Advanced Complex Analysis

Course Notes — Harvard University — Math 213a

Fall 2000, 2006, 2010
C. McMullen

January 22, 2017

Contents
1 Basic complex analysis . . . . . . . . . . . . . . . . . . . . . . 1
2 The simply-connected Riemann surfaces . . . . . . . . . . . . 27
3 Entire and meromorphic functions . . . . . . . . . . . . . . . 39
4 Conformal mapping . . . . . . . . . . . . . . . . . . . . . . . 56
5 Elliptic functions and elliptic curves . . . . . . . . . . . . . . 79
 Forward

Complex analysis is a nexus for many mathematical fields, including:

1. Algebra (theory of fields and equations);

2. Algebraic geometry and complex manifolds;

3. Geometry (Platonic solids; flat tori; hyperbolic manifolds of dimen-

sions two and three);

4. Lie groups, discrete subgroups and homogeneous spaces (e.g. H/ SL2 (Z);

5. Dynamics (iterated rational maps);

6. Number theory and automorphic forms (elliptic functions, zeta func-

tions);

7. Theory of Riemann surfaces (Teichmüller theory, curves and their Ja-

cobians);

8. Several complex variables and complex manifolds;

9. Real analysis and PDE (harmonic functions, elliptic equations and

distributions).
This course covers some basic material on both the geometric and analytic
aspects of complex analysis in one variable.
Prerequisites: Background in real analysis and basic differential topology
(such as covering spaces and differential forms), and a first course in complex
analysis.

1 Basic complex analysis

We begin with a quick review of elementary facts about the complex plane
and analytic functions.
Some notation. The complex numbers will be denoted C. We let ∆, H
and Cb denote the unit disk |z| < 1, the upper half plane Im(z) > 0, and the
Riemann sphere C ∪ {∞}. We write S 1 (r) for the circle |z| = r, and S 1 for
the unit circle, each oriented counter-clockwise. We also set ∆∗ = ∆ − {0}
and C∗ = C − {0}.
Algebra of complex numbers. The complex numbers are formally de-
fined as the field C = R[i], where i2 = −1. They are represented in the

1
Euclidean plane by z = (x, y) = x + iy. There are two square-roots of −1
in C; the number i is the one with positive imaginary part.
An important role is played by the Galois involution z 7→ z. We define
2
|z| = N (z) =√zz = x2 + y 2 . (Compare the case of a real quadratic field,
where N (a + b d) = a2 − db2 gives an indefinite form.) Compatibility of |z|
with the Euclidean metric justifies the identification of C and R2 . We also
see that z is a field: 1/z = z/|z|.
It is also convenient to describe complex numbers by polar coordinates

z = [r, θ] = r(cos θ + i sin θ).

Here r = |z| and θ = arg z ∈ R/2πZ. (The multivaluedness of arg z requires

care but is also the ultimate source of powerful results such as Cauchy’s
integral formula.) We then have

[r1 , θ1 ][r2 , θ2 ] = [r1 r2 , θ1 + θ2 ].

In particular, the linear maps f (z) = az + b, a 6= 0, of C to itself, preserve

angles and orientations.
This formula should be proved geometrically: in fact, it is a consequence
of the formula |ab| = |a||b| and properties of similar triangles. It can then
be used to derive the addition formulas for sine and cosine (in Ahflors the
reverse logic is applied).
Algebraic closure. A critical feature of the complex numbers is that
they are algebraically closed; every polynomial has a root. (A proof will be
reviewed below).
Classically, the complex numbers were introducing in the course of solv-
ing real cubic equations. Staring with x3 + ax + b = 0 one can make a
Tschirnhaus transformation soPa = 0.PThis is done by introducing a new
variable y = cx2 + d such that yi = yi2 = 0; even when a and b are real,
it may be necessary to choose c complex (the discriminant of the equation
for c is 27b2 + 4a3 .) It is negative when the cubic has only one real root;
this can be checked by looking at the product of the values of the cubic at
its max and min.
Polynomials and rational functions. Using addition Pn andn multiplication
we obtain naturally the polynomial functions f (z) = 0 an z : C → C. The
ring of polynomials C[z] is an integral domain and a unique factorization
domain, since C is a field. Indeed, since C is algebraically closed, fact every
polynomial factors into linear terms.
It is useful to add the allowed value ∞ to obtain the Riemann sphere
b
C = C ∪ {∞}. Then rational functions (ratios f (z) = p(z)/q(z) of relatively

2
prime polynomials, with the denominator not identically zero) determine
rational maps f : C → C. The rational functions C(z) are the same as the
field of fractions for the domain C[z]. We set f (z) = ∞ if q(z) = 0; these
points are called the poles of f .
Analytic functions. Let U be an open set in C and f : U → C a function.
We say f is analytic if
f (z + t) − f (z)
f ′ (z) = lim
t→0 t
exists for all z ∈ U . It is crucial here that t approaches zero through arbitrary
values in C. Remarkably, this condition implies that f is a smooth (C ∞ )
function. For example, polynomials are analytic, as are rational functions
away from their poles.
Note that any real linear function φ : C → C has the form φ(v) = av+bv.
The condition of analytic says that Dfz (v) = f ′ (z)v; in other words, the v
part is absent.
To make this point systematically, for a general C 1 function F : U → C
we define
   
dF 1 dF 1 dF dF 1 dF 1 dF
= + and = − ·
dz 2 dx i dy dz 2 dx i dy
We then have
dF DF
DFz (v) = v+ v.
dz dz
We can also write complex-valued 1-form dF as
dF dF
dF = ∂F + ∂F = dz + dz
dz dz
Thus F is analytic iff ∂F = 0; these are the Cauchy-Riemann equations.
We note that (d/dz)z n = nz n−1 ; a polynomial p(z, z) behaves as if these
variables are independent.
Sources of analytic functions.
Algebraic functions. Beyond the rational and polynomial functions, √ the
analytic functions include algebraic functions suchP that f (z) = z 2 + 1.
A general algebraic function f (z) satisfies P (f ) = N n
0 an (z)f (z) = 0 for
some rational functions an (z); these arise, at least formally, when one forms
algebraic extension of C(z). Such functions are generally multivalued, so we
must choose a particular branch to obtain an analytic function.
Differential equations. Analytic functions also arise when one solves
differential equations. Even equations with constant coefficients, like y ′′ +y =

3
0, can give rise to transcendental functionsRsuch as sin(z), cos(z) and ez . A
special case of course is integration. While (x2 + ax + b) 1/2
R 3 dx can be 1/2given
explicitly in terms of trigonometric functions, already (x + ax + b) dx
leads one into elliptic functions; and higher degree polynomials lead one to
hyperelliptic surfaces of higher genus.
Power series. Analytic
P functions can be given concretely, locally, by
power series such as an z n . Conversely,
P suitable coefficients determine
analytic functions; for example, ez = z n /n!.
Riemann surfaces and automorphy. A third natural source of complex
analytic functions is functions that satisfy invariant properties such as f (z +
λ) = f (z) for all λ ∈ Λ, a lattice in C; or f (g(z)) = f (z) for all g ∈ Γ ⊂
Aut(H).
The elliptic modular functions f : H → C with have the property that
f (z)
= f (z + 1) = f (1/z), and hence f ((az + b)/(cz + d)) = f (z) for all
a b ∈ SL (Z).
c d 2
Geometric function theory. Finally a complex analytic function can be
specified by a domain U ⊂ C; we will see that for simply-connected domains
(other than C itself), there is an essentially unique analytic homeomorphism
f : ∆ → U . When U tiles H or C, this is related to automorphic functions;
and when ∂U consists of lines or circular arcs, one can also give a differential
equation for f .
Example. We can also define analytic functions by taking limits of poly-
nomials or other known functions. For example, consider the formula:
ez = lim(1 + z/n)n .
The triangle with vertices 0, 1 and 1 + iθ has a hypotenuse of length 1 +
O(1/n2 ) and an angle at 0 of θ + O(1/n2 ). Thus one finds geometrically that
zn = (1 + iθ/n)n satisfies |zn | → 1 and arg zn → θ; in other words,
eiθ = cos θ + i sin θ.
In particular, eπi = −1 (Euler).
Exponential and trigonometric functions in C. Here are some useful
facts about these familiar functions when extended to C:
| exp(z)| = exp Re z
cos(iz) = cosh(z)
sin(iz) = i sinh(z)
cos(x + iy) = cos(x) cosh(y) − i sin(x) sinh(y)
sin(x + iy) = sin(x) cosh(y) + i cos(x) sinh(y).

4
In particular, the apparent boundedness of sin(z) and cos(z) fails badly as
we move away from the real axis, while |ez | is actually very small in the
halfplane Re z ≪ 0.
Complex integration; Cauchy’s theorem. Now suppose U is a com-
pact, connected, smoothly bounded region in C, f : U → C is continuous
and f : U → C is analytic. We then have:

Theorem
R 1.1 (Cauchy) For any analytic function f : U → C, we have
∂U f (z) dz = 0.

Remark. It is critical to know the definition of such a path integral.

(ForR example, f (z) = 1 is analytic, its average over the circle is 1, and
yet S 1 1 dz = 0; why is this?)
If γ : [a, b] → C parameterizes an arc, then we define
Z Z b
f (z) dz = f (γ(t)) γ ′ (t) dt.
γ a

Alternatively, we choose a sequence of points z1 , . . . , zn close together along

γ, and then define
Z n−1
X
f (z) = lim f (zi )(zi+1 − zi ).
γ 1

(If the loop is closed, we choose zn = z1 ).

This should not be confused with the integral with respect to arclength:
Z X
f (z) |dz| = lim f (zi )|zi+1 − zi |.
γ

Note that the former depends on a choice of orientation of γ, while the latter
does not.
Proof of Cauchy’s formula: (i) observe that d(f dz) = (∂f )dz dz = 0
and apply Stokes’ theorem. (ii — Goursat). Cut the region U into small
Rsquares, observe that on these squares f (z) ≈ az + b, and use the fact that
∂U (az + b) dz = 0.

5
Aside: distributions. The first proof implicitly assumes f is C 1 , while
the second does not. (To see where C 1 is used, R suppose α = u dx + v dy
and dα = 0 on a square S. In the proof that ∂S α = 0, we integrate vdy
over the vertical sides of S and observe that this is the same as integrating
dv/dx dx dy over the square. But if α is not C 1 , we don’t know that dv/dx
is integrable.)
Notes. More on the Cauchy–Riemann equations and with minimal smooth-
ness assumptions can be found in [GM].
More generally,
R we say a distribution (e.g. an L1 function f ) is (weakly)
analytic if f ∂φ = 0 for every φ ∈ Cc∞ (U ). By convolution with a smooth
function (a mollifier), any weakly analytic function is a limit of C ∞ analytic
functions. We will see below that uniform limits of C ∞ analytic functions
are C ∞ , so even weakly analytic functions are actually smooth.
Cauchy’s integral formula: Differentiability and power series. Because
of Cauchy’s theorem, only one integral has to be explicitly evaluated in
complex analysis (hence the forgetability of the definition of the integral).
Namely, setting γ(t) = eit we find, for any r > 0,
Z
1
dz = 2πi.
S (r) z
1

By integrating between ∂U and a small loop around p ∈ U , we then obtain

Cauchy’s integral formula:
Theorem 1.2 For any p ∈ U we have
Z
1 f (z) dz
f (p) = ·
2πi ∂U z − p
Now the integrand depends on p only through the rational function
1/(z − p), which is infinitely differentiable. (It is the convolution of f |∂U
and 1/z.) So we conclude that f (p) itself is infinitely differentiable, indeed,
it is approximated by a sum of rational functions with poles on ∂U . In
particular, differentiating under the integral, we obtain:
Z
f (k) (p) 1 f (z) dz
= · (1.1)
k! 2πi ∂U (z − p)k+1
If d(p, ∂U ) = R, the length of ∂U is L and sup∂U |f | = M , then this gives
the bound:
f (k) (p) ML
|ak | = ≤ ·
k! 2πRk+1

6
 P
In particular, ak (z − p)k has radius of convergence at least R, since
lim sup |ak |1/k < 1/R. This suggests that f is represented by its power
series, and indeed this is the case:
P
Theorem 1.3 If f is analytic on B(p, R), then f (z) = ak (z − p)k on this
ball.

Proof. We can reduce to the case z ∈ B(0, 1). Then for w ∈ S 1 and fixed
z with |z| < 1, we have
1 1

= 1 + (z/w) + (z/w)2 + · · · ,
w−z w
converging uniformly on the circle |w| = 1. We then have:
Z Z
1 f (w) dw X k 1 f (w) dw X
f (z) = = z = ak z k
2πi S 1 w − z 2πi S 1 wk+1

as desired.

Corollary 1.4 An analytic function has at least one singularity on its circle
of convergence.

That is, if f can be extended analytically from B(p, R) to B(p, R′ ), then

the radius of convergence is at least R′ . So there must be some obstruction
to making such an extension, if 1/R = lim sup |ak |1/k .
P
Example: Fibonacci numbers. Let f (z) = an z n , where an is the nth
Fibonacci number. We have (a0 , a1 , a2 , a3 , . . .) = (1, 1, 2, 3, 5, 8, . . .). Since
an = an−1 + an−2 , except for n = 0, we get

f (z) = (z + z 2 )f (z) + 1

and so f (z) = 1/(1 − z − z 2 ). This has

√ a singularity at z = 1/γ and thus
lim sup |an |1/n = γ, where γ = (1 + 5)/2 = 1.618 . . . is the golden ratio
(slightly more than the number of kilometers in a mile).

Theorem 1.5 A power series represents a rational function iff its coeffi-
cients satisfy a recurrence relation.

7
Aside: Pisot numbers. The golden ratio is an example of a Pisot number;
it has the property that d(γ n , Z) → 0 as n → ∞. It is an unsolved problem
to show that if α > 1 satisfies d(αn , Z)
P→ 0,i then α is an algebraic number.
Kronecker’s theorem asserts that ai z is a rational function iff deter-
minants of the matrices ai,i+j , 0 ≤ i, j ≤ n are zero for all n sufficiently
large [Sa, §I.3]
Question: why are 10:09 and 8:18 such pleasant times? [Mon].
P
Isolation of zeros. If f (z) = an (z − p)n vanishes at p but is not
identically zero, then we can factor out the leading term and write:

f (z) = (z − p)n (an + an+1 (z − p) + · · · ) = (z − p)n g(z)

where g is analytic and g(p) 6= 0. This is the simplest case of the Weier-
strass preparation theorem: it shows germs of analytic functions behave like
polynomials times units (invertible functions).
In particular, we find:

Theorem 1.6 The zeros of a nonconstant analytic function are isolated.

Warning: we are assuming the domain is connected!

Proof. Let U0 ⊂ U be the largest open set with f |U0 = 0. Let U1 = U − U0 .
Then by the factorization theorem above, the zeros of f in U1 are isolated.
Thus U1 is open as well. So either U = U0 — in which case f is constant —
or U = U1 — in which case f has isolated zeros.

Corollary 1.7 An analytic function which is constant along an arc, or even

on a countable set with an accumulation point, is itself constant.

Corollary 1.8 The extension of a function f (x) on [a, b] ⊂ R to an analytic

function f (z) on a connected domain U ⊃ [a, b] is unique — if it exists.

Mean value and maximum principle. On the circle |z| = r, with

z = reiθ , we have dz/z = i dθ. Thus Cauchy’s formula gives
Z Z
1 dz 1
f (0) = f (z) = f (z) dθ.
2πi S 1 (r) z 2π S 1 (r)

In other words, analytic functions satisfy:

Theorem 1.9 (The mean-value formula) The value of f (p) is the av-
erage of f (z) over S 1 (p, r).

8
Corollary 1.10 (The Maximum Principle) A nonconstant analytic func-
tion does not achieve its maximum in U .

Proof. Suppose f (z) achieves its maximum at p ∈ U . Then f (p) is the

average of f (z) over a small circle S 1 (p, r). Moreover, |f (z)| ≤ |f (p)| on this
circle. The only way the average can agree is if f (z) = f (p) on S 1 (p, r). But
then f is constant on an arc, so it is constant in U .

Corollary 1.11 If U is compact, then supU |f | = sup∂U |f |.

Cauchy’s bound and algebraic completeness of C. Suppose f is

analytic on B(p, R) and let M (R) = sup|z−p|=R |f (z)|; then Cauchy’s bound
(1.1) becomes:
|f (n) (p)| M (R)
≤ .
n! Rn
On the other hand, if U = C — so f is an entire function — then the bound
above forces M (R) to grow unless some derivative vanishes identically. Thus
we find:

Theorem 1.12 A bounded entire function is a constant. More generally, if

M (R) = O(Rn ), then f is a polynomial of degree at most n.

Corollary 1.13 Any polynomial f ∈ C[z] of degree 1 or more has a zero in

C.

Otherwise 1/f (z) would be a nonconstant, bounded entire function. Al-

ternatively, 1/f (z) → ∞ as |z| → ∞, so we obtain a violation of the maxi-
mum principle.

Corollary 1.14 There is no conformal homeomorphism between C and ∆.

(An analytic map with no critical points is said to be conformal, because

it preserves angles.)
Aside: quasiconformal maps. A diffeomorphism f : U → V between
domains in C is quasiconformal if supU |∂f /∂f | < ∞. Many qualitative the-
orems for conformal maps also hold for quasiconformal maps. For example,
there is no quasiconformal homeomorphism between C and ∆.

9
Parseval’s theorem. The power series of an analytic function on the ball
B(0, R) also contains
P information about its L2 -norm on the circle |z| = R:
n
namely if f (z) = an z , then we have:
X Z
2 2n 1
|an | R = |f (z)|2 dθ.
2π |z|=R

This comes from the fact that the functions z n are orthogonal in L2 (S 1 ).
It also gives another important perspective on holomorphic functions: they
are the elements in L2 (S 1 ) with positive Fourier coefficients, and hence give
a half–dimensional subspace of this infinite–dimensional space.
Compactness of bounded functions. Cauchy’s bound on a disk also
implies that if f is small, then f ′ is also small, at least if we are not too near
the edge of U .

Theorem 1.15 Let f (z) be analytic on U and bounded by M . Then |f ′ (z)| ≤

M/d(z, ∂U ).

Corollary 1.16 If fn are analytic functions and fn → f uniformly, then

f ′ exists and fn′ → f ′ locally uniformly.

Corollary 1.17 A uniform limit of analytic functions is analytic.

Note: Many fallacies in real analysis become theorems in complex analysis.

Note that fn (z) = z n /n tends to zero uniformly on ∆, but fn′ (z) = z n−1
only tends to zero locally uniformly.

Corollary 1.18 Let fn be a sequence of functions on U with |fn | ≤ M .

Then after passing to a subsequence, there is an analytic function g such
that fn → g locally uniformly on U .

Proof. For any compact set K ⊂ U , the restrictions fn |K are equicontinu-

ous. Apply the Arzela–Ascoli theorem.

P n
Laurent series. Using again the basic series 1/(1−z) = z and Cauchy’s
formula over the two boundary components of the annulus A(r, R) = {z :
r < |z| < R}, we find:

10
Theorem 1.19 If f (z) is analytic on A(r, R), then in this region we have
∞
X
f (z) = an z n .
n=−∞

The positive terms converge for |z| < R, and the negative terms converge
for |z| > r.

Corollary 1.20 An analytic function on the annulus r < |z| < R can be
expressed as the sum of a function analytic on |z| < R and a function
analytic on |z| > r.

PU −p, with p n∈
Isolated singularities. As a special case, if f is analytic on
U , then near p we have a Laurent series expansion f (z) = ∞ −∞ an (z − p) .
We say f has an isolated singularity at p.
We write ord(f, p) = n if an 6= 0 but ai = 0 for all i < 0. The values
n = −∞ and +∞ are also allowed. If n ≥ 0, the apparent singularity at p
is removable and f has a zero of order n at p. If −∞ < n < 0, we say f has
a pole of order −n at p. In either of these cases, we can write

f (z) = (z − p)n g(z),

where g(p) 6= 0 and g(z) is analytic near p (so O(g, p) = 0).

If ord(f, p) > −∞, then f has a finite Laurent series
X ∞
a−n a−1
f (z) = + · · · + + an z n
(z − p)n z−p
n=0

near p. The germs of functions at p with finite Laurent tails form a local field,
with ord(f, p) as its discrete valuation. (Compare Qp , where vp (pn a/b) = n.)
If ord(f, p) = −∞ we say f has an essential singularity at p. (Example:
f (z) = sin(−1/z) at z = 0.)
The residue. A critical role is played by the residue of f (z) at p, defined
by Res(f, p) = a−1 . It satisfies
Z
f (z) dz = 2πi Res(f, p)
γ

for any small loop encircling the point p in U . Thus the residue is intrin-
sically an invariant of the 1-form f (z) dz, not the function f (z). (If we
regard f (z) dz as a 1-form, then its residue is invariant under change of
coordinates.)

11
Theorem 1.21 (The residue theorem) Let f : U → C be a function
which is analytic apart from a finite set of isolated singularities. We then
have: Z X
f (z) dz = 2πi Res(f, p).
∂U p∈U

We will develop two applications of the residue theorem: the argument

principle, and the evaluation of definite integrals.
The argument principle. The previous argument for algebraic complete-
ness of C is clever and nonconstructive. One way to make the proof more
transparent and constructive is to employ the argument principle.
We first observe that if f has at worst a pole at p, then its logarithmic
derivative
d log f = f ′ (z)/f (z) dz
satisfies
Res(f ′ /f, p) = ord(f, p).
We thus obtain:
Theorem 1.22 (Argument principle) If f is analytic in U and f |∂U is
nowhere zero, then the number of zeros of f in U is given by
X Z
1 f ′ (z) dz
N (f, 0) = ord(f, p) = ·
2πi ∂U f (z)
p∈U

Corollary 1.23 (Rouché’s Theorem) If |f | > |g| along ∂U , then f and

f + g have the same number of zeros in U .

Proof. The continuous, integer-valued function N (f + tg, U ) is constant

for t ∈ [0, 1].

Example. Consider p(z) = z 5 +14z+1. Then all its zeros are inside |z| < 2,
since |z|5 = 32 > |14z + 1| when |z| = 2; but only one inside |x| < 3/2, since
|z 5 + 1| ≤ 1 + (3/2)5 < 9 < |14z| on |z| = 3/2. (Intuitively, the zeros of p(z)
are close to the zeros of z 5 + 14z which are z = 0 and otherwise 4 points on
|z| = 141/4 ≈ 1.93.
Corollary 1.24 Let p(z) = z d + a1 z d−1 + . . . + ad , and suppose |ai | < Ri /d.
Then p(z) has d zeros inside the disk |z| < R.

Proof. Write p(z) = f (z) + g(z) with f (z) = z d , and let U = B(0, R); then
on ∂U , we have |f | = Rd and |g| < Rd ; now apply Rouché’s Theorem.

12
Corollary 1.25 A nonconstant analytic function is an open mapping.

Proof. Suppose f (p) = q. Then p is an isolated zero of f (z) − q. Choose

a ball B(p, r) so there are no other zeros inside this ball, and let s by the
minimum of |f (z) − q| over ∂B(p, r). Then if |q − q ′ | < s, we find f (z) − q ′
also has a zero in B(p, r), and thus f (B(p, r)) contains B(q, s). This shows
f (U ) is open whenever U is open.

The open mapping theorem gives an alternate proof of the maximum

principle and its strict version:

Corollary 1.26 If |f | achieves its maximum in U , then f is a constant.

Geometric picture. Suppose for simplicity that U is a disk, and p 6∈

f (∂U ). The the number of solutions to f (z) = p in U , counted with multi-
plicity, is:
Z Z
1 1
N (f, p) = d log(f (z) − p) = d arg(f (z) − p)·
2πi ∂U 2π ∂U

This is nothing more than the winding number of f (∂U ) around p. Ob-
serve that N (f, p) is a locally constant function on C − f (∂U ), zero on the
noncompact component. Summing up:

Theorem 1.27 For any p ∈ C−f (∂U ), the number of solutions to f (z) = p
in U is the same as the winding number of f (∂U ) around p.

Using isolation of zeros, we have:

Corollary 1.28 If f ′ (p) 6= 0, then f is a local homeomorphism at p.

In fact for r sufficiently small and q close to p, we have

Z
−1 1 zf ′ (z) dz
f (q) = ·
2π B(p,r) f (z) − q

Aside: the smooth case. These results also hold for smooth mappings f
once one finds a way to count the number of solutions to f (z) = p correctly.
(Some may count negatively, and the zeros are only isolated for generic
values of p.)

13
Aside: Linking numbers and intersection multiplicities of curves
in C2 . Counting the number of zeros of y = f (x) at x = 0 is the same
as counting the multiplicity of intersection between the curves y = 0 and
y = f (x) in C2 , at (0, 0).
In general, to an analytic curve C defined by F (x, y) = 0 passing through
(0, 0) ∈ C2 , we can associate a knot by taking the intersection of C with the
boundary S 3 (r) of a small ball centered at the origin.
A pair of distinct, irreducible, curves C1 and C2 passing through (0, 0)
(say defined by fi (x, y) = 0, i = 1, 2), have a multiplicity of intersection
m(C1 , C2 ). This can be defined geometrically by intersection with a small
sphere S 3 (r) in C2 centered at (0, 0). Then we get a pair of knots, and their
linking number is the same as this multiplicity.
Examples. It is often simpler the use S 3 = ∂∆2 instead of |x|2 + |y|2 = 1.
Then S 3 is the union of two solid tori, ∆ × S 1 and S 1 × ∆. The axes x = 0
and y = 0 are the core curves of these tori; they have linking number one.
The line y = x is a (1, 1) curve on S 1 × S 1 ; more generally, for gcd(a, b) = 1,
y a = xb is an (a, b) curve, parameterized by (x, y) = eita , eitb ). In particular,
the cusp y 2 = x3 meets S 3 in a trefoil knot. It links x = 0 twice and y = 0
three times.
Problem. Show that the figure-eight knot cannot arise from an analytic
curve.
Multivalued functions. It is useful to have a general discussion of the
sometimes confusing notion of ‘branch cuts’ and ‘multivalued functions’.
Here are 2 typical results.

Theorem 1.29 Let U ⊂ C∗ be a simply-connected region, and suppose ea =

b ∈ U . Then there is a unique analytic function L : U → C such that
L(b) = a and exp L(z) = z for all z ∈ U .

This function is a ‘branch’ of log(z). To define it, we simply set

Z z
L(z) = a + dt/t.
b

The integral is over any path in U connecting b to z. Since U is simply-

connected and does not contain zero, this integral is path independent, and
we have L′ (z) = 1/z and L(b) = a. Thus f (z) = eL (z)/z satisfies f ′ (z) = 0,
and hence f (z) is a constant. Since f (b) = ea /b = 1, we conclude that
f (z) = log z.

14
Corollary 1.30 If an = b then there is a unique analytic function R : U →
C such that R(b) = a and R(z)n = z for all z ∈ U .

Of course near z = 1 with the usual normalizations these functions can

be written down explicitly. Integrating 1/(1 + z), we find:

log(1 + z) = z − z 2 /2 + z 3 /3 − · · ·
and, by the binomial theorem,

(1 + z)α = 1 + αz + α(α − 1)z 2 /2 + · · ·

Functional factorization. Here is an alternate proof of the open mapping

theorem. It is clear that pn (z) = z n is an open map, even at the origin; and
(by the inverse function theorem) that an analytic function is open at any
point p where f ′ (p) 6= 0. The general case follows from these via:

Theorem 1.31 Let f be an analytic map at p with ord(f ′ , p) = n ≥ 0. Then

up to a local chance of coordinates in domain and range, f (z) = z n+1 . That
is, there exist analytic diffeomorphisms with h1 (0) = p, and h2 (0) = f (p),
such that
h2 ◦ f ◦ h−1
1 =z
n+1
.

Proof. We may assume p = f (p) = 0 and f (z) = z n g(z) where g(0) 6=

0. Then h(z) = g(z)1/n is a well-defined analytic function near z = 0,
once we have chosen a particular value for g(0)1/n . It follows that f (z) =
(zh(z))n = pn (h1 (z)) where h′1 (z) 6= 0. Thus h1 is a local diffeomorphism,
and f ◦ h−1
1 (z) = z .
n

Bounded functions and essential singularities.

Theorem 1.32 Every isolated singularity of a bounded analytic function is
removable.

Proof.P Suppose the singularity is at z = 0, and write f (z) as a Laurent

series an z n . Then for any r > 0 we have
Z
1
Res(f, 0) = a−1 = f (z) dz.
2πi S 1 (r)
But if |f | ≤ M then this integral tends to zero as r → 0, and hence a−1 = 0.
Similarly a−(n+1) = Res(z n f (z), 0) = 0 for all n ≥ 0. Thus the Laurent
power series gives an extension of f to an analytic function at z = 0.

15
Corollary 1.33 (Weierstrass-Casorati) If f (z) has an essential singu-
larity at p, then there exist zn → p such that f (zn ) is dense in C.

Proof. Otherwise there is a neighborhood U of p such that f (U −{p}) omits

some ball B(q, r) in C. But then g(z) = 1/(f (z) − q) is bounded on U , and
hence analytic at p, with a zero of finite order. Then f (z) = q + 1/g(z) has
at worst a pole at p, not an essential singularity.

Aside: several complex variables. Using the same type of argument

and some basic facts from several complex variables, it is easy to show that
if V ⊂ U ⊂ Cn is an analytic hypersurfaces, and f : U −V → C is a bounded
analytic function, then f extends to all of U .
More remarkably, if V has codimension two or more, then every ana-
lytic function extends across V . For example, an isolated point is always a
removable singularity in C2 .
Residue calculus and definite integrals. The residue theorem can be
used to systematically evaluate various definite integrals.
Definite integrals 1: rational functions on R. R Whenever a rational
function R(x) = P (x)/Q(x) has the property that R |R(x)| dx is finite, we
can compute this integral via residues: we have
Z ∞ X
R(x) dx = 2πi Res(R, p).
−∞ Im p>0

(Of course we can also compute this integral by factoring Q(x) and using
partial fractions and trig substitutions.)
Example. Where does π come from? It emerges naturally from rational
functions by integration — i.e. it is a period. Namely, we have
Z ∞
dx
2
= 2πi Res(1/(1 + z 2 ), i) = 2πi(−i/2) = π.
−∞ 1 + x
R
Of course this can also be done using the fact that dx/(1+x2 ) = tan−1 (x).
More magically, for f (z) = 1/(1 + z 4 ) we find:
Z ∞ √ √
dx π
4
= 2πi(Res(f, (1 + i)/ 2) + Res(f, (1 + i)/ 2) = √ ·
−∞ 1 + x 2

Both are obtain by closing a large interval [−R, R] with a circular arc in the
upper halfplane, and then taking the limit as R → ∞.

16
 We can even compute the general case, f (z) = 1/(1 + z n ), with n even.
For this let ζk = exp(2πi/k), so f (ζ2n ) = 0. Let P be the union of the paths
[0, ∞)ζn and [0, ∞), oriented so P moves positively on the real axis. We can
then integrate over the boundary of this pie-slice to obtain:
Z ∞ Z
dx n−1
(1 − ζn ) n
= f (z) dz = 2πi Res(f, ζ2n ) = 2πi/(nζ2n ),
0 1 + x P

which gives Z ∞
dx 2πi π/n
n
= −1 +1 = ·
0 1+x n(−ζ2n + ζ2n ) sin π/n
Here we have used the fact that ζ2nn = −1. Note that the integral tends to

1 as n → ∞, since 1/(1 + xn ) converges to the indicator function of [0, 1].

Definite integrals 2: rational functions of sin(θ) and cos(θ). Here
is an even more straightforward application of the residue theorem: for any
rational function R(x, y), we can evaluate
Z 2π
R(sin θ, cos θ) dθ.
0

The method is simple: set z = eiθ and convert this to an integral of an

analytic function over the unit circle. To do this we simple observe that
cos θ = (z + 1/z)/2, sin θ = (z − 1/z)/(2i), and dz = iz dθ. Thus we have:
Z 2π Z     
1 1 1 1 dz
R(sin θ, cos θ) dθ = R z− , z+ ·
0 S1 2i z 2 z iz
For example, for 0 < a < 1 we have:
Z 2π Z
dθ i dz 2π
2
= = 2πi(i/(a2 − 1)) = ·
0 1 + a − 2a cos θ S 1 (z − a)(az − 1) 1 − a2
R∞
Definite integrals 3: fractional powers of x. 0 xa R(x)dx, 0 < a < 1,
R a rational function.
For example, consider
Z ∞
xa
I(a) = dx.
0 1 + x2
Let f (z) = z a /(1 + z 2 ). We integrate out along [0, ∞) then around a large
circle and then back along [0, ∞). The last part gets shifted by analytic
continuation of xa and we find

(1 − 1a )I(a) = 2πi(Res(f, i) + Res(f, −i))

17
and Res(f, i) = ia /(2i), Res(f, −i) = (−i)a /(−2i) (since xa /(1 + x2 ) =
xa /(x − i)(x + i)). Thus, if we let ia = ω = exp(πia/2), we have
π(ia − (−i)a ) ω − ω3 π π
I(a) = a
= π 4
= −1
= ·
(1 − 1 ) 1−ω ω+ω 2 cos(πa/2)
For example, when a = 1/3 we get
√
I(a) = π/(2 cos(π/6)) = π/ 3.

Residues and infinite sums. The periodic function f (z) = π cot(πz) has
the following convenient properties: (i) It has residues 1 at all the integers;
and (ii) it remains bounded as Im z → ∞. From these facts we can deduce
some remarkable properties: by integrating over a large rectangle S(R), we
find for k ≥ 2 even,
Z X∞
1 f (z) dz k
0 = lim = Res(f (z)/z , 0) + 2 1/nk .
R→∞ 2πi S(R) zk
1
P
Thus we can evaluate the sum 1/n2 using the Laurent series
cos(z) 1 − z 2 /2! + z 4 /4! − · · ·
cot(z) = =
sin(z) z(1 − z 2 /3! + z 4 /5! − · · · )
= z −1 (1 − z 2 /2! + z 4 /4! − · · · )(1 + z 2 /6 + 7z 4 /360 + · · · )
= z −1 − z/3 − z 3 /45 − · · ·
(using the fact thatP(1/(3!)2 − 1/5! = 7/360). This shows Res(f (z)/z 2 , 0) =
−π 2 /3 and hence 1/n2 = π 2 /6. Similarly, 2ζ(2k) = − Res(f (z)/z 2k , 0).
For example, this justifies ζ(0) = 1 + 1 + 1 + · · · = −1/2.
Little is known about ζ(2k + 1). Apéry showed that ζ(3) is irrational,
but it is believed to be transcendental.
P
We note that ζ(s) = 1/ns is analytic for Re s > 1 and extends an-
alytically to C − {1} (with a simple pole at s = 1). Q In particular ζ(0) is
well-defined. Because of the factorization ζ(z) = (1 − 1/ps )−1 , the be-
havior of the zeta function is closely related to the distribution of prime
numbers. The famous Riemann hypothesis states that any zero of ζ(s) with
0 < Re s < 1 satisfies Re s = 1/2. It implies a sharp form of the prime
number theorem, π(x) = x/ log x + O(x1/2+ǫ ).
The zeta function also has trivial zeros at s = −2, −4, −6, . . ..
R
Hardy’s paper on sin(x)/x dx. We claim
Z ∞
sin x dx π
I= = ·
0 x 2

18
Note that this integral is improper, i.e. it does not converge absolutely.
Also, the function f (z) = sin(z)/z has no poles — so how can we apply the
residue calculus?
The trick is to observe that
Z
eix dx
−2iI = lim ·
r→0 r<|x|<1/r x

We now use the fact that |eix+iy | ≤ e−y to close the path in the upper
halfplane, and conclude that
Z
eiz dz
2iI = lim ·
r→0 S 1 (r)
+
z

(Here the semicircle is oriented counter-clockwise, as usual.) Since Res(eiz /z, 0) =

1, we find 2iI = (2πi)(1/2) and hence I = π/2.
Harmonic functions. A C 2 real-value function u(z) is harmonic if
d2 u d2 u d2
∆u = + = 4 = 0.
dx2 dy 2 dz dz
Equivalently, we have
d ∗ du = 0,
where ∗ is the Hodge star operator (given by ∗dx = dy and ∗dy = −dx). In
terms better adapted to complex analysis, u is harmonic iff
∂∂u = 0.
In physical terms, ∇ · ∇u = 0, and d2 u = 0 implies ∇ × ∇u = 0; i.e., u
generates a volume preserving flow with zero curl.
Note. The operators ∂ and ∂ extend naturally to maps from 1-forms to
2-forms; they satisfy ∂(f (z) dz) = (∂f ) dz, ∂(f (z) dz) = 0. Then d = ∂ + ∂
on 1-forms as well.
Basic facts:
1. If f = u + iv is analytic, then f , f , u and v are harmonic.
2. A function u is harmonic iff du/dz is holomorphic.
3. Any real-valued harmonic function is locally the real part of a holo-
morphic function.
(Integrate ∂u to obtain an analytic function with ∂f = ∂u. Then
∂f = ∂u; thus d(f + f ) = du and so u = f + f up to an additive
constant.)

19
 4. Thus any C 2 harmonic function is actually infinitely differentiable.

5. A harmonic function satisfies the mean-value theorem: u(p) is the

average of u(z) over S 1 (z, p). This implies:

6. A harmonic function satisfies the maximum principle.

7. If u is harmonic and f is analytic, then u ◦ f is also harmonic.

8. If f = u + iv is analytic, we say v is a harmonic conjugate of u. Then

−u is a harmonic conjugate of v.

9. A uniform limit of harmonic functions is harmonic.

-4 -2 2 4

-1

-2

-3

Figure 1. Orthogonal level sets.

Examples. The function Re(z 3 ) = x3 − 3xy 2 is a harmonic polynomial.

The function arg z is the harmonic conjugate of log |z|. This shows the
harmonic conjugate may be multivalued.
Flows. The level sets of u and v are orthogonal. Thus the area-preserving
flow generated by ∇u follows the level sets of v, and vice-versa. A simple
example is provided by polar coordinates, which give the level sets of u and
v where log |z| = u + iv. The area-preserving flows are rotation around the
origin, and the radial flow at rate 1/r through circles of radius r.
See Figure 1 for another example, this time on U = C − {−2, 2}, where
the level sets are conics. These conics are the images of radial lines and
circles under f (z) = z + 1/z, so they are also locally level sets of harmonic
functions.

20
Harmonic extension. Here is one of the central existence theorems for
harmonic functions.

Theorem 1.34 There is a unique linear map P : C(S 1 ) → C(∆) such that
u = P (u)|S 1 and P (u) is harmonic on ∆.

Proof. Uniqueness is immediate from the maximum principle. To see

existence, observe that we must have P (z n ) = z n and P (z n ) = z n . Thus P
is well-defined on the span S of polynomials in z and z, and satisfies there
kP (u)k∞ = kuk∞ . Thus P extends continuously to all of C(S 1 ). Since the
uniform limit of harmonic functions is harmonic, P (u) is harmonic for all
u ∈ C(S 1 ).

Poisson kernel. The map P can be given explicit by the Poisson kernel.
For example, u(0) is just the average of u over S 1 . We can also say u(p) is
the expected value of u(z) under a random walk starting at p that exits the
disk at z.
To find the Poisson kernel explicitly, suppose we have a δ-mass at z = 1.
Then it should extend to a positive harmonic function u on ∆ which vanishes
along S 1 except at 1, and has u(0) = 0. In turn, u should be the real part
of an analytic function f : ∆ → C such that f (0) = 1 and Re f |S 1 = 0
and f has a pole at z = 1. Such a function is given simply by the Möbius
transformation f : ∆ → U = {z : Re z > 0}:
1+z
f (z) = ·
1−z
Convolving, we find the analytic function with Re f = u for a given u ∈
C(S 1 ) is given by Z
1
F (z) = f (z/t)u(t)|dt|,
2π S 1
and thus Z
1
u(r, α) = Pr (α − θ)u(θ) dθ,
2π S1

where, for z = reiθ , we have

1 − |z|2 1 − r2
Pr (θ) = Re f (z) = 2
= ·
|1 − z| 1 − 2r cos θ + r 2

Relation to Fourier series. The above argument suggests P∞that, to define

the harmonic extension of u, we should just write u(z) = −∞ an z on S 1 ,
n

21
 8

-3 -2 -1 0 1 2 3

Figure 2. The Poisson kernel Pr (θ) for r = 0.5, 0.7, 0.8.

and then replace z −n by z n to get its extension to the disk. This actually
works, and gives another approach to the Poisson kernel. P
Given f ∈ C(S 1 ), it is not true, in general, that its Fourier series an z n
converges for all z ∈ S 1 . However, it is true that this series converges in
the unit disk and defines a harmonic function there. As we have seen, this
harmonic function provides a continuous extension of f . This shows:

Theorem 1.35 If f ∈ C(S 1 ) has Fourier coefficients an , then for all z ∈ S 1

we have
X∞
f (z) = lim an r |n| z n .
r→1−
−∞
R
(Here an = (1/2π) S 1 f (z)z n |dz|.)
P
Abel summation.
P In general, we say S is the Abel sum of the series bn
if S = limr→1− bn r n . For example,

1 − 2 + 3 − 4 + · · · = lim(1 − 2r + 3r 2 − · · · ) = lim −1/(1 + r)2 = −1/4.

The result above shows the Fourier series of f is Abel summable to the
original function f .
Laplacian as a quadratic form, and physics. Suppose u, v ∈ Cc∞ (C)
– so u and v are smooth, real-valued functions vanishing outside a compact
set. Then, by integration by parts, we have
Z Z Z
h∇u, ∇vi = − hu, ∆vi = − hv, ∆ui.
∆ ∆ ∆

22
To see this using differential forms, note that:
Z Z Z
0= d(u ∗ dv) = (du)(∗dv) + u(d ∗ dv).
∆ ∆ ∆

In particular, we have
Z Z
|∇u|2 = − u∆u.
∆ ∆

Compare this to the fact that hT x, T xi = hx, T ∗ T xi on any inner product

space. Thus −∆ defines a positive-definite quadratic form on the space of
smooth functions.
The extension
R of u from S 1 to ∆ is a ‘minimal surface’ in the sense that
it minimizes ∆ |∇u|2 over all possible extensions. Similarly, minimizing
the energy in an electric field then leads to the condition ∆u = 0 for the
electrical potential.
2.0

1.5

1.0

0.5

0.0
-2 -1 0 1 2

Figure 3. Streamlines around a cylinder.

Probabilistic interpretation. Brownian motion is a way of constructing

random paths in the plane (or Rn ). It leads to the following simply inter-
pretation of the extension operator P . Namely, given p ∈ ∆, one considers
a random path pt with p0 = p. With probability one, there is a first T > 0
such that |pT | = 1; and then one sets u(p) = E(u(pT )). In other words, u(p)
is the expected value of u(pT ) at the moment the Brownian path exits the
disk.
Using the Markov property of Brownian motion, it is easy to see that u(p)
satisfies the mean-value principle, which is equivalent to it being harmonic.

23
It is also easy to argue that |p0 − pT | tends to be small when p0 is close to
S 1 , and hence u(p) is a continuous extension of u|S 1 .
The Poisson kernel (1/2π)Pr (θ)dθ gives the hitting density on S 1 for a
Brownian path starting at (r, 0).
Hyperbolic geometry interpretation. Alternatively, u(z) is the ex-
pected value of u(p) and the endpoint of a random hyperbolic geodesic ray
γ in ∆ with one vertex at z. (The angle of the ray in Tz ∆ is chosen at
random in S 1 .)
Example: fluid flow around a cylinder. We begin by noticing that
f (z) = z + 1/z gives a conformal map from the region U ⊂ H where |z| > 1
to H itself, sending the circular arc to [−2, 2]. Thus the level sets of Im f =
y(1 − 1/(x2 + y 2 )) describe fluid flow around a cylinder. Note that we are
modeling incompressible fluid flow with no rotation, i.e. we are assuming
the curl of the flow is zero. This insures the flow is given by the gradient of
a function.
Harmonic functions and the Schwarz reflection principle. Here is an
application of harmonic functions that will be repeatedly used in geometric
function theory.
Let U ⊂ C be a region invariant under z 7→ z. Suppose f : U → C is
continuous,
f (z) = f (z), (1.2)
and f is analytic on U+ = U ∩ H. We can then conclude that f is analytic
on U . In particular, f is analytic at each point of U ∩ R.
Here is a stronger statement:

Theorem 1.36 Suppose f : U+ → C is analytic and Im f (z) → 0 as Im z →

0. Then f extends to an analytic function on U satisfying (1.2).

Proof. The statement is local, so we can assume U = B(p, r) where p ∈ R.

Let f (z) = u(z)+iv(z); then v is harmonic on U+ and v extends continuously
to the real axis, with v(x) = 0. Extend v to U by v(z) = −v(z).
Now by the Poisson integral, v|S 1 (p, r) extends to a unique harmonic
function h on B(p, r). By uniqueness, h(z) = −h(z), and hence h also
vanishes on the real axis. Thus h = v on ∂U+ . By the maximum principle,
h = v on U+ , and hence on U .
We are now done: by the existence of harmonic conjugates, there is some
analytic function on U with Im F = v. But then Im F = Im f on U , so F
and f differ by a real constant, which can be normalized to be zero.

24
Remark. One can replace z 7→ z with reflection through any circle, or
more generally with local reflection through a real-analytic arc.
Example. Suppose f (z) is analytic on H and f (z) → 0 along an interval
in R. Then f is identically zero. This extends the ‘isolated zero’ principle
to the boundary of a region.
Additional topics.
The Phragmen–Lindelöf Theorems. These theorems address the fol-
lowing question. Suppose f (z) is an analytic function on the horizontal
strip U = {x + iy : a < y < b}, and continuous on U . Can we assert that
supU |f | = sup∂U |f |?
The answer is no, in general. However, the answer is yes if f (x+ iy) does
not grow too rapidly as |x| → ∞. In fact, this is a property of harmonic
functions u(z). The point is that if we truncate the strip to a rectangle by
cutting along the lines where |x| = R, then the harmonic measure of the
ends (as seen from a fixed point z ∈ U ) tends to zero exponentially fast.
Thus if |u(x + iy)| = O(|x|n ) for some n, we get the desired control.
Runge’s theorem. Here is an interesting and perhaps surprising applica-
tion of Cauchy’s formula.
Let K ⊂ C be a compact set, let C(K) denote the Banach space of
continuous functions with the sup-norm, and let A(K), R(K) and P (K) ⊂
C(K) denote the closures of the analytic, rational and polynomial functions.
Note that all three of these subspaces are algebras.
Example. For K = S 1 we R have R(K) = A(K) = C(K) by Fourier series,
but P (K) 6= C(K), since S 1 p(z) dz = 0 for any polynomial. In particular,
1/z 6∈ P (S 1 ).
Remarkably, if we remove a small interval from the circle to obtain an
arc K = exp[0, 2π − ǫ], then 1/z can be approximated by polynomials on K.

Theorem 1.37 (Runge) For any compact set K ⊂ C we have R(K) =

A(K), and P (K) = A(K) provided C − K is connected.

Proof. For the first result, suppose f (z) is analytic on a smoothly bounded
neighborhood U of K. Then we can write
Z Z
1 f (t) dt
f (z) = = Fz (t) dt.
2πi ∂U t − z ∂U

Since d(z, ∂U ) ≥ d(K, ∂) > 0, the functions {Fz } range in a compact subset
of C(∂U ). Thus we can replace this integral with a finite sum at the cost

25
of an error that is small independent of z. But the terms f (ti )/(ti − z)
appearing in the sum are rational functions of z, so R(K) = A(K).
The second result is proved by pole–shifting. By what we have just
done, it suffices to show that fp (z) = 1/(z − p) ∈ P (K) for every p 6∈ K.
Let E ⊂ C − K denote the set of p for which this is true.
Clearly E contains all p which are sufficiently large, because then the
power series for fp (z) converges uniformly on K. Also E is closed by defini-
tion. To complete the proof, it suffices to show E is open.
The proof that E is open is by ‘pole shifting’. Suppose p ∈ E, q ∈ B(p, r)
and B(p, r) ∩ K = ∅. Note that fq (z) is analytic on C − B(p, r), and tends to
zero as |z| → ∞. Thus fq (z) can be expressed as a power series in 1/(z − p):
∞
X X
−1
(z − q) = an (z − p)−n = an fp (z)n ,
0

convergent for |z − p| > |z − q|, and converging uniformly on K. (Compare

the expression
∞
X
1 1
= ,
z − 1 n=1 z n

valid for |z| > 1.) Since (z − q)−1 → 0 as |z| → ∞, only terms with n ≥ 0
occur on the right. But fp ∈ A(K) by assumption, and A(K) is an algebra,
so it also contains fpn . Thus fq ∈ A(K) as well.

Aside: Lavrentiev’s Theorem. It can be shown that if C−K is connected

and K has no interior, then A(K) = C(K). This is definitely R false for a fat
Swiss cheese: if ∂K is rectifiable and of finite length, then ∂K f (z) dz = 0
for all f in A(K), and so A(K) 6= C(K). For more details see [Gam].
Applications: pointwise convergence. Runge’s theorem can be used
to show easily that there is a sequence of polynomials fn (z) that converge
pointwise, but whose limit is not even continuous. Indeed, let An = [0, n]
and let B1 ⊂ B2 ⊂ · · · be an increasing sequence of compact sets such that
S
Bn = C − [0, ∞). Then every z ∈ C eventually belongs to An or Bn . Let
fn (z) be a polynomial, whose existence is guaranteed by Runge’s theorem,
such that |fn (z)| < 1/n on An and |fn (z) − 1| < 1/n on Bn . Then clearly
lim fn (z) = 0 on [0, ∞) and lim fn (z) = 1 elsewhere.
Applications: embedding the disk into affine space. Runge’s theorem
can also be used to show there is a proper embedding of the unit disk into
C3 . See [Re, §12.3].

26
2 The simply-connected Riemann surfaces
Riemann surfaces. A Riemann surface X is a connected complex 1-
manifold. This means X is a Hausdorff topological space equipped with
charts (local homeomorphisms) fi : Ui → C, and the transition functions
fij = fi ◦ fj−1 are analytic where defined.
It then makes sense to discuss analytic functions on X, or on any open
subset of X. Technically we obtain a sheaf of rings OX with OX (U ) con-
sisting of the analytic maps f : U → C.
Aside from C and connected open sets U ⊂ C, the first interesting Rie-
mann surface (and the basic example of a compact Riemann surface) is the
Riemann sphere C. b The map f : C b − {0} → C given by f (z) = 1/z provides
a chart near infinity.
Meromorphic functions. There is also a natural notion of holomorphic
(analytic) maps between Riemann surfaces. A meromorphic function on X
is an analytic map f : X → C b that is not identically ∞.
Since X is connected, the ring K(X) of all meromorphic functions on X
forms a field, and the ring O(X) of all analytic functions forms an integral
domain.
We will see that K(C)b = C(z), while O(C) b = C, so not every meromor-
phic function is a quotient of holomorphic functions. We will see that O(C)
is much wilder (it contains exp(exp z), etc.), and yet K(C) is the field of
fractions of O(C).
Classification. In principle the classification of Riemann surfaces is com-
pleted by the following result:
Theorem 2.1 (The Uniformization Theorem) Every simply-connected
b
Riemann surface is isomorphic to H, C or C.
Then by the theory of covering spaces, an arbitrary Riemann surface sat-
isfies X ∼ b X = C/Γ, or X = C/Γ,
= C, b where Γ is a group of automorphisms.
Such Riemann surfaces are respectively elliptic, parabolic and hyperbolic.
Their natural metrics have curvatures 1, 0 and −1.
In this section we will discuss each of the simply-connected Riemann
surfaces in turn. We will discuss their geometry, their automorphisms, and
their proper endomorphisms.

2.1 The complex plane

An analytic map f : X → Y is proper if f −1 sends compact sets to compact
sets. This is equivalent to the condition that f (zn ) → ∞ whenever zn → ∞.

27
 Proper analytic maps f : X → Y are the ‘tamest’ maps between Rie-
mann surfaces. For example, they have the following properties:

1. If f is not constant, it is surjective.

2. If f ′ never vanishes, then f is a covering map.

3. If f ′ never vanishes and Y is simply-connected, then f is an isomor-

phism.

Example. The entire function exp : C → C is not proper, since it omits

the point 0.

Theorem 2.2 An analytic function f : C → C is proper iff f (z) is a non-

constant polynomial.

Proof. Clearly a polynomial of positive degree is proper. Conversely, if f

is proper, then f has finitely many zeros, and so no zeros for |z| > some
R. Then g(z) = 1/f (1/z) is a nonzero, bounded analytic function for 0 <
|z| < R. Consequently g(z) has a zero of finite order at z = 0. This
gives |g(z)| ≥ ǫ|z|n , and hence |f (z)| ≤ M |z|n . By Cauchy’s bound, f is a
polynomial.

To say f : C → C is proper is the same as to say that f extends to a

continuous function F : Cb → C.b The same argument shows, more generally,
that if f : X → Y is a continuous map between Riemann surfaces, and f is
analytic outside a discrete set E ⊂ X, then f is analytic.

Corollary 2.3 The automorphisms of C are given by the affine maps of the
form f (z) = az + b, where a ∈ C∗ and b ∈ C.

Thus Aut(C) is a solvable group. If f ∈ Aut(C) is fixed-point free, then

it must be a translation. This shows:

Corollary 2.4 Any Riemann surface covered by C has the form X = C/Λ,
where Λ is a discrete subgroup of (C, +).

Example. We have C/(z 7→ z + 1) ∼

= C∗ ; the isomorphism is given by
π(z) = exp(2πiz).
Degree. For a proper map f : X → Y , the number of points in f −1 (z),
counted with multiplicity, is independent of z. This number is called the

28
degree of f . In the case X = Y = C, it is the same as the degree of the
polynomial f .
Metrics. A conformal metric on a Riemann surface is given in local coor-
dinates by ρ = ρ(z) |dz|. We will generally assume that ρ(z) ≥ 0 and ρ is
continuous, although metrics with less regularity are also useful.
A conformal metric allows one to measure lengths of arcs, by
Z Z b
L(γ, ρ) = ρ = ρ(γ(t)) |γ ′ (t)| dt;
γ a

and areas of regions, by

Z Z Z
A(U, ρ) = ρ2 = ρ2 (z) |dz|2 = ρ2 (z) dx dy.
U U U

Metrics pull back under analytic maps by the formula:

f ∗ (ρ) = ρ(f (z)) |f ′ (z)| |dz|,

and satisfy natural formulas such as

L(γ, f ∗ ρ) = L(f (γ), ρ)

(so long as f (γ) is understood as the parameterized path γ ◦ f ).

A metric allows us to intrinsically measure the size of the derivative a
map f : (X1 , ρ1 ) → (X2 , ρ2 ): it is given by

|f ′ (z)|ρ2 (f (z)) f ∗ ρ2
kDf k = |f ′ |ρ = = ·
ρ1 (z) ρ1

We have |f ′ | = 1 iff f is a local isometry.

Flat metrics. The Euclidean metric on the plane is given by ρ = |dz|.
Since |dz| is Λ-invariant, we find:

Theorem 2.5 Every Riemann surface covered by C admits a complete flat

metric, unique up to scale.

For example, on C∗ ∼ = C/Z we have the cylindrical metric ρ = |dz|/|z|.

Every circle |z| = r has length 2π in this metric.
Cone metrics. Consider the metric ρ = |z|α |dz|/|z| on C. We claim this
is a flat metric, making the origin into a cone point of total angle θ = 2πα.
R 1 To
α
see this is plausible, note that the unit ball B(0, 1) has radius R =
0 t dt/t = 1/α, and circumference C = 2π, so C/R = 2πα.

29
 Alternatively, let f (z) = z n . Then we find

f ∗ (ρ) = |z|nα n|dz|/|z| = n|dz|

if α = 1/n. Thus the case α = 1/n gives the quotient metric on (C, n|dz|)/hζn i,
where ζn = exp(2πi/n).
Orbifold quotients. We can also take the quotient of C by the infinite
dihedral group, D∞ = hz + 1, −zi ⊂ Aut C. The result is C itself, which the
quotient map given by cos(2πz).
Closely related is the important map

π : C∗ → C

given by π(z) = z + 1/z. This degree two map gives the orbifold quotient of
C∗ by z 7→ 1/z. It gives an intermediate covering space to the one above, if
we regard C∗ as C/Z.
The map π sends circles to ellipse and radial lines to hyperboli, with foci
[−2, 2]. In fact all conics with these foci arise in this way. (See Figure 1.)
By attaching cone angles to ±2, one turns C into the (2, 2, ∞) orbifold
X. A loop around a cone point of order n has order n in the orbifold
fundamental group, so π1 (X) = Z/2 ∗ Z/2. In general, by broadening the
scope of covering spaces and deck groups to include orbifolds and maps
with fixed points, we enrich the supply of Euclidean (and other) Riemann
surfaces. For example, the (3, 3, 3) orbifold is also Euclidean — it is the
double of an equilateral triangles.
Chebyshev polynomials. Now let Sn (z) = z n . Since Sn (1/z) = 1/Sn (z),
there is a sequence of polynomials P1 (z) = z, P2 (z) = z 2 −2, P3 (z) = z 3 −3z,
etc. satisfying
Pn (z + 1/z) = (z + 1/z)n ,
or equivalently
Pn (π(z)) = π(Sn (z)).
These Chebyshev polynomials are related to multiple angle formulas for co-
sine, since for z = eiθ we have π(z) = 2 cos θ, and thus:

Pn (2 cos θ) = 2 cos(nθ).

Since Sn (z) preserves circles and radial lines, we have:

Theorem 2.6 The Chebyshev polynomials preserve the hyperbolas and el-
lipses with foci at ±2.

30
 In particular Pn : [−, 2, 2] → [−2, 2] by degree n, and Pnk (z) → ∞ for all
other z ∈ C. (Thus the Julia set of Pn is [−2, 2].)
Solving the cubic. Complex algebra finds its origins in the work of Car-
dano et al on solving cubic polynomial equations. Remarkably, complex
numbers intervene even when the root to be found is real.
One can always make a simple transformation of the form x 7→ x + c to
reduce to the form
x3 + ax + b = 0.
One can further replace x with cx to reduce to the form

x3 − 3x = b.

Thus the solution to the cubic involves inverting a single cubic function
P3 (z) = z 3 − 3z.
But to solve Pn (x) = a, we just write a = y + 1/y (by solving a quadratic
equation), and then we have x = y 1/n + y −1/n . In particular, this method
can be used to solve x3 − 3x = b.
Classification of polynomials. Let us say p(z) is equivalent to q(z) is
there are A, B ∈ Aut(C) such that Bp(Az) = q(z). Then every polynomial
is equivalent to one which is monic and centered (the sum of its roots is
zero). Every quadratic polynomial is equivalent to p(z) = z 2 .
The reasoning above shows, every cubic polynomial with distinct critical
points is equivalent to P3 (z) = z 3 − 3z. Otherwise it is equivalent to z 3 .
But for degree 4 polynomials we are in new territory: the cross-ratio of the
3 critical points, together with infinity, is an invariant.
It is a famous fact (proved using Galois theory) that a general quintic
polynomial (with integral coefficients) cannot be solved by radicals.

2.2 The Riemann sphere

We now turn to the Riemann sphere C b = C ∪ {∞}. We will examine the
sphere from several perspectives: as a Riemann surface, as a round sphere,
as the projectively line, and as the boundary of hyperbolic space.
The projective line. We have a natural projection
b∼
π : C2 − {(0, 0)} → C = P1

given by π(z0 , z1 ) = z0 /z1 ; it records the slope of each line. This gives a
natural identification of Cb with the projective line P1 , i.e. the space of lines
in C2 .

31
Aside: Some topology of projective spaces. The real projective plane RP2 is
the union of a disk and a Möbius band. The natural map p : C2 − {(0, 0)} →
b factors through the Hopf map S 3 → S 2 , whose fibers have linking number
C
one. This map generates π3 (S 2 ).
Möbius transformations. So long as ad − bc 6= 0, the map f (z) =
(az + b)/(cz + d) defines an b
a b

automorphism of C. Its inverse can be found by
inverting the matrix c d . In fact, the map π above transports the linear
b
action of SL2 (C) on C2 to the fractional linear action of PSL2 (C) on C.

b = PSL2 (C).
Theorem 2.7 We have Aut(C)

Proof. Given f ∈ Aut(C), b pick g ∈ PSL2 (C)b so g(f (∞)) = ∞. Then

g ◦ f ∈ Aut(C) ⊂ PSL2 (C), so f ∈ PSL2 (C).

Corollary 2.8 The action of Aut(C) b is uniquely triply-transitive: any dis-

tinct triple of points can be sent to (0, 1, ∞) by a unique Möbius transfor-
mation.

Corollary 2.9 The group PSL2 (C) is isomorphic, as a complex manifold,

b
to the space of distinct triples of points on C.

Classification of automorphisms up to conjugacy. There is a natural

map tr : PSL2 (C) → C/h±1i. This is clearly a class function (it is constant
on conjugacy classes), and in fact the conjugation class is almost determined
by this map. Namely we have the following classes:

(1) The identity map, tr(A) = ±2.

(2) Parabolics, A(z) = z + a; tr(A) = ±2.
(3) Elliptics: A(z) = eiθ z; tr(A) = 2 cos θ/2 ∈ [−2, 2].
(4) Hyperbolics: A(z) = et z, Re t 6= 0; tr(A) = 2 cosh(t/2).

Note that a Möbius transformation (other than the identity) either has
two simple fixed points, or a single fixed point of multiplicity two. The fixed
points of A correspond to its eigenvectors on C2 .
Note also that all these elements, except for irrational elliptics, generate
b
discrete subgroups of Aut(C).
±1
If A has eigenvalues λ , then tr(A) = λ + 1/λ. Our previous analysis
of this map shows, for example, that the traces of matrices a given value for
|λ| correspond to an ellipse with foci ±2.

32
Cross-ratios. The cross-ratio is an invariant of ordered 4-tuples of points,
characterized by the conditions that [a : b : c : d] = [ga : gb : gc : gd]
b and [0 : 1 : ∞ : λ] = λ ∈ C
for all g ∈ Aut(C), b − {0, 1, ∞}. The cross-
ratio gives an explicit isomorphism between the moduli space M0,4 and the
triply-punctured sphere.
Stereographic projection. There is a geometric identification between
the unit sphere S 2 ⊂ R3 and the Riemann sphere C b in which the north pole
N becomes ∞ and the rest of the sphere is projected linearly to C = R2 ×{0}.

Theorem 2.10 Stereographic projection is a conformal map that sends cir-

cles to circles.

Proof of conformality. Consider two vectors at a point p 6= N on S 2 .

Construct a pair of circles tangent to these vectors at p and passing through
N . Then these circles meet in the same angles at p and N . On the other
hand, each circle is the intersection of the sphere with a plane. These planes
meet C in the same angle they meet a plane tangent to the sphere at the
north pole N . Thus stereographic projection preserves angles.

θ/2
1

x
θ

Figure 4. Stereographic projection on S 1 .

Theorem 2.11 The Euclidean metric on S 2 is transported, by stereographic

b
projection, to the metric 2|dz|/(1 + |z|2 ) on C.

Proof. Reducing dimensions by one, we obtain the stereographic projection

p : S 1 − N → R. If θ ∈ [−π, π] is the angle on the sphere normalized so

33
θ(N ) = π, then we find x = p(θ) = tan(θ/2). Thus dx = 2 sec2 (θ/2) dθ =
2(1 + x2 ), which gives dθ = 2dx/(1 + x2 ). The case of S 2 follows by confor-
mality and rotation invariance.

The unitary point of view. Here is an alternative perspective on the

spherical metric. Let hv, wi = v0 w 0 + v1 w1 be the usual Hermitian metric
on C2 . Define the angle θ between two lines Cv, Cw ⊂ C2 by
|hv, wi|
cos(θ/2) = ·
|v| |w|
This θ agrees with the spherical distance on C. b (To check this, consider
1
the points 1 and exp(iα) in S . Their spherical distance is α, and if we set
v = [1, 1] and w = [eiθ/2 , e−iθ/2 ], then |v||w| = 2 and hv, wi = 2 cos(α/2).)
Given this agreement, it is now clear that the group SU(2) ⊂ SL2 (C)
b In fact this group is the full group of orientation-
acts isometrically on C.
preserving isometries.
b in the spherical metric is given by
Theorem 2.12 The isometry group of C
( ! )
a b
Isom+ (S 2 ) = : |a|2 + |b|2 = 1 ⊂ PSL2 (C).
−b a
Note that these matrices satisfy AA∗ = I.
Exercise. What is the distance, in the spherical metric, from 1 to 1 + i?
For [1 : 1] and [1 + i : 1] we get
√ √
cos θ/2 = 5/ 6,
p
and hence θ = 2 arccos( 5/6) ≈ 0.841069 . . ..
Circles and lines. A circle C ⊂ C b is either a line through ∞, or an
ordinary Euclidean circle in C.
Theorem 2.13 Möbius transformations send circles to circles, and any two
circles are equivalent.

Proof 1. A circle x2 + y 2 + Ax + By + C = 0 is also given by r 2 + r(A cos θ +

B sin θ) + C = 0, and it is easy to transform the latter under z 7→ 1/z, which
replaces r by 1/r and θ by −θ.
Proof 2. A circle is the same, projectively, as the set of null vectors for
a Hermitian form on C2 of signature (1, 1). Any two such forms are, up to
scale, related by an element of SL2 (C).

34
Area of triangles. The geodesics on S 2 are arcs of great circles. Their
angle sum always exceeds π, and in fact we have:

Theorem 2.14 The area of a spherical triangle is equal to its excess angle,
α + β + γ − π.

Proof. Since the sphere has area 4π, a lune of angle θ has area 4θ. The
three lunes coming from a given triangle T cover the whole sphere, with
points inside T and its antipode triply covered and the rest simply covered.
Thus 4π + 4 area(T ) = 4(α + β + γ).

This result is a special case of the Gauss-Bonnet formula:

Z Z
2πχ(X) = K+ k.
X ∂X

The general formula can be deduced from the constant curvature formulas
(for K = ±1) by subdivision.

2.3 The hyperbolic plane

We now turn to a discussion of the hyperbolic plane H = {z ∈ C : Im(z) >
0}. This is the most versatile of the simply-connected Riemann surfaces.
Note that H is isomorphic to ∆, e.g. by the Möbius transformation z 7→
z−i
z+i which sends (i, 0, ∞) to (0, −1, 1). One often uses these two models
interchangeably.
First, observe that the Schwarz reflection principle implies:

Theorem 2.15 Every automorphism of H or ∆ extends to an automor-

b
phism of C.

Corollary 2.16 Aut(H) corresponds to the subgroup SL2 (R) ⊂ SL2 (C).

Proof. Suppose g(z) = (az + b)/(cz + d) preserves H. Then g preserves

∂H = R b = R ∪ {∞}, so we can assume the coefficients of g are real. Then
Im g(i) = Im(ai + b)(−ci + d)/|ci + d|2 = (ad 2
p − bc)/|ci + d| > 0,
so det(g) >
0. Thus we can be further rescale by 1/ det(g) so that ac db ∈ SL2 (R).
Conversely, every Möbius transformation represented by a matrix in SL2 (R)
preserves H.

35
Corollary 2.17 Aut(∆) corresponds to the subgroup SU(1, 1) of isometries
of the form |Z|2 = |Z0 |2 − |Z1 |2 .

Proof. This can be deduced from the preceding result by a change of

coordinates. Note that the vectors in C2 with |Z|2 = 0 correspond to ∂∆,
and ∆ itself corresponds to the cone |Z|2 < 0.

Corollary 2.18 The automorphism group of the disk is given by the sub-
group ( ! )
a b 2 2
Aut(∆) = : |a| − |b| = 1 ⊂ PSL2 (C).
b a


Note that these matrices satisfy AQA∗ = Q, where Q = 1 0
0 −1 .
Hyperbolic geometry. The hyperbolic metric on H is given by ρ =
|dz|/ Im z = |dz|/y. On the unit disk ∆, the hyperbolic metric becomes
ρ = 2|dz|/(1 − |z|2 ).

Theorem 2.19 Every automorphism of H is an isometry for the hyperbolic

metric.

Proof. This is clear for automorphisms of the form g(z) = az + b, and

can be easily checked by g(z) = −1/z. These two types of automorphisms
generate Aut(H).

The geodesics for the hyperbolic metric are given by circles orthogonal
to the boundary (of H or ∆). For example, in the case of the imaginary
axis γ = iR+ ⊂ H, it is easy to see that the projection π : H → γ given
by π(x, y) = (0, y) is distance-decreasing, and so γ gives a shortest path
between any two of its points. The general case follows from this one, since
any circle orthogonal to the boundary of H is equivalent, under Aut(H), to
γ.
These geodesics satisfy all of Euclid’s postulates except the fifth. Thus
if we declare them to be straight lines, we find:

Theorem 2.20 Euclid’s fifth postulate cannot be deduced from the other
axioms of geometry.

36
Triangles. Gauss-Bonnet for hyperbolic triangles reads:

area(T ) = π − α − β − γ.

For example, an ideal triangle in H has vertices at infinity and internal angles
of 0. Its area is π.
Curvature. We remark that the Gauss curvature of a conformal metric
ρ(z) |dz| is readily computed in terms of the Laplacian of ρ: we have

K(ρ) = −ρ−2 ∆ log ρ.

Note that this expression is invariant under change of coordinates, since an

analytic function satisfies ∆ log |f ′ | = 0. It is particularly easy to check that
K(1/y) = −1, so (H, |dz|/y) has constant curvature −1.
Classification of isometries. Let f : H → H be an automorphism. Then
f is an isometry, and we can define its translation distance by

τ (f ) = inf d(x, f (x)).

H

This invariant is useful for the classification of isometries. Here are the
possibilities:
1. (Elliptic) τ (f ) = 0, achieved. Then f is conjugate to a rotation fixing
i, of the form ! !
a b cos θ − sin θ
= .
c d sin θ cos θ
(Note that when θ = π above, the matrix is −I and so f (z) = z.) In
the disk model, we can simply conjugate so f (z) = e2iθ z.

2. (Hyperbolic) τ (f ) > 0. In this case τ (f ) is achieved along a unique

geodesic γ ⊂ H which is translated distance τ (f ) by f . Up to conju-
gacy, f (z) = et z where t = τ (f ).

3. (Parabolic) τ (f ) = 0, not achieved. Then f is conjugate to f (z) =

z + 1.

The Schwarz Lemma. The following fundamental fact is central to the

theory of analytic maps between hyperbolic Riemann surfaces.

Theorem 2.21 Let f : ∆ → ∆ be an analytic function such that f (0) = 0.

Then |f ′ (0)| ≤ 1 and |f (z)| ≤ |z| for all z.
Equality holds (for some z 6= 0) iff f (z) = eiθ is a rotation.

37
Proof. The maximum principle implies that the analytic function g(z) =
f (z)/z satisfies |g(z)| ≤ 1. If equality holds, then g is constant.

Corollary 2.22 A holomorphic map f : ∆ → ∆ is either an isometry, or it

is distance decreasing for the hyperbolic metric. (This means |f ′ |ρ < 1 and
d(f (x), f (y)) < d(x, y) if x 6= y.)

Remark. The Schwarz lemma, instead of Schwarz reflection, can also be

b
used to show every automorphism of H or ∆ extends to C.
The hyperbolic metric on other Riemann surfaces. Now suppose,
more generally, that X = H/Γ is a hyperbolic Riemann surface. (Almost all
Riemann surfaces are of this type — only C, b C, C∗ and C/Λ not hyperbolic.
In particular, every region U ⊂ C other than C itself is hyperbolic.)
Since ρ = |dz|/y is invariant under Aut(H), it descends to give a hyper-
bolic metric on X itself. (Also called the Poincaré metric.)
Example. The hyperbolic metric on the strip S = {z = x+iy : 0 < |y| < π}
is given by ρ = |dz|/ sin(y). Indeed, the map f (z) = ez sends S conformally
to H, and so

|ez | |dz| ex |dz| |dz|

ρ = f ∗ (|dz|/y) = z
= x
= .
Im e e sin(y) sin(y)

The more general version of the Schwarz lemma then reads:

Theorem 2.23 Let f : X → Y be an analytic map between hyperbolic

Riemann surfaces. Then f is either a covering map, or it is a contraction
for the hyperbolic metric.

Proof. Pass to the universal covers of the domain and range and apply the
usual Schwarz lemma.

Proper maps. A Blaschke product B : ∆ → ∆ is a rational map of the

form
d
Y z − ai
f (z) = eiθ
1
1 − ai z

with zeros satisfying |ai | < 1. Using the fact that 1 = zz on S 1 , it is easy
to show that |B(z)| = 1 on S 1 , and thus B : ∆ → ∆ is proper. Conversely,
we have:

38
Theorem 2.24 Every proper analytic map f : ∆ → ∆ is a Blaschke prod-
uct.

Proof. A proper map is surjective, so f has at least one zero, say a. Let
M (z) = (z − a)/(1 − az). Then g = f (z)/M (z) : ∆ → ∆ is analytic, with
one fewer zero than f . The proof now follows by induction on the degree of
f.

Dynamics and the Schwarz Lemma.

Theorem 2.25 Let f : C b→C b be a rational map of degree d > 1. Then the
immediate basin of any attracting cycle contains a critical point.

Corollary 2.26 The map f has at most 2d − 2 attracting cycles.

Theorem 2.27 (Denjoy-Wolff ) Let f : ∆ → ∆ be a holomorphic map

which is not conjugate to z 7→ eiθ z. Then there exists a point p ∈ ∆ such
that f n (z) → p for all z ∈ ∆.

3 Entire and meromorphic functions

This section discusses general constructions of functions on C with given ze-
ros and poles, and analyzes special cases such as the trigonometric functions
and Γ(z).

3.1 Infinite products

Q ∗
P
To start we recall that (1Q+ an ) converges (to an element of C ) iff |an |
converges. (Note:
Q we say (1 − 1/n) diverges to zero.) P
Similarly, (1 + fn (z)) defines an entire function on C iff |fn (z)|
converges uniformly on compact sets.
Functions with infinitely many zeros. Given an → ∞ in C, can one
find an analytic function f (z) whose zeros are exactly these points? Is this
function essentially unique, given e.g. some constraints on its growth?
For example, is sin(πz) simply a ‘polynomial’ with roots at all points of
Z?
Polynomials work for finitely many zeros. We Q could try to address the
case of infinitely many zeros by defining f (z) = (z − ai ). But this product
Q no chance of converging. A better choice – assuming ai 6= 0 — is
has
(1 − z/ai ). And this works: we have

39
 P Q
Theorem 3.1 If 1/|an | is finite, then f (z) = (1 − z/an ) defines an
entire function with zeros exactly at these points.

Multiplicities of zeros can also be specified – they just correspond to

repetitions of the same number in the sequence ai .
However this result is still too weak to address the case an = n needed
for constructing sin(πz).
Weierstrass factors. Here is an elegant expression for an entire function
with a zero only at z = 1, which is also close to 1 for |z| < 1. It is called the
Weierstrass factor of order p:
 
z2 zp
Ep (z) = (1 − z) exp z + + ··· + .
2 p

By convention, E0 (z) = (1 − z).

The idea behind this expression is that log(1/(1− z)) = z + z 2 /2+ z 3 /3+
· · · , and hence the two terms ‘almost cancel’ to give (1 − z)/(1 − z) = 1.
(For more insight, compute the logarithmic derivative Ep′ /Ep or see the proof
below.)
Since we have truncated at the term z p , it is easy to see that for |z| < 1/2
(say) we have
|Ep (z) − 1| = O(|z|p+1 ).
For many purposes this bound is sufficient, however it is sometimes useful
to have a bound which works for any z ∈ ∆ and where the implicit constant
(which might depend on p) is explicit. Such a bound is provided by:

Theorem 3.2 For |z| < 1, we have |Ep (z) − 1| ≤ |z|p+1 .

Proof. The negative of the logarithmic derivative of Ep , we obtain:

−Ep′ (z)/Ep (z) = 1/(1 − z) − 1 − z − . . . z p−1 = z p /(1 − z).

Hence for all z we have

∞
X
−Ep′ (z) p 2 p
= z exp(z + z /2 + · · · z /p) = z p
ak z k ,
0

with ak ≥ 0 for all k. Integrating term by term and using the fact that
Ep (0) = 1, we find
∞
X
p+1
1 − Ep (z) = z bk z k
0

40
 P
with bk ≥ 0. We also have bk = 1 − Ep (1) = 1, and hence for |z| < 1 we
have X
|1 − Ep (z)| ≤ |z|p+1 bk = |z|p+1 .

Theorem 3.3 For Q any sequence of nonzero complex numbers an → ∞,

the formula f (z) = ∞ 1 En (z/an ) converges for all z and defines an entire
analytic function with zero set exactly (an ).

Proof. The previous estimate yields convergence of the tail of the series:
for all z ∈ B(0, R), we have:

X ∞
X ∞
X
|1 − En (z/an )| ≤ (|z|/|an |)n+1 < (R/2R)n+1 < ∞.
|an |>2R 1 1

Canonical products. In general using En (z/an ) is overkill. Then the

same argument shows:
P Q
Theorem 3.4 If 1/|an |p+1 < ∞, then P (z) = Ep (z/an ) defines an
entire analytic function.

Proof. For |z| < R we have

X ∞
X X
|1 − Ep (z/an )| (|z|/|an |)p+1 < ∞. ≤ Rp+1 1/|an |p+1 < ∞.
|an |>2R 1

P
≥ 0 is the least integer such that
If p Q 1/|an |p+1 < ∞, then we say
P (z) = Ep (z/an ) is the canonical product associated to (an ).
The counting function. It is also useful to introduce the counting function
−β
P(r) = |{n β: |an | < r}|. Then r N (r) is a rough approximation to
N
|an |<r 1/|an | . Consequently we have:

log N (r)
α = lim sup ·
r→∞ log r

41
In more detail: suppose N (r) ≤ r β ; then we can collect the points an into
groups where 2n < |an | ≤ 2n+1 ; then
X X X
|an |−α ≤ (2n )−α N (2n+1 ) ≤ 2β 2n(β−α) < ∞

if α > β. So β is an upper bound for the P critical exponent. Similarly, if

N (r) ≥ r β , then β is a lower bound, since |an |−α ≥ r −αN (r) ≥ r β−α → ∞
if α < β.
Observe that knowledge of N (r) is the same as knowledge of rn = |an |
for all n. Thus we can also express functions of rn in terms of N (r). A
typical example is:
X Z ∞
−α dr
|an | = N (r)αr −α ·
0 r

A similar expression will arise in connection with Jensen’s formula.

Entire functions of finite order. An entire function f : C → C is of
finite order if there is a ρ > 0 such that |f (z)| = O(exp |z|ρ ). The infimum
of all such ρ is the order ρ(f ). It is standard to denote the maximum and
minimum of |f | on |z| = r by M (r) and m(r). Thus the order f is given by

log log M (r)

ρ(f ) = lim sup .
r→∞ log r

Examples: Polynomials have order 0; sin(z), cos(z) and exp(z) have order
√
1; cos( z) has order 1/2; Ep (z) has order p; exp(exp(z)) has infinite order.
Hadamard’s 3-circles theorem. Here is useful general property of the
function M (r) = sup|z|=r |f (z)|. The following result applies not just to
entire functions, but also to functions analytic in an annulus of the form
r1 < |z| < r2 .

Theorem 3.5 For any analytic function f (z), the quantity log M (r) is a
convex function of log r.

Proof. A function φ(s) of one real variable is convex if and only if φ(s) +
ar satisfies the maximum principle for any constant a. This holds for
log M (exp(s)) by considering f (z)z a locally.

42
 √ p
Corollary 3.6 We have M ( rs ≤ M (r)M (s).

The convex functions satisfying F (log r) = log M (r) look roughly linear
for polynomials, e.g. like F (x) = deg(f )x + c, and look roughly exponential
for functions of finite order, e.g. F (x) = exp(ρ(f )x).
Hadamard’s factorization theorem. We can now state a formula that
describes every entire function of finite order in terms of its zeros and an
additional polynomial.

Theorem 3.7 (Hadamard) A entire function f (z) 6= 0 of finite order ρ

can be uniquely expressed in the form:
Y
f (z) = z m Ep (z/an )eQ(z) ,
P
where (an ) are the zeros of f , p ≥ 0 is the least integer such that 1/|an |p+1 <
∞, and Q(z) is a polynomial of degree q. We have p, q ≤ ρ.

The number p is called the genus of f . Note that ordinary polynomials

arise as a special case, where p = q = ρ = 0.
Remark. This theorem shows that the zeros of f determine f up to finitely
many additional constants, namely the coefficients of Q(z). It is tempting to
conclude then that if f (z) has no zeros, it is determined by its values at any
⌊ρ + 1⌋ points. This is not quite true, however, since f (z) only determines
Q(z) mod 2πiZ; for example, exp(2πiz) and the constant function 1 agree
on the integers.
It is, however, true that f ′ /f = Q′ in this case, so knowing the logarith-
mic derivative at enough points almost determines Q. This shows:

Corollary 3.8 Suppose f (z) and g(z) are entire functions of order ρ with
the same zeros, and f ′ /f = g ′ /g at ⌊ρ⌋ distinct points (where neither func-
tion vanishes). Then f is a constant multiple of g.

I. Entire functions without zeros. The simplest case of Hadamard’s

theorem in which there is no canonical product, is handled by:

Theorem 3.9 Let f (z) be an entire function of finite order with no zeros.
Then f (z) = eQ(z) , where Q(z) is a polynomial of degree d, and ρ(f ) = d.

To prove it we strengthen our earlier characterization of polynomials by

M (r) = O(r d ).

43
Lemma 3.10 Let Q(z) be an entire function satisfying Re Q(z) ≤ A|z|d +B
for some A, B > 0. Then Q is a polynomial of degree at most d.

Proof. There is a constant C > 0 such that for R > 1, Q maps ∆(2R) into
the half-plane U (R) = {z : Re z < CRd }. By the Schwarz Lemma, Q is
distance-decreasing from the hyperbolic metric on ∆(2R) to the hyperbolic
metric on U (R). Since ∆(R) ⊂ ∆(2R) has bounded hyperbolic diameter,
the same is true for Q(∆(R)) ⊂ U (R). Therefore in the Euclidean metric,

diam Q(∆(R)) = O(d(Q(0), ∂U (R))) = O(Rd ).

This shows |Q(z)| = O(|z|d ) for |z| > 1, and hence Q is a polynomial of
degree at most d.

Proof of Theorem 3.9. Since f has no zeros, f (z) = eQ(z) for some entire
d
function Q(z). Since f has finite order, |f (z)| = O(e|z| ) for some d, and
thus Re Q(z) ≤ |z|d + O(1); now apply the Lemma above.

II. Functions with zeros. Next we analyze entire functions with zeros.
We will show:

Theorem
P 3.11ρ+ǫLet f (z) be an entire function of order ρ with zeros (an ).
Then 1/|an | < ∞ for all ǫ > 0.

In general, a P sequence an → ∞ has a critical exponent α, the least

number such that 1/|an |α+ǫ < ∞. The result above states that the critical
exponent of the zeros of f satisfies
P α ≤ ρ(f ). Thus if p = ⌊ρ(f )⌋, then
p + 1 > ρ(f ) ≥ α, and hence |an |p+1 < ∞. This gives:

Corollary 3.12 The genus of f satisfies p ≤ ρ(f ).

Jensen’s formula. Informally, the result above says that if f has many
zeros, then M (r) must grow rapidly. The relation between the size of f and
its zeros is made precise by the following important result:

Theorem 3.13 (Jensen’s formula) Let f (z) be a holomorphic function

on B(0, R) with zeros a1 , . . . an . Then:
X R
avgS 1 (R) log |f (z)| = log |f (0)| + log .
|ai |

44
Proof. We first note that if f has no zeros, then log |f (z)| is harmonic and
the formula holds. Moreover, if the formula holds for f and g, then it holds
for f g; and the case of general R follows from the case R = 1.
Next we verify that the formula holds when f (z) = (z − a)/(1 − az) on
the unit disk, with |a| < 1. Indeed, in this case log |f (z)| = 0 on the unit
circle, and log |f (0)| + log(1/|a|) = log |a/a| = 0 as well.
The general case now follows, since a general
Q function f (z) on the unit
disk can be written in the form f (z) = g(z) (z − ai )/(1 − ai z), where g(z)
has no zeros.

Remark. The physical interpretation of Jensen’s formula is that log |f | is

the potential for a set of unit point charges at the zeros of f .
Counting zeros. Here is another way to write Jensen’s formula. Let N (r)
be the number of zeros of f inside the circle of radius r. Then:
Z R
dr
N (r) = avgS 1 (R) log |f (z)| − log |f (0)|.
0 r

We can now show α ≤ ρ.

Proof of Theorem 3.11. Since N (r) is an increasing function, by inte-
grating from r/2 to r we find

N (r/2) log(r/2) ≤ log M (r) + O(1),

and hence α = lim sup log N (r)/ log r ≤ lim sup log M (2r)/ log(2r) = ρ.

Of course functions can also grow rapidly without having any zeros. But
then Jensen’s formula shows the average of log |f | is constant over every
circle |z| = R; so if f is large over much of the circle, it must also be close
to zero somewhere on the same circle. (For example, ez is very large over
half the circle |z| = R, and very small over the rest.)
III. Canonical products. Our next task is to determine the order of a
canonical product. It will also be useful, to complete the proof of Hadamard’s
theorem, to obtain lower bounds for such a product.

Theorem 3.14 Let α be the criticalQ exponent of the sequence an → ∞.

Then the canonical product P (z) = Ep (z/an ) has order ρ(P ) = α.

Proof.
P Let rp+1n = |an | and r = |z|. Recall that
P p pis the least integer such
that (1/rn ) < ∞. So we also have 1/rn = +∞. This implies

45
 P
p ≤ α ≤ p + 1. For convenience we will assume (1/rn )α < ∞. (For the
general case, just replace α with α + ǫ.)
As we have seen, the Weierstrass factor
 
z2 zp
Ep (z) = (1 − z) exp z + + ··· +
2 p

satisfies, for ‘small z’ meaning |z| < 1/2, the inequality |1−Ep (z)| ≤ |z|p+1 ≤
1/2, and hence also the inequality

| log Ep (z)| = O(|z|p+1 ). (3.1)

On the other hand, for ‘large z’, meaning |z| ≥ 1/2, we have

| log Ep (z)| = O(| log(1 − z)| + |z|p ). (3.2)

(To be precise — since the logarithm is multivalued — these inequalities

mean the logarithm can be chosen so the desired bound holds.) The log(1−z)
can term be ignored unless z is very close to 1, in which case Ep (z) is close
to zero. So it can also be ignored when boundingQ Ep (z) from above.
Combining these estimates for P (z) = Ep (z/an ), we get the upper
bound:  
X 1 X 1
log |P (z)| ≤ O r p+1 p+1 + r
p
p
.
r >2r rn r n
n rn ≤2r

Now in the second sum, since p ≤ α, we have

X 1 X rnα−p X 1
α−p
p = ≤ (2r) = O(r α−p ).
rn rnα rnα
rn ≤2r

Similarly in the first sum, since α ≤ p + 1, we have

X 1 X 1 1 X 1
p+1 = p+1−α r α ≤ (2r)α−p−1 = O(r α−p−1 ).
rnα
rn >2r rn rn >2r rn n

Altogether this gives

log |P (z)| ≤ O(r α ).
Thus log M (r) = O(r α ) and hence ρ(P ) ≤ α. By Jensen’s theorem, the
order of P (z) is exactly α.

46
The minimum modulus. To control the result of division by a canonical
product, we now estimate its minimum modulus. As an example — for
sin(z), we have m(r) ≍ 1 for infinitely many r.

Theorem 3.15 The minimum modulus of the canonical product P (z) above
satisfies m(r) ≥ exp(−r α+ǫ ) for arbitrarily large r.

Proof. We must show | log m(r)| = O(r α+ǫ ). The follows the same lines as
the bound log M (r) = O(r α ) just obtained, since (3.1) and (3.2) give bounds
for | log Ep (z)|. The only nuance is that we cannot ignore the logarithmic
term in (3.2). We must also decide which values of r to choose, since m(r) =
0 whenever r = |an |.
To this end, we fix ǫ > 0 and exclude from consideration the balls Bn
defined by |z − an | < 1/rnα+ǫ . Since the sum of the radii of the S excluded
balls in finite, there are plenty of large circles |z| = r which avoid Bn .
To complete the proof, it suffices to show that for z on such circles, we
have X
| log(1 − z/an )| = O(r α+ǫ ).
|z−an |<rn

Note that the number of terms in the sum above is at most N (2r) = O(r α ).
Because we have kept z away from an , we have

| log(1 − z/an )| = O(log r).

Consequently
X
| log(1 − z/an )| = O(N (2r) log r) = O(r α+ǫ ),
|z−an |<rn

as desired.

Proof of the Hadamard factorization theorem. Let f (z) be an en-

tire function of order ρ with zeros (ai ), and let P (z) be the corresponding
canonical product. Then P also has order ρ, as we have just seen. Since f
and P have the same zeros, the quotient f /P is an entire function with no
zeros. The lower bound on m(r) just established implies that f /P also has
order ρ, and hence f /P = exp Q(z) where deg Q ≤ ρ.

47
Trigonometric functions. As a first example, we determine the canonical
factorization of the sine function:

Theorem 3.16 We have

Y z2

sin(πz) = πz 1− 2 .
n
n6=0

Proof. Indeed, the right hand side is a canonical product, and sin(πz) has
order one, so the formula is correct up to a factor exp Q(z) where Q(z) has
degree one. But since sin(πz) is odd, we conclude Q has degree zero, and
by checking the derivative at z = 0 of both sides we get Q = 0.

Use of the logarithmic derivative. For later application, we record some

useful properties of the logarithmic derivative f ′ /f of an entire function f (z).

1. We have (f g)′ /f g = f ′ /f + g ′ /g.

2. If f ′ /f = g′ /g, then f = Cg for some constant C 6= 0.

3. We have f ′ (az + b)/f (az + b) = a(f ′ /f )(az + b).

Sine, cotangent and zeta. The product formula above gives, under log-
arithmic differentiation,
∞
(sin(πz))′ 1 X 1 1
= π cot(πz) = + + ·
sin(πz) z z−n z+n
1

This formula is useful in its own right: it shows π cot(πz) has simple poles at
all points of Z withPresidue2kone. This property can be used, for example, to
evaluate ζ(2k) = ∞ 1 1/n and other similar sums by the residue calculus
— see the examples in Chapter 1.
We note that the product formula for sin(z) can also be used to prove
ζ(2) = π 2 /6, by looking at the coefficient
P of z 2 on bothPsides of the equa-
tion. The sine formula also shows a<b 1/(ab) = π /5!, a<b<c 1/(abc)2 =
2 4

π 6 /7!, etc., so with some more work it can be used to evaluate ζ(2k). For
example, we have
X  X  !
1 1 X 1 π4 π4 π4
ζ(4) = − 2 = − = ·
a2 b2 (ab)2 36 60 90
a<b

48
Translation and duplication formulas. Many of the basic properties
of the sine and cosine functions can be derived from the point of view of
the uniqueness of an odd entire function with zeros at Zπ. For example,
equations sin(z + π) = sin(z) and sin(2z) = 2 sin(z) sin(z + π/2) hold up to
a factor of exp(az + b) as a consequence of the fact that both sides have the
same zero sets.

3.2 The Gamma Function

We now turn to a discussion of the extension of the factorial function n 7→ n!
to the complex numbers.
We begin by studying the canonical product associated to the negative
integers: namely
∞ 
Y z  −z/n
G(z) = 1+ e .
n
1

It has half the terms that appear in the factorization of sin(πz); indeed, we
have
sin(πz)
G(z)G(−z) = · (3.3)
πz
Euler’s constant. We have G(0) = 1, but what is G(1)? To answer this,
we define Euler’s constant by:

γ = lim (1 + 2 + · · · + 1/n) − log(n + 1).

n→∞

This
R n+1 expression gives the limiting error obtained when one approximates
1 dx/x by n unit rectangles lying above the graph of y = 1/x. (The sum
is finite because these areas can all be slid horizontally to lie disjointly inside
a fixed rectangle of base one.)
Then, using the fact that (1 + 1)(1 + 1/2) · · · (1 + 1/n) = (n + 1), we find
G(1) = exp(−γ).
Functional equation. The functions G(z − 1) and zG(z) have the same
zeros. How are they related? By Hadamard’s theorem, we have G(z − 1) =
z exp(az + b)G(z) for a, b. In fact:

Theorem 3.17 We have G(z − 1) = zeγ G(z).

Proof. The logarithmic derivative of zG(z) is given by:

∞
1 X 1 1
+ − ,
z z+n n
1

49
while the logarithmic derivative of G(z − 1) is just
∞
X X 1 ∞
1 1 1 1
− = −1+ − .
1
z+n−1 n z 1
z+n n+1
P∞
Since 1 (1/n − 1/(n + 1)) telescopes to 1, these series are equal, and hence
G(z − 1) and zG(z) agree up to a constant. The value of this constant is
determined by setting z = 1.

10

-3 -2 -1 1 2 3

-5

-10

Figure 5. The Γ function.

The Γ function. We now define

1
Γ(z) = .
zeγz G(z)

Note especially that we have added a factor of exp(γz) to compensate for

the multiplicative factor of eγ . This equation is sometimes written more
explicitly as
∞ 
Y z −1 z/n
zΓ(z) = e−γz 1+ e .
n
n=1
From the discussion above we find:

1. Γ(z + 1) = zΓ(z); Γ(1) = 1; and hence

2. Γ(n + 1) = n! for n ≥ 0;

3. Γ(z) has no zeros; and

4. Γ(z) has simple poles at 0, −1, −2, −3, . . . .

50
Using the fact that
Γ(z + 1)
Γ(z − n) = ,
(z − n) · · · (z − 1)z
we find
(−1)n
Res−n (Γ(z)) = ·
n!
From (3.3) we obtain Euler’s supplementj
π
= Γ(z)Γ(1 − z),
sin(πz)
which implies, for example:
√
Γ(1/2) = π.

Variants of the sine formula are:

   
1 1 π π
Γ +z Γ −z = and Γ(z)Γ(−z) = − ·
2 2 cos πz z sin(πz)

Using the fact that Γ(z) = Γ(z), we find from this last that
π
|Γ(iy)|2 = ·
y sinh(πy)

We also note that by the functional equation, Γ(z) has constant sign on
each interval of the form (n, n + 1); in fact the sign is positive for n > 0 and
(−1)n for n < 0.
Gauss’s formula. The formula above for the Γ function would not be easy
to discover from scratch. Here is a formula that could also be taken as a
simpler definition of the Gamma function.

Theorem 3.18 We have

n!nz
Γ(z) = lim ·
n→∞ z(z + 1) · · · (z + n)

−1
As motivation, we note that for integers z we have z! = lim nz z+n n ,
which is consistent with the formula above if we multiply both sides by z.
Indeed, formally we have
 
z+n (z + 1) · · · (z + n) (n + 1) · · · (n + z) nz
= = ≈ ,
n n! z! z!

51
and solving for z! and taking the limit gives the formula above.
Proof. By definition, we have:
n 
1 γz
Y z  −z/k
= lim ze 1+ e .
Γ(z) n→∞ k
k=1

Now γ − 1 − 1/2 − · · · − 1/n ≈ − log(n), and (1 + z/k) = (k + z)/k, so this

becomes
1 z(z + 1) · · · (z + n)n−z
= lim ,
Γ(z) n→∞ n!
which gives the formula above.

Corollary 3.19 We have |Γ(z)| ≤ |Γ(Re z)| for all z.

Proof. This inequality holds for each term in the product above.

This Corollary gives us good estimates on Γ(z): it is uniformly bounded

in any region of the form 0 < a < Re(z) < b, and for −n < Re z < b the
function z(z + 1) · · · (z + n)Γ(z) is uniformly bounded.
Characterizing Γ(z). In fact these properties are enough to characterize
the Γ function.
Note that exp(2πiz)Γ(z) satisfies the same functional equation as Γ(z),
and its reciprocal still has order 1; but it violates the growth condition. The
boundedness is also needed.

Theorem 3.20 (Wielandt) If F (z + 1) = zF (z) for Re z > 0, F (1) = 1

and F (z) is bounded on the strip {z : Re z ∈ [1, 2]}, then F (z) = Γ(z).

Proof. The functional equation allows one to extend F (z) to a meromorphic

function on the whole plane, whose poles and their residues agree with those
of Γ(z). Thus G(z) = F (z) − Γ(z) is entire, G(0) = G(1) = 0 and G(z + 1) =
zG(z). Our boundedness assumptions now imply that G(z) = G(z + 1)/z is
bounded in the strip S = {Re z ∈ [0, 1]}, since it has a removable singularity
at z = 0. Thus H(z) = G(z)G(1 − z) is also bounded in S. The functional
equation for G implies H(z + 1) = −H(z) (as in the sine formula), and thus
H is a constant, which must be zero.

52
A duplication formula. As an application one can prove ‘multiple angle’
formulas for Γ(nz). The simplest is:
√
Corollary 3.21 We have 2 πΓ(2z) = 22z Γ(z)Γ(z + 1/2).

Proof. Let F (z) = Γ(z/2)Γ(z/2 + 1/2). Then F (z + 1) = Γ(z/2 +

1/2)Γ(z/2+ 1) = zF (z)/2. So if write instead F (z) = 2z Γ(z/2)Γ(z/2+ 1/2),
then F (z + 1) = zF (z). Since F (z) is also bounded for Re z ∈ [1, 2],
√ we have
F (z) = CΓ(z) for some C; and indeed, C = F (1) = 2Γ(1/2) = 2 2π. The
result above now follows, upon replacing z with 2z.

The integral representation: the Mellin transform. We now turn to

a second motivation for introducing the Γ function.

Theorem 3.22 For Re(z) > 0, we have

Z ∞
dt
Γ(z) = e−t tz .
0 t

In other words, Γ(z) is the Mellin transform of the function e−t on R∗ .

The Mellin transform is an integral against characters χ : R∗ → C∗ (given
by χ(t) = tz ), and as such it can be compared to the Fourier transform (for
the group R under addition) and to Gauss sums. Indeed the Gauss sum
X
σ(χ) = χ(n)e2πin/p
(n,p)=1

is the analogue of the Gamma function for the group (Z/p)∗ .

Proof. Apply the uniqueness theorem above.

Stirling’s formula. Here is a very brief introduction to the method of

steepest descent, with the aim of explaining Stirling’s formula: as s → ∞,
we have: r  
2π s s
Γ(s) ∼ ·
s e
R∞
The idea is to model Γ(s) = 0 e−t ts−1 dt by the more easily understood
Gaussian integral:
Z ∞ p
2
ea−b(t−t0 ) /2 dt = ea 2π/b,
−∞

53
which is itself proved using the identity
Z ∞ 2 Z
−x2 /2 2
e dx = e−r /2 r dr dθ = 2π.
−∞
R∞
To this end we rewrite Γ(s) = 0 eφ(t) dt, where φ(t) = −t + (s − 1) log t.
Then φ′ (t) = −1+(s−1)/t vanishes at t0 = s−1, with φ′′ (t0 ) = −(s−1)/t20 =
−1/(s − 1). So we have

φ(t) ≈ φ(t0 ) + φ′′ (t0 )(t − t0 )2 /2 = a − b(t − t0 )/ 2

with a = 1 − s + (s − 1) log(s − 1) and b = 1/(s − 1). This gives

p
Γ(s) ∼ e1−s (s − 1)s−1 2π(s − 1)·
√ √
Using the fact that (s − 1)s ∼ ss /e and s − 1 ∼ s, this gives Stirling’s
formula.
As a simple check, we note that Stirling’s formula implies:

lim (n!/nn )1/n = 1/e.

This can be seen by interpreting (1/n)(2/n)(3/n)

R1 · · · (n/n) as the exponen-
tial of an integral, and using the fact that 0 log x dx = −1.
The name steepest descent comes from the fact that the real axis passes
through the saddle pass of the absolute value of the integrand at its criti-
cal point, and proceeds in the direction where the absolute value decreases
fastest. Compare [MH, p. 438].

3.3 Meromorphic functions: The Mittag-Leffler theorem

We now turn to the problem of describing meromorphic functions with pre-
scribed principal parts. This means we specify a sequence of distinct points
an → ∞, and a finite Laurent tail
bk b1
pn (z) = k
+ ··· +
(z − an ) z − an
at each point. (The values of k and b1 , . . . , bk depend on n.) We say pn (z)
is the principal part of a meromorphic function f (z) if f (z) − pn (z) is holo-
morphic at z = an .

Theorem 3.23 For any sequence of points an → ∞ and principal parts

pn (z), there exists a meromorphic function f (z) with poles exactly at an ,
and with the prescribed principal parts.

54
Proof. Let qn (z) be the polynomial given by truncating the power series
−n when |z| < |a |/2. Then f (z) =
P pn (z) so that |pn (z) − qn (z)| < 2
for n
(pn (z) − qn (z)) is the desired function.

Singly-periodic functions. (This is a hint of the theory of elliptic or

doubly-periodic functions to come.)
Any meromorphic function satisfying f (z+1) = f (z) has the form f (z) =
F (e2πiz ) for some meromorphic function F : C∗ → C. b This leads to formulas
such as:
∞
X 1 π2
= ·
−∞
(z − n)2 sin2 (πz)
Note that the right hand side is a sum of principal parts. Also, compare
this to the series for π cot(πz).
To verify this equation, first note that |z − n|2 ≥ | Im z|2 + n2 − O(1) for
0 ≤ Re z ≤ 1, and hence the sum on the left converges uniformly in this strip
so long as | Im z| ≥ 1. Thus both sides tends to 0 as | Im z| → ∞. But their
difference is holomorphic and periodic, and so it must vanish identically.
On the other hand, by evaluating theP constant term in the Laurent series
on both sides, we obtain the formula ∞ 1 1/n 2 = π 2 /6.

From Mittag–Leffler to Weierstrass. The Weierstrass Theorem — that

there exists a function g(z) with prescribed zeros — is a Corollary of the
Mittag–Leffler theorem (which historically came later).
Indeed, to construct an entire function with zeros of multiplicities mn
at distinct points an → ∞, we can simply first construct a meromorphic
R with principle part pn (z) = mn /(z − an ) at an , and then take
function f (z)
g(z) = exp( f (z) dz). Since the residues of f (z) are integers, the integral is
only well-defined modulo 2πiZ, but this is sufficient to make its exponential
well-defined.

...this topsy–turvy way of doing things (die Dinge auf dem Kopf
zu stellen) should be not sanctioned by anyone who sees mathe-
matics as something other than a disordered heap of mathemat-
ical results.
—A. Pringsheim, 1915.

(See [Re, p.131].)

55
4 Conformal mapping
We now turn to the theory of analytic functions as mappings. Here the
dominant operation is composition, rather than addition or multiplication.

4.1 The Riemann mapping theorem

A conformal homeomorphisms between plane domains is an isomorphism
in the category of Riemann surfaces. The first main result classifies these
domains when they are simply-connected.

Theorem 4.1 (Riemann mapping theorem) Let U ⊂ C be a simply

connected domain, other than C itself. Then there exists a conformal home-
omorphism f : U → ∆.

The same result holds for any simply-connected region U ⊂ C b such that
b b
|C−U | ≥ 1. Note that π1 (U ) = (1) iff C−U is connected, so the complement
is either empty, one point or a nontrivial continuum.
Since Aut(∆) is large, the Riemann map is not unique. It can be made
unique by picking a point p ∈ ∆ and requiring f (p) = 0 and f ′ (p) > 0. The
value of f ′ (p) is an interesting invariant of (U, p); its reciprocal is called the
conformal radius of U with center p.
Examples of Riemann maps.
1. We first note that ∆ ∼= H, e.g. by the Möbius transformation A(z) =
−i(z −1)/(z +1). We also remark that A : S 1 → R b becomes in angular
coordinates,
2 eiθ/2 − e−iθ/2
A(θ) = = tan(θ/2).
2i eiθ/2 + e−iθ/2
This is the change of variables used in calculus to integrate rational
functions of sine and cosine.

2. We also remark that (∆, 0) ∼= (∆, p), and (H, i) ∼

= (H, p), as can be
seen using the Möbius transformations B(z) = (z + p)/(1 + pz) and
C(z) = az + b where p = ai + b.
b bounded by a circle (or line) is also isomor-
3. Of course any region in C
phic to ∆.

4. Let L ⊂ Cb be a lune – the region between two circular arcs. Apply a

Möbius transformation so the vertices of L are at 0 and ∞; then after
a rotation, we can assume L = {z : 0 < arg z < α}, and f (z) = z π/α

56
 send L to H. (This illustrates the idea that the Riemann map U ∼
= ∆ is
often constructed as a composition of several simple transformations.)

5. The map f (z) = z + 1/z sends the region Ω = H − ∆ to H. This is a

particular case of a lune.

6. The strip S = {z : 0 < Im z < π} is a lune with angle zero; it is

sent to H by f (z) = ez . The region between any two tangent circles is
isomorphic to S by a Möbius transformation.
This is related to the problem of extending the harmonic function
u(z) = 1 for Im z > 0 and 0 for Im z < 0 from S 1 to ∆; we can take
u = Re f (z)/π where f : ∆ → S sends ±1 to the two ends of S.

7. Let us write S = S+ and S− according to the sign of Re(z). Note that

f (z) = ez sends the imaginary axis to S 1 . Thus f sends S− to ∆ ∩ H,
and sends S+ to H − ∆. Both of these regions are lunes. The map
g(z) = (z + 1/z)/2 sends f (S− ) to −H and f (S+ ) to +H.

8. As a variation on the discussion above, we observe that f (z) = cos z

maps the half-strip T ⊂ H defined by 0 < Re z < π to −H. Indeed,
cos(z) = (eiz + e−iz )/2 is simply the composition of the map z 7→ eiz ,
which sends T to ∆ ∩ H, with the map g(z) = (z + 1/z)/2.
If we reflect T through its sides, we obtain a tiling of C. The map
cos(z) sends half the tiles to H and the other half to −H, as can be
seen by Schwarz reflection.

9. One last trigonometric example: the map f (z) = sin(z) sends T =

{z ∈ H : 0 < Re z < π/2} to the first quadrant. As an alternative way
to see this, use the fact that sin(iy) = sinh(iy) and sin(π/2 + iy) =
cosh(y).
Consequently sin2 (z) maps T to H. This is consistent with the previous
example, because sin2 (z) = (1 − cos(2z))/2.

10. We note that ∆∗ is isomorphic to C − ∆ by z 7→ 1/z. The Riemann

mapping theorem shows there is a conformal map

C−∆→C−K

for any connected set K with |K| > 1.

Univalence and compactness. To begin the proof of the Riemann map-

ping theorem, we recall a few fundamental facts.

57
 Let us say f : U → C is univalent if f is injective and analytic. By
injectivity, its inverse is analytic, and f provides a homeomorphism between
U and V = f (U ). (It need not provide a homeomorphism between their
closures.) We then have:

Lemma 4.2 Let fn : U → C be a sequence of univalent maps, converging

locally uniformly to f : U → C. If f is nonconstant, it too is univalent.

Proof. Suppose f (a) = f (b) with a 6= b, but f is nonconstant. Then

g(z) = f (z) − f (b) has a zero at z = a and g(z) = lim gn (z) = fn (z) − fn (b).
It follows by Rouché’s theorem that for n ≫ 0 we have gn (an ) = 0 with
an → a. But then fn (an ) = fn (b), contrary to univalence of fn .

Lemma 4.3 The space of analytic maps f : U → ∆ is compact in the

topology of locally uniform convergence.

Proof. This is a consequence of Arzela-Ascoli and the fact that |f ′ (z)| ≤

1/d(z, ∂U ) by Cauchy’s integral formula.

Lemma 4.4 If U ⊂ C is simply-connected and 0 6∈ U , then there exists a

univalent map f : U → C such that f (z)2 = z.

Proof of the Riemann mapping theorem. Pick p ∈ U and let F denote

the space of univalent maps f : (U, p) → (∆, 0). For convenience, we will
also impose the condition that f ′ (p) > 0.
We first observe that F is nonempty. This is clear if U is bounded. It
is also true if U 6= C, for in this case we can translate to 0 6∈ U ; then the
√
image V of a particular branched of z : U → C is disjoint from −V , so by
composing with 1/(z − a), a ∈ −V , we obtain a univalent map from U to a
bounded region.
Note that if B(p, r) ⊂ U , then by the Schwarz lemma we have |f ′ (p)| ≤
1/r for all f ∈ F. Thus M = supF f ′ (p) is finite. By the preceding lemmas,
the bound M is actually realized: there exists an f ∈ F such that f ′ (p) = M .
To complete the proof, we need to show f (U ) = ∆. If not, there is an
a ∈ ∆ − f (U ). Let A : ∆ → ∆ be an automorphism such that A(a) = 0. Let
s(z) = z 2 , and construct a branched of s−1 on the simply-connected region
A(f (U )) ⊂ ∆∗ . Let B : ∆ → ∆ be an automorphism such that

g = B ◦ s−1 ◦ A ◦ f : U → ∆

58
satisfies g(p) = 0 and g ′ (p) > 0. Then g ∈ F as well.
Now notice that f = (A ◦ s ◦ B) ◦ g = h ◦ g. By construction, h : ∆ → ∆
is a proper map of degree two, with h(0) = 0. Thus |h′ (0)| < 1 by the
Schwarz lemma. But g′ (0) = h′ (0)f ′ (0), so g′ (0) > f ′ (0) = M , contrary to
the definition of M since g ∈ F.
Thus f must have been surjective after all, so it provides the desired
conformal map between U and ∆.

Boundary behavior. It is now convenient to reverse domain and range

and consider a Riemann map f : ∆ → U ⊂ C. We will investigate the
question of extending f at least to a continuous map on ∂U .
Here is one of the main results. Recall that a compact set J ⊂ C is a
Jordan curve if it is homeomorphic to a circle.

Theorem 4.5 If ∂U is a Jordan curve, then any conformal map f : ∆ → U

extends to a homeomorphism between ∆ and U .

Length–area. Here is the basic argument, exploiting the fact that f

stretches all directions by the same factor, |f ′ (z)|.
Let R(a, b) = [0, a] × [0, b] ⊂ C. For each y ∈ [0, b], let Lx = [0, a] × {y}.
Then we have

Lemma 4.6 For any conformal map f : R(a, b) → U ⊂ C, there exists a

y ∈ [0, b] such that

b2 length(f (Ly ))2 ≤ area(U ) area(R).

In other words,
p the length of the image of some horizontal line is bounded
above by area(U )a/b.

Corollary 4.7 If area(U ) = area(R(a, b)), then the image of some horizon-
tal line is shorter in U (or at least no longer).

Proof of Lemma 4.6 The average length of f (Ly ), squared, satisfies

 Z b 2 Z !2
1 1
length(Ly ) dy = 2 |f ′ (z)| |dz|2
b 0 b R(a,b)
Z Z
1 2 area(U ) area(R(a, b))
≤ 2 1 |f ′ |2 = ·
b b2
Since the average exceeds the minimum, the result follows.

59
Jordan curves. Here is a basic fact about a Jordan curve J ⊂ C. Given
a, b ∈ J, let [a, b] ⊂ J be the subarc joining these two points with smallest
diameter. Then diam[a, b] → 0 if |a − b| → 0.
Proof of Theorem 4.5. Given a point z ∈ ∂∆, map ∆ to an infinite strip,
sending z to one end. Then there is a sequence of disjoint squares in the strip
tending towards that end. The images of these squares have areas tending to
zero, so there are cross-cuts γn ⊂ ∆ enclosing z such that length(f (γn )) → 0.
Thus the endpoints of f (γn ) converge to points an , bn ∈ J, with z ∈ [an , bn ]
for n ≫ 0. Since diam[an , bn ] → 0, the disk bounded by f (γn ) ∪ [an , bn ]
shrinks to z, and this implies f (zn ) → f (z) whenever zn → z in ∆.
Consequently f extends to a continuous map S 1 → J. Since f |∆ is a
homeomorphism, f |S 1 is monotone; thus it is injective unless it is constant
on an arc. But it cannot be constant on an arc, since f is nonconstant.

Local connectivity. The key point in the proof above is the following
property of a Jordan domain U: if a, b ∈ ∂U are joined by a short arc
α ⊂ U , then the disk cut off by α has small diameter. This principle is
called ‘short dam, small lake’.
Recall that a compact set K is locally connected if for every open set
U ⊂ K and x ∈ U there is a connected open set with x ∈ V ⊂ U .
Exercise. If there exists a continuous surjective map f : S 1 → K, then K
is locally connected.
The argument in the proof of Theorem 4.5 furnishes short cross-cuts for
any Riemann map, so it also shows:

Theorem 4.8 The Riemann map extends continuously to S 1 iff ∂U is lo-

cally connected.

Finally, using Schwarz reflection one has the useful statement:

Theorem 4.9 The Riemann map extends analytically across an arc α ⊂ S 1

iff f (α) is a real-analytic arc in ∂U .

Remark. The length–area method also easily shows:

Theorem 4.10 Any Riemann mapping f : ∆ → C has radial limits almost

everywhere; that is, limr→1 f (reiθ ) exists for almost every θ.
R ′2
Proof. We may assume the imageR of′ f is bounded; then |f | < ∞, which
implies by Cauchy-Schwarz that |f | < ∞ along almost every ray.

60
 More generally, it is known that any bounded analytic function has radial
limits a.e. By [Ko], not much more can be said — given a Gδσ set A ⊂ S 1 of
measure zero, there exists a bounded function that fails to have radial limits
exactly on the set A. For Riemann mappings, radial limits exist except
outside a very small set — the exceptional set A has capacity zero and in
particular H. dim(A) = 0.
Harmonic functions and boundary values. Recalling that if u is har-
monic and f is analytic then u◦f is also harmonic, it is now a simple matter
to solve the Dirichlet problem (at least implicitly) on any Jordan domain.

Theorem 4.11 For any Jordan domain U ⊂ C, there exists a unique map
P : C(∂U ) → C(U ) such that P u|∂U = u and P u is harmonic in U .

Proof. Let f : ∆ → U be the Riemann map, and let P u = P0 (u ◦ f ) ◦ f −1 ,

where P0 is given by the Poisson kernel on the unit disk.

Annuli. A domain U ⊂ C b is an annulus if C

b − U = K1 ⊔ K2 has exactly
two connected components. Equivalently, π1 (U ) ∼ = Z. Note: in general
2 ∼
H0 (S − U ) = H1 (U ); this is an example of Alexander duality.
The standard annulus is A(R) = {z : 1 < |z| < R}. Other examples of
annuli are C∗ and ∆∗ . Exercise: none of these annuli are isomorphic.

b is isomorphic
Theorem 4.12 The universal cover of any annulus U ⊂ C
to C or H.

Proof. We may suppose the two complementary components of U contain

0 and ∞ respectively. Then the map U ֒→ C∗ induces an isomorphism
π1 (U ) ∼
= π1 (C∗ ) ∼
= Z. The universal cover of C∗ is given by π : C → C∗ ,
π(z) = ez , so
V = log U = π −1 (U ) ⊂ C
gives the universal cover of U . By the Riemann mapping theorem, V itself
is isomorphic to H or C.

Remark. Clearly U ∼ = ∆∗ or C∗ iff one or both of its complementary com-

ponents are singletons. This can be proved e.g. by applying the removable
singularities theorem to f : ∆∗ → U .

Theorem 4.13 Any annulus U is isomorphic to C∗ , ∆∗ , or A(R) for a

unique R > 1.

61
Proof. Let g : V → V denote a generator for π1 (U ) ∼ = Z acting on its
universal cover. If V ∼= C then g is conjugate to g(z) = z + 1 and we have
U ∼= C/hgi ∼ = C∗ by the map π(z) = exp(2πiz). The same reasoning shows
U∼ = ∆∗ if V ∼ = H and g is parabolic.
Otherwise V ∼ = H and we can assume g(z) = λz, λ > 1. This means the
core curve of U is a geodesic of length L = log λ. Then π(z) = z α maps V
to A(R) if we choose α correctly. We want π to map [1, λ] onto the unit
circle, so we want π(λ) = λα = exp(αL) = 1; so we take α = −2πi/L. Note
that this is a purely imaginary number, so π sends R+ to the unit circle and
R− to a circle of radius R = (−1)α = exp(απi) = exp(2π 2 /L). (We put a
minus sign in α so that R > 1.)

Modulus of an annulus. Two natural invariants of an annulus are the

number R such that U ∼ = A(R), and the hyperbolic length L of its core
geodesic. These are related by log R = 2π 2 /L as we have just seen. For a
direct proof, consider the metric |dz|/|z| on A(R). This makes it into a right
cylinder with height over circumference given by h/c = log R/2π. On the
other hand, the same metric makes H/(z 7→ eL z) into a right cylinder with
h/c = π/L. Equating these two expressions gives the desired relation.
The quantity h/c is often called the modulus of A, written mod(A).
The space of all Riemann mappings. It is traditional to denote the set
of univalent mappings f : ∆ → C such that f (0) = 0 and f ′ (0) = 1 by S,
for schlicht (plain, simple). So

f (z) = z + a2 z 2 + z3 z 3 · · ·

We give S the topology of uniform convergence on compact sets. We will

soon show:

Theorem 4.14 The space S is compact.

As a consequence, the coefficients an are bounded. The deeper Bieberbach

Conjecture, now a theorem, asserts that |an | ≤ n. This is also more than
sufficient to show that S is compact.
Remark: no nesting. There are no nesting relations among the images
of f ∈ S; that is, the Schwarz lemma implies that if f (∆) = g(∆) then
f = g. So if f (∆) contains some points outside the unit disk, it must also
omit some points inside the unit disk.
Maps to the outside. It is somewhat easier to start with the space Σ of
b−∆ → C
all univalent maps F : C b with F (∞) = ∞, normalized so that

62
their Laurent series has the form:
X∞
bn
F (z) = z + ·
n=1
zn

Since F is an open mapping near ∞, the complement of its image is a

compact set K(F ) ⊂ C.
Example. If we set b2 = 1 we get F (z) = z + 1/z which satisfies K(F ) =
[−2, 2].
P
Theorem 4.15 The area of K(F ) is given by A = π(1 − n|bn |2 ).

Proof. Integrate F dF over the unit circle and observe that, since |dz|2 =
(−i/2)dz dz, the area A of K(F ) is given by:
Z  Z dz
i i X
2
 X 
A=− F dF = − 1 − n|bn | =π 1− n|bn |2 .
2 S1 2 S1 z

P
Corollary 4.16 We have n|bn |2 ≤ 1. In particular, we have |bn | ≤
√
1/ n.

Corollary 4.17 The space Σ is compact.

Proof. This follows from the fact that |bn | ≤ n−1/2 . So if we fix R > 1,
then for |z| ≥ R we have
X
|F (z) − z| ≤ n|bn |/Rn+1 = C(R) < ∞.

This uniform bound implies every sequence has a subsequence converging

uniformly on compact sets. As usual, univalence is also preserved in the
limit.

Remark. Little is known about optimal bounds for |bn | over Σ. It is

conjectured that |bn | = O(1/n3/4 ); see [CJ].
We can now use the statement |b1 | ≤ 1 to prove the first case of the
Bieberbach conjecture. Note that the case of equality, we must have all other
bn = 0, and hence up to a rotation, F (z) = z + 1/z and K(F ) = [−2, 2].

Theorem 4.18 For all f ∈ S with have |a2 | ≤ 2.

63
Proof. We first note that F (z) = 1/f (1/z) + a2 is in Σ. Indeed, we find:
z
1/f (1/z) =
1 + a2 /z + a3 /z 2 + · · ·


= z 1 − (a2 /z + a3 /z 2 + · · · ) + (a2 /z + a3 /z 2 + · · · )2 − · · ·
a22 − a3
= z − a2 + + ···
z
and thus F (z) has b1 = a22 − a3 . So we find |a22 − a3 | ≤ 1, which is sort of
interesting, but not what we are aiming for yet.
Next we use a nice trick to make p f (∆) more symmetrical: we consider,
instead of f (z), the new map g(z) = f (z 2 ). This map is also in S and it
is given by
p
g(z) = z 1 + a2 z 2 + a3 z 2 + · · · = z + (a2 /2)z 3 + · · ·

Thus |a2 /2| ≤ 1 and hence |a2 | ≤ 2.

Corollary 4.19 S is compact.

Proof. Suppose fi ∈ S and let Fi (z) = 1/f (1/z) + a2 (i) ∈ Σ. Pass to a

subsequence so Fi (z) converges in Σ, and so a2 (i) converges. Then Gi (z) =
Fi (z) − a2 (i) converges outside the disk, and hence fi (z) = 1/Gi (1/z) con-
verges on the unit disk.

Here is a useful geometric consequence of the bound on a2 .

Theorem 4.20 (Koebe 1/4 theorem) For any f ∈ S we have B(0, 1/4) ⊂
f (∆).

Proof. Suppose p 6∈ f (∆). Note that A(z) = z/(1 − z/p) has A(0) = 0 and
A′ (0) = 1. Thus

f (z)
g(z) = = (z + a2 z 2 + · · · )(1 + z/p + · · · ) = z + (a2 + 1/p)z 2 + · · ·
1 − f (z)/p

belongs to S. Thus 2 ≥ |a2 + 1/p| ≥ |1/p| − 2 and hence |p| ≥ 1/4.

64
 From the Koebe theorem we get an important comparison between the
hyperbolic metric ρ and the ‘1/d’ metric δ = δ(z)|dz| = |dz|/d(z, ∂U ).

Corollary 4.21 The hyperbolic metric ρ(z) |dz| on a simply-connected re-

gion U ⊂ C is comparable to the 1/d metric: we have ρ(z)/δ(z) ∈ [1/2, 2]
for all z ∈ U . Equivalent, ρ(z)d(z, ∂U ) ∈ [1/2, 2].

Proof. We will check this at a given p ∈ U . Let f : (∆, 0) → (U, p) be a

Riemann mapping. We may suppose p = 0 and f ′ (0) = 1; then f ∈ S, and
ρU (p) = ρ∆ (p) = 2. By the Schwarz lemma we have d(p, ∂U ) ≤ 1 and by
Koebe we have d(p, ∂U ) ≥ 1/4; hence the result.

Comparison of S and Σ. The most important mapping in Σ is F (z) =

z + 1/z; the most important one in S is
X d 1 z
f (z) = nz n = z = ·
dz 1 − z (1 − z)2

These maps are equivalent: F (z) = 1/f (1/z) + 2. We have K(F ) = [−2, 2]
and K(f ) = (−∞, −1/4]. The map f (z) shows the Bieberbach conjecture
is sharp. We also note the problem with trying to find a ‘Bieberbach con-
jecture’ for Σ: there is no map which simultaneously maximizes all the bn ’s.
Indeed, by the area theorem, if b1 = 1 then the rest of the bn ’s are zero.
The distortion theorems. Given f ∈ S, think of f (∆) as a splattered
egg; then one finds that no matter what, the yolk f (B(0, r)), r < 1, is still
good (not too distorted). For example, the curvature of f (S 1 (r)) is bounded
by a constant Kr independent of f . Also, f (S 1 (r)) is convex if r is small
enough.
Qualitative theorems of this type can be easily deduced from compact-
ness of S. The Koebe distortion theorems make these results more precise.
They state:

Theorem 4.22 For all f ∈ S and z ∈ ∆ with |z| = r, we have

(1 − r) (1 + r)
3
≤ |f ′ (z)| ≤
(1 + r) (1 − r)3

and
r r
2
≤ |f (z)| ≤ · (4.1)
(1 + r) (1 − r)2

65
 The proof can be made rather conceptual. Let U ⊂ C be a proper
simply-connected region, and let f : U → C be a univalent map. We will
use the hyperbolic metric ρ(z) |dz| on U and the Euclidean metric |dz| on
C. It is then natural to study how these metrics compare under f . To this
end we define
δ(z) = log(|f ′ (z)|/ρ(z))
Note that if we replace f (z) with af (z) + b, it only changes δ(z) to δ(z) +
log |a|.

Lemma 4.23 The gradient of δ(z) in the hyperbolic metric on U satisfies

|dδ|/ρ ≤ 2.

Proof. We can assume U = ∆, z = 0 and f ∈ S by the Riemann mapping

theorem and by the observations above. Then ρ(z) = 2/(1 − |z|2 ) is station-
ary at z = 0. Thus |dδ| = |f ′′ (0)|/|f ′ (0)| = |2a2 | ≤ 4. Since ρ(0) = 2 we get
the bound above.

Proof of Theorem 4.22. We continue to assume U = ∆ and f ∈ S. Let

D(z) = exp δ(z) = |f ′ (z)|/ρ(z). Integrating along the hyperbolic geodesic
from 0 to z, we find
Z z Z z
|δ(z) − δ(0)| ≤ |dδ| ≤ 2ρ = 2 d(0, r),
0 0

and thus D(z)/D(0) ≤ exp(2d(0, r)), r = |z|. But exp(d(0, r)) = (1 +

r)/(1 − r), as can be seen by using i(1 − z)/(1 + z) to map to H. Thus
D(z)/D(0) ≤ (1 + r)2 /(1 − r)2 . It follows that for f ∈ S, since f ′ (0) = 1,
we have:
|f ′ (z)| (1 + r)2 ρ(r) (1 + r)
|f ′ (z)| = ′ ≤ 2
· = ·
|f (0)| (1 − r) ρ(0) (1 − r)3
The reverse inequality is similar: we have

(1 − r)2 ρ(r) (1 − r)
|f ′ (z)| ≥ 2
· = ·
(1 + r) ρ(0) (1 + r)3

Integrating these bounds gives (4.1).

66
 These results also show that S is compact.
Multiply-connected regions. We briefly mention one of the standard
forms for a region that has ‘more than two holes’.
Suppose U = C − (K1 ∪ · · · ∪ Kn ), where the Ki are disjoint compact
connected sets, none of which is a single point. We then have:

Theorem 4.24 There exists a unique conformal map F : U → C of the

form F (z) = z + b1 /z + · · · such that K(F ) consists of n disjoint horizontal
segments.

Remark: smoothing the boundary. As a preliminary remark, we note

b
that by applying the Riemann mapping to C−K i for each i, one can arrange
(if desired) that each Ki is a Jordan domain with real-analytic boundary.
An extremal problems for slits. For the proof of Theorem 4.24, it is
useful to introduce the family F of all univalent conformal maps f : U → C
of the form above. We then have the following complement:

The map F maximizes Re b1 (f ) over all f ∈ F.

The proof depends on the following observations. (Cf. [Gol, V.2].)

1. The theorem is true in the case n = 1. Indeed, in this case we can

assume U = C − ∆, and then F = Σ, and F (z) = z + 1/z. This is the
unique map maximizing Re b1 , by the area theorem.

2. For any pair of maps defined near ∞ by z+b1 /z+. . ., we have b1 (f ◦g) =
b1 (f ) + b1 (g).

3. Thus the complement is also true in the case n = 1. For in this case
we can assume (after translation) that U is the image of g ∈ Σ. Then
F = f ◦ g −1 , and Re b1 (F ) = Re b1 (f ) − b1 (g) = 1 − Re b1 (g).
But even more is true! Unless U is already a slit domain, Re b1 (F ) > 0,
by the area theorem applied to g — since |b1 (g)| ≤ 1.

4. Now return to the case of general U , and suppose F ∈ F maximizes

Re b1 . We have K(F ) = L1 ∪· · ·∪Ln . Suppose one of these components
— say L1 — is not a horizontal slit. Then we can find a map G :
C − L1 → C − [a, b] such that Re b1 (G) > 0 by what we have just
observed. But then G ◦ F ∈ F and b1 (G ◦ F ) > b1 (F ), contrary to the
extremal property of F .

67
 5. We should also check uniqueness. For this it is useful to think of
U as the interior of a compact, smoothly bounded domain in C, and
b as normalized maps of the form fi (z) = 1/(z −p)+gi (z)
f1 , f2 : U → C
whose images are horizontal slit domains. Then the bounded analytic
function h(z) = f1 (z) − f2 (z) on U also sends each component of ∂U
into a horizontal line. Thus h(∂U ) has winding number zero about
points not on these lines, so h must be constant.

Number of moduli.
S The number of n-slit regions in C, i.e. those of the
form S = C − n1 [zi , zi + ai ] with ai ∈ R, has real dimension 3n. We can
normalize by an affine transformation so the first slit is, say, [−2, 2]; so up
to isomorphism, the number of such slit regions is 3n − 3.
Now to take an n-connected region U and produce a slit region, we
need to choose a point p ∈ U to send to infinity and we need to choose a
‘horizontal’ direction at p. That gives 3 more real parameters. But of course
these choices are only relevant up to the action of the automorphism group
of U , which has dimension 3, 1 and 0 for n = 1, 2 and n ≥ 3. Altogether we
find:

dn = dim{moduli space of n-connected regions}

= 3n − 6 + dim Aut(U )

and hence d1 = 0, d2 = 1 (the modulus of an annulus), d3 = 3, d4 = 6, etc.

Rigidity. We remark that although an n-connected region U can have
automorphisms, it cannot have a positive–dimensional set thereof, when n >
2. To see this, we can assume U is a plane region bounded by real-analytic
Jordan curves. Then dim Aut(U ) > 0 would imply there is a holomorphic
vector field v = v(z)(d/dz) on U tangent to ∂U . If v 6= 0, then the Euler
characteristic of U can be expressed as sum of positive contributions, one
for each zero of v on ∂U or in U . But χ(U ) < 0, so v = 0.
A more geometric proof can be given once one knows that U ∼ = ∆/Γ
(a proof will be sketched below). Then any automorphism must be an
isometry, and hence send geodesics to geodesics. By looking at intersections
of geodesics one can easily conclude that f is the identity.

4.2 Conformal mappings of polygons

To get explicit forms for Riemann mappings, it is useful to initially try
to determine, not f itself, but the deviation of f from a simpler class of
mappings. This deviation will be a form in general, so we first make some
remarks on forms.

68
 Let f : X → C b be a meromorphic function on a Riemann surface. Then
′
ω = df = f (z) dz is naturally a 1-form. That is, if we change coordinates
by setting z = φ(w), then in these new coordinates we have

φ∗ (df ) = d(f (φ(w)) = f ′ (φ(w))φ′ (w) dw.

We can similarly define quadratic differentials q(z) dz 2 which transform by

φ∗ q = q(φ(z))φ′ (z)2 dz 2 .
In more intrinsic terms, these forms are sections of the complex cotangent
bundle T ∗ X and its tensor product with itself.
For a meromorphic 1-form, the residue ResRp (ω) is also coordinate inde-
pendent, since it can be expressed as (2πi)−1 γ ω for a small loop around
p. Stokes’ theorem implies:

Theorem 4.25 (The Residue Theorem) If ω Pis a nonzero meromorphic

1-form on a compact Riemann surface X, then p∈X Resp (ω) = 0.

Let’s check this in an example: if we take ω(z) = dz/P (z) where P (z)
is a polynomial of degree d ≥ 2, then it says:
X 1
= 0.
P ′ (z)
P (z)=0
P
ThisP can be verified using partial fractions: we have P (z) = ai /(z − bi )
and ai = 0 because |P (z)| = O(|z|−2 ) for large z. It can also be proved
by integrating dz/P (z) around a large circle and taking the limit.
If ω = ω(z) dz is a meromorphic 1-form on the sphere, we can set π(z) =
1/z and form π ∗ ω to find:

Res∞ (ω) = Res0 (−ω(1/z) dz/z 2 ).

In particular Res∞ (dz/z) = −1.

Theorem 4.26 Every meromorphic 1-form on the sphere has 2 more poles
than zeros.

Proof. This is true for ω = dz/z, which has a simple poles at 0 and ∞
and no zeros. It then follows for any other 1-form η, since f = η/ω is a
meromorphic function, which must have the same number of poles as zeros.

69
Corollary 4.27 A holomorphic 1-form on C b must be zero. In particular,
if ω1 and ω2 are 1-forms with the same principal parts, then ω1 = ω2 .

Remark: one forms and vector fields in higher genus. The first
assertion can also be seen by integrating ω to obtain a global holomorphic
b On a Riemann surface of genus g, a meromorphic 1-form has
function on C.
2g − 2 more zeros than poles, and dim Ω(X) = g. The space of holomorphic
b = 3, dim Θ(E) = 1 on a
vector fields, on the other hand, satisfies dim Θ(C)
torus, and dim Θ(X) = 0 in higher genus. This is because a meromorphic
vector field has 2g − 2 more poles than zeros.
One can use the vanishing of Θ(X) is see an n-connected plane region U
has a zero-dimensional automorphism group for n 6= 3; if not, there would
be a boundary-parallel holomorphic vector field on U , but then the double
X of U would satisfy dim Θ(X) > 0; while its genus is given by g = n − 1.
Measuring distortion: the 3 great cocycles. Now let C(f ) be a dif-
ferential operator that sends meromorphic functions to meromorphic forms.
We say C is a cocycle if it satisfies:

C(f ◦ g) = C(g) + g ∗ C(f ).

This formula implies the maps satisfying C(f ) = 0 form a group.

There are 3 important cocycles in complex analysis:
1. The derivative Df (z) = log f ′ (z).
2. The nonlinearity N f (z) dz = d Df = (f ′′ /f ′ )(z) dz.
3. The Schwarzian derivative
" ′  2 #
1 f ′′ 1 f ′′
Sf (z) dz = (N f ) − (N f )2 =
2 ′
− dz 2 .
2 f′ 2 f′

Their values are functions, 1-forms and 2-forms respectively. The groups
they annihilate are transformations of the form f (z) = z + a, f (z) = az + b
and f (z) = (az + b)/(cz + d).
Let us check this for Sf : if f (z) = (az + b)/(cz + d) and ad − bc = 1,
then f ′ (z) = 1/(cz + d)2 , hence N f (z) = −2c/(cz + d), which satisfies
2c2 1
(N f )′ = 2
= (N f )2 .
(cz + d) 2

Higher cocycles? There are in fact operators Td f generalizing Sf with

the property that Td f = 0 iff f is a rational map of degree d > 1. But these

70
cannot be cocycles, because the rational maps of degree d > 1 do not form
a group. P
To get some insight into Sf (and
P T d ), note that if f (z) = ai z i =
(az + b)/(cz + d), then (cz + d) ai zi = az + b and hence there must
be a linear relation between most of the coefficients of zf (z) and f (z); in
particular, we must have
!
a1 a2
det = a22 − a1 a3 = 0,
a2 a3

which just says that f ′ f ′′′ − (3/2)(f ′′ )2 vanishes. This expression is also the
denominator of Sf . Simiarly a rational map of degree two is characterized
by the property that  
a1 a2 a3
 
det 
a2 a 3 a4
 = 0.

a3 a4 a5

It is natural to divide this expression by a31 just as we have divided Sf by

a21 = (f ′ )2 , so we have Td (af + b) = Td (f ).
The Schwarz-Christoffel formula. We can now use the nonlinearity to
derive a formula for the Riemann mapping to a polygonal region. It is based
on the fact that
dz
N (z α ) = (α − 1) .
z
Theorem 4.28 Let f : H → U be the Riemann mapping to a polygon with
vertices pi , i = 1, . . . , n and exterior angles πµi . Then
Z
dζ
f (z) = α Qn dζ + β,
1 (ζ − qi )µi

where f (qi ) = pi .

Proof. We will show that:

X −µi dz
Nf = ·
(z − qi )

We first observe that N f extends to a holomorphic 1-form on C−{p1 , . . . , pn }.

This follows by Schwarz reflection across each complementary interval on
the real axis. These reflected maps gi do not agree, but they differ by linear

71
maps: gj = Aij ◦ gi , and thus N (gi ) = N (gj ). It remains to show N f has
simple poles with the given residues at the pi , and at infinity.
Let αi = 1 − µi . The idea of the proof is that f (z) behaves like (z − pi )αi
near pi , and thus N f has the same residue, which is αi − 1 = −µi . To check
this, we note that g(z) = (f (z)−pi )1/αi extends by Schwarz reflection across
qi , to give a conformal map near qi . Thus we can locally write

g(z)αi = f (z) − pi .

Taking the nonlinearity of both sides, and using the fact that the residue is
preserved under pullback, we find Resqi (N f ) = 1 − αi = −µi .
Near infinity, f (z) behaves like 1/z which has nonlinearity −2dz/z and
hence residue
P 2 at infinity. This is consistent with the residue theorem,
because πµi = 2π. Since a 1-form is determined by its singularities, we
have justified the formula for N f .
Integrating, we find
X
log f ′ = −µi log(z − qi ) + C;

by exponentiating and integrating again, we get the desired formula for f .

Two Examples. This formula explains the geometric connection between

Riemann maps to polygons and transcendental functions we have seen in
two examples, namely: Z
dz
log(z) =
z
maps H to a bigon with external angles of π, namely the strips 0 < Im z < π;
and Z
dz
sin−1 (z) = √
1 − z2
maps H to a triangle with external angles π/2, π/2 and π, namely the half-
strip defined by Im z > 0, | Re(z)| < π/2.
Unit disk version. The key point in the proof above was that the various
results of extending f by Schwarz reflection through ∂H agree up to compo-
sition with linear maps. The same is true if H is replaced by the unit disk.
Thus if we choose the domain of f to be ∆, with qi ∈ S 1 , we find exactly
the same formula for f .
Regular polygons. As a particular example, we find:

72
 R
Theorem 4.29 The map f (z) = (1 − z n )−2/n dz maps the unit disk to
a regular n-gon, sending 0 to its center and the nth roots of unity to its
vertices.

Note that f ′ (0) = 1, and the distance from the center of the polygon to
one of its vertices is given by
Z 1
Rn = (1 − z n )−2/n dz.
0

The Beta function and conformal radius. The quantity Rn can be

computed explicitly. To this end, we recall the Euler Beta-function
Z 1
du Γ(α)Γ(β)
B(α, β) = uα (1 − u)β = ·
0 u(1 − u) Γ(α + β)
Note that it is elementary to prove the integral above satisfies B(α, 1) = 1/α,
and integration by parts gives B(α, β) = ((β + 1)/α)B(α + 1, β − 1); this
easily implies
α!β!
B(α + 1, β + 1) =
(α + β + 1)!
for integral values of α, β. We also note that |dx|/x, which appears in
the integral definition of the Γ function, is the hyperbolic metric on R+ ,
and |dx|/(x(1 − x)) is the hyperbolic metric on (0, 1). Note that formally
these one forms have residue 1 at the endpoints of the interval, just as
2|dx|/(1 − |x|2 ) does on (−1, 1).
Making the substitution u = 1−z n , du = −nz n−1 dz, dz = −(1/n)u1/n−1 du,
we find
Z
1 1 −2/n Γ(1 − 2/n)Γ(1/n)
Rn = u (1 − u)1/n−1 du = ·
n 0 nΓ(1 − 1/n)
Note that Rn → 1 as n → ∞, since Γ(1/n) ∼ n, and Rn < 1 by the Schwarz
lemma.
Now if we let Sn be the side length of the polygon, then Sn = 2 sin(π/n)Rn .
But Γ(1/n)Γ(1 − 1/n) = π/ sin(π/n), so we get
2πΓ(1 − 2/n)
Sn = ·
nΓ(1 − 1/n)2
Note that nSn → 2π as n → ∞, since Γ(1) = 1. The factor Pn by
with the perimeter exceeds the perimeter of the unit circle is given by
(P3 , P4 , P5 , P6 ) ≈ (1.461, 1.180, 1.098, 1.043).

73
 In particular, the conformal radius of the unit square (sides of length
2) is 2/S4 = 4Γ[3/4]2 /π 3/2 = 1.078705 . . .. (It
√ is clear that the conformal
radius of the unit square lies between 1 and 2.)
Quadrilaterals. A quadrilateral Q is a Jordan domain with 4 distinguished
points on its boundary. A conformal map between quadrilaterals is required
to preserve these points as a set.

Theorem 4.30 Any quadrilateral is conformally equivalent to a rectangle

of the form R(a) = [0, a] × [0, 1], a > 0. This rectangle is unique up to
a 7→ 1/a.

Proof. By the Riemann mapping theorem, any quadrilateral is isomorphic

to H with 4 distinguished
RQ points, q1 , . . . , q4 ∈ R. By the Schwarz-Christoffel
formula, f (z) = (z −qi )−1/2 dz maps H to a Euclidean rectangle, sending
qi to its vertices.

Besikovitch lemma. As an exercise, one can now use the method of

extremal length to show, for any quadrilateral Q ⊂ C, we have area(Q) ≥
AB where A and B are the minimum distances between opposite sides of
Q.
The Schwarzian derivative and hypergeometric functions. Just as
the nonlinearity can be used to find a formula for the Riemann mapping to
a polygon with linear sides, the Schwarzian derivative can be used to find
the Riemann mapping to a polygon with circular sides.
To describe a result in this direction, we first recall that for f (z) = z α
we have N f (z) = (α − 1)/z, and thus

(N f )2 1 − α2 dz 2
Sf (z) = (N f )′ − = ·
2 2 z2
The quantity 1 − α2 is a version of the residue for the quadratic differen-
tial Sf (z) dz 2 . Now given α, β, γ there is a unique quadratic differential
with double poles at 0, 1, ∞ with the corresponding residues, and no other
singularities. It is given by
 
1 1 − α2 1 − β2 1 − γ 2 − (2 − α2 − β 2 )
Q(α, β, γ) = + + dz 2 .
2 z2 (z − 1)2 z(z − 1)

The third term is chosen to that Q ∼ (1 − γ 2 )/z 2 as z → ∞.

74
Theorem 4.31 Let P be a circular triangle with interior angles πα, πβ and
πγ. Then the Riemann mapping f : H → P sending 0, 1 and ∞ to these
vertices satisfies Sf = Q(α, β, γ).

Proof. By Schwarz reflection, Sf extends to a meromorphic quadratic

b whose residues at 0, 1 and ∞ are determined by the angles
differential on C
of P .

Ideal triangles. As an example, if P is an ideal triangle, then we have

z2 − z + 1
Q = Q(0, 0, 0) = ·
2z 2 (z − 1)2
This differential is symmetric under the action of S3 on H, and it has zeros
at the primitive sixth roots of unity.
Second order differential equations. To find f itself, we remark that
any function f satisfying Sf = Q can be expressed as the ratio f = u1 /u2
of two independent solutions to the differential equation

u′′ + (Q/2)u = 0.

In the case at hand, these solutions in turn can be expressed in terms of

solutions to the hypergeometric equation

z(1 − z)u′′ + Au′ + Bu = 0

for suitable constants A and B.

4.3 The Picard theorems

We conclude this section with two further results concerning entire functions
that can be related to geometric function theory.

Theorem 4.32 (Little Picard Theorem) An entire function f : C → C

which omits two values must be constant.

Corollary 4.33 A meromorphic function on C can omit at most two values

b
on C.

The Little Picard Theorem is equivalent to the assertion that there is no

solution to the equation ef + eg = 1 where f and g are nonconstant entire
functions. Similarly, it implies there is no solution to Fermat’s equation
f n + gn = 1, n ≥ 3, unless the entire functions f and g are constant.

75
Theorem 4.34 (Great Picard Theorem) An analytic function f : U →
C takes on every value in C, with at most one exception, in every neighbor-
hood of an essential singularity p.

The first theorem follows from the second by considering the essential
singularity at z = 0 of f (1/z). These results generalize Liouville’s theorem
and the Weierstrass-Casorati theorem respectively.
Rescaling arguments. Here is a third result that initially seems unrelated
to the first two.

Theorem 4.35 (Bloch’s Theorem) There exists a universal R > 0 such

that for any f : ∆ → C with |f ′ (0)| = 1, not necessarily univalent, there
is an open set U ⊂ ∆ (perhaps a tiny set near S 1 ) such that f maps U
univalently to a ball of radius R.

Corollary 4.36 For any analytic map on ∆, the image f (∆) contains a
ball of radius R|f ′ (0)|.

Note that the ball usually cannot be centered at f (0); for example,
f (z) = exp(nz)/n satisfies f ′ (0) = 1 but the largest ball about f (0) = 1/n
in f (∆) ⊂ C∗ has radius 1/n.
√ The optimal value of R is known as Bloch’s constant. It satisfies 0.433 <
3/4 ≤ R < 0.473. The best-known upper bound comes from the Riemann
surface branched with order 2 over the vertices of the hexagonal lattice.
These apparently unrelated theorems can both be proved using the same
idea. (Cf. [BD] and references therein; for another way to relate these
theorems, see [Re, Ch. 10].)
Proof of Bloch’s theorem. Given f : ∆ → C, let
kf ′ (z)k = kf ′ (z)k∆,C = (1/2)|f ′ (z)|(1 − |z|2 )
denote the norm of the derivative from the hyperbolic metric to the Eu-
clidean metric. By assumption, kf ′ (0)k = 1/2. We can assume (using
f (rz)) that f is smooth on S 1 ; then kf ′ (z)k → 0 as |z| → 1, and thus
sup |kf ′ (z)k is achieved at some p ∈ ∆.
Now replace f with f ◦ r where r ∈ Aut(∆) moves p to zero. Replacing
f with af + b with |a| < 1, we can also arrange that f (0) = 0 and kf ′ (0)k =
1; this will only decrease the size of its unramified disk. Then kf ′ (z)k ≤
kf ′ (0)k = 1, and thus f |∆(1/2) ranges in a compact family of nonconstant
analytic functions. Thus the new f has an unramified disk of definite radius;
but then the old f does as well.

76
Rescaling proof of Picard’s theorem. The proof will use the following
remarkably general rescaling theorem. This argument is related to Bloch’s
proof, to Brody’s reparameterization theorem and to other results in com-
plex analysis.

Theorem 4.37 Let fn : C → C be a sequence of nonconstant entire func-

tions. There after passing to a subsequence, there is a sequence of Möbius
transformations An and a nonconstant entire function g : C → C such that
g = lim fn ◦ An uniformly on compact subsets of C.

We note that the An need not fix infinity, so fn ◦ An is undefined at some

b but we will have pn → ∞.
point pn ∈ C,
Example. For fn (z) = z n we can take fn ◦ An (z) = (1 + z/n)n → ez .
Metrics. As a preliminary to the proof, for g : C → C b we define

|g′ (z)|
kg′ (z)k∞ = ,
(1 + |g(z)|2 )
and kg′ k∞ = sup kg ′ (z)k over all z ∈ C. This is the norm of the derivative
from the scaled Euclidean metric ρ∞ = 2|dz| to the spherical metric. Note
that g(z) = exp(z) has kg′ k∞ = 1/2; a function with bounded derivative
can be rather wild.
Similarly, for g : ∆(R) → C,b we define

1 − |z/R|2
kg′ (z)kR = |g′ (z)| ·
1 + |g(z)|2
This is the derivative from a suitably rescaled hyperbolic metric ρR on ∆(R)
b Clearly ρR → ρ∞ uniformly on compact sets. Its key property is that
to C.
b stabilizing ∆(R).
k(g ◦ A)′ kR = kg ′ kR for all A ∈ Aut(C)
We also note that the set of maps with uniformly bounded derivatives
in one of these norms is compact.
Proof of Theorem 4.37. Let us first consider an arbitrary nonconstant
analytic function f (z) and a radius R > 0. We claim there exists an S ≥ R
and an A ∈ Aut(C) b such that g = f ◦ A is analytic on ∆(S), and

kg′ (0)kS = kg′ kS = 1.

Indeed, by replacing f with f (az + b), we can assume kf ′ (0)kR = 1. Then

kf ′ kR ≥ 1. On the other hand, the R-norm of the derivative of f tends to
zero at the boundary of ∆(R). Thus we can choose B ∈ Aut(∆(R)) such

M = k(f ◦ B)′ (0)kR = k(f ◦ B)′ kR ≥ 1.

77
Now just let g(z) = (f ◦ B)(z/M ), and S = RM .
Applying this claim to fn and a sequence Rn → ∞, we obtain Sn → ∞
and maps gn = fn ◦ An with kgn′ (0)k∞ = 1 and kgn′ kSn ≤ 1. Now pass to a
convergent subsequence.

To complete the proof of Picard’s theorem, we observe:

Lemma 4.38 If fn → f and f is nonconstant, then any value omitted by

all fn is also omitted by f .

Proof of the Little Picard Theorem. Suppose if f : C → C is non-

1/n
constant and omits 0 and 1. Then fn (z) = fn (z) omits more and more
points on the unit circle. We can of rescale in the domain so the spherical
derivative satisfies kfn′ (0)k∞ → ∞. Passing to a subsequence and reparam-
eterizing, we obtain in the limit a nonconstant entire function that omits
the unit circle. This contradicts Liouville’s theorem.

Classical proof. The classical proof of the Little Picard Theorem is based
on the fact that the universal cover of C − {0, 1} can be identified with the
upper halfplane.
To see this, it is useful to start by considering the subgroup Γ0 ⊂ Isom(∆)
generated by reflections in the sides of the ideal triangle T with vertices
{1, i, −1}. For example, z 7→ z is one such reflection, sending T to −T . By
considering billiards in T , one can see that its translates tile the disk and
thus T is a fundamental domain for Γ0 . Thus the quadrilateral F = T ∪(−T )
is a fundamental domain for the orientation-preserving subgroup Γ ⊂ Γ0 ,
and the edges of −T are glued to the edges of T to give a topological triply-
punctured sphere as quotient.
Now let π : T → H be the Riemann mapping sending T to H and its
vertices to {0, 1, ∞}. Developing in both the domain and range by Schwarz
reflection, we obtain a covering map π : ∆ → C b − {0, 1, ∞}.
Given this fact, we lift an entire function f : C → C − {0, 1} to a map
fe : C → H, which is constant by Liouville’s theorem.
Uniformization of planar regions. Once we know that C b − {0, 1, ∞} is
uniformized by the disk, it is straightforward to prove:

b with |C
Theorem 4.39 The universal cover of any region U ⊂ C b − U| ≥ 3
is isomorphic to the unit disk.

78
Sketch of the proof. Consider a basepoint p in the abstract universal
e → U , and let F be the family of all holomorphic maps
cover π : U
e , p) → (∆, 0)
f : (U

that are covering maps to their image. Using the uniformization of the
triply-punctured sphere, we have that F is nonempty. It is also a closed,
normal family of functions in O(U e ); and by the classical square-root trick,
it contains a surjective function (which maximizes |f ′ (p)|). By the theory
of covering spaces, this extremal map must be bijective.

Proof of the Great Picard Theorem. Let f : ∆∗ → C b − {0, 1, ∞} be

an analytic function. We will show f does not have an essential singularity
at z = 0.
Consider a loop γ around the puncture of the disk. If f sends γ to a
contractible loop on the triply-punctured sphere, then f lifts to a map into
the universal cover H, which implies by Riemann’s removability theorem
that f extends holomorphically over the origin.
Otherwise, by the Schwarz lemma, f (γ) is a homotopy class that can be
represented by an arbitrarily short loop. Thus it corresponds to a puncture,
which we can normalize to be z = 0 (rather than 1 or ∞). It follows that f
is bounded near z = 0 so again the singularity is not essential.

5 Elliptic functions and elliptic curves

In this section we discuss the field of meromorphic functions on the first
compact Riemann surfaces after C, b namely complex tori of the form E =
C/Λ.
These functions arise naturally when trying to solve one of the simplest
calculus integrals: Z
dx
F (x) = √ ·
3
x + ax + b
Just as the inverses of the
√ exponential and trigonometric functions arise from
integrating 1/x and 1/ x2 + ax + b, the integrals of cubic polynomials in-
volve inverses of elliptic functions. We will ultimately see this geometrically,
using the Schwarz-Christoffel formula.
From the perspective of algebraic geometry, elliptic functions arise if one
tries to sweep out or parameterize all solutions of a cubic equation such as

y 2 = x3 + ax + b.

79
A special feature of the case of cubics is that the solutions (in projective
space) form a group.

5.1 Doubly–periodic functions

Let Λ ⊂ C be a lattice – meaning a discrete subgroup such that E = C/Λ
is compact. Then Λ ∼
= Z2 as an abstract group. We can choose a basis such
that
Λ = Zα ⊕ Zβ.
Every point of E is represented by an essentially unique point in the period
parallelogram with vertices 0, α, β and α + β.
Meromorphic functions F : E → C b are the same thing as doubly periodic
b
functions f : C → C, i.e. meromorphic functions satisfying

f (z + α) = f (z + β) = f (z)

(and hence f (z + λ) = f (z) for all λ ∈ Λ. For a fixed Λ, these functions

form a field, denoted K(E).
Note that E forms a group under addition, and hence E ⊂ Aut(E). We
also have an involution η : E → E given by η(z) = −z. This is the full
automorphism group of E provided Λ is not a square or hexagonal lattice.
In the latter cases, Aut(E) also contains a cyclic group Z/4 or Z/6.
Because of these automorphisms, if F ∈ K(E) then so is F (−z) and
F (z + p), p ∈ E.
We note that a holomorphic doubly-periodic function must be constant,
by Liouville’s theorem, because the period parallelogram is compact. So to
find other functions we must allow poles. The simplest are constructed for
each k ≥ 3 by setting
X 1
ζk (z) = ·
(z − λ)k
Λ

This sum does not quite converge for k = 2; it can be made to converge by
writing
1 X′ 1 1
℘(z) = 2 + 2
− 2·
z (z − λ) λ
Λ

The critical properties of ℘(z) are that is it even, and it has a unique pole
of order 2 on E.

Theorem 5.1 The Weierstrass ℘-function is doubly-periodic.

80
 Figure 6. The Weierstrass ℘-function for the hexagonal lattice.

Proof. Note that ℘′ (z) = −2ζ3 (z) is doubly-periodic. Thus ℘(z + 1) =

℘(z) + A and ℘(z + λ) = ℘(z) + B for some constants A and B. But we also
have ℘(−z) = ℘(z); setting z = −1/2 and −λ/2, we find A = B = 0.

Uniformization of cubic curves. We can now state the main theorems

regarding elliptic functions.
Theorem 5.2 The map π : C → P2 given by
π(z) = (℘(z), ℘′ (z))
gives an isomorphism between the Riemann surface E = C/Λ and a smooth
cubic curve of the form y 2 = (4x3 + ax + b).
Theorem 5.3 The field of all doubly-periodic functions for a given lattice
Λ satisfies
K(C/Λ) ∼ = C[x, y]/(y 2 − (4x3 + ax + b)),
where (x, y) = (℘, ℘′ ). In particular, every doubly-periodic function is a
rational function of ℘ and ℘′ .
We remark that any smooth cubic curve C ⊂ P2 can be put into the
form above by applying a change of coordinates (an automorphism of P2 ).
Later, will see that every cubic curve occurs for suitable choice of Λ.
Basic properties of elliptic functions. Let f (z) be a nonconstant
doubly-period function for a lattice Λ ⊂ C with period parallelogram P .
Then f defines a meromorphic function on the compact Riemann surface
E = C/Λ. We also note that the form dz is invariant under Λ, so we also
get a meromorphic 1-form ω = f (z) dz on E. Here are some basic facts.

81
 1. The sum of the residues of f (z) dz over E, or over points in P , is zero.

2. The function f has the same number of zeros as poles. The number
of each is called the degree d = deg(f ).
P P
3. If f has poles p1 , . . . , pd and zeros a1 , . . . , ad in E, then pi = ai
in the group law on E.

Proofs. (1) This follows applying Stokes’ theorem to the closed form f (z) dz
on E − {p1 , . . . , pd }, or by integrating f (z) dz around the boundary of P (we
may assume f has no poles on ∂P .)
(2) This is a general property of proper maps between Riemann surfaces.
For a direct proof, one can also apply the residue theorem to df /f .
(3) This property is special to elliptic curves. Let P = [0, α] × [0, β], and
assume f has no zeros or poles in P . Then we have
X X Z
−1 zf ′ (z) dz
ai − pi = (2πi) ·
∂P f (z)
We wish to show that this quantity lies in Λ. The R integrals over opposite
edges cancel, up to a terms of the form λ(2πi)−1 e f ′ (z)/f (z) dz with λ ∈ Λ.
Since the f is periodic, it has an integral winding number N (e) on each edge,
and these terms have the form N (e)λ ∈ Λ.

We will later see that we may construct an elliptic function with given
zeros and poles subject only to constraint (3).
Pushforward. Here is another way to see property (3). Let f : E → C b
b
be a nonconstant meromorphic function. Then f∗ (dz) = 0, since Ω(C) = 0.
Now choose a path C ⊂ C b running from 0 to ∞ and avoiding the critical
values of f . Then C e = f −1 (C) ⊂ E gives a collection of arcs connecting
(ai ) and (pi ) in pairs, which we can assume have the same indices. We then
have Z Z X
0= f∗ ω = dz = pi − ai mod Λ.
C e
C

The differential equation. We can now give a cubic equation relating

x = ℘(z) and y = ℘′ (z).
Since ℘′ (z) is an odd function of degree three, its zeros coincide with the
points E[2] of order two on E. Using our chosen basis Λ = Zα ⊕ Zβ, we can
explicitly label representatives of these points:
 
α β α+β
E[2] = {0, c1 , c2 , c3 } = 0, , , .
2 2 2

82
Let ei = ℘(ci ).
We note that the critical values ei are distinct; indeed, since ℘′ (zi ) = 0,
the function ℘(z) − ei has a double zero at zi , so it cannot vanish anywhere
else. This shows:

The zeros of ℘′ (z) coincide with the points of order two c1 , c2 , c3 .

(Morally there is also a critical point at z = 0.) Consequently:

The function ℘(z)− ei has a double zero at ci and no other zeros.

(Note: it is not at all easy to say where the two zeros of ℘(z) lie, except in
the case of symmetric lattices.)

Theorem 5.4 For all z ∈ C, we have

℘′ (z)2 = 4(℘(z) − e1 )(℘(z) − e2 )(℘(z) − e3 ).

Proof. From what we have seen above, the two sides are doubly-periodic
functions with the same zeros and poles. Thus they are multiples of one
another. They are equal since near zero they are both asymptotic to 4/z 6 .

Corollary 5.5 The map ℘ : E → C b presents E as a 2-sheeted covering

b branched over e1 , e2 , e3 and ∞. It gives the quotient of E by the
space of C,
involution z 7→ −z.

Proof. By construction ℘ is even and 2-to-1, and we have just identified

its critical points and values.

Corollary 5.6 The map π(z) = (℘(z), ℘′ (z)) maps E = C/Λ bijectively to
the smooth projective cubic curve E defined by

y 2 = 4(x − e1 )(x − e2 )(x − e3 ).

Proof. Since E is compact, the function x = ℘(z) maps E onto C; b and for
a given x, the two possible values of y solving the equation above are given
by y = ℘′ (±z).

83
 A B
A B

Figure 7. Visualizing the degree two map from E to E/η ∼ b where

= C,
η(z) = −z.

Power series expansion. Let O(d) ⊂ K(E) be the vector space of doubly–
periodic functions with a pole of order d at z = 0 and no other poles.

Theorem 5.7 We have dim O(d) = d for d ≥ 2, and dim O(0) = 1.

Proof. Clearly a function in O(d) is uniquely determined by its polar part

up to the constant term, ad /z d + · · · + a0 . Moreover a1 = 0 by the residue
theorem. Thus dim O(d) ≤ d. Using ℘(z) and ζk (z) we obtain the reverse
inequality.

Which element of O(2) is ℘? In fact, it is the unique element with a2 = 1

and a0 = 0. This is evident from its definition, since each term in the series
for ℘(z), apart from the initial term 1/z 2 , is normalized to vanish at the
origin.
More precisely, using the expansion
 
1 1 1 2z 3z 2 2z 3z 2 4z 3
− = + + · · · = + + + ··· ,
(z − λ)2 λ2 λ2 λ λ2 λ3 λ4 λ5
we obtain:

Theorem 5.8 The Weierstrass ℘-function is given near z = 0 by:

∞X
1 2 4 1
℘(z) = + 3G2 z + 5G3 z + · · · = + (2n + 1)z 2n Gn+1
z2 z2
1

where X′ 1
Gn = Gn (Λ) = .
λ2n
Λ

Corollary 5.9 We have ℘′ (z)2 = 4℘(z)3 − g2 ℘(z) − g3 , where g2 = 60G2

and g3 = 140G3 .

84
Proof. Neglecting terms of order O(z 2 ), we have:
1 9G2
℘(z)3 = 6
+ 2 + 15G3 and
z z
 2
−2 4 24G2
℘′ (z)2 = 3
3
+ 6G2 z + 20G3 z + · · · = 6 − 2 − 80G3 .
z z z

Thus 4℘(z)3 − ℘′ (z)2 = 60G2 /z 2 + 140G3 + O(z 2 ) = g2 ℘(z) + g3 .

P
Corollary 5.10 We have ei = 0.

Remark: other constructions of elliptic functions. To construct el-

liptic functions of degree two: if Λ = Z ⊕ Zτ , one can first sum over Z to
get:
∞
X 1 π2
f1 (z) = = ;
−∞
(z − n)2 sin(πz)2
then
∞
X
f (z) = f1 (z − nτ )
−∞

converges rapidly, and defines an elliptic function of degree two. Similarly,

if E = C∗ /hz 7→ αzi, with |α| =
6 0, 1, then we
∞
X αn w
F (w) =
−∞
(αn w − 1)2

defines a function on C∗ with a double pole at w = 1 satisfying F (αw) =

F (w); thus f (z) = F (ez ) is a degree two doubly-periodic function for the
lattice Λ = Z2πi ⊕ Z log α. (Note that z/(z − 1)2 has simple zeros at 0, ∞
and a double pole at z = 1.)
These functions are not quite canonical; there is a choice of direction in
the lattice to sum over first. As a consequence they agree with ℘(z) only
up to a constant. This constant is a multiple of the important quasimodular
form !
X X 1
G1 (τ ) = .
m n
(m + nτ )2

The real case. Now suppose Λ = Zα ⊕ Zβ with α ∈ R+ and β ∈ iR+ .

Then the critical points (0, c1 , c2 , c3 ) of ℘(z) bound a rectangle S ⊂ C (see
Figure 8).

85
Theorem 5.11 The value ℘(z) is real if and only if z lies on one of the
vertical or horizontal lines through (1/2)Λ.

Proof. Since Λ is invariant under both negation and complex conjugation,

we have
℘(z) = ℘(−z) = ℘(z).
Thus ℘(z) is real on the locus R ⊂ E which is fixed under z 7→ z or z 7→ −z.
This locus is covered in C by the vertical and horizontal lines through the
points of (1/2)Λ. (E.g. the locus fixed by z 7→ z consists of two horizontal
loops on E, because x + β/2 7→ x − β/2 ∼ x + β/2.)
Now as z traverses ∂S, it moves monotonically along R b passing through
b
∞ just once. Thus ℘ maps S bijectively to R, and the same for −S. Since
b
℘ has degree two, there are no other preimages of R.

Corollary 5.12 The map ℘|S gives the unique conformal map from S to
−H such that ℘(ci ) = ei for i = 1, 2, 3 and ℘(0) = ∞.

Proof. The region S is a component of the complement of ℘−1 (R), b con-

taining no critical points of ℘, so it maps homeomorphically to H or −H. In
fact the image is −H because Im ℘(z) ∼ Im 1/z 2 < 0 for small z in S.

β α+β

c3
c2

0 c1 α

Figure 8. The shaded rectangles map under ℘ to H; the unshaded ones, to

−H.

Theorem 5.13 Any branch f : (±H) → C of ℘−1 satisfies

Z
dζ
f (z) = p ·
3
4ζ − g2 ζ − g3

86
Proof. If ζ = ℘(z) then f (ζ) = z, and hence f ′ (℘(z))℘′ (z) = 1. Conse-
quently
1 1 1
f ′ (ζ) = =p =p ·
℘′ (z) 4℘(z)3 − g2 ℘(z) − g3 4ζ 3 − g2 ζ − g3

Thus the theory of elliptic functions emerges as the special case of the
Schwarz-Christoffel formula giving the Riemann map for a rectangle.
4

-2 -1 1 2

-2

-4

Figure 9. The elliptic curve y 2 = 4x3 − 8x + 2.

Location of roots. Note that the argument shows the roots of 4x3 − g2 x −
g3 = 0 are real, and satisfy e2 < e3 < e1 . We also note that ℘′ (z) ∈ R
along the horizontal lines forming ℘−1 (R), and ℘′ (z) ∈ iR along the vertical
lines, since ℘ most rotate the latter by 90◦ to make them real. Thus only
the horizontal lines map to the real points of the cubic y 2 = 4x3 − g2 x − g3 .
Periods. This picture makes clear the close relationship between the gen-
erators α, β of Λ and the cubic polynomial

4x3 − g2 x − g3 = 4(x − e1 )(x − e2 )(x − e3 );

namely, we have
Z ∞ Z e3
α 3 −1/2
= (4x − g2 x − g3 ) dx = (4x3 − g2 x − g3 )−1/2 dx,
2 e1 e2

and
Z e2 Z e1
β 3 −1/2
= (g3 + g2 x − 4x ) dx = (g3 + g2 x − 4x3 )−1/2 dx.
2i −∞ e3

87
In all 4 integrals we take the positive square-root; thus, these are simply
integrals of |(℘′ )−1 (x)|.
In each case, the fact that the two integrals agree follows from Cauchy’s
theorem (applied to two homotopic loops in C − {e1 , e2 , e3 }), or via a change
of variables coming from a Möbius transformation that swaps one interval
for the other.
More generally, we have:

Theorem 5.14 The lattice Λ with invariants g2 and g3 is generated by the

values of Z
dz
± p ,
γ 4z 3 − g2 z − g3
where γ ranges over all oriented loops in C which enclose exactly two roots
of the cubic in the denominator.

The condition on the roots insures that the integrand can be defined
continuously on a neighborhood of γ.
Evaluation of g2 and g3 . The preceding formula sometimes permits the
evaluation of g2 and/or g3 . For example, we have

X′ Z ∞ 6
−6 dx 256π 3 Γ[7/6]6
140 λ =4 √ = ·
1 x3 − 1 Γ[2/3]6
Z[ρ]

This follows from the fact that for Λ = Z[ρ], we have g2 = 0 and g3 > 0
satisfies Z ∞
1 dx
= p ,
2 e1 4x3 − g3
where e1 = (g3 /4)1/3 . Similarly, we have
X′ 64π 2 Γ[5/4]4
60 λ−4 = ·
Γ[3/4]4
λ∈Z[i]

Function fields. We now return to the case of general elliptic curves.

Theorem 5.15 The function field of E = C/Λ is generated by x = ℘ and

y = ℘′ ; more precisely, we have

K(E) = C(x, y)/(y 2 − 4x3 + g2 x + g3 ).

88
Proof. To see that ℘ and ℘′ generate K(E) is easy. Any even function
f :E→C b factors through ℘: f (z) = F (℘(z)), and so lies in C(℘). Any odd
function becomes even when multiplied by ℘′ ; and any function is a sum of
one even and one odd.
To see that the field is exactly that given is also easy. It amounts to
showing that K(E) is of degree exactly two over C(℘), and ℘ is transcen-
dental over C. The first assertion is obvious (else ℘ would be constant), and
if the second fails we would have K(E) = C(℘), which is impossible because
℘ is even and ℘′ is odd.

The addition law on an elliptic curve. Consider the curve E ⊂ P2

defined by y 2 = 4x3 − g2 x − g3 and parameterized by the Weierstrass ℘-
function via (x, y) = (℘(z), ℘′ (z)).

Theorem 5.16 For any line L, the intersection L ∩ E = {a, b, c} where

a + b + c = 0 on E = C/Λ.

Proof. The intersection L ∩ E is simply the zero set of A℘′ + B℘ + C for

some (A, B, C). This function has all its poles at z = 0. Since the sum of
the zeros and poles is zero, its zeros (a, b, c) also sum to zero.

Corollary 5.17 For any z, w ∈ C we have

1 1 1
℘(−z) ℘(−w) ℘(z + w) = 0.
℘′ (−z) ℘′ (−w) ℘(z + w)

Corollary 5.18 The map p 7→ −p on E is given by (x, y) 7→ (x, −y).

Proof. Then the line passes through ∞ which is the origin of E, consistent
with the equation p + (−p) + 0 = 0. (Alternatively, observe that x = ℘(z)
is even and y = ℘(z) is odd.)

Corollary 5.19 The point c = a+b is constructed geometrically by drawing

the line L through (a, b), finding its third point of intersection (−c) = (x, −y)
on E and then negating to get c = (x, y).

89
 Figure 10. Image of 2−5 Λ under ℘, for Λ = Z[i] and Z[ρ].

Corollary 5.20 The point 2a can be constructed by taking the line tangent
to E at a, finding its other point of intersection with E, and then negating
its y coordinate.

Dynamics of rational maps. There is a rational function f (z) such that

℘(2z) = f (℘(z)).

The preimages of the critical points of f lie along the images of the horizontal
and vertical lines in C under ℘, so they give a way of visualizing the ℘
function. For more details, see [Mil].
An elliptic function with given poles and zeros. It is natural to try
to construct an elliptic function by forming the Weierstrass product for the
lattice Λ:  
Y′  z z z2
σ(z) = z 1− exp + .
λ λ 2λ2
Λ

Then (log σ)′′ = −℘(z), from which it follows that

σ(z + λ) = σ(z) exp(aλ + bλ z)

for some aλ , bλ ∈ C. From this it is easy to see that

σ(z − a1 ) . . . σ(z − an )
σ(z − p1 ) . . . σ(z − pn )
P P
defines an elliptic function whenever ai = pi . This demonstrates:
P P
Theorem 5.21 A divisor D = ai − pi P on E is principal (it arises from
a meromorphic function) iff deg D = 0 and (ai − pi ) = 0 in the group law
on E.

90
5.2 Aside: Conics and singly-periodic functions
As a point of reference, we describe the parallel theory for conics and rank
one lattices in C.
Let Λ ⊂ C be a discrete subgroup isomorphic to Z. Then we can rescale
so Λ = Z, and form, for k ≥ 2, the singly-periodic functions
∞
X 1
Zk (z) =
−∞
(z − n)k

and the function

∞
1 X ′ 1 1
P (z) = + − ·
z −∞
(z − n) n
Note that we have the Laurent series (with ζ(0) = −1/2),
X′ 1 X′ 1 ∞
1 3 2X
P (z) = − z − z − · · · = ζ(2k)z 2k ,
z n2 n4 z
0

i.e. the coefficients of P (z) carry interesting invariants of the lattice Λ = Z.

We then find:
b
Theorem 5.22 The map P : C/Z → C−(±πi) is an isomorphism, sending
[0, 1] to [∞, −∞].

Proof. As we have seen,

e2πiz + 1
P (z) = π cot πz = πi = A(e2πiz ),
e2πiz − 1
where A(z) = πi(z + 1)/(z − 1) sends the omitted values 0 and ∞ for e2πiz
to ±i.

Just as we did for elliptic functions, we next note that the power series
1 π2z
P (z) = − + ···
z 3
gives
1 π2 1 2π 2
−P ′ (z) = + + ··· and P (z)2 = − + ···
z2 3 z2 3
yielding the differential equation
−P ′ (z) = P (z)2 + π 2 .
Here we have used the fact that both P (z) and P ′ (z) are bounded as
| Im z| → ∞. Put differently, we have:

91
Theorem 5.23 The map π : C → P2 given by

π(z) = (x, y) = (P (z), −P ′ (z))

gives an isomorphism between the C/Z ∼

= C∗ and the smooth projective conic
2 2
defined by y = x + π with two points removed.

The omitted values on the parabola y = x2 + π 2 are y = 0, where x = ±iπ.

We could also verify the final identity using the fact that

π2
−P ′ (z) = ·
sin2 πz
The fundamental period of Λ can now be expressed as
Z ∞ Z ∞ Z 1
dx dx
= 2 2
= dz = 1.
−∞ y −∞ π + x 0

I.e. the change of variables x = P (z) transforms the integrand into the
standard form dz on C/Z.
Remark. Any smooth conic in P2 is equivalent to the parabola above, so
we have uniformized all conics. The familiar conics x2 + y 2 = 1, x2 − y 2 = 1
and xy = 1 are isomorphic to C/2πZ or C/2πiZ, and are uniformized by
(cos(t), sin(t)), (cosh(t), sinh(t)) and (et , e−t ) respectively. Note that all 3
curves have, over C, two asymptotes, corresponding to the ends of C∗ .

5.3 Moduli spaces and elliptic curves

We now turn to an important complement to the results above.

Theorem 5.24 Suppose the cubic polynomial 4x3 +ax+b has distinct roots.
Then there exists a lattice Λ such that (x, y) = (℘(z), ℘′ (z)) satisfies y 2 =
4x3 + ax + b.

Corollary 5.25 Any smooth cubic curve in P2 is isomorphic to C/Λ for

some Λ ⊂ C.

Moduli spaces. To approach this ‘metatheory’ of doubly-periodic func-

tions — where the lattice is not fixed but allowed to vary — it is useful to
discuss M1 , the moduli space of lattices, and the Teichmüller space T1 ∼
=H
of marked lattices.

92
 As a set, we let

M1 = {all lattices Λ ⊂ C}/(Λ ∼ αΛ)

denote the space of lattices up to similarity. By associating to Λ the complex

torus E = C/Λ, we find

M1 = {all Riemann surfaces of genus one}/(isomorphism).

This is because any isomorphism E1 ∼

= E2 lifts to an automorphism f (z) =
αz + β of the universal cover C.
There is a natural map π : H → M1 , given by

π(τ ) = [Z ⊕ Zτ ] ∈ M1 .

This map is surjective, since Zα⊕Zβ ∼ Z⊕Z(β/α). We can think of a point

τ ∈ H as providing both a lattice Λ = Z ⊕ Zτ and a marking isomorphism,

φ : Z2 → Λ,

coming from the basis (τ, 1). All other bases for Λ with the
same orientation
as this one are given by (aτ + b, cτ + d), where g = ac db ∈ SL2 (Z). The
marked lattice with this new basis is equivalent to (g(τ ), 1) where
aτ + b
g(τ ) = ·
cτ + d
Thus forgetting the marking altogether is the same as taking the quotient
of H by the action of all g ∈ SL2 (Z), and hence:

Theorem 5.26 The moduli space M1 is naturally isomorphic to the com-

plex orbifold H/ SL2 (Z).

Here SL2 (Z) acts by Möbius transformations. The quotient is an orbifold

because the action is not quite free, e.g. τ 7→ −1/τ fixes τ = i.
Lie groups approach. More generally, we can specify a marked lattice
Λ ⊂ Rn , normalized so Rn /Λ has volume one, by a linear map

T : Zn → Λ ⊂ Rn

with T ∈ SLn (R). Two such lattices are similar iff they differ by a rotation,
i.e. T1 = R ◦ T2 where R ∈ SOn (R). Thus the Teichmüller space of lattices
in Rn is the homogeneous space

H = SOn (R)\ SLn (R).

93
For n = 2 we have H ∼ = H because SO2 (R) is the stabilizer of τ = i for the
usual action of SL2 (R) on H.
Finally T1 (Zn ) = T2 (Zn ) iff T1 ◦T2−1 gives an isomorphism of Zn to itself,
which shows
Theorem 5.27 The moduli space of lattices in Rn is isomorphic to

L(Rn ) = SOn (R)\ SLn (R)/ SLn (Z).

Fundamental domains. We can always normalize a lattice Λ by C∗ so

that its shortest nonzero vector is z = 1 and the shortest vector in Λ − Z is
τ ∈ H. Then |τ | ≥ 1, and Z + τ ⊂ Λ so | Re τ | ≤ 1/2. Moreover Z ⊕ Zτ = Λ;
otherwise there would be a vector v ∈ Λ − R of the form v = a + bτ with
a, b ∈ [0, 1/2]; but then

|v| < 1/2 + |τ |/2 ≤ |τ |,

contrary to our assumption that the shortest vector in Λ − Z has length |τ |.

The converse holds as well, and we have:

Theorem 5.28 The region | Re τ | ≤ 1/2, |τ | > 1 in H is a fundamental

domain for the action of SL2 (Z) on H.

The subgroup Γ(2). To study M1 further, we now introduce the space

of cross-ratios
Mf0,4 = C
b − {0, 1, ∞}

and its quotient orbifold

f0,4 /S3 .
M0,4 = M

Here M f0,4 is the moduli space of ordered quadruples of distinct points on C,

b
b
up to the action of Aut(C). Any such quadruple has a unique representative
of the form (∞, 0, 1, λ), giving a natural coordinate for this moduli space.
If we reorder the quadruple, the cross-ratio changes, ranging among the
six values
1 1 λ λ−1
λ, , 1 − λ, , , .
λ 1−λ λ−1 λ
(There is a natural action of S4 , but the Klein 4-group Z/2 × Z/2 acts
trivially.) There are 5 points in Mf0,4 with nontrivial stabilizers under the
action of S3 : the points

{−1, 1/2, 2} = zeros of 2z 3 − 3z 2 − 3z + 2,

94
which each have stabilizer Z/2 and correspond to the vertices of a square;
and the points
√
{ρ, ρ} = {1/2 ± −3/2} = zeros of z 2 − z + 1,

which have stabilizer Z/3 and correspond to the vertices of a tetrahedron.

The degree 6 rational map
4 (z 2 − z + 1)3
F (z) =
27 z 2 (1 − z)2
is invariant under S3 , and gives a natural bijection
b − {0, 1, ∞})/S3 ∼
F : (C =C

satisfying

F (0, 1, ∞) = ∞, F (ρ, ρ) = 0, and F (−1, 1/2, 2) = 1.

(The last fact explains the 4/27.) We should really think of the image as
M0,4 and in particular remember the orbifold structure: Z/2 at F = 1 and
Z/3 at F = 0.
The modular function. We now define a map J : M1 → M0,4 by
associating to any complex torus, the four critical values of the Weierstrass
℘-function. (Note: any degree two map f : X = C/Λ → C b is equivalent, up
to automorphisms of domain and range, to the Weierstrass ℘ function, so
the associated point in M0,4 is canonically determined by X.)
More concretely, given τ ∈ H we define the half-integral points (c1 , c2 , c3 ) =
(1/2, τ /2, (1 + τ )/2) and the corresponding critical values by ei = ℘(ci ).
Then their cross-ratio (together with the critical value at infinity) is given
by:
e3 − e2
λ(τ ) = ·
e1 − e2
For example, we have seen that if τ = iy ∈ iR+ then e2 < e3 < e1 , so
λ(iy) ∈ (0, 1); moreover e3 = 0 and e2 = −e1 for τ = i, and thus λ(i) = 1/2.
The value of λ(τ ) depends only on the ordering of E(2)∗ , the three
nontrivial points of order two on E. Now SL2 (Z) = Aut(Λ) acts on E(2) ∼ =
(Z/2)2 through the natural quotient

0 → Γ(2) → SL2 (Z) → SL2 (Z/2) → 0.

In particular, λ is invariant under the subgroup Γ(2) of matrices equivalent

to the identity modulo two.

95
 Now any elliptic element in SL2 (Z) has trace −1, 0 or 1, while the trace
of any element

in Γ(2) must be even. Moreover, trace zero cannot arise:
if g = ac −a
b ∈ Γ(2) then −a2 − bc = 1 implies −a2 = 1 mod 4 which is
impossible.
By assembling 6 copies of the fundamental domain for SL2 (Z) (some cut
into two pieces), one can then show:

Theorem 5.29 The group Γ(2) is torsion-free, with fundamental domain

the ideal quadrilateral with vertices {∞, −1, 0, 1}.

We may now state the main result relating lattices and cross-ratios.

Theorem 5.30 The natural map

f1 = H/Γ(2) → M
λ:M b − {0, 1, ∞}
f0,4 = C

sending a torus to the cross-ratio of the critical values of ℘ is a holomorphic

bijection, respecting the action of S3 .

Proof. We first check that λ is injective. If z = λ(τ1 ) = λ(τ2 ), then

the corresponding complex tori E1 , E2 both admit degree two maps to C b
branched over (0, 1, ∞, z). Lifting one map composed with the inverse of
the other gives an isomorphism E1 → E2 , and hence τ1 ∈ SL2 (Z) · τ2 .
We have also seen that there exists a continuous map µ : M0,4 → M1
which satisfies µ(λ(τ )) = [τ ]; it sends z to the lattice spanned by integrals of
the form dw/(w(w − 1)(w − z)). The existence of this map shows the image
of λ is closed. It follows that λ is a bijection.

As we have seen, if τ ∈ iR then e1 , e2 , e3 ∈ R. This shows:

Theorem 5.31 The function λ is real on the orbit of the imaginary axis
under SL2 (Z).

This orbit gives the edges of a tiling of H by ideal triangles; see Figure
11.
Corollary 5.32 The map λ gives an explicit Riemann map from the ideal
triangle spanned by 0, 1 and ∞ to H, fixing these three points.

Thus Sλ−1 is given by the quadratic differential Q(0, 0, 0) discussed ear-

lier, and λ−1 itself can be given as a ratio of solutions to a linear differential
equation of order two.
Picard’s theorem revisited. This gives the one line proof of the Little
Picard Theorem: ‘consider λ−1 ◦ f : C → H.’

96
 Figure 11. Tiling for Γ(2).

The question recently arose in conversation whether a disserta-

tion of 2 lines could deserve and get a Fellowship... in mathe-
matics the answer is yes....
(Theorem.) An integral function never 0 or 1 is a constant.
(Proof.) exp{iΩ(f (z))} is a bounded integral function.
...But I can imagine a referee’s report: ‘Exceedingly striking and
a most original idea. But, brilliant as it undoubtedly is, it seems
more odd than important; an isolated result, unrelated to any-
thing else, and not likely to lead anywhere.’
—J. E. Littlewood.

Cf. [Bol, p.39–40]: Here Ω is our λ−1 .

The J-function. To remove the ambiguity of ordering, we now define

4 (λ2 − λ + 1)3
J(τ ) = F (λ(τ )) = ·
27 λ2 (1 − λ)2
The S3 -equivariance of λ then implies:

Theorem 5.33 The map

b − {0, 1, ∞})/S3
J : M1 = H/ SL2 (Z) → M0,4 = (C

is a bijection, and a isomorphism of orbifolds.

Forgetting the orbifold structure, we get an isomorphism J : M1 ∼=

C. In other words, J(τ ) is a complex number which depends only on the

97
isomorphism class of E = C/(Z ⊕ Zτ ); we have J(τ1 ) = J(τ2 ) iff E1 ∼= E2 ;
and every complex number arises as J(τ ) for some τ .
Classical proof. Here is the classical argument that J is a bijection. The
value of J(τ ) determines a quadruple B ⊂ C b which in turn determines
b
a unique Riemann surface X → C of degree two, branched over B, with
X∼ = C/Z ⊕ τ Z. Thus J(τ1 ) = J(τ2 ) iff the corresponding complex tori are
isomorphic iff τ1 = g(τ2 ) for some g ∈ SL2 (Z). Thus J is injective.
To see it is surjective, we first observe that J(τ + 1) = J(τ ). Now if
Im τ = y → ∞, then on the region | Im z| < y/2 we have

1 X′ 1 1 π2
℘(z) = + − + ǫ(z) = + C + ǫ(z),
z2 (z − n)2 n2 sin2 (πz)
where ǫ(z) → 0 as y → ∞. (In fact we have
X′ π2
ǫ(z) = = O(e−πy ).
n
sin2 (π(z + nτ ))

We really only need that it tends to zero; and we will not need the exact
value of the constant C = −π 2 /3).
Since sin(z) grows rapidly as | Im z| grows, we have

(e1 , e2 , e3 ) = (π 2 + C, C, C) + O(ǫ)

and thus λ(τ ) = (e3 − e2 )/(e1 − e2 ) → 0 and hence J(τ ) → ∞.

Thus J is a proper open map, which implies it is surjective (its image is
open and closed).

5.4 Modular forms

In this section we will find an explicit formula for J(τ ), and give a com-
plete analysis of the holomorphic forms and functions on H (with controlled
growth) that are invariant under SL2 (Z).
Modular forms: analytic perspective. A holomorphic function f :
H → C is said to be a modular form of weight 2k for SL2 (Z) if it has the
following properties:
1. f (τ + 1) = f (τ );

2. f (−1/τ ) = τ 2k f (τ ); and

3. supIm τ >1 |f (τ )| < ∞.

98
The vector space of all such forms will be denoted by Mk . The product
forms of weight 2k and 2ℓ has weight 2(k + ℓ), so ⊕Mk forms a graded ring.
The first two properties imply that

f (g(τ )) = (cτ + d)2k f (τ )

a b


for all g = c d ∈ SL2 (Z). The first and third properties imply that we can
write
∞
X
f (τ ) = an q n ,
0

where q = exp(2πiτ ). In other words, f (τ ) descends to a holomorphic

function on ∆∗ = H/Z with a removable singularity at q = 0. In particular,
f (τ ) → a0 as Im τ → ∞. One sometimes writes f (i∞) = a0 .
If f (i∞) = 0, one says that f (τ ) is cusp form. The cusp forms (Spitzen-
formen) give a natural subspace Sk ⊂ Mk .
Examples.

1. Any holomorphic invariant F (Λ) of lattices that is homogeneous of

degree −2k and satisfies the right growth conditions defines a modular
form of weight 2k. That is, if F (tΛ) = t2k Λ, then the function

f (τ ) = F (Z ⊕ τ Z)

satisfies f (τ + 1) = f (τ ) and

f (−1/τ ) = F (Z ⊕ τ −1 Z) = F (τ −1 (Z ⊕ τ Z) = τ 2k f (τ ).

2. In particular, for k ≥ 2 the Eisenstein series

X′
Gk (τ ) = (n + mτ )−2k

are modular forms of weight 2k. Evidentally Gk converges to 2ζ(2k)

as Im τ → ∞, so these are holomorphic at infinity but not cusp forms.

3. Similarly for the normalized functions g2 = 60G2 , g3 = 140G3 .

Modular forms: algebraic perspective. If we associate to f (τ ) the

form
ω = f (τ ) dτ k ,

99
then the first two conditions above just say that g∗ ω = ω for all g ∈ SL2 (Z).
(Recall that g′ (z) = (cz +d)−2 .) Since dz = (2πi)−1 dq/q, the third condition
says that ω descends to a form
 k
−k dq
ω = (2πi) f (q)
q

on ∆∗ = H/Γ with at worst a pole of order k at q = 0.

Thus ω also descends to an S3 -invariant holomorphic k-form on
b − {0, 1, ∞} = H/Γ(2)
f0,4 = C
M

with poles of order ≤ k at 0, 1 and ∞. The converse is also true. This shows:

Theorem 5.34 The space Mk is naturally isomorphic to the space of S3 -

b with poles of order ≤ k at 0, 1, ∞
invariant rational forms ω(z) dz k on C
and no other poles.

This provides a purely algebraic perspective on the initially transcendental–

looking theory of automorphic forms.
We now wish to find all the rational holomorphic k-forms ω ∈ Mk . Here
are 2 key properties of any nonzero ω ∈ Mk :

(i) The zeros Z(ω) are invariant under S3 and satisfy |Z(ω)| ≤ k,
when counted with multiplicity, and determine ω up to constant
multiple.
(ii) The poles of ω at 0, 1, ∞ are all of the same order, which can
be k, k − 2, k − 4, etc. We have ω ∈ Sk iff the order of pole is
k − 2 or less.

Part (i) comes from the fact that ω has at most 3k poles, hence at most
3k − 2k = k zeros. The parity constraint in (ii) comes from the fact that
g(z) = 1 − z leaves ω invariant, and g ′ (∞) = −1.
Examples.

1. We have M0 = C; it consists of the constant functions.

2. We have dim M1 = 0; the group S3 has no fixed point, so there is no

candidate for Z(ω).

100
3. We have dim M2 = 1. The only possibility is Z(ω) = {ρ, ρ}; thus the
quadratic differential
(z 2 − z + 1) dz 2
F2 =
z 2 (z − 1)2
spans M2 . (We have met this differential before in the study of Schwarz
triangle functions.)
4. Similarly, dim M3 = 1; it is spanned by the cubic differential
(z − 2)(z − 1/2)(z + 1) dz 3
F3 = ,
z 3 (z − 1)3
which has zeros at −1, 1/2, 2.
5. The products F22 and F3 F2 span M4 and M5 . This is because S3 has
unique invariant sets with |Z| = 4 and 5.
6. We have dim M6 = 2; it is spanned by F23 and F32 . These two forms
are linearly independent because they have different zero sets.
7. The discriminant
4 dz 6
D6 = (F23 − F32 ) = 4
27 z (z − 1)4
is the first nontrivial cusp form; it has poles of order 4 at (0, 1, ∞) and
no zeros. It spans S6 .
8. Ratios of forms of the same weight give all S3 -invariant rational func-
tions. Indeed, as we have seen, an isomorphism C/S b 3 ∼ = Cb sending
(∞, ρ, 2) to (∞, 0, 1),is given by
F23 (z) 4 (z 2 − z + 1)3
F (z) = = ·
F2 (z)3 − F3 (z)2 27 z 2 (z − 1)2
More generally, any S3 -invariant rational function of degree d can be
expressed as a ratio of modular forms of degree d.
b 3 by substituting
9. If desired, we can regard F2 and F3 as forms on C/S
w = J(z). They then come:
dw2 dw3
F2 = and F3 = ·
4w(w − 1) 8w2 (w − 1)
We also find D6 = (1/432)(w−4 (w − 1)−3 ) dw6 .

101
Cusp forms. We let Sk ⊂ Mk denote the space of cusp forms with poles
of order ≤ k − 2 at 0, 1, ∞.

Theorem 5.35 The map ω 7→ D6 ω gives an isomorphism from Mk to Sk+6 .

Proof. Any F ∈ Mk has poles of order ≤ k, so F D6 has poles of order

≤ k + 4 and hence lies in Sk+6 . Conversely, if G ∈ Sk+6 then F = G/D6 has
poles of order at most k at 0, 1, ∞, and it is otherwise holomorphic since D6
has no zeros.

Every space Mk , k ≥ 2 contains a form of the type F2i F3j which is not a
cusp form; thus Mk ∼
= C ⊕ Sk . By inspection, dim Sk = 0 for k ≤ 5. On the
other hand, any cusp form is divisible by D6 . This shows:

Corollary 5.36 We have dim M6n+1 = n, and dim M6n+i = n + 1 for

i = 0, 2, 3, 4, 5.

Corollary 5.37 The forms F2 and F3 generate the ring M = ⊕Mk .

Corollary 5.38 The forms F2i F3j with 2i + 3j = k form a basis for Mk .

Proof. By the preceding Corollary these forms span Mk , and the number
of them agrees with dim Mk as computed above.

Corollary 5.39 The ring of modular forms M is isomorphic to the poly-

nomial ring C[F2 , F3 ].

Proof. The preceding results shows the natural map from the graded ring
C[F2 , F3 ] to M is bijective on each graded piece.

We have already seen that, under the isomorphism M1 ∼ = M0,4 , an

analytic modular form is the same as an algebraic modular form. Thus we
also have:

Corollary 5.40 The ring of modular forms is isomorphic to C[g2 , g3 ].

X′
Corollary 5.41 For every k ≥ 2, the quantity λ−2k can be expressed
X′ X′
as a polynomial in λ−4 and λ−6 .

102
Values of g2 and g3 . We now note that for τ = i, the zeros of 4℘3 −g2 ℘−g3
must look like (−1, 0, 1) and thus g3 (i) = 0. Similarly for τ = ρ the zeros
are arrayed like the cube roots of unity and hence g3 (ρ) = 0.
To determine the values at infinity, we observe that as τ → ∞ we have

G2 (τ ) → 2ζ(4) = π 4 /45

and
G3 (τ ) → 2ζ(6) = 4π 6 /945.
This gives the values g2 (∞) = 60G2 (∞) = (4/3)π 4 and g3 (∞) = 140G3 (∞) =
(8/27)π 6 , and thus
(g23 /g32 )(i∞) = 27.

The cusp form ∆ and the modular function J. By the preceding

calculation, the discriminant

∆(τ ) = g23 (τ ) − 27g3 (τ )2

is a cusp form of weight 12. We have seen such a form is unique up to

a scalar multiple, and is nonvanishing everywhere except for a simple zero
at infinity. This form has a natural meaning: up to a multiple, it is the
discriminant of the cubic polynomial 4x3 − g2 x − g3 . As we have seen, this
polynomial has distinct roots whenever g2 and g3 come from a lattice; this
explains why ∆(τ ) has no zeros in H.

Theorem 5.42 We have J(τ ) = g23 (τ )/∆(τ ).

Proof. First note that this is a ratio of forms of weight 12 and hence a
modular function, i.e. it is invariant under SL2 (Z). Since g2 (∞) 6= 0, it
has a simple pole at infinity, and thus has degree one on M1 . We also have
J(ρ) = 0 since g2 (ρ) = 0, and J(i) = 1 since g3 (i) = 0.

Remarks. One can use residues to determine the exact relationship be-
tween F2 and g2 . Namely, F2 ∼ (1/4) dw2 /w2 , while for q = exp(2πiτ ) we
have dq = 2πiq dτ , and hence

4π 4 dq 2 π 2 dq 2
g2 = 60G2 (τ ) dτ 2 ∼ = − ·
3 (2πiq)2 3 q2

Thus g2 = −(4π 2 /3)F2 . One can similarly calculate g3 /F3 .

103
 If we let j(τ ) = 1728J(τ ), then

X ∞
1
j(τ ) = + 744 + an q n
q 1

with an ∈ Z.
Connections with additive and multiplicative number theory. Incredibly,
we have
∞
Y
∆(q) = (2πi) q (1 − q n )24 .
12

This gives a close connection between the theory of modular forms and the
partition function p(n), since
Y X
(1 − q n )−1 = p(n)q n .

P Thek coefficients power series for Gk (τ ) involves the function σk (n) =

d|n d .
The Riemann function ζ(s) arises as the Mellin transform of a theta
function, and then modularity translates into the functional equation.

References
[BD] F. Berteloot and J. Duval. Sur l’hyperbolicité de certain
complèmentaires. Enseign. Math. 47(2001), 253–267.

[Bol] B. Bollabas, editor. Littlewood’s Miscellany. Cambridge University

Press, 1986.

[CJ] L. Carleson and P. Jones. On coefficient problems for univalent func-

tions. Duke Math. J. 66(1992), 169–206.

[Gam] T. W. Gamelin. Uniform Algebras. Prentice–Hall, 1969.

[Gol] G. M. Goluzin. Geometric Theory of Functions of a Complex Vari-

able. Amer. Math. Soc, 1969.

[GM] J. D. Gray and S. A. Morris. When is a function that satisfies the

Cauchy-Riemann equations analytic? The American Mathematical
Monthly 85(1978), 246–256.

104
[Ko] S. V. Kolesnikov. On the sets of nonexistence of radial limits of
bounded analytic functions. Russian Acad. Sci. Sb. Math. 81(1995),
477–485.

[MH] J. E. Marsden and M. J. Hoffman. Basic Complex Analysis. W. H.

Freeman, 1999.

[Mil] J. Milnor. On Lattès maps. In P. G. Hjorth and C. L. Petersen,

editors, Dynamics on the Riemann sphere, pages 9–44. European
Math. Soc., 2006.

[Mon] M. G. Monzingo. Why are 8:18 and 10:09 such pleasant times? Fi-
bonacci. Quart. 2(1983), 107–110.

[Re] R. Remmert. Classical Topics in Complex Function Theory.

Springer, 1998.

[Sa] R. Salem. Algebraic Numbers and Fourier Analysis. Wadsworth,

1983.

105