Internet Supplement for

Basic Complex Analysis

Third Edition

Jerrold E. Marsden and Michael J. Hoffman

November 3, 1998
ii Preface

Preface
This document consists of a series of supplements to the third edition of our book
Basic Complex Analysis. Some of the topics give additional technical details of
results that were stated but not proved in the textbook while others treat additional
topics and applications that may be of interest to some readers.

Pasadena, CA Jerry Marsden and Mike Hoffman

Fall, 1998
Contents

Preface ii

1 Analytic Functions 1

2 Cauchy’s Theorem 7

3 Series Representation of Analytic Functions 15

4 Calculus of Residues 21

5 Conformal Mappings 27

6 Further Development of the Theory 29

7 Asymptotic Methods 39

8 Laplace Transform and Applications 49

iii
iv Contents
Chapter 1

Analytic Functions

Uniqueness of the Complex Numbers

In the text we constructed the field C of complex numbers, which contains the reals
and in which every quadratic equation has a solution. It is only natural to ask
whether there are any other such fields. We shall address this question somewhat
informally. A precise explanation would be tantamount to a short course in abstract
algebra. However, the student should nevertheless be able to grasp the important
points.
The answer is that the complex numbers are the smallest field containing R in
which all quadratic equations are solvable and any two such fields are “the same”.
In that sense it is unique. The reason for this is quite simple. Let F be a field
containing R and in which quadratic equations are solvable. Let j be any solution
in F to the equation z 2 + 1 = 0. Consider, in F , all numbers of the form a + jb
for real numbers a and b. This set is, algebraically, “the same” as C , because of
the simple fact that since j 2 = −1, j plays the role of and may be identified with
i. We must also check that a + jb = c + jd implies that a = c and b = d to be sure
that equality in this set coincides with that in C. Indeed, (a − c) + j(b − d) = 0, so
we must prove that e + if = 0 implies that e = 0 and f = 0 (where e = a − c and
f = b − d). If f = 0, then clearly e = 0 as well. But if f 6= 0, then j = −e/f , which
is real. However, no real number satisfies j 2 = −1 because the square of any real
number is nonnegative. Therefore, f must be zero. This proves our claim. We can
rephrase our result by saying that C is the smallest field extension of R in which
all quadratic equations are solvable.1

1 Another question arises at this point. We made R 2 into a field. For what other n can R n be

made into a field? Let us demand at the outset that the algebraic operations agree with those on
R, assuming that R is the x axis. The answer is, only in the case n = 2. A fieldlike structure,
called the quaternions, can be obtained for n = 4, except that the rule zw = wz fails. Such a
structure is called a noncommutative field. The proof of these facts can be found in an advanced
abstract algebra text.

1
2 Chapter 1 Analytic Functions

Solution of a Cubic Equation

Complex numbers are often introduced by appealing to the quadratic formula,
√
−b ± b2 − 4ac
x= ,
2a
which supplies roots for the quadratic polynomial

ax2 + bx + c = 0.
√
The text points out how leads to solutions such as −1 ± −3 to the equation
x2 +x+1, which has no real number solutions. This is done by creating “imaginary”
square roots for negative numbers, and then all such quadratic polynomials will
have roots. This is certainly very elegant and may be aesthetucally pleasing, but
it may be less clear that anything really important has actually occurred. After
all, the only situations in which one needs this extension to complex numbers are
those for which there are no real solutions anyway. Even geometrically the graph
of y = x2 + x + 1 = (x + (1/2))2 + (3/4) is a parabola that never crosses the x-axis.
The situation for cubic, or third degree, equations may be more striking. Every
cubic polynomial
y = x3 + Ax2 + Bx + C
with real coefficients must have at least one real root and perhaps as many as
three. Its graph must go up for large positive x and down for large negative x. By
continuity it must cross the axis at least once somewhere in between. The fact that
complex numbers are deeply involved in an effective way of finding these solutions
may make it clearer that there is something meaningful and important going on
with them.
The solution to the cubic equation

x3 + Ax2 + Bx + C = 0 (1)

was discovered by Scipione del Ferro and Niccolo Tartaglia in the 1500s and pub-
lished by Giordano Cardano in 1545. To see how to get at the solution, consider
how the coefficients of the equation are related to its roots.

(x − α)(x − β)(x − γ) = x3 − (α + β + γ)x2 + (αβ + αγ + βγ)x − αβγ.

That is,

A = −(α + β + γ), B = αβ + αγ + βγ, and C = −αβγ.

If we make the change of variables t = x + A/3, then the corresponding roots are

A A A
t=α+ ,β + , and γ + .
3 3 3
Chapter 1 Analytic Functions 3

Since these add up to 0, there should be no t2 term in the transformed equation.

Indeed, substitution of x = t − A/3 into (1) produces
µ ¶ µ ¶
A2 AB 2A3
t + B−
3
t+ C − + = 0.
3 3 27

Thus we really need to be able to solve only the more special cubic equations of
the form

t3 + pt + q = 0. (2)

For this we need some cleverness.

Letting t = u + v, equation (2) becomes

(u3 + v 3 ) + (3uv + p)(u + v) + q = 0.

We will have a solution if we can pick u and v with

u3 + v 3 = −q and 3uv = −p. (3)

To solve (3), we begin by eliminating v. Since v = −p/3u, we have

p3
u3 − = −q
27u3
or
p3
(u3 )2 + (u3 )q − = 0.
27
Solving this quadratic for u3 gives
r r
q 1 4p3 q q2 p3
u =− ±
3 2
q + =− ± + .
2 2 27 2 4 27
Thus r
q q2 p3
v = −q − u = − ∓
3 3
+ .
2 4 27
We can take
s r s r
3 q q2 p3 3 q q2 p3
u= − + + and v= − − + .
2 4 27 2 4 27
This gives a solution
s r s r
3 q q2 p3 3 q q2 p3
t= − + + + − − + .
2 4 27 2 4 27
This looks innocent enough, but there are a few interesting things lurking under
all those radicals. First, a cubic with real coefficients must have at least one real
4 Chapter 1 Analytic Functions

root. What happens if the quantity under the inner square root signs is negative?
Second, every cubic has three solutions (if we count multiplicity and allow complex
roots), and all three could be real. Where are the other two roots hiding? In the
algebra section of Standard Mathematical Tables published by the Chemical Rubber
Company, the roots are given as
u + v u − v√ u + v u − v√
u+v ; − + −3 ; − − −3.
2 2 2 2
Where do these formulas come from?
Every real number has a real cube root, but, like every nonzero complex number,
it actually has three complex cube roots. They are distributed at 120◦ intervals
around a circle whose radius is the real cube root of the absolute value of the
number. Thus there are three possibilities for each of the cube roots, u, u0 , u00
and v, v 0 , v 00 . These must be combined in appropriate pairs guided by the relation
3uv = −p to get the three roots.
It is instructive to consider an example with three real roots. The equation

0 = (t − 3)(t + 1)(t + 2) = t3 − 7t − 6

has roots −1, −2, and 3. Here we have p = −7 and q = −6. Our equations become
u3 + v 3 = 6 and 3uv = 7. We find
r r √
343 100 10 3
u =3+ 9−
3
=3+ − =3+ i ≈ 3 + 1.9245 i
27 27 9
r r √
343 100 10 3
v3 = 3 − 9 − =3− − =3− i ≈ 3 − 1.9245 i.
27 27 9
We have
¯ 3 ¯2 ¯ 3 ¯ 2
¯u ¯ = ¯v ¯ = 9 + 11 = 343 ≈ 12.7037,
27 27
so
|u| = |v| ≈ 1.5275.
Also
à √ !
−1 10 3
3
arg(u ) = tan ≈ 0.57rad ≈ 32.68o
27
arg(v 3 ) = − arg(u3 )
1
arg(u) = arg(u3 ) ≈ 10.89p ± 120o
3
1
arg(v) = − arg(u3 ) ≈ −10.89p ± 120o .
3
We get one pair from

Re(u) = |u| cos(arg(u)) ≈ 1.5 Im(u) = |u| sin(arg(u)) ≈ 0.28867

Re(v) = |v| cos(arg(v)) ≈ 1.5 Im(v) = |u| sin(arg(v)) ≈ −0.28867.
Chapter 1 Analytic Functions 5

This pair gives the root u + v = 3. These and the other two pairs are plotted in
Figure 1.1. Notice that none of these are real. Nevertheless, when combined in
the proper pairs, they produce the three real roots, −1, −1, and 3, of our equa-
tion. This ability of complex numbers to produce the real roots for a polynomial
equation with real coefficients was more convincing to many that there really was
something important going on here than were the purely formal complex solutions
to quadratics which did not have any real roots anyway.

v''
u'
u

–2 –1 1 2 3 x
v' v

u''

Figure 1.1: Solution of a cubic polynomial.

6 Chapter 1 Analytic Functions
Chapter 2

Cauchy’s Theorem

Riemann Sums
The theory of complex contour integrals can be based directly on a definition in
terms of approximation by Riemann sums, as in calculus. If γ is a curve from
a to b in the complex plane and f is a function defined along γ, we can choose
intermediate points a = z0 , z1 , z2 , . . . , zn−1 , zn = b on γ and form the sum

X
n
f (zk )(zk − zk−1 )
k=1

(see Figure 2.1). As in calculus, if these sums approach a limit as the maximum of
the oriented arc length fromR zk−1 to zk tends toward 0, we take that limit to be
the value of the integral γ f (z)dz.
The properties of the integral given in Proposition 2.1.3 follow from this ap-
proach much as the corresponding properties in real-variable calculus. To see that
this leads to the same result as Definition 2.1.1 when γ is a C 1 curve, suppose
that z(t) = u(t) + iv(t) is a continuously differentiable parametrization of γ with
z(tk ) = zk . The mean value theorem guarantees numbers t0k and t00k between tk−1
and tk such that zk − zk−1 = [u0 (t0k ) + iv 0 (t00k )](tk − tk−1R ). Thus, the Riemann sums
P
f (zk )(zk − zk−1 ) correspond to Riemann sums for f (γ(t))γ 0 (t)dt after sorting
out real and imaginary parts.
This approach to the integral allows the use of more general curves and is some-
times useful in writing approximations to the integral. For example, Proposition
2.1.6 may be established by using the triangle inequality: for any approximating
Riemann sum, we have
¯X ¯ X
¯ ¯
¯ f (zk )(zk − zk−1 )¯ ≤ |f (zk )||zk − zk−1 |
X
≤ M |zk − zk−1 |
≤ M l(γ).

7
8 Chapter 2 Cauchy’s Theorem

z3
z2 g
z1

Figure 2.1: A polygonal approximation of γ.

The last step uses the fact that |zk − zk−1 | is the length of the line segment from
zk−1 to zk , which is no greater than the distance between them along γ. Since the
estimate holds for each approximating sum, it must hold for the integral, which is
their limit.

A More General Definition of the Integral

The topics here are separated from the body of the section because they are es-
sential to neither an understanding of Cauchy’s Theorem nor of the material in
subsequent chapters. We supply the material promised in the text’s discussion of
the Deformation Theorem. The Smooth Deformation Theorem is used to show how
the integral of an analytic function may be defined along a curve which is continu-
ous but not necessarily piecewise C 1 . This definition and the Smooth Deformation
Theorem itself are used to finish the proof of the Deformation Theorem. Following
this, we explore, without proof, the relationship of Cauchy’s Theorem to a geomet-
ric result known as the Jordan Curve Theorem, which discusses what we mean by
the inside and outside of a simple continuous closed curve.

Integrals along Continuous Curves In the proof of the Deformation Theo-

rem, that is, the homotopy version of Cauchy’s Theorem, we made the provisional
assumption that the deformation was smooth in the sense that each intermediate
curve γs (t) = H(s, t) and each cross curve λt (s) = H(s, t), thought of as curves
traced out by the point H(s, t) as either s or t, respectively, is held constant, are
piecewise C 1 . It was stated that the condition of being C 1 is not actually neces-
§2.1 Contour Integrals 9

sary. We really need only to assume that H(s, t) is a continuous function of s and
t (which implies that each γs (t) is a continuous curve).
For the time being we will refer to the theorem with the C 1 assumption as the
“Smooth Deformation Theorem.” The main reason for the assumption was that
our whole definition of contour integrals was based on piecewise C 1 curves—after
all, the derivative of the curve appears explicitly in the definition! In general we
do not know what the integral of a function along a curve which is continuous but
not piecewise C 1 really is. In fact, such a general theory is not within our grasp.
However, the situation is saved by the fact that we are interested not in general
functions but in analytic functions. This extra assumption about the function to
be integrated makes up for the weaker information about the curve along which it
is to be integrated.
The approach taken here to overcome this difficulty may not be the most direct
route to the Deformation Theorem, but it has the advantage of showing how we
can make sense of the integral of an analytic function along a continuous curve. It
also has the interesting feature of using the Smooth Deformation Theorem in the
process of showing that the smoothness assumption is not really needed.1
Suppose f is an analytic function on an open set G and that γ : [0, 1] → G is a
continuous (but not necessarily piecewise
R C 1 ) curve from z0 to z1 in G. We want
to find a reasonable way to define γ f . The outline of the program is this:
R
(i) We know what λ f means if λ is a piecewise C 1 curve in G from z0 to z1 .
(ii) We show that there is at least one such λ that is “close to” γ by using the
Path-covering Lemma 1.4.24.
(iii) We show that if λ0 and λ1 are two such curves that are “close to” γ, then
they are “close
R to” each
R other, and we use the Smooth Deformation Theorem
to show that λ0 f = λ1 f .
R
(iv) Because of (iii), λ f is the same for all the piecewise C 1 curves λ that are
“close to” γ with the same Rendpoints, and we can take that common value as
a reasonable definition for γ .
To carry out this program, we must first define “close to”. To do this, we define a
type of distance between two parametrized curves with the same parameter interval
by moving out along both curves, recording at each parameter value t the distance
between the corresponding points in the curves and then taking the largest of these
distances. This is illustrated in Figure 2.2.

Definition of the “distance” between curves If λ : [0, 1] → C and γ : [0, 1] →

C are parametrized curves in C, let
dist(λ, γ) = max{|λ(t) − γ(t)| such that 0 ≤ t ≤ 1}.
1 Many of the ideas here are presented more completely and a bit differently in the paper by

R. Redheffer, “The homotopy theorems of function theory,” American Mathematical Monthly, 76

(1969), 778–787, and are used there to do several other interesting things.
10 Chapter 2 Cauchy’s Theorem

l(1)
l(t) g(1)

l(0) g(t)

g(0)
x

Figure 2.2: A “distance” between parametrized curves.

Now suppose G is an open set in C and γ : [0, 1] → G is a continuous curve from

z0 to z1 in G. By the Distance Lemma 1.4.21, there is a positive distance ρ between
the compact image of γ and the closed complement of G, that is, |γ(t) − w| ≥ ρ
for w ∈ C\G, so |γ(t) − z| < ρ implies that z is in G. The Path-covering Lemma
1.4.24 provides a covering of the curve γ by a finite number of disks centered at
points γ(tk ) along the curve in such a way that each disk is contained in G and
each contains the centers of the succeeding and preceding disks. The radius of these
disks may be taken to be ρ for purposes of this proof.
We construct a piecewise C 1 curve λ in G by putting λ(tk ) = γ(tk ) for k =
0, 1, 2, . . . , n and then connecting these points by straight-line segments. More
precisely, for tk−1 ≤ t ≤ tk , we put

(t − tk−1 )λ(tk ) + (tk − t)λ(tk−1 )

λ(t) = .
tk − tk−1

Since the numbers (t − tk−1 )/(tk − tk−1 ) and (tk − t)/(tk − tk−1 ) are positive and
add up to 1, the point λ(t) traces out the straight line segment from γ(tk−1 ) to
γ(tk ) as t goes from tk−1 to tk , as in Figure 2.3.
The function λ(t) is linear and therefore is a differentiable function of t between
tk−1 and tk , so λ is a piecewise C 1 path from z0 to z1 . Furthermore, for each t,
the points λ(t) and γ(t) both lie in the disk D(γ(tk−1 ; ρ), so the curve λ lies in the
set G and dist(λ, γ) ≤ 2ρ. In fact, since λ(t) is on the line between the centers and
γ(t) is in both disks D(γ(tk−1 ); ρ) and D(γ(tk ); ρ), we have dist(λ, γ) ≤ ρ. Since
all three sides of the triangle shown have length less than ρ, the distance from λ(t)
to γ(t) is also less than ρ. (See Figure 2.4.)
This gives us the existence of at least one piecewise C 1 path that is “close to”
γ. Step (iii) of the program outlined earlier is to show that the integrals along all
such paths are the same. Suppose λ0 and λ1 are piecewise C 1 paths from z0 to
z1 such that dist(λ0 , γ) < ρ and dist(λ1 , γ) < ρ. Then both λR0 and λR1 lie in G.
The Smooth Deformation Theorem can be used to show that λ0 f = λ1 f . The
required homotopy between the two curves can be accomplished by following the
straight line from λ0 (t) to λ1 (t). (See Figure 2.5.) For s and t between 0 and 1,
§2.1 Contour Integrals 11

l
g
G

Figure 2.3: A piecewise smooth, in fact linear, approximation to a continuous curve.

g(t)

l(t)

Figure 2.4: dist(λ, γ) < ρ.

define
H(s, t) = sλ1 (t) + (1 − s)λ0 (t).
The function H(s, t) is a piecewise C 1 function of s and of t. Trouble can occur
only when t = tk , k = 0, 1, 2, . . . n, so we need only check that the image always lies
in G. But
|H(s, t) − γ(t)| = |sλ1 (t) + (1 − s)λ0 (t) − γ(t)|
= |s[λ1 (t) − γ(t)] + (1 − s)[λ0 (t) − γ(t)]|
≤ s|λ1 (t) − γ(t)| + (1 − s)|λ0 (t) − γ(t)|
≤ sρ + (1 − s)ρ = ρ.
Thus H(s, t) ∈ D(γ(t); ρ) R⊂ G, soRthe Smooth Deformation Theorem applies to
λ0 and λ1 and shows that λ0 f = λ1 f . This completes step (iii) of the program
and shows that it makes sense to define the integral of an analytic function along
a continuous curve as follows.

Definition of the integral along continuous curves Suppose that f is an-

alytic on an open set G and that γ : [0, 1] → G is a continuous curve in G. If
12 Chapter 2 Cauchy’s Theorem

l0(t)

H(s, t)

l1(t)

Figure 2.5: Smooth homotopy from λ0 to λ1 .

R R
the distance from γ to the complement of G is ρ, let γ f = λ f , where λ is any
piecewise C 1 curve in G that has the same endpoints as γ and that is “close to” γ
in the sense that dist(λ, γ) < ρ.

The Deformation Theorem With a bit of care, essentially the same idea used
in the proof of step (iii) can be used to obtain the deformation theorem (for both
fixed endpoints and closed curves) from the Smooth Deformation Theorem. If H is
a continuous homotopy from γ0 to γ1 , then for s∗ close to s, γs∗ (t) is close to γs (t),
so γs∗ is “close to”γs . If we choose piecewise C 1 curves λ and µ sufficiently “close
to” γs and γs∗ , respectively, then λ will be “close to” µ, and following along the
short straight-line segment between λ(t) and µ(t) will provide a smooth deformation
from λ to µ. (See Figure 2.6.)

l(t)
G
gs(t)
m(t)
gs*(t)

Figure 2.6: The deformation theorems can be obtained from the Smooth Deforma-
tion Theorem.
R R
The Smooth Deformation Theorem says that λ f = µ f , so the integral along
γs is the same as that along γs∗ . Thus if we shift s from 0 to 1 in steps sufficiently
§2.1 Contour Integrals 13

small that this argument applies at every step, the integral will never change and
the integral along γ0 will be the same as that along γ1 . That this actually can be
done in a finite number of sufficiently small steps follows because H is a continuous
function from the compact square [0, 1] × [0, 1], so its image is a compact subset of
G and lies at a positive distance from the closed complement of G.

The Jordan Curve Theorem

An understanding of the Jordan Curve Theorem is not essential to an understanding
of Cauchy’s Theorem or of the material in subsequent chapters. However, the
Jordan Curve Theorem is closely related to the hypotheses in Cauchy’s Theorem,
and therefore it is briefly considered here. In many practical examples the result
of the Jordan Curve Theorem is geometrically obvious and can usually be proven
directly. The general case of the theorem is quite difficult and is not be proven
here.

Jordan Curve Theorem Let γ : [a, b] → C be a simple closed continuous curve

in C. Then C\γ([a, b]) can be written uniquely as the disjoint union of two regions
I and O such that I is bounded (that is, lies in some large disk). The region I is
called the inside of γ and O is called the outside. Region I is simply connected
and γ is contractible to any point in I ∪ γ([a, b]). The boundary of each of the two
regions is γ([a, b]).

The proof of this theorem uses more advanced mathematics and is beyond the
scope of this book.2
Thus the Jordan Curve Theorem, combined with Cauchy’s Theorem, yields the
following: If f is analytic on
R a region A and γ is a simple closed curve in A and the
inside of γ lies in A, then γ f = 0. This is one classical way of stating Cauchy’s
Theorem. Although convenient in practice, it is theoretically awkward for two
reasons: (1) It depends on the Jordan Curve Theorem for defining the concept
of “inside”; (2) γ is restricted to being a simple curve. The versions of Cauchy’s
Theorem stated in §2.3 do not depend on the difficult Jordan Curve Theorem, are
more general, and are just as easy to apply. On the other hand, the Jordan Curve
Theorem reassures us that regions we intuitively expect to be simply connected
indeed are. (There is another way to describe the inside of a simple closed curve
using the index, or winding number, of a curve; this method is discussed in the
next section.)
By applying the Jordan Curve Theorem, one can prove that a region is simply
connected iff, for every simple closed curve γ in A, the inside of γ also lies in A.
This conclusion should seem reasonable. We can also apply the theorem to prove
that the inside of a simple closed curve is simply connected.
2 See, for example, G. T. Whyburn, Topological Analysis (Princeton, N.J.: Princeton University

Press, 1964).
14 Chapter 2 Cauchy’s Theorem

The general philosophy of this text is that we should use our geometric intuition
to justify that a given region is simply connected or that two curves are homotopic—
but with the realization that such knowledge is based on intuition and that to
attempt to make it precise could be tedious. On the other hand, a precise argument
should be used whenever possible and practical (see, for instance, the argument in
the text that a convex region is simply connected).
Chapter 3

Series Representation of
Analytic Functions

The following two results illustrate how the Cauchy integral formula can sometimes
be used to obtain uniformity where it might not be expected. We begin with some
useful terminology.

Definition 3.1 A family of functions S defined on a set G is said to be uniformly

bounded on closed disks in G if for each closed disk B ⊂ G there is a number
M (B) such that |f (z)| ≤ M (B) for all z in B and for all f in S.

The word “uniformly” refers to the fact that the constant M (B) does not depend
on the particular function used from the family, but may depend on the family S
itself and on the disk B chosen.

Theorem 3.2 If f1 , f2 , f3 , . . . is a sequence of functions analytic on a region G

that is uniformly bounded on closed disks in G, then the sequence of derivatives
f10 , f20 , f30 , . . . is also uniformly bounded on closed disks in G.

Proof Suppose B = {z such that |z − z0 | ≤ r} is a closed disk in G. Since B is

closed and G is open, Worked Example 1.4.27 shows that there is a number ρ with
B ⊂ D(z0 ; ρ) ⊂ G. Let R = (r + ρ)/2 and D = {z such that |z − z0 | ≤ R}. By
hypothesis, there is a number N (D) such that |fn (z)| ≤ N (D) for all n and all z
in D. If Γ is the boundary circle of D, the Cauchy integral formula for derivatives
gives, for any z in B,
¯ Z ¯ · ¸
¯ 1 fn (ζ) ¯ 1 N (D)
0 ¯
|fn (z)| = ¯ ¯
dζ ≤ 2πR.
2πi Γ (ζ − z)2 ¯ 2π (R − r)2

Thus, if we put M (B) = N (D)R/(R − r)2 , we will have |fn0 (z)| ≤ M (B) for all n
and for all z in B, as desired. ¥

15
16 Chapter 3 Series Representation of Analytic Functions

Definition 3.3 A family of functions S defined on a set B is said to be uniformly

equicontinuous on B if for each ² > 0 there is a number δ > 0 such that |f (ζ) −
f (ξ)| < ² for all f in S whenever ζ and ξ are in B and |ζ − ξ| < δ. That is, for each
², the same δ can be made to work for all functions in the family S and everywhere
in the set B.

The following shows how one can use Cauchy’s Theorem to verify that a given
family is uniformly equicontinuous. Why one would want to do so will become
clear in the supplementary material for Chapter 6.
Theorem 3.4 Prove that if f1 , f2 , f3 , . . . is a sequence of functions analytic on
a region G that is uniformly bounded on closed disks in G, then this family of
functions is uniformly equicontinuous on every closed disk in G.

Proof Let B be a closed disk in G. By the last example, there is a number M (B)
such that |fn0 (z)| ≤ M (B) for every n and for all z in B. Let γ be the straight line
from ζ to ξ in B. Since that straight line is contained in B, we have
Z Z
|fn (ζ) − fn (ξ)| = | fn0 (z)dz| ≤ |fn0 (z)||dz| ≤ M (B)|ζ − ξ|.
γ γ

Thus, given ² > 0, we can satisfy the definition of uniform equicontinuity on B by

setting δ = ²/M (B). ¥
Some applications of these results are given in the supplementary material for
Chapter 6.

Power Series via Hadamard’s Formula

In this section, the basic facts about convergence of a power series are proved
by directly involving a formula due to Hadamard for the radius of convergence.
A sketch is given showing how these ideas can be applied to operator theory, in
particular to the spectral radius of a matrix or continuous linear operator. This
can be used to link the material here with other analysis courses the student may
be taking.
The basic facts about power series can be developed in a way that also supplies a
formula for the radius of convergence with the help of a few facts from intermediate
analysis. The notion needed is that of the limit superior of a sequence of real
numbers. If c1 , c2 , c3 , . . . is a sequence in R, then the largest cluster point (possibly
infinite) of the sequence is called the limit superior , and the smallest is called the
limit inferior . More precisely,

lim sup ck = lim [sup{ck+1 , ck+2 , ck+3 , . . . }]

k→∞ k→∞

lim inf ck = lim [inf{ck+1 , ck+2 , ck+3 , . . . }] .

k→∞ k→∞

These could be infinite. The facts we need are

§3.1 Convergent Series of Analytic Functions 17

(i) If B < lim supk→∞ ck , then ck > B for infinitely many values of k.

(ii) If B > lim supk→∞ ck , then there is an index N such that ck < B whenever
k ≥ N.

These facts are proved in, for example, J. Marsden and M. Hoffman Elementary
Classical Analysis, Second Edition (New York: W. H. Freeman and Company,
1993).
To apply this concept, we use the following:

(iii) If a sequence f0 , f1 , f2 , f3 , . . . of functions with values in a complete space

such as R or C is uniformly Cauchy on a domain A in the sense that for each
² there is an index N (²) such that

|fn+p (z) − fn (z)| < ²

for every z in A whenever n ≥ N (²) and p > 0, then there is a function f to

which the sequence converges uniformly on A.

(iv) If a series whose terms are functions on a domain A with values in a complete
space such as R or C is such that the series of absolute values converges
uniformly on A, then the series itself converges uniformly on A.

Fact (iii) is used to obtain (iv) by taking the partial sums of the series as the fn .
They are then used to obtain the Weierstrass M test.

Proposition 3.5 Suppose g0 , g1 , g2 , g3 , . . . are functions on a domain A with

P∞ space such as R or C. If there is a sequence of positive constants
values in a complete
Mk such that k=0 PM k converges and |gk (z)| ≤ Mk for every z in A and for every
∞
k, then the series k=0 gk (z) converges uniformly and absolutely on A.

P
Proof Let ² > 0. The partial sums of the seriesP Mk form a uniformly Cauchy
n+p
sequence in R, so there is an index N (²) such that k=n+1 Mk < ² whenever n ≥ N
and p > 0. For such n and for z in A, we have both
¯n+p ¯ ¯ n+p ¯
¯X Xn ¯ ¯ X ¯ X
n+p X
n+p
¯ ¯ ¯ ¯
¯ gk (z) − gk (z)¯ = ¯ gk (z)¯ ≤ |gk (z)| ≤ Mk < ²
¯ ¯ ¯ ¯
k=0 k=0 k=n+1 k=n+1 k=n+1

and
X
n+p X
n X
n+p X
n+p
|gk (z)| − |gk (z)| = |gk (z)| ≤ Mk < ².
k=0 k=0 k=n+1 k=n+1

The series of absolute values and the series itself are uniformly Cauchy and hence
uniformly convergent on the domain A. ¥
We are now ready to obtain the fundamental theorem about power series.
18 Chapter 3 Series Representation of Analytic Functions

P∞
Theorem 3.6 Suppose k=0 ak (z − z0 )k is a power series in a complez variable z
1/k
with complex coefficients ak . Let L = lim supk→∞ |ak | and define R by


0 if L = +∞
R = 1/L if 0 < L < +∞ .


+∞ if L = 0

Then the series converges absolutely if |z − z0 | < R and diverges if |z − z0 | > R.

Furthermore,

(i) If R = 0, the series converges only for z = z0 .

(ii) If R = +∞, the convergence is absolute and uniform on each closed disk
Dr = {z ∈ C | |z − z0 | ≤ r} with 0 < r < ∞.

(iii) If 0 < R < +∞, the convergence is absolute and uniform on each closed disk
Dr = {z ∈ C | |z − z0 | ≤ r} with 0 < r < R.

Proof If z = z0 , the only non-zero term is a0 , and the series certainly converges.
Consider divergence first. If |z − z0 | > R, we can select a nonzero number ρ
with |z − z0 | > ρ > R such that

1 1 1/k
< < lim sup |ak | .
|z − z0 | ρ k→∞

Thus
1 1 1/k
< < |ak |
|z − z0 | ρ
k
for infinitely many values of k. Taking kth powers and multiplying by |z − z0 |
gives
¯ ¯
1 < ¯ak (z − z0 )k ¯
for those values of k. The terms cannot converge to 0 and the series must diverge.
This settles the divergence and case (i).
For the general convergence claim and (ii) and (iii), it suffices to show uniform
absolute convergence on each disk Dr with 0 < r < R since any z with |z − z0 | < R
is contained in such a disk. If |z − z0 | ≤ r < R, we can select a finite nonzero
number ρ with |z − z0 | ≤ r < ρ < R. Then

1 1 1/k
> > lim sup |ak | .
r ρ k→∞

There is an index N such that

1 1 1/k
> > |ak |
r ρ
§3.1 Convergent Series of Analytic Functions 19

whenever k ≥ N . Multiplying by r and taking kth powers gives

r 1/k 1/k
1> > |ak | r ≥ |ak | |z − z0 |
ρ
µ ¶k
r ¯ ¯
≥ |ak | |z − z0 | = ¯ak (z − z0 )k ¯ .
k
1>
ρ
P∞ P∞
Sine r/ρ < 1, the series k=N (r/ρ)k converges. Thus, k=N ak (z − z0 )k converges
uniformly and absolutely on Dr by the Weierstrass M test. Adding in the finitely
many terms for k = 0, 1, 2, 3, . . . does not change this conclusion. ¥.
The formula
1
R= 1/k
lim supk→∞ |ak |
P∞
for the radius of convergence of the power series k=0 ak (z − z0 )k is usually called
Hadamard’s Formula after the French mathematician Jacques Hadamard who
lived from 1865 to 1963. For any particular series, direct appeal to the ratio test
or the root test may well be a more practical way of actually finding the radius of
convergence, but this formula is often valuable, particularly as a theoretical tool.

Application in Operator Theory: The Spectral Radius The preceding

argument does not really demand that the coefficients ak be numbers. What is
required is something corresponding to the absolute value. This can be the norm
or length of a vector much as the absolute value of a complex number is the same
as its length as a vector in the plane. We also want the terms of the series to
be in a complete space, so the statements about the Cauchy property implying
convergence apply. A setting in which this is particularly useful is that of the space
of square matrices or of continuous linear operators on a normed vector space. In
this setting one wants to know as much as possible about invertibility of the matrices
or operators. If T is a matrix or operator, I is the identity matrix or operator, and
µ is a number, one studies the invertibilty of µI − T . The set of numbers for which
this is not invertible is called the spectrum of T , and for square matrices it is the
same as the set of eigenvalues. It is tempting to try to use a geometric series. Will
this work?
µ ¶−1 µ ¶
−1 1 1 1 1 1 2 1 3
(µI − T ) = I− T = I + T + 2T + 3T + ... .
µ µ µ µ µ µ
To study this, one can define the operator norm of a matrix or operator T as the
maximum amount it can stretch a unit vector.

k T k = sup{k T u k | u is a unit vector}

This turns out to have the usual properties of a norm or length for a vector, namely,
1. k T k ≥ 0
2. k T k = 0 if and only if T = 0
20 Chapter 3 Series Representation of Analytic Functions

3. k αT k = |α| k T k for every number α

4. k S + T k ≤ k S k + k T k
in additional to a particularly useful property relating it to products;

5. k ST k ≤ k S k · k T k

It is not too hard to show that if the series converges with respect to this norm,
then it gives the desired inverse. Furthermore, properties 3. and 5. show that
° ° µ ¶k
° 1 k° ° °
° T ° ≤ 1 ° T k ° ≤ 1 k T kk = 1 k T k .
° µk ° k k |µ|
|µ| |µ|

Thus, if |µ| > k T k, the series of norms converges by comparison to a geometric

series of numbers. Once we know that the space of matrices or of operators is
complete, we can use the fact that absolute convergence implies convergence in
any complete space to conclude that our series converges. However, if we use
Hadamard’s Formula or the techniques that lead to it, we get a much sharper
result. The series converges if
³° °1/k ´
|µ| > lim sup ° T k ° .

As a consequence we have the following.

Proposition 3.7 If µ is an eigenvalue of a square matrix T (or is in the spectrum

of a continuous linear operator T ), then
³° °1/k ´
|µ| ≤ lim sup ° T k ° .

Further use of ideas from complex analysis such as the Laurent series expansion
and Liouville’s theorem show that this estimate is precise. The number
³° °1/k ´
ρ(T ) = lim sup ° T k °

is called the spectral radius of T and is equal to the largest absolute value of
points in the spectrum.
Chapter 4

Calculus of Residues

Technical Lemma In the text, the following technical lemma was of interest in
the evaluation of definite integrals along the whole real line. We provide a proof
here.
Lemma 4.1 If
Z B
lim f (x) dx
A→∞,B→∞ −A

exists, then
Z ∞
f (x) dx
−∞

exists and is equal to this limit.

Proof To say that the limit exists and is L is to say that for each ² > 0 there is
an R(²) such that
¯ Z B ¯
¯ ¯
¯ ¯
¯L − f (x) dx¯ < ²
¯ −A ¯

whenever A ≥ R(²) and B ≥ R(²). The assertion is that this implies the indepen-
R0 RB
dent existence of the two limits limA→∞ −A f (x) dx and limB→∞ 0 f (x) dx. We
show how to do the second of these. The first is similar. Notice that if α and β are
both larger than R(²), then
¯Z ¯ ¯ÃZ ! ÃZ !¯
¯ β ¯ ¯ β α ¯
¯ ¯ ¯ ¯
¯ f (x) dx¯ = ¯ f (x) dx − L − f (x) dx − L ¯
¯ α ¯ ¯ −R(²) −R(²) ¯
¯ Z ¯ ¯ Z ¯
¯ β ¯ ¯ α ¯
¯ ¯ ¯ ¯
≤ ¯L − f (x) dx¯ + ¯L − f (x) dx¯ < 2².
¯ −R(²) ¯ ¯ −R(²) ¯

21
22 Chapter 4 Calculus of Residues

We use this observation twice. First suppose b1 , b2 , b3 , . . . is any sequence tend-

Rb
ing to +∞. The observation shows that the integrals 0 k f (x) dx form a Cauchy
sequence and must converge to some limit. Our desired conclusion follows if the
value of that limit is independent of the particular sequence used. Suppose
Z bk
b1 , b2 , b3 , · · · → +∞ and f (x) dx → Lb
0
Z βk
β1 , β2 , β3 , · · · → +∞ and f (x) dx → Lβ .
0

If k is large enough such that bk and βk are both larger than R(²) and each of the
integrals is within ² of its respective limit, then
¯ Z bk ¯ ¯Z ¯
¯ ¯ ¯ βk ¯
¯ ¯ ¯ ¯
|Lb − Lβ | ≤ ¯Lb − f (x) dx¯ + ¯ f (x) dx − Lβ ¯
¯ 0 ¯ ¯ 0 ¯
¯Z Z ¯
¯ bk βk ¯
¯ ¯
+¯ f (x) dx − f (x) dx¯
¯ 0 0 ¯
¯Z ¯
¯ bk ¯
¯ ¯
≤²+²+¯ f (x) dx¯ < 4².
¯ βk ¯

Since this is true for every positive ², we must have Lb = Lβ as required. ¥

Fresnel Integrals Next we treat some special integrals that can be evaluated
using the methods of contour integrals. These types of integrals are useful in optics.

Example 4.2 (Fresnel Integrals) Show that

Z ∞ Z ∞
cos(x2 )dx and sin(x2 )dx
−∞ −∞
p
both exist and equal π/2.

Solution
√ First we show the integrals exist. Observe that sin(x2 ) has zeros at
xn = πn for integers n. Since
√ √ √ √
n + 1 − n = 1/( n + 1 + n),

the distance between these zeros shrinks to zero as n increases, so the quantities
¯Z ¯
¯ xn ¯
¯ 2 ¯
an = ¯ sin(x )dx¯
¯ xn−1 ¯
Chapter 4 Calculus of Residues 23

P∞
decrease monotonically to 0. Thus, 0 (−1)n an converges by the alternating series
test to some number A. If R is any real number, then xN −1 ≤ R < xN for a unique
RR
N , and 0 sin(x2 )dx is between the partial sums

X
N −1 X
N
(−1)n an and (−1)n an .
0 0
RR RR
Thus, limR→∞ 0 sin(x2 )dx exists and is equal to A. Similarly, limR→∞ 0 cos(x2 )dx
exists. √
2
Consider the integral of f (z) = eiz / sin( πz) around the contour

γ = I + II + III + IV

shown in Figure 4.1.

Figure 4.1: Contour used for evaluating the Fresnel integrals.

√
The function f has a simple pole at 0 inside γ with residue 1/ π, so
Z
√
f = 2 πi.
γ

Along I, z = x − Ri, so
2 2
−2Rix−R2 )
|eiz | = |ei(x | = e2Rx
24 Chapter 4 Calculus of Residues

and
√ 1 √ √ √ √ 1 √
| sin πz| = |ei πx−R π − e−i πx+R π | ≥ (eR π − 1).
2 2

Thus, along I we have

¯Z ¯ Z √π/2 √ √
¯ ¯ 2 1 eR π − e−R π
¯ f¯ ≤ √ e2Rx
dx = √ ,
¯ ¯ eR π − 1 √ R eR π − 1
I − π/2

which goes to 0 as R → ∞. Similarly

Z
f →0 as R → ∞.
III

The contribution from the vertical sides is

Z Z Z R
√ 2 Z −R
√ 2
ei( π/2+iy) ei(− π/2+iy)
f+ f = √ idy + √ idy
II IV −R sin(π/2 + πyi) R sin(−π/2 + πyi)
Z R i(π/4−y2 ) −√πy √ Z R
e (e + e πy ) 2
= √ idy = 2i ei(π/4−y ) dy
−R cos(i πy) −R
Z R
e−iy dy
2
= 2e3πi/4
−R
"Z Z R #
√ R
= 2(−1 + i) cos(x )dx − i
2 2
sin(x )dx .
−R −R

Letting R → ∞, we obtain
·Z Z ¸
√ √ ∞ ∞
2 πi = 2(−1 + i) cos(x )dx − i2 2
sin(x )dx
−∞ −∞

and
· Z Z ¸ ·Z Z ¸
√ ∞ ∞ ∞ ∞
2πi = − 2
cos(x )dx + 2
sin(x )dx + i 2
cos(x )dx + 2
sin(x )dx .
−∞ −∞ −∞ −∞

The real part of this equation shows that√our integrals

p are equal, while the imaginary
part shows that their common value is 2π/2 = π/2.

Example 4.3 Show that

Z ∞ √
e−x dx =
2
π.
−∞
Chapter 4 Calculus of Residues 25

R∞
e−x dx.
2
Figure 4.2: Contour for −∞

Solution1 Let f (z) = e−z and consider the integral of f along the contour γ =
2

I + II + III shown in Figure 4.2.

Notice that
Z Z R
e−x dx
2
f=
I 0

and
Z Z 0 Z R
−ir 2 πi/4
f= e e dr = e 5πi/4
(cos r2 − i sin r2 )dr.
III R 0

Along II, z = Reiθ , so

|f (z)| = |e−R | = e−R

2 2
(cos 2θ+i sin 2θ) cos 2θ
.

But for 0 ≤ θ ≤ π/4, we have cos 2θ ≥ 1 − 4θ/π (see Figure 4.3).

Therefore,

|f (z)| ≤ e−R e4R

2 2
θ/π
,

and thus
¯Z ¯ Z
¯ ¯ π/4
π π
¯ f¯ ≤ e−R e4R
2 2
θ/π
|iReiθ |dθ = Re−R
2 2
(eR − 1) = (1 − e−R ).
2

¯ ¯ 4R 2 4R
II 0

This goes to 0 as R → ∞. Since f is entire,

Z Z Z
0= f+ f+ f.
I II III
1 The method that follows is usually attributed to R. Courant.
26 Chapter 4 Calculus of Residues

Figure 4.3: Proof that cos 2θ ≥ 1 − 4θ/π.

Letting R → ∞, we obtain
Z ∞ ·Z ∞ Z ∞ ¸
1+i
e−x dx − √
2
0= cos(x2 )dx − i sin(x2 )dx ,
0 2 0 0

since we already know from the last example that both of these integrals
√ exist.
√
Both integrands are even, and by the last example, both integrals equal π/2 2.
We are left with
Z ∞ √ √
−x2 1+i π π
e dx = √ (1 − i) √ = .
0 2 2 2 2

Again, the integrand is even, so

Z ∞ Z ∞ √
e−x dx = 2 e−x dx =
2 2
π.
−∞ 0
Chapter 5

Conformal Mappings

The main supplementary material for this chapter is the proof of the Riemann
Mapping Theorem. However, the proof we use requires some tools from Chapter 6,
so it is deferred to the next chapter.

27
28 Chapter 5 Conformal Mappings
Chapter 6

Further Development of the

Theory

Normal Families and the

Riemann Mapping Theorem
The main objective of this supplement is to outline a proof of the Riemann Mapping
Theorem. The material is separated from the main text since it is somewhat more
advanced than the rest of the chapter and is not needed for understanding or using
the theorem in succeeding chapters. However, it does illustrate several powerful
tools and techniques of complex analysis.
Throughout this section, G represents a connected, simply connected open
set properly contained in the complex plane C, and D, the open unit disk D =
D(0; 1) = {z such that |z| < 1}. Given z0 ∈ C, the Riemann Mapping Theo-
rem asserts:

There is a function f that is analytic on G and maps G one-to-one onto

D with f (z0 ) = 0. Furthermore, if it is required that f 0 (z0 ) > 0, then
there is exactly one such function.

The uniqueness has already been established in Chapter 5; that is, there can be no
more than one such function. We still need to show there is at least one. The idea
of the proof is to look at all the analytic functions that map G one-to-one into D
taking z0 to 0 with positive derivative at z0 , find one among them that maximizes
f 0 (z0 ), and show that this function must take G onto D.

Montel’s Theorem on Normal Families The proof of the existence of a func-

tion that maximizes f 0 (z0 ) rests on the material of §3.1 concerning uniform con-
vergence on closed disks. We learned there that if a sequence of analytic functions
on a region converges uniformly on closed disks contained in the region, then the

29
30 Chapter 6 Further Development of the Theory

limit function must be analytic. The existence of such sequences is addressed by

the theorem of Montel on normal families.

Definition 6.1 (Definition of Normal Family) If A is an open subset of C, a

set S of functions analytic on A is called a normal family if every sequence of
functions in S has a subsequence that converges uniformly on closed disks in A.

By the Analytic Convergence Theorem, the limit of such a subsequence must

be analytic on A.

Theorem 6.2 (Montel’s Theorem) If A is an open subset of C and S is a set of

functions analytic on A that is uniformly bounded on closed disks in A, then every
sequence of functions in S has a subsequence that converges uniformly on closed
disks in A. That is, S is a normal family.

Proof The plan of attack is as follows:1

(i) Select a countable set of points C = {z1 , z2 , z3 , . . . } that are scattered densely
throughout A in the sense that A ⊂ cl (C).
(ii) Show that there is a subsequence of the original sequence of functions that
converges at all of these points.
(iii) Show that convergence on this dense set of points is enough to force the
subsequence to converge at all points of A.
(iv) Check that this convergence is uniform on every closed disk in A.
The first step may be accomplished by taking those points whose real and imag-
inary parts are both rational numbers. There are only countably many of these,
so they may be arranged in a sequence, and they are scattered densely in A in the
sense that some of them are arbitrarily close to anything in A.
Let f1 , f2 , f3 , . . . be a sequence of functions in S. The assumption of uniform
boundedness on closed disks is that for each closed disk B ⊂ A, there is a number
M (B) such that |fn (z)| < M (B) for all n and for all z in B. In particular, the
numbers f1 (z1 ), f2 (z1 ), f3 (z1 ), . . . are all smaller than M ({z1 }). Thus there must be
a subsequence of them that converges to a point w1 with |w1 | ≤ M ({z1 }). Relabel
this subsequence as

f1,1 (z1 ), f1,2 (z1 ), f1,3 (z1 ), . . . → w1 .

Evaluating these functions at z2 gives another sequence of numbers,

f1,1 (z2 ), f1,2 (z2 ), f1,3 (z2 ), . . . ,

1 The student who has seen the Arzela-Ascoli theorem (see, for example, J. Marsden and M.

Hoffmann, Elementary Classical Analysis, Second Edition (New York: W. H. Freeman and Com-
pany, 1993)) can give a quick proof of Montel’s theorem by using the assumed uniform boundedness
and Worked Example 3.1.19 of this book to prove equicontinuity.
Chapter 6 Further Development of the Theory 31

which are bounded by M ({z2 }). Some subsequence of these must converge to a
point w2 . Relabel this subsubsequence as
f2,1 (z2 ), f2,2 (z2 ), f2,3 (z2 ), . . . → w2 .
It is important to notice that the functions f2,1 , f2,2 , f2,3 , . . . are selected from
among f1,1 , f1,2 , f1,3 , . . . . Continuing in this way, selecting subsequences of subse-
quences, produces an array,
f1,1 (z1 ), f1,2 (z1 ), f1,3 (z1 ), . . . → w1
f2,1 (z2 ), f2,2 (z2 ), f2,3 (z2 ), . . . → w2
f3,1 (z3 ), f3,2 (z3 ), f3,3 (z3 ), . . . → w3
f4,1 (z4 ), f4,2 (z4 ), f4,3 (z4 ), . . . → w4
.. .. .. .. ..
. . . . .
in which the kth horizontal row converges to some complex number wk and the
functions used in each row are selected from among those in the row above. The
proof uses a procedure, called the diagonal construction, which is sometimes
useful in other contexts. Let gn = fn,n . Then g1 , g2 , g3 , . . . is a subsequence of
the original sequence of functions, and liml→∞ gl (zk ) = wk for each k. This is
because gn = fn,n is a subsequence of fk,1 , fk,2 , fk,3 , . . . as soon as n > k. Thus the
subsequence gn converges at a set of points that are scattered densely throughout
A. Steps (iii) and (iv) of the program are to show that the fact that the gn ’s
are uniformly bounded on closed disks in A is enough to force them to converge
everywhere in G and in fact to do so uniformly on closed disks in A. We accomplish
this by showing that the sequence satisfies the Cauchy condition uniformly on closed
disks.
Let B be a closed disk contained in A, and let ² > 0. By the supplementary
results for Chapter 3 (see Theorem 3.4 in this Supplement), the functions gn are
uniformly equicontinuous on B; that is, there is a number δ > 0 such that |gl (ζ) −
gl (ξ)| < ²/3 for all l whenever ζ and ξ are in B and |ζ − ξ| < δ. By using only
finitely many of the points zk we can guarantee that everything in B is within a
distance δ of at least one of them. That is, there is an integer K(B) such that
for each z ∈ B there is at least one k ∈ {1, 2, 3, . . . , K(B)} with |z − zk | < δ and
hence |gl (z) − gl (zk )| < ²/3 for all l. One way to do this would be to take a square
grid of points with rational coordinates and separation less than δ (see Figure 6.1).
Since liml→∞ gl (zk ) = wk for each k, each of these sequences satisfies the Cauchy
condition, and as there are only finitely many of them, there is an integer N (B) such
that |gn (zk ) − gm (zk )| < ²/3 whenever n ≥ N (B), m ≥ N (B), and 1 ≤ k ≤ K(B).
Putting all this together, suppose n ≥ N (B) and m ≥ N (B). If z ∈ A, then z
is within δ of zk for some k ≤ K(B), so
|gn (z) − gm (z)| ≤
|gn (z) − gn (zk )| + |gn (zk ) − gm (zk )| + |gm (zk ) − gm (z)|
² ² ²
≤ + + = ².
3 3 3
The sequence gn thus uniformly satisfies the Cauchy condition on B, so converges
uniformly on B to some limit function, as desired. ¥
32 Chapter 6 Further Development of the Theory

Figure 6.1: Finitely many of the zk ’s give one within δ of anything in B.

Proof of the Riemann Mapping Theorem We are now in a position to prove

the Riemann Mapping Theorem. Let G be a connected, simply connected, open
set properly contained in the complex plane C. Let z0 ∈ G, and let D = D(0; 1)
be the open unit disk. We must show that there is a function f analytic on G that
maps G one-to-one onto D with f (z0 ) = 0 and f 0 (z0 ) > 0. To do this, let

S = {f : G → D | f is analytic and one-to-one on G, f (z0 ) = 0, and f 0 (z0 ) > 0}.

The main steps of the proof are:

(i) Show that S is not empty.

(ii) Show that the numbers {f 0 (z0 ) | f ∈ S} are bounded above, so have a finite
least upper bound M .

(iii) Use Montel’s theorem to extract from a sequence of functions in S whose

derivatives at z0 converge to M a subsequence that converges uniformly on
closed disks in G. The limit function f is analytic in G and f 0 (z0 ) = M .

(iv) Show that f ∈ S.

(v) Show that f must map G onto D.

To show that S is not empty, it is enough to show that we can map G analytically
into the unit disk. Once that is done, we need only compose with a linear fractional
transformation of the disk onto itself, which takes z0 to 0, and then multiply by a
constant eiθ , chosen so that the derivative of the resulting map at z0 is positive. If
G is bounded, for example, if |z − z0 | < R for all z in G, the map z 7→ (z − z0 )/R
does the job. If G is not bounded, it at least omits a point a. The translation
Chapter 6 Further Development of the Theory 33

z 7→ z − a takes G to a simply connected region G1 not containing 0. By Theorem

2.2.6, there is a branch of logarithm defined on G1 , which we will call F . Then
the map g defined by z 7→ e(1/2)F (z) is a branch of the square root function; by
the open mapping or inverse mapping theorem, one sees that G2 = g(G1 ) contains
some disk D(b; r). By properties of the square root function, D(−b; r) fails to meet
G2 . The map f (z) = r/[b + z] then maps G2 into the unit disk. See Figure 6.2.

Figure 6.2: Mapping G into the unit disk.

Having shown that S is not empty, we must establish step (ii). The family S is
uniformly bounded by 1 on G, so by the supplementary Theorem 3.2, the derivatives
are uniformly bounded on closed disks in G. In particular, there is a finite number
M ({z0 }) such that f 0 (z0 ) ≤ M ({z0 }) for all f in S. Let M be the least upper
bound of these derivatives. There must be a sequence f1 , f2 , f3 , . . . of functions
in S with the property that limn→∞ fn0 (z0 ) = M . Since the family S is uniformly
bounded, it is normal by Montel’s Theorem and there must be a subsequence that
converges uniformly on closed disks in G. We may as well throw away the functions
we don’t need and assume that we have a sequence that converges uniformly on
closed disks in G. By the Analytic Convergence Theorem 3.1.8 they converge to a
limit function f , which is analytic on G and f 0 (z0 ) = M .
We next want to know that f is a member of S. Each of the functions fn maps
G into the open unit disk, so f certainly maps G into the closed unit disk. Since
f is not constant, the Maximum Modulus Principle says that |f (z)| cannot have a
maximum anywhere in G, so the image never touches the boundary of the disk and
34 Chapter 6 Further Development of the Theory

f maps G into D. Certainly f (z0 ) = limn→∞ fn (z0 ) = 0. Finally, the corollary of

Hurwitz’ Theorem 6.2.8 shows that f must be one-to-one since it is a nonconstant
limit of one-to-one functions that converge uniformly on closed disks. Thus, f ∈ S.
The final step, (iv), is to show that f must actually map G onto D. This follows
from the following assertion.

Claim If A is a connected and simply connected open set properly contained in

D and 0 ∈ A, then there is a function F analytic on A that maps A one-to-one
into D with F (0) = 0 and F 0 (0) > 1.
To see how (iv) follows from this assertion, suppose that f does not map G onto
D. Then A = f (G) satisfies the conditions of the claim. (That A is open follows
from the Open Mapping Theorem 6.3.3. Consider g(z) = F (f (z)). Then g ∈ S,
but g 0 (z0 ) = F 0 (f (z0 ))f 0 (z0 ) = F 0 (0)M > M , contradicting the maximality of M .
Thus it remains to check the claim. The construction is a bit like that used in
step (i) and is seen by following the diagrams in Figure 6.3.
The region A is shaded by diagonal lines in the first diagram. It misses a point
a indicated by an open circle in the diagram. The successive images of a and 0
are indicated by open dots and solid dots, respectively, in each of the following
diagrams. Map F1 is a linear fractional transformation of the disk to itself taking a
to 0 and 0 somewhere. The purpose of map F2 is to guarantee a situation in which
the image of A misses a neighborhood of a point on the boundary circle. This is
done just as in step (i) by using a branch of logarithm on the simply connected
region F1 (A) that misses 0. Map F3 is another linear fractional transformation that
returns the image of 0 to 0. At this stage the image of A misses a small circle γ
that intersects the unit circle C at right angles at two points. An appropriate linear
fractional transformation F4 taking these points to 0 and ∞ will take the circles to
lines through 0 and ∞ and the region between them to a quarter plane. Squaring
F5 opens this up to a half plane. Finally another linear fractional transformation
takes the half plane to the unit disk with the black dot going to 0 and the correct
rotation making the derivative of the whole thing at 0 positive. The function F is

F (z) = F6 (F5 (F4 (F3 (F2 (F1 (z)))))) = w.

The inverse function g(w) = F −1 (w) = z satisfies the conditions of the Schwarz
Lemma. Since it is not a rotation, we have strict inequality |g 0 (0)| < 1 by the
Schwarz Lemma, but F 0 (0) = 1/g 0 (0). Therefore, F 0 (0) > 1, as required. All the
pieces have been assembled, so the proof of the Riemann Mapping Theorem is now
complete. ¥
 Chapter 6 Further Development of the Theory 35

,

,  
,
,
, 
, ,










,























,






Figure 6.3: Construction for the claim in the proof of the Riemann Mapping The-
orem.

Dynamics of Complex Analytic Mappings

The pictures shown in Figure 6.4 are representations of the dynamics of complex
analytic mappings. The purpose of this section is to provide a brief introduction to
this subject—mainly to inspire the reader to find out more by consulting a reference
on the subject.2
The subject we will be looking at has to do with the way points in the complex
plane behave under iteration of an analytic function. It has its origins in classical

2 Such as R. L. Devany, An Introduction to Chaotic Dynamical Systems (Reading, Mass.:

Addison-Wesley, 1985); P. Blanchard, Complex dynamics on the Riemann sphere, Bulletin of the
American Mathematical Society, 11 (1984), 85–141; or B. Mandelbrot, The Fractal Geometry of
Nature (New York: W. H. Freeman and Company, 1982).
36 Chapter 6 Further Development of the Theory

(a) (b)

Figure 6.4: The different shadings represent the rate of approach of points to infinity
under iteration of the mapping; the black region (assuming that you are viewing
the figure in color) consists of “stable” points that remain bounded under iteration.
In part (a) the mapping is (1+0.1i) sin z, while in (b) it is (1+0.2i) sin z. (Courtesy
of R. Devany of Boston University, with the assistance of C. Mayberry, C. Small,
and S. Smith)

and beautiful work of G. Julia3 and P. Fatou.4 In this study normal families play
an important role. In fact, Montel himself was interested in these questions.5
Let us fix an entire function f : C → C. We need a little terminology to get
going. Given a point z ∈ C, the orbit of z is the sequence of points

z, f (z), f (f (z)), f (f (f (z))), . . . ,

which we also write as z, f (z), f 2 (z), f 3 (z), . . . . We think of the point z as moving
successively under the mapping f to new locations. A fixed point is a point z such
that f (z) = z, that is, a point z that does not move when we apply f . A periodic
point is a point z such that f n (z) = z for some integer n (called the period), where
f n means f composed with itself n times.
A fixed point z is called an attracting fixed point if |f 0 (z)| < 1. The reason
for this terminology is that the orbits of nearby points converge to z; this is so
because near z, f behaves like a mapping that rotates by an amount arg f 0 (z) and
magnifies by an amount |f 0 (z)|, so every time f is applied, points will be pulled
toward z by a factor |f 0 (z)|, so as this is repeated, the point tends to z. Likewise,
3 Memoiré sur l’itérations des fonctions rationelles, J. Math., 8 (1918), 47–245.
4 Sur
l’itérations des fonctions transcendantes entière, Acta Math., 47 (1926), 337–370.
5 See his Leçons sur les familles normales de fonctions analytiques et leurs applications (1927;

reprinted New York: Chelsea, 1974), Chapter VIII.

Chapter 6 Further Development of the Theory 37

a point z is called a repelling fixed point if |f 0 (z)| > 1; points near repelling
points will be pushed away under iteration of the function f . Similarly, a periodic
point z with period n is called an attracting periodic point if |(f n )0 (z)| < 1;
such points have the property that the orbits of points close to z tend to the orbit
of z. Likewise, a repelling periodic point has the property that |(f n )0 (z)| > 1;
orbits of points near such points will be shoved away from the orbit of z.
The Julia set J(f ) of f is defined to be the closure of the set of repelling
periodic points of f . This set can have remarkable and beautiful complexity usually
called a fractal ; in fact, in the picture in Figure 6.4 the nonblack region is the Julia
set. This statement rests on a theorem, which we shall not prove, stating that the
Julia set is the closure of the points that go to infinity under iteration of f . It is
this characterization that is useful for computational purposes. Figure 6.5 shows
two more Julia sets for quadratic maps.

Figure 6.5: (a) Julia set of f (z) = z 2 + 12 i, which is a simple closed curve but is
nowhere differentiable. (b) Julia set of f (z) = z 2 − 1, which contains infinitely
many closed curves.

As far as complex analysis is concerned, one of the most important results is

the following:
The Julia set of f is the set of points at which the family of functions
f n is not normal.
This result can be used as an alternative definition for the Julia set, and in fact
this was the original definition of Fatou and Julia. We will prove only the following
statement here to give a flavor of how the arguments go:
If f is a repelling fixed point of f (and therefore is in the Julia set),
then the family of iterates f n fails to be normal at z.
Let us assume that this family is normal at z and derive a contradiction. Normal at
z means normal in a neighborhood of z, in the same way we used the terminology
“analytic at z.” By Definition 6.1, the family f n has a subsequence that converges
38 Chapter 6 Further Development of the Theory

uniformly on a neighborhood of z. Since f (z) = z and |f 0 (z)| > 1, it follows from

the chain rule that
0
|f n (z)| = |f 0 (z)|n → ∞,

that is, that the sequence of derivatives of f n evaluated at z must tend to infinity
as n → ∞. However, the sequence of derivatives must converge to the derivative
of the limit function by the Analytic Convergence Theorem 3.1.8, which is finite,
giving us the required contradiction.
This discussion represents only the tip of a large collection of very interesting
and beautiful results. We hope the reader will be inspired to look up some of
the references on the subject we have given as well as further references found in
those sources and will explore the subject further. We hasten to point out that the
iteration of complex mappings is just one part of a larger and growing field called
chaotic dynamics. For the more general aspects, the reader can consult Devany’s
book cited in footnote 2 or the book Nonlinear Oscillations, Dynamical Systems,
and Bifurcation of Vector Fields, by J. Guckenheimer and P. Holmes (New York:
Springer-Verlag, 1983).
Chapter 7

Asymptotic Methods

Proof of the Steepest Descent Theorem

The goal of this section is to provide the proof of the Steepest Descent Theorem
that we omitted in the textbook. We first recall the statement of the theorem.

Theorem 7.1 (Steepest Descent Theorem) Let γ : ] − ∞, ∞[ → C be a C 1

curve. (γ may also be defined only on a finite interval.) Let ζ0 = γ(t0 ) be a point
on γ and let h(ζ) be a function continuous along γ and analytic at ζ0 . Make the
following hypotheses: For |z| ≥ R and arg z fixed,

(i) The integral

Z
f (z) = ezh(ζ) dζ
γ

converges absolutely.

(ii) h0 (ζ0 ) = 0; h00 (ζ0 ) 6= 0.

(iii) Im[zh(ζ)] is constant for ζ on γ in some neighborhood of ζ0 .

(iv) Re[zh(ζ)] has a strict maximum at ζ0 along the entire curve γ.

Then
√
ezh(ζ0 ) 2π
f (z) ∼ √ p
z −h00 (ζ0 )

as z → ∞, arg z fixed. The sign of the square root is chosen such that
√ p 00
z −h (ζ0 ) · γ 0 (t0 ) > 0.

39
40 Chapter 7 Asymptotic Methods

Proof We begin by breaking up the curve γ into three portions, γ1 , C, and γ2 , as

illustrated in Figure 7.1. Choose C such that it lies in a neighborhood of ζ0 small
enough that h(ζ) is analytic and that condition (iii) holds. Clearly,
f (z) = I1 (z) + I2 (z) + J(z),
where we use the notations
Z Z
J(z) = ezh(ζ) dζ and Ik (z) = ezh(ζ) dζ, k = 1, 2.
C γk

Figure 7.1: Method of steepest descent.

We will show that for large z the part of the integral that really matters is J(z)
so that an asymptotic approximation for J(z) will also give one for f (z). Worked
Example 7.2.12 says that to do this, it is enough to show that
µ ¶
Ik (z) 1
=O
J(z) zn
for all positive n. To prove this, note that
¯Z ¯ Z
¯ ¯
¯
|Ik (z)| = ¯ ezh(ζ)
dζ ¯¯ ≤ eRe zh(ζ) |dζ|.
γk γk

However,
Z Z
J(z) = zh(ζ)
e dζ = eRe zh(ζ) ei Im zh(ζ) dζ.
C C

Since Im[zh(ζ)] is constant on C, we get

¯Z ¯
¯ ¯
¯
|J(z)| = ¯ eRe zh(ζ)
dζ ¯¯ .
C
0
If C is short enough that arg(γ ) changes by less than π/4 along C, then we obtain
√ Z Re zh(ζ)
|J(z)| > (1/ 2) e |dζ|.
C

Thus,
¯ ¯ R
¯ Ik (z) ¯ eRe zh(ζ) |dζ| √
¯ ¯ ≤ Rγk 2.
¯ J(z) ¯ eRe zh(ζ) |dζ|
C
§7.1 Infinite Products 41

Let C̃ be a strictly smaller subinterval of C, centered at ζ0 . Then

¯ ¯ R
¯ Ik (z) ¯ eRe zh(ζ) |dζ| √
¯ ¯ ≤ Rγk 2.
¯ J(z) ¯ eRe zh(ζ) |dζ|
C̃

Fix z0 and let α be the minimum of Re[(zh(ζ)] on C̃. There is an ² > 0 such that
Re[(z0 h(ζ)] ≤ α − ² for all ζ ∈ γk . Thus, using the fact that z lies on the same ray
as z0 ,
R R
eRe zh(ζ) |dζ| eRe z0 h(ζ) eRe(z−z0 )h(ζ) |dζ|
R γk
= Rγk
eRe zh(ζ) |dζ| eRe z0 h(ζ) eRe(z−z0 )h(ζ) |dζ|
C̃
³C̃
R ´
γk
e Re z0 h(ζ)
|dζ| e|z−z0 |(α−²)/|z0 |
≤ ¡R ¢ .
C̃
eRe z0 h(ζ) |dζ| e|z−z0 |α/|z0 |

This expression is a constant factor, say M , times e−|z−z0 |²/|z0 | . The latter is
certainly O(1/z) (and in fact is O(1/z n ) for all n ≥ 1), so we have proved that
Ik (z)/J(z) = O(1/z n ) for all n ≥ 1. This localizes the problem to a neighborhood
around ζ0 where the bulk of the contribution to the integral is made. Also, we can
shrink the length of C without affecting the conclusion that f (z) ∼ J(z) as z → ∞.
Next, we write

h(ζ) = h(ζ0 ) − w(ζ)2 ,

where w(ζ) is analytic and invertible (abusing notation, we denote the inverse by
ζ(w)), where w(ζ0 ) = 0, and where

−h00 (ζ0 )
[w0 (ζ0 )]2 =
2
(see Worked Example 6.3.7). Since

Im(zh(ζ)) = Im[zh(ζ0 )]

on C and

Re[zh(ζ)] < Re[zh(ζ0 )],

we see that z[w(ζ)]2 √

is real and greater than zero on C; also, by our choice of branch
for the square root, zw(ζ) is real and, as a function of the curve parameter t, has
positive derivative at t0 . Thus, by shrinking C if necessary, we can assume that
√
zw(ζ) is increasing along C.
Note that
Z Z Z
ezh(ζ0 ) · e−zw(ζ) dζ = ezh(ζ0 ) e−z[w(ζ)] dζ.
2 2
J(z) = ezh(ζ) dζ =
C C C
42 Chapter 7 Asymptotic Methods

√
We change variables by setting zw(ζ) = y, and we get
Z √ |z|²2 Z √|z|²2
−y 2 dζ dy ezh(ζ0 ) dζ
e−y
2
J(z) = e zh(ζ0 )
√ e √ = √ dy, √
− |z|²1 dw z z
− |z|²1 dw
p p
since y is real on C; we choose positive numbers ²1 and ²2 such that [− |z|²1 , |z|²2 ]
is in the range of y corresponding to ζ on C. Next we write
ζ = ζ0 + a1 w + a2 w2 + . . . ,
so
dζ
= a1 + 2a2 w + 3a3 w2 + . . . ,
dw
√
where w = y/ z. Thus,
√ Z √|z|²2 " ∞
X µ ¶k−1 #
J(z) z y
e−y
2
= √ kak √ dy
ezh(ζ0 ) − |z|²1 k=1
z
Z √|z|²2 "
X
N µ ¶k #
y
e−y
2
= √ (k + 1)ak+1 √ dy
− |z|²1 z
k=0
Z √|z|²2 õ ¶N +1 !
−y 2 y
+ √ e O √ dy
− |z|²1 z
XN Z √|z|²2
(k + 1)ak+1
e−y y k dy
2
= √ k √
( z) − |z|²1
k=0
√
Z |z|²2 õ ¶N +1 !
−y 2 y
+ √ e O √ dy.
− |z|²1 z

By Exercise 7,
Z ∞ √
(2m)! π
e−y y k dy =
2

−∞ m!22m
if k = 2m is even and it is zero if k = 2m + 1 is odd, so we are led to the series
X (k + 1)ak+1 Z ∞ X∞ √
(2m)! π (2m + 1)a2m+1
e−y y k dy =
2
S≡ √ k .
( z) −∞ m=0
m!22m zm

This gives
ÃZ √ Z ∞ !
J(z) X
2M
(k + 1)ak+1 |z|²1
−y 2 k −y 2 k
√ − SM = − √ e y dy + √ e y dy
ezh(ζ0 ) / z ( z)k −∞ |z|²2
k=0
Z √|z|²2 õ ¶2M +1 !
−y 2 y
+ √ e O √ dy.
− |z|²1 z
§7.1 Infinite Products 43

√
The first two integrals are o(1/( z)2M ) by Proposition 7.2.3(v), since e−y y k =
2

o(1/y 2M +1 ). In the third, there is a constant BM such that the integrand is bounded
by
µ ¶Ã !
BM e−y |y|2M +1
2
BM 1
e−y |y|2M +1 .
2
√ 2M +1 = p
| z| |z|M
|z|

Since
Z ∞
e−y |y|2M +1 dy < ∞,
2

−∞

this term is also o(1/|z|M ). Thus,

ezh(ζ0 ) S
J(z) ∼ √ ,
z
and by Worked Example 7.2.12, the same is true of f (z). Thus,
r µ ¶
π 1 · 3a3 1 · 3 · 5a5
f (z) ∼ ezh(ζ0 )
a1 + + + ... .
z z z2
To complete the proof, note that
√
dζ 1 2
a1 = (0) = =p ,
dw dw 00
−h (ζ0 )
dζ (ζ0 )

so
√
ezh(ζ0 ) 2π
f (z) ∼ √ p ,
z −h00 (ζ0 )
as desired. ¥

Bounded Variation and the Stationary Phase Formula

We saw in the method of stationary phase that we needed to impose a condition
on the amplitude that limits the amount of high-frequency oscillation. This type
of condition is often needed in theory involving integrals; the notion of bounded
variation provides the appropriate tools. We will then use it to prove the Stationary
Phase Formula.

Definition 7.2 Suppose f : [a, b] → R.

(i) If P is a partition of [a, b] given by a = t0 < t1 < . . . < tn = b, then the
variation of f on [a, b] relative to P is defined to be
X
n
VP f = |f (tk ) − f (tk−1 )|.
k=1
44 Chapter 7 Asymptotic Methods

(ii) The total variation of f on [a, b] is V[a,b] f = sup{VP f }, where the least
upper bound is taken over all possible partitions. (It might be +∞.)
(iii) If V[a,b] f < ∞ we say that f is of bounded variation and write
f ∈ BV ([a, b]).

Some important examples of such functions are included in the following.

Proposition 7.3
(i) If f is monotone and bounded on [a, b], then f ∈ BV ([a, b]) and V[a,b] f =
|f (b) − f (a)|.
(ii) If f is differentiable on a bounded interval [a, b] and |f 0 (x)| < M for all
x ∈ [a, b], then f ∈ BV ([a, b]) and V[a,b] f ≤ |b − a|M .
(iii) If f has a continuous derivative on the bounded interval [a, b]—that is, if
f ∈ C 1 ([a, b])—then f ∈ BV ([a, b]).

Proof The first result holds since the succeeding differences from point to point
along any partition are all of the same sign and values at intermediate points cancel
out. The second is shown by applying the mean value theorem to each subinterval of
any partition, and the third follows from it since if f 0 is continuous on the compact
interval [a, b], then it is bounded. ¥

It is possible for a continuous function not to have bounded variation. On [−1, 1]

set f (0) = 0 and f (x) = x cos(1/x) for x 6= 0. (See Figure 7.2.) Then we have
|f (1/nπ) − f (1/(n + 1)π)| = (2n + 1)/n(n + 1)π > 1/nπ.
Since the harmonic series diverges, partitions may be created using these points
that give arbitrarily large variation.

Figure 7.2: The continuous function x cos(1/x) has unbounded variation.

Some of the important properties of functions of bounded variation are outlined

in the following proposition.
§7.1 Infinite Products 45

Proposition 7.4 Suppose f ∈ BV ([a, b]).

(i) If [c, d] ⊂ [a, b], then f ∈ BV ([c, d]) and V[c,d] f ≤ V[a,b] f .
(ii) V[a,c] f + V[c,b] f = V[a,b] f if a < c < b.
(iii) (V f )(x) = V[a,x] f is a bounded increasing function on [a, b] with (V f )(a) = 0
and (V f )(b) = V[a,b] f .
(iv) If a ≤ x ≤ y ≤ b, then (V f )(y) − (V f )(x) = V[x,y] f .
(v) f is the difference of two bounded increasing functions: f = f1 − f2 with
f1 = (V f + f )/2 and f2 = (V f − f )/2.

Proof The first assertion follows since any partition of [c, d] can be extended by
the intervals [a, c] and [d, b] to obtain a partition of [a, b] offering a larger candidate
for V[a,b] f . For the second, adjoin partitions of [a, c] and [c, b] to get a partition of
[a, b] and show V[a,c] f + V[c,b] f ≤ V[a,b] f . For the opposite inequality let a = t0 <
t1 < . . . < tn = b be any partition of [a, b] with
X
n
|f (tk ) − f (tk−1 )| > V[a,b] f − ².
k=1

Pick N with tN ≤ c ≤ tN +1 . Then

X
N X
N +1
V[a,b] f < |f (tk ) − f (tk−1 )| + |f (tk ) − f (tk−1 )| + ²
k=0 k=N
X
N
< |f (tk ) − f (tk−1 )| + |f (c) − f (tN )| + |f (tN +1 ) − f (c)|
k=0
X
n
+ |f (tk ) − f (tk−1 )| + ²
k=N +2
≤ V[a,c] f + V[c,b] f + ².

Since this holds for any ² ≥ 0, we have the desired inequality. The third assertion
is clear and the fourth follows from it and the second. For the last assertion, use
(iv) to show that the functions indicated are increasing. ¥

The last property will be the one directly utilized in the proof of Theorem 7.2.10.
The tool by which we will use it is the second mean value theorem for integrals.

Theorem 7.5 (Second Mean Value Theorem for Integrals) If f is bounded

and increasing on [a, b] and g is integrable, then there is a point c in [a, b] such that
Z b Z c Z b
f (t)g(t)dt = f (a) g(t)dt + f (b) g(t)dt.
a a c
46 Chapter 7 Asymptotic Methods

Proof Let
Z x Z b
F (x) = f (a) g(t)dt + f (b) g(t)dt.
a x

Then
Z b Z b
F (a) = f (b) g(t)dt and F (b) = f (a) g(t)dt
a a

and
Z b Z b Z b
f (a) g(t)dt ≤ f (t)g(t)dt ≤ f (b) g(t)dt.
a a a

Since F is continuous on [a, b], the conclusion follows from the intermediate value
theorem. ¥

We will also need the following estimates.

Lemma 7.6 If f is differentiable on [a, b] and |f 0 (x)| ≤ M for all x in [a, b],
then |f (y) − f (x)|, |(V f )(y) − (V f )(x)|, |f1 (y) − f1 (x)|, and |f2 (y) − f2 (x)| are each
bounded above by M |y − x|.

Proof The first conclusion follows from the mean value theorem and the second
from Proposition 7.3(ii) and Proposition 7.4(iv). The last two follow from the first
two and the formulas for f1 and f2 given in Proposition 7.4(v). ¥

The first step in the intuitive derivation of Theorem 7.2.10 was that contribu-
tions to the integral from parts of the interval away from t0 tended to cancel out
and could be neglected by comparison with the contribution from a short interval
near t0 . The only critical point of h was at t0 and h0 and h00 were continuous,
so away from t0 the derivative stays away from 0 and we can apply the following
lemma.
Lemma 7.7 Suppose h has a continuous second derivative on [a, b], that h0 (x) is
never 0 in [a, b], and that g has a continuous derivative on [a, b]. Then
Z b
eizh(t) g(t)dt = O(1/z).
a

Proof The function ψ(x) = g(x)/h0 (x) has a continuous derivative on [a, b] and
thus has bounded variation and may be written as a difference of two increasing
functions, ψ = ψ1 − ψ2 . Then
Z b Z Z
1 b 0 i b
eizh(t)
g(t)dt = ψ1 (t) cos(zh(t))zh (t)dt + ψ1 (t) sin(zh(t))zh0 (t)dt
a z a z a
Z Z
1 b i b
− ψ2 (t) cos(zh(t))zh0 (t)dt − ψ2 (t) sin(zh(t))zh0 (t)dt.
z a z a
§7.1 Infinite Products 47

Each of these integrals may be estimated using the second mean value theorem for
integrals. There is a point x between a and b with
¯Z ¯ ¯ Z x
¯ b ¯ ¯
¯ 0 ¯ ¯
¯ ψ1 (t) cos(zh(t))zh (t)dt¯ = ¯ψ1 (a) cos(zh(t))zh0 (t)dt
¯ a ¯ a
Z b ¯
¯
0 ¯
+ψ1 (b) cos(zh(t))zh (t)dt¯
x ¯
= |ψ1 (a)[sin(zh(x)) − sin(zh(a))]
+ ψ1 (b)[sin(zh(b)) − sin(zh(x))]|
≤ 2|ψ1 (a)| + 2|ψ1 (b)|.

The others are treated similarly to obtain

¯Z ¯
¯ b ¯ 4
¯ ¯
¯ eizh(t)
g(t)dt¯ ≤ [|ψ1 (a)| + |ψ1 (b)| + |ψ2 (a)| + |ψ2 (b)|],
¯ a ¯ z

as needed. ¥

We now complete the proof of Theorem 7.2.10. Since t0 is the only critical point
of h in [a, b], we know that for any δ > 0, h0 (t) is never 0 on [a, t0 − δ] or [t0 + δ, b],
√ so
by the last lemma, the integrals of eizh(t) g(t) over each are O(1/z), so are o(1/ z).
Thus, to establish Theorem 7.2.10 it is enough to show that
Z √
√ −ih(t0 ) 2π
lim ze eizh(t)
g(t)dt = p e±πi/4 g(t0 ),
z→∞ J ±h00 (t0 )

where J = [t0 − δ, t0 + δ]. We may fix δ as small as we please as long as its choice
does not depend on z. In the course of the proof we shall find conditions for that
choice.
We know h is analytic in a neighborhood of t0 . By Worked Example 6.3.7 there
is an analytic function w(t) such that h(t) = h(t0 ) ± [w(t)]2 for t near t0 and w is
locally one-to-one. We may choose w to be real and strictly increasing on J if δ is
selected small enough. This is our first criterion for δ. We choose the plus sign if
h00 (t0 ) > 0 and the minus sign if h00 (t0 ) < 0. Since w(t0 ) = 0 and w is continuous,
w(t0 + δ) = c and w(t0 − δ) = d, where c < 0 < d. The change of variables x = w(t)
gives
Z Z d
e±izx ψ(x)dx,
2
izh(t) izh(t0 )
e g(t)dt = e
J c

where ψ(x) = g(w−1 (x))/(w−1 )0 (x). The function ψ has a continuous derivative
on [c, d]. The point x = 0 corresponds to t = t0 , and

h00 (t0 ) = ±2w(t0 )w00 (t0 ) ± 2[w0 (t0 )]2 .

48 Chapter 7 Asymptotic Methods
p
Thus, (w−1 )0 (0) = 1/w0 (t0 ) = ±h00 (t0 )/2. Since ψ 0 is continuous, ψ has bounded
variation and can be written as a difference ψ1 − ψ2 of two increasing functions.
Let ² > 0. Since c and d got to 0 as δ → 0, we can use Lemma 7.6 to select δ small
enough so that the quantities |ψ1 (c) − ψ1 (0)|, |ψ1 (d) − ψ1 (0)|, |ψ2 (c) − ψ2 (0)|, and
|ψ2 (d) − ψ2 (0)| are all smaller than ². Thus,
Z Z
√ √ d
ze−izh(t0 ) e±izx ψ(x)dx
2
eizh(t) g(t)dt = z
J c
Z Z
d √ d √
= cos(zx )ψ1 (x) zdx ± i
2
sin(zx2 )ψ1 (x) zdx
c c
Z Z
d √ d √
− cos(zx2 )ψ2 (x) zdx ∓ i sin(zx2 )ψ2 (x) zdx.
c c

As in the proof of Lemma 7.7, each integral may be handled by the second mean
value theorem for integrals and the first is typical. There is a point y between c
and d such that
Z d Z y Z d
√ √ √
cos(zx2 )ψ1 (x) zdx = ψ1 (c) cos(zx2 ) zdx + ψ1 (d) cos(zx2 ) zdx
c c y
Z √
y z Z √
d z
= ψ1 (c) √
cos(u2 )du + ψ1 (d) √
cos(u2 )du.
c z y z

Using the Fresnel integrals from the above supplementary material for Chapter p 4,
these integrals converge
p as z goes to +∞. Since c < 0 < d, the
p limit is ψ 1 (d) π/2
if y < 0, is ψ1 (c) π/2pif y > 0, andpis {[ψ1 (c) + ψ1 (d)]/2} π/2 if y = 0. But
each of these is within ² π/2 of ψ1 (0) π/2. Similar arguments for the other three
integrals show that the whole sum converges to a limit that is
r r r r
π π π π
ψ1 (0) ± iψ1 (0) − ψ2 (0) ∓ iψ2 (0) ,
2 2 2 2
p
with an error of √ no more than ² π/2 in each term. Thus, we do get a limit that is
no more than 2² 2π away from the point
p √ √ p
[ψ1 (0) − ψ2 (0)](1 ± i) π/2 = ψ(0) πe±πi/4 = 2πg(t0 )e±πi/4 / ±h00 (t0 ) ,

just as desired. This completes the proof of Theorem 7.2.10. ¥

Chapter 8

Laplace Transform and

Applications

Fourier Transform and Wave Equation

The Fourier transform, which was introduced in §4.3, provides an alternative to the
Laplace transform for solving differential equations. We illustrate this use and the
role of complex variables by focusing on the wave equation. Our discussion will be
somewhat informal and we shall forgo the rigorous formulation of theorems.

Wave Equation The wave equation is the equation of motion that describes
the development of a wave disturbance propagating in a medium. It describes,
for example, the vertical displacement of a vibrating string (see Figure 8.1), the
propagation of an electromagnetic wave through space and of a sound wave in a
concert hall, and some types of water wave motion.

velocity = c

Figure 8.1: φ is the wave amplitude.

First consider the homogeneous problem, the simplest case of which is a wave
traveling down a string of constant density ρ and under constant tension T . The
vertical displacement φ(x, t) at position x and time t satisfies the wave equation

1 ∂2φ ∂2φ
2
· 2 = ,
c ∂t ∂x2
p
where c = T /ρ is the velocity of propagation, a constant. We accept this fact
from elementary physics. (The derivation assumes that the amplitude is small.)

49
50 Chapter 8 Laplace Transform and Applications

√
Note that if we were to have c = −1 in the wave equation, we would recover
the Laplace equation (see §2.5√ and §5.3). Indeed, just as that equation admitted
solutions of the form f (x ± −1y), the solutions to the wave equation take the
form f (x ± ct). The fact that the wave equation is of second order in the t variable
suggests that a solution is uniquely given when two pieces of initial data at t = 0
are specified. These data consist of φ(x, 0) and dφ/dt at (x, 0); the wave equation
then gives the development of φ(x, t) for subsequent t.
To solve the wave equation, we perform a transform on the x variable to obtain
a simpler equation involving the transform variable k. However, here x runs from
−∞ to +∞, so instead of using the Laplace transform we use the Fourier transform.
Let f : R → C; the Fourier transform fˆ of f is defined by
Z +∞
ˆ
f (k) = e−ikx f (x) dx.
−∞

There is an inversion formula that is analogous to the Laplace inversion formula:

Z +∞
1
f (x) = eikx fˆ(k) dk.
2π −∞

The Fourier transform of the function φ(x, t) is defined by

Z +∞
φ̂(k, t) = e−ikx φ(x, t) dx.
−∞

Here we perform the integral with respect to the x variable, regarding t as a fixed
parameter. The Fourier inversion formula now reads
Z +∞
1
φ(x, t) = eikx φ̂(k, t) dk.
2π −∞

We are now ready to solve the wave equation. Taking the Fourier transform
and differentiating under the integral, we obtain

1 ∂ 2 φ̂
· (k, t) + k 2 φ̂(k, t) = 0.
c2 ∂t2
In other words, our transformation technique has replaced the partial differential
equation for φ(x, t) with an ordinary differential equation for φ̂(k, t), which is easily
solved. The solution is

φ̂(k, t) = A(k)eikct + B(k)e−ikct ,

where A(k), B(k) are two constants of integration that may depend on the param-
eter k. Applying the inversion formula, we get
Z +∞
1
φ(x, t) = [A(k)eik(x+ct) + B(k)eik(x−ct) ] dk.
2π −∞
Laplace Transform and Applications 51

This is our solution to the wave equation. The functions A(k), B(k) are determined
by the initial data φ(x, 0) and ∂φ(x, 0)/∂t.
Note that the first integral in this solution depends only on the variable x + ct,
whereas the second depends only on x − ct; that is, φ(x, t) has the form

φ(x, t) = f (x + ct) + g(x − ct),

where the functions, f, g are again determined by φ and ∂φ/∂t at t = 0. We can

verify by substitution that this formula for φ does indeed give a solution of the
wave equation.
Some special solutions deserve separate attention. These are monochromatic
(single-frequency) waves and are of the form

φ(x, t) = ei(x/c−t)ω ,

where ω is the frequency. This φ represents a wave of frequency ω traveling to the

right down the string. Generally, f (x + ct) is a wave moving to the left whose shape
is that of the graph of f , with velocity c. Similarly, g(x − ct) is a wave moving to
the right.
Next we shall deal with the inhomogeneous problem, which occurs when an
external force is applied to the wave. For example, suppose that the string illus-
trated in Figure 8.1 were given a constant charge density q and then placed in an
external electric field E(x, t) pointing in the y direction. This would result in the
application to the string of a force F (x, t) proportional to qE(x, t). We must then
solve the following equation of motion for the displacement φ(x, t), which is called
the inhomogeneous wave equation:
1 ∂2φ ∂2φ
2
· 2 = + F (x, t).
c ∂t ∂x2
For simplicity, we take F (x, t) to be periodic with frequency ω:

F = f (x, ω)eiωt .

This allows us to consider the simpler problem

1 ∂2φ ∂2φ
· = + f (x, ω)eiωt .
c2 ∂t2 ∂x2
We write the solution as φ(x, t, ω). Once we have solved this simpler problem, we
can deal with the problem of a general force F (x, t). First we “Fourier-analyze” it;
that is, we write F as follows:
Z +∞
1
F (x, t) = f (x, ω)e−iωt dω,
2π −∞
where
Z +∞
f (x, ω) = F (x, t)eitω dt.
−∞
52 Chapter 8 Laplace Transform and Applications

We then superimpose the solutions of the simpler problem to obtain

Z +∞
1
φ(x, t) = φ(x, t, ω) dω.
2π −∞

We solve the simpler inhomogeneous wave equation by taking the Fourier trans-
form with respect to the variable x to get

1 ∂ 2 φ̂
· + k 2 φ̂ = fˆ(k, ω)eiωt ,
c2 ∂t2
where
Z +∞
fˆ(k, ω) = eikx f (x, ω) dx.
−∞

This is a simple inhomogeneous second-order differential equation. Adding the

particular solution

fˆ(k, ω)eiωt
k 2 − (ω/c)2

to the homogeneous solutions obtained earlier, we obtain the general solution

fˆ(k, ω)
φ̂(t, ω) = A(k)eikct + B(k)e−ikct + eiωt .
k2 − (ω/c)2

The solution is thus

φ(x, t, ω) = h(x + ct) + g(x − ct) + (G ∗ f )eiωt .

The terms in this equation are explained as follows. The first two terms are
solutions to the homogeneous wave equation, and again they are to be chosen so
that the initial data at t = 0 are satisfied. The last term, a particular solution to
the inhomogeneous equation, is given by taking the inverse Fourier transform of
the last term in the expression for φ̂:
Z +∞ " #
1 fˆ(k, ω)
eikx 2 dk.
2π −∞ k − (ω/c)2

As with the Laplace transform, this term is the convolution of G and f where
Ĝ = 1/[k 2 − (ω/c)2 ]. This function Ĝ plays a central role in the theory of partial
differential equations. Its transform,
Z +∞ · ¸
1 eikx
G(x, ω) = dk,
2π −∞ k 2 − (ω/c)2
Laplace Transform and Applications 53

is called the Green’s function1 , and we can use contour integration to evaluate
it in closed form as follows.
The integrand of G(x, ω) has simple poles at k = ±(ω/c). In its present form,
the integral is not convergent. To specify its value, we use the Cauchy principal
value. Several possible values may be obtained depending on how we interpret our
integrals. To select the value we want, we shall evaluate the integral by closing the
contour of integration in the upper half of the complex k plane for x > 0 and in the
lower half of the plane for x < 0. This is necessary if the integral over the semicircle
is to approach zero as the radius approaches infinity. By Cauchy’s Theorem, we
pick up the residues of the enclosed poles. We still must specify how we are to go
around the singularities at k = ±ω/c. Different choices will lead to different but
still mathematically acceptable values of G. Our final choice is determined by the
asymptotic behavior we want G to have as x → ∞. The homogeneous solutions to
the wave equation in which we are interested behave like exp(±ikx) as a function
of x, and we will require the same behavior of G. This can be specified by the “i²
prescription”:
Z +∞ · ¸
1 eikx
G(x, ω) = lim dk,
²→0,²>0 2π −∞ k 2 − (ω/c − i²)2

in which we still close the contour (as shown in Figure 8.2) according to the sign of
x.

Figure 8.2: Contours for G(x, ω).

We can now evaluate G(x, ω). From the preceding equation and the Residue
Theorem we obtain, for x > 0,

· ¸
1 eix(i²+ω/c) ic
G(x, ω) = lim 2πi · = − eiωx/c .
²→0,²>0 2π −2(ω/c − i²) 2ω
1 Gramatically, the use of the term “the Green’s function” is incorrect, just as it would be to

say “the Cauchy’s Theorem”, but it is, unfortunately, how it is commonly expressed.
54 Chapter 8 Laplace Transform and Applications

Making a similar computation for x < 0 and using the contour at the right in
Figure 8.2, we obtain

 −ic eiωx/c

x>0
G(x, ω) = 2ω .
 −ic
 e−iωx/c x < 0
2ω

Equivalently,
c iω|x|/c
G(x, ω) = e .
2iω
In textbooks on differential equations, G is often obtained as the solution to

d2 G
+ ω 2 G = δ(x − y),
dx2
where δ is the “Dirac δ function.” The solution is found by the general formula
½
−u(x)v(y)/w x>y
G(x, y, ω) = .
−u(y)v(x)/w x−y

Here u, v are solutions of the corresponding homogeneous equations and w is their

Wronskian, w = uv 0 − vu0 . In this case, we have u = eiωx and v = e−iωx . We
recover the formula for G(x, ω) by setting y = 0.

Scattering Problem When the medium through which the wave propagates
is not homogeneous, we encounter the scattering problem. For example, suppose
that the vibrating string of Figure 8.1 now consist of three pieces smoothly joined
together, with one piece, of length a, having a density of ρ2 (region II in Figure 8.3)
and the other two pieces each having a density of ρ1 (region I, III in Figure 8.3).

Figure 8.3: One-dimensional scattering.

Let c1 and c2 denote the corresponding velocities of propagation. Assume that

ρ2 > ρ1 . Imagine that an incident wave ei(x/c1 −t)ω from the left travels down the
string. As the wave moves onto the denser material at x = 0, part of it will be
reflected backward, while part will be transmitted onward. At x = a, some of the
wave will again be reflected backward while the rest travels forward (see Figure 8.4).
Laplace Transform and Applications 55

Figure 8.4: Reflection in one-dimensional scattering.

We must solve the wave equation in each region:

∂2φ 1 ∂2φ
= · for regions I, III
∂x2 c21 ∂t2
∂2φ 1 ∂2φ
= 2· 2 for region II.
∂x2 c2 ∂t
It is not unreasonable to expect solutions of the following forms:
φI (x, t) = ei(x/c1 −t)ω + Rei(−x/c1 −t)ω
φII (x, t) = Aei(x/c2 −t)ω + Bei(−x/c2 −t)ω
φIII (x, t) = T ei(x/c1 −t)ω .
We shall require that, at x = 0 and x = a, the solutions continuously join each
other and have no sharp bend. Mathematically, this means that we impose the
boundary conditions that φ and ∂φ/∂x be continuous at x = 0, a. In other words,
φI (0, t) = φII (0, t) φII (a, t) = φIII (a, t)
∂φI ∂φIII
(0, t) = ∂φII
∂x (0, t) ∂φII
∂x (a, t) = (a, t).
∂x ∂x
These four equations allow us to solve for the coefficients R, A, B, T . We are par-
ticularly interested in T . After performing some algebraic manipulations we find
that
4c1 c2 eit[(1/c2 )−(1/c1 )]ωa
T =
(c1 + c2 )2 − (c1 − c2 )2 e2iaω/c2
(see Exercise 5). The quantity T is called the scattering amplitude, and the
square of its absolute value represents the intensity of the wave transmitted into
region III.
We now allow ω in this equation to become a complex variable, and we see that
T (ω) has the following property:
(i) T is meromorphic in ω and has poles in the lower half plane at ω = (c2 /a)(nπ−
iρ), where ρ is determined by e2ρ = (c1 + c2 )2 /(c1 − c2 )2 .
(ii) T has absolute value 1 at those values of ω for which e2iω/c2 = 1.
(iii) As ω → +i∞, T → 0.
(iv) As ω → −i∞, T → ∞.
56 Chapter 8 Laplace Transform and Applications

Dispersion Relations When a function of a complex variable f (z) shares the

same four properties as T (ω), the Cauchy theorem can be used to obtain an inter-
esting and useful representation for f (z), as shown in the following.
Theorem 8.1 (Hilbert Transform Theorem) If f (z) is analytic for Im(z) ≥ 0
and f (z) → 0 uniformly as z → ∞ in the half plane 0 < arg z < π, then f (z)
satisfies the following integral relationships.
(i) If z0 = x0 + iy0 with y0 > 0, then
Z
y0 +∞ f (x, 0)
f (z0 ) = dx
π −∞ (x0 − x)2 + y02
and
Z
1 +∞
f (x, 0)(x − x0 )
f (z0 ) = dx.
πi −∞ (x − x0 )2 + y02

(ii) If z0 = x0 is real, then

Z +∞
1 f (x, 0)
f (z0 ) = dx.
πi −∞ x − x0

Proof Because of the assumptions, we can apply Cauchy’s Theorem, using a large
semicircle in the upper half plane, to give
Z +∞
1 f (x, 0)
f (z0 ) = dx
2πi −∞ x − z0
(see §4.3). We also have
Z +∞
1 f (x, 0)
0= dx,
2πi −∞ x − z0
where z 0 lies in the lower half plane. If we subtract the last two equations we obtain
the first equation for f (z0 ); if we add them we get the second. The third follows
from formula 6 of Table 4.2.1. ¥
As a corollary, by taking real and imaginary parts of each side of the first
equation for f (z0 ), we get
Z
y0 +∞ u(x, 0)
u(x0 , y0 ) = dx,
π −∞ (x − x0 )2 + y02
where f = u + iv; a similar equation holds for v(x, y). From the second we have
Z
1 +∞ (x − x0 )v(x, 0)
u(x0 , y0 ) = dx
π −∞ (x − x0 )2 + y02
Z
1 +∞ (x − x0 )u(x, 0)
v(x0 , y0 ) = − dx.
π −∞ (x − x0 )2 + y02
Laplace Transform and Applications 57

Finally, from the third, we obtain

Z +∞
1 v(x, 0)
u(x0 , 0) = P. V. dx
π −∞ x − x0
Z +∞
−1 u(x, 0)
v(x0 , 0) = P. V. dx.
π −∞ x − x0

The first formula for u(x0 , y0 ) gives us the values of a harmonic function in the upper
half plane, in terms of its boundary values on the real axis, and thus provides a
solution to the Laplace equation in the upper half plane.
Note that if f (z) satisfies the symmetry property f (−x) = f (x), then we can
write
Z ∞
2 Im f (x)
Re f (x0 ) = P. V. x dx,
π 0 x2 − x20

whereas if f (z) satisfies f (−x) = −f (x), then

Z ∞
2x0 Im f (x)
Re f (x0 ) = P. V. dx.
π 0 x2 − x20

These equations for u and v can be regarded as integral versions of the Cauchy-
Riemann equations; they simply tell us, for example, the values that the real part
of an analytic function must take when the imaginary part is specified. When func-
tions u, v satisfy these equations, we say that u and v are “Hilbert transforms” of
each other. Historically, the Hilbert transforms were the forerunners of a series of
such relations called dispersion relations. They were first observed to hold for the
complex dielectric constant as a function of incident frequency by H. A. Kramers
and R. de L. Kronig in 1924. Since approximately 1950, they have been system-
atically studied and applied to the scattering amplitude T (ω) and to quite general
classes of scattering problems for which this amplitude is defined. The extension of
these relations to three-dimensional scattering problems will be considered later in
this supplement.
The relations derived assumed that f (z) is analytic only for Im z ≥ 0. However,
there is a second class of dispersion relations for functions that are analytic in the
z plane except for a branch line along the real axis.

Proposition 8.2 If f (z) is analytic in the z plane with a branch line from z = a
to ∞, and if |f (z)| = O(1/z), then
Z ∞
1 1
f (z) = lim [f (x + i²) − f (x − i²)] dx.
²→0+ 2πi a x−z

(The notation O(1/z) is explained in §7.2 and lim²→0+ means the limit is taken
through ² > 0.)
58 Chapter 8 Laplace Transform and Applications

Figure 8.5: Contour used in the proof of the preceding proposition.

Proof Take the contour of Figure 8.5 and apply the Cauchy Theorem to the
function f (ζ)/(ζ − z). ¥
If, in addition to the hypotheses of the preceding proposition, f (z) also satisfies
the relation f (z) = f (z), that is, if u(x, ²) + iv(x, ²) = u(x, −²) − iv(x, −²), so
the real part of f is continuous across the real axis whereas the imaginary part is
discontinuous, then we obtain
Z
1 ∞ 1
f (z) = lim Im f (x + i²) dx.
²→0+ π a x−z
When z actually moves onto the real axis, we can take real parts of each side of the
preceding equation to obtain
Z
P.V. +∞ 1
Re f (x0 ) = lim Im f (x + i²) dx.
²→0+ π a x − x0

Wave Equation in Three Dimensions The ideas that have been developed
thus far in this section for wave motion in one dimension can easily be extended
to higher-dimensional problems. In two dimensions the vibrating string is replaced
by a vibrating membrane. In three dimensions we can think of sound waves prop-
agating in air. The pressure φ(r, t) then satisfies the equation of motion,
1 ∂2φ
· (r, t) = ∇2 φ(r, t) + F (r, t),
c2 ∂t2
where F represents some external source of waves, r = (x, y, z), and
∂2 ∂2 ∂2
∇2 = + +
∂x2 ∂y 2 ∂z 2
is the Laplace operator. When the equation of motion is expressed in terms of rect-
angular coordinates, its homogeneous and inhomogeneous solutions are obtained
Laplace Transform and Applications 59

in much the same way they were previously. The really new and exciting features
not present in one dimension arise in the scattering problem, and these are most
interesting and tractable when the scattering medium has spherical symmetry.
Consider an incident plane wave ei(x/c−ω)t traveling from the left down the x
axis that impinges on a ball located at the origin (Figure 8.6). Part of the wave
may penetrate the ball, part of the wave is scattered by the surface of the ball and
then travels radially outward, and the remainder of the wave simply bypasses the
ball. To solve for φ, we proceed as previously. We first obtain the solutions for φ in
region I and region II separately and then require that φ and the radial derivative
∂φ/∂r be continuous at the surface of the ball. This procedure eventually specifies
the total wave in region I that results from the “impurity” of the medium in region
II.

Figure 8.6: Scattering of a wave by a spherical obstacle.

These calculations are not detailed here, because such a task would take us too
far afield into the subject of partial differential equations. However, the form of the
final result is not too difficult to anticipate. The wave in region I will be a sum of
incident wave and the outgoing radial wave, and this will take the asymptotic form
· ¸
eir/c
φI (r, t) ∼ eix/c
+ f (ω, θ) e−iωt
r

as |r| = r → ∞. Here f (ω, θ) is the amount of scattering wave that is trav-

eling outward at an angle θ to the axis of symmetry (see Figure 8.6); it is the
three-dimensional analogue of the scattering amplitude T (ω) of the one-dimensional
problem discussed earlier in the section.
Note that f is now a function of two complex variables, ω and θ. Physically
observable scattering occurs, of course, only when ω and θ are real with 0 ≤ θ ≤ π.
But by studying the properties of f for complex values of ω and θ, as we will do in
60 Chapter 8 Laplace Transform and Applications

the following paragraphs, we do gain a deeper understanding of the characteristics

of f .
It will be convenient to change variables as follows:

4ω 2 ω2
s= t = −2 (1 − cos θ).
c2 c2
Define the funtion A of the two complex variables s, t by A(s, t) = f (ω, θ). For
a large class of scattering problems, it can be shown that A has the following
properties:

(i) A(s, t) is analytic in the two complex variables s and t with branch lines from
s = a to ∞ and t = b to ∞.

(ii) A(s, t) = O(1/s) as s → ∞ for each t.

(iii) A(s, t) = A(s, t) for t real.

Applying the preceding proposition,

 Z ∞
 1 1
 lim²→0+ [A(s + i², t0 ) − A(s − i², t0 )] ds s0 , t0 in C
A(s0 , t0 ) = Z ∞a s − s0
2πi .

 lim²→0+ 1 1
Im A(s + i², t0 ) ds t0 real
π b s − s0

In the first equation for A(s0 , t0 ), we are integrating A(ζ, t) for ζ slightly above
and below the real axis. Now consider the integrand. By property (i) of A(s, t) we
can write a dispersion relation in the t variable as follows:
Z ∞
1 1
A(s + i², t0 ) = lim [A(s + i², t + iδ) − A(s + i², t − iδ)] dt.
δ→0+ 2πi b t − t0

Using a similar representation for the second term in the first equation for A(s0 , t0 ),
we finally obtain the double dispersion relation
Z ∞ ·Z ∞ ¸
1 1 1
A(s0 , t0 ) = 2 ρ(s, t) dt ds,
π a s − s0 b t − t0

where
µ ¶2
1
ρ(s, t) = lim [A(s + i², t + iδ) − A(s + i², t − iδ)
²,δ→0+ 2i
−A(s − i², t + iδ) + A(s − i², t − iδ)].

The equation for A(s0 , t0 ) is a representation first obtained by S. Mandelstam in

1958.
Laplace Transform and Applications 61

Exercises for Supplement to Chapter 8

1. Apply the boundary conditions
φI (0, t) = φII (0, t) φII (a, t) = φIII (a, t)
∂φI ∂φIII
(0, t) = ∂φ
∂x (0, t)
II ∂φII
∂x (a, t) = (a, t)
∂x ∂x
to obtain expressions for T (ω), R(ω), A(ω), B(ω). Verify that when ω is real,
|T (ω)|2 + |R(ω)|2 = 1.
(This relation expresses “conservation of intensity”: In a medium without
dissipative loss, the unit intensity of the incident wave of Figure 8.4 is equal
to the sum of the intensity of the wave R reflected back into region I and
intensity of the wave T transmitted into region III.)
2. Since
Z +∞
y0 f (x, 0)
f (z0 ) = dx
π −∞ (x0 − x)2 + y02
solves the Dirichlet problem for the upper half plane, apply it to the problem
illustrated at the right in Figure 5.3.5 of the text and obtain the same solu-
tion as in Example 5.3.1. Thus, we have two methods for solving Laplace’s
equation: conformal mapping and integral relations of the type described by
the preceding displayed equation.
3. As an application of the Hilbert Transform, let
1
f (z) = √ = r−1/2 e−iθ/2 for 0 < θ < 2π
z
be defined on the z plane with a branch line along the positive real axis. What
equality results?
4. As yet another way to solve the Laplace equation, take Fourier transforms of
∂2u ∂2u
2
(x, y) + 2 (x, y) = 0 (8.1)
∂x ∂y
with respect to the variable x (use the method described in the Wave Equation
subsection); show that this method gives the ordinary differential equation
∂ 2 û
(k, y) − k 2 û(k, y) = 0. (8.2)
∂y 2
By solving, summing over all solutions, and keeping only those that decay
exponentially in the limit y → +∞, obtain the formula
Z +∞
1
u(x, y) = A(k)eikx−k/y dk. (8.3)
2π −∞
62 Chapter 8 Laplace Transform and Applications

With relative ease, A(k) can now be evaluated in terms of the boundary data
at y = 0. Since at y = 0, equation (8.2) reduces to
Z +∞
1
u(x, 0) = A(k)eikx dk,
2π −∞

A(k) must in fact be the Fourier transform of u(x, 0). Using this result,
substituting in equation (8.2), and interchanging orders of integration, obtain
Z +∞ µ Z +∞ ¶
1 ik(x−z)−|k|/y
u(x, y) = u(z, 0) e dk dz.
−∞ 2π −∞
Perform the integral on k to get
Z +∞
1 1
u(x, y) = u(z, 0) · · dz. (8.4)
−∞ π (x − z)2 + y 2

Referring to equation (8.4), show that

¯
y −∂ ¯
= G(x, y|z, w)¯ ,
(x − z) + y
2 2 ∂w ¯
w=0

where

G(x, y|z, w) = log |r − r0 |

with r = x + iy, r0 = z + iw. (The function G is the Green’s function and

equals the potential at r caused by a unit charge at the point r0 in the plane.)
5. Many problems in applied mathematics include an infinite series
∞
X
F (z) = fn (z)
n=0

in which the function F (z) has singularities in addition to all those of each
fn (z). These singularities are introduced by the failure of the series to con-
verge. The simplest example of this type of series is
1
= 1 + z + z2 + . . . .
1−z
The individual terms on the right of the equation are entire functions; the
function their sum defines has a simple pole at z = 1. As a second example,
consider
z 1 1
= 1 + + 2 + ... .
z−1 z z
The terms on the right of this equation have singularities at z = 0; their sum
is singular at z = 1.
Laplace Transform and Applications 63

Consider
∞
X
F (z) = [g log(z − α)]n ,
n=0

where α, g are constants.

(a) What are the analytic properties of the individual terms on the right
side of the last equation?
(b) By summing the series in closed form, verify that the sum in this equation
has a pole at z = α + e1/g .

Comments on Selected Exercises

1. Substituting the expressions for φI , φII , φIII into the boundary conditions that
follow, we have

1+R=A+B (8.5)

c1
1−R= (A − B) (8.6)
c2

Aeiaω/c2 + Be−iaω/c2 = T eiaω/c1 (8.7)

c2 iaω/c1
Aeiaω/c2 − Be−iaω/c2 = Te . (8.8)
c1
Solve Equations (8.7) and (8.8) to get
µ ¶
c1 + c2
A=T eiaω(1/c1 −1/c2 ) (8.9)
2c1

µ ¶
c1 − c2
B=T eiaω(1/c1 +1/c2 ) . (8.10)
2c1

Adding equations (8.5) and (8.6) gives

µ ¶ µ ¶
c1 + c2 c2 − c1
2=A +B . (8.11)
2 c2

Now substitute (8.9), and (8.10) into (8.11) to obtain the expression for T in the
text. If we subtract (8.5) and (8.6), we have
µ ¶ µ ¶
c2 − c1 c2 + c1
2R = A +B . (8.12)
c2 c2
64 Chapter 8 Laplace Transform and Applications

Substituting (8.9) and (8.10) into (8.12) gives R in terms of T ,

µ ¶ µ ¶
T aω
R= (c22 − c21 )eiaω/c1 sin .
2ic1 c2 c2
We can now obtain the desired result. We have
· µ ¶¸
(c21 − c22 )2 aω
|T | + |R| = |T | 1 +
2 2 2
sin2
. (8.13)
4c21 c22 c2
The denominator of |T |2 is
(c1 + c2 )2 + (c1 − c2 )4 − (c21 − c22 )2 (e2iaω/c2 + e−2iaω/c2 )
µ µ ¶¶
aω
= (c1 + c2 )4 + (c1 − c2 )4 − (c21 − c22 )2 2 · 1 − 2 sin2
c2
µ ¶
aω
= [(c1 + c2 )2 − (c1 − c2 )2 ]2 + 4(c21 − c22 )2 sin2
c2
µ µ ¶¶
aω
= 4 4c21 c22 + (c21 − c22 )2 sin2 .
c2
Thus,
4c21 c22
|T |2 = ³ ´.
4c21 c22 + (c21 − c22 )2 sin2 aω
c2

This, with Equation (8.13), gives the desired result.

3. By Proposition 8.3.5,
Z ∞
1 dx
f (z0 ) = lim [f (x + i²) − f (x − i²)],
²→0 2πi 0 x − z0
where
1 e−iθ/2
f (z) = √ = √ , 0 < θ < 2π.
z r
Now
1
f (x + i²) → √ as ² → 0+
x
and
1
f (x − i²) → − √ as ² → 0+,
x
so the identity which results is
Z ∞ Z ∞
1 1 dx 1 i dx 1
√ = f (x0 + iy0 ) = √ =− √ .
z0 πi 0 x − z0 x π 0 x − z0 x
Laplace Transform and Applications 65

This can be verified directly by changing variables using w2 = x.

Z Z Z
i ∞ dx 1 2i ∞ dw i ∞ dw
− √ =− =− .
π 0 x − z0 x π 0 w2 − z0 π −∞ w2 − z0
√
This has poles at ± z0 . If z0 is not positive real, one is in the upper half plane
√
and one in the lower with residues ±1/(2 z0 ). The last expression becomes
Z µ ¶ µ ¶
i ∞ dw i 1 1
− = − (2πi) √ =√
π −∞ w − z0
2 π 2 z0 z0

as before.

4. We have

G(x, y|z, w) = log[(x − z)2 + (y − w)2 ]1/2 .

Then
∂G w−y
=
∂w (x − z)2 + (y − w)2
¯
∂G ¯
and hence − ∂w w=0 is the desired expression. Thus,
Z +∞
y0 u(x, 0)
u(x0 , y0 ) = dx,
π −∞ (x − x0 )2 + y02

becomes
Z +∞
1 ∂G
u(x, y) = − u(z, 0) (x, y|z, 0).
−∞ π ∂w

This expression for the solution to Laplace’s equation is a special case of the Green’s
function solution obtained in courses on partial differential equations.

5. Each term of the series has a logarithmic singularity at z = α. Provided

|g log(z − α)| < 1, we can use the geometric series to exactly sum the series as
1
F (z) = .
1 − g log(z − α)

The denominator is 0 at z = α + e1/g , but its derivative, −g/(z − α), is not 0 there.
Thus, there is a pole of order 1 at z = α + e1/g .