~//
, , valley /' /'....-, '"
..... /' 6' (asymptote)
"" ""

ridge u _-00
v = constant =O
u -;.. 00

Figure 4-12 Topography of the surface u = Ref(w) in the w-plane

near the saddle points w1,2 = + h/ -x, for the
f(w) = -i[w+w3 j3(-x)]
Ifw=réo, u = Ref(w)=r sin 8+ [r 3 j3(-x)] sin 30, and
v = lmf(m) = -r cos V- [r 3 j3( -x)] cos 3V

References 119

The topography of the surface u = Ref(w) is indicated in Figure 4-12. w,

go through the saddle point at w 2 = - iJ -
x, with tP = O. Then

y(x) - J2nJ - x oxp [_( -x)

2 ( x) 3
~ ~]
y(x)"'" (_X)1/4 exp
-"3 (-x)
01'] (4-67)

We see that our two asymptotic formulas (4-66) and (4-67) are in fact
the respective WKB solutions, and that they are indeed related by the con-
nection formula (l-113), derived previously in Chapter 1.

In Table 4-1 we list a few of the most frequently occurring Laplace trans-
forms; the results enumerated as 1, 2,"', 7 at the beginning of Section 4--5
may be used to extend this list.
Table 4-1 Laplace Transforms

f(x) F(s) = f " e-'~f(x)


1 ,-
o(x - xo} (Xo > O) e-'~O

sinÀx $2 +À2
cos "-x 82 +À2

x" $"+ i


Rigorous discussions of Fourier series may be found .in ApostoI (A5)
and in Whittaker and Watson (W5) as well as in other mathematical texts.
Fourier integrais are treated by Morse and Feshbach (M9) and by
Titchmarsh (T3). Do not be deceived by the title of the last-named refer-
ence; it is not for the casual reader.
In recent years, various new formalisms have been developed, among
whose advantages is rigorous treatment of delta functions and similar
functions. Readahle discussions of some of this new material are given in
120 Integral Transforms

Lighthill (L8); Mikusinski (M5); and Erdelyi (E4). The first of these three
is especially recommended.
The reference work of Erdelyi et aI. (E6) contains useful tables of Laplace
and Fourier transforms as weIl as several other integral transforms.
Many examples of the application of Laplace transforms to physical
problems can be found in Carslaw and Jaeger (C2).

4-1 Expand the functionf(x) shown below, in a Fourier series.

f(x + L) ~ f(x)

L x

4-2 (a) Consider the Fourier series for

f(O)~(+1 O<O<:n
-1 :n <O <21t

f(8 + 2n) ~ f(O)

Just to the right of O = O, the sum of the first n terms looks as
shown below.

o 8

Find bn , the "overshoot" of the first maximum.

(b) Show that
lim bn ~ 0.18

(Evaluate it precisely.) This is known as Gibbs' phenomenon.


Problems 121

4-3 A functionf(x) equals e-x for O < x < l.

(a) Expandf(x) as a Fourier series af the form Ln Bn sin nnx.
(b) Expandf(x) as a Fourier series af period 1.
4-4 A linear system is driven by a periodic inputf(t), such that/(t + T) =
f(t). The response af the systern is such that a sinusoidal input
af angular frequeney ú) is multiplied by (000/00)2, unless (I) = O, in
which case no output Dccurs. The output can be written in the form

g(t) ~ -1
T ,
f' G(r - r')f(r') dr'

Find the funetioo G(t).

4-5 or what [uDetion is
LI cos 38 cos 50
COSu+ 9 + 25 +, ..
the Fourier series? Work it out; don't look it up!
4-6 Show that

SC,. ' f(x,)

l:~ H<) f
• f(x) ó[g(x)] dx ~ Ig'(x,)1

provided g(x) = O has a single root Xo in the interval a < x < b.

4-' Evaluate fI< dx

, f, dy b(sin x) ó(x 2 _ y2).

4-8 Pind lhe Fourier transform af the wave function for a 2p electron
in hydrogen:
rJt(x) = 1 ze-r/1ao
J32na õ
where ao = radius of first Bohr orbit and z is a rectangular coordinate.
4-9 A linear system has a response G(w)e- illl ' to the input signal e-Illl'
(warbitrary). If the inputf(t) has the particular form
O (r <O)
f(t) = (e-À' (I > O)

where À. is some fixed constant, the output is observed to be

(t < O)
(t > O)
where Cl is another fixed constant.
(a) Find G(w).
(b) Find the reponse of the system to the inputf(t) = A Ó(I).
122 Integral Transforms

4-10 Find the Laplace transform .2"[f(x)] of the function sketched below.


I •••

I 2 3 4 5 6 ••• x

4-11 A functionf(x) has the series expansion

!(x) ~ L• ,.x

~_o n!
Write the function g(y) = I:'=o cn y~ in elosed form in terms of f(x).
4-12 By using the integral representation

Jo(x) = -1 f" cos (x cos fl) de

2, o
find the Laplace transform of Jo(x).
4-13 Ofwhat fllnction is

(S2 + 1)(s 1)
the Laplace transform?
4-14 Three radioactive nuelei decay sllccessively in series, so that the
numbers N 1 (t) of the three types obey the equations
- - = -).tNt
- - = ).lN 1 - A.2 N 2
Tt=A.2N2 -À 3 N 3

If initially N 1 = N, N 2 = 0, N) = n, find N 3 (t) by usmg Laplace