Complex Analysis Lecture Notes

Dan Romik
About this document. These notes were created for use as primary reading
material for the graduate course Math 205A: Complex Analysis at UC Davis.

The current 2020 revision (dated June 6, 2020) updates my earlier version
of the notes from 2018. With some exceptions, the exposition follows the
textbook Complex Analysis by E. M. Stein and R. Shakarchi (Princeton Uni-
versity Press, 2003).

The notes are typeset in the Bera Serif font.

Acknowledgements. I am grateful to Christopher Alexander, Jennifer Brown,

Brynn Caddel, Keith Conrad, Bo Long, Anthony Nguyen, Jianping Pan, and
Brad Velasquez for comments that helped me improve the notes. Figure 5
on page 27 was created by Jennifer Brown and is used with her permission.
An anonymous contributor added an index and suggested the Bera Serif font
and a few other improvements to the document design.

You too can help me continue to improve these notes by emailing me at

[email protected] with any comments or corrections you have.

Complex Analysis Lecture Notes

Document version: June 6, 2020

Copyright © 2020 by Dan Romik

Cover figure: a heat map plot of the entire function z 7→ z(z − 1)π −z/2 Γ(z/2)ζ(z).
Created with Mathematica using code by Simon Woods, available at
http://mathematica.stackexchange.com/questions/7275/how-can-i-generate-this-domain-coloring-plot
Contents
1 Introduction: why study complex analysis? 1

2 The fundamental theorem of algebra 3

3 Analyticity 7

4 Power series 13

5 Contour integrals 16

6 Cauchy’s theorem 21

7 Consequences of Cauchy’s theorem 26

8 Zeros, poles, and the residue theorem 35

9 Meromorphic functions and the Riemann sphere 38

10 The argument principle 41

11 Applications of Rouché’s theorem 45

12 Simply-connected regions and Cauchy’s theorem 46

13 The logarithm function 50

14 The Euler gamma function 52

15 The Riemann zeta function 59

16 The prime number theorem 71

17 Introduction to asymptotic analysis 79

Problems 92

Suggested topics for course projects 119

References 121

Index 122
1 1 INTRODUCTION: WHY STUDY COMPLEX ANALYSIS?

1 Introduction: why study complex analysis?

These notes are about complex analysis, the area of mathematics that studies
analytic functions of a complex variable and their properties. While this
may sound a bit specialized, there are (at least) two excellent reasons why
all mathematicians should learn about complex analysis. First, it is, in my
humble opinion, one of the most beautiful areas of mathematics. One way of
putting it that has occurred to me is that complex analysis has a very high
ratio of theorems to definitions (i.e., a very low “entropy”): you get a lot more
as “output” than you put in as “input.”
The second reason is complex analysis has a large number of applications
(in both the pure math and applied math senses of the word) to things that
seem like they ought to have little to do with complex numbers. For example:

• Solving polynomial equations: historically, this was the motivation for

introducing complex numbers by Cardano, who published the famous
formula for solving cubic equations in 1543, after learning of the solu-
tion found earlier by Scipione del Ferro. An important point to keep in
mind is that Cardano’s formula sometimes requires taking operations
in the complex plane as an intermediate step to get to the final answer,
even when the cubic equation being solved has only real roots.

Example 1. Using Cardano’s formula, it can be found that the solutions

to the cubic equation
z 3 + 6z 2 + 9z + 3 = 0
are

z1 = 2 cos(2π/9) − 2,
z2 = 2 cos(8π/9) − 2,
z3 = 2 sin(π/18) − 2.
√
• Proving Stirling’s formula: n! ∼ 2πn(n/e)n . Here, an ∼ bn is the stan-
dard “asymptotic to” relation, defined to mean limn→∞ an /bn = 1.

• Proving the prime number theorem: π(n) ∼ logn n , where π(n) denotes
the number of primes less than or equal to n (the prime-counting func-
tion).
2 1 INTRODUCTION: WHY STUDY COMPLEX ANALYSIS?

• Proving many other asymptotic formulas in number theory and combi-

natorics, e.g. (to name one other of my favorite examples), the Hardy-
Ramanujan formula
1 √
p(n) ∼ √ eπ 2n/3 ,
4 3n
where p(n) is the number of integer partitions of n.

• Evaluation of complicated definite integrals, for example

∞
Z r
2 1 π
sin(t ) dt = .
0 2 2
(This application is strongly emphasized in older textbooks, and has
been known to result in a mild case of post-traumatic stress disorder.)

• Solving physics problems in hydrodynamics, heat conduction, electro-

statics and more.

• Analyzing alternating current electrical networks by extending Ohm’s

law to electrical impedance. Complex analysis also has many other
important applications in electrical engineering, signals processing and
control theory.

• Probability and combinatorics, e.g., the Cardy-Smirnov formula in per-

colation theory and the connective constant for self-avoiding walks on
the hexagonal lattice.

• It was proved in 2016 that the optimal densities for sphere packing in
8 and 24 dimensions are π 4 /384 and π 12 /12!, respectively. The proofs
make spectacular use of complex analysis (and more specifically, a
part of complex analysis that studies certain special functions known
as modular forms).

• Nature uses complex numbers in Schrödinger’s equation and quantum

field theory. This is not a mere mathematical convenience or sleight-of-
hand, but in fact appears to be a built-in feature of the very equations
describing our physical universe. Why? No one knows.

• Conformal maps, which come up in purely geometric applications where

the algebraic or analytic structure of complex numbers seems irrele-
vant, are in fact deeply tied to complex analysis. Conformal maps were
used by the Dutch artist M.C. Escher (though he had no mathematical
training) to create amazing art, and used by others to better under-
stand, and even to improve on, Escher’s work. See Fig. 1, and see [10]
for more on the connection of Escher’s work to mathematics.
3 2 THE FUNDAMENTAL THEOREM OF ALGEBRA

Figure 1: Print Gallery, a lithograph by M.C. Escher which was discovered to

be based on a mathematical structure related to a complex function z 7→ z α
for a certain complex number α, although it was constructed by Escher
purely using geometric intuition. See the paper [8] and this website, which
has animated versions of Escher’s lithograph brought to life using the math-
ematics of complex analysis.

• Complex dynamics, e.g., the iconic Mandelbrot set. See Fig. 2.

There are many other applications and beautiful connections of complex

analysis to other areas of mathematics. (If you run across some interesting
ones, please let me know!)
In the next section I will begin our journey into the subject by illustrating
a few beautiful ideas and along the way begin to review the concepts from
undergraduate complex analysis.

2 The fundamental theorem of algebra

One of the most famous theorems in complex analysis is the not-very-aptly
named Fundamental Theorem of Algebra. This seems like a fitting place to
start our journey into the theory.

Theorem 1 (The Fundamental Theorem of Algebra.). Every nonconstant

polynomial p(z) over the complex numbers has a root.
4 2 THE FUNDAMENTAL THEOREM OF ALGEBRA

Figure 2: The Mandelbrot set. [Source: Wikipedia]

The fundamental theorem of algebra is a subtle result that has many

beautiful proofs. I will show you three of them. Let me know if you see
any “algebra”. . .

First proof: analytic proof. Let

p(z) = an z n + an−1 z n−1 + . . . + a0
be a polynomial of degree n ≥ 1, and consider where |p(z)| attains its infi-
mum.
First, note that it can’t happen as |z| → ∞, since
|p(z)| = |z|n · (|an + an−1 z −1 + an−2 z −2 + . . . + a0 z −n |),
|p(z)|
and in particular lim|z|→∞ |z|n = |an |, so for large |z| it is guaranteed that
|p(z)| ≥ |p(0)| = |a0 |. Fixing some radius R > 0 for which |z| > R implies
|p(z)| ≥ |a0 |, we therefore have that
m0 := inf |p(z)| = inf |p(z)| = min |p(z)| = |p(z0 )|
z∈C |z|≤R |z|≤R

where z0 = arg min |p(z)|, and the minimum exists because p(z) is a continu-
|z|≤R
ous function on the disc DR (0).
Denote w0 = p(z0 ), so that m0 = |w0 |. We now claim that m0 = 0. As-
sume by contradiction that it doesn’t, and examine the local behavior of p(z)
around z0 ; more precisely, expanding p(z) in powers of z − z0 we can write
n
X
p(z) = w0 + cj (z − z0 )j = w0 + ck (z − z0 )k + . . . + cn (z − z0 )n ,
j=1
5 2 THE FUNDAMENTAL THEOREM OF ALGEBRA

where k is the minimal positive index for which cj 6= 0. (Exercise: why

can we expand p(z) in this way?) Now imagine starting with z = z0 and
traveling away from z0 in some direction eiθ . What happens to p(z)? Well, the
expansion gives

p(z0 + reiθ ) = w0 + ck rk eikθ + ck+1 rk+1 ei(k+1)θ + . . . + cn rn einθ .

When r is very small, the power r k dominates the other terms r j with k < j ≤
n, i.e.,

p(z0 + reiθ ) = w0 + rk (ck eikθ + ck+1 rei(k+1)θ + . . . + cn rn−k einθ )

= w0 + ck rk eikθ (1 + g(r, θ)),

where limr→0 |g(r, θ)| = 0. To reach a contradiction, it is now enough to

choose θ so that the vector ck r k eikθ “points in the opposite direction” from
w0 , that is, such that
ck rk eikθ
∈ (−∞, 0).
w0
Obviously this is possible: take θ = k1 (arg w0 − arg(ck ) + π). It follows that, for
r small enough,
|w0 + ck rk eikθ | < |w0 |
and for r small enough (possibly even smaller than the previous small r )

|p(z0 + reiθ )| = |w0 + ck rk eikθ (1 + g(r, θ))| < |w0 |,

a contradiction. This completes the proof.

Exercise 1. Complete the last details of the proof (for which r are the in-
equalities valid, and why?) Note that “complex analysis” is part of “analysis”
— you need to develop facility with such estimates until they become second
nature.

Second proof: topological proof. Let w0 = p(0). If w0 = 0 , we are done.

Otherwise consider the image under p of the circle |z| = r . Specifically:

1. For r very small the image is contained in a neighborhood of w0 , so it

cannot “go around” the origin.

2. For r very large we have

 
iθ an−1 −1 −iθ
n inθ a0 −n −inθ
p(re ) = an r e 1+ r e + ... + r e
an an
n inθ
= an r e (1 + h(r, θ))
6 2 THE FUNDAMENTAL THEOREM OF ALGEBRA

where limr→∞ h(r, θ) = 0 (uniformly in θ ). As θ goes from 0 to 2π , this is

a closed curve that goes around the origin n times (approximately in a
circular path, that becomes closer and closer to a circle as r → ∞).

As we gradually increase r from 0 to a very large number, in order to tran-

sition from a curve that doesn’t go around the origin to a curve that goes
around the origin n times, there has to be a value of r for which the curve
crosses 0. That means the circle |z| = r contains a point such that p(z) = 0,
which was the claim.

Remark 1. The argument presented in the topological proof is imprecise.

It can be made rigorous in a couple of ways — one way we will see a bit
later is using Rouché’s theorem and the argument principle. This already
gives a hint as to the importance of subtle topological arguments in complex
analysis.

Remark 2. The topological proof should be compared to the standard calcu-

lus proof that any odd-degree polynomial over the reals has a real root. That
argument is also “topological” (based on the mean value theorem), although
much more trivial.

Third proof: standard textbook proof (or: “hocus-pocus” proof). Recall:

Theorem 2 (Liouville’s theorem.). A bounded entire function is constant.

Assuming this result, if p(z) is a polynomial with no root, then 1/p(z)

is an entire function. Moreover, it is bounded, since as we noted before
lim|z|→∞ |p(z)|
|z|n
= |an |, so lim|z|→∞ 1/p(z) = 0. It follows that 1/p(z) is a constant,
which then has to be 0, which is a contradiction.

To summarize this section, we saw three proofs of the fundamental the-

orem of algebra. They are all beautiful — the “hocus-pocus” proof certainly
packs a punch, which is why it is a favorite of complex analysis textbooks
— but personally I like the first one best since it is elementary and doesn’t
use Cauchy’s theorem or any of its consequences, or subtle topological con-
cepts. Moreover, it is a “local” argument that is based on understanding how
a polynomial behaves locally, where by contrast the other two proofs can be
characterized as “global.” It is a general philosophical principle in analysis
(that has analogies in other areas, such as number theory and graph theory)
that local arguments are easier than global ones.
7 3 ANALYTICITY

3 Analyticity, conformality and the Cauchy-Riemann

equations
In this section we begin to build the theory by laying the most basic corner-
stone of the theory, the definition of analyticity, along with some of the useful
ways to think about this fundamental concept.

3.1 Definition and basic meanings of analyticity

Definition 1 (analyticity). A function f (z) of a complex variable is holomor-
phic (a.k.a. complex-differentiable, analytic 1 ) at z if the limit

f (z + h) − f (z)
f 0 (z) := lim
h→0 h
exists. In this case we call f 0 (z) the derivative of f at z .

In the case when f 0 (z) 6= 0, the existence of the derivative has a geometric
meaning: if we write the polar decomposition f 0 (z) = reiθ of the derivative,
then for points w that are close to z , we will have the approximate equality

f (w) − f (z)
≈ f 0 (z) = reiθ ,
w−z
or equivalently

f (w) ≈ f (z) + reiθ (w − z) + [lower order terms],

where “lower order terms” refers to a quantity that is much smaller in mag-
nitude that |w − z|. Geometrically, this means that to compute f (w), we start
from f (z), and move by a vector that results by taking the displacement vec-
tor w − z , rotating it by an angle of θ , and then scaling it by a factor of r
(which corresponds to a magnification if r > 1, a shrinking if 0 < r < 1, or
doing nothing if r = 1). This idea can be summarized by the slogan:

“Analytic functions behave locally as a rotation composed with a

scaling.”
1
Note: some people use “analytic” and “holomorphic” with two a priori different defini-
tions that are then proved to be equivalent; I find this needlessly confusing so I may use
these two terms interchangeably.
8 3 ANALYTICITY

The local behavior of analytic functions in the case f 0 (z) = 0 is more

subtle; we will consider that a bit later.
A further interpretation of the meaning of analyticity is that analytic func-
tions are conformal mappings where their derivatives don’t vanish. More
precisely, if γ1 , γ2 are two differentiable planar curves such that γ1 (0) =
γ2 (0) = z , f is differentiable at z and f 0 (z) 6= 0, then, denoting v1 = γ10 (0),
v2 = γ20 (0), w1 = (f ◦ γ1 )0 (0), w2 = (f ◦ γ2 )0 (0), we can write the inner prod-
ucts (in the ordinary sense of vector geometry) between the complex number
pairs v1 , v2 and w1 , w2 as
hv1 , v2 i = Re(v1 v2 ),
hw1 , w2 i = h(f 0 (γ1 (0))γ10 (0)), (f 0 (γ2 (0))γ20 (0))i
= f 0 (z)f 0 (z)hv1 , v2 i = |f 0 (z)|2 hv1 , v2 i.
If we denote by θ (resp. ϕ) the angle between v1 , v2 (resp. w1 , w2 ), it then
follows that
hw1 , w2 i |f 0 (z)|2 hv1 , v2 i hv1 , v2 i
cos ϕ = = 0 = = cos θ.
|w1 | |w2 | |f (z)v1 | |f 0 (z)v2 | |v1 | |v2 |
That is, the function f maps two curves meeting at an angle θ at z to two
curves that meet at the same angle at f (z). A function with this property is
said to be conformal at z .
Conversely, if f is conformal in a neighborhood of z then (under some
additional mild assumptions) it is analytic — we will prove this below after
discussing the Cauchy-Riemann equations. Thus the theory of analytic func-
tions contains the theory of planar conformal maps as a special (and largely
equivalent) case, although this is by no means obvious from the purely geo-
metric definition of conformality.
Let us briefly review some properties of derivatives.
Lemma 1. Under appropriate assumptions, we have the relations
(f + g)0 (z) = f 0 (z) + g 0 (z),
(f g)(z) = f 0 (z)g(z) + f (z)g 0 (z),
 0
1 f 0 (z)
=− ,
f f (z)2
 0
f f 0 (z)g(z) − f (z)g 0 (z)
= ,
g g(z)2
(f ◦ g)0 (z) = f 0 (g(z))g 0 (z).
Exercise 2. Explain precisely what the assumptions in the lemma are (see
Proposition 2.2 on page 10 of [11]).
9 3 ANALYTICITY

3.2 The Cauchy-Riemann equations

In addition to the geometric picture associated with the definition of the
complex derivative, there is yet another quite different but also extremely
useful way to think about analyticity, that provides a bridge between com-
plex analysis and ordinary multivariate calculus. Remembering that complex
numbers are vectors that have real and imaginary components, we can de-
note z = x + iy , where x and y will denote the real and imaginary parts of the
complex number z , and f = u + iv , where u and v are real-valued functions
of z (or equivalently of x and y ) that return the real and imaginary parts,
respectively, of f . Now, if f is analytic at z then

f (z + h) − f (z)
f 0 (z) = lim
h→0 h
u(x + h + iy) − u(x + iy) v(x + h + iy) − v(x + iy)
= lim +i
h→0, h∈R h h
∂u ∂v
= +i .
∂x ∂x
On the other hand also

f (z + h) − f (z)
f 0 (z) = lim
h→0 h
u(x + h + iy) − u(x + iy) v(x + h + iy) − v(x + iy)
= lim +i
h→0, h∈iR h h
u(x + iy + ih) − u(x + iy) v(x + iy + ih) − v(x + iy)
= lim +i
h→0, h∈R ih ih
∂u ∂v ∂v ∂u
= −i −i·i = −i .
∂y ∂y ∂y ∂y
Since these limits are equal, by equating their real and imaginary parts we
get a famous system of coupled partial differential equations, the Cauchy-
Riemann equations:
∂u ∂v ∂v ∂u
= , =− .
∂x ∂y ∂x ∂y
We have proved that if f is analytic at z = x + iy then the components
u, v of f satisfy the Cauchy-Riemann equations. Conversely, we now claim
if f = u + iv is continuously differentiable at z = x + iy (in the sense that
each of u and v is a continuously differentiable function of x, y as defined in
ordinary real analysis) and satisfies the Cauchy-Riemann equations there, f
is analytic at z .
10 3 ANALYTICITY

Proof. The assumption implies that f has a differential at z , i.e., in the no-
tation of vector calculus, denoting f = (u, v), z = (x, y)> , ∆z = (h1 , h2 )> , we
have ! 
  ∂u ∂u
u(z) ∂x ∂y h1
f (z + ∆z) = + ∂v ∂v + E(h1 , h2 ),
v(z) ∂x ∂y
h2
where E(h1 , h2 ) = o(|∆z|) as |∆z| → 0. Now, by the assumption that the
Cauchy-Riemann equations hold, we also have
∂u ∂u
!  !
∂u
∂x ∂y h1 h + ∂u
∂x 1
h
∂y 2
∂v ∂v = ∂u ∂u ,
∂x ∂y
h2 − ∂y h1 + ∂x h2
which is the vector calculus notation for the complex number
   
∂u ∂u ∂u ∂u
−i (h1 + ih2 ) = −i ∆z.
∂x ∂y ∂x ∂y
So, we have shown that (again, in complex analysis notation)
 
f (z + ∆z) − f (z) ∂u ∂u E(∆z) ∂u ∂u
lim = lim −i + = −i .
∆z→0 ∆z ∆z→0 ∂x ∂y ∆z ∂x ∂y
This proves that f is holomorphic at z with derivative given by f 0 (z) = ∂u
∂x
−
i ∂u
∂y
.

With the help of the Cauchy-Riemann equations, we can now prove our
earlier claim that conformality implies analyticity.
Theorem 3. If f = u + iv is conformal at z , continuously differentiable in the
real analysis sense, and satisfies det Jf > 0 (i.e., f preserves orientation as a
planar map), then f is holomorphic at z .

Proof. In the notation of the proof above, we have as before that

  ∂u ∂u
! 
u(z) ∂x ∂y h1
f (z + ∆z) = + ∂v ∂v + E(h1 , h2 ),
v(z) ∂x ∂y
h2

where E(h1 , h2 ) = o(|∆z|) as |∆z| → 0. The assumption is that the differential

map ! ∂u ∂u
∂x ∂y
Jf = ∂v ∂v
∂x ∂y
preserves orientation and is conformal; the conclusion is that the Cauchy-
Riemann equations are satisfied (which would imply that f is holomorphic at
z by the result shown above). So the theorem will follow once we prove the
simple claim about 2 × 2 matrices contained in Lemma 2 below.
11 3 ANALYTICITY

 
a b
Lemma 2 (Conformality lemma.). Assume that A = is a 2 × 2 real
c d
matrix. The following are equivalent:

(a) A preserves orientation (that is, det A > 0) and is conformal, that is

hAw1 , Aw2 i hw1 , w2 i

=
|Aw1 | |Aw2 | |w1 | |w2 |

for all w1 , w2 ∈ R2 .
 
a b
(b) A takes the form A = for some a, b ∈ R with a2 + b2 > 0.
−b a
 
cos θ − sin θ
(c) A takes the form A = r for some r > 0 and θ ∈ R. (That
sin θ cos θ
is, geometrically A acts by a rotation followed by a scaling.)

Proof that (a) =⇒ (b). Note that both columns of A are nonzero vectors by
the assumption that det A > 0. Now applying the conformality assumption
with w1 = (1, 0)> , w2 = (0, 1)> yields that (a, c) ⊥ (b, d), so that (b, d) = κ(−c, a)
for some κ ∈ R \ {0}. On the other hand, applying the conformality assump-
tion with w1 = (1, 1)> and w2 = (1, −1)> yields that (a + b, c + d) ⊥ (a − b, c − d),
which is easily seen to be equivalent to a2 + c2 = b2 + d2 . Together with the
previous
  relation
  implies that κ = ±1. So A is of one of the two forms
that
a −c a c
or . Finally, the assumption that det A > 0 means it is the
c a c −a
first of those two possibilities that must occur.

Exercise 3. Complete the proof of the lemma above by showing the implica-
tions (b) ⇐⇒ (c) and that (b) =⇒ (a).

Another curious consequence of the Cauchy-Riemann equations, which

gives an alternative geometric picture to that of conformality, is that analyt-
icity implies the orthogonality of the level curves of u and of v . That is, if
f = u + iv is analytic then

h∇u, ∇vi = (ux , uy ) ⊥ (vx , vy ) = ux vx + uy vy = vy vx − vx vy = 0.

Since ∇u (resp. ∇v ) is orthogonal to the level curve {u = c} (resp. the level

curve {v = d}, this proves that the level curves {u = c}, {v = d} meet at right
angles whenever they intersect.
Yet another important and remarkable consequence of the Cauchy-Riemann
equations is that, at least under mild assumptions (which we will see later
12 3 ANALYTICITY

4 4 4

2 2 2

-4 -2 2 4 -4 -2 2 4 -4 -2 2 4

-2 -2 -2

-4 -4 -4

(a) (b) (c)

Figure 3: The level curves for the (a) real and (b) imaginary parts of z 2 =
(x2 − y 2 ) + i(2xy). (c) shows the superposition of both families of level curves.

2
1.0 1.0

1
0.5 0.5

0.0 0 0.0

-0.5 -1 -0.5

-1.0 -2 -1.0
-1.0 -0.5 0.0 0.5 1.0 -2 -1 0 1 2 -1.0 -0.5 0.0 0.5 1.0

(a) (b) (c)

Figure 4: The level curves for the real and imaginary parts of z −1 = x
x2 +y 2
−
y
i x2 +y 2.

can be removed) the functions u, v are harmonic functions. Assume that f is

analytic at z and twice continuously differentiable there. Then
∂ 2u ∂ 2u
   
∂ ∂u ∂ ∂u
+ = +
∂x2 ∂y 2 ∂x ∂x ∂y ∂y
∂ 2v ∂ 2v
   
∂ ∂v ∂ ∂v
= − = − = 0,
∂x ∂y ∂y ∂x ∂x∂y ∂y∂x
i.e., u satisfies Laplace’s equation

4u = 0,
∂ ∂ 2 2
where 4 = ∂x 2 + ∂y 2 is the two-dimensional Laplacian operator. Similarly

(check), v also satisfies

∂ 2v ∂ 2v
4v = + = 0.
∂x2 ∂y 2
13 4 POWER SERIES

That is, we have shown that u and v are harmonic functions. This is an
extremely important connection between complex analysis and the theory of
partial differential equations, which also relates to many other areas of real
analysis.
We will later see that the assumption of twice continuous differentiability
is unnecessary, but proving this requires some subtle complex-analytic ideas.
A final remark related to analyticity and the Cauchy-Riemann equation is
the observation that if f = u + iv is analytic then its Jacobian (in the sense of
multivariate calculus when we consider it as a map from R2 to R2 ) is given
by  
ux uy
Jf = det = ux vy − uy vx = u2x + vx2 = |ux + ivx | = |f 0 (z)|2 .
vx vy
This can also be understood geometrically. (Exercise: how?)

4 Power series
Until now we have not discussed any specific examples of functions of a com-
plex variable. Of course, there are the standard functions that you probably
encountered already in your undergraduate studies: polynomials, rational
functions, ez , the trigonometric functions, etc. But aside from these exam-
ples, it would be useful to have a general way to construct a large family
of functions. Of course, there is such a way: power series, which—non-
obviously—turn out to be essentially as general a family of functions as one
could hope for.
To make things precise, a power series is a function of a complex variable
z that is defined by
∞
X
f (z) = an z n
n=0

where (an )∞
n=0 is a sequence of complex numbers, or more generally by

∞
X
g(z) = f (z − z0 ) = an (z − z0 )n
n=0

where (an )∞
n=0 is again a a sequence and z0 is some fixed complex number.
These functions are defined wherever the respective series converge.
For which values of z does this formula make sense? It is not hard to see
that it converges absolutely precisely for 0 ≤ |z| < R, where the value of R is
14 4 POWER SERIES

given by
 −1
1/n
R = lim sup |an | .
n→∞
R is called the radius of convergence of the power series.

Proof. Assume 0 < R < ∞ (the edge cases R = 0 and R = ∞ are left as
an exercise).
The defining property of R is that for all  > 0, we have that
n
|an | < R1 +  if n is large enough, and R is the minimal number with that
1


property. Let z ∈ DR (0). Since |z| < R, we have |z| R +  < 1 for some
fixed  > 0 chosen small enough. That implies that for n > N (for some large
enough N as a function of ),
∞ ∞   n
X
n
X 1
|an z | < +  |z| ,
n=N n=N
R
so the series is dominated by a convergent geometric series, and hence con-
verges.
Conversely, if |z| > R, then, |z| R1 −  > 1 for some small enough fixed



nk
 > 0. Taking a subsequence (ank )∞ 1
k=1 for which |ank | > R −  (guaranteed
to exist by the definition of R), we see that
∞ ∞   nk
X
n
X 1
|an z | ≥ |z| − = ∞,
n=0 k=1
R
so the power series diverges.
Exercise 4. Complete the argument in the extreme cases R = 0, ∞.
Another important theorem is:
Theorem 4. Power series are holomorphic functions in the interior of the
disc of convergence and can be differentiated termwise.

Proof. Denote
∞
X
f (z) = an z n = SN (z) + EN (z),
n=0
XN
SN (z) = an z n ,
n=0
X∞
EN (z) = an z n ,
n=N +1
∞
X
g(z) = nan z n−1 .
n=1
15 4 POWER SERIES

The claim is that f is differentiable on the disc of convergence and its deriva-
tive is the power series g . Since n1/n → 1 as n → ∞, it is easy to see that f (z)
and g(z) have the same radius of convergence. Fix z0 with |z| < r < R. We
f (z0 +h)−f (z0 )
wish to show that h
converges to g(z0 ) as h → 0. Observe that
 
f (z0 + h) − f (z0 ) SN (z0 + h) − SN (z0 ) 0
− g(z0 ) = − SN (z0 )
h h
EN (z0 + h) − EN (z0 ) 0
+ + (SN (z0 ) − g(z0 ))
h
The first term converges to 0 as h → 0 for any fixed N . To bound the second
term, fix some  > 0, and note that, if we assume that not only |z0 | < r but
also |z0 + h| < r (an assumption that’s clearly satisfied for h close enough to
0) then
∞
 (z0 + h)n − z0n 
   
 EN (z0 + h) − EN (z0 )  X
 ≤ |an |  
 h 
n=N +1
 h 
∞
 P 
 h n−1 hk (z + h)n−1−k 
0
X
k=0
= |an | 
 
h

 
n=N +1
∞
X
≤ |an |nrn−1 ,
n=N +1

where we use the algebraic identity

an − bn = (a − b)(an−1 + an−2 b + . . . + abn−2 + bn−1 ).

The last expression in this chain of inequalities is the tail of an absolutely

convergent series, so can be made <  be taking N large enough (before
taking the limit as h → 0).
0
Third, when choosing N also make sure it is chosen so that |SN (z0 ) −
0
g(z0 )| < , which of course is possible since SN (z0 ) → g(z0 ) as N → ∞.
Finally, having thus chosen N , we get that
 
 f (z0 + h) − f (z0 ) 
lim sup  − g(z0 ) ≤ 0 +  +  = 2.
h→0 h
f (z0 +h)−f (z0 )
Since  was an arbitrary positive number, this shows that h
→ g(z0 )
as h → 0, as claimed.

The proof above can be thought of as a special case of the following more
conceptual result: if gn is a sequence of holomorphic functions on a region Ω,
16 5 CONTOUR INTEGRALS

and gn → g uniformly on closed discs in Ω, gn0 → h uniformly on closed discs

on Ω, and h is continuous, then g is holomorphic and g 0 = h on Ω. (Exercise:
prove this, and explain the connection to the previous result.)

Corollary 1. Analytic functions defined as power series are differentiable (in

the complex-analytic sense) infinitely many times in the disc of convergence.

P∞
Corollary 2. For a power series g(z) = n=0 an (z −z0 )n with a positive radius
of convergence, we have
g (n) (z0 )
an = .
n!
In other words g(z) satisfies Taylor’s formula
∞
X g (n) (z0 )
g(z) = (z − z0 )n .
n=0
n!

5 Contour integrals
We now introduce contour integrals, which are another fundamental building
block of the theory.
Contour integrals, like many other types of integrals, take as input a func-
tion to be integrated and a “thing” (or “place”) over which the function is
integrated. In the case of contour integrals, the “thing” is a contour, which
is (for our current purposes at least) a kind of planar curve. We start by de-
veloping some terminology to discuss such objects. First, there is the notion
of a parametrized curve, which is simply a continuous function γ : [a, b] → C.
The value γ(a) is called the starting point and γ(b) is called the ending point.
Two curves γ1 : [a, b] → C, γ2 : [c, d] → C are called equivalent, which is
denoted γ1 ∼ γ2 , if γ2 (t) = γ1 (I(t)) where I : [c, d] → [a, b] is a continuous,
one-to-one, onto, increasing function. A “curve” is an equivalence class of
parametrized curves with respect to this equivalence relation.
In practice, we will usually refer to parametrized curves as “curves”,
which is the usual abuse of terminology that one sees in various places in
mathematics, in which one blurs the distinction between equivalence classes
and their members, remembering that various definitions, notation, and proof
arguments need to “respect the equivalence” in the sense that they do not
depend of the choice of member. (Meta-exercise: think of 2–3 other examples
of this phenomenon.)
17 5 CONTOUR INTEGRALS

For our present context of developing the theory of complex analysis, we

shall assume all our curves are piecewise continuously differentiable. More
generally, one can assume them to be rectifiable, but we will not bother to
develop that theory. There are yet more general contexts in which allowing
curves to be merely continuous is beneficial (and indeed some of the ideas
we will develop in a complex-analytic context can be carried over to that
more general setting), but we will not pursue such distractions either.
You probably encountered curves and parametrized curves in your earlier
studies of multivariate calculus, where they were used to define the notion
of line integrals of vector and scalar fields. Recall that there are two types
of line integrals, which are referred to as line integrals of the first and sec-
ond kind. The line integral of the first kind of a scalar (usually real-valued)
function u(z) over a curve γ is defined as
Z n
X
u(z)ds = lim u(zj )∆sj (line integral of the first kind),
γ max ∆sj →0
j j=1

where the limit is a limit of Riemann sums with respect to a family of parti-
tions of the interval [a, b] over which the curve γ is defined, as the norm of the
partitions shrinks to 0. Here the partition points are a = t0 < t1 < . . . < tn = b,
the points zj = f (tj ) are their images on the curve γ , and the symbols ∆sj
refer to finite line elements, namely ∆sj = |zj − zj−1 |.
The line integral of the second kind is defined for a vector field F = (P, Q)
(the more traditional notation from calculus for what we would denote in the
current context as the complex-valued function F = P + iQ) by
Z Z n
X
F · ds = P dx + Q dy = lim P (zj )∆xj + Q(zj )∆yj ,
γ γ max ∆sj →0
j j=1

where zj are as before and xj = Re(zj ), yj = Im(zj ).

It is well-known from calculus that line integrals can be expressed in
terms of ordinary (single-variable) Riemann integrals. Take a couple of min-
utes to remind yourself of why the following formulas are true (assuming all
the functions involved are piecewise continuously differentiable):
Z Z b
u(z) ds = u(γ(t))|γ 0 (t)| dt,
γ a
Z Z b
F · ds = F(γ(t)) · γ 0 (t) dt.
γ a
18 5 CONTOUR INTEGRALS

As a further reminder, the basic result known as the fundamental theorem

of calculus for line integrals states that if F = ∇u then
Z
F · ds = u(γ(b)) − u(γ(a)).
γ

Definition 2 (contour integrals and arc length intervals). For a function

f = u + iv of a complex variable z and a curve γ , define
Z Z
f (z) dz = “ (u + iv)(dx + idy)”
γ γ
Z  Z 
= u dx − v dy + i v dx + u dy
γ γ
Z b
f (γ(t))γ 0 (t) dt
= (contour integral),
Z Za Z Z
f (z) |dz| = f (z) ds = u ds + i v ds (arc length integral).
γ γ γ γ

If γ is a closed curve (the two endpointsHare the same, i.e., it satisfies

H γ(a) =
γ(b)), we denote the contour integral as f (z) dz , and similarly f (z) |dz| for
γ γ
the arc length integral.

A special case of an arc length integral is the length of the curve, defined
as the integral of the constant function 1:
Z Z b
len(γ) = |dz| = |γ 0 (t)| dt.
γ a

As mentioned above, our convention of mildly abusing terminology puts

on us the burden of having to remeber to check that these definitions do
not depend on the parametrization of the curve. Indeed: if γ1 ∼ γ2 are
representatives of the same equivalence class of parametrized curves, that
is, γ2 (t) = γ1 (I(t)) for some nicely-behaved function, then using a standard
change of variables in single-variable integrals we see that
Z Z d Z d
f (z) dz = f (γ2 (t))γ20 (t)dt = f (γ1 (I(t)))(γ1 ◦ I)0 (t) dt
γ2 c c
Z d Z b
= f (γ1 (I(t)))γ10 (I(t))I 0 (t) dt = f (γ1 (τ ))γ10 (τ ) dτ
Zc a

= f (z) dz.
γ1
19 5 CONTOUR INTEGRALS

Exercise 5. Show that the integral with respect to arc length similarly does
not depend on the parametrization.

Contour integrals have many surprising properties, but the ones on the
following list of basic properties are not of the surprising kind:

Proposition 1 (properties of contour integrals). Contour integrals satisfy

the following properties:
R R
(a) Linearity as an operator on functions: γ
(αf (z)+βg(z)) dz = α γ
f (z) dz+
R
β γ
g(z) dz .
(b) Linearity as an operator on curves: if a contour Γ is a “composition” of
two contours γ1 and γ2 (in a sense that is easy to define graphically, but
tedious to write down precisely), then
Z Z Z
f (z) dz = f (z) dz + f (z) dz.
Γ γ1 γ2

Similarly, if γ2 is the “reverse” contour of γ1 , then

Z Z
f (z) dz = − f (z) dz.
γ2 γ1

(c) Triangle inequality:

Z  Z
 
 f (z) dz  ≤ |f (z)| |dz| ≤ len(γ) · sup |f (z)|.
  z∈γ
γ

Exercise 6. Prove this result (part of the exercise is to define precisely the
notions of “composition of curves” and “reverse curve”).

Contour integrals have their own version of the fundamental theorem of

calculus.

Theorem 5 (The fundamental theorem of calculus for contour integrals.). If

γ is a curve connecting two points w1 , w2 in a region Ω on which a function F
is holomorphic, then
Z
F 0 (z) dz = F (w2 ) − F (w1 ).
γ

R Equivalently, the theorem says that to compute a general contour integral

γ
f (z) dz , we try to find a primitive (a.k.a. anti-derivative) of f , that is, a
function F such that F 0 (z)R = f (z) on all of Ω. If we found such a primitive
then the contour integral γ f (z) dz is given by F (w2 ) − F (w1 ).
20 5 CONTOUR INTEGRALS

Proof. For smooth curves, an easy application of the chain rule gives
Z Z b Z b
0 0 0
F (z) dz = F (γ(t))γ (t) dt = (F ◦ γ)0 (t) dt = (F ◦ γ)(t)|t=b
t=a
γ a a
= F (γ(b)) − F (γ(a)) = F (w2 ) − F (w1 ).
For piecewise smooth curves, this is a trivial extension that is left as an
exercise.

Many of our discussions of contour integrals will involve the behavior of

integrals over closed contours, and the interplay between the properties of
such integrals and integrals over general contours. As an example of this
interplay, the above result has an easy — but important — consequence for
integrals over closed contours.
Corollary 3. If f = F 0 where F is holomorphic on a region Ω (in that case
we say that f has a primitive), γ is a closed curve in Ω, then
I
f (z) dz = 0.
γ

This last result has a partial converse:

Proposition 2. if f : Ω → C is a continuous function on a region Ω such that
I
f (z) dz = 0
γ

holds for any closed contour in Ω, then f has a primitive.

Proof. Fix some z0 ∈ Ω. For any z ∈ Ω, there is some path γ(z0 , z) connecting
z0 and z (since Ω is connected and open, hence pathwise-connected — a
standard exercise in topology, see the exercises in Chapter 1 of [11]). Define
Z
F (z) = f (w) dw.
γ(z0 ,z)

By the assumption, this integral does not depend on which contour γ(z0 , z)
connecting z0 and z was chosen, so F (z) is well-defined. We now claim that
F is holomorphic and its derivative is equal to f . To see this, note that
F (z + h) − F (z)
− f (z)
h Z Z 
1
= f (w) dw − f (w) dw − f (z)
h γ(z0 ,z+h) γ(z0 ,z)
Z Z
1 1
= f (w) dw − f (z) = (f (w) − f (z)) dw
h γ(z,z+h) h γ(z,z+h)
21 6 CAUCHY’S THEOREM

where γ(z, z + h) denotes a contour connecting z and z + h. When |h| is

sufficiently small so that the disc D(z, h) is contained in Ω, one can take
γ(z, z + h) as the straight line segment connecting z and z + h. For such h we
get that
 
 F (z + h) − F (z)  1
 − f (z) ≤ len(γ(z, z + h)) sup |f (w) − f (z)|
 h  h
w∈D(z,h)

= sup |f (w) − f (z)| −−→ 0,

w∈D(z,h) h→0

by continuity of f .
Remark 3. Note that with the last result, if we knew that holomorphic func-
tions are differentiable infinitely many times (the so-called regularity theo-
rem), we could conclude that a function that satisfies the assumption that
all its contour integrals on closed contours were 0 is holomorphic. This is in
fact true, and is called Morera’s theorem (and is an important fact in com-
plex analysis), but we won’t be able to prove it until we’ve proved Cauchy’s
theorem.
H
Example 2. Compute |z|=1 z n dz for n ∈ Z. What do we learn from the fact
that the integral is not zero for n = −1? (Hint: something; but what?) And
what do we learn from the fact that it’s 0 when n 6= −1? (Hint: nothing; but
why?)
Lemma 3. If f is holomorphic on Ω and f 0 ≡ 0 then f is a constant.

Proof. Fix some z0 ∈ Ω. For any z ∈ Ω, as we discussed above there is a path

γ(z0 , z) connecting z0 and z . Then
Z
f (z) − f (z0 ) = f 0 (w) dw = 0,
γ(z0 ,z)

hence f (z) ≡ f (z0 ), so f is constant.

6 Cauchy’s theorem
One of the central results in complex analysis is Cauchy’s theorem.
Theorem 6 (Cauchy’s theorem.). If f is holomorphic on a simply-connected
region Ω, then for any closed curve in Ω we have
I
f (z) dz = 0.
γ
22 6 CAUCHY’S THEOREM

The challenges facing us are: first, to prove Cauchy’s theorem for curves
and regions that are relatively simple (where we do not have to deal with
subtle topological considerations); second, to define what simply-connected
means; third, which will take a bit longer and we won’t do immediately, to
extend the theorem to the most general setting.
Two other theorems that are closely related to Cauchy’s theorem are
Goursat’s theorem, a relatively easy special case of Cauchy’s theorem, and
Morera’s theorem which is a kind of converse to Cauchy’s theorem.
Theorem 7 (Goursat’s theorem).H If f is holomorphic on a region Ω, and T
is a triangle contained in Ω, then ∂T f (z) dz = 0 (where T refers to the “full”
triangle, and ∂T refers to its boundary considered as a curve oriented in the
usual positive direction).
Theorem 8 (Morera’s theorem). If f : Ω → C is a continuous function on a
region Ω such that I
f (z) dz = 0
γ
holds for any closed contour in Ω, then f is holomorphic on Ω.

Proof of Goursat’s theorem. The proof can be summarized with a slogan:

“localize the damage.” Namely, try to translate a global statement about
the integral around the triangle to a local statement about behavior near a
specific point inside the triangle, which would become manageable since we
have a good understanding of the local behavior of a holomorphic function
near a point. If something goes wrong with the global integral, something
has to go wrong at the local level, and we will show that can’t happen (al-
though technically the proof is not a proof by contradiction, conceptually I
find this a helpful way to think about it).
The idea can be made more precise using triangle subdivision. Specifi-
cally, let T (0) = T , and define a hierarchy of subdivided triangles

order 0 triangle: T (0) ,

(1)
order 1 triangles: Tj , 1 ≤ j ≤ 4,
(2)
order 2 triangles: Tj,k , 1 ≤ j, k ≤ 4
(3)
order 3 triangles: Tj,k,` , 1 ≤ j, k, ` ≤ 4,
..
.
(n)
order n triangles: Tj1 ,...,jn , 1 ≤ j1 , . . . , jn ≤ 4.
..
.
23 6 CAUCHY’S THEOREM

(n)
Here, the triangles Tj1 ,...,jn for jn = 1, 2, 3, 4 are obtained by subdividing the
(n−1)
order n − 1 triangle Tj1 ,...,jn−1 into 4 subtriangles whose vertices are the ver-
(n−1)
tices and/or edge bisectors of Tj1 ,...,jn−1 (see Figure 1 on page 35 of [11]).
Now, given the way this subdivision was done, it is clear that we have the
equality
I 4 I
X
f (z) dz = f (z) dz
(n−1) (n)
∂Tj jn =1 ∂Tj
1 ,...,jn−1 1 ,...,jn

due to cancellation along the internal edges, and hence

I 4
X I
f (z) dz = f (z) dz.
(n)
∂T (0) j1 ,...,jn =1 ∂Tj
1 ,...,jn

That is, the integral along the boundary of the original triangle is equal to
the sum of the integrals over all 4n triangles of order n. Now, the crucial
observation is that one of these integrals has to have a modulus that is at
least as big as the average. That is, we have
I  I 
I  4
X    
 
f (z) dz  ≤ f (z) dz  ≤ 4n  f (z) dz 
    
 (n)

∂T (0)  ∂Tj ,...,j   ∂T (n) 
j1 ,...,jn =1 1 n j(n)

(n) (n)
where j(n) = (j1 , . . . , jn ) is some n-tuple chosen such that the second in-
equality holds. Moreover, we can choose j(n) inductively in such a way that
(n) (n) (n−1)
the triangles Tj(n) are nested — that is, Tj(n) ⊂ Tj(n−1) for n ≥ 1, or equiva-
(n−1) (n−1)
lently j(n) = (j1  , k) for some 1 ≤ k ≤ 4 — to make this happen,
, . . . , jn−1
H 
choose k to be such that  ∂T (n) f (z) dz  is greater than (or equal to) the
(j(n−1),k)
average I 
4
1 X  
f (z) dz  ,

4
 (n)
 ∂T 
d=1 (j(n−1),d)

which in turn is (by induction) greater than or equal to

 4 I  I  I
1 X   
−(n−1)
f (z) dz  =  f (z) dz  ≥ 4 f (z) dz.
   
4
 (n)
∂T   ∂T (n−1)  ∂T
d=1 (j(n−1),d) j(n−1)

Now observe that the sequence of nested triangles shrinks to a single point.
That is, we have
∞
\ (n)
Tj(n) = {z0 }
n=0
24 6 CAUCHY’S THEOREM

for some point z0 ∈ T . This is true because the diameter of the triangles goes
to 0 as n → ∞, so certainly there can’t be two distinct points in the inter-
section; whereas, on the other hand, the intersection cannot be empty, since
the sequence (zn )∞
n=0 of centers (in some obvious sense, e.g., intersection of
the angle bisectors) of each of the triangles is easily seen to be a Cauchy se-
quence (and hence a convergent sequence, by the completeness property of
the complex numbers), whose limit must be an element of the intersection.
Having defined z0 , write f (z) for z near z0 as

f (z) = f (z0 ) + f 0 (z0 )(z − z0 ) + ψ(z)(z − z0 ),

where
f (z) − f (z0 )
ψ(z) = − f 0 (z0 ).
z − z0
The holomorphicity of f at z0 implies that ψ(z) → 0 as z → z0 . Denote by d(n)
(n)
the diameter of Tj(n) and by p(n) its perimeter. Each subdivision shrinks both
the diameter and perimeter by a factor of 2, so we have

d(n) = 2−n d(0) , p(n) = 2−n p(0) .

It follows that
Z  Z 
   
0
 (n) f (z) dz  =  (n) f (z0 ) + f (z0 )(z − z0 ) + ψ(z)(z − z0 ) dz 
   
 ∂T   ∂T 
j(n)
Z j(n) 
 
= ψ(z)(z − z0 ) dz  ≤ p(n) d(n) sup |ψ(z)|
 
 ∂T (n)  (n)
j(n) z∈T j(n)

H  R
Finally, combining this with the relationship between  ∂T (0)
f (z) dz  and | ∂T (n) f (z) dz|,
j(n)
we get that Z 
 
 f (z) dz  ≤ p(0) d(0) sup |ψ(z)| −−−→ 0,

∂T (0) (n) n→∞
z∈Tj(n)

which finishes the proof.

In the next section we will explore some of the amazing consequences

of this ultimately quite simple result. But first, let us note a few not-very-
amazing consequences..

Corollary 4 (Goursat’s theorem for rectangles.). Theorem 7 is also true

when we replace the word “triangle” with “rectangle.”
25 6 CAUCHY’S THEOREM

Proof. Obvious: a rectangle can be decomposed as the union of two trian-

gles, with the contour integral around the rectangle being the sum of the
integrals around the two triangles due to cancellation of the integrals going
in both directions along the diagonal (see Prop 1(b)).

Corollary 5 (existence of a primitive for a holomorphic function on a disc).

If f is holomorphic on a disc D , then f = F 0 for some holomorphic function
F on D.

Proof. The idea is similar to the proof of Proposition 2 above. If we knew

that all contour integrals of f around closed contours vanished, that result
would give us what we want. As it is, we know this is true but only for trian-
gular contours. How can we make use of that information? The textbook [11]
gives a clever approach in which the contour γ(z0 , z) is comprised of a hori-
zontal line segment followed by a vertical line segment. Then one shows in
three steps, each involving a use of Goursat’s theorem (see Figure 4 on page
38 of [11]), that F (z0 + h) − F (z0 ) is precisely the contour integral over the
line segment connecting z0 and z0 + h. From there the theorem proceeds in
exactly the same way as before.

Corollary
H 6 (Cauchy’s theorem for a disc.). If f is holomorphic on a disc,
then γ f dz = 0 for any closed contour γ in the disc.

Proof. f has a primitive, and we saw that that implies the claimed conse-
quence.

Theorem 9 (Cauchy’s
R theorem for a region enclosed by a “toy contour”).
The statement γ f (z)dz = 0 is also true for a function that’s analytic in a
region enclosed by a contour that is simple enough that the method of proof
used for the disc above can be extended to it.2

Proof. Repeat the same ideas, going from Goursat’s theorem, to the fact
that the function has a primitive, to the fact that its contour integrals along
closed curves vanish. The difficulty as the toy contour gets more complicated
is to make sure that the geometry works out when proving the existence of
the primitive — see for example the (incomplete) discussion of the case of
“keyhole contours” on pages 40–41 of [11].
2
The book [11] calls such a contour a “toy contour”, leaving the term as a somewhat
vaguely defined meta-concept.
26 7 CONSEQUENCES OF CAUCHY’S THEOREM

7 Consequences of Cauchy’s theorem

Theorem 10 (Cauchy’s integral formula). If f is holomorphic on a region Ω,
and C = ∂D is a circular contour contained in Ω, then



 f (z) if z ∈ D,
I
1 f (w) 
dw = 0 if z ∈ Ω \ D,
2πi C w − z 

undefined if z ∈ C.


Proof. The case when z ∈ / D is covered by Cauchy’s theorem in a disc, since

in that case the function w 7→ f (w)/(w − z) is holomorphic in an open set
containing D . It remains to deal with the case z ∈ D . In this case, denote by
z0 the center of the circle C . The idea is now to consider instead the integral
I I
f (w)
Fz (w) dw = dw,
Γ,δ Γ,δ w−z

where Γ,δ is a so-called keyhole contour, namely a contour comprised of

a large circular arc around z0 that is a subset of the circle C , and another
smaller circular arc of radius  centered at z , with two straight line segments
connecting the two circular arcs to form a closed curve, such that the width
of the “neck” of the keyhole is δ (think of δ as being much smaller than );
see Fig. 5. Note that the function Fz (w) is holomorphic inside the region
enclosed by Γ,δ , so Cauchy’s theorem for toy contours (assuming you can be
convinced that the keyhole contour is a toy contour) gives that
I
Fz (w) dw = 0.
Γ,δ

As δ → 0, the two parts of the integral along the “neck” of the contour
Γ,δ cancel out in the limit because Fz is continuous, and hence uniformly
continuous, on the compact set D \ D(z, ). So we can conclude that
I I
Fz (w) dw = Fz (w) dw.
C |w−z|=

The next, and final, step, is to take the limit as  → 0 of the right-hand side
of this equation, after first decomposing Fz (w) as

f (w) − f (z) 1
Fz (w) = + f (z) · ,
w−z w−z
27 7 CONSEQUENCES OF CAUCHY’S THEOREM

z0


z

Figure 5: The keyhole contour used in the proof of Cauchy’s integral formula.

Integrating each term separately, we have for the first term

I 
 f (w) − f (z)
 ≤ 2π · sup |f (w) − f (z)|

 dw

C w−z 
|w−z|= 
= 2π sup |f (w) − f (z)| −−→ 0,
|w−z|= →0

by continuity of f ; and for the second term,

I I
1 1
f (z) · dw = f (z) dw = 2πif (z)
|w−z|= w−z |w−z|= w−z

(by aHstandard calculation, see Example 2 above). Putting all together gives
1
that C 2πi Fz (w) dw = f (z), which was the formula to be proved.
Example 3. in the case when z is the center of the circle C = {w : |w − z| =
r}, Cauchy’s formula gives that
I Z 2π
1 dw 1
f (z) = f (w) = f (z + reit )dt.
2π |w−z|=r i(w − z) 2π 0

In other words, we have proved:

Theorem 11 (the mean value property for holomorphic functions). The value
of a holomorphic function f at z is equal to the average of its values around
a circle |w − z| = r (assuming it is holomorphic on an open set containing the
disc |w − z| ≤ r ).
28 7 CONSEQUENCES OF CAUCHY’S THEOREM

Considering what the mean value property means for the real and imag-
inary parts of f = u + iv , which are harmonic functions, we see that they in
turn also satisfy a similar mean value property:
Z 2π
1
u(x, y) = u(x + r cos t, y + r sin t) dt.
2π 0

This is in fact true for all harmonic functions — a fact, known as the mean
value property for harmonic functions, that can be proved separately us-
ing PDE/real analysis methods, or derived from the above considerations by
proving that every harmonic function in a disc is the real part of a holomor-
phic function.

Theorem 12 (Cauchy’s integral formula, extended version). Under the same

assumptions as in Theorem 10, f is infinitely differentiable, and for z ∈ D its
derivatives f (n) (z) are given by
I
(n) n! f (w)
f (z) = dw.
2πi C (w − z)n+1

The fact that holomorphic functions are differentiable infinitely many

times is referred to by [11] as the regularity theorem.

Proof. The key observation is that the expression on the right-hand side of
Cauchy’s integral formula for f (z) (which is the case n = 0 of the “extended”
version) can be differentiated under the integral sign. To make this precise,
let n ≥ 1, and assume inductively that we proved

(n − 1)!
I
(n−1) f (w)
f (z) = dw.
2πi C (w − z)n

Then

f (n−1) (z + h) − f (n−1) (z)

h  
(n − 1)!
I
1 1 1
= f (w) · − dw.
2πi C h (w − z − h)n (w − z)n
(w−z−h)−n −(w−z)−n
It is easily seen that as h → 0, the divided difference h
con-
−n−1
verges to n(w − z) , uniformly over w ∈ C . (The same claim without the
uniformity is just the rule for differentiation of a power function; to get the
uniformity one needs to “go back to basics” and repeat the elementary al-
gebraic calculation that was originally used to derive this power rule — an
29 7 CONSEQUENCES OF CAUCHY’S THEOREM

illustration of the idea that in mathematics it is important not just to under-

stand results but also the techniques used to derive them.) It follows that we
can go to the limit h → 0 in the above integral representation, to get

(n − 1)!
I
f (n)
(z) = f (w)n(w − z)−n−1 dz,
2πi C

which is precisely the nth case of Cauchy’s integral formula.

Proof of Morera’s theorem. We already proved that if f is a function all of

whose contour integrals over closed curves vanish, then f has a primitive F .
The regularity property now implies that the derivative F 0 = f is also holo-
morphic, hence f is holomorphic, which was the claim of Morera’s theorem.

As another immediate corollary to Cauchy’s integral formula, we now get

an extremely useful family of inequalities that bounds a function f (z) and its
derivatives at some specific point z ∈ C in terms of the values of the function
on the boundary of a circle centered at z .

Theorem 13 (Cauchy inequalities). For f holomorphic in a region Ω that

contains the closed disc DR (z), we have

|f (n) (z)| ≤ n!R−n sup |f (z)|

z∈∂DR (z)

(where ∂DR (z) refers to the circle of radius R around z ).

Theorem 14 (Analyticity of holomorphic functions). If f is holomorphic in

a region Ω that contains a closed disc DR (z0 ), then f has a power series
expansion at z0
∞
X
f (z) = an (z − z0 )n ,
n=0

that is convergent for all z ∈ DR (z0 ), where (of course) an = f (n) (z0 )/n!.

Proof. The idea is that Cauchy’s integral formula gives us a representation of

f (z) as a weighted “sum” (=an integral, which is a limit of sums) of functions
of the form z 7→ (w − z)−1 . Each such function has a power series expansion
since it is, more or less, a geometric series, so the sum also has a power
series expansion.
30 7 CONSEQUENCES OF CAUCHY’S THEOREM

To make this precise, write

1 1 1 1
= = ·  
w−z (w − z0 ) − (z − z0 ) w − z0 1 − z−z0
w−z0
∞  n ∞
1 X z − z0 X
= = (w − z0 )−n−1 (z − z0 )n .
w − z0 n=0 w − z0 n=0

This is a power series in z − z0 that, assuming w ∈ CR (z0 ), converges abso-

lutely for all z such that |z − z0 | < R (that is, for all z ∈ DR (z0 )). Moreover
the convergence is clearly uniform in w ∈ CR (z0 ). Since infinite summations
that are absolutely and uniformly convergent can be interchanged with in-
tegration operations, we then get, using the extended version of Cauchy’s
integral formula, that
I
1 f (w)
f (z) = dw
2πi CR (z0 ) w − z
I ∞
1 X
= f (w) (w − z0 )−n−1 (z − z0 )n dw
2πi CR (z0 ) n=0
∞  I 
X 1 n−1
= f (w)(w − z0 ) dw (z − z0 )n
n=0
2πi CR (z0 )
∞
X f (n) (z0 )
= (z − z0 )n ,
n=0
n!

which is precisely the expansion we were after.

Remark 4. In the above proof, if we only knew the simple (n = 0) case of

Cauchy’s integral formula (and in particular didn’t know the regularity the-
orem that follows from the extended case of this formula), we would still
conclude from the penultimate expression in the above chain of equalities
n
P
that f (z) hasR a power series expansion of the form n an (z − z0 ) , with
an = (2πi)−1 CR (z0 ) f (w)(w − z)−n−1 . It would then follow from earlier re-
sults we proved that f (z) is differentiable infinitely many times, and that
an = f (n) (z0 )/n!, which would again give the extended version of Cauchy’s
integral formula.

Theorem 15 (Liouville’s theorem). A bounded entire function is constant.

Proof. An easy application of the (case n = 1 of the) Cauchy inequalities

gives upon taking the limit R → ∞ that f 0 (z) = 0 for all z , hence, as we
already proved, f must be constant.
31 7 CONSEQUENCES OF CAUCHY’S THEOREM

Theorem 16. If f is holomorphic on a region Ω, and f = 0 for z in a set

containing a limit point in Ω, then f is identically zero on Ω.

Proof. If the limit point is z0 ∈ Ω, that means there is a sequence (wk )∞

k=0
of points in Ω such that f (wk ) = 0 for all n, wk → z , and wk 6= z0 for all
k . We know that in a neighborhood of z0 , f has a convergent power series
expansion. If we assume that f is not identically zero in a neighborhood of
z0 , then we can write the power series expansion as
X ∞
X
f (z) = k
an (z − z0 ) = an (z − z0 )∞
n=0 n=m
∞
X an+m
= am (z − z0 )m (z − z0 )n = am (z − z0 )m (1 + g(z)),
n=0
am
P∞ a
where m is the smallest index such that am 6= 0, and where g(z) = n=1 n+m am
(z−
z0 )n is a holomorphic function in the neighborhood of z0 that satisfies g(z0 ) =
0. It follows that for all k ,

am (wk − z0 )m (1 + g(wk )) = f (wk ) = 0,

but for large enough k this is impossible, since wk − z0 6= 0 for all k and
g(wk ) → g(z0 ) = 0 as k → ∞.
The conclusion is that f is identically zero at least in a neighborhood of
z0 . But now we claim that that also implies that f is identically zero on all of
Ω, because Ω is a region (open and connected). More precisely, denote by U
the set of points z ∈ Ω such that f is equal to 0 in a neighborhood of z . It is
obvious that U is open, hence also relatively open in Ω since Ω itself is open;
U is also closed, by the argument above; and U is nonempty (it contains z0 ,
again by what we showed above). It follows that U = Ω by the well-known
characterization of a connected topological space as a topological space that
has no “’clopen” (closed and open) sets other than the empty set and the
entire space.
An alternative way to finish the proof is the following. For every point
z ∈ Ω, let r(z) be the radius of convergence of the power series expansion
of f around z . Thus the discs {Dr(z) (z) : z ∈ Ω} form an open covering of
Ω. Take w ∈ Ω (with z0 being as above), and take a path γ : [a, b] → Ω con-
necting z0 and w (it exists because Ω is open and connected, hence pathwise-
connected). The open covering of Ω by discs is also an open covering of the
compact set γ[a, b] (the range of γ ). By the Heine-Borel property of compact
sets, it has a finite subcovering {Dr(zj ) (zj ) : j = 0, . . . , m} (where we take
w = zm+1 . The proof above shows that f is identically zero on Dr(z0 ) (z0 ), and
32 7 CONSEQUENCES OF CAUCHY’S THEOREM

also shows that if we know f is zero on Dr(zj ) (zj ) then we can conclude that
it is zero on the next disc Dr(zj+1 ) (zj+1 ). It follows that we can get all the way
to the last disc Dr(w) (w). In particular, f (w) = 0, as claimed.
Remark 5. The above result is also sometimes described under the heading
zeros of holomorphic functions are isolated, since it can be formulated
as the following statement: if f is holomorphic on Ω, is not identically zero on
Ω, and f (z0 ) = 0 for z0 ∈ Ω, then for some  > 0, the punctured neighborhood
D (z0 ) \ {z0 } of z0 contains no zeros of f . In other words, the set of zeros of f
contains only isolated points.
Remark 6. The condition that the limit point z0 be in Ω is needed. Note that
it is possible to have a sequence zn → z0 of points in Ω such that f (zn ) = 0
for all n. For example, consider the function e1/z − 1 — it has zeros in every
neighborhood of z0 = 0.
Corollary 7. If f, g are holomorphic on a region Ω, and f (z) = g(z) for z in a
set with limit point in Ω (e.g., an open disc, or even a sequence of points zn
converging to some z ∈ Ω), then f ≡ g everywhere in Ω.

Proof. Apply the previous result to f − g .

The previous result is usually reformulated slightly as the following con-

ceptually important result:
Theorem 17 (Principle of analytic continuation). If f is holomorphic on a
region Ω, and f+ is holomorphic on a bigger region Ω+ ⊃ Ω and satisfies
f (z) = f+ (z) for all z ∈ Ω, then f+ is the unique such extension, in the sense
that if f˜+ is another function with the same properties then f+ (z) = f˜+ (z) for
all z ∈ Ω+ .
Example 4. In real analysis, we learn that “formulas” such as
1
1 − 1 + 1 − 1 + 1 − 1 + ... = ,
2
1 + 2 + 4 + 8 + 16 + 32 + . . . = −1
don’t have any meaning. However, in the context of complex analysis one can
in fact make perfect sense of such identities, using the principle of analytic
continuation! Do you see how? We will also learn later in the course about
additional such amusing identities, the most famous of which being
1
1 + 2 + 3 + 4 + ... = − ,
12
1
1 − 2 + 3 − 4 + ... = .
4
33 7 CONSEQUENCES OF CAUCHY’S THEOREM

Such supposedly “astounding” formulas have attracted a lot of attention

in recent years, being the subject of a popular Numberphile video, a New
York Times article, a discussion on the popular math blog by Terry Tao, a
Wikipedia article, a discussion on Mathematics StackExchange, and more.

A “toy” (but still very interesting) example of analytic continuation is the

case of removable singularities. A point z0 ∈ Ω is called a removable singu-
larity of a function f : Ω → C∪{undefined} if f is holomorphic in a punctured
neighborhood of Ω, is not holomorphic at z0 , but its value at z0 can be rede-
fined so as to make it holomorphic at z0 .

Theorem 18 (Riemann’s removable singularities theorem). If f is holomor-

phic in Ω except at a point z0 ∈ Ω (where it may be undefined, or be defined
but not known to be holomorphic or even continuous). Assume that f is
bounded in a punctured neighborhood Dr (z0 ) \ {z0 } of z0 . Then f can be ex-
tended to a holomorphic function f˜ on all of Ω by defining (or redefining) its
value at z0 appropriately.

Proof. Fix some disc D = DR (z0 ) around z0 whose closure is contained in Ω.

The idea is to prove that the Cauchy integral representation formula
I
1 f (w)
f (z) = dw =: f˜(z)
2πi CR (z0 ) w−z

is satisfied for all z ∈ D \ {z0 }. Once we show this, we will set f˜(z0 ) to be
defined by the same integral representation, and it will be easy to see that
that gives the desired extension.
To prove that the representation above holds, consider a “double keyhole”
contour Γ,δ that surrounds most of circle C = ∂D but makes diversions to
avoid the points z0 and z , circling them in the negative direction around most
of a circle of radius . After applying a limiting argument similar to the one
used in the proof of Cauchy’s integral formula, we get that
I I I
1 f (w) 1 f (w) 1 f (w)
= + .
2πi C w−z 2πi C (z) w − z 2πi C (z0 ) w−z

On the right-hand side, the first term is f (z) by Cauchy’s integral formula
(since f is known to be holomorphic on an open set containing D (z)). The
second term can be bounded in magnitude using the assumption that f is
bounded in a neighborhood of z0 ; more precisely, we have
I 
 f (w)  1
 ≤ 2π sup |f (w)| · −−→ 0.

C (z0 ) w−z 
w∈C (z0 ) |z − z0 | −  →0
34 7 CONSEQUENCES OF CAUCHY’S THEOREM

Thus by taking the limit as  → 0 we obtain precisely the desired representa-

tion for f .
Finally, once we have the integral representation f˜ (defined only in terms
of the values of f (w) for w ∈ CR (z0 )), the fact that this defines a holomor-
phic function on all of DR (z0 ) is easy to see, and is something we implicitly
were aware of already. For example, the relevant argument (involving a
direct manipulation of the divided differences h1 (f˜(z + h) − f˜(z))) appeared
in the proof of the extended version of Cauchy’s integral formula. Another
approach is to show that integrating f˜ over closed contours gives 0 (which
requires interchanging the order of two integration operations, which will
not be hard to justify) and then use Morera’s theorem. The details are left
as an exercise.

Definition 3 (Uniform convergence on compact subsets). If f, (fn )∞ n=0 are

holomorphic functions on a region Ω, we say that the sequence fn converges
to f uniformly on compact subsets if for any compact set K ⊂ Ω, fn (z) →
f (z) uniformly on K .

Theorem 19. If fn → f uniformly on compact subsets in Ω and fn are holo-

morphic, then f is holomorphic, and fn0 → f 0 uniformly on compact subsets
in Ω.

Proof. The fact that f is holomorphic is an easy consequence of a combina-

tion of Cauchy’s theorem and Morera’s theorem. More precisely, note that
for each closed disc D = Dr (z0 ) ⊂ Ω we have fn (z) → f uniformly on D .
In particular, for each curve γ whose image is contained in the open disc
D = Dr (z0 ), Z Z
fn (z) dz −−−→ f (z) dz.
γ n→∞ γ
R
By Cauchy’s theorem, the integrals in this sequence are all 0, so γ f (z) dz
is also zero. Since this is true for all such γ , by Morera’s theorem f is holo-
morphic on D . This was true for any disc in Ω, and holomorphicity is a local
property, so in other words f is holomorphic on all of Ω.
Next, let D = Dr (z0 ) be a disc whose closure D satisfies D ⊂ Ω. for z ∈ D
we have by Cauchy’s integral formula that
I I
1 fn (w) 1 f (w)
fn0 (z) 0
− f (z) = dw − dw
2πi ∂D (w − z)2 2πi ∂D (w − z)2
fn (w) − f (w)
I
1
= dw.
2πi ∂D (w − z)2
35 8 ZEROS, POLES, AND THE RESIDUE THEOREM

It is easy to see therefore that fn0 (z) → f 0 (z) as n → ∞, uniformly as z ranges

on the disc Dr/2 (z0 ), since fn (w) → f (w) uniformly for w ∈ ∂D ⊂ D , and
|w − z|−2 ≤ (r/2)−2 for z ∈ Dr/2 (z0 ), w ∈ ∂D.
Finally, let K ⊂ D be compact. For each z ∈ K let r(z) be the radius of
a disc Dr(z) (z) around z whose closure is contained in Ω. The family of discs
{Dz = Dr(z)/2 (z) : z ∈ Ω} is an open covering of K , so by the Heine-Borel
property of compact sets it has a finite subcovering Dz1 , . . . , Dzn . We showed
that fn0 (z) → f 0 (z) uniformly on every Dzj , so we also have uniform conver-
gence on their union, which contains K , so we get that fn0 → f 0 uniformly on
K , as claimed.

8 Zeros, poles, and the residue theorem

Definition 4 (zeros). z0 is a zero of a holomorphic function f if f (z0 ) = 0.

Definition(+lemma) 5. If f is a holomorphic function on a region Ω that is

not identically zero and z0 is a zero of f , then f can be represented in the
form
f (z) = (z − z0 )m g(z)
in some neighborhood of z0 , where m ≥ 1 and g is a holomorphic function in
that neighborhood such that g(z) 6= 0. The number m is determined uniquely
and is called the order of the zero z0 , i.e, z0 will be described as a zero of
order m.
A zero of order 1 is called a simple zero.

Remark 7. In the case when z0 is not a zero of f , the same representation

holds with m = 0 (and g = f ), so in certain contexts one may occasionally say
that z0 is a zero of order 0.

Proof of the definition-lemma. Power series expansions: this is similar to the

argument used in the proof that zeros of holomorphic functions are isolated.
That is, write the power series expansion (known to converge in a neighbor-
hood of z0 )
∞
X ∞
X
n
f (z) = an (z − z0 ) = an (z − z0 )n
n=0 n=m
∞
X
= (z − z0 )m am+n (z − z0 )n =: (z − z0 )m g(z),
n=0
36 8 ZEROS, POLES, AND THE RESIDUE THEOREM

where m is the smallest index ≥ 0 such that am 6= 0. This gives the desired
representation. On the other hand, given a representation of this form, ex-
panding g(z) as a power series shows that m has to be the smallest index of
a nonzero coefficient in the power series expansion of f (z), which proves the
uniqueness claim.

Definition 6 (poles). If f is defined and holomorphic in a punctured neigh-

borhood of a point z0 , we say that it has a pole of order m at z0 if the function
h(z) = 1/f (z) (defined to be 0 at z0 ) has a zero of order m at z0 . A pole of
order 1 is called a simple pole.

Remark 8. As with the case of zeros, one can extend this definition in an
obvious way to define a notion of a “pole of order 0”. If f (z) is actually
holomorphic and nonzero at z0 (or has a removable singularity at z0 and can
be made holomorphic and nonzero by defining its value at z0 appropriately),
we define the order of the pole as 0 and consider f to have a pole of order 0
at z0 .

Lemma 4. f has a pole of order m at z0 if and only if it can be represented

in the form
f (z) = (z − z0 )−m g(z)
in a punctured neighborhood of z0 , where g is holomorphic in a neighborhood
of z0 and satisfies g(z0 ) 6= 0.

Proof. Apply the previous lemma to 1/f (z).

Theorem 20. If f has a pole of order m at z0 , then it can be represented

uniquely as
a−m a−m+1 a−1
f (z) = + + . . . + + G(z)
(z − z0 )m (z − z0 )m−1 z − z0

where G is holomorphic in a neighborhood of z0 .

Proof. This follows immediately on expressing f (z) as (z − z0 )−m g(z) as in the

previous lemma and separating the power series expansion of g(z) into the
powers (z − z0 )k with 0 ≤ k ≤ m − 1 and the powers with k ≥ m.
a−m a−m+1
Definition 7 (principal part, residue). The expansion + (z−z
(z−z0 )m 0)
m−1 +. . .+
a−1
z−z0
in the above representation is called the principal part of f at the pole
z0 . The coefficient a−1 is called the residue of f at z0 and denoted Resz0 (f ).
37 8 ZEROS, POLES, AND THE RESIDUE THEOREM

Exercise 7. The definitions of the order of a zero and a pole can be con-
sistently unified into a single definition of the (generalized) order of a zero,
where if f has a pole of order m at z0 then we say instead that f has a zero
of order −m. Denote the order of a zero of f at z0 by ordz0 (f ). With these
definitions, prove that

ordz0 (f + g) ≥ min (ordz0 (f ), ordz0 (g))

(can you give a useful condition when equality holds?), and that

ordz0 (f g) = ordz0 (f ) + ordz0 (g).

Theorem 21 (The residue theorem (simple version)). Assume that f is holo-

morphic in a region containing a closed disc D , except for a pole at z0 ∈ D .
Then I
f (z) dz = 2πi Resz0 (f ).
∂D

Proof. By the standard argument involving a keyhole contour, we see that

the circle C = ∂D in the integral can be replaced with a circle C = C (z0 ) of
a small radius  > 0 around z0 , that is , we have
I I
f (z) dz = f (z) dz.
∂D C

When  is small enough, inside C we can use the decomposition

−1
X
f (z) = ak (z − z0 )k + G(z)
k=−m

for f into its principal part and a remaining holomorphic part. Integrating
termwise gives 0 for the integral of G(z), by Cauchy’s theorem; 0 for the
integral powers (z − z0 )k with −m ≤ k ≤ −2, by a standard computation;
and 2πia−1 = 2πi Resz0 (f ) for the integral of r(z − z0 )−1 , by the same standard
computation. This gives the result.

Theorem 22 (The residue theorem (extended version)). Assume that f is

holomorphic in a region containing a closed disc D , except for a finite num-
ber of poles at z1 , . . . , zN ∈ D . Then
I N
X
f (z) dz = 2πi Reszk (f ).
∂D k=1
38 9 MEROMORPHIC FUNCTIONS AND THE RIEMANN SPHERE

Proof. The idea is the same, except one now uses a contour with multiple
keyholes to deduce after a limiting argument that
I N I
X
f (z) dz = f (z) dz
∂D k=1 C (zk )

for a small enough , and then proceeds as before.

(Note: The above argument seems slightly dishonest to me, since it re-
lies on the assertion that a multiple keyhole contour with arbitrary many
keyholes is a “toy contour”; while this is intuitively plausible, it will be un-
doubtedly quite difficult to think of, and write, a detailed proof of this argu-
ment.)

Theorem 23 (The residue theorem (version for general toy contours).). As-
sume that f is holomorphic in a region containing a toy contour γ (oriented
in the positive direction) and the region Rγ enclosed by it, except for poles
at the points z1 , . . . , zN ∈ Rγ . Then

I N
X
f (z) dz = 2πi Reszk (f ).
γ k=1

Proof. Again, construct a multiple keyhole version of the same contour γ

(assuming that one can believably argue that the resulting contour is still a
toy contour), and then use a limiting argument to conclude that
I N I
X
f (z) dz = f (z) dz,
γ k=1 C (zk )

for a small enough . Then proceed as before.

9 Meromorphic functions, holomorphicity at ∞

and the Riemann sphere
Definition 8 (meromorphic function). A meromorphic function on a region
Ω is a function f : Ω → C ∪ {undefined} such that f is holomorphic except
for an isolated set of poles.
39 9 MEROMORPHIC FUNCTIONS AND THE RIEMANN SPHERE

Definition 9 (holomorphicity at ∞). Let U ⊂ C be an open set containing

the complement C \ DR (0) of a closed disc around 0. A function f : U → C
is holomorphic at ∞ if g(z) = f (1/z) (defined on a neighborhood D1/R (0) of
0) has a removable singularity at 0. In that case we define f (∞) = g(0) (the
value that makes g holomorphic at 0).

Definition 10 (order of a zero/pole at ∞). Let U ⊂ C be an open set con-

taining the complement C \ DR (0) of a closed disc around 0. We say that a
function f : U → C has a zero (resp. pole) of order m at ∞ if g(z) = f (1/z)
has a zero (resp. pole) at z = 0, after appropriately defining the value of g at
0.

Conceptually, it is useful to think of meromorphic functions as holomor-

phic functions with range in the Riemann sphere C b . Let’s define what that
means.

Definition 11 (Riemann sphere). The Riemann sphere (a.k.a. the extended

b = C ∪ {∞}, equipped with the following
complex numbers) is the set C
additional structure:

• Topologically, we think of C
b as the one-point compactification of C; that
is, we add to C an additional element ∞ (called “the point at infinity”)
and say that the neighborhoods of ∞ are the complements of compact
sets in C. This turns C
b into a topological space in a simple way.

• Geometrically, we can identify C

b with an actual sphere embedded in
3
R , namely
n
2 o
S 2 = (x, y, z) ∈ R3 : x2 + y 2 + z − 21 = 12 .

The identification is via stereographic projection, given explicitly by the

formula
(
X Y
x + iy = + i 1−Z if (X, Y, Z) 6= (0, 0, 1),
(X, Y, Z) ∈ S 2 −
7 → 1−Z
∞ if (X, Y, Z) = (0, 0, 1).

See page 88 in [11] for a more detailed explanation. One can check that
this geometric identification is a homeomorphism between S 2 (equipped
with the obvious topology inherited from R3 ) and C
b (with the one-point
compactification topology defined above).

• Analytically, the above definition of what it means for a function on a

neighborhood of ∞ to be holomorphic at ∞ provides a way of giving C b
40 9 MEROMORPHIC FUNCTIONS AND THE RIEMANN SPHERE

the structure of a Riemann surface (the simplest case of a manifold

with a complex-analytic structure). The details can be found in many
textbooks and online resources, and we will not discuss them in this
course.

With the above definitions, the concept of a meromorphic function f :

Ω → C∪{undefined} can be seen to coincide with the notion of a holomorphic
function f : Ω → C
b — that is, the underlying concept of the definition is still
holomorphicity, but it pertains to functions taking values in C
b , a different
Riemann surface, instead of C.

Definition 12 (classification of singularities). If a function f : Ω → C b ∪

{undefined} is holomorphic in a punctured neighborhood Dr (z0 ) \ {z0 } of z0 ,
we say that f has a singularity at z0 if f is not holomorphic at z0 . We classify
singularities into three types, two of which we already defined:

• Removable singularities: when f can be made holomorphic at z0 by

defining or redefining its value at z0 ;

• poles;

• any singularity that is not removable or a pole is called an essential

singularity.

For a function defined on a neighborhood of ∞ that is not holomorphic

at ∞, we say that f has a singularity at ∞, and classify the singularity as
a removable singularity, a pole, or an essential singularity, according to the
type of singularity that z 7→ f (1/z) has at z = 0.

Theorem 24 (Casorati-Weierstrass theorem on essential singularities). If

f is holomorphic in a punctured neighborhood Dr (z0 ) \ {z0 } of z0 and has
an essential singularity at z0 , the image f (Dr (z0 ) \ {z0 }) of the punctured
neighborhood under f is dense in C.

Proof. Assume that the closure f (Dr (z0 ) \ {z0 }) does not contain a point w ∈
1
C. Then g(z) = f (z)−w is a function that’s holomorphic and bounded in Dr (z0 )\
{z0 }. By Riemann’s removable singularity theorem, its singularity at z0 is
removable, so we can assume it is holomorphic at z0 after defining its value
there. It then follows that
1
f (z) = w +
g(z)
has either a pole or a removable singularity at z0 , depending on whether
g(z0 ) = 0 or not.
41 10 THE ARGUMENT PRINCIPLE

10 The argument principle

Definition 13. The logarithmic derivative of a holomorphic function f (z)
is f 0 (z)/f (z).
Lemma 5. The logarithmic derivative of a product is the sum of the loga-
rithmic derivatives. That is,
0 n
( nk=1 fk ) fk0 (z)
Q X
Qn = .
k=1 fk k=1
f k (z)

Proof. Show this for n = 2 and proceed by induction.

Theorem 25 (the argument principle). Assume that f is meromorphic in a
region Ω containing a closed disc D , such that f has no poles on the circle
∂D. Denote its zeros and poles inside D by z1 , . . . , zn , where zk is a zero
of order mk = ordzk (f ) (in the sense mentioned above, where mk = m is a
positive integer if zk is a zero of order m, and mk = −m is a negative integer
if zk is a pole of order m). Then
n
f 0 (z)
I
1 X
dz = mk
2πi ∂D f (z) k=1
= [total number of zeros of f inside D]
− [total number of poles of f inside D]. (1)
Proof. Define
n
Y
g(z) = (z − zk )−mk f (z).
k=1
Then g(z) is meromorphic on Ω, has no singularities zeros on ∂D , and inside
D it has no poles or zeros, only removable singularities at z1 , . . . , zn (so after
redefining its values at these points we can assume it is holomorphic on D ).
It follows that
n
Y
f (z) = (z − zk )mk g(z).
k=1
Taking the logarithmic derivative of this equation gives that
n
f 0 (z) X mk g 0 (z)
= + .
f (z) k=1
z − zk g(z)
The result now follows by integrating this equation and using the residue
theorem (the term g 0 (z)/g(z) is holomorpic on D so does not contribute any-
thing to the integral).
42 10 THE ARGUMENT PRINCIPLE

Remark 9. By similar reasoning, the theorem also holds when the circle is
replaced by a toy contour γ .

The proof above hides, as some slick mathematical proofs have a way of
doing, the fact that the formula (1) has a fairly simple intuitive explanation.
Start by noting that the integral in the argument principle can be repre-
sented as
Z b 0 Z b
f 0 (z) f (γ(t))γ 0 (t) (f ◦ γ)0 (t)
I
1 1 1
dz = dt = dt
2πi γ f (z) 2πi a f (γ(t) 2πi a (f ◦ γ)(t)
Z
1 1
= dw,
2πi f ◦γ w

that is, an integral of dw/w over the contour f ◦ γ — the image of γ under f .
Now note that the differential form dw/w has a special geometric meaning
in complex analysis, namely we have

dw
= “d (log w) ” = “d (log |w| + i arg w) ”.
w
We put these expressions in quotes since the logarithm and argument are
not single-valued functions so it needs to be explained what such formulas
mean. However, at least log |w| is well-defined for a curve that does not cross
0, so when integrating over the closed curve f ◦ γ , the real part is zero by
the fundamental theorem of calculus. The imaginary part (which becomes
real after dividing by 2πi) can be interpreted intuitively as the change in the
argument over the curve — that is, initially at time t = a one fixes a specific
value of arg w = arg γ(a); then as t increases from t = a to t = b, one tracks the
increase or decrease in the argument as one travels along the curve γ(t); if
this is done correctly (i.e., in a continuous fashion), at the end the argument
must have a well-defined value. Since the curve is closed, the total change in
the argument must be an integer multiple of 2π , so the division by 2πi turns
it into an integer.
Of course, this explanation also explains the name “the argument princi-
ple,” which may sound arbitrary and uninformative when you first hear it.
Connection to winding numbers. What the above reasoning shows is
that in general, an integral of the form
I
1 f (w)
dw
2πi γ w
over a closed curve γ that does not cross 0 carries the meaning of “the total
number of times the curve γ goes around the origin,” with the number being
43 10 THE ARGUMENT PRINCIPLE

positive if the curve goes in the positive direction around the origin; negative
if the curve goes in the negative direction around the origin; or zero if there
is no net change in the argument. This number is more properly called the
winding number of f around w = 0 (also sometimes referred to as the
index of the curve around 0), and denoted
I
1 f (z)
Ind0 (f ) = dz.
2πi γ z

More generally, one can define the winding number at z = z0 as the number
of times a curve γ winds around an arbitrary point z0 , which (it is easy to
see) will be given by I
1 f (z)
Indz0 (f ) = dz,
2πi γ z − z0
assuming that γ does not cross z0 .
Note that winding number is a topological concept of planar geometry
that can be considered and studied without any reference to complex anal-
ysis; indeed, in my opinion that is the correct approach. It is possible, and
not especially difficult, to define it in purely topological terms without men-
tioning contour integrals, and then show that the complex analytic and topo-
logical definitions coincide. Try to think what such a definition might look
like.

Theorem 26 (Rouché’s theorem). Assume that f, g are holomorphic on a

region Ω containing a circle γ = C and its interior (or, more generally, a toy
contour γ and the region U enclosed by it). If |f (z)| > |g(z)| for all z ∈ γ then
f and f + g have the same number of zeros inside the region U .

Proof. Define ft (z) = f (z) + tg(z) for t ∈ [0, 1], and note that f0 = f and
f1 = f + g , and that the condition |f (z)| > |g(z)| on γ implies that ft has no
zeros on γ for any t ∈ [0, 1]. Denote

ft0 (z)
I
1
nt = dz,
2πi γ ft (z)

which by the argument principle is the number of “generalized zeros” (zeros

or poles, counting multiplicities) of ft in U . In particular, the function t 7→ nt
is integer-valued. If we also knew that it was continuous, then it would have
to be constant (by the easy exercise: any integer-valued continuous function
on an interval [a, b] is constant), so in particular we would get the desired
conclusion that n1 = n0 .
44 10 THE ARGUMENT PRINCIPLE

To prove continuity of nt , note that the function g(t, z) = ft0 (z)/ft (z) is
continuous, hence also uniformly continuous, on the compact set [0, 1] × γ .
For s, t ∈ [0, 1] satisfying |t − s| < δ , we can write
I
1
|nt − ns | ≤ |g(t, z) − g(s, z)| · |dz|
2πi γ
1
≤ len(γ) sup{|g(u, z) − g(v, z)| : z ∈ γ, u, v ∈ [0, 1], |u − v| < δ}.
2πi
Given  > 0, we can choose δ that ensures that this expression is <  if
|t − s| < δ , by the uniform continuity. This is precisely what is needed to
show that t 7→ nt is continuous.

As with the argument principle, Rouché’s theorem also has a rather amus-
ing intuitive explanation, which I learned from the book Visual Complex Anal-
ysis by Tristan Needham. The slogan to remember is “walking the dog”.
Imagine that you are walking in a large empty park containing at some “ori-
gin” point 0 a large pole (in the English sense of a metal post sticking out
of the ground, not the complex analysis sense). You start at some point X
and go for a walk along some curve, ending back at the same starting point
X . Let N denote your winding number around the pole at the origin — that
is, the total number of times you went around the pole, with the appropriate
sign.
Now imagine that you also have a dog that is walking alongside you in
some erratic path that is sometimes close to you, sometimes less close. As
you traverse your curve C1 , the dog walks along on its own curve C2 , which
also begins and ends in the same place. Let M denote the dog’s winding
number around the pole at the origin. Can we say that N = M ? The answer
is: yes, we can, provided that we know the dog’s distance to you was always
less than your distance to the pole. To see this, imagine that you had the dog
on a leash of variable length; if the distance condition was not satisfied, it
would be possible for the dog to reach the pole and go in a short tour around
it while you were still far away and not turning around the pole, causing an
entanglement of the leash with the pole.
Amazingly, the above scenario maps in a precise way to Rouché’s theo-
rem, using the following dictionary: the curve f ◦ γ represents your path; the
curve (f +g)◦γ represents the dog’s path; g ◦γ represents the vector pointing
from you to the dog; the condition |f | > |g| along γ is precisely the correct
condition that the dog stays closer to you than your distance to the pole; and
the conclusion that the two winding numbers are the same is precisely the
theorem’s assertion that f and f + g have the same number of generalized
45 11 APPLICATIONS OF ROUCHÉ’S THEOREM

zeros in the region U enclosed by γ (see the

H 0discussion above regarding the
−1
connection between the integral (2πi) γ
f /f dz and the winding number
of f ◦ γ around 0).

Exercise 8. Spend a few minutes thinking about the above correspondence

and make sure you understand it. You will probably forget the technical
details of the proof of Rouché’s theorem in a few weeks or months, but I
hope you will remember this intuitive explanation for a long time.

As another small cryptic remark to think about, the proof of Rouché’s

theorem given above can be thought of as an argument about the invariance
of a certain integral under the homotopy between two curves. Can you see
how?

11 Applications of Rouché’s theorem

Rouché’s theorem is an important tool in estimating the numbers of roots of
polynomials and other functions in regions of interests (see Exercises 28–29
on page 109). The next results show how Rouché’s theorem can also be used
to deduce interesting theoretical results.

Theorem 27 (the open mapping theorem). Holomorphic functions are open

mappings, that is, they map open sets to open sets.

Proof. Let f be holomorphic in a region Ω, z0 ∈ Ω, and denote w0 = f (z0 ).

What we need to show is that the image of any neighborhood D (z0 ) for  > 0
contains a neighborhood Dδ (z0 ) of w0 for some δ > 0. Fixing w (visualized as
being near w0 ), denote

h(z) = f (z) − w = (f (z) − w0 ) + (w0 − w) =: F (z) + G(z).

The idea is now to apply Rouché’s theorem to F (z) and G(z). Fix  > 0
small enough so that the disc D (z0 ) is contained in Ω and does not contain
solutions of the equation f (z) = w0 other than z0 (this is possible, by the
property that zeros of holomorphic functions are isolated). Defining

δ = inf{|f (z) − w0 | : z ∈ D (z0 )},

we therefore have that δ > 0 and |f (z) − w0 | ≥ δ for z on the circle |z − z0 | = .

That means that under the assumption that |w − w0 | < δ (i.e., if w is assumed
to be close enough to w0 ), the condition |F (z)| > |G(z)| in Rouché’s theorem
will be satisfied for z ∈ D (z0 ). The conclusion is that the equation h(z) = 0
46 12 SIMPLY-CONNECTED REGIONS AND CAUCHY’S THEOREM

(or equivalently f (z) = w) has the same number in solutions (in particular,
at least one solution) as the equation f (z) = w0 in the disc D (z0 ). This was
precisely the claim to be proved.

Corollary 8 (the maximum modulus principle). If f is a non-constant holo-

morphic function on a region Ω, then |f | cannot attain a maximum on Ω.

Proof. Trivial exercise.

At the beginning of the course we discussed several proofs of the fun-

damental theorem of algebra. One of them, the topological proof, was only
sketched. As a final demonstration of the power of Rouché’s theorem, prob-
lem 30 on page 109 asks you to use the theorem to make precise the idea of
the topological proof.

12 Simply-connected regions and the general

version of Cauchy’s theorem
Definition 14 (homotopy of curves). Given a region Ω ⊂ C, two parametrized
curves γ1 , γ2 : [0, 1] → Ω (assumed for simplicity of notation to be defined on
[0, 1]) are said to be homotopic (with fixed endpoints) if γ1 (0) = γ2 (0),
γ1 (1) = γ2 (1), and there exists a function F : [0, 1] × [0, 1] → Ω such that

i) F is continuous.

ii) F (0, t) = γ1 (t) for all t ∈ [0, 1].

iii) F (1, t) = γ2 (t) for all t ∈ [0, 1].

iv) F (s, 0) = γ1 (0) for all s ∈ [0, 1].

v) F (s, 1) = γ1 (1) for all s ∈ [0, 1].

The map F is called a homotopy between γ1 and γ2 . Intuitively, for each

s ∈ [0, 1] the function t 7→ F (s, t) defines a curve connecting the two endpoints
γ1 (0), γ1 (1). As s grows from 0 to 1, this family of curves transitions in a
continuous way between the curve γ1 and γ2 , with the endpoints being fixed
in place.

Exercise 9. Prove that the relation of being homotopic is an equivalence

relation.
47 12 SIMPLY-CONNECTED REGIONS AND CAUCHY’S THEOREM

Definition 15 (simply-connected region). A region Ω is called simply-connected

if any two curves γ1 , γ2 in Ω with the same endpoints are homotopic.

Remark 10. A common alternative way to define the notion of homotopy

of curves is for closed curves, where the endpoints are not fixed but the
homotopy must keep the curves closed as it is deforming them. The definition
of a simply-connected region then becomes a region in which any two closed
curves are homotopic. It is not hard to show that those two definitions are
equivalent.

Theorem 28. If f is a holomorphic function on a region Ω, and γ0 ,γ1 are two

curves on Ω with the same endpoints that are homotopic, then
Z Z
f (z) dz = f (z) dz.
γ0 γ1

Proof. This proof is based on the idea of translating the global statement
about the equality of the two contour integrals into a local statement, simi-
larly to the proof of Goursat’s theorem but in a more general setting. (See
also pages 93–95 in [11] for a variant of the proof presented below.)
Denote by F : [0, 1] × [0, 1] → Ω the homotopy between γ0 and γ1 , and for
any s ∈ [0, 1] denote by γs : [0, 1] → C the curve γs (t) = F (s, t). The strategy
of the proof is to show that there are values 0 = s0 < s1 < s2 < . . . < sn = 1
such that
Z Z Z Z
f (z) dz = f (z) dz = . . . = f (z) dz = f (z) dz.
γs0 γs1 γsn−1 γsn

That is, we will show that a slight perturbation of the s parameter does not
change the value of the integral. To this end, for two fixed values 0 ≤ s <
s0 ≤ 1 that are close to each other (in a sense we will make precise shortly),
we break up the t-interval [0, 1] over which the curves s, s0 are defined into
very small segments by fixing points 0 = t0 < t1 < . . . < tm = 1 that are very
close together (in a sense that, again, we will need to make precise below
once we understand what is needed to make the argument work), and then
write
Z m Z
X Z m Z
X
f (z) dz = f (z) dz, f (z) dz = f (z) dz.
γs j=1 γs ([tj−1 ,tj ]) γs0 j=1 γs0 ([tj−1 ,tj ])

We will now show that these two integrals are equal by exploiting our knowl-
edge of local properties of f that follow from its analyticity. Specifically,
assume that for each 1 ≤ j ≤ m we know that there exists an open disk Dj
48 12 SIMPLY-CONNECTED REGIONS AND CAUCHY’S THEOREM

containing the convex hull of the union of the two curve segments γs ([tj−1 , tj ])
and γs0 ([tj−1 , tj ]). For each 0 ≤ j ≤ m, let ηj denote a straight line segment
(considered as a parametrized curve) from γs (tj ) to γs0 (tj ), and for each 1 ≤
j ≤ m let Γj denote the closed curve γs ([tj−1 , tj ]) + ηj − γs0 ([tj−1 , tj ]) − ηj−1 (the
concatenation of the four curves γs ([tj−1 , tj ]), “the reverse of ηj ”, γs0 ([tj−1 , tj ]),
and “the reverse of ηj−1 ”). By Cauchy’s theorem on a disc, we have
I
f (z)dz = 0,
Γj

or, more explicitly,

Z Z Z Z
f (z) dz − f (z) dz = f (z) dz − f (z) dz.
γs ([tj−1 ,tj ]) γs0 ([tj−1 ,tj ]) ηj−1 ηj

Summing this relation over j , we get that

Z Z m Z Z !
X
f (z) dz − f (z) dz = f (z) dz − f (z) dz
γs γs0 j=1 γs ([tj−1 ,tj ]) γs0 ([tj−1 ,tj ])
m Z Z !
X
= f (z) dz − f (z) dz
j=1 ηj−1 ηj
Z Z
= f (z) dz − f (z) dz = 0,
η0 ηm

since in the next-to-last step the sum is telescoping, and in the last step
we note that η0 and ηm are both degenerate curves each of which simply
stays at a single point (γs (0) = F (s, 0) = γs0 (0), and γs (1) = F (s, 1) = γs0 (1),
respectively). This is precisely the equality we wanted.
We still need to justify the assumption about the discs Dj . This can be
made to work if s and s0 are sufficiently close to each other and the points
0 < t0 < t1 < . . . < tm = 1 are sufficiently densely spaced, using an argument
involving continuity and compactness. Here is one way to make the argu-
ment: fix 0 ≤ s ≤ 1. At each 0 ≤ t ≤ 1, by continuity of the homotopy function
F there exists a number δ > 0 and a disc Ds,t centered at γs (t) = F (s, t) such
that for any s0 , t0 ∈ [0, 1] with |s0 − s| < , |t0 − t| < δ , we have γs0 (t0 ) ∈ Ds,t . The
family of discs Ds,t for 0 ≤ t ≤ 1 are an open cover of the curve γs (or more
precisely of its image γs ([0, 1]), which is a compact set), so by the Heine-Borel
property we will have a finite subcovering Ds,t0 , . . . , Ds,tm . This is enough to
prove the assumption for s0 sufficiently close to s.
Finally, for each s denote by δ(s) the value of δ chosen above as a function
of s. The collection of open intervals {(s − δ(s), s + δ(s)) : s ∈ [0, 1]} form an
49 12 SIMPLY-CONNECTED REGIONS AND CAUCHY’S THEOREM

open covering of the interval [0, 1], so again using the Heine-Borel property,
we can extract a finite subcovering. This enables us to find a sequence 0 =
s0 < s1 < . . . < sn = 1 that we claimed exist at the beginning of the proof,
namely where the relation
Z Z
f (z) dz = f (z) dz
γsj−1 γsj

holds for each j = 1, . . . , n (with sj−1 playing the role of s and sj playing the
role of s0 in the discussion above).

Theorem 29 (Cauchy’s theorem (general version)). If f is holomorphic on a

simply-connected region Ω, then for any closed curve in Ω we have
I
f (z) dz = 0.
γ

Proof. Assume for simplicity that γ is parametrized as a curve on [0, 1]. Then
it can be thought of as the concatenation of two curves γ1 and −γ2 , where
γ1 = γ|[0,1/2] and γ2 is the “reverse” of the curve γ|[1/2,1] . Note that γ1 and γ2
have the same endpoints. By the invariance property of contour integrals
under homotopy proved above, we have
Z Z Z Z
f (z) dz = f (z) dz = f (z) dz − f (z) dz = 0.
γ γ1 −γ2 γ1 γ2

Corollary 9. Any holomorphic function on a simply-connected region has a

primitive.

Exercise 10. The proof of Theorem 28 above still involves a minor amount of
what I call “dishonesty”; that is, the proof is not actually formally correct as
written but contains certain inconsistencies between what the assumptions
of the theorem are and what we end up actually using in the body of the
proof. Can you identify those inconsistencies? What additional work might
be needed to fix these problems? And why do you think the author of these
notes, and the authors of the textbook [11], chose to present things in this
way rather than treat the subject in a completely rigorous manner devoid of
any inaccuracies? (The last question is a very general one about mathemati-
cal pedagogy; coming up with a good answer might help to demystify for you
a lot of similar decisions that textbook authors and course instructors make
in the teaching of advanced material, and make the study of such topics a bit
less confusing in the future.)
50 13 THE LOGARITHM FUNCTION

13 The logarithm function

The logarithm function can be defined as

log z = log |z| + i arg z

on any region Ω that does not contain 0 and where one can make a consistent,
smoothly varying choice of arg z as z ranges over Ω. It is easy to see that this
formula gives an inverse to the exponential function.
For example, if
Ω = C \ (−∞, 0]
(the “slit complex plane” with the negative real axis removed), we can set

Log z = log |z| + i Arg z

where Arg z is set to take values in (−π, π). This is called the principal branch
of the logarithm — basically a kind of standard version of the log function
that complex analysts have agreed to use whenever this is convenient (or
not too inconvenient). However, sometimes we may want to consider the
logarithm function on more strange or complicated regions. When can this
be made to work? The answer is: precisely when Ω is simply-connected.

Theorem 30. Assume that Ω is a simply-connected region with 0 ∈ / Ω, 1 ∈ Ω.

Then there exists a function F (z) = logΩ (z) with the properties:

i) F is holomorphic in Ω.

ii) eF (z) = z for all z ∈ Ω.

iii) F (r) = log r (the usual logarithm for real numbers) for all real numbers
r ∈ Ω sufficiently close to 1.

Proof. We define F as a primitive function of the function z 7→ 1/z , that is, as

Z z
dw
F (z) = ,
1 w
where the integral is computed along a curve γ connecting 1 to z . By the
general version of Cauchy’s theorem for simply-connected regions, this in-
tegral is independent of the choice of curve. As we have already seen, this
function is holomorphic and satisfies F 0 (z) = 1/z for all z ∈ Ω. It follows that

d
ze−F (z) = e−F (z) − zF 0 (z)e−F (z) = e−F (z) (1 − z/z) = 0,


dz
51 13 THE LOGARITHM FUNCTION

so ze−F (z) is a constant function. Since its value at z = 1 is 1, we see that

R r dw
F (z)
e = z , as required. Finally, for real r close to 1 we have that F (z) = 1 w =
log r, which can be seen by taking the integral to be along the straight line
segment connecting 1 and r .

Exercise 11. Prove that the principal branch of the logarithm has the Taylor
series expansion
∞
X (−1)n−1
Log z = (z − 1)n (|z| < 1).
n=1
n

Exercise 12. Modify the proof above to prove the existence of a branch of
the logarithm function in any simply-connected region Ω not containing 0,
without the assumption that 1 ∈ Ω. In what way is the conclusion weakened
in that case?

Exercise 13. Explain in what sense the logarithm functions F (z) = logΩ (z)
satisfying the properties proved in the theorem above (and its generalization
described in the previous exercise) are unique.

Exercise 14. Prove the following generalization of the logarithm construc-

tion above: if f is a holomorphic function on a simply-connected region Ω,
and f 6= 0 on Ω, then there exists a holomorphic function g on Ω, referred to
as a branch of the logarithm of f , satisfying

eg(z) = f (z).

Definition 16 (power functions and nth roots). On a simply-connected re-

gion Ω we can now define the power function z 7→ z α for an arbitrary α ∈ C
by setting
z α = eα logΩ z .
In the special case α = 1/n this has the meaning of the nth root function
z 7→ z 1/n , which satisfies
 1 n 1
1/n n logΩ z
(z ) = e n = en n logΩ z = elogΩ z = z.

Note that if f (z) = z 1/n is one choice of an nth root function, then for any 0 ≤
k ≤ n − 1, the function g(z) = e2πik/n f (z) will be another function satisfying
g(z)n = z . Conversely, it is easy to see that those are precisely the possible
choices for an nth root function.
52 14 THE EULER GAMMA FUNCTION

14 The Euler gamma function

The Euler gamma function (often referred to simply as the gamma function)
is one of the most important special functions in mathematics. It has ap-
plications to many areas, such as combinatorics, number theory, differential
equations, probability, and more, and is probably the most ubiquitous tran-
scendental function after the “elementary” transcendental functions (the ex-
ponential function, logarithms, trigonometric functions and their inverses)
that one learns about in calculus. It is a natural meromorphic function of a
complex variable that extends the factorial function to non-integer values.
In complex analysis it is particularly important in connection with the theory
of the Mellin transform (a version of the Fourier transform associated with
the multiplicative group of positive real numbers in the same way that the
ordinary Fourier transform is associated with the additive group of the real
numbers).
Most textbooks define the gamma function in one way and proceed to
prove several other equivalent representations of it. However, the truth is
that none of the representations of the gamma function is more fundamental
or “natural” than the others. So, it seems more logical to start by simply
listing the various formulas and properties associated with it, and then prov-
ing that the different representations are equivalent and that the claimed
properties hold.

Theorem 31 (the Euler gamma function). There exists a unique function

Γ(s) of a complex variable s that has the following properties:

1. Γ(s) is a meromorphic function on C.

2. Connection to factorials: Γ(n + 1) = n! for n = 0, 1, 2, . . ..

√
3. Important special value: Γ(1/2) = π.

4. Integral representation:
Z ∞
Γ(s) = e−x xs−1 dx (Re s > 0).
0

5. Hybrid series-integral representation:

∞ Z ∞
X (−1)n
Γ(s) = + e−x xs−1 dx (s ∈ C).
n=0
n!(n + s) 1
53 14 THE EULER GAMMA FUNCTION

6. Infinite product representation:

∞ 
−1 γs
Y s  −s/n
Γ(s) = se 1+ e (s ∈ C),
n=1
n

1

.
where γ = limn→∞ 1 + 2
+ 13 + . . . + 1
n
− log n = 0.577215 is the Euler-
Mascheroni constant.

7. Limit of finite products representation:

n! ns
Γ(s) = lim (s ∈ C).
n→∞ s(s + 1) · · · (s + n)

8. Zeros: the gamma function has no zeros (so Γ(s)−1 is an entire func-
tion).

9. Poles: the gamma function has poles precisely at the non-positive inte-
gers s = 0, −1, −2, . . ., and is holomorphic everywhere else. The pole at
s = −n is a simple pole with residue

(−1)n
Ress=−n (Γ) = (n = 0, 1, 2, . . .).
n!

Functional equation:

Γ(s + 1) = s Γ(s) (s ∈ C).

The reflection formula (a surprising connection to trigonometry):

π
Γ(s)Γ(1 − s) = (s ∈ C).
sin(πs)

To begin the proofs, we do have to define the function we are claiming

exists somehow, so let’s take the formula
Z ∞
Γ(s) = e−x xs−1 dx
0

as our working definition of Γ(s). This improper integral is easily seen to

converge absolutely for Re(s) > 0, since
∞ ∞ ∞
Z  Z Z
−x s−1 −x
e−x xRe(s)−1 dx.
  s−1

 e x dx ≤ e |x | dx =
0 0 0
54 14 THE EULER GAMMA FUNCTION

I leave it as an exercise to check (or read the easy explanation in [11]) that
the function it defines is holomorphic in that region.
Next, perform an integration by parts, to get that, again for Re(s) > 0, we
have
Z ∞ Z ∞
−x s x=∞
−e−x xs x=0 e−x sxs−1 dx = s Γ(s),

Γ(s + 1) = e x dx = +
0 0

which is the functional equation.

R∞
Combining the trivial evaluation Γ(1) = 0 e−x dx = 1 with the functional
equation shows by induction that Γ(n + 1) = n!.
√
The special value Γ(1/2) = π follows immediately by a change of vari-
Rable x = u2 in√the integral and an appeal to the standard Gaussian integral
∞ 2
−∞
e−u du = π :
Z ∞ Z ∞ Z ∞ √
−x −1/2 −u2 2
Γ(1/2) = e x dx = e 2 du = e−u du = π.
0 0 −∞

The functional equation can now be used to perform an analytic contin-

uation of Γ(s) to a meromorphic function on C: for example, we can define

Γ(s + 1)
Γ1 (s) =
,
s
which is a function that is holomorphic on Re(s) > −1, s 6= 0 and coincides
with γ(s) for Re(s) > 0. By the principle of analytic continuation this provides
a unique extension of Γ(s) to the region Re(s) > −1. Because of the factor
1/s and the fact that Γ(1) = 1 we also see that Γ1 (s) has a simple pole at s = 0
with residue 1.
Next, for Re(s) > −2 we define

Γ1 (s + 1) Γ(s + 2)
Γ2 (s) = = ,
s s(s + 1)

a function that is holomorphic on Re(s) > −2, s 6= 0, −1, and coincides with
Γ1 (s) for Re(s) > −1, s 6= 0. Again, this provides an analytic continuation of
Γ(s) to that region. The factors 1/s(s + 1) show that Γ2 (s) has a simple pole
at s = −1 with residue −1.
Continuing by induction, having defined an analytic continuation Γn−1 (s)
of Γ(s) to the region Re(s) > −n + 1, s 6= 0, −1, −2, . . . , −n + 2, we now define

Γn−1 (s + 1) Γ(s + n)
Γn (s) = = ... = .
s s(s + 1) · · · (s + n − 1)
55 14 THE EULER GAMMA FUNCTION

By inspection we see that this gives a meromorphic function in Re(s) > −n

whose poles are precisely at s = −n + 1, . . . , 0 and have the claimed residues.
An alternative way to perform the analytic continuation is to separate the
integral defining Γ(s) into
Z 1 Z ∞
−x s−1
Γ(s) = e x dx + e−x xs−1 dx
0 1

and to note that the integral over [1, ∞) converges (and defines a holomor-
phic function of s) for all s ∈ C, and the integral over [0, 1] can be computed
by expanding e−x as a power series in x and integrating term by term. That
is, for Re(s) > 0 we have

1 ∞
1X ∞ 1
(−1)n n+s−1 X (−1)n
Z Z Z
−x s−1
e x dx = x dx = xn+s−1 dx
0 0 n=0
n! n=0
n! 0
∞ n
X (−1)
=
n=0
n!(n + s)

The justification for interchanging the summation and integration operations

is easy and is left as an exercise. Thus, we have obtained not just an alterna-
tive proof for the meromorphic continuation of Γ(s), but a proof of the hybrid
series-integral representation of Γ(s), which also clearly shows where the
poles of Γ(s) are and that they are simple poles with the correct residues.

Lemma 6. For Re(s) > 0 we have

Z n  x n s−1
Γ(s) = lim 1− x dx.
n→∞ 0 n

Proof. As n → ∞, the
integrand converges to e−x xs−1 pointwise. Further-

n
more, the factor 1 − nx is bounded from above by the function e−x (because
of the elementary inequality 1 − t ≤ e−t that holds for all real t). The claim
therefore follows from the dominated convergence theorem.

Lemma 7. For Re(s) > 0 we have

n
n! ns
Z  x n s−1
1− x dx = .
0 n s(s + 1) · · · (s + n)
Proof. For n = 1, the claim is that
Z 1
1
(1 − x) xs−1 dx = ,
0 s(s + 1)
56 14 THE EULER GAMMA FUNCTION

which is easy to verify directly. For the general claim, using a linear change
of variables and an integration by parts we see that
Z n Z 1
 x n s−1
1− x dx = ns (1 − t)n ts−1 dt
0 n
0 s t=1 Z 1 s

s nt  n−1 t
= n (1 − t)  + n(1 − t) dt
s t=0 0 s
Z 1
s n
=n · (1 − t)n−1 t(s+1)−1 dt,
s 0
so the claim follows by induction on n.

Corollary 10. For Re(s) > 0 we have

n! ns
Γ(s) = lim .
n→∞ s(s + 1) · · · (s + n)

Proof of the infinite product representation for Γ(s). For Re(s) > 0 we have

s(s + 1) · · · (s + n)
Γ(s)−1 = lim
n→∞ n!ns
s  s  s
= s lim e−s log n 1 + 1+ ··· 1 +
n→∞ 1 2 n
n 
Pn 1
Y s −s/k

= s lim es( k=1 k −log n) 1+ e
n→∞
k=1
k
∞ 
Y s −s/n

= seγs 1+ e .
n=1
n

We now check that the infinite product actually converges absolutely and
uniformly on compact subsets in all of C, so defines an entire function. Let’s
start with some preliminary elementary observations on infinite products.

Q∞Recall that for a sequence of complex numbers cn , the infinite Qnproduct

n=1 cn is defined as the limit of finite (partial) products limn→∞ k=1 ck , if
the limit exists. This is analogous to the notation for infinite series. One
point of terminology that differs slightly from the corresponding terminology
Q∞
for infinite series, is that we say the product n=1 cn converges if the limit of
the finite products exists and is non-zero.

Lemma 8. For a sequence of complex numbers (an )∞ n=1 , if an 6= −1 for all n

P∞ Q∞
and n=1 |an | < ∞ then the infinite product n=1 (1 + an ) converges.
57 14 THE EULER GAMMA FUNCTION

Proof. Under the assumption, there exists some large enough N0 ≥ 1 such
that |an | < 1/2 for all n ≥ N0 , which in particular implies that 1 + an =
exp(Log(1 + an )), where Log(z) is the principal branch of the logarithm func-
tion. Now, by the assumption, clearly it is enough to prove the convergence
Q∞
of the infinite product n=N0 (1 + an ), and this can be written as
∞
Y n
Y n
Y
(1 + an ) = lim (1 + ak ) = lim exp (Log(1 + ak ))
n→∞ n→∞
n=N0 k=N0 k=N0
n
!
X
= lim exp Log(1 + ak ) .
n→∞
k=N0
P∞ Pn
If we knew that the infinite series n=N0 Log(1 + an ) = limn→∞ k=N0 Log(1 +
ak ) converged, we could continue the above chain of equalities as
n
! ∞
!
X X
= exp lim Log(1 + ak ) = exp Log(1 + an ) ,
n→∞
k=N0 n=N0
Q∞
and since ez is never 0, that would mean that the product n=N0 (1+an ) exists
and is non-zero (that is, it converges), hence as was mentioned above the
Q∞
infinite product n=1 (1 + an ) also converges and the claim would be proved.
P∞
To see the convergence of the series n=N0 Log(1 + an ), recall that the
function z 7→ Log(z) has the convergent Taylor expansion
∞
X (−1)m−1
Log(z) = (z − 1)m (|z − 1| < 1).
m=1
m

In particular, there is some constant C > 0 such that

| Log(1 + w)| ≤ |w| + C|w|2 if |w| < 1/2.

P∞
Now using the assumption that n=1 |an | < ∞ it follows that

∞ ∞ ∞ ∞ ∞
!2
X X X X X
| Log(1 + an )| ≤ |an | + C |an |2 ≤ |an | + C |an | < ∞,
n=N0 n=N0 n=N0 n=N0 n=N0

which is what we needed.

Lemma 9. Let (fn )∞ n=1 be a sequence of functions on some region Ω such

that the functions 1 + fn are everywhere holomorphic and nonzero. If the
P∞
series
Q∞ n=1 |fn | converges uniformly on compacts in Ω, the infinite product
n=1 (1 + fn ) also converges uniformly on compact subsets in Ω to a nonzero
holomorphic function.
58 14 THE EULER GAMMA FUNCTION

Q∞
Proof. Lemma 8 above implies that the infinite product n=1 (1 + fn (z)) con-
verges to a nonzero limit for any z ∈ Ω. By repeating the same estimates
in the proof of that lemma in the context of z being allowed to range on
a compact subset K ⊂ Ω, one sees that the sequence of partial products
Qn
k=1 (1 + fn ) actually converges uniformly on compacts, so the limiting func-
tion is holomorphic.
Q∞ z
e−z/n is an entire function.


Proof that n=1 1+ n

∞  ∞    2  
X z  −z/n  X  z z z 
 1+ e − 1 = 1+ 1− +O − 1
  
n  n n n 2
n=1 n=1
∞   2 
X  z 
=  O <∞
n=1
 n2 

(where the big-O notation hides a universal constant — the dependence on

z is encapsulated in the z 2 factor). In particular, the convergence is uniform
on compacts on C. So we are almost in the setting of Lemma 9, except that
in order to apply that result, which requires the functions participating in
the product to be nonzero, one needs to be a bit more careful and separate
out the zeros: for a fixed disc DN +1/2 (0) of radius N + 1/2 around 0, consider
only the product starting at n = N + 1 — those functions are nonzero in
the disc so the previous result applies to give a function that’s holomorphic
and nonzero in DN (0). Then separately the factors (1 + z/n), n = 1, . . . , N
contribute simple zeros at z = −1, . . . , −N .
π
Corollary 11 (the reflection formula). Γ(s)Γ(1 − s) = .
sin πs
Proof.
1
= Γ(s)−1 (−s)−1 Γ(−s)−1
Γ(s)Γ(1 − s)
∞  ∞ 
−1 γs
Y s  −s/n −γs
Y s  s/n
= · se 1+ e · (−s)e 1− e
s n=1
n n=1
n
∞  2

Y s sin(πs) sin(πs)
=s 1− 2 =s = ,
n=1
n πs π
Q∞
where we used the product representation sin(πz) = πz n=1 (1 − z 2 /n2 ) for
the sine function derived in a homework problem.
59 15 THE RIEMANN ZETA FUNCTION

Alternative derivation of the reflection formula ([11], page 164). By analytic

continuation, it is enough to prove the formula for real s in (0, 1). For such s
we have
Z ∞
Γ(s)Γ(1 − s) = e−t t−s Γ(s) dt
Z0 ∞  Z ∞ 
−t −s −vt s−1
= e t t e (vt) dv dt
0 0
Z ∞Z ∞ Z ∞ Z ∞ 
−t(1+v) s−1 −t(1+v)
= e v dv dt = e dt v s−1 dv
Z0 ∞ 0s−1 Z ∞ sx
0 0
v e
= dv = x
dx (by setting v = ex ).
0 1 + v −∞ 1 + e
So it is enough to prove that for 0 < s < 1 we have
∞
esx
Z
π
x
dx = .
−∞ 1+e sin(πs)
This integral can be evaluated using residue calculus; see Example 2 in Sec-
tion 2.1, Chapter 3, pages 79–81 of [11] for the details.

Note that by combining the alternative derivation of the reflection for-

mula given above with the infinite product representation for the gamma
function, we get a new proof of the infinite product representation for sin(πz).

15 The Riemann zeta function

15.1 Definition and basic properties

The Riemann zeta function (often referred to simply as the zeta function
when there is no risk of confusion), like the gamma function is considered
one of the most important special functions in “higher” mathematics. How-
ever, the Riemann zeta function is a lot more mysterious than the gamma
function, and remains the subject of many famous open problems, including
the most famous of them all: the Riemann hypothesis, considered by many
(including myself) as the most important open problem in mathematics.
The main reason for the zeta function’s importance is its connection with
prime numbers and other concepts and quantities from number theory. Its
study, and in particular the attempts to prove the Riemann hypothesis, have
also stimulated an unusually large number of important developments in
many areas of mathematics.
60 15 THE RIEMANN ZETA FUNCTION

As with the gamma function, the Riemann zeta function is usually defined
on only part of the complex plane and its definition is then extended by an-
alytic continuation. Again, I will formulate this as a theorem asserting the
existence of the zeta function and its various properties.
Theorem 32 (Riemann zeta function). There exists a unique function, de-
noted ζ(s), of a complex variable s, having the following properties:

1. ζ(s) is a meromorphic function on C.

2. For Re(s) > 1, ζ(s) is given by the series

∞
X 1 1 1
ζ(s) = s
= 1 + s + s + ....
n=1
n 2 3

3. Euler product formula: for Re(s) > 1, ζ(s) also has an infinite product
representation
Y 1
ζ(s) = ,
p
1 − p−s
where the product ranges over the prime numbers p = 2, 3, 5, 7, 11, . . ..

4. ζ(s) has no zeros in the region Re(s) > 1.

5. ζ(s) has no zeros on the line Re(s) = 1 (this requires a separate proof
from the previous claim).

6. The “trivial” zeros: the zeros of ζ(s) in the region Re(s) ≤ 0 are pre-
cisely at s = −2, −4, −6, . . ..

7. ζ(s) has a unique pole located at s = 1. It is a simple pole with residue

1.
8. The “Basel problem” and its generalizations: the values of ζ(s) at
even positive integers are given by Euler’s formula
(−1)n−1 (2π)2n
ζ(2n) = B2n (n = 1, 2, . . .),
2(2n)!
where (Bm )∞m=0 are the Bernoulli numbers, defined as the coefficients
in the Taylor expansion
∞
z X Bm m
z
= z .
e − 1 m=0 m!
Many of the properties of these amazing numbers were discussed in
our homework problem sets.
61 15 THE RIEMANN ZETA FUNCTION

9. Values at negative integers: we have

Bn+1
ζ(−n) = − (n = 1, 2, 3, . . .).
n+1
(Note that for negative even integers, this coincides with the property
stated above about the trivial zeros at s = −2, −4, −6, . . ., since it is an
easy fact that the Bernoulli numbers satisfy B2k+1 = 0 for integer k ≥ 1.
But this formula adds information about the values of ζ(s) at negative
odd integers.)

10. Functional equation: the zeta function satisfies

ζ ∗ (1 − s) = ζ ∗ (s),

where we denote by ζ ∗ (s) the symmetrized zeta function

s
ζ ∗ (s) = π −s/2 Γ ζ(s).
2

11. Mellin transform representation: an expression for ζ(s) valid for all
s ∈ C is
s
π −s/2 Γ ζ(s)
2 Z ∞
1 1 1  s+1 s−2

=− − + t− 2 +t 2 (ϑ(t) − 1) dt,
1−s s 2 1

where the function ϑ(t) is one of Jacobi theta series, defined as

∞ ∞
−πn2 t 2
X X
ϑ(t) = e =1+2 e−πn t .
n=−∞ n=1

12. Contour integral representation: another expression for ζ(s) valid

for all s ∈ C is
s
Γ(1 − s)
Z
(−x) dx
ζ(s) = ,
2πi C ex − 1 x
where C is a keyhole contour coming from +∞ to 0 slightly above the
positive x-axis, then circling the origin in a counterclockwise direction
around a circle of small radius, then going back to +∞ slightly below
the positive x-axis.
62 15 THE RIEMANN ZETA FUNCTION

13. Connection to prime number enumeration — the “explicit for-

mula of number theory”: define Von Mangoldt’s weighted prime
counting function X
ψ(x) = log p,
pk ≤x

where the sum is over all prime powers less than or equal to x. Then
for non-integer x > 1,
X xρ
ψ(x) = x − − log(2π),
ρ
ρ

where the sum ranges over all zeros ρ of the Riemann zeta function
counted with their respective multiplicities. (In most textbooks the sum
is separated into two sums, one ranging over the trivial zeros which can
be evaluated explicitly, and the other ranging over the much less trivial
zeros in the strip 0 < Re(s) < 1. Also the sum is only conditionally
convergent; refer to a book on analytic number theory for the proper
way to interpret it to get a convergent sum.)

The explicit formula of number theory illustrates that knowing where the
zeros of ζ(s) are has important consequences for prime number enumeration.
In particular, proving that Re(s) has no zeros in Re(s) ≥ 1 will enable us to
prove one of the most famous theorems in mathematics.

Theorem 33 (Prime number theorem). Let π(x) denote the number of prime
numbers less than or equal to x. Then we have

π(x)
lim = 1.
x→∞ x/ log x

Conjecture 1 (The Riemann hypothesis). All the nontrivial zeros of ζ(s) are
on the “critical line” Re(s) = 1/2.

To begin the proof of Theorem 32, again, let’s take as the definition of
ζ(s) the standard representation
∞
X 1
ζ(s) = .
n=1
ns
−s − Re(s)
P P
Since n |n | = nn , we see that the series converges absolutely
precisely when Re(s) > 1, and that the convergence is uniform on any half-
plane of the form Re(s) > α where α > 1. In particular, it is uniform on
compact subsets, so ζ(s) is holomorphic in this region.
63 15 THE RIEMANN ZETA FUNCTION

− p−s )−1 converges absolutely

Q
Similarly, the Euler product Z(s) := p (1
−s − Re(s)
P P
if and only if the series p |p | = pp converges, and in particular if
Re(s) > 1. It follows that Z(s) is well-defined, holomorphic and nonzero for
Re(s) > 1.

Proof of the Euler product formula. We now prove that Z(s) = ζ(s). This can
be done by manipulating the partial products associated with the infinite
product defining Z(s), as follows:
Y 1 Y
ζN (s) := = (1 + p−s + p−2s + p−3s + . . .)
p≤N
1 − p−s p≤N
X 1
= ,
j j
ns
n=p11 ···pkk
p1 ,...,pk primes ≤N

where the last equality follows from the fundamental theorem of arithmetic,
together with the fact that when multiplying two (or a finite number of) in-
finitely convergent series, the summands can be rearranged and summed in
any order we desire. So, we have represented ζN (s) as a series of a simi-
lar form as the series defining ζ(s), but involving terms of the form n−s only
for those positive integers n whose prime factorization contains only primes
≤ N . It follows that
X 1
|ζ(s) − ζN (s)| ≤ s
.
n>N
n

Taking the limit as N → ∞ shows that Z(s) = limN →∞ ζN (s) = ζ(s). This
proves the validity of the Euler product formula.

Corollary 12. ζ(s) has no zeros in the region Re(s) > 1.

Proof. The Euler product formula gives a convergent product for ζ(s) in this
region where each factor (1 − p−s )−1 has no zeros.

To prove the functional equation, we need a somewhat powerful tool from

harmonic analysis, the Poisson summation formula.

Theorem 34 (the Poisson summation formula). For a sufficiently well-behaved

function f : R → C, we have
∞
X ∞
X
f (n) = fˆ(k),
n=−∞ k=−∞
64 15 THE RIEMANN ZETA FUNCTION

where Z ∞
fˆ(k) = f (x)e−2πikx dx.
−∞
is the Fourier transform of f .

Proof. Define a function g : [0, 1] → C by

∞
X
g(x) = f (x + n),
n=−∞

the “periodiciziation” of f . Assume that f (x) is sufficiently well-behaved

(i.e., decays fast enough as x → ±∞ so that g(x) is in turn well-behaved, and
has reasonable smoothness properties). In that case, a standard result from
harmonic analysis states that g(x) will have a convergent Fourier expansion
of the form
∞
X
g(x) = ĝ(k)e2πikx ,
k=−∞

where the Fourier coefficients ĝ(k) can be computed as

Z 1
ĝ(k) = g(x)e−2πikx dx.
0

In particular, setting x = 0 in the formula for g(x) gives the basic fact that
∞
X
g(0) = ĝ(k).
k=−∞
P∞
However, note that g(0) = n=−∞ f (n), the quantity on the left-hand side of
the Poisson summation formula. On the other hand, the Fourier coefficient
ĝ(k) can be expressed in terms of the Fourier coefficients of the original
function f (x):
Z 1 Z 1 ∞
X
−2πikx
ĝ(k) = g(x)e dx = f (x + n)e−2πikx dx
0 0 n=−∞
∞
X Z 1 ∞ Z
X n+1
−2πikx
= f (x + n)e dx = f (u)e−2πiku du
n=−∞ 0 n=−∞ n
Z ∞
= f (u)e−2πiku du = fˆ(k).
−∞

Combining these observations gives the result, modulo a few details we’ve
glossed over concerning the precise assumptions that need to be made about
f (x) (we will only apply the Poisson summation formula for one extremely
well-behaved function, so I will not bother discussing those details).
65 15 THE RIEMANN ZETA FUNCTION

Theorem 35. The Jacobi theta function ϑ(t) satisfies the functional equation

1
ϑ(t) = √ ϑ(1/t) (t > 0).
t

Remark 11. Equations of this form are studied in the theory of modular
forms, an area of mathematics combining number theory, complex analysis
and algebra in a very surprising and beautiful way.

Proof. The idea is to apply the Poisson summation formula to the function
2
f (x) = e−πtx ,

for which it can be checked that

2
fˆ(k) = t−1/2 e−πk /t ,

using a simple change of variables from the standard integral evaluation

Z ∞
2 2
e−πx e−2πixu du = e−πu
−∞

2
(that is, the fact that the function e−πx is its own Fourier transform); this
evaluation appears in Example 1, Chapter 2, pages 42–44 in [11]. With
the above substitution for f (x) and fˆ(k), the Poisson summation formula be-
comes precisely the functional equation for ϑ(t).

Exercise 15. (a) Use the residue theorem to evaluate the contour integral
2
e−πz t
I
dz,
γN e2πiz − 1

where γN is the rectangle with vertices ±(N + 1/2) ± i (with N a positive

integer), then take the limit as N → ∞ to derive the integral representation
∞−i 2 ∞+i 2
e−πz t e−πz t
Z Z
ϑ(t) = dz − dz
−∞−i e2πiz − 1 −∞+i e2πiz − 1
for the Jacobi theta function.
(b) In this representation, expand the factor (e2πiz −1)−1 as a geometric series
in e−2πiz (for the first integral) and as a geometric series in e2πiz (for the
second integral). Evaluate the resulting infinite series, rigorously justifying
all steps, to obtain an alternative proof of the functional equation for ϑ(t).
66 15 THE RIEMANN ZETA FUNCTION

Lemma 10. The asymptotic behavior of ϑ(t) near t = 0 and t = +∞ is given

by

1
ϑ(t) = O √ (t → 0+),
t
ϑ(t) = 1 + O(e−πt ) (t → ∞).
Proof. The asymptotics as t → ∞ is immediate from
∞ ∞
X
−πn2 t
X 2e−πt
ϑ(t) − 1 = 2 e ≤2 e−πnt = ,
n=1 n=1
1 − e−πt
−πt
which is bounded by Ce if t > 10. Using the functional equation now gives
−1/2 −π/t
that ϑ(t) = t (1 + O(e )) = O(t−1/2 ) as t → 0+.

Proof of the analytic continuation of ζ(s). Start with the formula

s Z ∞
Γ = e−x xs/2−1 dx,
2 0

valid for Re(s) > 0. A linear change of variable x = πn2 t brings this to the
form s Z ∞
2
π −s/2 Γ n−s = e−πn t ts/2−1 dt.
2 0
Summing the left-hand side over n = 1, 2, . . . gives π −s/2 Γ 2s ζ(s) — the func-

tion we denoted ζ ∗ (s) — adding the stronger assumption that Re(s) > 1. For
the right-hand side we have that
∞ Z ∞
!
∞ Z ∞
2 2t
X X
e−πn t ts/2−1 dt. = e−πn ts/2−1 dt
n=1 0 0 n=1
∞
ϑ(t) − 1 s/2−1
Z
= t dt,
0 2
where the estimates in the lemma are needed to justify interchanging the or-
der of the summation and integration, and show that the integral converges
for Re(s) > 1. Thus we have obtained the representation
Z ∞ Z ∞
∗ 1 s/2−1
ζ (s) = (ϑ(t) − 1)t dt = ϕ(t)ts/2−1 dt,
2 0 0
1
where we denote ϕ(t) = − 1). Next, the idea is to use the functional
2
(ϑ(t)
equation for ϑ(t) to bring this to a new form that can be seen to be well-
defined for all s ∈ C except s = 1. Specifically, we note that the functional
equation for can be expressed in the form
ϕ(t) = t−1/2 ϕ(1/t) + 12 t−1/2 − 12 .
67 15 THE RIEMANN ZETA FUNCTION

We can therefore write

Z 1 Z ∞
∗ s/2−1
ζ (s) = ϕ(t)t dt + ϕ(t)ts/2−1 dt
Z0 1 1
Z ∞
−1/2 1 −1/2 s/2−1
ϕ(t)ts/2−1 dt
1


= t
ϕ(1/t) + − t dt +
2
t 2
0 1
Z ∞
1 1
t−s/2−1/2 + ts/2−1 ϕ(t) dt.


=− − +
1−s s 1

We have derived a formula for ζ ∗ (s) (one of the formulas claimed in the main
theorem above) that is now seen to define a meromorphic function on all of C
— the integrand decays rapidly as t → ∞ so actually defines an entire func-
tion, so the only poles are due to the two terms −1/s and 1/(s − 1). We have
therefore proved that ζ(s) can be analytically continued to a meromorphic
function on C.

Corollary 13. The zeta function satisfies the functional equation

ζ ∗ (1 − s) = ζ ∗ (s).

Equivalently, because of the reflection formula satisfied by the gamma func-

tion, it is easy to check that the functional equation can be rewritten in the
form   πs
ζ(s) = 2s π s−1 sin Γ(1 − s)ζ(1 − s).
2
Proof. The representation we derived for ζ ∗ (s) is manifestly symmetric with
respect to replacing each occurrence of s by 1 − s.

Corollary 14. The only pole of ζ(s) is a simple pole at s = 1 with residue 1.

Proof. Our representation for ζ ∗ (s) expresses it as a sum of − 1s , s−1

1
, and
∗
an entire function. Thus the poles of ζ (s) are simple poles at s = 0, 1 with
residues −1 and 1, respectively. It follows that

ζ(s) = π s/2 Γ(s/2)−1 ζ ∗ (s)

has a pole at s = 1 with residue π 1/2 Γ(1/2)−1 = 1, and a pole (that turns out
to be a removable singularity) at s = 0 with residue π 0 Γ(0)−1 = 0. (That is,
the pole of ζ ∗ (s) at s = 0 is cancelled out by the zero of Γ(s/2).)

Corollary 15. ζ(−n) = −Bn+1 /(n + 1) for n = 1, 2, 3, . . ..

68 15 THE RIEMANN ZETA FUNCTION

Proof. Using the functional equation, we have that

ζ(−n) = 2−n π −n−1 sin(−πn/2)Γ(n + 1)ζ(n + 1)

= 2−n π −n−1 sin(−πn/2)n!ζ(n + 1).

If n = 2k is even, then sin(−πn/2) = sin(−πk) = 0, so we get that ζ(−2k) = 0

(that is, n = 2k is one of the so-called “trivial zeros”). We also know that
B2k+1 = 0 for k = 1, 2, 3, . . ., so the formula ζ(−n) = Bn+1 /(n + 1) is satisfied
in this case.
If on the other hand n = 2k − 1 is odd, then sin(−π(2k − 1)/2) = (−1)k ,
and therefore we get, using the formula expressing ζ(2k) in terms of the
Bernoulli numbers (derived in the homework and in the textbook), that

ζ(−n) = (−1)k 2−2k+1 π −2k (2k − 1)!ζ(2k)

(−1)k−1 (2π)2k
= (−1)k 2−2k+1 π −2k (2k − 1)! B2k
2(2k)!
B2k Bn+1
=− =− ,
2k n+1
so again the formula is satisfied.

Corollary 16. The zeros of ζ(s) in the region Re(s) < 0 are precisely the
trivial zeros s = −2, −4, −6, . . ..

Proof. We already established the existence of the trivial zeros. The fact that
there are no other zeros also follows easily from the functional equation and
is left as an exercise.

Remark 12 (alternative approaches to the analytic continuation of ζ(s)).

There is a more “down-to-earth” approach to the analytic continuation of
ζ(s) based on the standard idea from numerical analysis of approximating an
integral by a sum (or in this case going in the other direction, approximating
a sum by an integral). The technical name for this procedure, when it is done
in a more systematic way, is Euler-Maclaurin summation.
∞ ∞ Z n+1  Z n+1 
X 1 X dx 1 dx
ζ(s) = = + −
n=1
ns n=1 n xs ns n xs
Z ∞ ∞ Z
dx X n+1 1
 
1
= + − dx
1 xs n=1 n ns xs
Z ∞
1
x−s − bxc−s dx.


= −
s−1 1
69 15 THE RIEMANN ZETA FUNCTION

This representation is certainly valid for Re(s) > 1. However, note that we
have the bound
 −s
x − bxc−s  ≤ |s| · bxc− Re(s)−1

(x ≥ 1)

by the mean value theorem. Thus, the integral is actually an absolutely con-
vergent integral in the larger region Re(s) > 0, and the representation we
derived gives an analytic continuation of ζ(s) to a meromorphic function on
Re(s) > 0, which has a single pole at s = 1 (a simple pole with residue 1) and
is holomorphic everywhere else.
An elaboration of this idea using what is known as the Euler-Maclaurin
summation formula can be used to perform the analytic continuation of ζ(s)
to a meromorphic function on C by extending it inductively from each region
Re(s) > −n to Re(s) > −n − 1, as we saw could be done for the gamma
function. Another approach is to use the analytic continuation for Re(s) > 0
shown above, then prove that the functional equation ζ ( 1 − s) = ζ ∗ (s) holds in
the region 0 < Re(s) < 1, and then use the functional equation to analytically
continue ζ(s) to Re(s) ≤ 0 (which is the reflection of the region Re(s) ≥ 1
under the transformation s 7→ 1 − s).

15.2 A theorem on the zeros of the Riemann zeta func-

tion
Next, we prove a nontrivial and very important fact about the zeta function
that will play a critical role in our proof of the prime number theorem.

Theorem 36. ζ(s) has no zeros on the line Re(s) = 1.

This theorem can also be thought of as a “toy” version of the Riemann

hypothesis. If you ever want to try solving this famous open problem, getting
a good understanding of its toy version seems like a good idea...

Proof. For this proof, denote s = σ + it, where we assume σ > 1 and t is
real and nonzero. The proof is based on investigating simultaneously the
behavior of ζ(σ + it), ζ(σ + 2it), and ζ(σ), for fixed t as σ & 1. Consider the
following somewhat mysterious quantity

X = log |ζ(σ)3 ζ(σ + it)4 ζ(σ + 2it)|.

70 15 THE RIEMANN ZETA FUNCTION

We can evaluate “X ” as

log |ζ(σ)3 ζ(σ + it)4 ζ(σ + 2it)|

= 3 log |ζ(σ)| + 4 log |ζ(σ + it)| + log |ζ(σ + 2it)|
! !
Y Y
= 3 log |1 − p−σ |−1 + 4 log |1 − p−σ−it |−1
p p
!
Y
+ log |1 − p−σ−2it |−1
p
X 
= − 3 log |1 − p−σ | − 4 log |1 − p−σ−it | − log |1 − p−σ−2it |
p
X
− 3 Re Log(1 − p−σ ) − 4 Re Log(1 − p−σ−it )
   
=
p
 −σ−2it

− Re Log 1 − p ,

where Log(·) denotes the principal branch of the logarithm function. Now
note that for z = a + ib with a > 1 and p prime we have |p−z | = p−a < 1, so
∞
X
−z p−mz
− Log(1 − p ) = ,
m=1
m
and
∞
h X
−zp−ma
i h i
− Re Log(1 − p ) = Re cos(mb log p) + i sin(mb log p)
m=1
m
∞
X p−ma
= cos(mb log p).
m=1
m
So we can rewrite X as
∞
X
X= cn n−σ (3 + 4 cos θn + cos(2θn ))
n=1

where θn = t log n and cn = 1/m if n = pm for some prime p. We can now use
a simple trigonometric identity

3 + 4 cos θ + cos(2θ) = 2(1 + cos θ)2 ,

to rewrite X yet again as
∞
X
X=2 cn n−σ (1 + cos θn )2 .
n=1
71 16 THE PRIME NUMBER THEOREM

We have proved a crucial fact, namely that X ≥ 0, or equivalently that

eX = |ζ(σ)3 ζ(σ + it)4 ζ(σ + 2it)| ≥ 1.

We now claim that this innocent-looking inequality is incompatible with the
existence of a zero of ζ(s) on the line Re(s) = 1. Indeed, assume by contra-
diction that ζ(1 + it) = 0 for some real t 6= 0. Then the three quantities ζ(σ),
ζ(σ + it) and ζ(σ + 2it) have the following asymptotic behavior as σ & 1:
1
|ζ(σ)| = + O(1) (since ζ(s) has a pole at s = 1),
σ−1
|ζ(σ + it)| = O(σ − 1) (since ζ(s) has a zero at s = 1 + it),
|ζ(σ + 2it)| = O(1) (since ζ(s) is holomorphic at s = 1 + 2it).

Combining these results we have that

eX = |ζ(σ)3 ζ(σ + it)4 ζ(σ + 2it)| = O((σ − 1)−3 (σ − 1)4 ) = O(σ − 1).
In particular, eX → 0 as σ & 1, in contradiction to the result we proved above
that eX ≥ 1. This proves the claim that ζ(s) cannot have a zero on the line
Re(s) = 1.
Exercise 16. The above proof that eX ≥ 1 (which immediately implied the
claim of the theorem) relied on showing that for any prime number p, the
corresponding factors in the Euler product formula satisfy the inequality

(1 − p−σ )−3 |1 − p−σ p−it |−4 |1 − p−σ p−2it |−1 ≥ 1,

and this was proved by taking the logarithm of the left hand-side, expanding
in a power series and using the elementary trigonometric identity 3+4 cos θ +
cos 2θ = 2(1 + cos θ)2 . However, one can imagine a more direct approach that
starts as follows: denote x = p−σ and z = p−it = e−it log p . Then the inequality
reduces to the claim that

(1 − x)3 |1 − zx|4 |1 − z 2 x| ≤ 1
for all x ∈ [0, 1] and z satisfying |z| = 1. Since this is an elementary inequality,
it seems like it ought to have an elementary proof (i.e., a proof that does not
involve logarithms and power series expansions). Can you find such a proof?

16 The prime number theorem

The prime number theorem was proved in 1896 by Jacques Hadamard and
independently by Charles Jean de la Vallée Poussin, using the groundbreak-
ing ideas from Riemann’s famous 1859 paper in which he introduced the use
72 16 THE PRIME NUMBER THEOREM

of the Riemann zeta function as a tool for counting prime numbers. (This
was the only number theory paper Riemann wrote in his career!) The his-
tory (including all the technical details) of these developments is described
extremely well in the classic textbook [4], which I highly recommend.
The original proofs of the prime number theorem were very complicated
and relied on the “explicit formula of number theory” (that I mentioned in
the previous section) and some of its variants. Throughout the 20th cen-
tury, mathematicians worked hard to find simpler ways to derive the prime
number theorem. This resulted in several important developments (such as
the Wiener tauberian theorem and the Hardy-Littlewood tauberian theorem)
that advanced not just the state of analytic number theory but also complex
analysis, harmonic analysis and functional analysis. Despite all the efforts
and the discovery of several paths to a proof that were simpler than the orig-
inal approach, all proofs remained quite difficult. . . until the year 1980, when
the mathematician Donald Newman discovered a wonderfully simple way to
derive the theorem using a completely elementary use of complex analysis.
It is Newman’s proof (as presented in the short paper [13] by D. Zagier) that
I present here.
Define the weighted prime counting functions
X
π(x) = #{p prime : p ≤ x} = 1,
p≤x
 
X X log x
ψ(x) = log p = log p ,
p≤x
log p
pk ≤x

with the convention that the symbol p in a summation denotes a prime num-
ber, and pk denotes a prime power, so that summation over p ≤ x denotes
summation over all primes ≤ x, and the summation over pk denotes summa-
tion over all prime powers ≤ x. Another customary way to write the function
ψ(x) is as X
ψ(x) = Λ(n),
n≤x

where the function Λ(n), called the von Mangoldt function, is defined by
Lambda (
log p if n = pk , p prime,
Λ(n) =
0 otherwise.
x
Lemma 11. The prime number theorem π(x) ∼ log x
is equivalent to the
statement that ψ(x) ∼ x.
73 16 THE PRIME NUMBER THEOREM

Proof. Note the inequality

  X
X log x log x X
ψ(x) = log p ≤ log p = log x = log x · π(x).
p≤x
log p p≤x
log p p≤x

In the opposite direction, we have a similar (but slightly less elegant) in-
equality, namely that for any 0 <  < 1 and x ≥ 2,
X X X
log x1−


ψ(x) ≥ log p ≥ log p ≥
p≤x x1− <p≤x x1− <p≤x
   
= (1 − ) log x π(x) − π(x1− ) ≥ (1 − ) log x π(x) − x1− .

Now assume that ψ(x) ∼ x as x → ∞. Then the first of the two bounds above
implies that
ψ(x)
π(x) ≥ ,
log x
so  
x
lim inf π(x)/ ≥ 1.
x→∞ log x
On the other hand, the second of the two bounds implies that

1 ψ(x)
π(x) ≤ · + x1− ,
1 −  log x
 
x 1 log x 1
which implies that lim supx→∞ π(x)/ log x ≤ 1− +lim supx→∞ x
= 1−
. Since
 was an arbitrary number in (0, 1), it follows that
 
x
lim sup π(x)/ ≤ 1.
x→∞ log x
Combining the two results about the lim inf and the lim sup gives that π(x) ∼
x/ log x.
Now assume that π(x) ∼ logx x , and apply the inequalities we derived above
in the opposite direction from before. That is, we have

ψ(x) ≤ log x · π(x),

so
lim sup ψ(x)/x ≤ 1.
x→∞
On the other hand,

ψ(x) ≥ (1 − ) log x(π(x) − x1− )

74 16 THE PRIME NUMBER THEOREM

implies that
 
log x
lim inf ψ(x)/x ≥ lim (1 − ) 1 −  = 1 − .
x→∞ x→∞ x
ψ(x)
Again, since  ∈ (0, 1) was arbitrary, it follows that lim inf x→∞ = 1. Com- x
ψ(x)
bining the two results about the lim inf and lim sup proves that limx→∞ x =
1, as claimed.
Lemma 12. For Re(s) > 1 we have
∞
ζ 0 (s) X
− = Λ(n)n−s .
ζ(s) n=1

Proof. Using the Euler product formula and taking the logarithmic deriva-
tive (which is an operation that works as it should when applied to infinite
products of holomorphic functions that are uniformly convergent on compact
subsets), we have
ζ 0 (s) X d
ds
(1− p−s ) X log p · p−s
− = =
ζ(s) p
1 − p−s p
1 − p−s
X ∞
X X
−s −2s −3s
= log p (p +p +p + . . .) = log p · p−ks
p p prime k=1
∞
X
= Λ(n)n−s .
n=1

Lemma 13. There is a constant C > 0 such that ψ(x) < Cx for all x ≥ 1.

Proof. The idea of the proof is that the binomial coefficient 2n

n
is not too
large on the one hand, but is divisible by many primes (all primes between n
and 2n) on the other hand — hence it follows that there cannot be too many
primes, and in particular the weighted prime-counting function ψ(x) can be
easily bounded from above using such an argument. Specifically, we have
that
2n  
  !
X 2n2n Y X
22n 2n
= (1 + 1) = > ≥ p = exp log p
k=0
k n n<p≤2n n<p≤2n
 X 
= exp ψ(2n) − ψ(n) − log p .
n<pk ≤2n, k>1
 √ 2

≥ exp ψ(2n) − ψ(n) − O( n log n) .
75 16 THE PRIME NUMBER THEOREM

√
(The estimate O( n log2 n) for the sum of log p for prime powers higher than
1 is easy and is left as an exercise.) Taking the logarithm of both sides, this
gives the bound
√
ψ(2n) − ψ(n) ≤ 2n log 2 + C1 n log n ≤ C2 n,

valid for all n ≥ 1 with some constant C2 > 0. It follows that

ψ(2m ) = (ψ(2m ) − ψ(2m−1 ))

+ (ψ(2m−1 ) − ψ(2m−2 )) + . . . + (ψ(21 ) − ψ(20 ))
≤ C2 (2m−1 + . . . + 20 ) ≤ C2 2m ,

so the inequality ψ(x) ≤ C2 x is satisfied for x = 2m . It is now easy to see that

this implies the result also for general x, since for x = 2m + ` with 0 ≤ ` < 2m
we have

ψ(x) = ψ(2m + `) ≤ ψ(2m+1 ) ≤ C2 2m+1 ≤ 2C2 2m ≤ 2C2 x.

Theorem 37 (Newman’s tauberian theorem). Let f : [0, ∞) → R be a bounded

function that is integrable on compact intervals. Define a function g(z) of a
complex variable z by Z ∞
g(z) = f (t)e−zt dt
0
(g is known as the Laplace transform of f ). Clearly g(z) is defined and holo-
morphic in the open half-plane Re(z) > 0. Assume that g(z) has an analytic
continuation
R∞ to an open region Ω containing the closed half-plane Re(z) ≥ 0.
Then 0 f (t) dt exists and is equal to g(0) (the value at z = 0 of the analytic
continuation of g ).

Proof. Define a truncated version of the integral defining g(z), namely

Z T
gT (z) = f (t)e−zt dt
0

for T > 0, which for any T is an entire function of z . Our goal is to show
that limT →∞ gT (0) = g(0). This can be achieved using a clever application of
Cauchy’s integral formula. Fix some large R > 0 and a small δ > 0 (which
depends on R in a way that will be explained shortly), and consider the con-
tour C consisting of the part of the circle |z| = R that lies in the half-plane
Re(z) ≥ −δ , together with the straight line segment along the line Re(z) = −δ
connecting the top and bottom intersection points of this circle with the line
(see Fig. 6(a)). Assume that δ is small enough so that g(z) (which extends an-
alytically at least slightly to the right of Re(z) = 0) is holomorphic in an open
76 16 THE PRIME NUMBER THEOREM

C C+ C- ' C+

C-

R R R

δ δ

(a) (b) (c)

0
Figure 6: The contours C , C+ , C− and C− .

set containing C and the region enclosed by it. Then by Cauchy’s integral
formula we have
z 2 dz
Z  
1 Tz
g(0) − gT (0) = (g(z) − gT (z))e 1+ 2
2πi C R z
z 2 dz
Z Z   
1 Tz
= + (g(z) − gT (z))e 1+ 2 ,
2πi C+ C− R z

where we separate the contour into two parts, a semicircular arc C+ that
lies in the half-plane Re(z) > 0, and the remaining part C− in the half-plane
Re(z) < 0 (Fig. 6(b)). We now bound the integral separately on C+ and on
C− . First, for z lying on C+ we have
∞ ∞
Be− Re(z)T
Z  Z
−zt
|e−zt | dt =
 
|g(z) − gT (z)| =  f (t)e dt ≤ B
 ,
T T Re(z)
where B = supt≥0 |f (t)|, and

z 2 
  
 Tz 2 Re(z)
e 1 + 2  = eRe(z)T
 R R

(by the trivial identity |1 + eit |2 = |eit (eit + eit )|2 = 2 cos(t), valid for t ∈ R). So
in combination we have
z 2 dz 
 Z   
 1 Tz 2B B

 2πi (g(z) − gT (z))e 1+ 2 ≤ (πR) 2
= .
C+ R z  2πR R

Next, for C− , we bound the integral by bounding the contributions from

g(z) and gT (z) separately. In the case of gT (z), the function is entire, so we
77 16 THE PRIME NUMBER THEOREM

0
can deform the contour, replacing it with the semicircular arc C− = {|z| =
R, Re(z) < 0} (Fig. 6(c)). On this contour we have the estimate
T T
Be− Re(z)T
Z  Z
−zt
|e−zt | dt =
 
|gT (z)| =  f (t)e dt ≤ B ,
0 −∞ | Re(z)|

which leads using a similar calculation as before to the estimate

2
  
 |dz|
Z
1  T z z B
gT (z)e 1 + 2
 ≤ .
2πi 0
C−
 R  |z| R

The remaining integral

2
  
 |dz|
Z
1 
g(z)e T z z
1 + 2 
2πi C−
 R |z|

tends to 0 as T → ∞, since the dependence on T is only through the factor

eT z , which converges to 0 uniformly on compact sets in Re(z) < 0 as T → ∞.
Combining the above estimates, we have shown that

2B
lim sup |g(0) − gT (0)| ≤ .
T →∞ R
Since R was an arbitrary positive number, the lim sup must be 0, and the
theorem is proved.

Consider now a very specific application of Newman’s theorem: take

f (t) = ψ(et )e−t − 1 (t ≥ 0),

which is bounded by the lemma we proved above, as our function f (t). The
associated function g(z) is then
Z ∞ Z ∞ 
−t −ztψ(x)
g(z) = t
(ψ(e )e − 1)e dt = − 1 x−z−1 dx
0 1 x
Z ∞ Z ∞ X !
1 1
= ψ(x)x−z−2 dx − = Λ(n) x−z−2 dx −
1 z 1 n≤x
z
∞ Z ∞ ∞ ∞
x−z−1 

X
−z−2 1 X 1
= Λ(n) x dx − = Λ(n) −
n=1 n z n=1
−z − 1 n
 z
∞
1 X 1 1 ζ 0 (z + 1) 1
= Λ(n)n−z−1 − = − · − (Re(z) > 0).
z + 1 n=1 z z + 1 ζ(z + 1) z
78 16 THE PRIME NUMBER THEOREM

Recall that −ζ 0 (s)/ζ(s) has a simple pole at s = 1 with residue 1 (because

ζ(s) has a simple pole at s = 1; it is useful to remember the more general
fact that if a holomorphic function h(z) has a zero of order k at z = z0 then
the logarithmic derivative h0 (z)/h(z) has a simple pole at z = z0 with residue
0 (z+1)
1
k ). So − z+1 · ζζ(z+1) has a simple pole with residue 1 at z = 0, and therefore
1 ζ 0 (z+1)
− z+1 · ζ(z+1)
− z1 has a removable singularity at z = 0. Thus, the identity
0 (z+1)
g(z) = − z+1 1
· ζζ(z+1) − z1 shows that g(z) extends analytically to a holomorphic
function in the set
{z ∈ C : ζ(z + 1) 6= 0}.
By the “toy Riemann Hypothesis” — the theorem we proved according to
which ζ(s) has no zeros on the line Re(s) = 1, g(z) in particular extends
holomorphically to an open set containing the half-plane Re(z) ≥ 0. Thus,
f (t) satisfies the assumption of Newman’s theorem. We conclude from the
theorem that the integral
Z ∞ Z ∞ Z ∞  
t −t ψ(x) dx
f (t) dt = (ψ(e )e − 1)dt = −1
0 0 1 x x
∞
ψ(x) − x
Z
= dx
1 x2
converges.

Proof of the prime number theorem. We will prove that ψ(x) ∼ x, which we
already showed is equivalent to the prime number theorem. Assume by con-
ψ(x) ψ(x)
tradiction that lim supx→∞ x > 1 or lim inf x→∞ x < 1. In the first case,
that means there exists a number λ > 1 such that ψ(x) ≥ λx for arbitrarily
large x. For such values of x it then follows that
λx λx λ
ψ(t) − t λ−t λx − t
Z Z Z
dt ≥ dt =: A > 0, dt =
x xt2 1 t2 t2
R∞
but this is inconsistent with the fact that the integral 1 (ψ(x) − x)x−2 dx
converges.
ψ(x)
Similarly, in the event that lim inf x→∞ x < 1, that means that there exists
a µ < 1 such that ψ(x) ≤ µx for arbitrarily large x, in which case we have
that
x x 1
ψ(t) − t λx − t λ−t
Z Z Z
dt ≤ dt = dt =: B < 0,
λx t2 λx t2 λ t2
again giving a contradiction to the convergence of the integral.
79 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

17 Introduction to asymptotic analysis

In this section we’ll learn how to use complex analysis to prove asymptotic
formulas such as
√  n n
n! ∼ 2πn (Stirling’s formula), (2)
e
1 √
p(n) ∼ √ eπ 2n/3 (the Hardy-Ramanujan formula), (3)
4 3n
 
1 −1/4 2 3/2
Ai(x) ∼ √ x exp − x (asymptotics for the Airy function), (4)
2 π 3

and more. At the heart of many such results is an important technique known
as the saddle point method. Some related techniques (that are all minor
variations on the same theme) are Laplace’s method, the steepest descent
method and the stationary phase method.

17.1 First example: Stirling’s formula

Our goal in this subsection is to prove a version of Stirling’s approximation
(2) for the factorial function n!. Let us start by simply having a bit of seem-
ingly aimless fun and asking ourselves, what are some really easy things we
can say about the magnitude of n! for large n?3 An obvious upper bound is
n! ≤ nn . As for a lower bound, one can say similarly trivial things such as
n! ≥ (n/2)n/2 , but that is quite far from the upper bound. We can do a bit
better by making the simple observation that for any real number x > 0, we
have the relations
∞
xn X xn
≤ = ex ,
n! n=0
n!
and this gives a lower bound
n! ≥ e−x xn .
It now makes sense to try to get the best lower bound possible by looking for
the x where the lower-bounding function is maximal. This happens when

d −x n

e x = e−x −xn + nxn−1 = e−x xn−1 (−x + n),


0=
dx
3
As a general rule of problem-solving, it’s often helpful to start attacking a problem by
thinking about really easy things you can say about it before moving on to advanced tech-
niques. It’s a great way to develop your intuition, and sometimes you discover that the easy
techniques solve the problem outright.
80 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

i.e., when x = n. Plugging this value into the inequality gives the bound

n! ≥ (n/e)n (n ≥ 1).

This is of course a standard and very easy result, but it’s intriguing to note
that it’s brought us quite close to the “true” asymptotics given by (2). The
point of this trivial calculation is that, as we shall see below, there is some-
thing special about the value x = n that resulted from this maximization op-
eration; when interpreted in the context of complex analysis, it corresponds
to a so-called “saddle point,” since it is a local minimum of ex /xn as one
moves along the real axis, but it will be a local maximum when one moves
in the orthogonal direction parallel to the imaginary axis. Thus, the trivial
approach has revealed the kernel of a deeper, much more powerful one.
Now let’s move on to the more powerful approach. Our “toy” approach
was based purely on real analysis, making use of the Taylor expansion of
the function x 7→ ex ; it turns out we could do better by thinking complex-
analytically. Start with the power series expansion
∞
z
X zn
e = .
n=0
n!

As we know very well from our study of Cauchy’s integral formula and the
residue theorem, the nth Taylor coefficient can be extracted from the func-
tion by contour integration, that is, by writing

ez
I
1 1
= dz,
n! 2πi |z|=r z n+1

where the radius r of the circle chosen as the contour of integration is an

arbitrary positive number. It turns out that some values of r are better than
others when one is trying to do asymptotics. We select r = n (I’ll explain
later where that seemingly inspired choice comes from), to get
Z π
ez
I
1 1 1 it

−n −int
= dz = exp ne n e i dt
n! 2πi |z|=n z n+1 2πi −π
Z π
1
exp n(eit − it) dt


= n
2π n −π
Z π
en
exp n(eit − 1 − it) dt,


= n
2π n −π
where we have strategically massaged the integrand (by pulling out the fac-
tor en ) to cancel out a term in the Taylor expansion of eit , in addition to a
81 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

term that was already canceled out. For convenience, rewrite this as
π
nn
Z
1
exp n(eit − 1 − it) dt.


n
=
e n! 2π −π

Now noting that

√  √ 3
it nt2

3 ( nt)2 ( nt)
n e − 1 − it = − + O(nt ) = +O √ ,
2 2 n
√
for |t| small, we see that a change of variable u = nt in the integral will
enable us to rewrite this as
u2
   3
√ iu u
iu/ n
n e −1− √ =− +O √ .
n 2 n
√
Performing the change of variable and moving a factor of n to the left-hand
side, the integral then becomes
√ √
π n
nnn
Z   
1 √ iu
iu/ n
n
= √
exp n e −1− √ du.
e n! 2π −π n n
2
The integrand converges pointwise to e−u /2 (for u fixed Rand n → ∞), so
√it’s
∞ 2
reasonable to guess that the integral should converge to −∞ e−u /2 du = 2π ,
which would lead to the formula
√ n
nn 1
n
≈√ ,
e n! 2π
or √  n n
n! ≈ 2πn ,
e
√
which is precisely Stirling’s formula. However, note that the O(u3 / n) esti-
√
mate holds whenever t = u/ n is in a neighborhood of 0, and since u actually
√ √
ranges in [−π n, π n], we need to be more careful to get a precise asymp-
totic result. To proceed, it makes sense to divide the integral into two parts.
Denote M = n1/10 , and let
√
Z π n 
√

iu/ n iu
I= √
exp n e −1− √ du = I1 + I2 ,
−π n n
Z M  
√

iu/ n iu
I1 = exp n e −1− √ du,
−M n
Z   
√ iu
iu/ n
I2 = √ √
exp n e −1− √ du.
[−π n,π n]\[−M,M ] n
82 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

We now estimate each of I1 and I2 separately. For I1 , we have

Z M  2  3 
u u
I1 = exp − + O √ du
−M 2 n
Z M   3 
−u2 /2 u
= e exp O √ du
−M n
Z M

M −u2 /2
 3  Z
u −u2 /2 −1/5
= 1+O √ e du = 1 + O(n ) e du
−M n −M
Z ∞ Z ∞
−1/5 2
e−u /2 du


= 1 + O(n ) −2
−∞ M
−1/5

√ −1/5



= 1 + O(n ) 2π − O exp −n
√
= 1 + O(n−1/5 ) 2π.

For I2 , we have
√
π n
Z    
 √ iu 
iu/ n
|I2 | ≤ 2 exp n e − 1 − √  du
M
 n 
Z π √n   √ 
=2 exp n Re eiu/ n − 1 du
M
Z π √n     
u
=2 exp n cos √ −1 du
M n

Now use the elementary fact that cos(t) ≤ 1 − t2 /8 for x ∈ [−π, π] (see Fig. 7)
to infer further that
√
Z π n √
exp −u2 /8 du ≤ 2π n exp −n1/5 = O(n−1/5 ).



|I2 | ≤ 2
M

Combining the above results, we have proved the following version of Stir-
ling’s formula with a quantitative (though suboptimal) bound:

Theorem 38 (Stirling’s approximation for n!). The asymptotic relation

√  n n
n! = 1 + O(n−1/5 ) 2πn
e
holds as n → ∞.
83 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

1.0

0.5

-3 -2 -1 1 2 3

-0.5

-1.0

Figure 7: Illustration of the inequality cos(t) ≤ 1 − t2 /8.

17.2 Second example: the central binomial coefficient

(2n)!
Let an = 2n


n
= (n!)2 . A standard way to find the asymptotic behavior for an
as n → ∞ is to use Stirling’s formula. This easily gives that

4n
 
2n
= (1 + o(1)) √ .
n πn

(Note that this is not too far from the trivial upper bound 2n


n
≤ (1 + 1)2n =
2n
2 .) It is instructive to rederive this result using the saddle-point method,
starting from the expansion
2n  
2n
X 2n n
(1 + z) = z ,
k=0
k

which in particular gives the contour integral representation

(1 + z)2n
  I
2n 1
= dz.
n 2πi |z|=r z n+1
By the same trivial method for deriving upper bounds that we used in the
case of the Taylor coefficients 1/n! of the function ez , we have that for each
x > 0,  
2n
≤ (1 + x)2n /xn = exp (log(1 + x) − n log x) .
n
We optimize over x by differentiating the expression log(1 + x) − n log x inside
the exponent and setting the derivative equal to 0. This gives x = 1, the
location of the

saddle point. For this value of x, we again recover the trivial
2n 2n
inequality n ≤ 2 .
84 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

Next, equipped with the knowledge of the saddle point, we set r = 1 in

the contour integral formula, to get
Z π
(1 + z)2n
  I
2n 1 1
= dz = (1 + eit )2n e−int dt
n 2πi |z|=r z n+1 2π −π
Z π
1 
it


= exp n 2 log(1 + e ) − t dt.
2π −π

Now note that the expression in the exponent has the Taylor expansion

1
n(2 log(1 + eit ) − t) = 2 log 2 − nt2 + O(nt4 ) as t → 0.
4
√
Again, we see that a change of variables u = t/ n will bring the integrand
to an asymptotically scale-free form. More precisely, we have
  Z π  !
2n 1 1 2
= exp n 2 log 2 − nt + O(nt4 ) dt
n 2π −π 4
Z π
4n
  4 
1 2 u
= √ exp − u + O du.
2π n −π 4 n

It is now reasonable to guess that in the limit as n → ∞, the pointwise limit

of the integrands translates to a limit of the integrals, so that we get the
approximation
Z ∞
4n 4n √ 4n
 
2n 2
≈ √ e−u /4 du = √ 2 π = √ ,
n 2π n −∞ 2π n πn

as required. Indeed, this is correct, but it remains to make this argument

precise by breaking up the integral into two parts, a “central part” where
the O(u4 /n) error term can be shown to be small, and the remaining part
that has to be bounded separately.

Exercise 17. Complete this analysis
to give a rigorous proof using this
√
method of the asymptotic formula 2n
n
= (1 + o(1))4n
/ πn.

Exercise 18. Repeat this analysis for the sequence (bn )∞ n=1 of central trino-
mial coefficients, where bn is defined as the coefficient of xn in the expansion
of (1+x+x2 )n , a definition that immediately gives rise to the contour integral
representation
(1 + z + z 2 )n
I
1
bn = dz.
2πi |z|=r z n+1
85 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

2 5

1 4

3
10 20 30 40

-1 2

-2 1

-3
10 20 30 40

-4 -1

-5 -2

(a) (b)

Figure 8: An illustration (with n = 40) of the random walks enumerated by

(a) the central binomial coefficients and (b) the central trinomial coefficients.

Like their more famous cousins the central binomial coefficients, these coef-
ficients are important in combinatorics and probability theory. Specifically,
an and bn correspond to the numbers of random walks on Z that start and end
at 0 and have n steps, where in the case of the central binomial coefficients
the allowed steps of the walk are −1 or +1, and in the case of the central
trinomial coefficients the allowed steps are −1, 0 or 1; see Fig. 8.
Using a saddle point analysis, show that the asymptotic behavior of bn as
n → ∞ is given by √
3 · 3n
bn ∼ √ .
πn

17.3 A conceptual explanation

In both the examples of Stirling’s formula and the central binomial coeffi-
cient we analyzed above, we made what looked like ad hoc choices regard-
ing how to “massage” the integrals, what value r to use for the radius of the
contour of integration, what change of variables to make in the integral, etc.
Now let us think more conceptually and see if we can generalize these ideas.
Note that the quantities we were trying to estimate took a particular form,
where for some function g(z) our sequence of numbers could be represented
86 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

in the form

e−ng(z) dz
I I
1 1   dz
a(n) = = exp − n(g(z) + log z)
2πi |z|=r z n z 2πi |z|=r z
Z π
1 it
= e−ng(re ) r−n e−int dt
2π −π
Z π
1 
it

= exp − n(g(re ) + it − log r) dt.
2π −π

(Sometimes g(z) would actually be gn (z), a sequence of functions that de-

pends on n.) The idea that is key to making the method work turns out to be
to choose the contour radius r as the solution to the equation

d 1
(g(z) + log z) = g 0 (z) + = 0.
dz z
This causes the first-order term in the Taylor expansion of g(z) + log z around
z = r to disappear. One is then left with a constant term, that can be pulled
outside of the integral; a second order term, which (in favorable circum-
stances where this technique actually works) causes the integrand to be
2
well-approximated by a Gaussian density function e−u /2 near z = r ; and
lower-order terms which can be shown to be asymptotically negligible.
Geometrically, if one plots the graph of |g(z) + 1/z| then one finds the
emergence of a saddle point at z = r , and this is the origin of the term
“saddle point method.” This phenomenon is illustrated with many beautiful
examples and graphical figures in the lecture slides [6] prepared by Flajolet
and Sedgewick as an online resource to accompany their excellent textbook
Analytic Combinatorics [5].

Exercise 19. It is instructive to see an example where the saddle point anal-
ysis fails if applied mindlessly without checking that the part of the integral
that is usually assumed to make a negligible contribution actually behaves
that way. A simple example illustrating what can go wrong is the function
∞ ∞
z2
X z 2n X
f (z) = e = = bn z n ,
n=0
n! n=0

where the Taylor coefficients are

(
1
(n/2)!
n even,
bn =
0 otherwise.
87 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

Clearly any analysis, asymptotic or not, needs to address and take into ac-
count the fact that bn behaves differently according to whether n is even or
odd. Try to apply the method we developed to derive an asymptotic formula
for bn . The method fails, but the failure can easily be turned into a success
by noting that there are actually two saddle points, each of which makes a
contribution to the integral, in such a way that for odd n the contributions
cancel and for even n they reinforce each other. This shows that periodicities
are one common pitfall to look out for when doing asymptotic analysis.

Exercise 20. As another amusing example, apply the saddle point method
P∞
to the function f (z) = 1/(1 − z) = n=0 dn z n , for which the Taylor coefficients
dn = 1 are all equal to 1. Can you succeed in deriving an asymptotic formula
for the constant function 1?

17.4 Third example: Stirling’s formula for the gamma

function.
Our next goal is to prove a stronger version of Stirling’s formula that gives
an asymptotic formula for Γ(t), the extension of the factorial function to non-
integer arguments. Specifically, we will prove.

Theorem 39 (Stirling’s approximation for Γ(t)). For a real-valued argument

t, the gamma function satisfies the asymptotic formula
r  t
2π t
Γ(t) = 1 + O(t−1/5 )


(t → ∞).
t e

Proof. We use a method called Laplace’s method, which is a variant of the

saddle-point method adapted to estimating real integrals instead of contour
integrals around a circle. Start with the integral formula
Z ∞
Γ(t) = e−x xt−1 dx
0

Performing the change of variables x = tu in the integral gives that

Z ∞ Z ∞
−tu t−1 t −t
Γ(t) = tt
e u e−tu+t ut−1 du
du = t e
0 0
Z ∞ Z ∞  t
t −t −tu+t t−1 t −t −tΦ(u) du t
=te e u du = t e e = I(t),
0 0 u e
88 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

where we define

Φ(u) = u − 1 − log u,
Z ∞
du
I(t) = e−tΦ(u) .
0 u
(Again, note that we massaged the integrand to cancel the Taylor expansion
of − log u around u = 1 up to the first order.) Our goal is to prove that
r
2π
I(t) = + O(t−7/10 ) as t → ∞.
t
As before, this will be done by splitting the integral into a main term and
error terms. The idea is that for large t, the bulk of the contribution to
the integral comes from a region very near the point where Φ(u) takes its
minimum. It is easy to check by differentiation that this minimum is obtained
at u = 1, and that we have

Φ(1) = 0, Φ0 (1) = 0, Φ00 (1) = 1,

and Φ(u) ≥ 0 for all u ≥ 0. See Fig. 9. Denote

Z 1/2
du
I1 = e−tΦ(u)
,
0 u
Z 2
du
I2 = e−tΦ(u) ,
1/2 u
Z ∞
du
I3 = e−tΦ(u) ,
2 u

so that I(t) = I1 + I2 + I3 . The main contribution will come from I2 , the part
of the integral that contains the critical point u = 1, so let us examine that
term first. Expanding Φ(u) in a Taylor series around u = 1, we have

(u − 1)2
Φ(u) = + O((u − 1)3 )
2
 
(u−1)2 
for u ∈ [1/2, 2] (in fact the explicit bound Φ(u) − 2  ≤ (u − 1)3 on this

interval can be easily checked). As before, noting that
 √
(u − 1)2 1 √ ( t(u − 1))3
  
3 2
t + O((u − 1) ) = ( t(u − 1)) + O √ ,
2 2 t
89 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

3.0

2.5

2.0

1.5

1.0

0.5

1 2 3 4

Figure 9: The function Φ(u) = u − 1 − log u.

√
we see that it is natural to apply a linear change of variables v = t(u − 1) to
bring the integrand to a scale-free, centered form. This results in
√
Z t  
1 v 1
I2 = √ √
exp −tΦ 1 + √ √ dv
t 1 t 1 + v/ t
− t
2
Z √t  2  3    
1 v v t
=√ √
exp − + O √ 1+O √ dv.
t − 12 t 2 t t
As before, we actually need to split
√ up this integral into two parts to take into
3
account the fact that the O(v / t) term can blow up when v is large enough.
Let M = t1/10 , and denote
Z M   
1 v 1
J1 = √ exp −tΦ 1 + √ √ dv,
t −M t 1 + v/ t
Z   
1 v 1
J2 = √ exp −tΦ 1 + √ √ dv,
t [− 21 √t,√t]\[−M,M ] t 1 + v/ t
so that I2 = J1 + J2 . For J1 we have
Z M   
1 v 1
J1 = √ exp −tΦ 1 + √ √ dv
t −M t 1 + v/ t
Z M   3    
1 −v 2 /2 v v
=√ e 1+O √ 1+O √ dv
t −M t t
Z M r
1 2 2π
= √ 1 + O(t−1/5 ) e−v /2 dv = 1 + O(t−1/5 ) ,



t −M t
in the last step using a similar estimate as the one we used in our proof of
Stirling’s approximation for n!. Next, for J2 we use the elementary inequality
90 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

(prove it as an exercise)
(u − 1)2
Φ(u) ≥ (0 ≤ u ≤ 1),
2
√ √ √
and the more obvious fact that 1/(1 + v/ t) ≤ 2 for v ∈ [− 12 t, t] to get that
Z Z ∞
2 −v 2 /2 4 2 /2
J2 ≤ √ e dv ≤ √ e−v dv
t 1√ √ t
[− t, t]\[−M,M ] M
2
1
= O(e−M ) = √ O(t−1/5 ).
t
as in our earlier proof. Combining the above results, we have shown that
r
−1/5

2π
I2 = 1 + O(t ) .
t
Next, we bound I1 . Here we use a different method since there is a different
source of potential trouble near the left end u = 0 of the integration inter-
val. Considering first a truncated integral over [ε, 1/2] and performing an
integration by parts, we have
Z 1/2 Z 1/2
−tΦ(u) du 1 d −tΦ(u)
1
e =− e du
ε u t ε du Φ0 (u)u
u=1/2
1 e−tΦ(u) 1 1/2 −tΦ(u) du
 Z
=− − e .
t u − 1 u=ε t ε (u − 1)2
Taking the limit as ε → 0 (and noting that Φ(ε) → +∞ in this limit) yields the
formula
Z 1/2  
2 1 du 1
I1 = e−tΦ(1/2) − e −tΦ(u)
=O as t → ∞.
t t 0 (u − 1)2 t

Finally, I leave it as an exercise to obtain a similar estimate I3 = O(1/t)

for the remaining integral on [2, ∞). Combining the various estimates yields
the claimed result that
r
2π
I(t) = I1 + I2 + I3 = 1 + O(t−1/5 )


.
t
The proof above is a simplified version of the analysis in Appendix A
of [11]. The more detailed analysis there shows that the asymptotic for-
mula we proved for Γ(t) remains valid for complex t. Specifically, they prove
that for complex s in the “Pac-Man shaped” region

Sδ = {z ∈ C : | arg z| ≥ π − δ}
91 17 INTRODUCTION TO ASYMPTOTIC ANALYSIS

(for each fixed 0 < δ < π ) the gamma function satisfies

√
Γ(s) = 1 + O(|s|−1/2 ) 2πss−1/2 e−s as |s| → ∞, s ∈ Sδ .

Here, ss−1/2 is defined as exp((s − 1/2) Log s), where Log denotes as usual
the principal branch of the logarithm function. This sort of approximation is
important in certain applications of complex analysis, for example to analytic
number theory.
92 PROBLEMS

Problems
1. Below is a list of basic formulas in complex analysis (think of them as
“formulas you need to know like the back of your hand”). Review each
of them, making sure that you understand what it says and why it is
true—that is, if it is a theorem, prove it, or if it is a definition, make
sure you understand that that is the case.

In the formulas below, a, b, c, d, t, x, y denote arbitrary real numbers; w, z

denote arbitrary complex numbers.

a. (a + bi)(c + di) l. |wz| = |w| · |z|

= (ac − bd) + (ad + bc)i  
m. |w| − |z| ≤ |w + z| ≤ |w| + |z|
 
b. i2 = −1
1 n. ex+iy = ex (cos(y) + i sin(y))
c. = −i
i o. |ez | = eRe(z)
d. z = Re(z) + i Im(z)
p. |ez | ≤ e|z|
e. z = Re(z) − i Im(z)
z+z q. eit = cos(t) + i sin(t)
f. Re(z) =
2 r. |eit | = 1
z−z eit + e−it
g. Im(z) = s. cos(t) =
2i 2
2
h. |z| = zz
e − e−it
it
1 z t. sin(t) =
i. = 2 2i
z |z| πi
u. e = −1
1 x − iy
j. = 2 v. e±πi/2 = ±i
x + iy x + y2
k. w·z =w·z w. e2πi = 1

2. Remind yourself of the definitions of the following terms in complex

analysis, referring to the textbook [11] (in particular, sections 1.2, 1.3,
2.1, 2.2) or online sources if necessary. Try to spend some time thinking
about the answers yourself before looking them up.

a. real part f. open set (in C)

b. imaginary part g. closed set
c. complex conjugate h. connected set
d. modulus i. bounded set
e. argument j. compact set
93 PROBLEMS

k. region complex variable)

l. convergent sequence r. holomorphic function
m. Cauchy sequence s. analytic function
n. limit point t. entire function
o. accumulation point u. meromorphic function
p. continuous function v. harmonic function (of two
q. differentiable function (of a variables)

3. For each of the following functions, determine for which z it is analytic

a. f (z) = z d. f (z) = |z|2

b. f (z) = Re(z) e. f (z) = z
c. f (z) = |z| f. f (z) = 1/z

4. For each of the following functions u(x, y), determine if there exists
a function v(x, y) such that f (x + iy) = u(x, y) + iv(x, y) is an entire
function, and if so, find it, and try to find a formula for f (z) directly in
terms of z rather than in terms of its real and imaginary parts. (Hint4 )

a. u(x, y) = x2 − y 2 c. u(x, y) = x4 −6x2 y 2 +3x+y 4 −2

b. u(x, y) = y 3 d. u(x, y) = cos x cosh y

5. Draw (approximately, or with as much precision as you can) the image

in the w-plane of the following figure in the z -plane
4
Use the well-known equations named after two pioneers of complex analysis.
94 PROBLEMS

-3 -2 -1 1 2 3

-1

-2

-3

under each of the following maps w = f (z):

a. w = 12 z d. w = (2 + i)z − 3
b. w = iz e. w = 1/z
c. w = z f. w = z 2 − 1

6. Prove that the complex numbers a, b, c form the vertices of an equilat-

eral triangle if and only if a2 + b2 + c2 = ab + ac + bc.

7. Illustrate the claim from page 11 regarding the orthogonality of the

level curves of the real and imaginary parts an analytic functions by
drawing (by hand after working out the relevant equations, or using a
computer) the level curves of Re(f ) and Im(f ) for f = z 2 , f = ez .

8. An immediate corollary of the Fundamental Theorem of Algebra (to-

gether with standard properties of polynomials, namely the fact that c
is a root of p(z) if and only if p(z) is divisible by the linear factor z − c)
is that any complex polynomial

p(z) = an z n + an−1 z n−1 + . . . + a0 ,

(where a0 , . . . , an ∈ C and an 6= 0), can be factored as

n
Y
p(z) = an (z − zk )
k=1
95 PROBLEMS

for some z1 , . . . , zn ∈ C (these are the roots of p(z) counted with multi-
plicities). Use this to prove that any such polynomial where the coeffi-
cients a0 , . . . , an are real has a factorization

p(z) = an Q1 (z)Q2 (z) . . . Qm (z)

where each Qk (z) is a linear or quadratic monic polynomial (i.e., is of
one of the forms z − c or z 2 + bz + c) with real coefficients.

9. Let p(z) = an z n + an−1 z n−1 + . . . + a0 be a complex polynomial of degree n

(that is, a0 , . . . , an ∈ C and an 6= 0), which as mentioned in question 1
above can be factored as
n
Y
p(z) = an (z − zk )
k=1

where z1 , . . . , zn are the roots of p(z) counted with multiplicities. As-

suming that n ≥ 2, the derivative p0 (z) can be similarly factored as
n−1
Y
0
p (z) = nan (z − wk )
k=1

where w1 , . . . , wn−1 denote the roots of p0 (z). Prove that w1 , . . . , wn−1 all
lie in the convex hull of z1 , . . . , zn (see Figure 1 for an illustration). That
is, each wk can be expressed as a convex combination

wk = α1 z1 + α2 z2 + . . . + αn zn ,
P
where α1 , . . . , αn are nonnegative real numbers and j αj = 1. (To be
clear, there are different coefficients for each k .)

Hint. A complex number z is a root of p0 (z) that is not also a root of

p(z) if and only if p0 (z)/p(z) = 0. (Note: the expression p0 /p is known as
the logarithmic derivative of p.) Find a way to make more explicit what
this equation says.

10. Cardano’s method for solving cubic equations. Let p(z) = az 3 +

bz 2 + cz + d, with a, b, c, d ∈ C, a 6= 0. We wish to solve the equation
p(z) = 0, i.e., find the roots of the cubic polynomial p(z).
b
(a) Show that the substitution w = z − 3a
brings the equation to the
simpler form
w3 + pw + q = 0 (5)
for some values of p, q (find them!) given as functions of a, b, c, d.
96 PROBLEMS

z3

z4
w3

w1

z1

w2

z2

Figure 10: An example of the roots of a complex polynomial and of its deriva-
.
tive. Here z1 = 0, z2 = 3 − i, z3 = 2 + 2i, z4 = 1+3i
2
and w1 = 0.375 + 0.586i,
. .
w2 = 2.336 − 0.335i, w3 = 1.414 + 1.624i.

(b) Show that assuming a solution to (5) of the form w = u + v , the

equation (5) for w can be solved by finding a pair u, v of complex
numbers such that the equations

p = −3uv, (6)
q = −(u3 + v 3 ) (7)

are satisfied.
(c) Explain why, in order to solve the pair of equations (6)–(7), one can
alternatively solve
p3
= −RS, (8)
27
q = −(R + S), (9)

where we now denote new unknowns R, S defined by R = u3 , S =

v 3 . More precisely, any solution of (6)–(7) can be obtained from
some (easily determined) solution of (8)–(9).
(d) Explain why the problem of solving (8)–(9) in the unknowns R, S is
equivalent to solving the quadratic equation
p3
t2 + qt − =0 (10)
27
in a (complex) unknown variable t.
97 PROBLEMS

(e) Using the above reductions, show that the three solutions of the
simplified cubic (5) can be expressed as

w1 = u + v,
w2 = ζ u + ζ v,
w3 = ζ u + ζ v,
√
where ζ = e2πi/3 = 21 (−1 + i 3) (a cube root of unity) and u, v are
properly chosen cube roots of R, S obtained as solutions to (10).
(f) Illustrate the above procedure by applying it to get formulas for
the roots of the cubic equation

z 3 + 6z 2 + 9z + 3 = 0.

Bring the formulas to a form that makes it clear that the roots are
real numbers.
 
a b
11. Let A = be a 2 × 2 real matrix. Prove the “conformality lemma”
c d
from page 11, which asserts the equivalence of the following three con-
ditions:

(a) A as a linear map preserves orientation (that is, det A > 0) and is
conformal, that is
hAw1 , Aw2 i hw1 , w2 i
=
|Aw1 | |Aw2 | |w1 | |w2 |
for all w1 , w2 ∈ R2 . (Here hw1 , w2 i denotes the standard inner prod-
uct in R2 , and |w| = hw, wi1/2 is the usual two-dimensional norm of
a vector in R2 .)
 
a b
(b) A takes the form A = for some a, b ∈ R.
−b a
 
cos θ − sin θ
(c) A takes the form A = r for some r > 0 and θ ∈ R.
sin θ cos θ
(That is, geometrically A acts by a rotation followed by a scaling.)

12. A function f = u + iv of a complex variable z = x + iy is traditionally

thought of as a function of the two coordinates x and y . However, if we
think of the equations

z = x + iy, z = x − iy
98 PROBLEMS

as representing a formal change of variables from the “real coordi-

nates” (x, y) to the “complex conjugate coordinates” (z, z), then it may
make sense to think of f as a function of the two variables z and z
(pretending that those are two independent variables). Thus we may
suggestively write u = u(z, z) and v = v(z, z), and consider operations
such as taking the partial derivatives of f, u, v with respect to z and z .

Show that, from this somewhat strange point of view, the Cauchy-
Riemann equations
∂u ∂v ∂u ∂v
= , =−
∂x ∂y ∂y ∂x
can be rewritten in the more concise equivalent form
∂f
= 0,
∂z
assuming that it is okay to apply the chain rule from multivariable cal-
culus; and moreover, that in this notation we also have the identity
∂f
f 0 (z) = .
∂z

13. Let f : Ω → C be a function defined on a region Ω such that both

the functions f (z) and zf (z) have real and imaginary parts that are
2 2
harmonic functions (i.e., satisfy the Laplace equation ∂∂xu2 + ∂∂yu2 = 0).
Prove that f (z) is holomorphic on Ω.

14. Uniform convergence on compact subsets. Given a sequence of

functions fn : Ω → C, n ≥ 1, defined on a region Ω ⊂ C, we say that fn
converges uniformly on compact subsets to a limiting function f : Ω →
C if for any compact subset K ⊂ Ω, fn (z) → f (z) as n → ∞, uniformly
on z ∈ K .

(a) (Warm-up) Write this definition more precisely in -δ language.

(b) Prove that if fn are holomorphic functions, fn → f uniformly on
compact subsets, fn0 → g uniformly on compact subsets, and g is
continuous, then f is holomorphic and f 0 = g .
Hint. Fix some z0 ∈ C. Start by proving that for z in a sufficiently
small neighborhood of z we have the two identities
Z z
fn (z) = fn (z0 ) + fn0 (w) dw,
z0
Z z
f (z) = f (z0 ) + g(w) dw,
z0
99 PROBLEMS

where the integral is over the line segment connecting z0 to z .

P∞ n
(c) Prove that a power series n=0 an z converges uniformly on com-
pact subsets in its disk of convergence, and that the function it
defines is continuous.
Hint. It is enough (explain why) to prove uniform convergence on
any closed disk of the form Dr (0) where 0 < r < R and R is the
radius of convergence of the series.
(d) Deduce that power series are holomorphic functions that can be
differentiated termwise (a fact we already proved in class in a
more direct way; the above approach provides an alternative proof).

Remark. This problem requires only elementary arguments, but using

more advanced material we will prove as a consequence of Cauchy’s
theorem that in part (b) above the assumption that the sequence of
derivatives fn0 converges to a limit can be dropped; that is, if a se-
quence of holomorphic functions converges uniformly on compact sub-
sets, then the limiting function is automatically a holomorphic function
whose derivative is the limit (in the sense of uniform convergence on
compacts) of the sequence of derivatives of the original sequence. This
is a surprising and nontrivial fact, as illustrated for example by the ob-
servation that the analogous statement in real analysis is false (e.g., by
the Weierstrass approximation theorem, any continuous function on a
closed interval is the uniform limit of a sequence of polynomials).

15. Cauchy’s theorem and irrotational vector fields. Recall from vec-
tor calculus that a planar vector field F = (P, Q) defined on some region

∂g ∂g
Ω ⊂ C = R2 is called conservative if it is of the form F = ∇g = ∂x , ∂y
(the gradient of g ) for some scalar function g : Ω → R. By the funda-
mental theorem of calculus for line integrals, for such a vector field we
have I
F · ds = 0
γ

for any closed curve γ . Recall also that (as is easy to check) any con-
servative vector field is irrotational, that is, it satisfies

curl F = 0

(where in the context of two-dimensional vector fields, the curl is simply

curl F = ∂Q
∂x
− ∂P
∂y
). The converse also holds under suitable conditions:
if the region Ω is simply-connected (a concept we will discuss later in
100 PROBLEMS

the course), then a theorem in vector calculus says that an irrotational

vector field is also conservative.

Use these background results to show that if f = u + iv is holomorphic

on a simply-connected domain Ω, then
I
f (z) dz = 0
γ

for any closed curve γ in Ω. (This is, of course, Cauchy’s theorem.)

16. The Bernoulli numbers. Define the function

z
z
 6 0,
if z =
f (z) = e − 1
1 if z = 0.

(a) Convince yourself that f (z) is analytic in a neighborhood of 0.

Where else is it analytic? In particular, find the maximal radius
R such that f (z) is analytic on the disk DR (0).
(b) One of the basic complex analysis theorems we will discuss is that
analytic functions have a power series expansion. The Bernoulli
numbers are the numbers (Bn )∞n=0 defined by the power series ex-
pansion
∞
X Bn
z n = f (z).
n=0
n!
For example, the first three Bernoulli numbers are B0 = 1, B1 =
−1/2, B2 = 1/6. Prove that the Bernoulli numbers satisfy the fol-
lowing identities:
i. B2k+1 = 0 for k = 1, 2, . . . (but not for k = 0).
P∞ n
Hint. A function g(z) = n=0 an z satisfies a1 = a3 = a5 =
. . . = 0 if and only if g(z) = g(−z), i.e., g(z) is an even function.
Pn−1 n+1


ii. (n + 1)Bn = − k=0 k
Bk , (n ≥ 2).
Pn−1 2n


iii. (2n + 1)B2n = − k=1 2k
B2k B2n−2k , (n ≥ 2).
Hint. Show that the function g(z) = f (z) + z/2 satisfies the
equation
g(z) − zg 0 (z) = g(z)2 − z 2 /4.
∞
z z  X B
2n 2n
iv. coth = z .
2 2 n=0
(2n)!
101 PROBLEMS

(c) Another general result is that the radius of convergence of the

power series of an analytic function around z = z0 is precisely the
radius of the maximal disk around z0 where f is analytic. Assuming
this, deduce that
 Bn 1/n
 
lim sup   = 1/R,
n→∞ n!
where R is the number you found in part (a). (Note: in a later
problem — problem 26 on pages 107–108 — we will derive a much
more precise estimate for the asymptotic rate of growth of the
Bernoulli numbers.)

17. Bessel functions. The Bessel functions are a family of functions (Jn )∞
n=−∞
of a complex variable,defined by
∞
X (−1)k  z 2k+n
Jn (z) = .
k=0
k!(k + n)! 2
√ x
P∞ k
(For example, note that J0 (−2 x) = k=0 (k!)2 , which is reminiscent of
the exponential function and already seems like a fairly natural function
to study.) Find the radius of convergence of the series defining Jn (z),
and prove that the Bessel functions satisfy the following properties:

i. J−n (z) = (−1)n Jn (z).

2n
ii. Recurrence relation: Jn+1 (z) = Jn (z) − Jn−1 (z).
z
iii. Bessel’s differential equation: z 2 Jn00 (z) + zJn0 (z) + (z 2 − n2 )Jn (z) = 0.
∞
X 1  z n
iv. Summation identity: Jn (z) = 1.
n=0
n! 2
v.* Other miscellaneous identities (for those who enjoy this sort of
thing—feel free to skip if you find these sorts of computations un-
102 PROBLEMS

interesting):
   ∞
z 1 X
exp t− = Jn (z)tn ,
2 t n=−∞
∞
X
cos(z sin t) = J0 (z) + 2 J2n (z) cos(2nt),
n=1
∞
X
sin(z sin t) = 2 J2n+1 (z) sin((2n + 1)t),
n=0
∞
X
cos(z cos t) = J0 (z) + 2 (−1)n J2n (z) cos(2nt),
n=1
∞
X
sin(z cos t) = 2 (−1)n J2n+1 (z) sin((2n + 1)t),
n=0
Z π
1
Jn (z) = cos (z sin t − nt) dt.
π 0

Hint for the last equation: cos(a − b) = cos(a) cos(b) + sin(a) sin(b).
Remark. The Bessel functions are very important functions in
mathematical physics, and appear naturally in connection with
various problems in diffusion, heat conduction, electrodynamics,
quantum mechanics, Brownian motion, probability, and more. More
recently they played an important role in some problems in com-
binatorics related to longest increasing subsequences (a subject I
wrote a book about, available to download from my home page).
Their properties as analytic functions of a complex variable are
also a classical, though no longer very fashionable, topic of study.

18. Show that Liouville’s theorem (“a bounded entire function is constant”)
can be proved directly using the “simple” (n = 0) case of Cauchy’s in-
tegral formula, instead of using the case n = 1 of the extended formula
as we did in the lecture.

Hint. For an arbitrary pair of complex numbers z1 , z2 ∈ C, show that

|f (z1 ) − f (z2 )| = 0.

19. Show that Liouville’s theorem can in fact be deduced even just from
the mean value property of holomorphic functions, which is the special
case of Cauchy’s integral formula in which z is taken as the center of
the circle around which the integration is performed.
103 PROBLEMS

Hint. Here it makes sense to consider a modified version of the mean

value property (that follows easily from the original version) that says
that f (z) is the average value of f (w) over a disc DR (z) (instead of a
circle CR (z)). That is,
ZZ
1
f (z) = f (x + iy) dx dy,
πR2 DR (z)

where the integral is an ordinary two-dimensional Riemann integral.

Explain why this formula holds, then use it to again bound |f (z1 )−f (z2 )|
from above by a quantity that goes to 0 as R → ∞.

20. Prove the following generalization of Liouville’s theorem: let f be an

entire function that satisfies for all z ∈ C the inequality

|f (z)| ≤ A + B|z|n

for some constants A, B > 0 and integer n ≥ 0. Then f is a polynomial

of degree ≤ n.

21. If p(z) = an z n + an−1 z n−1 + . . . + . . . + a0 is a polynomial of degree n such

that
n−1
X
|an | > |aj |,
j=0

prove that p(z) has exactly n zeros (counting multiplicities) in the unit
disc |z| < 1.

Hint. Use the fundamental theorem of algebra.

Note. This is a special case of a less elementary fact that can be proved
using Rouché’s theorem; see problem 28

22. The Cauchy integral formula is intimately connected to an important

formula from the theory of the Laplace equation and harmonic func-
tions called the Poisson integral formula. Solve exercises 11–12 (pages
66–67) in Chapter 2 of [Stein-Shakarchi], which explore this connec-
tion, and more generally the connection between holomorphic and har-
monic functions.

23. Spend at least 5–10 minutes thinking about the concept of a toy con-
tour. Specifically, for the case of a keyhole contour we discussed in
the context of the proof of Cauchy’s integral formula, think carefully
about the steps that are needed to get a proof of Cauchy’s theorem for
104 PROBLEMS

the region enclosed by such a contour. Even better, sketch a proof of

the key result that a function holomorphic in such a region (and there-
fore having the property that its contour integral along triangles and
rectangles vanish) has a primitive.

24. Characterizing some important families of holomorphic func-

tions on C and C
b . For the problems below, denote

b = C ∪ {∞} = the Riemann sphere,

C
K = the set of constant functions z 7→ c ∈ C,
L = the set of linear functions z 7→ az + b, a, b ∈ C,
X n
P = the set of complex polynomials z 7→ ak z k ,
k=0
p(z)
R = the set of rational functions z 7→ , p, q ∈ P,
q(z)
az + b
M = the set of Möbius transformations z 7→ , a, b, c, d ∈ C.
cz + d
Note the containment relations K ⊂ L ⊂ P ⊂ R ⊃ M ⊃ L ⊃ K.

(a) (Warm-up problem) Prove that an entire function has a removable

singularity at ∞ if and only if it is constant.
(b) Prove that the set of entire functions f : C → C that have a
nonessential singularity at ∞ is P , the polynomials.
(c) Prove that the set of meromorphic functions f : C → C b that have a
nonessential singularity at ∞ is R, the rational functions.
(d) Prove that the set of meromorphic, one-to-one and onto functions
b→C
f :C b is M\K, the set of nonconstant Möbius transformations.

Hint. Use the characterization in problem 24c above. Specifically,

show that a rational function f (z) = p(z)/q(z) that is one-to-one
must be a Möbius transformation. For example (I’m not sure if
this is the simplest argument): argue that if z0 is a complex number
such that q(z0 ) 6= 0 and such that p0 (z0 )q(z0 ) − p(z0 )q 0 (z0 ) 6= 0, and
w0 = f (z0 ), then the equation f (z) = w0 must have more than one
solution in z , unless p(z), q(z) are linear functions.
(e) Prove that set of entire functions f : C → C that are one-to-one
and onto is precisely L \ K, the set of nonconstant linear functions.
105 PROBLEMS

Hint. This is a slightly more advanced problem since it relies on

both the Casorati-Weierstrass theorem and the open mapping the-
orem (although a more elementary solution may exist). See the
guidance for exercise 14 on page 105 of [Stein-Shakarchi].

Remarks. Given a region Ω ⊂ C, or more generally a Riemann surface

Σ, complex analysts are interested in understanding the structure of
its set of holomorphic functions (C-valued holomorphic functions on Σ);
its set of meromorphic functions (C b -valued holomorphic functions on
Σ); and its set of holomorphic automorphisms (holomorphic, one-to-one
and onto mappings from Σ to itself). Although we won’t get into the
general theory of Riemann surfaces, once one defines these concepts
it easy to see that the above exercises essentially prove the following
conceptually important results:

(i) The constant functions are the only holomorphic functions on C

b.
(ii) The rational functions are the meromorphic functions on C
b.
(iii) The nonconstant linear functions are the holomorphic automor-
phisms of C.
(iv) The nonconstant Möbius transformations are the holomorphic au-
tomorphisms of C
b.

Another related result that is not very difficult to prove is:

(v) The holomorphic automorphisms of the upper half-plane H = {z :

Im z > 0} are the Möbius transformations z 7→ az+b
cz+d
with a, b, c, d ∈
R and ad − bc > 0. (Try to prove that any such map is indeed an au-
tomorphism of H; the reverse implication that all automorphisms
of H are of this form is a bit more difficult and requires a result
known as the Schwarz lemma.)
Note that the set of holomorphic functions on C (a.k.a. entire func-
tions) and the set of meromorphic functions on C are much larger
families of functions that do not have such a simple description
as the functions in the relatively small families L, P, R, M. This is
related to the fact that C is a non-compact Riemann surface.

25. The partial fraction expansion of the cotangent function. In this

multipart question you are asked to prove a well-known infinite sum-
mation identity (equation (12) below), known as the partial fraction ex-
pansion of the cotangent function. In the next problem we will explore
some of the consequences of this important result.
106 PROBLEMS

(a) Let z ∈ C \ Z. Use the residue theorem to evaluate the contour

integral I
π cot(πw)
IN := dw
γN (w + z)2
over the contour γN going in the positive direction around the rectangle
with the four vertices (±(N + 1/2), ±N ). Take the limit as N → ∞ to
deduce the identity
∞
π2 X 1
= (z ∈ C \ Z). (11)
(sin πz)2 n=−∞ (z + n)2

Guidance. This is not a trivial exercise, but is not very difficult when
broken down into the following elementary substeps:

i. Start by identifying the location of the singularities of the func-

π cot(πw)
tion w 7→ fz (w) = (w+z)2 (considered as a function of w for a fixed
z whose value is not an integer), and their residues. This provides
some good practice with residue computations.
ii. Use the residue theorem to obtain an expression for the con-
tour integral IN defined above.
iii. Separately, obtain estimates for IN that can be used to show
that IN → 0 as N → ∞. Specifically, show using elementary ma-
nipulations that

| sin(x + iy)|2 = sin2 x + sinh2 y, | cos(x + iy)|2 = cos2 x + sinh2 y,

use this to conclude that when x = π(N + 1/2) and y is arbitrary,

sinh2 y
| cot(x + iy)| = ≤ 1,
1 + sinh2 y
and that when y = N and x is arbitrary,
1 + sinh2 N
| cot(x + iy)| ≤ ≤ 2 (if N > 10);
sinh2 N
then use these estimates to bound the integral.
iv. By comparing the two results about IN , deduce (11).

(b) Integrate the identity (11) to deduce (using some additional fairly
easy reasoning) the formulas
N ∞
X 1 1 X 2z
π cot(πz) = lim = + (z ∈ C \ Z). (12)
N →∞
n=−N
z+n z n=1 z 2 − n2
107 PROBLEMS

26. Consequences of the partial fraction expansion of the cotangent

function.

(a) Show that (12) implies the following infinite-product representa-

tion for the sine function:
∞ 
z2
Y 
sin(πz) = πz 1− 2 (z ∈ C). (13)
n=1
n

Note that the function on the right-hand side is (or can be easily
checked to be) an entire function of z with a simple zero at any
integer z = n ∈ Z, and whose Taylor expansion around z = 0 starts
with πz + O(z 3 ); thus it is a natural guess for an infinite product
expansion of sin(πz), although the fact that this guess is correct is
far from obvious; for example one can multiply the right-hand side
by an arbitrary function of the form eg(z) and still have an entire
function with the same set of zeros.
Hint. Compute the logarithmic derivatives of both sides of (13).
You may want to review some basic properties of infinite products,
as discussed for example on pages 140–142 of [Stein-Shakarchi].
(Spoiler alert: pages 142–144 contain a solution to this subexer-
cise, starting with an independent proof of (12) and proceeding
with a derivation of (13) along the same lines as I described above.)
(b) By specializing the value of z in (13) to an appropriate specific
value, obtain the following infinite product formula for π , known
as Wallis’ product (first proved by John Wallis in 1655):

π 2 2 4 4 6 6 8 8
= · · · · · · · ··· .
2 1 3 3 5 5 7 7 9
(c) By comparing the first terms in the Taylor expansion around z = 0
of both sides of (13), derive the well-known identities
∞ ∞
X 1 π2 X 1 π4
2
= , 4
= .
n=1
n 6 n=1
n 90

(d) More generally, one can use (13), or more conveniently (12), to
obtain closed formulas for all the series
∞
X 1 1 1 1
ζ(2k) = (k = 1, 2, . . .) = 1 + + + + ...,
n=1
n2k 22k 32k 42k
108 PROBLEMS

that is, the special values of the Riemann zeta function ζ(s) =
P∞ 1
n=1 ns at the positive even integers. To see this, first, rewrite
(12) as
∞  
1 X 1 1
π cot(πz) = + − (z ∈ C \ Z). (14)
z n=−∞ z + n n
n6=0

Expand both sides of (14) in a Taylor series around z = 0, making

use of the expansion
∞
z z  X B
2n 2n
coth = z
2 2 n=0
(2n)!
we proved in an earlier homework exercise (where (Bn )∞
n=0 are the
Bernoulli numbers). Compare coefficients and simplify to get the
formula
(−1)k+1 (2π)2k
ζ(2k) = B2k .
2(2k)!
1 1
For example, using the first few values B2 = 6
, B4 = − 30 , B6 =
1 1
42
, B8 = − 30 , we get
∞
X 1 π2
ζ(2) = = ,
n=1
n2 6
∞
X 1 π4
ζ(4) = 4
= ,
n=1
n 90
∞
X 1 π6
ζ(6) = = ,
n=1
n6 945
∞
X 1 π8
ζ(8) = = ,
n=1
n8 9450
where of course the first two values coincide with those found ear-
lier.
(e) Show that ζ(2k) = 1 + O(2−2k ) as k → ∞, and deduce that the
asymptotic behavior of the Bernoulli numbers is given by
2(2k)!
B2k = (1 + O(2−2k ))(−1)k+1 , k → ∞.
(2π)2k
Note that this is consistent with our earlier (and much weaker)
result that  1/2k
 B2k  1
lim sup   = .
k→∞ (2k)!  2π
109 PROBLEMS

27. Let f (z) = p(z)/q(z) be a rational function such that deg q ≥ deg p + 2
(where deg p denotes the degree of a polynomial p). Prove that the sum
of the residues of f (z) over all its poles is equal to 0.

28. (A generalization of the result from problem 21) If p(z) = an z n +an−1 z n−1 +
. . . + . . . + a0 is a polynomial of degree n such that for some 0 ≤ k ≤ n
we have X
|ak | > |aj |,
0≤j≤n
j6=k

prove that p(z) has exactly k zeros (counting multiplicities) in the unit
disk |z| < 1.

29. Suggested reading: go to the Mathematics Stack Exchange website

(https://math.stackexchange.com) and enter "Rouche" into the search
box, to get an amusing list of questions and exercises involving appli-
cations of Rouché’s theorem to count zeros of polynomials and other
analytic functions.

30. Show how Rouché’s theorem can be used to give yet another proof
of the fundamental theorem of algebra. This proof is one way to make
precise the intuitively compelling “topological” proof idea we discussed
at the beginning of the course.

31. (a) Draw a simply-connected region Ω ⊂ C such that 0 ∈ / Ω, 1, 2 ∈ Ω,

and such that there exists a branch F (z) of the logarithm function on Ω
satisfying
F (1) = 0, F (2) = log 2 + 2πi
(where log 2 = 0.69314 . . . is the ordinary logarithm of 2 in the usual
sense of real analysis).

(b) More generally, let k ∈ Z. If we were to replace the above condition

F (2) = log 2+2πi with the more general condition F (2) = log 2+2πik but
keep all the other conditions, would an appropriate simply-connected
region Ω = Ω(k) exist to make that possible? If so, what would this
region look like, roughly, as a function of k ?

32. Prove the following properties satisfied by the gamma function:

i. Values at half-integers:

1

(2n)! √
Γ n+ 2
= π (n = 0, 1, 2, . . .).
4n n!
110 PROBLEMS

ii. The duplication formula:

√
Γ(s)Γ(s + 1/2) = 21−2s πΓ(2s).

iii.* The multiplication theorem:

Γ s Γ s + k1 Γ s + k2 · · · Γ s + k−1
= (2π)(k−1)/2 k 1/2−ks Γ(ks).





k

33. For n ≥ 1, let Vn denote the volume of the unit ball in Rn . By evaluating
the n-dimensional integral
ZZ Z n
!
X
An = ... exp − 12 x2j dx1 dx2 . . . dxn
Rn j=1

in two ways, prove the well-known formula

π n/2
Vn =
.
Γ n2 + 1

Note. This problem requires applying a small amount of geometric in-

tuition (or, alternatively, having some technical knowledge of spherical
coordinates in Rn ). The solution can be found on this Wikipedia page.

34. The beta function is a function B(s, t) of two complex variables, defined
for Re(s), Re(t) > 0 by
Z 1
B(s, t) = xs−1 (1 − x)t−1 dx.
0

(a) (Warm-up) Convince yourself that the improper integral defining

B(s, t) converges if and only if Re(s), Re(t) > 0.
(b) Show that B(s, t) can be expressed in terms of the gamma function
as
Γ(s)Γ(t)
B(s, t) = .
Γ(s + t)
Hint. Start by writing Γ(s)Γ(t) as a double integral on the positive
quadrant [0, ∞)2 of R2 (with integration variables, say, x and y );
then make the change of variables u = x + y , v = x/(x + y) and
use the change of variables formula for two-dimensional integrals
to show that the integral evaluates as Γ(s + t)B(s, t).
Remark. Note the similarity of the
identity relating the gamma
and beta functions to the formula nk = k!(n−k)!
n
; indeed, using the
111 PROBLEMS

relation Γ(m+1) = m! and the functional equation Γ(s+1) = s Γ(s),

we see using the above relation that for nonnegative, integer-
valued arguments we have
 
−1 nm Γ(n + m + 1) nm n + m
B(n, m) = · = .
n + m Γ(n + 1)Γ(m + 1) n+m n
nm
In other words, except for the correction factor n+m , the inverse
of the beta function can be thought of as a natural extention of
binomial coefficients to real-valued arguments.

35. The digamma function ψ(s) is the logarithmic derivative

Γ0 (s)
ψ(s) =
Γ(s)
of the gamma function, also considered as a somewhat important spe-
cial function in its own right.

(a) Show that ψ(s) has the convergent series expansions

∞
1 X s
ψ(s) = −γ − +
s n=1 n(n + s)
∞  
X 1 1
= −γ + − (s 6= 0, −1, −2, . . .).
n=0
n+1 n+s

where γ is the Euler-Mascheroni constant.

(b) Equivalently, show that ψ(s) can be expressed as
n
!
X 1
ψ(s) = − lim − log n .
n→∞
k=1
k+s

(c) Show that ψ(s) satisfies the functional equation

1
ψ(s + 1) = ψ(s) + .
s
(d) Show that
n
X 1
ψ(n + 1) = −γ + (n = 0, 1, 2, . . .).
k=1
k

That is, ψ(x) + γ can be thought of as extending the definition of

Pn
the harmonic numbers Hn = k=1 k1 to non-integer arguments.
112 PROBLEMS

(e) Show that ψ(s) satisfies the reflection formula

ψ(1 − s) − ψ(s) = π cot(πs).

(f)* Here is an amusing application of the digamma function. Consider

the sequence of polynomials

Pn (x) = x(x − 1) . . . (x − n) (n = 0, 1, 2, . . .)

and their derivatives

Qn (x) = Pn0 (x).
Note that by Rolle’s theorem, Qn (x) has precisely one root in each
interval (k, k + 1) for 0 ≤ k ≤ n − 1. Denote this root by k + αn,k ,
so that the numbers αn,k (the fractional parts of the roots of Qn (x))
are in (0, 1).
A curious phenomenon can now be observed by plotting the points
αn,k , k = 0, . . . , n − 1 numerically, say for n = 50 (Figure 11(a)). It
appears that for large n they approximate some smooth limiting
curve. This is correct, and in fact the following precise statement
can be proved.
Theorem. Let t ∈ (0, 1). Let k = k(n) be a sequence such that
0 ≤ k(n) ≤ n − 1, k(n) → ∞ as n → ∞, n − k(n) → ∞ as n → ∞,
and k(n)/n → t as n → ∞. Then we have
  
1 1 1−t
lim αn,k(n) = R(t) := arccot log .
n→∞ π π t

In the above formula, arccot(·) refers to the branch of the inverse

cotangent function taking values between 0 and π . The limiting
function R(t) is shown in Figure 11(b).

Prove this.

Guidance. Take the logarithmic derivative of Pn (x) to see when

the equation Qn (x)/Pn (x) = 0 (which is equivalent to Qn (x) = 0)
holds. This will give an equation with a sum of terms. Find a
way to separate them into two groups such that the sum in each
group can be related, in an asymptotic sense as n → ∞, to the
digamma function evaluated at a certain argument (using prop-
erty (b) above). Take the limit as n → ∞, then simplify using the
reflection formula (part (c)).
113 PROBLEMS

1.0 1.0

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 10 20 30 40 50 0.0 0.2 0.4 0.6 0.8 1.0

(a) (b)
1.0

0.8

1 2 3 4 5 6 7
0.6

-200
0.4

-400
0.2

-600
0.0 0.2 0.4 0.6 0.8 1.0

(c) (d)

Figure 11: (a) A plot of the fractional parts of the roots of Qn (x) for n = 50.
(b) The limiting function R(t). (c) The two previous plots combined. (d) The
polynomial P7 (x). Note that the roots of Q7 (x) correspond to the local minima
and maxima of P7 (x), which are highlighted.

36. Given two integrable functions f, g : R → C (of a real variable), their

convolution is the new function h = f ∗ g defined by the formula
Z ∞
h(x) = (f ∗ g)(x) = f (t)g(x − t) dt (x ∈ R).
−∞

As you might be aware, the convolution operation is extremely impor-

tant in harmonic analysis, since it corresponds to a simple multiplica-
tion operation in the Fourier domain; in probability theory, where it
corresponds to the addition of independent random variables; and in
many other areas of mathematics, science and engineering.

For α > 0 define the gamma density with parameter α, denoted γα :

R → R, to be the function
1 −x α−1
γα (x) = e x 1[0,∞) (x) (x ∈ R)
Γ(α)

(where 1A (x) denotes the characteristic function of a set A ⊂ R, equal

to 1 on the set and 0 outside it). Note that γα (x) is the nonnegative
114 PROBLEMS

1.0

0.8 α=1

0.6

0.4 α=2
α=3
α=4
0.2 α=5

2 4 6 8 10 12

Figure 12: The gamma densities γα (x) for α = 1, 2, 3, 4, 5.

function whose integral equals Γ(α), except that it is divided by Γ(α)

so that it becomes a probability density function. See Figure 12 for an
illustration.

Show that for each α, β > 0 we have

γα ∗ γβ = γα+β .
That is, the family of density functions (γα )α>0 is closed under the con-
volution operation. This fact is one of the reasons why the family of
gamma densities plays a very important role in probability theory and
appears in many real-life applications.

37. (a) Show that the Laurent expansion of Γ(s) around s = 0 is of the form
1
Γ(s) = − γ + O(s)
s
(where γ is the Euler-Mascheroni constant). If you’re feeling especially
energetic, derive the more detailed expansion
γ 2 π2
 
1
Γ(s) = − γ + + s + O(s2 )
s 2 12
and proceed to derive (by hand, or if you prefer using a symbolic math
software application such as SageMath or Mathematica) as many addi-
tional terms in the expansion as you have the patience to do.

(b) Show that the Laurent expansion of ζ(s) around s = 1 is of the form
1
ζ(s) = + γ + O(s − 1).
s−1
115 PROBLEMS

38. Show that the symmetric version of the functional equation for the zeta
function
ζ ∗ (1 − s) = ζ ∗ (s),
where ζ ∗ (s) = π −s/2 Γ(s/2)ζ(s), can be rewritten in the equivalent form
 πs 
ζ(s) = 2s π s−1 sin Γ(1 − s)ζ(1 − s).
2

Γ0 (s)
39. Show that the Taylor expansion of the digamma function ψ(s) = Γ(s)
around s = 1 is given by
∞
X
ψ(s) = −γ − (−1)n−1 ζ(n + 1)(s − 1)n (|s − 1| < 1).
n=1

40. Define a function D(s) of a complex variable s by

∞
X (−1)n−1 1 1 1
D(s) = =1− + − + ....
n=1
ns 2s 3s 4s

(a) Prove that the series defining D(s) converges uniformly on any
half-plane of the form Re(s) ≥ α where α > 0, and conclude that
D(s) is defined and holomorphic in the half-plane Re(s) > 0.
(b) Show that D(s) is related to the Riemann zeta function by the for-
mula
D(s) = (1 − 21−s )ζ(s) (Re(s) > 1).
(c) Using this relation, deduce a new proof that the zeta function can
be analytically continued to a meromorphic function on Re(s) > 0
that has a simple pole at s = 1 with residue 1 and is holomorphic
everywhere else in the region.
P
41. Let ψ(x) = pk ≤x log p denote von Mangoldt’s weighted prime count-
ing function. Show that ψ(n) = log lcm(1, 2, . . . , n), where for integers
a1 , . . . , ak , the notation lcm(a1 , . . . , ak ) denotes the least common multi-
ple of a1 , . . . , ak .

Note that this implies that an equivalent formulation of the prime num-
ber theorem is the interesting statement that

lcm(1, . . . , n) = e(1+o(1))n as n → ∞.
116 PROBLEMS

42. (a) Prove that for all x ≥ 1,

Y 1
1 ≥ log x
p≤x
1− p

(where the product is over all prime numbers p that are ≤ x).

(b) Pass to the logarithm and deduce that for some constant K > 0 we
have the bound
X1
≥ log log x − K (x ≥ 1).
p≤x
p
1
P
That is, the harmonic series of primes p p diverges as log log x, in
contrast to the usual harmonic series which diverges as log x.

43. An alternative proof of the functional equation of the Jacobi

theta function. Recall that in Subsection 15.1 we defined the Jacobi
theta function by
∞
2t
X
ϑ(t) = e−πn (t > 0),
n=−∞

and showed that it satisfies the functional equation

√
 
1
ϑ = tϑ(t). (15)
t

(a) Use the residue theorem to evaluate the contour integral

2
e−πz t
I
dz,
γN e2πiz − 1

where γN is the rectangle with vertices ±(N +1/2)±i (with N a positive

integer), then take the limit as N → ∞ to derive the integral represen-
tation 2 2
∞−i ∞+i
e−πz t e−πz t
Z Z
ϑ(t) = dz − dz
−∞−i e2πiz − 1 −∞+i e2πiz − 1
for the function ϑ(t).

(b) In this representation, expand the factor (e2πiz − 1)−1 as a geometric

series in e−2πiz (for the first integral) and as a geometric series in e2πiz
(for the second integral). Evaluate the resulting infinite series, rigor-
ously justifying all steps, to obtain an alternative proof of the functional
equation (15).
117 PROBLEMS

44. (a) Reprove Theorem 36 (the “toy Riemann hypothesis” — the result
that the Riemann zeta function has no zeros on the line Re(s) = 1) by
considering the behavior of

ζ 0 (σ) ζ 0 (σ + it) ζ 0 (σ + 2it)

 
Y = Re −3 −4 −
ζ(σ) ζ(σ + it) ζ(σ + 2it)

for t ∈ R \ {0} fixed and σ & 1, instead of the quantity

X = log |ζ(σ)3 ζ(σ + it)4 ζ(σ + 2it)|.

Use the series expansion

∞
ζ(s) X
− = Λ(n)n−s ,
ζ(s) n=1

where Λ(n) is von Mangoldt’s function (equal to log p if n = pk is a prime

power, and 0 otherwise).

(b) Try to reprove the same theorem in yet a third way by considering

Z = log |ζ(σ)10 ζ(σ + it)15 ζ(σ + 2it)6 ζ(σ + 3it)|,

and attempting to repeat the argument involving expanding the loga-

rithm in a power series and deducing that Z ≥ 0. Does this give a proof
of the theorem? If not, what goes wrong?

Hint. (a + b)6 = a6 + 6a5 b + 10a4 b2 + 15a3 b3 + 10a2 b4 + 6ab5 + b6 .

45. Define arithmetic functions taking an integer argument n, as follows:

(
(−1)k if n = p1 p2 · · · pk is a product of k distinct primes,
µ(n) =
0 otherwise,
(the Möbius µ-function),
X
d(n) = 1, (the number of divisors function),
d|n
X
σ(n) = d, (the sum of divisors function),
d|n

φ(n) = #{1 ≤ k ≤ n − 1 : gcd(k, n) = 1}, (the Euler totient function),

(
log p if n = pk , p prime,
Λ(n) = (the von Mangoldt Λ-function).
0 otherwise,
118 PROBLEMS

We saw that the zeta function and its logarithmic derivative have the
Dirichlet series representations
∞
X
ζ(s) = n−s ,
n=1
0 ∞
ζ (s) X
− = Λ(n)n−s .
ζ(s) n=1
Use the Euler product formula for the zeta function or other elementary
manipulations to prove the following identities (valid for Re(s) > 1):
∞
X
0
ζ (s) = − log n · n−s ,
n=1
∞
1 X
= µ(n)n−s ,
ζ(s) n=1
∞
ζ(s) X
= |µ(n)|n−s .
ζ(2s) n=1
Other famous Dirichlet series representations you may want to think
about or look up are
∞
X
2
ζ(s) = d(n)n−s ,
n=1
∞
ζ(s − 1) X
= φ(n)n−s ,
ζ(s) n=1
∞
X
ζ(s)ζ(s − 1) = σ(n)n−s .
n=1

46. Sendov’s conjecture, an elementary statement in complex analysis pro-

posed by the mathematician Blagovest Sendov in 1959 and still open
today, is the claim that if p(z) = (z − z1 ) . . . (z − zn ) is a complex poly-
nomial whose roots zj , j = 1, . . . , n all lie in the closed unit disc |z| ≤ 1,
then for each root zj there is a root α of the derivative p0 (z) for which
|zj − α| ≤ 1.
(a) Prove the conjecture for the case n = 2 of quadratic polynomials.
(b) Prove that in the inequality |zj − α| ≤ 1, if the number 1 is replaced
by any smaller number then the claim is false.
(c) Prove the conjecture for the case n = 3 of cubic polynomials. (This
is not a trivial result; for one possible proof, see the paper [2].
119 PROBLEMS

Suggested topics for course projects

1. Complex analysis and self-avoiding random walks. Let an denote
the number of self-avoiding walks of length n in the square lattice Z2
(that is, lattice paths with n steps) starting from (0, 0). It is known that
1/n
an −−−→ µ for some number µ, approximately equal to 2.638. No closed
n→∞
form formula is known, or even suspected to exist, for µ. However, for
the hexagonal lattice, a precise analogous result is known: if bn denotes
p of self-avoiding walks of length n starting at the origin, then
the number
√
bn −−−→ 2+ 2 ≈ 1.8477. This is proved in the paper [3]. The proof is
n→∞
elementary and uses ideas from complex analysis in a crucial way. See
this Wikipedia article for more details.

2. Complex analysis and the art of M. C. Escher. See Fig. 1 on page 3

and the references [8], [10].

3. Elliptic functions and the characterization of meromorphic maps

on the complex torus. In Problem 24 we were interested in classify-
ing the holomorphic functions f : M → N when M, N are Riemann
surfaces; we answered this question for several interesting pairs of
surfaces (such as M = C, N = Ĉ , etc.). The case when N is the Rie-
mann sphere Ĉ and M is a complex torus C/Λ where Λ = Z + τ Z for
some complex number τ ∈ C \ R is an especially interesting example
of such a question: in that case the relevant family turns out to be the
family of elliptic functions (also known as doubly periodic functions).
Some places where you can read about this topic are [11, Ch. 9] and [1,
Ch. 1].

4. Sendov’s conjecture. This conjecture is described in Problem 46, and

can be a suitable topic for a project.

5. Advanced topic in analytic combinatorics. In the area of analytic

combinatorics, one uses asymptotic methods in complex analysis (as
well as real and harmonic analysis) to study the asymptotic behavior of
integer sequences (or more generally integer arrays, probability distri-
butions and other quantities) arising in combinatorics, similarly to the
discussion in Section 17. The book [5] is an excellent reference for this
subject (see also [6]), and provides many topics suitable for a project.

6. The Jordan curve theorem. The Jordan curve theorem states that a
continuous, simple planar curve separates the plane into two connected
components, precisely one of which is unbounded. The theorem plays
120 PROBLEMS

an important role in complex analysis, discussed in Appendix B of [11].

Despite its intuitive nature it is not trivial to prove, and has several
different proofs described in different sources, for example [11, Ap-
pendix B], [9], and [12]. A project could present one of those proofs,
and discuss additional related topics such as the significance of the
theorem for complex analysis, various generalizations of it (for exam-
ple its higher-dimensional analogues), and related theorems and open
problems in topology. For more details, see this Wikipedia article and
this MathOverflow discussion.
121 PROBLEMS

References
[1] T. Apostol. Modular Functions and Dirichlet Series in Number Theory.
Springer, 1990.

[2] G. L. Cohen, G. H. Smith.. A simple verification of Ilieff’s conjecture for

polynomials with three zeros. Amer. Math. Monthly 95 (1988), 734–737.

[3] H. Duminil-Copin,
p S. Smirnov. The connective constant of the honeycomb
√
lattice equals 2+ 2. Ann. Math. 175 (2012), 1653–1665.

[4] H. M. Edwards. Riemann’s Zeta Function. Dover Publications, 2001.

[5] P. Flajolet, R. Sedgewick. Analytic Combinatorics. Cambridge University

Press, 2009.

[6] P. Flajolet, R. Sedgewick. Analytic Combinatorics — Online Course Ma-

terials. Online resource:
https://ac.cs.princeton.edu/online/. Accessed March 20, 2020.

[7] A. Ivić. The Riemann Zeta-Function: Theory and Applications. Dover Pub-
lications, 2003.

[8] B. de Smit, H. W. Lenstra Jr. Artful mathematics: the heritage of M.C.

Escher. Notices Amer. Math. Soc. 50 (2003), 446–457.

[9] R. Maehara. The Jordan curve theorem via the Brouwer fixed point theo-
rem.. Amer. Math. Monthly 91 (1984), 641–643.

[10] D. Schattsschneider. The mathematical side of M. C. Escher. Notices

Amer. Math. Soc. 57 (2010), 706–718.

[11] E. M. Stein, R. Shakarchi, Complex Analysis. Princeton University

Press, 2003.

[12] H. Tverberg. A proof of the Jordan curve theorem. Bull. London

Math. Soc. 12 (1980), 34–38.

[13] D. Zagier. Newman’s short proof of the prime number theorem. Amer.
Math. Monthly 104 (1997), 705–708.
Index
absolutely convergent, 13 closed
H curve, 18
Airy function, 79 , 18
an -formula, 16 closed disc, 39
analytic continuation, 32 compact set, 31
analytic function, 7 compactification
geometric view, 7 one point, 39
angle preserving map, 8 complex differentiable, 7
arc length integral complex dynamics, 3
definition, 18 complex-valued function, 17
iϕ
argument: arg w ; w = re , arg w = ϕ complicated integrals, 2
, 42 conformal map, 8
argument principle, 41 conformal maps, 2
Arg z , 50 conformality lemma, 11
asymptotic analysis, 79 constant function
criterion, 21
Basel problem, 60 contour integral, 16
Bernoulli numbers, 60 definition, 18
contour integrals
calculus
Fundamental Theorem of Calcu-
fundamental theorem, 18
lus, 19
Cardano
properties, 19
cubic equation, 1 0
contours C , C+ , C− , C− , 76
Cardy-Smirnov formula, 2
convergence
Casorati-Weierstrass
uniform, 28
theorem, 40
convergence disc, 15
Cauchy inequalities, 29
convergent
Cauchy’s integral formula, 26
absolutely, 13
extended, 28
covering, 31
Cauchy’s theorem, 21
cubic equation, 1
consequences, 26
curve
for a disc, 25
closed, 18
general version, 49
equivalence class, 16
toy contour (?), 25
equivalent, 16
Cauchy-Riemann equations, 9
length len γ , 18
central binomial coefficient, 83
parametrization, 18
change of variables, 18
piecewise continuously differentiable,
clopen
17
connected and open set, 31

122
123 INDEX

derivative, 7 exercise 20, 87

logarithmic, 41
properties, 8 Fourier transform, 52
differentiable fundamental theorem
infinitely often, 16 calculus, 18
differential form dw/w, 41 Fundamental Theorem of Algebra, 3
differentiation 3 proofs, 3
termwise, 14 analytic proof, 4
disc Liouville’s proof, 6
D(z, h), 21 topological proof, 5
closed, 39
gamma function
convergence, 15
Γz , 52
double keyhole contour, 33
analytical continuation, 54
dw/w as improper intetral, 53
meaning, 42
axiomatic definition, properties, ...,
electrical networks, 2 52
M.C. Escher, 2 equivalent definitions, 52
essential singularity, 40 interpolating factorials, 52
Euler product formula, 60 reflection formula, 58
Euler-Maclaurin summation, 68 Stirling’s formula, 87
Eulers gamma function, 52 geometric series, 29
exercise 1, 5 Goursat’s theorem, 22
exercise 2, 8 rectangles, 24
exercise 3, 11
Hardy-Ramanjuan formula, 2
exercise 4, 14
Hardy-Ramanjuan’s formula, 79
exercise 5, 19
harmonic function, 12
exercise 6, 19
Heine-Borel-property, 31
exercise 7, 37
holomorphic
exercise 8, 45
at ∞, 39
exercise 9, 46
holomorphic function, 7
exercise 10, 49
mean value property, 27
exercise 11, 51
holomorphic implies analytic, 29
exercise 12, 51
homotopic curves, 46
exercise 13, 51
homotopy, 46
exercise 14, 51
homotopy invariance of integrals, 47
exercise 15, 65
exercise 16, 71 index of a curve, 42
exercise 17, 84 inequality, 15
exercise 18, 84 infinite products
exercise 19, 86 what you need to know, 56
124 INDEX

infinte products open mapping theorem, 45

convergence criteria, 56 order
inner product, 8 of pole, zero, at ∞, 39
integral orthogonal level curves, 11
homotopy-invariant, 47
isolated point, 32 partial differential equation, 9
isolated zero, 32 physics problems, 2
π(x) ∼ logx x , 72
Jacobi ϑ-function, 65 Poisson summation formula, 63
Jacobian matrix, 13 polar decomposition, 7
pole
keyhole contour, 26 of order m, 36
representation Lemma, 36
Λ(n), 72
simple, 36
Laplace’s equation, 12
polynomial equation, 1
Laplace’s method, 79
polynomial root
limit points
existence, 3
common, 32
Fundamental Theorem of Algebra,
lim sup, 14
3
line integral
power function, 51
two types, 17
power series, 13
logarithm function, 50
divergent, 14
logarithmic derivative, 41
prime counting functions
of a product, 41
weighted, 72
Log z , 50
prime number theorem, 1, 62
log0 z , 50P
equivalences, 72
− ζζ(s)
(s)
= ∞ −s
n=1 Λ(n)n , 74
history, seealso [4], 71
M.C. Escher, 2 Newman’s proof, 72
Mandelbrot set, 4 prime number theory
Mangoldt function Λ(n), 72 proof, 78
maximum modulus principle, 46 primitive
mean value property, 27 existence criteria, 20
Mellin transform, 52 existence on a disc, 25
meromorphic function, 38 of holomorphic function, 49
Morera’s theorem, 21, 22 synonym of anti-derivative, 49
proof, 29 with log-properties, 50
primitive, anti-derivative, 19
Newman’s tauberian theorem, 75 principal branch
n! of logarithm, 50
asymptotic relation, 82 principal part, 36
n-th root, 51 ψ(x) ∼ x, 72
125 INDEX

punctured neighborhood, 32 classification, 40

removable, 33
radius of convergence, 14 removable, pole, essential, 40
region slit complex plane, 50
open connected set, 19 sphere packing, 2
simply-connected open set, 22 stationary phase method, 79
regularity theorem, 28 steepest descent method, 79
removable singularity, 33 stereographic projection, 39
theorem, 33 Stirling’s formula, 1, 79
residue of ..., 36 gamma function, 87
residue theorem
extended version, 37 Taylor’s formula, 16
general version, 38 termwise differentiation, 14
simple version, 37 textbook
Riemann compromises, 49
hypothesis, 59, 62 toy contour
zeta function, 59 strange object, 38
Riemann sphere, 39 toy example, 33
Riemann surface, 40
Rouché uniform convergence, 28
dog and pole example, 44 uniform convergent
Rouché’s theorem, 43 derivates, 34
applications, 45 on compact subset, 34
b , 39
C
1
H f (w) vector field, 17
2πi γ w
dw, 42
walking the dog
S2 winding number, 44
geometric view, 39 weighted sum, 29
saddle point method, 79 winding number, 42
Schrödinger equation, 2 walking the dog, 44
sequence
of complex numbers, 13 zero
of holomorphic functions, 16 isolated, 32
series of holomorphic function, 35
absolutely convergent, 14 of order ..., 35
dominated, 14 simple, 35
geometric, 14 zeta function
simple pole, 36 as contour-integral, 61
simple zero, 35 axiomatic definition, properties, 60
simply connected region, 47 Basel problem, 60
singularity Bernoulli numbers, 60
126 INDEX

Euler’s product formula, 60

Euler-Maclaurin summation, 68
functional equation, 61, 67
in prime counting, 62
location of zeroes, 60
Mellin transformed, 61
no zeroes with Re(s) = 1, 69
pole, 67