Complex Analysis:

The Geometric Viewpoint

Second Edition
 ©2004 by
The Mathematical Association of America (Incorporated)
Library of Congress Catalog Card Number 2003114309
2nd edition Print ISBN 978-0-88385-035-0
Electronic ISBN 978-0-88385-968-1
Printed in the United States of America
 The Carus Mathematical Monographs

Number Twenty-Three

Complex Analysis:
The Geometric Viewpoint
Second Edition

Steven G. Krantz
Washington University in St. Louis

Published and Distributed by

THE MATHEMATICAL ASSOCIATION OF AMERICA
 THE
CARUS MATHEMATICAL MONOGRAPHS

Published by
THE MATHEMATICAL ASSOCIATION OF AMERICA

Committee on Publications
Gerald L. Alexanderson, Chair
Editorial Board
Kenneth A. Ross, Chair
Joseph Auslander
Harold P. Boas
Robert E. Greene
Roger Horn
Jeffrey Lagarias
Barbara Osofsky
 The following Monographs have been published:

1. Calculus of Variations, by G. A. Bliss (out of print)

2. Analytic Functions of a Complex Variable, by D. R. Curtiss (out of
print)
3. Mathematical Statistics, by H. L. Rietz (out of print)
4. Projective Geometry, by J. W. Young (out of print)
5. A History of Mathematics in America before 1900, by D. E. Smith
and Jekuthiel Ginsburg (out of print)
6. Fourier Series and Orthogonal Polynomials, by Dunham Jackson
(out of print)
7. Vectors and Matrices, by C. C. MacDuffee (out of print)
8. Rings and Ideals, by N. H. McCoy (out of print)
9. The Theory of Algebraic Numbers, second edition, by Harry Pollard
and Harold G. Diamond
10. The Arithmetic Theory of Quadratic Forms, by B. W. Jones (out of
print)
11. Irrational Numbers, by Ivan Niven
12. Statistical Independence in Probability, Analysis and Number The-
ory, by Mark Kac
13. A Primer of Real Functions, third edition, by Ralph P. Boas, Jr.
14. Combinatorial Mathematics, by Herbert J. Ryser
15. Noncommutative Rings, by I. N. Herstein (out of print)
16. Dedekind Sums, by Hans Rademacher and Emil Grosswald
17. The Schwarz Function and Its Applications, by Philip J. Davis
18. Celestial Mechanics, by Harry Pollard
19. Field Theory and Its Classical Problems, by Charles Robert Had-
lock
20. The Generalized Riemann Integral, by Robert M. McLeod
21. From Error-Correcting Codes through Sphere Packings to Simple
Groups,
by Thomas M. Thompson
22. Random Walks and Electric Networks, by Peter G. Doyle and J.
Laurie Snell
23. Complex Analysis: The Geometric Viewpoint, second edition, by
Steven G. Krantz
24. Knot Theory, by Charles Livingston
25. Algebra and Tiling: Homomorphisms in the Service of Geometry,
by Sherman Stein and Sándor Szabó
26. The Sensual (Quadratic) Form, by John H. Conway assisted by
Francis Y. C. Fung
27. A Panorama of Harmonic Analysis, by Steven G. Krantz
28. Inequalities from Complex Analysis, by John P. D’Angelo
29. Ergodic Theory of Numbers, by Karma Dajani and Cor Kraaikamp

MAA Service Center

P. O. Box 91112
Washington, DC 20090-1112
800-331-1MAA FAX: 301-206-9789
To my parents
Acknowledgments

This book owes its existence to many people. I thank Don Albers,
Robert E. Greene, and Paul Halmos for convincing me to write it. I am
grateful to Marco Abate, Harold Boas, Ralph Boas, David Drasin, Paul
Halmos, Daowei Ma, David Minda, Marco Peloso, John Stapel, and
Jim Walker for reading various versions of the manuscript and making
valuable comments and suggestions. The Carus monograph committee
of the MAA helped me to find the right level and focus for the book.
Ralph Boas, the chairman of the Carus monograph committee,
played a special role in the development of this book. In addition to
shepherding the project along, he provided necessary prodding and ca-
joling at crucial stages to keep the project on track. Paul Halmos also
provided expert and sure counsel; he has been a good friend and mentor
for many years. To both men I express my sincere gratitude.
For the new edition, Ken Ross as Chair of the Carus Monographs
Editorial Board provided a strong and sure guiding hand. My friends
Harold Boas, Daowei Ma, David Minda, Jeff McNeal, Marco Peloso,
and Jim Walker offered valuable suggestions for the form of the new
edition. Robert Burckel, Robert E. Greene, and John P. D’Angelo were
especially generous in offering detailed commentary which led to deci-
sive improvements. Finally, the Carus Monograph Committee provided
yeoman service with many careful readings, sharp but constructive crit-
icisms, and helpful suggestions. Surely a better book has been the re-
sult.
ix
x Acknowledgments

I thank Jim Milgram for creating TECPRINT, the world’s first

technical word processor. This has been a valuable tool to me for many
years. I am grateful to Micki Wilderspin for performing the tedious task
of converting the manuscript from TECPRINT into TEX. Beverly Joy
Ruedi did a splendid job of making my project into a finished book.
A special vote of thanks goes to Robert Burckel for reading most
of the manuscript inch by inch and correcting both my mathematics and
my writing. His contributions have improved the quality of the book
decisively. Responsibility for all remaining errors of course resides en-
tirely with me.

—S.G.K.
Preface to the Second Edition

The warm reception with which the first edition of this book has been
received has been a source of both pride and pleasure. It is a special
privilege to have created a “for the record” version of Ahlfors’s seminal
ideas in the subject. And the geometric viewpoint continues to develop.
In the intervening decade, this author has learned a great deal more
about geometric analysis, and his view of the subject has developed and
broadened. It seems appropriate, therefore, to bring some new life to
these pages, and to set forth a fresh enunciation of the role of curvature
in basic complex function theory.
In this new edition, we explain how, in a natural and elementary
manner, the hyperbolic disc is a model for the non-Euclidean geometry
of Bolyai and Lobachevsky. Later on, we explain the Bergman kernel
and provide an introduction to the Bergman metric.
I have many friends and colleagues to thank for their incisive re-
marks and suggestions about the first edition of this book. I hope that I
do them justice in my efforts to implement a second edition. As always,
the Mathematical Association of America has been an exemplary pub-
lisher and has provided all possible support in the publication process.
I offer my humble thanks.

xi
Preface to the First Edition

The modern geometric point of view in complex function theory began

with Ahlfors’s classic paper [AHL]. In that work it was demonstrated
that the Schwarz lemma can be viewed as an inequality of certain differ-
ential geometric quantities on the disc (we will later learn that they are
curvatures). This point of view—that substantive analytic facts can be
interpreted in the language of Riemannian geometry—has developed
considerably in the last fifty years. It provides new proofs of many
classical results in complex analysis, and has led to new insights as
well.
In this monograph we intend to introduce the reader with a stan-
dard one semester background in complex analysis to the geometric
method. All geometric ideas will be developed from first principles,
and only to the extent needed here. No background in geometry is as-
sumed or required.
Chapter 0 gives a bird’s eye view of classical function theory of
one complex variable. We pay special attention to topics which are de-
veloped later in the book from a more advanced point of view. In this
chapter we also sketch proofs of the main results, with the hope that
the reader can thereby get a feeling for classical methodology before
embarking on a study of the geometric method.
Chapter 1 begins a systematic treatment of the techniques of Rie-
mannian geometry, specially tailored to the setting of one complex vari-
able. In order that the principal ideas may be brought out most clearly,
xiii
xiv Preface to the First Edition

we shall concentrate on only a few themes: the Schwarz lemma, the

Riemann mapping theorem, normal families, and Picard’s theorems.
For many readers this will be a first contact with the latter two results.
The geometric method provides a particularly cogent explanation of
these theorems, and can be contrasted with the more classical proofs
which are discussed in Chapter 0. We shall also touch on Fatou-Julia
theory, a topic which is rather technical from the analytic standpoint
but completely natural from the point of view of geometry.
In Chapter 3 we introduce the Carathéodory and Kobayashi met-
rics, a device which is virtually unknown in the world of one complex
variable. This decision allows us to introduce invariant metrics on ar-
bitrary planar domains without resort to the uniformization theorem.
We are then able to give a “differential geometric” interpretation of the
Riemann mapping theorem.
The last chapter gives a brief glimpse of several complex vari-
ables. Some of the themes which were developed earlier in the book
are carried over to two dimensions. Biholomorphic mappings are dis-
cussed, and the inequivalence of the ball and bidisc is proved using a
geometric argument.
The language of differential geometry is not generally encoun-
tered in a first course in complex analysis. It is hoped that this volume
will be used as a supplement to such a course, and that it may lead to
greater familiarity with the fruitful methodology of geometry.
Contents

Acknowledgments ix
Preface to the Second Edition xi
Preface to the First Edition xiii

0 Principal Ideas of Classical Function Theory 1

1. A Glimpse of Complex Analysis . . . . . . . . . . . . . . . . . . . . . . 1
2. The Maximum Principle, the Schwarz Lemma, and
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3. Normal Families and the Riemann Mapping Theorem . . . 16
4. Isolated Singularities and the Theorems of Picard . . . . . . . 22

1 Basic Notions of Differential Geometry 29

0. Introductory Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1. Riemannian Metrics and the Concept of Length . . . . . . . . 30
2. Calculus in the Complex Domain . . . . . . . . . . . . . . . . . . . . . 38
3. Isometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4. The Poincaré Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5. The Schwarz Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6. A Detour into Non-Euclidean Geometry . . . . . . . . . . . . . . . 58

xv
xvi Contents

2 Curvature and Applications 67

1. Curvature and the Schwarz Lemma Revisited . . . . . . . . . . 67
2. Liouville’s Theorem and Other Applications . . . . . . . . . . . 73
3. Normal Families and the Spherical Metric . . . . . . . . . . . . . 79
4. A Generalization of Montel’s Theorem and the
Great Picard Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3 Some New Invariant Metrics 89
0. Introductory Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
1. The Carathéodory Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2. The Kobayashi Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3. Completeness of the Carathéodory and Kobayashi
Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4. An Application of Completeness: Automorphisms . . . . . . 121
5. Hyperbolicity and Curvature . . . . . . . . . . . . . . . . . . . . . . . . . 133
4 Introduction to the Bergman Theory 137
0. Introductory Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
1. Bergman Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
2. Invariance Properties of the Bergman Kernel . . . . . . . . . . . 140
3. Calculation of the Bergman Kernel . . . . . . . . . . . . . . . . . . . . 143
4. About the Bergman Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5. More on the Bergman Metric . . . . . . . . . . . . . . . . . . . . . . . . . 155
6. Application to Conformal Mapping . . . . . . . . . . . . . . . . . . . 156

5 A Glimpse of Several Complex Variables 161

0. Functions of Several Complex Variables . . . . . . . . . . . . . . . 161
1. Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
2. The Automorphism Groups of the Ball and Bidisc . . . . . . 170
3. Invariant Metrics and the Inequivalence of the Ball
and the Bidisc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Contents xvii

Appendix 191
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
2. Expressing Curvature Intrinsically . . . . . . . . . . . . . . . . . . . . 191
3. Curvature Calculations on Planar Domains . . . . . . . . . . . . . 200

Symbols 205
References 209

Index 213
CHAPTER 0
Principal Ideas of Classical
Function Theory

1. A Glimpse of Complex Analysis

The purpose of this book is to explain how various aspects of complex
analysis can be understood both naturally and elegantly from the point
of view of metric geometry. Thus, in order to set the stage for our work,
we begin with a review of some of the principal ideas in complex analy-
sis. A good companion volume for this introductory material is [GRK].
See also [BOAS] and [KR3].
Central to the subject are the Cauchy integral theorem and the
Cauchy integral formula. From these follow the Cauchy estimates,
Liouville’s theorem, the maximum principle, Schwarz’s lemma, the
argument principle, Montel’s theorem, and most of the other power-
ful and elegant results which are basic to the subject. We will discuss
these results in essay form. The proofs which we provide are more con-
ceptual than rigorous: the aim is to depict a flow of ideas rather than
absolute mathematical precision.
We let z ∈ C denote a complex variable. If P ∈ C and r > 0, then
we use the standard notation
1
2 Principal Ideas of Classical Function Theory

D(P, r ) = {z ∈ C : |z − P| < r } ,
D(P, r ) = {z ∈ C : |z − P| ≤ r } ,
∂ D(P, r ) = {z ∈ C : |z − P| = r } .

We often use the lone symbol D to denote the unit disc D(0, 1). A
connected open set U ⊆ C is called a domain.
Complex analysis consists of the study of holomorphic functions.
Let F be a complex-valued continuously differentiable function (in the
sense of multivariable calculus) on a domain U in the complex plane.
We write F = u + iv to distinguish the real and imaginary parts of F.
Then F is said to be holomorphic, or analytic, if it satisfies the Cauchy–
Riemann equations:

∂u ∂v ∂u ∂v
= and =− .
∂x ∂y ∂y ∂x

This definition is equivalent to other familiar definitions, such as that

in terms of the complex derivative, which we now discuss. If F is a
function on a domain U in the complex plane and if P ∈ U , then F is
said to possess a complex derivative, or to be complex differentiable, at
P if
∂F F(z) − F(P)
F  (P) = (P) = lim
∂z z→P z−P

exists. The function F is holomorphic on U if F possesses the complex

derivative at each point of U .
Another useful approach to complex analytic (or holomorphic)
functions is by way of power series: a function on a domain U is holo-


morphic if it has a convergent power series expansion j a j (z − P) j
about each point P of U .
The complex derivative definition of “holomorphic” is of great
historical interest. Much effort was expended in the early days of the
subject in proving that a function which is complex differentiable at
each point of a domain U is in fact automatically continuously dif-
ferentiable (in the usual sense of multivariable calculus), and from
A Glimpse of Complex Analysis 3

that point it is routine to check that the function satisfies the Cauchy–
Riemann equations. The converse implication is a straightforward ex-
ercise. So, in the end, either definition is correct. From our perspective
the Cauchy–Riemann equations provide the most useful point of view.
This assertion will become more transparent as we develop the notion
of complex integration.

Definition 1. A C 1 , or continuously differentiable, curve in a domain

U ⊆ C is a function γ : [a, b] → U from an interval in the real
line into U such that γ  exists at each point of [a, b] (in the one-sided
sense at the endpoints) and is continuous on [a, b]. When there is no
danger of confusion, we sometimes use the symbol γ to denote the set
of points {γ (t) : t ∈ [a, b]} as well as the function from [a, b] to U .

A piecewise continuously differentiable curve is a single continu-

ous curve which can be written as a finite union of continuously differ-
entiable curves—Figure 1. A curve is called closed if γ (a) = γ (b).
It is called simple closed if it is closed and not self-intersecting:
γ (s) = γ (t) and s = t together imply either that s = a and t = b
or that s = b and t = a.

Figure 1.

A simple closed curve γ is said to be positively oriented if the

region interior to the curve is to the left of the curve while it is being
traversed from t = a to t = b. See Figure 2. Otherwise it is called neg-
atively oriented. If F is a continuous function on our open set U , then
4 Principal Ideas of Classical Function Theory

Figure 2.

we define its complex line integral over a continuously differentiable

curve γ in U to be the quantity
  b
F(z) dz = F(γ (t)) · γ  (t) dt.
γ a

Here the dot · denotes multiplication of complex numbers.

Notice that, in analogy with the study of directed curves in
Stokes’s formula, the derivative of the curve is incorporated into the
integral. In case γ is only piecewise continuously differentiable, we
define

F(z) dz
γ

by integrating along each of the continuously differentiable pieces and

adding.
Now we may formulate the Cauchy integral theorem. A rigorous
treatment of this result requires a discussion of deformation of curves.
However, since this is only a review, we may be a bit imprecise. Let γ
be a closed curve in a domain U and suppose that γ (more precisely,
the image of γ ) can be continuously deformed to a point within U . We
shall call such a curve “topologically trivial (with respect to U ).” In
Figure 3, γ2 is topologically trivial but γ1 is not. We have:
A Glimpse of Complex Analysis 5

Figure 3.

Theorem 2. Let F be a holomorphic function on a domain U and let γ

be a topologically trivial, piecewise continuously differentiable, closed
curve in U . Then

F(z) dz = 0.
γ

This theorem may be proved, using the Cauchy–Riemann equations, as

a direct application of Stokes’s theorem (see [GRK]). It will tell us, in
effect, that a holomorphic function is strongly influenced on an open
set by its behavior on the boundary of that set.
Now fix a point P in U and let γ be a positively oriented, topo-
logically trivial, simple closed curve in U with P in its interior. Let F
be holomorphic on U . By suitable limiting arguments, we may apply
the Cauchy integral theorem to the function

 F(ζ ) − F(P)
 if ζ = P,
G(ζ ) ≡ ζ−P

 
F (P) if ζ = P.
6 Principal Ideas of Classical Function Theory

After some calculations, the result is that


1 F(ζ )
F(P) = dζ.
2πi γ ζ − P

This is the Cauchy integral formula. It shows that a holomorphic func-

tion is completely determined in the interior of γ by its behavior on
the boundary curve γ itself. From this there quickly flows a wealth of
information.

Theorem 3. Let F be holomorphic on a domain U and let P ∈ U .

Assume that the closed disc

D(P, r ) ≡ {z : |z − P| ≤ r }

is contained in U . Then F may be written on D(P, r ) as a convergent

power series:

∞
F(z) = a j (z − P) j .
j=0

The convergence is absolute and uniform on D(P, r ).

Thus we see that, in a natural sense, holomorphic functions are gener-

alizations of complex polynomials. The power series expansion is, in
general, only local. But for many purposes this is sufficient.

Proof of Theorem 3. Observe that, for |z − P| < r and |ζ − P| = r ,

we may write

1 1 1
= · .
ζ −z ζ − P 1 − ζz−P
−P

Since |z − P| < r = |ζ − P|, we have that

z − P

ζ − P < 1.
A Glimpse of Complex Analysis 7

Thus
∞

 
1 1 z−P j
= · .
ζ −z ζ − P j=0 ζ − P

Substituting this power series expansion for the Cauchy kernel into the
Cauchy integral formula on D(P, r ) gives the desired power series ex-
pansion for the holomorphic function F.

As an added bonus, the proof gives us a formula for the series

coefficients a j :

1 F(ζ )
aj = dζ.
2πi γ (ζ − P) j+1
Just as in the theory of Taylor series, it turns out that the coefficients a j
must also be given by


1 ∂jF
aj = (P).
j! ∂z j
We conclude that

j  
∂ F j! F(ζ )
(P) = dζ. (∗)
∂z j 2πi γ (ζ − P) j+1

Corollary 3.1. (Riemann removable singularities theorem). We let

 be a holomorphic function on a punctured disc D  (P, r ) ≡ D(P, r )\
F
 is bounded, then F
{P}. If F  continues analytically to the entire disc
D(P, r ). That is, there is a holomorphic function F on D(P, r ) such
that F  
= F.
D (P,r )

Sketch of Proof. Assume without loss of generality that P = 0. Con-

 on D  (P, r ) and
sider the function G(z) that is defined to equal z 2 · F
to equal 0 at P = 0. Then G is continuously differentiable on D(P, r )
and satisfies the Cauchy–Riemann equations.
8 Principal Ideas of Classical Function Theory

The leading term of the power series expansion of G about 0 is of

the form a2 z 2 . Thus the holomorphic function G may be divided by z 2
to define a holomorphic function F on D(P, r ) which agrees with F 

on D (P, r ).

It is a standard fact from the theory of power series that the zeros
of a function given by a power series expansion cannot accumulate in
the interior of the domain of that function. Thus we have:

Theorem 4. If F is holomorphic on a domain U, then

{z ∈ U : F(z) = 0} has no accumulation point in U .

This theorem once again bears out the dictum that holomorphic func-
tions are much like polynomials: The zero set of a polynomial a0 +
a1 z + a2 z 2 + · · · + an z n is discrete, indeed it is finite.
The Cauchy estimates on the derivatives of a holomorphic func-
tion in terms of the supremum of the function follow from direct esti-
mation of the formula (∗):

Theorem 5. Let F be a holomorphic function on a domain U that

contains the closed disc D(P, R). Let M be the supremum of |F| on
D(P, R). Then the derivatives of F satisfy the estimates

j 
∂ j! · M

∂z j F(P) ≤ R j .

An immediate corollary of the Cauchy estimates is the fact that

if a sequence of holomorphic functions converges then so does the se-
quence of its derivatives:

Corollary 5.1. Let {F j } be a sequence of holomorphic functions on

a domain . Suppose that the sequence converges uniformly on com-
pact subsets of . Then the sequence {F j } also converges uniformly on
compact subsets of .
A Glimpse of Complex Analysis 9

Notice that the Cauchy estimates tell us that if F is bounded on a

large disc, then its derivatives are relatively small at the center of the
disc; this assertion is exploited in the next result.

Theorem 6. (Liouville). Let F be a holomorphic function on the com-

plex plane (an entire function) which is also bounded. Then F must be
a constant.

Proof. Assume without loss of generality that |F| is bounded by 1. Fix

a point P in the plane. Applying the Cauchy estimates to F on the disc
D(P, R) yields that
1
|F  (P)| ≤ .
R1
Letting R tend to +∞ yields that F  (P) = 0. Since P was arbitrary,
we see that F  ≡ 0. A simple calculus exercise now shows that F must
be constant.

One of the most dramatic applications of Liouville’s theorem is in

the proof of the fundamental theorem of algebra. That is our next task:

Theorem 7. Let p(z) = a0 +a1 z +a2 z 2 +· · ·+ak z k be a non-constant

polynomial. Then there is a point z at which p vanishes.

Proof. Suppose not. Then F(z) = 1/ p(z) is an entire function. Since a

non-constant polynomial blows up at infinity, F must be bounded. By
Liouville’s theorem, F is a constant. Hence p is constant, and therefore
has degree zero. This contradiction completes the proof.

Let k be the degree of the polynomial p. Notice that if the poly-

nomial p vanishes at the point r1 , then the Euclidean algorithm implies
that p is divisible by (z − r1 ): that is to say, p(z) = (z − r1 ) · p1 (z)
for a polynomial p1 of degree k − 1. If k − 1 ≥ 1, then we may apply
the preceding result to p1 . Continuing in this fashion, we obtain that p
may be expressed as a product of linear factors:

p(z) = (z − r1 ) · (z − r2 ) · · · (z − rk ).
10 Principal Ideas of Classical Function Theory

We conclude this brief overview of elementary complex analysis

by recalling the argument principle and Hurwitz’s theorem.

Theorem 8. (The Argument Principle). Let F be holomorphic on

a domain U and let γ be a topologically trivial, positively oriented,
simple closed curve in U . Assume that F does not vanish on γ . We can
be sure, by Theorem 4, that there are at most finitely many, say k, zeros
of F inside γ (counting multiplicity). Then we have that

1 F  (ζ )
k= dζ.
2πi γ F(ζ )

Sketch of Proof. By an easy reduction, it is enough to prove the result

when k = 1. A second reduction allows us to consider the case when
γ is a positively oriented circle. After a change of coordinates, let us
suppose that F has a simple zero at the point P = 0 inside γ . By
writing out the power series expansions for F and for F  , we find that

F  (ζ ) 1
= + h(ζ ),
F(ζ ) ζ
where h is holomorphic near 0. Of course, h integrates to 0, by the
Cauchy integral theorem. And it is easily calculated that

1 1
dζ = 1,
2πi γ ζ

completing the proof.

Let U be a domain and {F j } a sequence of holomorphic functions

on U that converges, uniformly on compact sets, to a limit function F.
It is an easy consequence of the Cauchy integral formula that the limit
function is also holomorphic. Let us now use the argument principle to
see how the zeros of F are related to the zeros of the F j ’s.

Theorem 9. (Hurwitz’s Theorem). With {F j } and F as above, if the

F j ’s are all zero-free, then either F is zero-free or F is identically zero.
The Maximum Principle, the Schwarz Lemma, and Applications 11

Proof. Assume that F is not identically 0. Then the zero set of F is

discrete, so we can find a simple, closed, topologically trivial curve γ
in U that misses the zero set completely. Therefore
 F  (ζ ) 
1 j 1 F  (ζ )
dζ −→ dζ
2πi γ F j (ζ ) 2πi γ F(ζ )
as j → ∞. But the expression on the left is zero for each j. Hence so
is the expression on the right. So F has no zeros inside the curve γ .
Since γ was chosen to be an “arbitrary” curve that misses the zeros of
F, we find that F is zero-free.

2. The Maximum Principle, the Schwarz

Lemma, and Applications
We begin with a brief treatment of the maximum principle. Let F be
holomorphic on an open set U that contains the closed disc D(P, r ).
Then the Cauchy integral formula says that

1 F(ζ )
F(P) = dζ.
2πi ∂ D(P,r ) ζ − P

Parametrizing the boundary of the disc by γ (t) = P + r eit and writing

out the definition of the line integral gives
 2π
1
F(P) = F(P + r eit ) dt.
2π 0
This is the mean value property for a holomorphic function. Now we
have:

Theorem 1. Let F be holomorphic on a domain U . If there is a point

P ∈ U such that

|F(P)| ≥ |F(z)|

for all z ∈ U , then F is a constant function.

12 Principal Ideas of Classical Function Theory

Sketch of Proof. Multiplying F by a constant, we may assume that

M ≡ F(P) is real and nonnegative. Let S = {z ∈ U : F(z) = F(P)}.
Observe that S is nonempty since P ∈ S. Moreover, since F is a con-
tinuous function, S is trivially closed in U . To see that S is open, let
w ∈ S and suppose that D(w, r ) ⊆ U . Now, for 0 < r  < r , we see
that

1  2π
 it
M = F(w) = |F(w)| = F(w + r e ) dt
2π 0
 2π
1
≤ F(w + r  eit ) dt ≤ M.
2π 0
Since the expressions on the extreme left and right are
equal, all the
inequalities must be equalities. Thus, F(w +r e ) = F(w +r e ) =
 it it

M for all values of t and all 0 < r  < r . Therefore


D(w, r ) = w + r  eit : 0 ≤ t ≤ 2π, 0 < r  < r ⊆ S.

This shows that S is open.

Since S is nonempty, closed, and open, and since U is connected,
it follows that S = U . That is what we are required to prove.

For us, the main application of the maximum principle will be to the
classical Schwarz lemma.

Theorem 2. (The Schwarz Lemma). If F : D → D is holomorphic

and if F(0) = 0, then

F(z) ≤ |z| and F (0) ≤ 1.

If either F(z) = |z| for some z = 0 or F  (0) = 1, then F is a
rotation: F(z) ≡ eiτ · z for some real number τ ∈ R.

Proof. The function



 F(z) , z = 0
G(z) = z

 F  (z), z=0
The Maximum Principle, the Schwarz Lemma, and Applications 13

is holomorphic on D. Applying the maximum principle to G on the

disc {z : |z| ≤ 1 − ε} for ε > 0 gives

G(z) ≤ (1 − ε)−1

when |z| ≤ 1 − ε. Letting ε → 0+ yields G(z) ≤ 1 on D, which is
equivalent to the desired conclusion.
For uniqueness, notice that |F(z)| = |z| for some z = 0 implies
that |G(z)| = 1. The maximum principle then forces G to be a uni-
modular constant, hence F is a rotation. The uniqueness statement for
the derivative is obtained similarly.

As a simple application of the Schwarz lemma, we classify the

conformal self-maps (that is, the one-to-one, onto holomorphic maps)
of the unit disc to itself. We begin by claiming that if a is a complex
number of modulus less than 1, then the Möbius transformation

ζ −a
φa (ζ ) =
1 − aζ

maps D(0, 1) to D(0, 1). [We shall use the phrase “Möbius transfor-
mation” specifically to mean a mapping of this form. Later on, we shall
consider compositions of rotations with Möbius transformations.] To
see this, notice that φa is certainly well defined and holomorphic in a
neighborhood of D(0, 1); moreover, φa (a) = 0. If we prove that φa
maps ∂ D(0, 1) to ∂ D(0, 1), then our claim will follow from the maxi-
mum principle. But for |ζ | = 1 we have that

ζ − a 1 ζ − a

|φa (ζ )| = = ·
1 − aζ ζ̄ 1 − aζ

ζ − a
= = 1,
ζ̄ − a

since ζ · ζ = 1. This proves that φa : D(0, 1) → D(0, 1). It is straight-

forward to check that (φa )−1 = φ−a , proving that φa is one-to-one and
onto.
14 Principal Ideas of Classical Function Theory

In addition to the functions φa we also have the rotations ρτ (ζ ) ≡

eiτ · ζ , τ ∈ R, which map the disc in a one-to-one fashion onto the
disc. The remarkable fact is that these two types of maps completely
characterize all conformal self-maps of the disc.

Theorem 3. Let F be a conformal self-map of the unit disc. Then there

exists a complex number a of modulus less than 1 and a real number τ
such that

F(ζ ) = φa ◦ ρτ (ζ ).

Proof. Let F(0) = b and consider the function G ≡ φb ◦ F. Then G

is a conformal self-map of the unit disc and G(0) = 0. The Schwarz
lemma tells us that |G  (0)| ≤ 1. But we
may also apply the Schwarz
lemma to G −1 to obtain that 1/G  (0) = (G −1 ) (0) ≤ 1. We con-
clude that |G  (0)| = 1. The uniqueness part of the Schwarz lemma then
says that G(ζ ) = ρτ (ζ ) for some τ . But this is equivalent to

F = φ−b ◦ ρτ .

Setting a = −b completes the proof.

Exercises. Give a similar proof to show that every conformal self-map

of the disc can be written in the form ρτ ◦ φa .
Verify directly that the composition of two functions of the form
φa ◦ ρτ (resp. ρτ ◦ φa ) is also a function of that form.
It is worth noting here that the Möbius transformations φa are in-
stances of linear fractional transformations. In general, a linear frac-
tional transformation has the form
az + b
z −→
cz + d
(with ad − bc = 0). It is elementary to see that any linear fractional
transformation is a composition of translations, dilations, and inversion
(z → 1/z). As a result, a linear fractional transformation will take
circles and lines to circles and lines.
The Maximum Principle, the Schwarz Lemma, and Applications 15

Our next step is to generalize the Schwarz lemma by removing the

requirement that F(0) = 0. We have:

Theorem 4. (Schwarz, Pick). If F : D → D is holomorphic,

F(z 1 ) = w1 and F(z 2 ) = w2 , then

w1 − w2 z 1 − z 2
≤
1 − w w 1 − z z
1 2 1 2

and
1 − |w1 |2
|F  (z 1 )| ≤ .
1 − |z 1 |2
If equality obtains in the first expression for some z 1 = z 2 or if equality
obtains in the second expression, then F must be a conformal self-map
of the disc.

Proof. Define
z + z1 z − w1
φ(z) = , ψ(z) = .
1 + z1 z 1 − w1 z
Then ψ ◦ F ◦ φ satisfies the hypotheses of Schwarz’s lemma. Therefore

(ψ ◦ F ◦ φ)(z) ≤ |z|, ∀z ∈ D.

Setting z = φ −1 (z 2 ) now gives the first inequality. Also the Schwarz

lemma says that

(ψ ◦ F ◦ φ) (0) ≤ 1.

Using the chain rule to write this out gives the second inequality.
The case of equality is analyzed as in Theorem 2.
16 Principal Ideas of Classical Function Theory

3. Normal Families and the Riemann

Mapping Theorem
One of the most important concepts in topology is compactness. Com-
pactness for a set of points in Euclidean space is, thanks to the Heine–
Borel theorem, easy to understand: a set is compact if and only if it
is closed and bounded. In modern times, compactness for families of
functions has proved to be a powerful tool. The notion of normal fam-
ilies, and in particular of Montel’s theorem, is historically one of the
first instances of this concept. In the present section we shall treat both
Montel’s theorem and its application to the Riemann mapping theorem.
In an effort to keep our exposition simple, we treat here a restric-
tive definition of normal family; later in the book (Section 2.3), normal
families will be given a more thorough treatment from a completely
different point of view.

Definition 1. A family F of functions on a domain U will be called

normal if every sequence in F has a subsequence which converges
uniformly on compact subsets of U .

Now Montel’s theorem says the following:

Theorem 2. (Montel). Let U be a domain in the complex plane. Let

F be a family of holomorphic functions on U . If there is a positive
constant M such that

|F(z)| ≤ M

for all z ∈ U and all F ∈ F, then F is a normal family.

For the proof of this theorem, we shall need an important result from
real analysis. We begin with a little terminology.

Definition 3. Let F be a family of functions on a common domain

S ⊆ Rn . We say that the family is equicontinuous if for each ε > 0
Normal Families and the Riemann Mapping Theorem 17

there is a δ > 0 such that whenever z, w ∈ S satisfy |z − w| < δ, then

|F(z) − F(w)| < ε

for all F ∈ F.
Notice that the property of equicontinuity is stronger than uniform
continuity: not only is the choice of δ independent of the points z and w
in S (i.e., it depends only on ε), but it is independent of which F from
the family F we are considering. Next we define a companion notion.

Definition 4. Let F be a family of functions on a common domain

S ⊆ Rn . We say that the family is equibounded if there is a number
M > 0 such that, whenever z ∈ S and F ∈ F, then

|F(z)| ≤ M.

We use the notions of equicontinuity and equiboundedness to for-

mulate the following essential result.

Proposition 5. (Ascoli/Arzelà). Let K be a compact set in Rn . Let

F be an equicontinuous family of functions on K which is also equi-
bounded. Then every sequence { f j } in F contains a subsequence { f jk }
which converges uniformly on K .

We shall not prove the Ascoli/Arzelà theorem here. Instead we

refer the interested reader to [RU1] for details. Note, however, the thrust
of the theorem: if a family of rather nice objects is bounded, then it has
a convergent subsequence. In other words, the theorem is a statement
about sequential compactness.
Proposition 5, as stated, is about scalar-valued functions. Later on
we shall have occasion to use the Ascoli/Arzelà theorem for metric-
space-valued functions. The proof is just the same.
Now we apply the Ascoli/Arzelà theorem to prove Montel’s theo-
rem:

Proof of Theorem 2. Let D(P, R) ⊆ U . Since U is open, the comple-

ment c U of U is closed. Since D(P, R) and c U are disjoint, there is
18 Principal Ideas of Classical Function Theory

a positive distance between the two sets: that is, there exists a k > 0
such that if z ∈ D(P, R) and u ∈ c U , then |z − u| > k. Thus, for any
z ∈ D(P, R) and any F ∈ F we may apply the Cauchy estimates to F
on the disc D(z, k). We conclude that
M · 1!
|F  (z)| ≤ ≡ C.
k1
It is now an easy exercise, using the fact that
 z
F(z) − F(w) = F  (ξ ) dξ,
w

to see that

F(z) − F(w) ≤ C|z − w|

for any z, w ∈ D(P, R). This estimate shows that the family F is
equicontinuous on D(P, R): given ε > 0, choose δ = ε/C.
Finally, if K is any compact subset of U , then K can be covered
by finitely many discs D(P, R) as above, and the equicontinuity on K
follows from the triangle inequality.
We may now apply the Ascoli/Arzelà theorem to F on K to
find a uniformly convergent subsequence {F jk } of any given sequence
{F j } ⊆ F. A standard diagonalization argument then yields a subse-
quence {F jkl } which converges uniformly on all compact subsets of U .

Notice that in real analysis there is no analogue for Montel’s theo-

rem. The family F = {sin j x}∞ j=1 is equibounded, as in the hypothesis
of Montel’s theorem, but there is no convergent subsequence. It is the
property of being holomorphic (and hence the resulting control on first
derivatives) that gives the required extra structure.
Now we turn our attention to the Riemann mapping theorem. Let
U and V be open sets in the complex plane. We say that U and V
are conformally equivalent if there is a one-to-one, onto holomorphic
function (i.e., a conformal map)

σ : U → V.
Normal Families and the Riemann Mapping Theorem 19

The motivation for this concept is clear: any holomorphic function F on

V gives rise to a holomorphic function F ◦ σ on U and any holomorphic
function G on U gives rise to a holomorphic function G ◦ σ −1 on V .
Thus complex analysis on the two domains is, in effect, equivalent.
How can we tell when two domains are conformally equivalent?
An obvious necessary condition is that they be topologically equiva-
lent, for any conformal mapping is certainly a homeomorphism. Rie-
mann’s astonishing theorem is that if a domain U is topologically
equivalent to the unit disc, and is not the entire plane, then it is con-
formally equivalent to the unit disc.

Theorem 6. (The Riemann Mapping Theorem). Let U be a proper

subdomain of C that is homeomorphic to the unit disc. Then U is con-
formally equivalent to the open unit disc D(0, 1).

The proof of this theorem is quite elaborate; we shall outline the

principal steps. To keep matters as simple as possible, we shall assume
that U is bounded.

S TEP 1. Fix a point P ∈ U . Let F = {σ : U → D(0, 1) such that

σ is one-to-one, holomorphic, and σ (P) = 0}. We claim that F is
nonempty. In fact, since U is bounded, there is an R > 0 such that
U ⊆ D(0, R). Then the function

1
σ : ζ → (ζ − P)
2R
certainly maps P to 0, is holomorphic, and satisfies

σ (ζ ) < 1 (R + R) = 1.
2R
Obviously σ is one-to-one. So σ ∈ F and F is nonempty.
S TEP 2. The family F is a normal family. In fact, the elements of F
are holomorphic, and they are all bounded by 1. Montel’s theorem
thus applies, and the family is normal.
20 Principal Ideas of Classical Function Theory

 
S TEP 3. Let us define M = sup |σ  (P) : σ ∈ F . We claim that M
is finite. In fact, let D(P, r ) be a closed disc about P that is con-
tained in U . Then the Cauchy estimates apply: for any σ ∈ F, we
have that


σ (P) ≤ 1! = 1 .
r r

Hence M does not exceed 1/r.

S TEP 4. We claim that there is a function σ0 ∈ F such that σ0 (P) =
M. This is where normal families come in. The present step is of
historical interest because Riemann asserted it without proof. Thus
his theorem was doubted for several years until the step was cor-
rected using the Dirichlet principle. Nowadays normal families are
more commonly used to verify the assertion. The point is that it is
fallacious to formulate an extremal problem and simply to assume
that it has a solution.
To see how the argument works, we use the definition of
“supremum”
 to conclude that there is a sequence {σ j } in F such that
σ (P) → M. But since, by Step 2, F is a normal family, we know
j
that there is a subsequence {σ jk } that converges on compact subsets
to a limit function σ0 . Now we certainly know that {|σ jk (P)|} con-
verges to M (Corollary 5.1 of Section 1). Hence |σ0 (P)| = M. Mul-
tiplying σ0 by a unimodular constant now gives the function which
we seek.
S TEP 5. We claim that σ0 is one-to-one on U . In fact, this follows from
a clever application of the argument principle. Fix distinct points
Q and R in U . Let 0 < s < |Q − R|. Consider the functions
ψk (z) ≡ σ jk (z) − σ jk (Q) on D(R, s). Since the σ j are all one-to-
one, the functions ψk are all nonvanishing on D(R, s). By Hurwitz’s
theorem the limit function σ0 (z) − σ0 (Q) is either identically zero
or is nonvanishing on D(R, s). It certainly is not identically zero,
for we verified in Step 4 that σ0 (P) = M > 0. Thus σ0 (z) = σ0 (Q)
for all z ∈ D(R, s), in particular σ0 (R) = σ0 (Q).
Since Q and R were arbitrary, we find that σ0 is one-to-one.
Normal Families and the Riemann Mapping Theorem 21

S TEP 6. We claim that σ0 maps U onto the unit disc. It is in this step
that the topology of U comes in. One proves, by an advanced cal-
culus argument, that if F is any nonvanishing holomorphic function
on U , then log F is a well defined holomorphic function on U . In-
deed, one can define this logarithm by the formula

F  (ζ )
log F(z) = dζ,
γz F(ζ )

where γz is any piecewise C 1 path connecting P to z. The topo-

logical hypothesis on U guarantees that this definition of log F is
independent of the path.

Knowing that log F exists allows us to see immediately that any

power of F also exists: indeed, if α ∈ C, then we define

F α (z) = exp (α log F(z)) .

Now, seeking a contradiction, assume that σ0 is not onto D(0, 1).

Say that a point β ∈ D(0, 1) is omitted from the image. We now com-
pose with certain Möbius transformations of the disc in order to nor-
malize our mapping.
Recall the Möbius transformations φa of Section 2. Our hypothe-
sis implies that φβ ◦ σ0 (ζ ) never vanishes on U . Thus the function
 1/2
µ(ζ ) = φβ ◦ σ0 (ζ )

is a well defined holomorphic function on U . Finally, we let τ = µ(P)

and set
|µ (P)|
ν(ζ ) = · (φτ ◦ µ(ζ )) .
µ (P)

Then ν ∈ F. By definition, ν(P) = 0 and a calculation shows that

1 + |β|
|ν  (P)| = · M > M.
2|β|1/2
22 Principal Ideas of Classical Function Theory

This contradicts the choice of σ0 as having maximal derivative at P.

Thus we conclude that σ0 is onto.
The function σ0 is the required conformal equivalence of U with
D(0, 1).
We close this section by recalling Carathéodory’s celebrated the-
orem on the boundary behavior of conformal mappings.

Theorem 7. (Carathéodory’s theorem). Let  ⊆ C be a simply con-

nected domain whose boundary consists of a single Jordan curve. Let
φ :  → D be a conformal mapping of  to D (as provided by the
Riemann mapping theorem). Then φ extends to a homeomorphism of 
to D.

It is interesting that Painlevé’s theorem about the smooth continu-

ation of conformal mappings to the boundary (which we treat in some
detail in Chapter 4) actually preceded Carathéodory’s theorem by sev-
eral years. This historical accident actually impeded the appreciation
of Painlevé’s result by the ranking experts of the time.

4. Isolated Singularities and the Theorems

of Picard
We begin with a brief discussion of the types of singularities which
a holomorphic function can have. Let F be holomorphic on the set
U = D(P, r ) \ {P}. Then there are three possible ways that F can
behave near P:

(i) F is bounded near P;

(ii) F is unbounded near P, but there exists a k > 0 such that (z −
P)k F(z) is bounded near P;
(iii) neither (i) nor (ii) holds.

It is plain that (i), (ii), (iii) are mutually exclusive and cover all possi-
bilities.
Isolated Singularities and the Theorems of Picard 23

In the first instance, the Riemann removable singularities theorem

guarantees that α ≡ limz→P F(z) exists and that the function

F(z), z ∈ U
F(z) =
α, z=P

is holomorphic on D(P, r ).
In the second case, we see that

lim |F(z)| = +∞.

z→P

The function F is said to have a pole at P. If k is the least positive

integer which satisfies (ii), then it turns out that F can be written in the
form

F(z) = (z − P)−k h(z)

with h holomorphic on D(P, r ) and nonvanishing near P.

The third case is the most complex, and is the one that we wish to
concentrate on here. In this circumstance F is said to have an essential
singularity at P. We begin by proving the Casorati–Weierstrass theo-
rem, which only begins to suggest the level of complexity involved.

Theorem 1. (Casorati–Weierstrass). Let F be holomorphic on U =

D(P, r ) \ {P} and suppose that F has an essential singularity at P.
Then, for any 0 < s < r , the set {F(z) : 0 < |z − P| < s} is dense in
the complex plane.

The theorem says that holomorphic functions have a very special

mode of behavior near an isolated singularity: either the function has a
finite limit, or an infinite limit, or else it takes values arbitrarily close
to all complex values.

Proof of Theorem 1. Suppose that the conclusion is false. Then there

is a complex number β and numbers s > 0, δ > 0 such that

F(z) − β > δ, for all z ∈ D(P, s) \ {P}.
24 Principal Ideas of Classical Function Theory

Define
1
h(z) = .
F(z) − β
Then |h(z)| is bounded near P by the constant 1/δ. By the Riemann re-
movable singularities theorem, h can be continued analytically through
the point P. If h(P) = 0, then we have a contradiction since then F
could be continued analytically through P. But h(P) = 0 would imply
that F has a pole at P, and that is also false. The resulting contradiction
completes the proof.

It is sometimes useful to think about the point at ∞ just like any point
of the finite complex plane. The map
1
z →
z
allows us to pass back and forth between neighborhoods of 0 and
neighborhoods of ∞. In particular, a function F holomorphic on a set
{z ∈ C : |z| > R} is said to have a removable singularity, a pole, or an
essential singularity at ∞ if the function F (1/z) has, respectively, a
removable singularity, a pole, or an essential singularity at the origin.
We now record some striking properties about singularities at infinity.

Proposition 2. If F is entire (holomorphic on all of C) and F has a

removable singularity at infinity, then F is constant.

Sketch of Proof. By examining F (1/z) , we see that F must have a

finite limit at ∞. Thus F is bounded. By Liouville’s theorem, F is
constant.

Proposition 3. If F is entire and F has a pole at ∞, then F is a poly-

nomial.

Sketch of Proof. Looking at F(1/z), we find for some k > 0 that

z k F(1/z) is bounded near 0. But then z −k F(z) is bounded near ∞.
Isolated Singularities and the Theorems of Picard 25

Let Pk be the kth degree Taylor polynomial of F about zero. Then the
function
F(z) − Pk (z)
h(z) =
zk
is holomorphic on the entire plane and bounded at ∞. Thus, by Liou-
ville, h is constant. It follows that F is a polynomial.

We learn from these two propositions that if F is entire, then F

is either a polynomial or F has an essential singularity at ∞. Since
the whole point of holomorphic function theory is to study a class of
functions which generalizes the properties of polynomials, it is clearly
important to understand essential singularities.
Picard’s theorems give very sharp results on the value distribution
of entire functions and, more generally, holomorphic functions near an
essential singularity. We now state them.

Theorem 4. (Picard’s little theorem). If F is a non-constant entire

function, then the image of F contains all complex numbers with the
possible exception of one value. In other words, if an entire function
omits two values, then it is constant.

Theorem 5. (Picard’s great theorem). Let F be holomorphic on an

open set U = D(P, r ) \ {P}. If F has an essential singularity at P,
then for any 0 < s < r we have {F(z) : 0 < |z − P| < s} contains all
complex numbers except possibly one value.

It should now be clear that the great theorem is a generalization of

the little theorem, since an entire function that is not a polynomial has
an essential singularity at infinity. Let us discuss some examples to put
the results in perspective.
If F(z) is a non-constant polynomial and α is any complex num-
ber, then the equation F(z) = α always has a solution, by the fun-
damental theorem of algebra. Thus F takes all complex values. If
26 Principal Ideas of Classical Function Theory

F(z) = e z , then F takes all complex values except 0. The little the-
orem allows non-constant entire functions to omit one complex value.
It does not allow omission of two.
Both of Picard’s theorems are quite difficult to prove by classi-
cal techniques. The original proof involved the construction of a spe-
cial holomorphic function from the upper half plane to C \ {0, 1}. This
function, known as the elliptic modular function, is a powerful tool in
function theory. Here we briefly discuss the construction of the modular
function and its application to the proof of the little theorem.
Begin with a conformal mapping h of the right half of the re-
gion , shown in Figure 1, to the upper half plane, normalized so that
h(0) = 0, h(1) = 1, h(∞) = ∞. By Schwarz reflection [GRK], the
map h extends to a map from the entire region  to C \ {0, 1}. By
more delicate reflection arguments, it is possible to continue the func-
tion analytically to a map from the entire upper half plane to C \ {0, 1}.
Denote the extended function by the symbol λ(z). Then λ is the elliptic
modular function.

Figure 1.
Isolated Singularities and the Theorems of Picard 27

Now if F is an entire function whose range omits two values, we

may arrange by composition with a linear transformation for those two
values to be 0 and 1. Then a function element of λ−1 ◦ F, initially
defined on a small disc, may be analytically continued to all of C. Thus
λ−1 ◦ F is an entire function taking values in the upper half plane.
Finally, the holomorphic function
i −z
ρ(z) =
i +z
maps the upper half plane to the unit disc. In summary, the composition
ρ ◦ λ−1 ◦ F is an entire function which takes values in the unit disc; in
other words, it is bounded. By Liouville’s theorem, the composition is
constant. Unravelling the composition, we find that F is constant. This
completes our sketch of the proof of Picard’s little theorem.
The construction of the elliptic modular function is both techni-
cal and complicated. The analytic continuation argument which is re-
quired to extend an analytic function element of λ−1 ◦ F to all of C
requires a great deal of theory. Other proofs of Picard’s theorem arise
from Nevanlinna theory, or the value distribution theory of entire func-
tions. One of the benefits of our geometric study of complex function
theory will be a rather simple and natural proof of both the little and
the great Picard theorems (Section 2.4).
CHAPTER 1
Basic Notions of Differential Geometry

0. Introductory Remarks
Differential geometry has developed into one of the most powerful
mathematical tools of modern mathematics. It has become an integral
part of the theories of differential equations, harmonic analysis, and
complex analysis, to name just a few examples.
In spite of their importance, the techniques of geometry have not
proliferated as much as they might have because of the complexity of
the language. The characterization of differential geometry as “that
portion of mathematics which is invariant under change of notation”
is unfortunately rather accurate.
The best way to learn new mathematics is in the context of what
is already familiar. In the milieu of one complex variable, the notions
of Riemannian metric, of geodesic, and of curvature become rather
simple. The standard geometric device of tensors and bundles is not
necessary. The result is that one can learn the flavor and some of the
methodology of differential geometry without being encumbered by its
notation and machinery.
In this chapter, therefore, we shall learn some of the basic ideas
of differential geometry, but only in the complex plane; we shall also
learn a few simple applications. In later chapters we shall learn about
curvature and more advanced applications.

29
30 Basic Notions of Differential Geometry

1. Riemannian Metrics and the Concept

of Length
In classical analysis a metric is a device for measuring distance. If X is
a set then a metric λ for X is a function

λ : X × X −→ R

satisfying, for all x, y, z ∈ X ,

1. λ(x, y) = λ(y, x)
2. λ(x, y) ≥ 0 and λ(x, y) = 0 iff x = y;
3. λ(x, y) ≤ λ(x, z) + γ (z, y).

The trouble with a metric defined in this generality is that it does not
interact well with calculus. What sort of interaction might we wish to
see?
Given two points P, Q ∈ X , one would like to consider the curve
of least length connecting P to Q. Any reasonable construction of such
a curve leads to a differential equation, and thus we require that our
metric lend itself to differentiation. Yet another consideration is curva-
ture: in the classical setting curvature is measured by the rate of change
of the normal vector field. The concepts of normal and rate of change
lead inexorably to differentiation. Thus we will now take a different
approach to the concept of “metric.”
Recall that in multivariable calculus we define the arc length of a
continuously differentiable curve γ : [a, b] → R2 by the formula

b
(γ ) = |γ̇ (t)| dt. (∗)
a

In point of fact, most calculus books give a heuristic which leads from
a reasonable notion of length to this formula. But, rigorously speaking,
(∗) must be considered to be the definition of arc length.
Notice that the definition (∗) hinges on the assumption that we
have previously defined |γ̇ (t)|, the length of the tangent vector γ̇ (t) to
Riemannian Metrics and the Concept of Length 31

γ at t. Riemann’s insight into differential geometry was that we can

define a new metric by specifying a new way to measure the lengths of
tangent vectors. The most interesting examples arise when the method
of measuring the length of a tangent vector varies from point to point.
Thus we now give the word “metric” a new meaning; of course this
new meaning will turn out to be intimately related to the older, more
classical one.

Definition 1. If  ⊆ C is a domain, then a metric on  is a continuous

function ρ(z) ≥ 0 in  that is twice continuously differentiable on
{z ∈  : ρ(z) > 0}. If z ∈  and ξ ∈ C is a vector then we define the
length of ξ at z to be

 ξ ρ,z ≡ ρ(z) · |ξ |,

where |ξ | denotes the Euclidean length of the vector ξ .

Remark. Later in the book, we will consider metrics which are only
continuous (or even less smooth). We want a smooth metric now be-
cause we will differentiate it to calculate curvature.

Remark. Usually our metrics ρ will be strictly positive, but it will

occasionally be convenient for us to allow a metric to have isolated
zeros. These will arise as zeros of holomorphic functions. The zeros of
ρ should be thought of as singular points of the metric.

For the record, the metrics we are considering here are a special
case of the type of differential metric called Hermitian. This terminol-
ogy need not concern us here. Classical analysts sometimes call these
metrics conformal metrics and write them in the form ρ(z)|dz|.
Technically speaking, our metric lives on the tangent bundle to the
domain . That is to say, the metric is a function of the variable (z, v),
where v is thought of as a tangent vector at the point z. This is just a
mathematical way of saying that the length of the vector v depends on
the point z at which it is positioned.
32 Basic Notions of Differential Geometry

Later, in Sections 4.4 and 4.5, we shall see a discussion of a differ-

ent type of metric—a Riemannian metric (or Kähler metric). This will
be part of our study of the Bergman kernel and metric. In that context
it will be even more convenient to think of the metric as living on the
tangent bundle. For most of this book, we may forego this formality.

Example 1. Let  be any domain in C. Define ρ(z) = 1 for all z ∈ .

Notice that this particular metric yields that if z ∈  and ξ ∈ C then

 ξ ρ,z ≡ ρ(z) · |ξ | = |ξ |.

In short, this choice of metric gives the standard Euclidean notion of

vector length—and this notion is independent of the base point z. This
metric is usually called the Euclidean metric.

Example 2. Let  = {z ∈ C : |z| < 1} ≡ D, the unit disc. Let

1
ρ(z) =
1 − |z|2

This is the Poincaré metric, which has been used to gain deep insights
into complex analysis on the disc. It will receive our detailed attention
later on in this book. For now we do some elementary calculations with
the Poincaré metric.
For any ξ ∈ C we may calculate that

4
 ξ ρ,(1/2+0i) = ρ(1/2 + 0i) · |ξ | = · |ξ |;
3
 ξ ρ,0 = ρ(0) · |ξ | = 1 · |ξ | = |ξ |;
1
 ξ ρ,(0+.9i) = ρ(.9i) · |ξ | = |ξ | = (5.2631578 . . . ) · |ξ |.
.19

The notion of the length of a vector varying with the base point is in
contradistinction to what we learn in calculus. In calculus, a vector has
direction and magnitude but not position. Now we declare that a vector
Riemannian Metrics and the Concept of Length 33

has position and the way that its magnitude is calculated depends on
that position. The next example shows how vectors with position arise
in practice.

Example 3. Let η(t) = t, 0 ≤ t ≤ 1/2 and µ(t) = 1/2 + it, 0 ≤ t ≤

1/2. Let us calculate

 η̇(t) ρ,η(t) and  µ̇(t) ρ,µ(t)

for the Poincaré metric

1
ρ(z) = .
1 − |z|2

We have η̇(t) = 1 for all t. Notice that the Euclidean length of

η̇(t) is 1. However,

1
 η̇(t) ρ,η(t) = ρ(η(t)) · |η̇(t)| = .
1 − t2
We also have |µ̇(t)| = 1 for all t and

1
 µ̇(t) ρ,µ(t) = ρ(µ(t)) · |µ̇(t)| = .
3/4 − t 2

Tangent vectors will arise for us most often as tangents to curves. It

makes sense to think of the tangent vector η̇(t) as living at η(t) and
µ̇(t) as living at µ(t) for arbitrary curves η, µ. See Figure 1.

(t)

(t)

Figure 1.
34 Basic Notions of Differential Geometry

Definition 2. Let  ⊆ C be a domain and ρ a metric on . If

γ : [a, b] → 

is a continuously differentiable curve then we define its length in the

metric ρ to be

b
ρ (γ ) =  γ̇ (t) ρ,γ (t) dt.
a

The length of a piecewise continuously differentiable curve is defined

to be the sum of the lengths of its continuously differentiable pieces.

Notice that this definition is in complete analogy with the classical

Euclidean (calculus) definition of arc length. In that more familiar set-
ting, we use the Euclidean definition of vector length, and that in turn is
the foundation for our idea of arc length. Now we have a more general
means to determine the length of a vector. As a result, our integrand
contains this more general notion of vector magnitude. This important
idea of Riemann has led to a complete rethinking of what a geometry
should be. In particular, it has led to a powerful marriage of geometric
ideas with calculus ideas.
Observe also that it follows from the change of variables formula
that the length of a curve is independent of its parametrization. This
assertion generalizes a familiar fact from calculus. In practice, we often
find it convenient to suppose that a curve being studied is parametrized
with respect to arc length. This means that one unit of the parameter
corresponds to one unit of length on the curve.
Classical sources write the arc length as


ρ (γ ) = ρ(z)|dz|,
γ

but we shall not use this notation.

Example 4. Let D ⊆ C be the unit disc and let ρ(z) = 1/(1 − |z|2 ),
the Poincaré metric on D. Fix  > 0. Let us calculate the length of the
curve γ (t) = t, 0 ≤ t ≤ 1 − . Now
Riemannian Metrics and the Concept of Length 35

1−
ρ (γ ) =  γ̇ (t) ρ,γ (t) dt
0

1− |γ̇ (t)|
= dt
0 1 − |γ (t)|2

1−
1
= dt
0 1 − t2
 
1 2−
= log .
2 

Observe that, when  is small, ρ (γ ) is large. In fact

lim ρ (γ ) = +∞.
→0+

This suggests that the boundary ∂ D is infinitely far from the origin, at
least along this particular path γ , in this the Poincaré metric.

Remark. Notice that if we consider the metric ρ(z) ≡ 1, as in Ex-

ample 1, then the length of a curve turns out to be the ordinary Eu-
clidean notion of length.

In a rigorous course on Riemannian geometry one proves that, for

a reasonable metric (the metric needs to be complete), curves of least
length always exist (the proper language for formulating this fact is that
of geodesics, curves that have least length in an infinitesimal sense).
We shall say more about completeness in Proposition 4 of Section 1.4
below. In the present concrete situation, we may bypass abstractions
like this and still come up with useful information:

Example 5. Equip the disc D with the Poincaré metric ρ(z) = 1/(1 −
|z|2 ). Fix  > 0. Let us prove that, among all continuously differen-
tiable curves of the form

µ(t) = t + iw(t), 0 ≤ t ≤ 1 − ,
36 Basic Notions of Differential Geometry

that satisfy µ(0) = 0 and µ(1 − ) = 1 −  + 0i, the one of least length
is γ (t) = t. Here w(t) is a continuously differentiable, real-valued
function.
In fact for any such candidate µ we have

1−
ρ (µ) =  µ̇(t) ρ,µ(t) dt
0

1− 1
= |µ̇(t)| dt
0 1 − |µ(t)|2

1− 1
= · (1 + [w (t)]2 )1/2 dt.
0 1 − t 2 − [w(t)]2
However,
1 1
≥ and (1 + [w (t)]2 )1/2 ≥ 1.
1 − t2 − [w(t)] 2 1 − t2
We conclude that

1− 1
ρ (µ) ≥ dt = ρ (γ ).
0 1 − t2
This is the desired result.
Notice that, with only small modifications, this argument can also
be applied to piecewise continuously differentiable curves t + iw(t).

In fact if a piecewise continuously differentiable curve connecting

the point 0 ∈ D to (1 − ) + 0i ∈ D is not of the form

µ(t) = t + iw(t), (∗)

then it may cross itself. Of course we can eliminate the loops and
thereby create a shorter curve. If the resulting curve is still not the graph
of a function, then elementary comparisons show that it will be longer
than a curve of the form (∗) (see Figure 2). We may conclude that the
curve γ in the example is the shortest of all curves connecting 0 to
(1 − ) + 0i.
Riemannian Metrics and the Concept of Length 37

(1 - ) + 0i
0

Figure 2.

The preceding discussion suggests that if a metric ρ is given on a

planar domain , and if P, Q are elements of , then the distance in the
metric ρ from P to Q should be defined as follows: Define C (P, Q)
to be the collection of all piecewise continuously differentiable curves
γ : [0, 1] →  such that γ (0) = P and γ (1) = Q. Now define the
ρ-metric distance from P to Q to be

 
dρ (P, Q) = inf ρ (γ ) : γ ∈ C (P, Q) .

Check for yourself that the resulting notion of distance satisfies the
classical metric axioms listed at the outset of this section.
There is some subtlety connected with defining distance in this
fashion. If ρ(z) ≡ 1, the Euclidean metric, and if  is the entire plane,
then dρ (P, Q) is the ordinary Euclidean distance from P to Q. But
if  and P and Q are as shown in Figure 3, then there is no shortest
curve connecting P to Q. The distance from P to Q is suggested by
the dotted curve, but notice that this curve does not lie in . The crucial
issue here is whether the domain is complete in the metric, and we shall
have more to say about this point later.
We may use the language and notation of distance to summarize
the meaning of Examples 4 and 5 and the discussion following them: If
ρ(z) = 1/(1 − |z|2 ) is a metric on the disc D, P = 0, and Q = R + i0,
then dρ (P, Q) = (1/2) log((1 + R)/(1 − R)).
38 Basic Notions of Differential Geometry

Figure 3.

2. Calculus in the Complex Domain

In order that we may be able to do calculus computations easily and
efficiently in the context of complex analysis, we recast some of the
basic ideas in new notation. We define the differential operators
   
∂ 1 ∂ ∂ ∂ 1 ∂ ∂
= −i and = +i .
∂z 2 ∂x ∂y ∂z 2 ∂x ∂y

In complex analysis it is more convenient to use these operators than to

use ∂/∂ x and ∂/∂y. The reason is as follows. Let f (z) = u(z) + iv(z)
be a complex-valued, continuously differentiable function (in the sense
of multivariable calculus) on a planar domain . Notice that

∂
f (z) = 0 on 
∂z
Calculus in the Complex Domain 39

if and only if
 
1 ∂ ∂ 
+i u(z) + iv(z) = 0 on .
2 ∂x ∂y
Taking real and imaginary parts of this last equation leads to the pair of
equations
∂ ∂ ∂ ∂
u= v and v = − u.
∂x ∂y ∂x ∂y
But these are just the Cauchy-Riemann equations. We conclude that
∂
f = 0 on  iff f is holomorphic on .
∂z
Further notice that
∂ ∂
z=1 z=0
∂z ∂z
∂ ∂
z=0 z = 1.
∂z ∂z
Thus ∂/∂z and ∂/∂z fit in a natural way into complex function theory.
Next observe that if f is holomorphic then
f (z + h) − f (z)
f  (z) = lim
Ch→0 h
f (z + s) − f (z)
= lim
Rs→0 s
∂f
=
∂x
∂f
= −i
∂y
∂f
= .
∂z
Thus we see that the notion of complex derivative, introduced in Sec-
tion 0.1, is for holomorphic functions the same as ∂ f /∂z.
40 Basic Notions of Differential Geometry

Finally observe that the Laplace operator

∂2 ∂2
= +
∂x2 ∂y 2
may now be written as
∂ ∂ ∂ ∂
=4 =4 .
∂z ∂z ∂z ∂z
If we are going to use complex derivatives, then we need the usual
tools associated to derivatives. The linear properties (that is, the sum
and scalar multiplication rules) of ∂/∂z and ∂/∂z are obvious. The
product rule is a bit more tedious, but is safely left as an exercise. How-
ever, the chain rule is more subtle; it now takes the following form:

Proposition 1. If f and g are continuously differentiable functions,

and if f ◦ g is well defined on some open set U ⊆ C, then we have
∂ ∂f ∂g ∂f ∂g
( f ◦ g)(z) = (g(z)) (z) + (g(z)) (z)
∂z ∂z ∂z ∂z ∂z
and
∂ ∂f ∂g ∂f ∂g
( f ◦ g)(z) = (g(z)) (z) + (g(z)) (z).
∂z ∂z ∂z ∂z ∂z

Proof. We will sketch the proof of the first identity and leave the sec-
ond as an exercise.
We have
 
∂ 1 ∂ ∂
( f ◦ g) = −i ( f ◦ g).
∂z 2 ∂x ∂y
We write g(z) = α(z) + iβ(z), with α and β real-valued functions, and
apply the usual calculus chain rule for ∂/∂ x and ∂/∂y. We obtain that
the last line equals
 
1 ∂ f ∂α ∂ f ∂β ∂ f ∂α ∂ f ∂β
+ −i −i . (∗)
2 ∂x ∂x ∂y ∂ x ∂ x ∂y ∂y ∂y
Calculus in the Complex Domain 41

Now, with the aid of the identities

 
∂ ∂ ∂ ∂ ∂ ∂
= + and =i − ,
∂x ∂z ∂z ∂y ∂z ∂z

we may reduce the expression (∗) (after some tedious calculations) to

the desired formula.

Corollary 1.1. If either f or g is holomorphic, then

∂ ∂f ∂g
( f ◦ g)(z) = (g(z)) (z).
∂z ∂z ∂z

Here is an example of the utility of our complex calculus notation.

Example 1. Let f be a nonvanishing holomorphic function on a planar

domain . Then


 log(| f |2 ) = 0.

In other words, log(| f |2 ) is harmonic.

To see this, fix P ∈  and let U ⊆  be a neighborhood of P on
which f has a holomorphic logarithm. Then on U we have


(log(| f |2 )) =  log f + log f
∂ ∂ ∂ ∂
=4 log f + 4 log f
∂z ∂z ∂z ∂z
= 0.

We conclude this section with some exercises for the reader:

Exercise.

1. Calculate that, for a > 0,

∂2 a 2 (zz)a−1
log(1 + (zz)a ) = .
∂z∂z (1 + (zz)a )2
42 Basic Notions of Differential Geometry

2. If g is holomorphic, then calculate that

( f ◦ g) = ( f ◦ g) · |g  |2 .

3. If f is holomorphic, then calculate that

( f ◦ g) = ( f  ◦ g)  g + ( f  ◦ g)[(Dx g)2 + (D y g)2 ].

3. Isometries
In any mathematical subject there are morphisms: functions which pre-
serve the relevant properties being studied. In linear algebra these are
invertible linear maps, in Euclidean geometry these are rigid motions,
and in Riemannian geometry these are “isometries.” We now define the
concept of isometry.

Definition 1. Let 1 and 2 be planar domains and let

f : 1 → 2

be a continuously differentiable mapping with isolated zeros. Assume

that 2 is equipped with a metric ρ. We define the pullback of the
metric ρ under the map f to be the metric on 1 given by
 
∂ f 
f ρ(z) = ρ( f (z)) ·   .
∗
∂z

Remark. The particular form that we use to define the pullback is mo-
tivated by the way that f induces mappings on tangent and cotangent
vectors, but this motivation is irrelevant for us here.
It should be noted that the pullback of any metric under a conju-
gate holomorphic f will be the zero metric. Thus we have designed our
definition of pullback so that holomorphic pullbacks will be the ones
of greatest interest. This assertion will be made substantive in Proposi-
tion 3 below.
Isometries 43

Definition 2. Let 1 , 2 be planar domains equipped with metrics ρ1

and ρ2 , respectively. Let

f : 1 → 2

be an onto, holomorphic mapping. If

f ∗ ρ2 (z) = ρ1 (z)

for all z ∈ 1 then f is called an isometry of the pair (1 , ρ1 ) with the
pair (2 , ρ2 ).
The differential definition of isometry (Definition 2) is very natu-
ral from the point of view of differential geometry, but it is not intuitive.
The next proposition relates the notion of isometry to more familiar
ideas.

Proposition 3. Let 1 , 2 be domains and ρ1 , ρ2 be metrics on these

respective domains. If

f : 1 → 2

is an isometry of (1 , ρ1 ) to (2 , ρ2 ), then the following three proper-

ties hold:

(a) If γ : [a, b] → 1 is a continuously differentiable curve, then so

is the push-forward f ∗ γ ≡ f ◦ γ and

ρ1 (γ ) = ρ2 ( f ∗ γ ).

(b) If P, Q are elements of 1 then

dρ1 (P, Q) = dρ2 ( f (P), f (Q)).

(c) Part (b) implies that the isometry f is one-to-one. Then f −1 is well
defined and f −1 is also an isometry.

Proof. Assertion (b) is an immediate consequence of (a). Also (c) is a

formal exercise in definition chasing. Therefore we shall prove (a).
44 Basic Notions of Differential Geometry

By definition,

b
ρ2 ( f ∗ γ ) = ( f ∗ γ ) (t) dt
ρ2 , f ∗ γ (t)
a


b ∂f
= (γ (t)) · γ̇ (t) dt.
∂z
a ρ2 , f ∗ γ (t)

With elementary manipulations, we see that the integrand equals

   
∂ f   
 (γ (t)) · γ̇ (t)ρ , f γ (t) =  ∂ f (γ (t)) · |γ̇ (t)| · ρ2 ( f (γ (t)))
 ∂z  2 ∗  ∂z 
= γ̇ (t) f ∗ ρ2 ,γ (t)
= γ̇ (t)ρ1 ,γ (t) ,

since f is an isometry. Substituting this back into the formula for the
length of f ∗ γ gives

b
ρ2 ( f ∗ γ ) = γ̇ (t)ρ1 ,γ (t) dt = ρ1 (γ ).
a

This ends the proof.

It is not difficult to see that any positively oriented rigid motion

of the Euclidean plane—rotations, translations, and their compositions
(but not reflections, because they are conjugate holomorphic)—is an
isometry of the Euclidean metric. In the next section we shall consider
the isometries of the Poincaré metric on the disc.
If you have previously studied metric space theory or Banach
space theory then you may have already encountered the term “isome-
try.” The essential notion is that an isometry should preserve distance.
In fact, you can prove as an exercise that if f is a holomorphic mapping
of (1 , ρ1 ) onto (2 , ρ2 ) that preserves distance then f is an isometry
according to Definition 2. (Hint: compose f with a curve and differen-
tiate.)
We close this section by noting an important technical fact about
isometries:
The Poincaré Metric 45

Proposition 4. Let ρ j be metrics on the domains  j , j = 1, 2, 3, re-

spectively. Let f : 1 → 2 and g : 2 → 3 be isometries. Then
g ◦ f is an isometry of the metric ρ1 to the metric ρ3 .

Proof. We calculate that

ρ3 (g([ f (z)]) · |g  ( f (z))| = ρ2 ( f (z))

hence

ρ3 (g([ f (z)]) · |g  ( f (z))| · | f  (z)| = ρ2 ( f (z)) · | f  (z)| = ρ1 (z).

In other words,

ρ3 (g ◦ f (z)) · |(g ◦ f ) (z)| = ρ1 (z),

as required by the definition of an isometry.

4. The Poincaré Metric

The Poincaré metric on the disc has occurred in many of the examples
in previous sections. This metric is the paradigm for much of what we
want to do in this book, and we want to treat it in some detail here. The
Poincaré metric on the disc D is given by
1
ρ(z) = .
1 − |z|2
(For the record, we note that there is no agreement in the literature as
to
√ what constant
√ goes in the numerator; many references use a factor of
2 or 1/ 2.)
In this and succeeding sections, we shall use the phrase “confor-
mal map” to refer to a holomorphic mapping of one planar region to
another which is both one-to-one and onto.

Proposition 1. Let ρ be the Poincaré metric on the disc D. Let h :

D → D be a conformal self-map of the disc. Then h is an isometry of
the pair (D, ρ) with the pair (D, ρ).
46 Basic Notions of Differential Geometry

Proof. We have that

h ∗ ρ(z) = ρ(h(z)) · |h  (z)|.

Because of Theorem 3 of Section 0.2 (and using Proposition 4 of Sec-

tion 1.3), we now have two cases:

(i) If h is a rotation, then h(z) = µ · z for some unimodular constant

µ ∈ C. So |h  (z)| = 1 and
1 1
h ∗ ρ(z) = ρ(h(z)) = ρ(µz) = = = ρ(z)
1 − |µz|2 1 − |z|2
as desired.
(ii) If h is a Möbius transformation (these were defined in Section 0.2,
following Theorem 2), then
z−a
h(z) = , some constant a ∈ D.
1 − az
But then
1 − |a|2
|h  (z)| =
|1 − az|2
and
 
z−a
h ∗ ρ(z) = ρ · |h  (z)|
1 − az
1 1 − |a|2
=  2 ·
1− z−a  |1 − az|2
1−az

1 − |a|2
=
|1 − az|2 − |z − a|2
1 − |a|2
=
1 − |z|2 − |a|2 + |a|2 |z|2
1
=
1 − |z|2
= ρ(z).
The Poincaré Metric 47

Since any conformal self-map of D is a composition of maps of the

form (i) or (ii), the proposition is proved.

We know from this result and from Proposition 3 in Section 3 that

conformal self-maps of the disc preserve Poincaré distance. To under-
stand what this means, consider the Möbius transformation
z+a
φ(z) = .
1 + az
Then φ maps the disc to the disc conformally. It does not preserve Eu-
clidean distance, but it does preserve Poincaré distance. See Figure 1.

Q (P)

(Q)
P

Figure 1.

We can use what we have learned so far to calculate the Poincaré

metric explicitly.

Proposition 2. If P and Q are points of the disc D, then the Poincaré

distance of P to Q is
  
1 1 +  P−Q 
dρ (P, Q) = log 
1−P Q 
  .
2 1 −  P−Q 
1−P Q

Proof. In case P = 0 and Q = R + i0, the result was already noted in

Section 1. In the general case, note that we may define
z−P
φ(z) = ,
1 − Pz
48 Basic Notions of Differential Geometry

a Möbius transformation of the disc. Then, by Proposition 1,

dρ (P, Q) = dρ (φ(P), φ(Q)) = dρ (0, φ(Q)).

Next we have

dρ (0, φ(Q)) = dρ (0, |φ(Q)|) (∗)

since there is a rotation of the disc taking φ(Q) to |φ(Q)| + i0. Finally,
 
 P−Q 
|φ(Q)| =   ,
1 − PQ
so that (∗) together with the special case treated in the the first sentence
gives the result.

One of the reasons that the Poincaré metric is so useful is that it

induces the same topology as the usual Euclidean metric topology. That
is our next result.

Proposition 3. The topology induced on the disc by the Poincaré met-

ric is the usual planar topology.

Proof. A neighborhood basis for the topology of the Poincaré metric

at the origin is given by the balls

B(0, r ) = {z : dρ (0, z) < r }.

However, a calculation using Proposition 2 yields that these balls are

the same as the Euclidean discs
 
e2r − 1
z : |z| < 2r .
e +1

These discs form a neighborhood basis for the origin in the Euclidean
topology. Thus we find that the two topologies are the same at the ori-
gin. Now the origin can be moved to any other point a ∈ D by the
Möbius transformation
The Poincaré Metric 49

z+a
z −→ .
1 + az
Since the Poincaré metric is invariant under Möbius transformations,
and since Möbius transformations take circles to circles (after all, they
are linear fractional), the two topologies are the same at every point.

One of the most striking facts about the Poincaré metric on the
disc is that it turns the disc into a complete metric space. How could this
be? The boundary is missing! The reason that the disc is complete in
the Poincaré metric is the same as the reason that the plane is complete
in the Euclidean metric—the boundary is infinitely far away. We now
prove this assertion.

Proposition 4. The unit disc D, when equipped with the Poincaré met-
ric, is a complete metric space.

Proof. Let p j be a sequence in D that is Cauchy in the Poincaré metric.

Then the sequence is bounded in that metric. So there is a positive,
finite M such that

dρ (0, p j ) ≤ M, all j.

With Proposition 2 this translates to

 
1 1 + |pj|
log ≤ M.
2 1 − |pj|

Solving for | p j | gives

e2M − 1
|pj| ≤ < 1.
e2M + 1
Thus our sequence is contained in a relatively compact subset of the
disc. A similar calculation yields that the sequence must in fact be
Cauchy in the Euclidean metric. Therefore it converges to a limit point
in the disc, as required for completeness.
50 Basic Notions of Differential Geometry

In the proof of Proposition 2 and in the subsequent remarks we

used the fact, whose verification was sketched earlier, that the curve
of least length (in the Poincaré metric) connecting 0 to a point of the
form R +i0 is in fact a Euclidean segment. More generally, the shortest
path from 0 to any point w is a rotation of the shortest path from 0 to
|w| + i0, which is a segment. Let us now calculate the “curve of least
length” connecting any two given points P and Q in the disc.

Proposition 5. Let P, Q be elements of the unit disc. The “curve of

least Poincaré length” connecting P to Q is
Q−P
t +P
1−Q P
γ P,Q (t) = Q−P
, 0 ≤ t ≤ 1.
1+tP ·
1−Q P

Proof. Define the Möbius transformation

z−P
φ(z) = .
1 − Pz
By what we already know about shortest paths emanating from the ori-
gin, the curve τ (t) ≡ t · φ(Q) is the shortest curve from φ(P) = 0 to
φ(Q), 0 ≤ t ≤ 1. Applying the isometry
z+P
φ −1 (z) =
1 + Pz
to τ we obtain that
(tφ(Q) + P)
φ −1 ◦ τ (t) =
(1 + Ptφ(Q))

is the shortest path from P to Q. Since φ −1 ◦ τ = γ P,Q , we are done.

Let us analyze the curves discovered in Proposition 5. First notice

that, since the curve γ P,Q is the image under a linear fractional trans-
formation of a part of a line, the trace of γ P,Q is, therefore, either a line
The Poincaré Metric 51

segment or an arc of a circle. In fact, if P and Q are collinear with 0,

then the formula for γ P,Q quickly reduces to that for a segment; other-
wise γ P,Q traces an arc of a Euclidean circle. Which circle is it?
Matters are simplest if we let t range over the entire real line and
look for the whole circle. We find that t = ∞ corresponds to the point
1/P. It is now a simple, but tedious, calculation to determine the Eu-
clidean center and radius of the circle determined by the three points
P, Q, 1/P (note that, by symmetry, the circle will also pass through
1/Q). The circle is depicted in Figure 2. Notice that since the segment
{t + i0 : −1 ≤ t ≤ 1} is orthogonal to ∂ D at the endpoints 1 and −1,
conformality dictates that the circular arcs of least length provided by
Proposition 5 are orthogonal to ∂ D at the points of intersection. Some
of these “geodesic arcs” are exhibited in Figure 3.

Figure 2.

Figure 3.
52 Basic Notions of Differential Geometry

A final note on this matter is that geodesics are particularly easy

to calculate in the upper-half-plane realization U of the disc. For the
geodesic circles will then have their centers on the boundary (i.e., the
 Q
real line). If P,  are points of U, then the perpendicular bisector of
the segment connecting these points will intersect the real axis at the
center C of the circular arc that forms the geodesic through P  and Q.

See Figure 4.

~
Q
X
~
P

Figure 4.

We next see that the Poincaré metric is characterized by its prop-

erty of invariance under conformal maps.

Proposition 6. If ρ(z) is a metric on D which is such that every con-

formal map of the disc is an isometry of the pair (D, ρ ) with the pair
(D, ρ
), then ρ
 is a constant multiple of the Poincaré metric ρ.

Proof. The hypothesis guarantees that if z 0 ∈ D is fixed and

z + z0
h(z) =
1 + z0z
The Poincaré Metric 53

then

h∗ρ
(0) = ρ
(0).

Writing out the left-hand side gives

|h  (0)|
ρ (h(0)) = ρ
(0)

or
1
(z 0 ) =
ρ ·ρ
(0) = ρ
(0) · ρ(z 0 ).
1 − |z 0 |2
Thus we have exhibited ρ
 as the constant ρ
(0) times ρ.

Now that we know that the Poincaré metric is the right metric for
complex analysis on the disc, a natural next question is to determine
which other maps preserve the Poincaré metric.

Proposition 7. Let f : D → D be continuously differentiable and

let ρ be the Poincaré metric. If f pulls back the pair (D, ρ) to the pair
(D, ρ), then f is holomorphic and is one-to-one. We may conclude then
that f is the composition of a Möbius transformation and a rotation
(see Theorem 3 of Section 0.2).

Proof. First suppose that f (0) = 0. For R > 0 let C R be the set of
points in D that have Poincaré distance R from 0. Since the Poincaré
metric is invariant under rotations (after all, rotations are holomorphic
self-maps), it follows that C R is a Euclidean circle (however this circle
will have Euclidean radius (e2R −1)/(e2R +1), not R). Since f (0) = 0
and f preserves the metric, it follows that f (C R ) = C R . The fact that
f is distance-preserving then shows that f is one-to-one on each C R .
As a result, f is globally one-to-one.
Let P ∈ C R . Then
| f (P) − f (0)| | f (P)|
= = 1.
|P − 0| |P|
54 Basic Notions of Differential Geometry

Letting R → 0+ we conclude that f is conformal at the origin (that

is, f preserves length infinitesimally). Since f pulls back the metric ρ,
we can be sure that ∂ρ/∂z = 0 at the origin.
Now we drop the special hypothesis that f (0) = 0. Pick an arbi-
trary z 0 ∈ D and set w0 = f (z 0 ). Define
z + z0 z − w0
φ(z) = , ψ(z) = .
1 + z0 z 1 − w0 z
Also define

g = ψ ◦ f ◦ φ.

Then g(0) = 0 and g is an isometry (since ψ, f , and φ are); therefore

the argument in the last paragraph applies and g is conformal at the
origin. It follows as before that ∂g/∂z = 0 at the origin.
We conclude that f is conformal at every point—since ψ and φ
are—and ∂ f /∂z = 0 at every point. Therefore f is holomorphic.

Propositions 1 and 7 demonstrate that there are plenty of isome-

tries of the unit disc in the Poincaré metric. On the other hand, isome-
tries are very rigid objects. They are completely determined by their
first-order behavior at just one point. While a proof of this assertion in
general is beyond us at this point, we can certainly prove the result for
the Poincaré metric on the disc.

Proposition 8. Let ρ be the Poincaré metric on the disc. Let f be an

isometry of the pair (D, ρ) with the pair (D, ρ). If f (0) = 0 and
∂ f /∂z(0) = 1 then f (z) ≡ z.

Proof. By Proposition 7, f must be holomorphic. Since f preserves

the origin and is one-to-one and onto, f must be a rotation. Since
f  (0) = 1, it follows that f is the identity.

Corollary 8.1. Let f and g be isometries of the pair (D, ρ) with

the pair (D, ρ). Let z 0 ∈ D and suppose that f (z 0 ) = g(z 0 ) and
(∂ f /∂z)(z 0 ) = (∂g/∂z)(z 0 ). Then f (z) ≡ g(z).
The Schwarz Lemma 55

Proof. We noted in Section 2 that g −1 is an isometry. If ψ is a Möbius

transformation that takes 0 to z 0 , then ψ −1 ◦ g −1 ◦ f ◦ ψ satisfies the
hypothesis of the Proposition. As a result ψ −1 ◦ g −1 ◦ f ◦ ψ(z) ≡ z or
g(z) ≡ f (z).

5. The Schwarz Lemma

One of the important facts about the Poincaré metric is that it can be
used to study not just conformal maps (as in Section 1.4) but all holo-
morphic maps of the disc. The key to this assertion is the classical
Schwarz lemma. We begin with an elegant geometric interpretation of
the Schwarz–Pick lemma (see Section 0.2).

Proposition 1. Let f : D → D be holomorphic. Then f is distance-

decreasing in the Poincaré metric ρ. That is, for any z ∈ D,

f ∗ ρ(z) ≤ ρ(z).

The integrated form of this assertion is that if γ : [0, 1] → D is a

continuously differentiable curve then

ρ ( f ∗ γ ) ≤ ρ (γ ).

Therefore, if P and Q are elements of D, we may conclude that

dρ ( f (P), f (Q)) ≤ dρ (P, Q).

Proof. Now

1
f ∗ ρ(z) ≡ | f  (z)|ρ( f (z)) = | f  (z)| ·
1 − | f (z)|2

and
1
ρ(z) =
1 − |z|2
56 Basic Notions of Differential Geometry

so the asserted inequality is just the Schwarz–Pick lemma. The inte-

grated form of the inequality now follows by the definition of ρ . The
inequality for the distance dρ follows from the definition of distance.

Notice that, in case f is a conformal self-map of the disc, we may apply

this corollary to both f and f −1 to conclude that f preserves Poincaré
distance, giving a second proof of Proposition 1 of Section 4.
We next give an illustration of the utility of the geometric point of
view. We will see that the proposition just proved gives a very elegant
proof of Theorem 2 below (see [EAH] for the source of this proof).

Theorem 2. (Farkas, Ritt). Let f : D → D be holomorphic and

assume that the image M = { f (z) : z ∈ D} of f has compact closure
in D. Then there is a unique point P ∈ D such that f (P) = P. We call
P a fixed point for f .

Proof. By hypothesis, there is an  > 0 such that if m ∈ M and |z| ≥ 1

then |m − z| > 2. See Figure 1. Fix z 0 ∈ D and define

g(z) = f (z) + ( f (z) − f (z 0 )).

Image of f

Figure 1.
The Schwarz Lemma 57

Then g is holomorphic and g still maps D into D. Also

g  (z 0 ) = (1 + ) f  (z 0 ).

By the preceding proposition, g is thus distance-decreasing in the

Poincaré metric. Therefore

g ∗ ρ(z 0 ) ≤ ρ(z 0 ).

Writing out the definition of g ∗ now yields

(1 + ) · f ∗ ρ(z 0 ) ≤ ρ(z 0 ).

Note that this inequality holds for any z 0 ∈ D. But now if γ : [a, b] →
D is any continuously differentiable curve, then we may conclude that

ρ ( f ∗ γ ) ≤ (1 + )−1 ρ (γ ).

If P and Q are elements of D and d is Poincaré distance then we have

that

d( f (P), f (Q)) ≤ (1 + )−1 d(P, Q).

We see that f is a contraction in the Poincaré metric. Recall that

in Section 4 we proved that the disc D is a complete metric space when
equipped with the Poincaré metric. By the contraction mapping fixed-
point theorem (see [LS]), f has a unique fixed point.

The next result shows how to find the fixed point, and in effect
proves the contraction mapping fixed-point theorem in this special case.

Corollary 2.1. If f is as in the theorem and P is the unique fixed point,

then the iterates f, f ◦ f, f ◦ f ◦ f, . . . converge uniformly on compact
sets to the constant function P.

Proof. Let f n denote the nth iterate of f . Let

B(P, R) = {z ∈ D : d(z, P) ≤ R}
58 Basic Notions of Differential Geometry

be the closed Poincaré metric ball with center P and Poincaré metric
radius R. Then the theorem tells us that

f (B(P, R)) ⊆ B(P, R/(1 + ))

and, more generally,

f n (B(P, R)) ⊆ B(P, R/(1 + )n ). (∗)

Observe that

∞
B(P, j) = D.
j=1

Now Proposition 2 of Section 4 tells us that these non-Euclidean balls

are in fact open discs in the usual Euclidean topology. Therefore every
(Euclidean) compact subset K of D lies in some B(P, j). Then line (∗)
implies that


f n (K ) ⊆ B P, j/(1 + )n .

The result follows.

6. A Detour into Non-Euclidean Geometry

The main purpose of this book is to show how geometry can influence
and explain complex function theory. But function theory also pays
back geometry in a variety of ways. Certainly the hyperbolic disc—
the unit disc equipped with the Poincaré metric—is one of the most
important fundamental examples in any basic course on Riemannian
geometry.
One added bonus for us is that the hyperbolic disc is also an el-
egant model for the classical non-Euclidean geometry of Bolyai and
Lobachevsky. Since we have done all the necessary work to appreciate
A Detour into Non-Euclidean Geometry 59

this construction, we take a little byway here to discuss some of the

details.
The reader may recall that János Bolyai (1802–1860) and Nicolai
Lobachevsky (1793–1856) are generally credited with the discovery
of non-Euclidean geometry. The problem of determining the role of
the parallel postulate in Euclidean geometry was two thousand years
old, and many good minds had wrestled with it. Passions in the matter
ran high. Bolyai’s father, a provincial mathematics teacher, had worked
hard on the problem. When he learned that his son, an army officer at
the time, was absorbed in the problem he wrote to him

For God’s sake, I beseech you, give it up. Fear it no less than sensual
passions because it, too, may take all your time, and deprive you of your
health, peace of mind, and happiness.

János Bolyai did ultimately crack the problem (a couple of years in ad-
vance of Lobachevsky), producing an independent geometry in which
the parallel postulate fails. The elder Bolyai conveyed the discovery
to his friend Carl Friedrich Gauss (1777–1855). One can only imagine
János’s chagrin to read this reply from Gauss:

If I begin with the statement that I dare not praise such a work, you
will of course be startled for a moment: but I cannot do otherwise; to
praise it would amount to praising myself; for the entire content of the
work, the path which your son has taken, the results to which he is led,
coincide almost exactly with my own meditations which have occupied
my mind for from thirty to thirty-five years. On this account I find myself
surprised to the extreme.
My intention was, in regard to my own work, of which very little
up to the present has been published, not to allow it to become known
during my lifetime. Most people have not the insight to understand our
conclusions and I have encountered only a few who received with any
particular interest what I communicated to them. In order to understand
these things, one must first have a keen perception of what is needed,
and upon this point the majority are quite confused. On the other hand,
it was my plan to put all down on paper eventually, so that at least it
would not finally perish with me.
60 Basic Notions of Differential Geometry

So I am greatly surprised to be spared this effort, and am overjoyed

that it happens to be the son of my old friend who outstrips me in such a
remarkable way.

Lobachevsky’s publication of his results in 1829 is historically

significant. In the subsequent two decades he published three addi-
tional major works on the subject. On the strength of this mathematics,
Gauss recommended that Lobachevsky be given a membership in the
Göttingen Scientific Society. Today non-Euclidean geometry is consid-
ered to be straightforward, and it is sometimes taught to high school
students. For us, it is a convenient and charming byproduct of our con-
struction of Poincaré geometry on the disc.
First recall the basic tenets of Euclidean geometry. There are eight
undefinable terms: point, line, segment, parallel, right angle, the phrase
“lie on,” the concept of “between,” and “congruent.” [With some effort,
the concept of “circle” can also be defined using just these undefin-
ables.] We understand that we can obtain an intuitive understanding of
these notions, but they cannot be defined rigorously (because there are
no earlier concepts in terms of which to define them). The axioms of
Euclidean geometry are now formulated using these undefinable terms.
We in fact will give a modern formulation of these axioms (this version
has been strongly influenced by the work of David Hilbert). See [KR5]
for more on these ideas.

P1 Through any pair of distinct points there passes a line.

P2 For each segment AB and each segment C D there is a unique point
E (on the line determined by A and B) such that B is between A
and E and the segment C D is congruent to B E (Figure 1).
P3 For each point C and each point A distinct from C there exists a
circle with center C and radius C A.
P4 All right angles are congruent.

These are the standard four axioms which give our Euclidean con-
ception of geometry. The fifth axiom, a topic of intense study for two
A Detour into Non-Euclidean Geometry 61

D
C

Figure 1.

thousand years, is the so-called parallel postulate (in Playfair’s formu-

lation):

P5 For each line  and each point P that does not lie on  there is a
unique line m through P such that m is parallel to .

A model for this geometry is a collection of mathematical objects

which can be assigned to the terms in the axiom system and which then
satisfy those axioms. The most familiar context for Euclidean geometry
is the familiar geometry of lines which can be drawn with a ruler and
circles which can be drawn with a compass. This model satisfies all five
of Euclid’s axioms.
A non-Euclidean geometry is a geometry that satisfies axioms P1–
P4 but not axiom P5. We now explain how the geometry of the hyper-
bolic disc is such a geometry.
The universe for our new geometry will be the open unit disc D
in the plane. We know that the segment  ≡ {x + i0 : −1 < x < 1} is
a “line,” in the sense that it is a geodesic in hyperbolic geometry (see
Example 5 of Section 1). Now any subset of D that can be obtained as
φ() for φ a conformal self-map of the disc will also be called a line.
We will say that two of these lines are parallel if they do not intersect.
Can we give an extrinsic description of these new objects which we call
lines?
Recall (Theorem 3 of Section 0.2) that the conformal self-maps
of the disc consist of the Möbius transformations, the rotations, and
62 Basic Notions of Differential Geometry

their compositions. Now it is easy to see that a rotation of  will just be

another diameter through the origin. More interesting is the image of 
under a Möbius transformation (Section 0.2).
Let us consider the specific instance of the Möbius transformation
z − ia
φia (z) =
1 − iaz
for −1 < a < 1 a real number. Then φia () is the set of points
  
x(1 − a 2 ) −a(1 + x 2 )
, : −1 < x < 1 .
1 + a2 x 2 1 + a2 x 2

In fact a tedious calculation shows that this is a portion of the circle of

center (0, [1+a 2 ]/[−2a]) and radius (1−a 2 )/2|a|. Of course, properly
speaking, we are interested in that arc of this circle that lies inside the
disc D.
By conformality, since  is orthogonal to the unit circle at its end-
points, it follows that the new “line” φia () will be orthogonal to the
unit circle at its endpoints (Figure 2).

Figure 2.

Conversely, if C is any arc of a circle that lies in D and is perpen-

dicular to the unit circle at its endpoints, then C arises as some Möbius
transformation of . For, after a rotation, we may as well assume that C
is symmetric with respect to the y-axis and lies in the upper half-plane
A Detour into Non-Euclidean Geometry 63

Figure 3.

(Figure 3). Then the endpoints of this arc will certainly have the form
   
−(1 − a 2 ) −2a (1 − a 2 ) −2a
, , ,
1 + a2 1 + a2 1 + a2 1 + a2

for some real a between −1 and 0. Then we see that our circular arc C
is just φia ().
Of course we take a point in our new geometry to be just a stan-
dard Euclidean point. Then it is straightforward to verify that the lines
satisfy axioms P1 and P2. Axiom P3 is not relevant for our discussion
of the parallel postulate since it relates to circles. Axiom P4 follows
from conformality.
The punch line, of course, is that Axiom P5 is not satisfied by the
lines in our new geometry. We repeat that two lines are parallel if they
do not intersect. Now fix the line  as before, and consider the point
p = (0, 1 − ) not on that line. There is certainly a circular arc m of
the kind we have been discussing that (i) passes through p and (ii) is
disjoint from  (Figure 4). Now any small rotation of m will also be
disjoint from  and will intersect m in a point (Figure 5). Call that point
of intersection q. Then q is a point that is not on the line , yet we
have produced two distinct lines that pass through it and are disjoint
64 Basic Notions of Differential Geometry

Figure 4.

pq

Figure 5.

from . Of course each of these two lines is, by definition, parallel to .

Thus we have produced a situation that is inconsistent with the parallel
postulate.
To summarize: We have a model of geometry (the classical Eu-
clidean rectilinear model) in which all five of Euclid’s axioms are true.
And we have another model in which P1, P2, (P3), and P4 are true but
P5 is false. This signals that Axiom P5 is independent of the other four.
What is elegant about this discussion is that it requires no real
work: conformality, and the computations that we have already per-
formed in another context, are sufficient to see that Euclid’s 2000-year-
old axiom is independent of the other four axioms of Euclid’s geometry.
We close this chapter with a picture (Figure 6) of a tesselation
made up of geodesics from hyperbolic geometry.
A Detour into Non-Euclidean Geometry 65

Figure 6.
CHAPTER 2
Curvature and Applications

1. Curvature and the Schwarz

Lemma Revisited
If U ⊆ C is a planar domain and ρ is a metric on U , then the curvature
of the metric ρ at a point z ∈ U (at which ρ(z) = 0) is defined to be

− log ρ(z)
κU,ρ (z) = κ(z) ≡ . (∗)
ρ(z)2

(Here zeros of ρ(z) will result in singularities of the curvature

function—κ is undefined at such points.) We will study this quantity
for a bit, and then summarize and provide motivation afterward.
Since ρ is twice continuously differentiable, this definition makes
sense. It assigns to each z ∈ U a numerical quantity. The most impor-
tant preliminary fact about κ is its conformal invariance:

Proposition 1. Let U1 and U2 be planar domains and h : U1 → U2

a conformal map (in particular, h
never vanishes). If ρ is a metric on
U2 , then

κU1 ,h ∗ ρ (z) = κU2 ,ρ (h(z)), ∀z ∈ U1 .

67
68 Curvature and Applications

Proof. We need to calculate:

− log[ρ(h(z)) · |h
(z)|]
κU1 ,h ∗ ρ (z) ≡
[ρ(h(z)) · |h
(z)|]2
− log[ρ(h(z))] − [log |h
(z)|]
= .
[ρ(h(z)) · |h
(z)|]2

Now the second term in the numerator vanishes because of Example 1

in Section 1.2. We may further simplify the numerator (use Proposi-
tion 1 in Section 1.2 as well as Exercise 2 at the end of that section) to
obtain that the last line


− ( log ρ)((h(z)) |h
(z)|2
=
[ρ(h(z)) · |h
(z)|]2
− log ρ|h(z)
=
ρ(h(z))2
= κU2 ,ρ (h(z)).

Remark. In fact the proof gives the slightly more general fact that if
U1 , U2 are domains and f : U1 → U2 is a holomorphic map (not nec-
essarily one-to-one or onto), then the following holds. If ρ is a metric
on U2 , then

κ f ∗ ρ (z) = κρ ( f (z))

at every point z ∈ U1 for which f

(z) = 0 and ρ( f (z)) = 0. Usually
this is all points except for a discrete set.

From the point of view of geometry, any differential quantity

which is invariant is automatically of great interest. But why do we call
κ “curvature”? Why don’t we call it “mass” or “humidity” or “color”?
The answer is that a standard construction in Riemannian geometry as-
signs to any Riemannian metric a quantity called Gaussian curvature.
In the setting of classical Euclidean geometry this quantity is consistent
with our intuitive notion of what curvature ought to be (i.e., the rate of
Curvature and the Schwarz Lemma Revisited 69

change of the normal). In more abstract settings, the structural equa-

tions of Cartan lead to a quantity called curvature which is invariant
but which has somewhat less intuitive content. In the complex plane,
the curvature calculation leads to the formula which we have used to
define κ. One of the pleasant facts about the use of elementary differ-
ential geometry in complex function theory is that one does not require
all the geometric machinery which leads to the formula for κ in order
to make good use of κ.
Because of the considerations in the last paragraph, we shall sim-
ply accept the definition of κ as given. The geometric calculations
which lead to the definition of κ are provided in the Appendix. Un-
derstanding of the material which follows in the text does not in any
way depend on the material in this Appendix.
Let us begin our study of curvature by calculating the curvature of
the Euclidean metric.

Example 1. Let U be a planar domain equipped with the Euclidean

metric ρ(z) ≡ 1. It follows from the definition that κ(z) ≡ 0. This
should be expected, for the Euclidean metric does not change from
point to point.

Example 2. The metric

2
σ (z) =
1 + |z|2
on C is called the spherical metric (see Section 3 below for details).
A straightforward calculation shows that the curvature of σ is identi-
cally 1. Manifolds of constant curvature are a matter of great interest
(see [WOL]).

Next we calculate the curvature of the Poincaré metric on the disc,

which calculation will be of great utility for us. We will learn that the
curvature is negative. Notice that since any point of the disc may be
moved to any other by a Möbius transformation, and since curvature
is a conformal invariant, we will expect the curvature function to be in
fact a negative constant.
70 Curvature and Applications

Proposition 2. Consider the disc D equipped with the Poincaré metric

ρ. For any point z ∈ D we have κ(z) = −4.

Proof. We notice that

− log ρ(z) =  log(1 − |z|2 ).

Now we write  = 4(∂/∂z)(∂/∂z) (see Section 1.2) and |z|2 = z · z to

obtain that this last
4
=− .
(1 − |z|2 )2

It follows that κ(z) ≡ −4.

It was Ahlfors [AHL1] who first realized that the Schwarz lemma
is really an inequality about curvature. In his annotations to his col-
lected works he modestly asserts that “This is an almost trivial fact
and anybody who sees the need could prove it at once.” However, he
goes on to say (most correctly) that this point of view has been a de-
cisive influence in modern function theory. It is a good place to begin
our understanding of curvature. Here is Ahlfors’s version of Schwarz’s
lemma:

Theorem 3. Let the disc D = D(0, 1) be equipped with the Poincaré

metric ρ and let U be a planar domain equipped with a metric σ . As-
sume that, at all points of U, σ has curvature not exceeding −4. If
f : D → U is holomorphic, then we have

f ∗ σ (z) ≤ ρ(z), ∀z ∈ D.

Proof following [MIS]. Let 0 < r < 1. On the disc D(0, r ) define the
metric
r
ρr (z) = .
r 2 − |z|2
Curvature and the Schwarz Lemma Revisited 71

Then straightforward calculations (or change of variable) show that ρr

is the analogue of the Poincaré metric for D(0, r ): it has constant cur-
vature −4 and is invariant under conformal maps. Define

f ∗σ
v= .
ρr
Observe that v is continuous and nonnegative on D(0, r ) and that
v → 0 when |z| → r (since f ∗ σ is bounded above on D(0, r ) ⊂ D
while ρr → ∞). It follows that v attains a maximum value M = M(r )
at some point τ ∈ D(0, r ). We will show that M ≤ 1, hence v ≤ 1 on
D(0, r ). Letting r → 1− then finishes the proof.
If f ∗ σ (τ ) = 0, then v ≡ 0. So we may suppose that f ∗ σ (τ ) > 0.
Therefore κ f ∗ σ is defined at τ . By hypothesis,

κ f ∗ σ ≤ −4.

Since log v has a maximum at τ , we have

0 ≥  log v(τ ) =  log f ∗ σ (τ ) −  log ρr (τ )

= −κ f∗ σ (τ ) · ( f ∗ σ (τ ))2
+ κρr (τ ) · (ρr (τ ))2
≥ 4( f ∗ σ (τ ))2 − 4(ρr (τ ))2 .

This gives

f ∗ σ (τ )
≤1
ρr (τ )
or

M≤1

as desired.

Observe that the usual Schwarz lemma is a corollary of Ahlfors’s

version: we take U to be the disc with σ being the Poincaré metric. Let
72 Curvature and Applications

f be a holomorphic mapping of D to U = D such that f (0) = 0. Then

σ satisfies the hypotheses of the theorem and the conclusion is that

f ∗ σ (0) ≤ ρ(0).

Unravelling the definition of f ∗ σ yields

| f
(0)| · σ ( f (0)) ≤ ρ(0).

But σ = ρ and f (0) = 0, so this becomes

| f
(0)| ≤ 1.

Exercise. The property

d( f (P), f (Q)) ≤ d(P, Q),

where d = dρ is the Poincaré distance, is called the distance-

decreasing property of the Poincaré metric. Use this property to give a
geometric proof of the other inequality in the classical Schwarz lemma.

With a little more notation we can obtain a more powerful version

of the Ahlfors/Schwarz lemma. Let D(0, α) be the open disc of radius
α and center 0. For A > 0 define the metric ραA on D(0, α) by

2α
ραA (z) = √ .
A(α 2 − |z|2 )

This metric has constant curvature −A. Then we have

Theorem 4. Let U be a planar domain that is equipped with a metric

σ whose curvature is bounded above by a negative constant −B. Then
every holomorphic function f : D(0, α) → U satisfies
√
∗ A
f σ (z) ≤ √ ραA (z), ∀z ∈ D(0, α).
B
Liouville’s Theorem and Other Applications 73

It is a good exercise to construct a proof of this more general result,

modeled of course on the proof of Theorem 3.
In the next three sections we shall see several elegant applications
of Theorems 3 and 4.

2. Liouville’s Theorem and

Other Applications
It turns out that the invariant given by curvature provides criteria for
when there do or do not exist non-constant holomorphic functions from
a domain U1 to a domain U2 . The most basic result along these lines is
as follows.

Theorem 1. Let U ⊆ C be an open set equipped with a metric σ (z)

having the property that its curvature κσ (z) satisfies

κσ (z) ≤ −B < 0

for some positive constant B and for all z ∈ U . Then any holomorphic
function

f :C→U

must be constant.

Proof. Fix A > 0. For α > 0 we consider the restricted mapping

f : D(0, α) → U.

Here D(0, α) is the Euclidean disc with center 0 and radius α, equipped
with the metric ραA (z) as in the last section. Theorem 4 of Section 1
yields, for any fixed z and α > |z|, that
√
∗ A
f σ (z) ≤ √ ραA (z).
B
74 Curvature and Applications

Letting α → +∞ yields

f ∗ σ (z) ≤ 0,

hence

f ∗ σ (z) = 0.

This can only be true if f

(z) = 0. Since z was arbitrary, we conclude
that f is constant.

An immediate consequence of Theorem 1 is Liouville’s theorem.

Recall that an entire function is a holomorphic function whose domain
is the entire complex plane.

Theorem 2. Any bounded entire function is constant.

Proof. Let f be such a function. After multiplying f by a constant,

we may assume that the range of f lies in the unit disc. However, the
Poincaré metric on the unit disc has constant curvature −4. Thus The-
orem 1 applies and f must be constant.

Picard’s theorem (Corollary 4 below) is a dramatic strengthening

of Liouville’s theorem. It says that the hypothesis “bounded” may be
weakened considerably, yet the same conclusion may be drawn.
Let us begin our discussion with a trivial example.

Example 1. Let f be an entire function taking values in the set

S = C \ {x + i0 : 0 ≤ x ≤ 1}.

Following f by the mapping

z
φ(z) = ,
z−1
Liouville’s Theorem and Other Applications 75

we obtain an entire function g = φ ◦ f taking values in C less the set

{x + i0 : x ≤ 0}. If r (z) is the principal branch of the square root
function, then h(z) = r ◦ g(z) is an entire function taking values in the
right half plane. Now the Cayley map

z−1
c(z) =
z+1

takes the right half plane to the unit disc. So u(z) = c ◦ h(z) is a
bounded entire function. We conclude from Theorem 1 that u is con-
stant. Unravelling our construction, we have that f is constant.

The point of this easy example is that, far from being bounded,
an entire function need only omit a segment from its values in order
that it be forced to be constant. And a small modification of the proof
shows that the segment can be arbitrarily short. Picard considered the
question of how large a set can be omitted from the image of a non-
constant entire function.
Let us pursue the same line of inquiry rather modestly by asking
whether a non-constant entire function can omit one complex value.
The answer is “yes,” for f (z) = e z assumes all complex values except
zero. It also turns out (see Section 3.5) that it is impossible to construct
a metric on the plane less a point that has negative curvature bounded
away from zero.
The next step is to ask whether a non-constant entire function f
can omit two values. The striking answer, discovered by Picard, is “no.”
Because of Theorem 1, it suffices for us to prove the following:

Theorem 3. Let U be a planar open set such that C \ U contains at

least two points. Then U admits a metric µ whose curvature κµ (z)
satisfies

κµ (z) ≤ −B < 0

for some positive constant B.

76 Curvature and Applications

Proof. After applying an invertible linear map to U we may take the

two omitted points to be 0 and 1 (if there are more than two omitted
points, we may ignore the extras). Thus we will construct a metric of
strictly negative curvature on C0,1 ≡ C \ {0, 1}.
Define
 1/2    1/2 
1 + |z|1/3 1 + |z − 1|1/3
µ(z) = · .
|z|5/6 |z − 1|5/6

(After the proof, we will discuss where this nonintuitive definition

came from.) Of course µ is positive and smooth on C0,1 . We proceed
to calculate the curvature of µ.
First notice that, away from the origin,
  5  
 log |z|5/6 =  log |z|2 = 0
12

by Example 1 of Section 1.2. Thus

 1/2 
1 + |z|1/3 1  
 log =  log 1 + |z|1/3
|z|5/6 2

∂2  
=2 log 1 + [z · z]1/6 .
∂z∂z

Now a straightforward calculation (use the Exercise at the end of

Section 1.2) leads to the identity
 1/2 
1 + |z|1/3 1 1
 log =   .
|z|5/6 18 |z|5/3 1 + |z|1/3 2

The very same calculation shows that

 
(1 + |z − 1|1/3 )1/2 1 1
 log =   .
|z − 1|5/6 18 |z − 1|5/3 1 + |z − 1|1/3 2
Liouville’s Theorem and Other Applications 77

The definition of curvature now yields that


1 |z − 1|5/3
κµ (z) = −    
18 1 + |z|1/3 3 1 + |z − 1|1/3

|z|5/3
+  3 .
1 + |z|1/3 1 + |z − 1|1/3

We record the following obvious facts:

(a) κµ (z) < 0 for all z ∈ C0,1 ;

1
(b) limz→0 κµ (z) = − ;
36
1
(c) limz→1 κµ (z) = − ;
36
(d) limz→∞ κµ (z) = −∞.

It follows immediately that κµ is bounded from above by a negative

constant.

Remark. Now we discuss the motivation for the construction of µ.

When you reach the discussion in Section 3.5, you will get a thorough
mathematical treatment of these issues.
On looking at the definition of µ, one sees that the first factor is
singular at 0 and the second is singular at 1. Let us concentrate on the
first of these.
Since the expression defining curvature is rotationally invariant, it
is plausible that the metric we define would also be rotationally invari-
ant about its singularities. Thus it should be a function of |z|. Hence one
would like to choose exponents α, β so that (1 + |z|α )β defines a met-
ric of negative curvature. However, a calculation reveals that the α, β
which are suitable for z large are not suitable for z small and vice-versa.
This explains why the expression has powers both of |z| (for behavior
near 0) and of (1 + |z|) (for behavior near ∞). A similar discussion
applies to the factors |z − 1|α .
78 Curvature and Applications

Our calculations thus lead us to design the metric so that it behaves

like |z|−4/3 near ∞ and behaves like |z|−5/6 (respectively, |z − 1|−5/6 )
near 0 (respectively, 1).
Of course there is more going on than these brief comments might
indicate. For the construction is not possible on C \ {0}. One must work
on a plane with two punctures. There is a delicate balance between the
factors (discussed above) which are centered at 0 and those which are
centered at 1.

We formulate Picard’s little theorem as a corollary of Theorem 1

and Theorem 3:

Corollary 4. (Picard). Let f be an entire analytic function taking val-

ues in a set U . If C \ U contains at least two points, then f is constant.

Proof. Since C \ U contains at least two points, Theorem 3 says that

there is a metric of strictly negative curvature on U . But then Theorem 1
implies that any entire function taking values in U is constant.

Entire functions are of two types: there are polynomials and non-
polynomials (transcendental entire functions). Notice that a polyno-
mial has a pole at infinity. Conversely, any entire function with a pole
at infinity is a polynomial (see Section 0.4). So a transcendental entire
function cannot have a pole at infinity and, being unbounded (by Liou-
ville), cannot have a removable singularity at infinity. It therefore has
an essential singularity at infinity.
Now notice that, by the fundamental theorem of algebra, a poly-
nomial assumes all complex values. For a transcendental function, we
analyze its essential singularity at infinity by recalling the Casorati-
Weierstrass theorem (see Section 0.4): if 0 is an essential singularity for
a holomorphic function f on a punctured disc D
(0, ) ≡ D(0, ) \ {0}
then f assumes values on D
(0, ) which are dense in the complex
plane. One might therefore conjecture that the essential feature of Pi-
card’s theorem is not that the function being considered is entire, but
Normal Families and the Spherical Metric 79

rather that it has an essential singularity at infinity. This is indeed the

case and is the content of the great Picard theorem, which we treat in
Section 4.

3. Normal Families and the Spherical Metric

Recall that in Section 0.3 we briefly considered the concept of normal
family. Now we go into the subject more deeply; in particular, we give
a more general definition of the concept of normal family. We shall
learn that the property of being normal is, when suitably formulated,
equivalent to an equicontinuity condition formulated in certain metrics.
We begin with our new definition:

Definition 1. Let  be a planar domain and let {g j } be a sequence of

complex-valued functions on . We say that the sequence is normally
convergent to a limit function g if, for every  > 0 and every compact
set K ⊆ , there is a finite, positive J such that whenever j > J and
z ∈ K , then

|g j (z) − g(z)| < .

In short, the sequence is normally convergent if it converges uniformly

on compact subsets of .

Definition 2. Let  be a planar domain and let {g j } be a sequence of

complex-valued functions on . We say that the sequence is compactly
divergent if, for every compact set K ⊆  and every compact set L ⊆
C, there is a finite, positive J such that if j > J and z ∈ K then
g j (z) ∈
/ L.

Briefly, the sequence is compactly divergent if it converges to ∞,

uniformly on compact sets.

Definition 3. Let  be a planar domain and let F be a family of

complex-valued functions on . We call F a normal family if every
80 Curvature and Applications

sequence of elements of F has either a subsequence which is normally

convergent or a subsequence which is compactly divergent.

Example 1. Let F = { f j }, with f j (z) = z j , j = 1, 2, . . . . Then F

is a normal family on U1 = {z : |z| < 1} because every subsequence
converges normally to the zero function. Also F is a normal family on
U2 = {z : |z| > 1} because every subsequence is compactly divergent.
However, F is not normal on any domain U3 which contains a
point of the unit circle. The reason is that such a U3 will contain points
inside the circle, on which { f j } converges, and points outside the circle,
on which { f j } diverges.

Notice that, contrary to many complex analysis texts, we do not re-

quire in our definition of normal family that the functions in question
be holomorphic. In point of fact, normal families are useful outside of
complex function theory and it is helpful to have some extra flexibility.
However, if we do assume in advance that the functions we are treating
are holomorphic, then there are some beautiful theorems about normal
families, as we shall now see.
The standard result in the theory of normal families of holomor-
phic functions is Montel’s theorem:

Theorem 4. Let  be a planar domain and let F be a family of holo-

morphic functions on . If for each compact K ⊆  there is a number
M K such that

| f (z)| ≤ M K , ∀z ∈ K , f ∈ F, (∗)

then F is a normal family.

Corollary 4.1. If , F are as in the theorem and if instead of (∗) we

have a single constant M such that

| f (z)| ≤ M, ∀z ∈ , f ∈ F,

then F is a normal family.

Normal Families and the Spherical Metric 81

Remark. Notice that no family F of holomorphic functions satisfy-

ing the hypotheses of Montel’s theorem could have a compactly diver-
gent subsequence. Thus such a family F has the property that every
sequence of elements of F has a subsequence that is normally conver-
gent.
After we apply some geometry to the study of normal families,
we will be able to derive a version of Montel’s theorem which is much
more general, and which is invariant in a natural sense; it will allow for
compact divergence as well.

The proof of Montel’s theorem is discussed in Section 0.3; further

details may be found, for instance, in [AHL2, p. 219, ff.] or [GRK].
Now we turn to some examples.

Example 2. Let F be the family of all holomorphic functions on a

domain  that take values in the right half plane. Define
z−1
φ(z) = .
z+1
Then G ≡ {φ ◦ f : f ∈ F} is a family of holomorphic functions taking
values in the unit disc. By the corollary to Montel’s theorem, G is a
normal family. It follows that F is a normal family.

Example 3. Let

U0 = C \ {x + i0 : 0 ≤ x ≤ 1}.

Let F be all those functions holomorphic in a given domain U and

taking values in U0 . We claim that F is a normal family.
In fact, in Example 1 of Section 2 we found a univalent (one-to-
one) holomorphic map, call it µ, of U0 into the unit disc. Therefore
G = {µ ◦ f : f ∈ F} is a normal family. It follows that F is a normal
family.

With these examples we are trying to suggest that the concept of

normal families is related to Liouville’s theorem and Picard’s theorem
82 Curvature and Applications

(this notion is sometimes called “Bloch’s principle”). In fact this is true,

and we can now begin to see the assertion by looking at matters geo-
metrically. We begin with the concept of the spherical metric.
The stereographic projection p gives a map from the complex
plane C to the Riemann sphere Ĉ (see [AHL2, p. 19]). This map is il-
lustrated in Figure 1, with a sphere having radius 1 and center (0, 0, 0).

p(z)

Figure 1.

We associate to each point z in the complex plane the point p(z)

of intersection of the line through z and the “north pole” P with the
sphere.
If we equip Euclidean 3-space with coordinates as indicated in the
figure, then the map p is given explicitly by

2x 2y 2
p(x, y) = , ,1 − .
1+x +y 1+x +y
2 2 2 2 1 + x 2 + y2

It is an elegant exercise in differential calculus to check that the metric

2
σ (z) =
1 + |z|2
Normal Families and the Spherical Metric 83

on C has the following property: if z, w ∈ C, then dσ (z, w) is precisely

the distance (along a great circular arc on the surface of the sphere)
from p(z) to p(w). In fact a calculation shows that this distance is
 

 z−w 
dσ (z, w) = 2 arctan    . (∗)
1 + wz 
Differentiation of the formula (∗) along a line in the plane gives rise
to the definition of σ . (Be reassured that σ can also be thought of as
the pullback of the surface Euclidean metric τ on the sphere by the
map p. However, a precise formulation of this assertion would take us
far afield and we omit it.) As an exercise, the reader may wish to derive
the formula for dσ (z, w) by imitating our derivation of the formula
for Poincaré distance (Proposition 2 of Section 1.4): first calculate the
special case z = 0, w = r + 0i, then use the invariance of the metric
under conformal maps of the Riemann sphere  C (i.e., under suitable
linear fractional transformations).
Now suppose that in our study of normal families we study not
complex-valued holomorphic functions but meromorphic functions. A
meromorphic function is nothing other than a holomorphic function
taking values in the Riemann sphere Ĉ. In order to make this last point
clear, let m(z) be a meromorphic function on a domain U with a pole at
P ∈ U . Set I (z) = 1/z. Then I ◦ m is holomorphic in the usual sense
in a neighborhood of P. This is precisely the definition of the property
“holomorphic” for a Ĉ-valued function. Note that if the pole at P is of
order k then I ◦ m has a zero at P of order k, and conversely.
With this increased generality we can now give a more elegant
definition of “normal family.” This new definition captures the spirit of
the first one, but is strictly more general.

Definition 3
. Let U ⊆ C be a domain. Let F be a family of functions
f : U →  C. Then F is called a normal family if every sequence of
elements in F has a subsequence that converges uniformly on compact
sets to a function from U to C. [Note that, in order to make sense of
this condition, one must use the provided metric on the range, which is
the Riemann sphere.]
84 Curvature and Applications

Exercises. The point of our new definition, and you should prove this
as an exercise, is that a compactly divergent sequence (according to
Definition 2 above) is actually normally convergent to the point ∞ ∈ Ĉ.
You should note that if we specialize down to holomorphic functions,
then Hurwitz’s theorem guarantees that any family which is normal
according to Definition 3
is also normal according to Definition 3.
As a second exercise, discuss Example 1 in light of this new defi-
nition.

Now we can give Marty’s characterization of normal families in

terms of the spherical metric.

Theorem 5. Let U be a planar domain and let F be a family of Ĉ-

valued holomorphic functions on U (equivalently, meromorphic func-
tions). Let τ be the standard induced Euclidean metric on the surface
of the Riemann sphere. The family F is normal if and only if the family
of pullback metrics

{ f ∗ τ : f ∈ F}

is uniformly bounded on compact subsets of U .

An equivalent, and more concrete, way of saying this is that F is
a normal family if and only if, for each compact K ⊆ U , there is a
constant M K such that for all z ∈ K and all f ∈ F we have

2| f
(z)|
≤ MK .
1 + | f (z)|2

Proof of the Theorem. The equivalence of the two statements follows

immediately from the discussion of the spherical metric and the defi-
nition of pullback metric. We now prove the second formulation of the
theorem.
First we fix a compact set K and assume that the inequality

2| f
(z)|
≤ MK , ∀ f ∈ F, z ∈ K , (∗)
1 + | f (z)|2
Normal Families and the Spherical Metric 85

holds. Without loss of generality we may assume that K is the closure

of a connected open set V . If γ : [a, b] → V is a continuously differ-
entiable curve then the spherical length of f ◦ γ is
b
 ( f ◦ γ )
(t) τ, f ◦γ (t) dt.
a

It follows from (∗) that this quantity does not exceed

b
M K · |γ
(t)| dt.
a

But this last is just M K times the Euclidean length of γ . Notice that
this estimate is independent of f ∈ F. By our definition of distance, it
follows that F is an equicontinuous family of functions when thought
of as mapping K equipped with the Euclidean metric to Ĉ equipped
with the spherical metric. A metric-space-valued version of the As-
coli/Arzelà theorem (really just the same as the one we discussed in
Section 0.3) now yields immediately that F is a normal family.
Now we treat the converse direction. Suppose that F is a normal
family. Define the expression (the spherical derivative)

2| f
(z)|
f # (z) = .
1 + | f (z)|2

Fix a compact set K ⊆ U . Seeking a contradiction, we suppose that the

functions { f # : f ∈ F} are not uniformly bounded on K . Then there
is a sequence f j ∈ F such that the maximum of ( f j )# on K tends to
infinity with j. By the normal family hypothesis, we may suppose that
each point P ∈ U has a neighborhood N P on which { f j } converges
normally to a holomorphic function f which takes values in Ĉ. But
then ( f j )# converges normally on each N P (it is convenient here to
note that ( f j )# = (1/ f j )# ). Since K is compact, we may cover K with
finitely many of the sets N P . It follows that ( f j )# is bounded on K ,
contradicting our assumption.
86 Curvature and Applications

Remark. If g is a meromorphic function with a pole at P then the

spherical derivative g # (P) is defined to be limz→P g # (z). It is an exer-
cise for the reader to confirm that this limit will always exist.

The thrust of Marty’s theorem (and its proof) is that the condition of
being a normal family is really an equicontinuity condition on the fam-
ily, when measured in the correct metric. That metric becomes plain
when the functions are viewed as taking values in Ĉ.
In the next section we will derive two striking consequences of the
work done here.

4. A Generalization of Montel’s Theorem and

the Great Picard Theorem
Now we reap the reward of our labors. We obtain simple proofs of two
of the great results of classical function theory, theorems whose tradi-
tional proofs (involving the elliptic modular function, for instance) are
quite elaborate. Our main tools here will be the generalized Schwarz
lemma and the work in Section 3. We begin with a more general version
(also due to Paul Montel) of the Montel theorem which we discussed
earlier:

Theorem 1. Let U be a domain in the complex plane. Let P, Q, and

R be distinct elements of C. Suppose that F is a family of holomorphic
functions taking values in Ĉ \ {P, Q, R}. Then F is a normal family in
the sense of Definition 3
of Section 3.

Proof. We may apply a linear fractional transformation to arrange for

P = 0, R = ∞, and Q = 1. Thus we need to show that a family of
holomorphic functions taking values in C0,1 ≡ C \ {0, 1} is normal.
Of course it is enough to show that F is normal on any disc
D(z 0 , α) = {z : |z − z 0 | < α} ⊆ U . We may assume that z 0 = 0.
Let µ be the metric of strictly negative curvature on C0,1 con-
structed in Theorem 3 of Section 2. After multiplying µ by a positive
A Generalization of Montel’s Theorem and the Great Picard Theorem 87

constant, we may assume that the curvature is bounded from above by

−4. By Theorem 4 of Section 1 (the generalized Schwarz lemma), with
A = 4 and B = 4, we have for any f ∈ F that

f ∗ µ(z) ≤ ραA (z), ∀z ∈ D(0, α).

Now let us compare the spherical metric σ (z) restricted to C0,1

with the metric µ. Notice that σ (z)/µ(z) tends to zero as z tends to
either 0, 1, or ∞. It follows that there is a constant M > 0 such that

σ (z) ≤ M · µ(z).

Therefore we have

f # ≡ f ∗ σ ≤ M · f ∗ µ ≤ M · ραA on D(0, α).

The constant M is independent of f ∈ F. Thus f # is bounded on each

compact subset of D(0, α), with bound independent of f ∈ F. Now
Marty’s theorem implies that F is a normal family.

We isolate the case R = ∞ of the theorem as a corollary.

Corollary 1.1. If F is a family of complex-valued holomorphic func-

tions on U all of which omit the same two complex values from their
images, then F is a normal family.

We can now use this generalized Montel theorem to obtain the refine-
ment of Picard’s theorem which was promised at the end of Section 2.
This result is known as Picard’s great theorem.

Theorem 2. (Picard). Let U = D

(0, α) ≡ D(0, α) \ {0} be a punc-
tured disc and let f be a holomorphic function on U which has an es-
sential singularity at 0. Then f restricted to any deleted neighborhood
of 0 omits at most one complex value from its image.
88 Curvature and Applications

Proof. We prove the contrapositive. After rescaling, we may assume

that we have an f on D
≡ D(0, 1) \ 0 such that the image of f omits
the values 0 and 1. We will prove that 0 is therefore either a pole or a
removable singularity for f , resulting in a contradiction.
Define f n (z) = f (z/n), 0 < |z| < 1, and let F = { f n }. Since all
elements of F take values in C0,1 , the family F is normal. Because the
family F consists of holomorphic functions, we conclude that there is
a subsequence f n k which either converges normally or diverges com-
pactly.
In the first instance, the subfamily { f n k } is bounded on compact
subsets of D
. In particular, it is bounded by some constant M on the
set {z : |z| = 1/2}. But this says that f itself is bounded by M on
each of the circles {z : |z| = 1/(2n k )}. By the maximum principle, f
is bounded by M on {z : 0 < |z| < 1/2(n 1 )}. So 0 is a removable
singularity for f .
In the second case, a similar argument applied to 1/ f shows that f
tends to ∞ as z → 0. Therefore, 0 is a pole for f .
We have shown that if f omits two values in a deleted neighbor-
hood of 0, then f has either a removable singularity or a pole at 0. This
completes the proof of the contrapositive of the theorem.
CHAPTER 3
Some New Invariant Metrics

0. Introductory Remarks
Refer to Section 0.3 for the statement and sketch of the proof of the
Riemann mapping theorem. The Riemann mapping function is the so-
lution to a certain extremal problem: to find a map of the given domain
U into the disc D which is one-to-one, maps a given point P to 0, and
has derivative of greatest possible modulus λ P at P. The existence of
the extremal function, which also turns out to be one-to-one, is estab-
lished by normal families arguments; the fact that the extremal function
is onto is established by an extra argument which is in fact the only step
of the proof where the topological hypotheses about U are used.
The point of the present discussion is to observe that the scheme
we just described can be applied even if U is not topologically equiv-
alent to the disc. Constantin Carathéodory’s brilliant insight was that
the quantity λ P can be used to construct a metric, now called the
Carathéodory metric. We maximize the derivative at P of maps φ of
U into D such that φ(P) = 0 but we no longer require the maps to
be injective. Of course the proof of the Riemann mapping theorem will
break down at the stage where we attempt to show that the limit map is
surjective; we will also be unable to prove that it is injective. All other
steps, including the existence of the extremal function, are correct and
give rise to a metric (as in Section 3.1 following) on the domain.

89
90 Some New Invariant Metrics

A dual construction, using maps φ : D → U , results in a met-

ric called the Kobayashi or Kobayashi/Royden metric. Both the
Carathéodory and the Kobayashi/Royden metrics are motivated by the
extremal problem which arises from the Riemann mapping theorem,
and are of interest because they endow virtually any domain with an
invariant metric. Moreover, these metrics turn out to be invariant un-
der conformal mappings (more generally, they are distance-decreasing
under arbitrary holomorphic mappings). But the richest benefit is that
which can be obtained from studying the interaction of the two metrics.
In this chapter we learn about the Carathéodory and Kobayashi
metrics and the insights they lend to the theory of holomorphic map-
pings. We will also begin to explore what can be learned by comparing
the two metrics.
Note that in Chapter 1 we defined a “metric” to be a twice continu-
ously differentiable function with certain other special properties. This
requirement was imposed in part so that we could study curvature. For
the Carathéodory and Kobayashi metrics, we shall not examine curva-
ture. So we may suspend the smoothness hypothesis.

1. The Carathéodory Metric

Fix a domain U ⊆ C. Recall that D ⊆ C is the unit disc.

Definition 1. If P ∈ U , define

(D, U ) P = {holomorphic functions f from U to D

such that f (P) = 0}.

The Carathéodory metric for U at P is defined to be

FCU (P) = sup{|φ (P)| : φ ∈ (D, U ) P }.

Remark. As predicted in Section 0, the quantity FCU measures, for

each P, the extreme value posited in the proof of the Riemann mapping
The Carathéodory Metric 91

theorem. In that proof, it was necessary only to know that the extreme
value existed and was finite. Now we shall gain extra information by
comparing this value to other quantities.

Clearly FCU (P) ≥ 0. (Moreover the Cauchy estimates imply that

FCU (P) < ∞.) Is FCU (P) > 0 for all P? If U is bounded, then the
answer is “yes.” For then

U ⊆ D(0, R) = {z ∈ C : |z| < R}

for some R > 0. But then the map

ζ−P
φ : ζ −→
2R
satisfies φ ∈ (D, U ) P . Thus

1
FCU (P) ≥ |φ (P)| = > 0.
2R

However, if U is unbounded, then FCU may degenerate. If U = C,

for instance, then any f ∈ (D, U ) P is constant, hence FCU ≡ 0. The
same holds for U equaling C less a discrete set of points (by the Rie-
mann removable singularities theorem). Which domains have nonde-
generate (i.e., non-vanishing) Carathéodory metric (i.e., Carathéodory
metric which actually gives a distance)? Analytic capacity is a device
for answering this question (see [GAR]), but we cannot pursue that
topic here.
The primary interest, at least right at the moment, of the
Carathéodory metric is that it generalizes Proposition 1 in Section 1.5
to arbitrary domains.

Proposition 2. Let U1 , U2 be domains in C. Let ρ j be the Cara-

théodory metric on U j . If h : U1 → U2 is holomorphic, then h is
distance-decreasing from (U1 , ρ1 ) to (U2 , ρ2 ). In other words,

h ∗ ρ2 (z) ≤ ρ1 (z), ∀z ∈ U1 .
92 Some New Invariant Metrics

Corollary 2.1. If γ : [0, 1] → U1 is a piecewise continuously differ-

entiable curve, then

ρ2 (h ∗ γ ) ≤ ρ1 (γ ).

Remark. Observe that the Carathéodory metric is not manifestly inte-

grable on curves. But if one thinks about the fact that it is constructed
as the supremum of continuous functions, then it follows that the met-
ric is semicontinuous (see [KR1]). So it is indeed integrable on curves
(to wit, a semi-continuous function is the monotone limit of continuous
functions).

Corollary 2.2. If P1 , P2 ∈ U1 then

dρ2 (h(P1 ), h(P2 )) ≤ dρ1 (P1 , P2 ).

Corollary 2.3. If h is conformal, then h is an isometry of (U1 , ρ1 ) to

(U2 , ρ2 ).

The corollaries are proved just as they were proved for the Poincaré
metric in Chapter 1. We concentrate on proving the Proposition.

Proof of Proposition 2. Fix P ∈ U1 and set Q = h(P). Notice that if

φ ∈ (D, U2 ) Q , then φ ◦ h ∈ (D, U1 ) P . Thus

FCU1 (P) ≥ |(φ ◦ h) (P)|

= |φ (Q)| · |h (P)|.

Taking the supremum over all φ ∈ (D, U2 ) Q yields

FC 1 (P) ≥ FC 2 (Q) · |h (P)|

U U

or

ρ1 (P) ≥ h ∗ ρ2 (P).
The Kobayashi Metric 93

Next we calculate the Carathéodory metric for the disc.

Proposition 3. The Carathéodory metric on the disc coincides with the

Poincaré metric.

Proof. First we calculate the metric at the origin. If φ ∈ (D, D)0 then,
by the Schwarz lemma, |φ (0)| ≤ 1. Therefore

FCD (0) ≤ 1.

But the map

φ(ζ ) = ζ

satisfies φ ∈ (D, D)0 and φ (0) = 1. Thus

FCD (0) = 1.

Because every conformal map of the disc is an isometry of the

Carathéodory metric, Proposition 6 of Section 1.4 now implies that the
Carathéodory metric and the Poincaré metric are equal.

Given what we have learned so far, we might suspect that any

isometry of the Carathéodory metric must be a conformal map. This
is indeed true—in fact a much stronger assertion holds—provided
the domains in question have nondegenerate (i.e., non-vanishing)
Carathéodory metrics. We postpone this topic until after we have in-
troduced the Kobayashi metric.

2. The Kobayashi Metric

Fix a domain U ⊆ C.

Definition 1. If P ∈ U , define

(U, D) P = {holomorphic functions f from D

to U such that f (0) = P}.
94 Some New Invariant Metrics

The Kobayashi (or Kobayashi/Royden) metric for U at P is defined to

be


1
FKU (P) = inf : φ ∈ (U, D) P
.
|φ (0)|

Remark. As we remarked in the last section about the Carathéodory

metric, the Kobayashi metric quantifies a certain extremal problem and
will be useful for comparison purposes. The particular form of the def-
inition of the Kobayashi metric will turn out to facilitate comparisons
with the Carathéodory metric; in particular, the two metrics will turn
out to interact nicely with the classical Schwarz lemma.

Clearly FKU (P) ≥ 0. In order to learn more about FKU , we now

compare it with the Carathéodory metric.

Proposition 2. For all P ∈ U ,

FCU (P) ≤ FKU (P).

Proof. Let φ ∈ (D, U ) P and ψ ∈ (U, D) P . Then φ ◦ ψ : D → D and

φ ◦ ψ(0) = 0. It follows from Schwarz’s lemma that

|(φ ◦ ψ) (0)| ≤ 1

or
1
|φ (P)| ≤ .
|ψ (0)|
Taking the supremum over all φ gives
1
FCU (P) ≤ .
|ψ (0)|
Now taking the infimum over all ψ yields

FCU (P) ≤ FKU (P).

The Kobayashi Metric 95

An immediate consequence of this last result is that if U is

bounded, then FKU is nondegenerate, i.e., non-vanishing (for then FCU
is). On the other hand, if U = C, then FKU ≡ 0; for given P ∈ C the
function

φ R (ζ ) = P + Rζ

satisfies φ R ∈ (U, D) P , any R > 0. Thus

1 1
FKU (P) ≤ = ;
|φ R (0)| R

letting R → ∞, we have FKU (P) = 0.

In analogy with Proposition 2 of the last section we now have:

Proposition 3. Let U1 , U2 be domains equipped with the Kobayashi

metric (denoted, respectively, by ρ1 and ρ2 ). If h : U1 → U2 is holo-
morphic, then h is distance-decreasing from (U1 , ρ1 ) to (U2 , ρ2 ). That
is,

h ∗ ρ2 (z) ≤ ρ1 (z), ∀z ∈ U1 .

Corollary 3.1. If γ : [0, 1] → U1 is a piecewise continuously differ-

entiable curve, then

ρ2 (h ∗ γ ) ≤ ρ1 (γ ).

Remark. Just as for the Carathéodory metric, the Kobayashi metric

will be integrable on curves just because it is semicontinuous. Again
see [KR1].

Corollary 3.2. If P1 , P2 ∈ U1 , then

dρ2 (h(P1 ), h(P2 )) ≤ dρ1 (P1 , P2 ).

Corollary 3.3. If h is conformal, then h is an isometry of (U1 , ρ1 ) to

(U2 , ρ2 ).
96 Some New Invariant Metrics

Proof of Proposition 3. Fix P ∈ U1 and put Q = h(P). Choose φ ∈

(U1 , D) P . Then h ◦ φ ∈ (U2 , D) Q . We have

1
FKU2 (Q) ≤
|(h ◦ φ) (0)|
1
= .
|h (P)||φ (0)|

Taking the infimum over all φ ∈ (U1 , D) P gives

1
FKU2 (Q) ≤ · FKU1 (P)
|h (P)|
or

(h ∗ FK 2 )(P) ≤ FK 1 (P).
U U

Using Proposition 3, we may give another example of domains

U ⊆ C with degenerate Kobayashi metric. Let C0 = C \ {0}. Then

h : C → C0
ζ → eζ

is holomorphic. The map h will be distance-decreasing in the

Kobayashi metric. Since FK is identically 0 on C (because there are
arbitrarily large analytic discs in C centered at any point), it follows
that FK is identically zero on C0 .
As an exercise, prove that we cannot continue this line of reason-
ing to see that FKU is degenerate when U = C \ {0, 1}. In fact, we shall
later see that FKU is nondegenerate on U ≡ C \ {P1 , . . . , Pk } provided
the P j ’s are distinct and k ≥ 2 (notice that this situation is in contrast to
that for the Carathéodory metric as discussed in Section 1). The matter
will be addressed in Section 5; there we will see it connected in an el-
egant way with normal families—recall the special role of C0,1 in that
theory. Meanwhile, we turn our attention to more elementary questions
regarding the Kobayashi metric.
The Kobayashi Metric 97

Proposition 4. For U = D ⊆ C, the Kobayashi metric equals the

Poincaré metric.

Proof. For P ∈ D we have

FKD (P) ≥ FCD (P) = ρ(P),

where ρ is the Poincaré metric on the disc. For an inequality in the

opposite direction, first take P = 0. Let φ ∈ (D, D)0 be given by

φ(ζ ) = ζ.

Then
1
FKD (0) ≤ = 1 = ρ(0).
|φ (0)|
It follows that

FKD (0) = 1 = ρ(0).

By Corollary 3.3 above and Proposition 6 of Section 1.4, we conclude

that

FKD (P) = ρ(P), ∀P ∈ D.

The first major result of this chapter is a characterization of the

disc in terms of the Carathéodory and Kobayashi metrics (at this point
have another look at Section 0 to see why this result is a metric version
of the Riemann mapping theorem):

Theorem 5. Let U ⊆ C be a domain. The domain U is conformally

equivalent to the disc if and only if there is a point P ∈ U such that

FCU (P) = FKU (P) = 0.

Proof. If U is conformally equivalent to the disc then let

h:U → D
98 Some New Invariant Metrics

be a conformal map. For any point P ∈ U we have that

FCU (P) = (h ∗ FCD )(P)

which, by Proposition 3 of Section 1,

= (h ∗ ρ)(P);

but by Proposition 4 this

= (h ∗ FKD )(P).

By Corollary 3 to Proposition 3 we conclude that this last equals

FKU (P), proving the easy half of the theorem.
For the converse, choose φ j ∈ (D, U ) P such that

|φ j (P)| → FCU (P)

and choose ψ j ∈ (U, D) P such that

1
→ FKU (P).
|ψ j (0)|

Since {φ j } is uniformly bounded above by 1, we may extract a

subsequence {φ jk } converging to a normal limit φ0 . Consider

h jk = φ jk ◦ ψ jk .

Passing to another subsequence, which we denote by h j , we may sup-

pose that h j converges normally to a limit h 0 . Notice that h 0 (0) = 0 so
that h 0 is not a unimodular constant; therefore h 0 maps D to D. After
renumbering, we call this last sequence

h  = φ ◦ ψ .

Recall that when a sequence of holomorphic functions converges nor-

mally, then so does the sequence of its derivatives (see Corollary 5.1 in
The Kobayashi Metric 99

Section 0.1). It follows that

 
|h 0 (0)| = lim (φ ◦ ψ ) (0)
→∞

= lim |φ (P)| · |ψ (0)|

→∞
1
= FCU (P) ·
FKU (P)
= 1.

By Schwarz’s lemma, h 0 (ζ ) = µ · ζ for some unimodular constant µ.

Thus we have

µ · ζ = h 0 (ζ ) = lim (φ ◦ ψ (ζ )). (∗)

→∞

After composing the functions φ with a rotation, we may assume that

µ = 1.
Now C \ U must contain at least two points (else FKU would be
identically 0, contradicting our hypothesis). Hence, by Theorem 1 of
Section 2.4, {ψ } forms a normal family. Say that a subsequence ψm
converges to some ψ0 . After renumbering, we rewrite (∗) as

ζ = h 0 (ζ )
= lim (φ ◦ ψ (ζ )) (∗∗)
→∞
= φ0 ◦ ψ0 (ζ ).

Since h 0 is surjective, we conclude that φ0 is surjective.

Because of (∗∗), the function ψ0 is a nonconstant holomorphic
function; hence its image is open. We claim that the image is also closed
(in the relative topology of U ). To see this, let ψ0 (ζ j ) be elements of
that image converging to a limit point q ∈ U . Since φ0 is continuous, it
follows that φ0 (ψ0 (ζ j )) converges to a limit point r . But then (∗∗) tells
us that ζ j → r . It follows from the continuity of ψ0 that ψ0 (r ) = q,
so q is in the image of ψ0 . We thus have that the image of ψ0 is both
open and closed, and it is nonempty. Since U is connected, it follows
that the image of ψ0 equals U , hence ψ0 is surjective. Thus, since h 0
100 Some New Invariant Metrics

is injective, so must φ0 be injective. We conclude that φ0 is the desired

conformal map of U to D.

Notice that the conclusion of the theorem is false if we allow the

metrics to degenerate. For example, if U = C \ 0, then FCU = FKU ≡ 0,
yet U is definitely not conformally equivalent to the disc.
As anticipated in the last section, we now prove that the only
isometries of the Carathéodory or Kobayashi metrics which fix a point
are conformal maps. In fact we shall prove something much stronger.

Theorem 6. Let U ⊆ C be a domain on which the Kobayashi metric is

nondegenerate (i.e., gives a genuine distance function) and fix P ∈ U .
Suppose that

f : U −→ U

is a holomorphic function such that f (P) = P. Assume that f is an

isometry of either the Carathéodory or the Kobayashi metric at P; that
is, suppose that either

f ∗ FCU (P) = FCU (P)

or

f ∗ FKU (P) = FKU (P)

and that the metric does not vanish at P. Then f is a conformal map
of U onto U .

Remark. This theorem is a remarkable rigidity statement: the global

condition that f map U to U , together with the differential condition
at P, forces f to be one-to-one and onto.

Proof of the Theorem. Since U has nondegenerate Kobayashi metric,

C \ U must contain at least two points. This implies that any family
of holomorphic functions {gα } taking values in U is normal (by Theo-
rem 1 of Section 2.4). We will use this observation repeatedly.
The Kobayashi Metric 101

Now the hypothesis that

f ∗ FCU (P) = FCU (P) or f ∗ FKU (P) = FKU (P)

implies that

| f (P)| = 1.

Define

f1 = f
f2 = f ◦ f
···
f j
= f j−1
◦ f, j ≥ 2.

Then { f j } is a normal family. Since the numbers ( f j ) (P) all have unit
modulus (note that ( f j ) (P) = [ f (P)] j ), there is a subsequence { f  }
j

j
such that ( f ) (P) → 1. (Exercise: if f (P) has argument which is
a rational multiple of π, then the assertion is clear; if it is an irrational
multiple of π, then the set of ( f j ) (P) forms a dense subset of the
circle.) Passing to another subsequence, which we still denote by { f  },
j

we may also suppose that f  converges normally to a holomorphic

j

function f˜. It follows that f˜ (P) = 1. We claim in fact that f˜(z) ≡ z.

To prove this assertion, suppose not. Assume for simplicity that
P = 0. Then, for z near 0,

f˜(z) = 0 + z + (higher order terms).

Let am z m , m ≥ 2, be the first nonvanishing power of z which appears

after z (since we are assuming that  f is not the function z, there must
be one). Notice that
f˜2 ≡ f˜ ◦ f˜ = z + 2am z m + · · ·
f˜3 ≡ f˜ ◦ f˜ ◦ f˜ = z + 3am z m + · · ·
(∗)
···
f˜k ≡ f˜ ◦ f˜ ◦ · · · ◦ f˜ = z + kam z m + · · ·
102 Some New Invariant Metrics

On the other hand, { 

f k } forms a normal family, so there is a subse-
quence f˜ converging normally on U . Hence
k q

 m
∂
f˜kq (P)
∂z

converges. But (∗) clearly shows that

 m
∂
f˜k (P) = m! · k · am
∂z

blows up with k. This contradiction can only be resolved if am = 0.

Therefore f˜(z) ≡ z.
Now we have that f j (z) → z normally. We claim that this im-
plies that f is a conformal map. Indeed, consider the family { f j −1 }. It
is normal, so there is a subsequence, call it { f jr −1 }, which converges
normally to a limit g. Notice that g is not constant since g (P) is not
zero. Then
jr −1
f ◦ g(z) = f ◦ lim f (z)
r →∞
j
= lim f r (z)
r →∞
≡ z.

Similarly,

g ◦ f (z) ≡ z.

Thus f is one-to-one and onto. It is certainly holomorphic, so it is a

conformal map as claimed.

It is instructive to view Theorem 6 as a generalization of the

uniqueness part of the Schwarz lemma. For simplicity take U to be
a bounded domain. Then U is contained in some disc, and a disc has
nondegenerate Kobayashi metric. Therefore, by the distance decreas-
ing property of the Kobayashi metric applied to the inclusion map, the
domain U also has nondegenerate Kobayashi metric. Suppose that f
Completeness of the Carathéodory and Kobayashi Metrics 103

is a holomorphic function mapping U to U that fixes a point P in U .

The distance-decreasing property of the Kobayashi metric tells us that
| f (P)| ≤ 1. Theorem 6 now says that equality obtains if and only if f
is one-to-one and onto.

3. Completeness of the Carathéodory

and Kobayashi Metrics
In this section we prove that if U is a domain with a reasonably nice
boundary then it is complete, as a metric space, when equipped with
either the Carathéodory or Kobayashi metric. A warning is in order
here: this is simultaneously the richest and the most difficult section
of the book; for it uses both the metric language and tricky construc-
tions from analysis. But the section is self-contained, and provides an
introduction to a lot of important and beautiful material. Moreover, the
payoff for the work, in both this section and the next, is more than ad-
equate compensation for the effort expended. As a bonus, we will give
a nice application of the ideas (in the context of Bergman geometry) in
Section 4.6.
We begin by noting that, on compact sets of a bounded domain,
all of our metrics are comparable:

Proposition 1. Let  ⊆ C be a bounded domain. Let L ⊆  be a

compact subset. Let ρC , ρ K , and ρ E be the Carathéodory, Kobayashi,
and Euclidean metrics on  respectively. Then there are constants
C1 , C2 , C3 , C4 such that, for z ∈ L,
ρC (z)
C1 ≤ ≤ C2
ρ E (z)
and
ρ K (z)
C3 ≤ ≤ C4 .
ρ E (z)
[We stress that these four constants certainly depend both on L and
on .]
104 Some New Invariant Metrics

Proof. We will prove the second set of inequalities. The proof of the
first inequalities is similar.
Now let P be a point in . Let r be a positive number that is
less than one third of the Euclidean distance of P to the boundary. Let
D(P, r ) ⊆ D(P, 2r ) be closed discs in . If z ∈ D(P, r ) then
D(P,2r )
ρK (z) ≥ ρ K (z).

Here we use a superscript on ρ to indicate with respect to which domain

the metric is being computed.
On the other hand,

D(P,2r ) 1 D(0,1)
ρK (z) = · ρK ((z − P)/[2r ]).
2r
And of course we know the Kobayashi metric on the unit disc explicitly.
D(0,1)
It equals the Poincaré metric. Hence ρ K ((z − P)/[2r ]) ≤ 4/3.
Putting our estimates together, we find that

4 1
ρ K (z) ≤ · .
3 2r
This is one half of our estimate.
For the other half, notice that since  is bounded it is contained in
some large disc open D(0, R). Then, for z ∈ D(P, r ),

ρ K (z) ≥ ρ K
D(0,R)
(z).
D(0,1)
Now, by conformal mapping, this last is equal to [1/R]ρ K (z/[R]).
Thus it is at least equal to 1/R.
In summary, we have proved for points z ∈ D(P, r ) that

4 1
≥ ρ K (z) ≥ .
6r R
The compact set L ⊂  can be covered by finitely many discs of the
form D(P, r ). So, in the end, we obtain a uniform estimate

C3 ≤ ρ K (z) ≤ C4
Completeness of the Carathéodory and Kobayashi Metrics 105

for all z ∈ L. But the Euclidean metric is constantly equal to 1. So this

last line just says

C3 ρ E (z) ≤ ρ K (z) ≤ C4 ρ E (z),

and that is what we wished to prove.

Corollary 1.1. Let  ⊆ C be a bounded domain. The topology in-

duced by any of the Carathéodory, Kobayashi, or Euclidean metrics is
equivalent to the other two.

Proof. Since the metrics are comparable, the induced balls are compa-
rable. The balls form a sub-basis for the topology. So the topologies are
comparable.

Now we turn to the treatment of completeness of our two new met-

rics on a reasonable class of domains. It is enough to prove complete-
ness for the Carathéodory metric, since any sequence which is Cauchy
in the Kobayashi metric is also Cauchy in the Carathéodory metric (ex-
ercise: provide the details using Proposition 2 of Section 2). We begin
by defining what we mean by “reasonably nice boundary.”

Definition 2. A curve γ : [a, b] → C is called a closed, twice con-

tinuously differentiable curve if γ is twice continuously differentiable
(with one-sided derivatives as usual at the endpoints), if γ̇ is never 0,
if γ (a) = γ (b), and if the one-sided derivatives of γ , up to and in-
cluding order 2, at a equal those at b. We write “γ is C 2 and closed.”
More generally, we may define γ to be closed, k times continuously
differentiable (for k ≥ 1) if γ is k times continuously differentiable,
(with one-sided derivatives as usual at the endpoints), if γ̇ is never 0, if
γ (a) = γ (b), and if the one sided derivatives of γ , up to and including
order k, at a equal those at b. We write “γ is C k and closed.”

Definition 3. A domain is said to have C k boundary if the boundary

consists of finitely many C k , closed curves.
106 Some New Invariant Metrics

In case a function is C k for every k, or a boundary is C k for ev-

ery k, then we say that the function or boundary is C ∞ or infinitely
differentiable.
In geometric analysis, an alternative definition of C k boundary
often proves to be convenient. We think of a domain U with C k (resp.
C k ) boundary as one which can be specified by

U = {z ∈ C : ρ(z) < 0},

with ρ a C k function such that ∇ρ = 0 on ∂U . We call ρ a defining

function for U . For example, the disc may be specified in this way as

D = {z ∈ C : ρ(z) = |z|2 − 1 < 0}.

A bounded domain with infinitely many holes will not have C k bound-
ary, k ≥ 1, since any defining function will fail to be smooth at an
accumulation point of the holes. See [KRP2] for a thorough treatment
of defining functions.

Example 1. The curve γ (t) = eit , 0 ≤ t ≤ 2π, is a closed, twice con-

tinuously differentiable curve (indeed, it is C k for every k, or infinitely
differentiable). See Figure 1. Notice that the agreement of the first and
second derivatives at the endpoints causes a smooth transition in the
figure at the point 1 + i0 where the two ends meet.

Figure 1.
Completeness of the Carathéodory and Kobayashi Metrics 107

The curve
π π
µ(t) = (cos 2t)(cos t) + i(cos 2t)(sin t), − ≤t≤
4 4

is closed, but not twice continuously differentiable (because the deriva-

tives at t = −π/4 and t = π/4 do not agree). The curve is shown in
Figure 2.

Figure 2.

Figure 3 exhibits a domain with C k boundary. Note that there are

finitely many boundary curves, and each is k times continuously dif-
ferentiable.

Figure 3.
108 Some New Invariant Metrics

Observe that, by the usual constructions of multi-variable calcu-

lus, a domain with twice continuously differentiable boundary pos-
sesses at each boundary point P a well defined unit outward normal
ν P and a well defined unit inward normal ν P . See Figure 4.

Figure 4.

The functions

∂U  P −→ ν P ∈ R2

and

∂U  P −→ ν P ∈ R2

are then continuously differentiable.

We shall need two geometric/analytic results about domains with
2
C boundary. They are formulated as Propositions 4 and 5.

Proposition 4. If U is a domain with C 2 boundary, then there is an

open neighborhood W of ∂U such that if z ∈ U ∩ W , then there is
a unique point P = P(z) ∈ ∂U which is nearest (in the sense of
Euclidean distance) to z. In other words,

inf{|z − Q| : Q ∈ ∂U } = |z − P|.

We call W a tubular neighborhood of ∂U (see [MUN] for more on

these matters). See Figure 5.
Completeness of the Carathéodory and Kobayashi Metrics 109

Figure 5.

Proof. Define

T : ∂U × (−1, 1) −→ C
(Q, t) −→ Q + tν Q .

Think of T as a map from the two-dimensional real space ∂U ×(−1, 1)

to the two-dimensional real space C ≈ R2 . See Figure 6. [In fact the
reader may find it convenient to consider this alternative definition of T .
Let γ : [0, 1] → C be a C 2 parametrization of ∂U . Then set

T : [0, 1] × (−1, 1) −→ C
(s, t) −→ γ (s) + tνγ (s) .]

-1

Figure 6.
110 Some New Invariant Metrics

Fix a point Q 0 ∈ ∂U . Assume without loss of generality that the

tangent line to ∂U at Q 0 is horizontal (this may be achieved by rotating
the coordinate system). We may also normalize the tangent vector to be
of unit length.
Then the Jacobian matrix of T at (Q 0 , 0) is
 
1 0
0 1

which is invertible. By the inverse function theorem (see [RU1] or

[MUN] or [KRP3]), there is a neighborhood V0 of (Q 0 , 0) in ∂U ×
(−1, 1) on which T is invertible. Then T (V0 ) ≡ W Q 0 is a neighbor-
hood of Q 0 with the property that if z ∈ W Q 0 , then there is a unique
point Q ∈ ∂U and a unique t ∈ (−1, 1) such that

z = Q + tν Q .

It follows that Q is (locally) the unique nearest point in ∂U to z and the

distance of z to Q is |t|. We set

W = WQ0
Q 0 ∈∂U

and we are finished.

Remark. The hypothesis of C 2 boundary was needed in order to ap-

ply the inverse function theorem. Explain why. The analogous result is
false as soon as the boundary smoothness is less than C 2 (see [KRP4]).

Proposition 5. Let U ⊆ C be a bounded domain with C 2 boundary.

Then there is an r0 > 0 such that for each P ∈ ∂U there is a disc
D(C(P), r0 ) of radius r0 which is externally tangent to ∂U at P. There
is also a disc D(C (P), r0 ) which is internally tangent to ∂U at P.
These discs have the further property that D(C(P), r0 ) ∩ ∂U = {P}
and D(C (P), r0 ) ∩ ∂U = {P}.
Completeness of the Carathéodory and Kobayashi Metrics 111

Figure 7.

Remark. Refer to Figure 7 and notice that a disc d is internally tangent

to ∂U at P if d ⊆ U , d  P, and ∂d and ∂U have the same tangent
line at P.
A similar remark applies to externally tangent discs.

Proof of Proposition 5. This follows from basic calculus: To each

point P ∈ ∂U there corresponds a radius of curvature r (P) and a
center of curvature c(P) (here we are using the word “curvature” in
the classical Euclidean sense). The point c(P) is r (P) units along the
principal normal from P.
The corresponding circle is called the circle of curvature. See
[THO] for details. If the circle of curvature is inside the domain,
take C (P) = c(P) and take C(P) to be the reflected point in ∂U .
See Figure 8. If the circle of curvature is outside the domain, take
C(P) = c(P) and take C (P) to be a reflected point in ∂U .
Now, referring to [THO], we see that r (P) depends on the second
derivative of the boundary curve γ . In particular r (P) is a continuous
function of P. We take r0 to be the (positive) minimum of r . We would
like to think that we are finished.
Unfortunately, in spite of our best intentions, we may have a situ-
ation as in Figure 9 or 10: the circle of curvature has the right behavior
at P, but it does not take into account the global behavior of U . As a
result, it may be neither completely internal to U nor completely ex-
112 Some New Invariant Metrics

Figure 8.

Figure 9.

Figure 10.
Completeness of the Carathéodory and Kobayashi Metrics 113

ternal to U . We therefore make the following adjustment. Let W be a

tubular neighborhood of ∂U and choose  > 0 such that if z ∈ C has
Euclidean distance less than  from ∂U , then z is in W .
Define

r ∗ (P) = min{/2, r (P)}.

This is a positive, continuous function of P ∈ ∂U . Hence there is a

number r0 > 0 such that

r ∗ (P) > r0 ∀P ∈ ∂U.

This is the r0 we seek. We redefine

C(P) = P + r0 ν P and C (P) = P + r0 ν P .

The definition of tubular neighborhood guarantees that a disc with cen-

ter C(P) or C (P) and radius r0 can intersect ∂U only at the point P.

We return to our discussion of metrics. The main result of this

section is the following.

Theorem 6. If U ⊆ C is a bounded domain with C 2 boundary, then U

is complete in the Carathéodory metric.

Corollary 6.1. The domain U is also complete in the Kobayashi met-

ric.

Proof of the Theorem. Let z ∈ U be in a tubular neighborhood W of

∂U , as provided by Proposition 4. Let P be the nearest boundary point
to z and D(C(P), r0 ) the externally tangent disc provided by Proposi-
tion 5. The map

iP : U −→ D(C(P), r0 )
(r0 )2
ζ −→ C(P) +
ζ − C(P)
114 Some New Invariant Metrics

is holomorphic and inverts U into the disc D(C(P), r0 ). The map

j P : D(C(P), r0 ) −→ D(0, 1)
ζ − C(P)
ζ −→
r0
is holomorphic. To estimate the Carathéodory metric at z, we use the
distance-decreasing property of the Carathéodory metric:

FCU (z) ≥ (j P ◦ i P )∗ FC
D(0,1)
(z)
(∗)
≡ |(j P ◦ i P ) (z)|FC
D(0,1)
(j P (i P (z))).

We now estimate the various terms in this last expression.

First,
 
   
(j P ◦ i P ) (z) =  1 · (i P ) (z)
r 
0
 
1 (r0 )2 
 
= · 
 r0 (z − C(P))2 
  (∗∗)
 r0 
=  
(z − C(P))  2
 
 r0 
=  ,
(δ + r ) 
0
2

where δ is the Euclidean distance of z to P. See Figure 11.

Also


(r0 )2
j P ◦ i P (z) = j P + C(P)
z − C(P)
r0
=
z − C(P)
hence
r0 δ
|j P ◦ i P (z)| = =1− .
δ + r0 r0 + δ
Completeness of the Carathéodory and Kobayashi Metrics 115

z P C(P)
r0

Figure 11.

Recall that, on the unit disc, the Carathéodory metric coincides with the
Poincaré metric. It follows from our calculation of the Poincaré metric
in Proposition 2 of Section 1.4 that

D(0,1) 1
FC (j P ◦ i P (z)) =
(δ/(r0 + δ))(2 − δ/(r0 + δ))
r0 + δ
≥ (∗ ∗ ∗)
2δ
r0 1
≥ · .
2 δ
In summary, using (∗), (∗∗), and (∗ ∗ ∗), we have
r0 r0 1 1
FCU (z) ≥ · · ≥ C0 · ,
(δ + r0 )2 2 δ δ
where C0 is a positive constant which depends only on r0 .
But this is precisely the estimate which enabled us, in Section 1.4,
to prove that the disc is complete in the Poincaré metric. We leave
it now as an exercise to provide the details which show that the
Carathéodory distance from any fixed interior point P0 ∈ U to a point
with Euclidean distance δ from the boundary has size C ·(1+| log 1/δ|)
and to conclude that U is complete in the Carathéodory metric.
116 Some New Invariant Metrics

Exercise. Exploit the internally tangent disc at each boundary point to

prove that there is a constant C1 such that
1
FCU (z) ≤ C1 · .
δ
(Hint: Use the inclusion map from the internal disc to the domain, to-
gether with the distance-decreasing property of the Carathéodory met-
ric.)

Exercise. Let

U = D \ {0}

be the punctured disc. See Figure 12. This domain does not have
C 2 boundary. Use the Riemann removable singularities theorem and
Cauchy estimates to determine the behavior of the Carathéodory met-
ric near the boundary point 0 of U . Conclude that U is not complete in
the Carathéodory metric. What can you say about the Kobayashi met-
ric?

Figure 12.

Corollary 6.2. Let U be any bounded, finitely connected region in C

(that is, the complement of U has finitely many connected components).
Further assume that each boundary component of U is a Jordan curve.
Then the Carathéodory metric on U is complete.
Completeness of the Carathéodory and Kobayashi Metrics 117

Proof. Since completeness of the Carathéodory metric is a conformal

invariant, it is enough to show that U is conformally equivalent to a
domain U  having the property that ∂ U  has C 2 boundary.
To achieve this end, imagine a planar region with a single hole.
Assume that the outer boundary is a Jordan curve, and the boundary
of the hole is also a Jordan curve. Fill in the hole to obtain a simply
connected region U , and use the Riemann mapping theorem to map
U to the disc. Now unfill the hole. See Figure 13. We now have a
region U with a circle as the outer boundary and a Jordan curve as the
inner boundary.

Figure 13.

Let A be the bounded component of the complement of U and

let p be a point in the interior of A. Do an inversion in p (i.e., z →
1/(z − p)). This maps ∂ A to the outer boundary of a new region, and
118 Some New Invariant Metrics

the former (outer) boundary of U to the boundary of an interior hole

(again see Figure 13). Now fill in the new hole and use the Riemann
mapping theorem to map the (filled in) new region to a disc. Finally,
perform another inversion (Figure 13). Thus the original hole (that is,
A) has been converted to a smooth, circular hole.
If ∂U has k components then k successive applications of in-
version and the Riemann mapping theorem, as above, give rise to a
conformal mapping of U to a region consisting of the unit disc with
k − 1 smoothly bounded regions removed. That completes the proof.
See [AHL2] for further details.

We now use the completeness of the Carathéodory metric to prove

a version of the Lindelöf principle using the geometric point of view.
First we need some terminology.
Let U be a domain with C 2 boundary. Choose a point P ∈ ∂U
and let ν P be the unit outward normal. If f is a continuous function on
U , we say that f has radial boundary limit  at P provided that

lim f (P − r ν P ) = .
r →0+

If α > 1, set

α (P) = {z ∈ U : |z − P| < α · dist (z, ∂U )}.

Here the expression dist(z, ∂U ) denotes Euclidean distance of z to ∂U .

We call the domain α a Stolz region or nontangential approach re-
gion at the point P. See Figure 14. We say that the function f has
nontangential limit  at P if for each α > 1 we have

lim f (z) = .
α (P)z→P

Clearly the possession of a nontangential limit at P is a stronger condi-

tion than the possession of a radial limit at P. For example, on the unit
disc the continuous function
y
f (z) =
1−x
Completeness of the Carathéodory and Kobayashi Metrics 119

(P)

Figure 14.

has radial limit 0 at the boundary point P = 1 + i0. However, it does

not possess a nontangential limit at P (exercise—try P j = (1 − 1/j)
+ i/j for j = 1, 2, . . . ). It is therefore surprising that the Lindelöf prin-
ciple tells us that for bounded holomorphic functions the two notions
of boundary limit coincide:

Theorem 7. Let U1 , U2 be bounded domains with C 2 boundary and

let

f : U1 −→ U2

be holomorphic. If P ∈ ∂U1 , Q ∈ ∂U2 , and f has radial limit Q at P,

then f has nontangential limit Q at P.

Proof. If z is an element of one of our domains U j and if s > 0 then

we let B(z, s) denote the metric ball with center z and radius s in the
Carathéodory metric for U j . For P ∈ ∂U1 , r0 > 0, and β > 0 fixed,
we define

Mβ (P) = B(P − r ν P , β).
0<r <r0

The estimate
C
FCU (z) ≈ (∗)
dist(z, ∂U )
120 Some New Invariant Metrics

makes it a tedious but not difficult exercise to calculate that the regions
Mβ are comparable to the regions α (in this last formula, and in what
follows, “dist” means Euclidean distance). In point of fact, suppose
that z lies in some α ( p), some p ∈ ∂U j . Let us denote by τ p the
inward normal segment emanating from that boundary point p. Using
the estimate (∗), one can then estimate that distC (z, τ p ) ≤ C · α. For
the converse estimate, assume that z ∈ α ( p) and the same estimate
shows that distC (z, τ p ) ≥ C · α.
Thus we see that

lim f (z) = , ∀α > 1

α (P)z∈P

iff

lim f (z) = , ∀β > 0. (∗∗)

Mβ (P)z→P

Thus it is enough to prove (∗∗).

Since

Mβ (P) = B(P − r ν P , β),
0<r <r0

the distance-decreasing property of f with respect to the Carathéodory

metric implies that

f (Mβ (P)) ⊆ B( f (P − r ν P ), β).
0<r <r0

Pick  > 0. By the radial limit hypothesis, there is a δ > 0 such

that if 0 < t < δ, then

| f (P − tν P ) − Q| < .

For such a t, if z ∈ B(P − tν P , β), then

f (z) ∈ B( f (P − tν P ), β).

But

dist( f (P − tν P ), ∂U2 ) ≤ dist( f (P − tν P ), Q) < .

An Application of Completeness: Automorphisms 121

Therefore the estimate (∗) implies that the metric ball B( f (P −

tν P ), β) has Euclidean radius not exceeding C · . Here C depends
on β, but β has been fixed once and for all. Thus

| f (z) − f (P − tν P )| < C, ∀z ∈ B(P − tν P , β).

We conclude that

| f (z) − Q| ≤ | f (z) − f (P − tν P )| + | f (P − tν P ) − Q|
≤ C +  = C .

This is the desired conclusion.

It is not difficult to see that nontangential approach is the broad-

est possible method for calculating boundary limits of bounded holo-
morphic functions. See [PRI] for explicit constructions of examples to
show that the nontangential approach is the sharp condition. This ref-
erence also contains a complete account of the positive results, as well
as the history, of the theory of boundary behavior of holomorphic func-
tions.

4. An Application of Completeness:
Automorphisms
This section is an introduction to a topic which is usually treated as
part of Riemann surface theory. (See [FK] for a nice treatment of the
Riemann surface point of view.) We shall not use any of this theory,
nor shall we make any reference to Riemann surfaces. Instead, we use
geometry.

Definition 1. Let U ⊆ C be a domain. We let Aut(U ) denote the fam-

ily of conformal self-maps of U (i.e., one-to-one, onto holomorphic
functions from U to U ).
122 Some New Invariant Metrics

Proposition 2. With the binary operation being the composition of

functions, the set Aut(U ) forms a group. We call this group the au-
tomorphism group of U .

Proof. First, the identity mapping id : U → U, id(z) = z, is the group

identity. Now if φ ∈ Aut(U ), then φ is one-to-one and onto so that the
inverse function φ −1 makes sense and is holomorphic; it is also one-to-
one and onto. And φ ◦ φ −1 = φ −1 ◦ φ = id. Thus the inverse function
φ −1 is the group inverse.
Next, let φ, ψ ∈ Aut(U ). Then, by definition, φ : U → U and ψ :
U → U so certainly ψ ◦ φ : U → U . Since φ and ψ are one-to-one,
so is the composition ψ ◦ φ. Likewise, the composition is onto because
each of the individual functions is. Thus Aut(U ) is closed under the
group operation.
Finally, the group operation is associative because composition of
functions is associative. Hence Aut(U ) is a group.

Example 1. Let U be the disc. By the exercise following Theorem 3

of Section 0.2, Aut(U ) consists of all maps of the form

ζ −a
ζ −→ µ ·
1 − aζ
where a, µ ∈ C, |a| < 1, and |µ| = 1.

Example 2. Let 0 < r < R < ∞ and let

Ar,R = {z : r < |z| < R}.

Then Aut(Ar,R ) consists only of the rotations

z −→ µ · z (for µ a unimodular constant),

the reflection
R ·r
σ : z −→ ,
z
An Application of Completeness: Automorphisms 123

and compositions of these two types of functions (see Theorem 1 of

Section 4.6).
Clearly each of these two types of functions is an element of
Aut(Ar,R ). For the converse, let φ ∈ Aut(Ar,R ). Since φ −1 is con-
tinuous, φ is proper. This means that the inverse image of any compact
set under φ is compact. Taking contrapositives, we see that if z j ∈ Ar,R
satisfies z j → ∂ Ar,R , then φ(z j ) → ∂ Ar,R . Thus we have confirmed
that the boundary goes to the boundary (and either Carathéodory’s the-
orem, or the maximum principle, or Brouwer invariance of domain are
other standard ways to see this assertion). A tricky but elementary topo-
logical argument (see [GRK, pp. 243–244]) then shows that either the
inner boundary of Ar,R is mapped to the inner boundary of Ar,R or
else the inner boundary of Ar,R is mapped to the outer boundary. After
applying an inversion, we may suppose the former.
Notice that the Schwarz reflection principle implies that φ contin-
ues analytically past ∂ Ar,R . We claim that φ must be a rotation. We may
reflect φ across both boundaries of Ar,R to obtain an automorphism of
the larger annulus {z : r 2 /R < |z| < R 2 /r }. Continuing this reflection
process countably many times, we can analytically continue φ to an au-
tomorphism of C ≡ C \ {0}. By the Riemann removable singularities
theorem, φ continues analytically to an automorphism of C which fixes
the origin. Thus there is a constant α such that φ(z) ≡ αz. Since φ is
the continuation of an automorphism of a proper annulus Ar,R , it must
be that α has modulus 1. Therefore φ is a rotation as claimed.

Given a domain U , we equip Aut(U ) with the topology of uniform

convergence on compact sets. That is,

Aut(U )  φ j −→ φ ∈ Aut(U )

provided that φ j → φ normally. We are particularly interested in the

question of when the topological group Aut(U ) is compact. Given that
we have defined the topology sequentially, it is natural to consider
sequential compactness: Aut(U ) is compact if and only if, whenever
{φ j } ⊆ Aut(U ), then there is a subsequence {φ jk } that is convergent
(uniformly on compact subsets of U ) to an element of Aut(U ).
124 Some New Invariant Metrics

Example 3. Let 0 < r < R < ∞ and let

Ar,R = {z : r < |z| < R}

be the corresponding annulus. Then Aut(Ar,R ) is compact. To see this,

let α j be a sequence of automorphisms of Ar,R . Now each of the α j is
either a rotation, the reflection σ (z) = r · R/z, or a composition of the
two. Thus infinitely many of the α j are of the same one of these three
types. If infinitely many are equal to the reflection σ , then a subse-
quence consisting of these elements (this will be a constant sequence)
converges to σ . If infinitely many are rotations, then let us write these
as

α jk (z) = µ jk · z,

where the µ jk are unimodular constants. Since the unit circle is com-
pact, there is a subsequence—call it µm —which converges to a uni-
modular constant µ0 . But then the corresponding automorphisms

αm (z) = µm z

converge normally to the limit automorphism α0 (z) = µ0 z. The third

possibility is that infinitely many of the α j consist of a rotation fol-
lowed by the reflection σ . Write these as

α jk (z) = σ (µ jk z),

with the µ jk being unimodular constants. We extract a convergent sub-

sequence {µm } with limit µ0 and define

α0 (z) = σ (µ0 z).

Then the subsequence

αm (z) = σ (µm z)

converges normally to α0 (z).

The three possibilities having been covered, we conclude that
Aut(Ar,R ) is compact.
An Application of Completeness: Automorphisms 125

Remark. It is an interesting fact that if the connectivity of a planar

domain is finite but greater than two then the automorphism group of
the domain is a finite group (see [FK] or [HEI]). That is to say, a domain
with at least two holes, but finitely many, has automorphism group that
is a finite group. It is an open problem to determine which finite groups
arise as automorphism groups of planar domains.

Example 4. Let U = D, the unit disc. Then the functions

ζ + (1 − 1/j)
β j (ζ ) =
1 + (1 − 1/j)ζ
are elements of Aut(D). However, {β j } converges normally to the func-
tion

β0 (ζ ) ≡ 1.

Notice that β0 ∈ Aut(D)—in fact it is a constant function. Therefore

Aut(D) is not compact.

Remark. Notice that, at least when U is bounded, the issue of whether

Aut(U ) is compact is not a question of whether {φ j } ⊆ Aut(U ) has
a subsequence that converges: by normal families this is essentially
automatic. Rather, the issue is whether the subsequence will converge
to an element of Aut(U). It is this last phenomenon which fails for the
disc, but works for the annulus.

The principal result of this section is that the disc is characterized

by the non-compactness of its automorphism group. More precisely,
we have

Theorem 3. Let U ⊆ C be a bounded domain with C 2 boundary. If

Aut(U ) is non-compact then U is conformally equivalent to the unit
disc.

We prove this theorem with a sequence of lemmas, each of which has

intrinsic interest. Take a moment to review the notions of C 2 bound-
126 Some New Invariant Metrics

ary and C k boundary which were discussed in Definitions 2 and 3 of

Section 3.

Lemma 4. Let U ⊆ C be bounded. The group Aut(U ) is compact if

and only if, for each P ∈ U , there is a compact K P ⊆ U such that
φ(P) ∈ K P for all φ ∈ Aut(U ).

Proof. Assume that Aut(U ) is compact. Fix P ∈ U . If there is no set

K P as claimed then there exist φ j ∈ Aut(U ) such that φ j (P) → w ∈
∂U , some w. But U is bounded so that {φ j } is a normal family; thus
there is a subsequence φ jk and a holomorphic limit function φ0 such
that

φ jk −→ φ0

normally.
Notice that the image of each φ j lies in U , hence the image of φ0
lies in the closure U of U . If φ0 is nonconstant then it satisfies the open
mapping principle. But

φ0 (P) = lim φ jk (P) = w,

k→∞

hence the image of φ0 contains the accumulation point w ∈ ∂U , so it

contains a neighborhood of w. This is impossible because w is in the
boundary of the image of φ0 . Therefore φ0 must be constantly equal to
w; thus φ0 ∈ Aut(U ). The sequence φ jk therefore violates the com-
pactness of Aut(U ). We conclude that K P must exist.
For the converse, fix P ∈ U and let K P be the corresponding
compact set in U whose existence we assume. Let {φ j } ⊆ Aut(U )
be any sequence. Since U is bounded, there is a normally converging
subsequence φ jk with holomorphic limit function φ0 . As in the first half
of the proof, if the image of φ0 contains any boundary point w then φ0
must be constantly equal to w. But the image of P under φ0 must lie in
K P , so this possibility is ruled out. We conclude that the image of φ0
lies in U .
An Application of Completeness: Automorphisms 127

Next notice that each φ jk has an inverse ψ jk . Passing to another

subsequence, we may suppose that the ψ jk converge to a limit function
ψ0 . For convenience, we denote this final subsequence by ψm , corre-
sponding to the automorphisms φm . Just as for φ0 , we can be sure that
the image of ψ0 lies in U . As in the proof of Theorem 5 of Section 2,
ψ0 is non-constant.
Now we have

z ≡ lim φm ◦ ψm (z) = φ0 ◦ ψ0 (z).

m→∞

Since i(z) ≡ z is onto, so is φ0 . Also, as in the proof of Theorem 5 in

Section 2, the image of ψ0 is open, closed, and nonempty. Therefore ψ0
is surjective. Since i(z) is injective, it now follows that ψ0 is injective.
Therefore ψ0 ∈ Aut(U ) and it follows that φ0 ∈ Aut(U ). We conclude
that

Aut(U )  φm −→ φ0 ∈ Aut(U ),

and Aut(U ) is compact.

Remark. It may be noted that, in the proof of the converse direction

of Lemma 4, only one compact set K P was needed.

Lemma 5. Let U ⊆ C be a bounded domain with C 2 boundary. Sup-

pose that P ∈ U, {φ j } are holomorphic maps from U to U , and

φ j (P) −→ w ∈ ∂U.

If K is compact in U and V is a neighborhood of w, then there exists a

positive number J such that if j ≥ J then

φ j (K ) ⊆ V.

See Figure 1.

Proof. Since U , when equipped with the Carathéodory metric, is a met-

ric space and since K is compact, there is a positive number R such that
128 Some New Invariant Metrics

K
(K)
j

Figure 1.

the metric ball B(P, R) contains K (see Proposition 1 in Section 3). Let
Q j = φ j (P). Since each φ j is distance-decreasing in the Carathéodory
metric, it follows that φ j (B(P, R)) ⊆ B(Q j , R). We claim that there is
a positive J such that whenever j ≥ J then B(Q j , R) ⊆ V . Assuming
the claim, we would then have

φ j (K ) ⊆ φ j (B(P, R)) ⊆ B(Q j , R) ⊆ V,

as required.
To prove the claim, recall from the proof of Theorem 6 in the last
section that (because the Carathéodory metric on U is complete) the
Euclidean radii of the metric balls B(Q j , R) must tend to 0. Choose
 > 0 such that the Euclidean disc of center w and radius 2 lies in V .
We select J so large that when j > J , then both the Euclidean distance
of Q j to w is less than  and the Euclidean radius of B(Q j , R) is less
than . The claim now follows from the triangle inequality.

Proof of Theorem 3. If Aut(U ) is not compact then, by Lemma 4, there

is a sequence φ j ∈ Aut(U ) and a P ∈ U such that

φ j (P) −→ w ∈ ∂U,

for some w ∈ ∂U .
An Application of Completeness: Automorphisms 129

Let

γ : [0, 1] −→ U

be any continuous, closed curve in U . Since ∂U is C 2 there is a neigh-

borhood V of w such that U ∩V is simply connected (see Figure 2—the
existence of the interior osculating circle, or the tubular neighborhood,
provided by Proposition 3 of Section 3 makes this assertion clear).

Figure 2.

Let

K = {γ (t) : 0 ≤ t ≤ 1}.

Then K is compact. By Lemma 5, there is a J ≥ 0 such that j ≥ J

implies φ j (K ) ⊆ U ∩ V . Thus φ j ◦ γ is a continuous, closed curve
in U ∩ V . The simple connectivity of U ∩ V implies that φ j ◦ γ may
be continuously deformed to the point φ j ◦ γ (0); that is, there is a
homotopy

 : [0, 1] × [0, 1] −→ U ∩ V

such that

(0, t) = φ j ◦ γ (t), ∀t ∈ [0, 1]

130 Some New Invariant Metrics

and

(1, t) = φ j ◦ γ (0), ∀t ∈ [0, 1].

But then

(φ j )−1 ◦ 

is a homotopy of the curve γ to the point γ (0). It follows that U is

simply connected. By the Riemann mapping theorem, U is conformally
equivalent to the disc.

Remark. It is possible to prove this theorem without resort to the Rie-

mann mapping theorem. Instead one uses a delicate approximation ar-
gument on a decreasing sequence of boundary neighborhoods V j ∩ U
of w to prove that

FCU (P) = FKU (P).

It then follows from Theorem 5 of Section 2 that U is conformally

equivalent to the disc. In the same vein, one can actually prove the
Riemann mapping theorem using FCU and FKU . Since the ideas involved
are rather complicated and would take us far afield, we omit them.

The notion of compact/non-compact automorphism group is a bit

abstract, so we conclude this section with an interpretation of Theo-
rem 3 which is more concrete. Let us say that Aut(U ) acts transitively
on the domain U if for any P, Q in U there is an element φ ∈ Aut(U )
such that φ(P) = Q. In other words, the automorphism group acts
transitively if it is sufficiently rich with elements that it can move any
point of U to any other.

Example 5. Let U = D, the unit disc. If P is any element of D then

the automorphism
ζ−P
φ P (ζ ) ≡
1 − Pζ
An Application of Completeness: Automorphisms 131

has the property that φ P (P) = 0. Now if P and Q are arbitrary ele-
ments of D then (φ Q )−1 ◦ φ P is an automorphism of D that maps P
to Q. Thus the automorphism group of the disc does act transitively.

Example 6. Let Ar,R be the annulus

{z : r < |z| < R}.

Then Aut(Ar,R ) does not act transitively on Ar,R . In fact if P, Q ∈

Ar,R , if |P| = |Q|, and if |P| = r · R/|Q| then Example 2 shows that
there is no automorphism of Ar,R which takes P to Q.

We can now give a consequence of Theorem 3, using the language

of transitivity.

Corollary 5.1. Let U be a bounded domain with C 2 boundary. The

group Aut(U ) acts transitively on U if and only if U is conformally
equivalent with the disc.

Proof. Fix a point P in U . If Aut(U ) acts transitively on U , then the

point P can be moved to any other point of U by some element of
Aut(U ). Thus {φ(P) : φ ∈ Aut(U )} can be contained in no compact
set K P ⊆ U . By Lemma 4, Aut(U ) is non-compact. By the theorem,
U must be conformally equivalent to the disc.
The converse is obvious from our explicit knowledge of the auto-
morphisms of the disc.

It is natural to wonder whether the hypothesis of the C 2 boundary

is really necessary for the results presented in this section. The answer
is that some regularity of the boundary is required, as the next example
illustrates.

Example 7. Define the mapping

z + 1/2
φ(z) = .
1 + (1/2)z
132 Some New Invariant Metrics

We know that φ is an automorphism of the disc. Define φ 2 = φ ◦ φ,

φ 3 = φ ◦ φ ◦ φ, etc. With φ −1 denoting the inverse of φ, we set φ −2 =
φ −1 ◦ φ −1 , φ −3 = φ −1 ◦ φ −1 ◦ φ −1 , etc. We let φ 0 (z) ≡ z.
For j ∈ Z we define

d j = {φ j (z) : |z| ≤ 1/10}.

Finally, let

U = D(0, 1) \ d j.
j∈Z

The domain U is displayed in Figure 3.

Figure 3.

It is not difficult to verify that Aut(U ) consists precisely of the

maps φ j , j ∈ Z, together with the map z → −z. Thus Aut(U ) is
non-compact (exercise). Yet U is certainly not conformally equivalent
to the disc.
Notice that the boundary of U is not C 2 , in the sense that it is im-
possible to write U = {z ∈ C : ρ(z) < 0} for any C 2 defining function
ρ with ∇ρ = 0 on ∂U ; so there is no contradiction with Theorem 3.

Exercise. Determine the automorphism group of the punctured disc.

Is it compact or not?
Hyperbolicity and Curvature 133

5. Hyperbolicity and Curvature

In this section we relate the concepts of curvature and normal families
to the property of nondegeneracy of the Kobayashi metric. A conse-
quence is that we can prove that the plane and C \ {0} cannot support
a metric of strictly negative curvature. Thus we see, from a geometric
point of view, why Picard’s theorem must specify the omission of two
points of the range of an entire function in order for that function to be
necessarily constant.
We begin with an elegant generalization of the fact that the
Carathéodory metric is never greater than the Kobayashi metric. Re-
call here that ρ denotes the Poincaré metric on the disc D.
Proposition 1. If U ⊆ C is a domain which is equipped with a metric
σ and if every holomorphic map f : D → U is distance-decreasing
from (D, ρ) to (U, σ ) then

σ ≤ FKU .

Proof. Abbreviate FKU by the symbol µ. Let z ∈ U . If φ : D → U and

φ(0) = z then, by hypothesis,
φ ∗ σ (0) ≤ ρ(0)

or
1 σ (z)
≥ = σ (z).
|φ (0)| ρ(0)
Taking the infimum over all φ we conclude, for any ξ ∈ C, that
|ξ |
#ξ #µ,z = inf ≥ |ξ | · σ (z) = #ξ #σ,z ,
φ∈(U,D)z |φ (0)|

as desired.
Definition 2. A domain U ⊆ C is called hyperbolic if the Kobayashi
distance on U is actually a distance, that is, if
dKob (P, Q) > 0

whenever P and Q are distinct points of U .

134 Some New Invariant Metrics

Proposition 3. Let U ⊆ C be a domain equipped with a C 2 metric σ

which has the property that its curvature κ = κσ satisfies

κ ≤ −B < 0

for some positive constant B. Then U is hyperbolic.

Proof. Let f be any map of the disc D into U . By Ahlfors’s version of

Schwarz’s lemma, we have that
√
∗ 4
f σ ≤ √ ρ,
B

√ √ρ is the Poincaré metric on the disc. Replace σ by σ̃ ≡

where
( B/ 4) · σ . Then our inequality is

f ∗ σ̃ ≤ ρ.

Thus f is distance-decreasing from (D, ρ) to (U, σ̃ ). By Proposition 1,

we may conclude that

σ̃ ≤ FKU

or

C · σ ≤ FKU ,
√ √
where C = B/ 4 is a positive constant. Thus the Kobayashi dis-
tance is bounded from below by a positive constant times the σ -
distance. Since σ is assumed to be nondegenerate, so is FKU .

Corollary 3.1. Neither the plane C nor the punctured plane C \ {0}
possesses a metric of strictly negative curvature.

Proof. Neither C nor C \ {0} is hyperbolic.

Corollary 3.2. If U = C \ {P1 , . . . , Pk }, with the P j ’s distinct and

k ≥ 2, then U is hyperbolic.
Hyperbolicity and Curvature 135

Proof. As we saw in Chapter 2, such a U has a metric of strictly nega-

tive curvature.

Remark. An alternative proof of Corollary 3.2 may be obtained by

noticing that a domain U which satisfies the hypothesis will imbed
into C \ {0, 1} and then using the distance-decreasing property of the
Kobayashi metric.

Now we will connect hyperbolicity with the concept of normal

families. We first need some new terminology.

Definition 4. A domain U is taut if the family F of holomorphic func-

tions from the disc D into U is normal.

Example 1. The disc D is taut. For if F is the family of holomorphic

functions from D to D then the elements of F are uniformly bounded
by 1. So any sequence of elements { f j } ⊆ F has a subsequence { f jk }
which converges to a limit function f 0 . The function f 0 is certainly
holomorphic, and there are two possibilities: i) the image of f 0 con-
tains a boundary element w, in which case the open mapping principle
implies that f 0 is constant (this is the compactly divergent case); (ii)
the image of f 0 lies entirely in the interior of D, in which case f 0 ∈ F
(this is the convergent case).

Example 2. The domain U = C \ {0} is not taut. The functions

f n (z) = enz

map the disc into U but there is neither a convergent subsequence of

{ f n } nor a compactly divergent one (the origin is mapped to 1 by every
f n but f n (1/2) → ∞ as n → ∞).

Proposition 5. A planar domain U is hyperbolic if and only if it is

taut.
136 Some New Invariant Metrics

Proof. We know that if U is hyperbolic then C\U must contain at least

two points. By the general version of Montel’s theorem in Section 2.4,
U must therefore be taut.
Conversely, assume that U is taut. Then, by Example 2, C \ U
must contain at least two points. But then U is hyperbolic.

Remark. One theme which comes through in this section is that the
only enemies are C and C \ {0}. As soon as a domain excludes at
least two points, then it has all relevant properties. In the theory of
Riemann surfaces it also holds that a domain is taut if and only if it is
hyperbolic (see [FK]). However, in several complex variables, matters
are no longer so simple. A detailed discussion of the relationship be-
tween tautness and hyperbolicity appears in [KOB2, p. 240]. See also
[KOB1].

Exercise. The domain

U = {z ∈ C : |z| < 1} \ {x + i0 : 0 ≤ x < 1}

is taut and hyperbolic.

CHAPTER 4
Introduction to the Bergman Theory

0. Introductory Remarks
It is a remarkable fact—discovered by Stefan Bergman in 1927—that
a bounded domain
in C or in Cn can be equipped with a “canoni-
cal” reproducing kernel. [Here we use the phrase “reproducing kernel”
to mean a function k(z, ζ ) of two variables—like the familiar Cauchy
kernel—with the property that integration against a holomorphic func-
tion f produces the value of f at z. In one complex variable such a
formula could take the form


f (z) = k(z, ζ ) f (ζ ) dζ
∂

or



f (z) = k(z, ζ ) f (ζ ) dξ dη

when ζ = ξ + iη. Such a formula means, in effect, that the kernel

k contains valuable information about holomorphic functions on
.]
In one complex variable, the existence of a reproducing kernel is not
great news. The Cauchy kernel, after all, works for any domain. But
in Bergman’s time there were no reproducing kernels in several com-
plex variable (except on very special domains like the ball and the
polydisc—see Chapter 5 below), and his goal was to produce a con-
struction of a kernel that worked on virtually any domain.

137
138 Introduction to the Bergman Theory

A particularly remarkable byproduct of Bergman’s construction is

that his kernel can in turn be used to produce a “canonical invariant
metric.” This metric has many special properties: it is smooth (indeed
real analytic), it is Hermitian, it is Kähler, and it is invariant under bi-
holomorphic mappings. As a result, this “Bergman metric” has been of
great interest to geometers for more than three quarters of a century.
In the present chapter we shall give a very brief, and occasionally
sketchy, introduction to the Bergman kernel and metric. The basic el-
ements of the Bergman theory rely on Hilbert space ideas, and we do
not wish to burden the reader with this extra formalism. So we shall
steer clear of most of Hilbert space, at the expense of not being able to
provide all proofs in detail. Still, we believe that this chapter will of-
fer many rewards to the interested reader. We shall be able to calculate
the Bergman kernel and metric on at least some domains, and we shall
be able to verify their invariance explicitly. We shall provide means
(in fact three different methods) for constructing the Bergman kernel.
And we will apply the new Bergman metric to a problem in conformal
mapping.
We encourage the reader to study [RUD3] for details on Hilbert
space theory. The books [GRK], [KR1], [FUK], [EPS] provide further
details of the Bergman theory.

1. Bergman Basics
The Bergman theory centers on a special function space. Let
⊆ C be
a bounded domain. Now define the Bergman space



A2 (
) = { f holomorphic on
: | f (z)|2 d A(z) < ∞}.

Here d A is the usual area measure on
. In other words, d A = d x dy.

We define a norm on A2 (
) by


1/2
f A2 (
) ≡ | f (z)| d A(z)
2
.


Bergman Basics 139

This norm satisfies the standard properties which we would expect of

such an operation:

1. f A2 (
) ≥ 0;
2. f + g A2 (
) ≤ f A2 (
) + g A2 (
) ;
3. c f A2 (
) = |c| · f A2 (
) for any complex constant c.

There is an inner product on A2 (
) which is naturally associated

to this norm. Namely, for f, g ∈ A2 (
),



 f, g A2 (
) =  f, g = f (z)g(z) d A(z).

Naturally we could give similar definitions on a bounded domain

⊆ Cn , with area measure d A replaced by volume measure d V . The
theory would go through without any change. For simplicity, however,
we shall concentrate in this chapter on domains in the complex plane.
It is important to observe that, as long as
is a bounded domain,
the space A2 (
) is certainly non-empty. For example, any holomorphic
polynomial (restricted to
) will be in A2 (
). If
is simply connected
√
and P ∈ ∂
then g(z) = 1/ z − P (using the principal branch of
square root, as we may) will be in A2 (
).
Now the Bergman kernel K (z, ζ ) is a function on the set
×

with these properties:

1. For each fixed ζ ∈
, K (·, ζ ) ∈ A2 (
).

2. For each fixed z ∈
, K (z, ·) ∈ A2 (
).
3. For z, ζ ∈
, K (z, ζ ) = K (ζ, z).
4. If f ∈ A2 (
) and z ∈
then



f (z) = K (z, ζ ) f (ζ ) d A(ζ ).

Observe that properties (1) and (3) already imply (2). Also (2)
and (3) imply (1). But it is useful to have these attributes all laid out
explicitly. It is important to note that properties (1)–(4) characterize the
140 Introduction to the Bergman Theory

Bergman kernel: Any function k(z, ζ ) satisfying (1)–(4) will satisfy

k = K . This characterization will prove useful below.
One can use abstract Hilbert space methods to prove the existence
of the Bergman kernel. We instead will present, in Section 3, three dif-
ferent methods for constructing the Bergman kernel. The third method
in effect proves the existence of the Bergman kernel, for it expresses
the Bergman kernel in terms of the well-known Green’s function.
Property (4) is the reproducing property of the Bergman kernel.
Compare, for example, the Cauchy integral formula for holomorphic
functions. At first blush, we might suspect that the Bergman kernel is
the Cauchy kernel. But bear in mind that the Bergman kernel is a kernel
for integration with respect to area while the Cauchy kernel is a kernel
for complex line integration. It turns out that this information alone
forces the two kernels to be different.
Another important point to notice is that the Cauchy kernel
C(z, ζ ) = 1/(ζ − z) is the same for all domains. In fact, Bergman
kernels for different domains are (generically) different.

2. Invariance Properties of the

Bergman Kernel
Suppose that
1 ,
2 are bounded domains in C. Let φ :
1 →
2 be
a conformal mapping. We might hope that the Bergman kernel K
1 for

1 and the Bergman kernel K
2 for
2 are related by way of φ.
To establish such a fact, we first need a technical result about the
way that area transforms under conformal mapping.

Lemma 1. (Lusin). Let φ :
1 →
2 be a conformal mapping of

domains. Then





area (
2 ) = 1 d A(z) = |φ  (ζ )|2 d A(ζ ).

2
1

Proof. The proof amounts to writing the standard change-of-variables

formula in complex notation. Write φ in terms of its real and imaginary
Invariance Properties of the Bergman Kernel 141

parts as

φ(ζ ) = φ1 (ζ ) + iφ2 (ζ ).

Also write ζ = τ + iη. Then the usual Jacobian matrix determinant

Jac φ of φ is
 
∂φ1 ∂φ1
 ∂τ ∂η 
det   ∂φ1 ∂φ2 ∂φ2 ∂φ1
 ∂φ2 ∂φ2  = ∂τ · ∂η − ∂τ · ∂η .
∂τ ∂η
We may apply the Cauchy-Riemann equations twice to convert this last
expression to


2

2
∂φ1 ∂φ1 ∂φ2 ∂φ2
∂φ

∂φ

· + · =

= |φ  (ζ )|2 .
∂τ ∂τ ∂τ ∂τ ∂τ ∂ζ

In the last step we have used the fact that ∂φ/∂τ = ∂φ/∂ζ = φ  for a
holomorphic function φ.
In conclusion,





area (
2 ) = 1 d A(z) = det Jac φ(ζ ) d A(ζ )

2
1



= |φ  (ζ )|2 d A(ζ ),

1

as the lemma asserts.

Now we have

Theorem 2. If
1 ,
2 , φ are as in the introductory paragraph, then,

for z, ζ ∈
1 ,

K
1 (z, ζ ) = φ  (z) · K
2 (φ(z), φ(ζ )) · φ  (ζ ).

Proof. Let f ∈ A2 (
1 ). Then (applying the change of variable ξ =

φ(ζ )—valid because φ and φ −1 are univalent and continuously differ-
142 Introduction to the Bergman Theory

entiable)



φ  (z) · K
2 (φ(z), φ(ζ )) · φ  (ζ ) f (ζ ) d A(ζ )

1



= φ  (z) · K
2 (φ(z), φ(φ −1 (ξ )))

2


2
× φ  (φ −1 (ξ )) f (φ −1 (ξ ))
[φ −1 ] (ξ )
d A(ξ )



= φ  (z) · K
2 (φ(z), ξ )

2

× φ  (φ −1 (ξ )) f (φ −1 (ξ ))
1
× d A(ξ )
φ  (φ −1 (ξ ))φ  (φ −1 (ξ ))



= φ  (z) · K
2 (φ(z), ξ )

2
 
−1 1
× f (φ (ξ )) · d A(ξ )
φ  (φ −1 (ξ ))



≡ φ  (z) · K
2 (φ(z), ξ ) · g(ξ ) d A(ξ ),

2

where g(ξ ) = f (φ −1 (ξ )) · 1
φ  (φ −1 (ξ ))
.
It is easy to verify, using a change of variable, that g ∈ A2 (
2 ). As
a result, the Bergman kernel K
2 for
2 will reproduce g. We conclude
that



φ  (z) · K
2 (φ(z), φ(ζ )) · φ  (ζ ) f (ζ ) d A(ζ )

1

= φ  (z) · g(φ(z)) = f (z).

We have verified that the kernel

k(z, ζ ) ≡ φ  (z) · K
2 (φ(z), φ(ζ )) · φ  (ζ )

satisfies the reproducing property (4) for the Bergman kernel.

Calculation of the Bergman Kernel 143

The kernel φ  (z)· K
2 (φ(z), φ(ζ ))·φ  (ζ ) plainly satisfies the con-
jugate symmetric property (3) of the Bergman kernel for
1 . We also
note that k is holomorphic in the z-variable and conjugate holomorphic
in the ζ -variable. Properties (1) and (2) may be checked by a change
of variables just like the one we performed to prove the reproducing
property. We omit the details.
We may conclude, by the uniqueness of the Bergman kernel, that

φ  (z) · K
2 (φ(z), φ(ζ )) · φ  (ζ ) = K
1 (z, ζ ).

This is what we wished to prove.

3. Calculation of the Bergman Kernel

The usual construction of the Bergman kernel is rather abstract, re-
lying as it does on the Riesz representation theorem of Hilbert space
theory (see [RU3] for the details). But this gives us little leverage for
actually calculating the kernel. In the present section we shall present
three different techniques for constructing the Bergman kernel of the
unit disc. Theorem 2 of Section 2 then gives, at least in principle, the
means to calculate the Bergman kernel for other simply connected do-
mains.

3.1 Construction of the Bergman Kernel for the Disc

by Conformal Invariance
Let D ⊆ C be the unit disc. First we notice that, if either f ∈ A2 (D)
or f ∈ A2 (D), then



1
f (0) = f (ζ ) d A(ζ ). (∗)
π D

This is the standard area form of the mean value property for holomor-
phic or harmonic functions.
144 Introduction to the Bergman Theory

Of course the constant function u(z) ≡ 1 is in A2 (D), so it is

reproduced by integration against the Bergman kernel. Hence, for any
w ∈ D,





1 = u(w) = K (w, ζ )u(ζ ) d A(ζ ) = K (w, ζ ) d A(ζ )
D D

or



1 1
= K (w, ζ ) d A(ζ ).
π π D

By (∗), we may conclude that

1
= K (w, 0)
π
for any w ∈ D.
Now, for a ∈ D fixed, consider the Möbius transformation
z−a
h(z) =
1 − az
that we studied in Section 1.4. We know that
1 − |a|2
h  (z) = .
(1 − az)2
We may thus apply Theorem 2 of Section 4.2 with φ = h to find that

K (w, a) = h  (w) · K (h(w), h(a)) · h  (a)

1 − |a|2 1
= · K (h(w), 0) ·
(1 − aw) 2 1 − |a|2
1 1
= ·
(1 − aw)2 π
1 1
= · .
π (1 − wa)2
Thus, assuming the existence of the Bergman kernel of the disc, we
have derived a closed-form formula for it.
Calculation of the Bergman Kernel 145

3.2 Construction of the Bergman Kernel by Means of

an Orthonormal System
One of the most standard means of calculating the Bergman kernel is
the following lemma:

Lemma 1. Let {φ j } be a complete orthonormal system for A2 (
).

Then the function

∞
k(z, ζ ) ≡ φ j (z) · φ j (ζ ) ()
j=1

is in fact the Bergman kernel for
.

This assertion requires a bit of explanation. Let us assume, for

simplicity, that our domain
is bounded. Then we note that A2 (
) is
a vector space over the field of complex numbers; and it is an infinite
dimensional vector space because all holomorphic polynomials are in
A2 (
). It can be checked that A2 (
), for
bounded, is separable (it
is, after all, a subspace of the separable space L 2 (
)).
The inner product on A2 , as already noted, is



 f, g = f (z)g(z) d A(z).


A collection of elements {φα }α∈A is called orthonormal if each φα sat-
isfies φα 2 = φα , φα  = 1 for all α ∈ A and φα , φβ  = 0 when-
ever α = β. The collection {φα }α∈A is termed complete if, whenever
x, φα  = 0 for all α, then x = 0.
It is an exercise with Zorn’s lemma to see that any Hilbert space,
and A2 in particular, has a complete orthonormal system. Because A2 is
separable, one can see that such a system must be countable. It is a fact
from Hilbert space theory that if {φ j } is a complete orthonormal system

for a separable Hilbert space H and if f ∈ H, then f = j a j φ j , for
some constants a j , with convergence in the Hilbert space norm.
In general, it can be rather difficult to actually write down a com-
plete orthonormal system for A2 (
). Fortunately, the unit disc D ⊆ C
has enough symmetry that we can actually pull this off.
146 Introduction to the Bergman Theory

It is not difficult to see that {z j }∞j=0 is an orthogonal system for

A (D). That is, the elements are pairwise orthogonal, but they are not
2

normalized to have unit length. We may confirm the first of these as-
sertions by noting that if j = k, then



z j , z k  = z j z k d A(z)
D

1
2π
= r j ei jθ r k e−ikθ dθr dr
0 0

1
2π
= r j+k+1
dr ei( j−k)θ dθ
0 0
= 0.

The system {z j } is complete: If  f, z j  = 0 for every j, then f will

have a null power series expansion and hence be identically zero. It
remains to normalize these monomials so that we have a complete or-
thonormal system.
We calculate that



1
2π
|z | d A(z) =
j 2
r 2 j dθr dr
D 0 0

1
= 2π r 2 j+1 dr
0
1
=π· .
j +1

We conclude that
√
π
z = √j
.
j +1

Therefore the elements of our orthonormal system are

√
j +1· zj
φ j (z) = √ .
π
Calculation of the Bergman Kernel 147

Now, according to formula (), the Bergman kernel is given by


∞
K (z, ζ ) = φ j (z) · φ j (ζ )
j=0

∞
( j + 1)z j ζ j
=
j=0
π

1 ∞
= · ( j + 1) · (zζ ) j .
π j=0

Observe that, in this instance, the convergence of the series is mani-

fest (for both |z| < 1 and |ζ | < 1). The convergence of the general
orthonormal expansion in Lemma 1 is tricky, and we omit that proof.

Of course we can easily sum j ( j + 1)α j by noticing that

∞
d ∞
( j + 1)α j = α j+1
j=0
dα j=0
 
d 1
= α·
dα 1−α
1
= .
(1 − α)2
Applying this result to our expression for K (z, ζ ) yields that
1 1
K (z, ζ ) = · .
π (1 − z · ζ )2
This is consistent with the formula which we obtained by conformal
invariance in Subsection 3.1 for the Bergman kernel of the disc.

3.3 The Bergman Kernel by Way of

Differential Equations
It is actually possible to obtain the Bergman kernel of a domain in
the plane from the Green’s function for that domain (see [EPS]). Let
148 Introduction to the Bergman Theory

us now summarize the key ideas. Unlike the first two Bergman ker-
nel constructions, the present one will work for any domain with C 2
boundary.
First, the fundamental solution for the Laplacian in the plane is
the function

1
(ζ, z) = log |ζ − z|.
2π

This means that ζ (ζ, z) = δz . [Observe that δz denotes the Dirac

“delta mass” at z and ζ is the Laplacian in the ζ variable.] Here the
derivatives are interpreted in the sense of distributions. In more prosaic
terms, the condition is that


(ζ, z) · φ(ζ ) dτ dη = φ(z)

for any C 2 function φ with compact support. We write, as usual, ζ =

τ + iη. The reference [KR1, p. 35] provides details of this assertion.
Given a domain
⊆ C, the Green’s function is posited to be a
function G(ζ, z) which satisfies

G(ζ, z) = (ζ, z) + H (ζ, z),

where H (ζ, z) is a particular harmonic function in the ζ variable.

Moreover, it is mandated that G(·, z) vanishes on the boundary of
.
One constructs the function H , for each fixed z, by solving a suitable
Dirichlet problem. Again, the reference [KR1, p. 40, ff.] has all the
details. Now we have

Proposition 2. Let
⊆ C be a bounded domain with C 2 bound-

ary. Let G(ζ, z) be the Green’s function for
and let K (z, ζ ) be the
Bergman kernel for
. Then

∂2
K (z, ζ ) = 4 · G(ζ, z). ()
∂ζ ∂z
Calculation of the Bergman Kernel 149

Proof. Our proof will use a version of Stokes’s theorem written in the
notation of complex variables. It says that if u ∈ C 1 (
), then



∂u
u(ζ ) dζ = 2i · dτ dη, (∗∗)
∂U U ∂ζ

where again ζ = τ + iη. The reader is invited to convert this formula

to an expression in τ and η and to confirm that the result coincides
with the standard real-variable version of Stokes’s theorem which can
be found in any calculus book (see, e.g., [THO]).
Now we already know that

1 1
G(ζ, z) = log(ζ − z) + log (ζ − z) + H (ζ, z). (†)
4π 4π
Here we think of the logarithm as a multivalued holomorphic function;
after we take a derivative, the ambiguity (which comes from an additive
multiple of 2πi) goes away.
Differentiating with respect to z (and using subscripts to denote
derivatives), we find that
1 −1
G z (ζ, z) = + Hz (ζ, z).
4π ζ − z
We may rearrange this formula to read

1
= −4π · G z (ζ, z) + 4π Hz (ζ, z).
ζ −z
We know that G, as a function of ζ , vanishes on ∂
. Hence so does
G z . Let f ∈ C 2 (
) be holomorphic on
. It follows that the Cauchy
formula

1 f (ζ )
f (z) = dζ
2πi ∂
ζ − z

can be rewritten as

2
f (z) = f (ζ )Hz (ζ, z) dζ
i ∂

150 Introduction to the Bergman Theory

or

−2i f (ζ )Hz (ζ, z) dζ = f (z).
∂

Now we apply Stokes’s theorem (in the complex form) to rewrite

this last as



f (z) = 4 · ( f (ζ )Hz )ζ dτ dη.

Since f is holomorphic and H is real-valued, we may conveniently

write this last formula as



f (z) = 4 · f (ζ )Hζ z dτ dη.

Now formula (†) tells us that Hζ z = G ζ z . Therefore we have

f (z) = f (ζ )4G ζ z dτ dη. (‡)

With a suitable limiting argument, we may extend this formula from

functions f which are C 2 (
) to functions in A2 (
).
It is straightforward now to verify that 4G ζ z satisfies the first three
characterizing properties of the Bergman kernel—just by examining
our construction. The crucial reproducing property is of course formula
(‡). Then it follows that

∂2
K (z, ζ ) = 4 · G(ζ, z).
∂ζ ∂z
That is the desired result.

It is worth noting that the theorem we have just established gives

a practical method to confirm the existence of the Bergman kernel—by
relating it to the Green’s function, whose existence is elementary.
Now let us calculate. Of course the Green’s function of the unit
disc D is
1 1
G(ζ, z) = log |ζ − z| − log |1 − ζ z| ,
2π 2π
About the Bergman Metric 151

as a glance at any classical complex analysis text will tell us (see, for
example, [COH] or [HIL]).
With formula () in mind, we can make life a bit easier by writing
1 1
G(ζ, z) = log(ζ − z) + log(ζ − z)
4π 4π
1 1  
− log (1 − ζ z) − log 1 − ζ z .
4π 4π
Here we think of the expression on the right as the concatenation of
four multi-valued functions, in view of the ambiguity of the logarithm
function. This ambiguity is irrelevant for us because the derivative of
the Green’s function is still well defined (i.e., the derivative annihilates
additive constants).
Now we readily calculate that
∂G 1 −1 1 ζ
= · + ·
∂z 4π ζ − z 4π 1 − ζ z
and
∂2G 1 1
= · .
∂ζ ∂z 4π (1 − ζ z)2
In conclusion, we may apply Proposition 2 to see that
1 1
K (z, ζ ) = · .
π (1 − z · ζ )2
This result is consistent with that obtained in the first two calculations
(Subsections 4.3.1, 4.3.2).

4. About the Bergman Metric

The Bergman kernel is an important conformal invariant. For the
purposes of the present book, this invariance manifests itself most
strikingly in the form of the Bergman (sometimes called the Poincaré-
Bergman) metric.
152 Introduction to the Bergman Theory

In fact we first notice that when
is a bounded domain, K (z, z)

is always positive. This assertion is most easily seen from the series
construction of the Bergman kernel. At any point z 0 we have

∞
0 = K (z 0 , z 0 ) = |φ j (z 0 )|2 .
j=1

It is immediate that K is non-negative on the diagonal. If there were a z 0

such that K (z 0 , z 0 ) = 0, then the last sum would be zero. It would then
follow that φ j (z 0 ) = 0 for all j. But, because the {φ j } are a complete
orthonormal system for A2 , and any f ∈ A2 can be expanded in terms
of the φ j , it would then follow that f (z 0 ) = 0 for all f ∈ A2 . This is
false for the function f ≡ 1.
Since K (z, z) is always positive, we may consider the quantity
log K (z, z). We define the Bergman metric to be that metric given by
the weight function

∂2
ρ
(z) = ρ(z) = log K
(z, z).
∂z∂z
Notice that the logarithm under the square root is certainly real, and
the double derivative is real since it does not change under conjuga-
tion. That the expression under the square root sign is positive follows

because we may differentiate the series expansion j φ j φ j of K :
  
∂2 ∂ 1 
log K (z, z) = · φjφj 

∂z∂z ∂z j φjφ j j

1  
= − 2 · φ j φ j · φ j φ j

j φjφ j j j

1 
+  · φ j φ j
j φjφ j j


2 
1  1

= |φ j |2 − ·
φ φ 

j
.
j |φ j | j |φ j |
2 2 j
j j
About the Bergman Metric 153

Of course the expression in square brackets is nonnegative by the

Cauchy-Schwarz inequality.
Our first claim is that the Bergman metric ρ
is in fact an invariant
metric:

Proposition 1. Let
1 ,
2 be bounded, conformally equivalent do-

mains. Let  :
1 →
2 be a conformal mapping. Then

∗ ρ
2 (z) = ρ
1 (z).

Proof. We calculate that

∗ ρ
2 (w) = | (w)| · ρ
2 ((w))

  2 

=  (w) log K
2 (z, z)

 (w)
∂z∂z z=(w)

∂2  
= log  (w) · K
2 ((w), (w)) (w)
∂w∂w
Here we have inserted terms with derivative zero, and have also applied
the chain rule. Now we have that this expression

∂2
= log K
1 (w, w)
∂w∂w
= ρ
1 (w).

Thus the Bergman metric is conformally invariant.

Clearly any metric that is invariant under conformal mappings is

a matter of great interest. This book provides ample proof of these as-
sertions. Let us next calculate the Bergman metric for the disc D.

Example 1. We know that the Bergman kernel for the disc D is

1 1
K (z, ζ ) = · .
π (1 − z · ζ )2
154 Introduction to the Bergman Theory

Therefore
1 1
K (z, z) = · .
π (1 − |z|2 )2

In particular, we see (as we already know a priori) that K restricted to

the diagonal is positive. Thus log K (z, z) makes sense and

log K (z, z) = − log π − 2 log(1 − |z|2 ).

As a result,

∂ −2 2z
log K (z, z) = · (−z) = .
∂z 1 − |z|2 1 − |z|2

Therefore

∂2 2
log K (z, z) = .
∂z∂z (1 − |z|2 )2

In conclusion,
√
2
ρ D (z) = .
1 − |z|2

This is just the same (up to a constant multiple) as the Poincaré metric
which we calculated for the disc in Section 1.4.

In fact it can be argued—using ideas from Section 1.4—that the

only invariant metric (up to a constant multiple) on the disc—of the
type we have been studying in this book—is the Poincaré metric. So
it is no surprise that the Poincaré and Bergman metrics coincide. By
contrast, the reader may check for himself that the metric

z−w

ψ(z, w) = log

1 − z · w

is an invariant metric on the disc (i.e., conformal transformations φ

of the disc preserve distances in this metric, so that ψ(φ(z), φ(w)) =
More on the Bergman Metric 155

ψ(z, w)) but that ψ is not of the form studied in the present book. There
is no contradiction because ψ is not obtained by integration against a
weight function ρ. [The metric defined here is sometimes termed the
“pseudohyperbolic metric.”]

5. More on the Bergman Metric

The fact that the Bergman metric is obtained as the Laplacian of a po-
tential function (in this case log K (z, z)) means that the metric is a
Kähler metric. This means, roughly speaking, that the geometric struc-
ture and the complex structure are compatible in a certain sense—see
[KOK] for the full story about Kähler metrics. The potential function
is automatically real analytic, hence so is the metric itself. These are
powerful facts when the Bergman metric is used in geometric calcula-
tions.
We know from classical function theory (see, for instance, [GRK]
and the discussion of Carathéodory’s theorem in Chapter 0) that a con-
formal mapping of two bounded domains, each simply connected and
having boundary consisting of a single Jordan curve, will extend uni-
valently and continuously to the boundary. So the conformal mapping
becomes a homeomorphism of the closures of the domains. This fact is
useful in transferring the function theory from one domain to the other.
It is useful in more advanced contexts to know that, when two domains
have smooth boundaries and are conformally equivalent, then a confor-
mal mapping of the domains will extend smoothly to the boundaries.
This problem was first explored by Painlevé in his Paris thesis [PAI] in
1888. He proved, in particular, that if each domain has C ∞ boundary
then the conformal mapping and its inverse each extend in a C ∞ man-
ner to the boundary. Painlevé’s result was generalized and refined by
Kellogg and Warschawski, among others.
In modern times, the Bergman kernel and metric have been used
to simplify and clarify (and also make more geometrical) the study
of boundary smoothness of conformal mappings. This idea was first
explored by C. Fefferman in [FEF1] in a much broader setting. It was
156 Introduction to the Bergman Theory

later simplified by S. Bell and his collaborators. A complete exposition,

in the context of one complex variable, appears in [BEK] or [GRK].

6. Application to Conformal Mapping

A classical result from complex function theory is this:

Theorem 1. Let A = {z ∈ C : 1 < |z| < R}. If φ is a conformal

mapping of A to itself, then φ is either a rotation or an inversion of the
form z → R/z.

In this section we would like to conduct a mathematically precise

but somewhat informal discussion, from the point of view of Bergman
geometry, of why this theorem is true. What is rewarding about this
treatment is that it illustrates how a geometer thinks. Moreover, all of
this purely geometric reasoning can, with some effort, be made ab-
solutely rigorous and analytical. We are pleased to thank Robert E.
Greene and John McCarthy for helpful discussions of this result.
It is interesting to note that it is quite difficult to compute the
Bergman kernel and metric for an annulus. One can certainly see (us-
ing Laurent series) that the monomials {z j }∞ j=−∞ form an orthogonal
system on an annulus centered at the origin. And one can use a little
calculus to normalize these to an orthonormal system. But actually per-
forming the necessary summation is virtually intractable (and involves
elliptic functions [BER, pp. 9–10]). Fortunately, the proof which we are
about to present requires no detailed knowledge of the Bergman metric
of the annulus. Indeed it uses only the fact that the metric blows up at
the boundary. We will get this information for free from the following
elegant and surprising result of K. T. Hahn:

Proposition 2. Let
⊆ C be a bounded domain. Let ρ
denote the

Bergman metric on
and let FC
be the Carathéodory metric. Then

ρ
(z) ≥ FC
(z)

for every z ∈
.
Application to Conformal Mapping 157

Of course we already know that the Carathéodory metric FC
is com-

plete, hence blows up at the boundary—see Theorem 6 of Section 3.3.
Proposition 2 (see [HAH] for the proof) tells us immediately that the
Bergman metric ρ is complete, hence blows up at the boundary of
.
See [APF] for more on the boundary behavior of the Bergman kernel
and its derivatives.
We use the Bergman metric to prove our results in this section,
rather than the Carathéodory or Kobayashi metrics, because we will
be doing calculations with geodesics. So we need a smooth, complete
metric (details on this point follow). We will also utilize the real analyt-
icity of the Bergman kernel and metric in an interesting and surprising
way.
We will use the language of geodesics. Going by the book, a
geodesic is defined by a differential equation. For our purposes here,
we may think of a geodesic as a locally length-minimizing curve. Now
fix an annulus A = {z ∈ C : 1 < |z| < R}. We will be using standard
polar coordinates (r, θ ). At any point p ∈ A, a vector in the tangent
space decomposes into a component in the ∂/∂r direction and a com-
ponent in the ∂/∂θ direction. One way to look at the metric is that
it assigns a length to ∂/∂r and to ∂/∂θ at each point of the annulus.
Consider the set M of points where the Bergman length of ∂/∂θ is
minimal. Such points exist just because the Bergman metric blows up
at the boundary of the domain. What geometric properties will the set
M have?
First, the set must be rotationally invariant—because the metric
will be rotationally invariant (i.e., the rotations are conformal self-maps
of A). Thus M is a union of circles centered at the origin. And M is
certainly a closed set by the continuity of the metric. The set has no in-
terior because the Bergman metric ρ is given by a real analytic function
(it is the second derivative of the logarithm of the real analytic Bergman
kernel), and the zero set of a non-trivial real analytic function can have
no interior—see [KRP1]. In fact we may also note that ∂/∂θ ρ is real
analytic in the radial r direction, hence (by the same reasoning) we
claim that M can have only finitely many circles in it. More precisely,
let g(r, θ ) = ∂/∂θ ρ,(r,θ) . We have already noted that this function is
158 Introduction to the Bergman Theory

independent of θ , so we may consider g(r ) = ∂/∂θ ρ,r . This func-

tion will assume its minimum value at points r where g  (r ) = 0.
Since the metric is complete, we know that the set of such points forms
a compact subset of the interval (0, R). If the set is infinite, then it has
an accumulation point and, therefore, the real analytic function g  is
identically zero. That is impossible, again by the completeness of the
metric. Therefore the set M is finite.
Now it is easy to see that any curve (circle) in M will also be one
of the curves (the curves that go once around the hole in the middle
of the annulus) that minimizes arc length in the Bergman metric; and
vice versa. This is so because we have already selected the curve to
have ∂/∂θ length as small as possible; a curve whose tangents have
components in the ∂/∂r direction will a fortiori be longer. In more
detail, the length of any curve γ (t), 0 ≤ t ≤ 1 is calculated by

1
1

ρ (γ ) = γ (t) ρ,γ (t) dt = γr (t) + γθ (t) ρ,γ (t) dt,
0 0

where γr and γθ are, respectively, the normal and tangential compo-

nents of γ . Clearly, if we construct a new curve  γ by integrating the
vector field γθ , then the result is a curve that is shorter than γ .
Thus any circle in M is a length-minimizing geodesic. And the
converse is true as well.
Now let φ be a conformal self-map of A. Then, as a result of the
considerations in the last paragraph, φ will map circles in M (concen-
tric with the annulus) to circles in M (concentric with the annulus). But
more is true: any circle c that is centered at 0 has constant Bergman dis-
tance from any one given circle C in M. Let τc be the Bergman distance
from the arbitrary circle c to the fixed circle C. Then c will be mapped
by φ to another circle in A that has Bergman distance τc from φ(C). In
short, φ maps circles to circles. And the same remark applies to φ −1 .
Now the orthogonal trajectories to the family of circles centered
at P will of course be the radii of the annulus A. Therefore (by confor-
mality) these radii will get mapped to radii. Let R be the intersection
of the positive real axis with the annulus A. After composition with a
rotation, we may assume that φ maps R to R.
Application to Conformal Mapping 159

If M has just one circle in it, then that circle C must be fixed.
Since each circle c in the annulus (centered at P) has a distance τc from
C, then its image under φ will have the same distance from C. Thus
either c is fixed or it is sent to its image under inversion. By continuity,
whatever choice is valid for c will be valid for all other circles. Thus,
in this case, the map φ is either the identity or inversion.
Now suppose that M contains at least two circles. The only possi-
bilities for the action of φ on M ∩ R are preservation of the order of the
points or inversion of the order (as any other permutation of the points
is ruled out by elementary topology of a conformal mapping). By com-
posing φ with an inversion (z → R/z), we may assume that φ does not
act as an inversion on R. But of course φ must preserve M ∩ R. We
conclude that φ must be the identity.
What we have just proved is that, after we normalize φ so that it
maps the positive real axis to itself, then in fact φ must be the identity
(or an inversion). In other words, the original map φ must be a rotation
(or an inversion).
CHAPTER 5
A Glimpse of Several Complex Variables

0. Functions of Several Complex Variables

At a naive level, the analysis of several real variables is much like the
real analysis of one variable; the main difference is that one deals with
n-tuples of reals instead of scalars and one needs matrices to keep track
of information. Of course deeper study reveals much complexity and
richness in analysis of several real variables. It is noteworthy that, for
the most part, this richness was discovered rather late in the history of
the subject—mostly in the last fifty years.
The history of analysis of several complex variables is quite differ-
ent. Early in the subject, in the first decade of the twentieth century, two
remarkable discoveries indicated that this area has an incredible depth
and variety which one complex variable does not even hint at. Even
today we have only scratched the surface of several complex variables.
Let us briefly discuss the two developments which established sev-
eral complex variables as a subject in its own right. The first is related
to the Riemann mapping theorem. As we have discussed throughout
the present volume, Riemann’s theorem asserts that, with the single ex-
ception of the plane, any domain topologically equivalent to the disc is
conformally equivalent to it. One might expect an analogue to this re-
sult in two complex variables. But, to begin with, what is the analogue
of the disc in C2 ? Two candidates come to mind: the ball

{(z 1 , z 2 ) ∈ C2 : |z 1 |2 + |z 1 |2 < 1}
161
162 A Glimpse of Several Complex Variables

and the bidisc

{(z 1 , z 2 ) ∈ C2 : |z 1 | < 1, |z 2 | < 1}.

Which of these should serve as the model domain for the multivariable
Riemann mapping theorem? Before answering that question, perhaps
one should ask whether there is a biholomorphic equivalence between
the ball and the bidisc. [According to a result of Liouville (see [DFN]),
there are none but trivial conformal mappings in dimensions three and
higher. As a result, in several complex variables we consider biholo-
morphic mappings. These are mappings which are holomorphic, one-
to-one, and onto. The holomorphicity of the inverse is automatic. See
further discussion below in Section 1.] If there is such an equivalence,
then the other question need not be answered; if there is no equivalence,
then perhaps there is no Riemann mapping theorem.
As Poincaré discovered, there is in fact no equivalence between
the ball and the bidisc. In the last two decades, it has been discovered
that, generically, two domains in C2 are not biholomorphically equiv-
alent. Thus the entire question of the Riemann mapping theorem turns
into a rather complex subject in itself—that of classifying domains ac-
cording to biholomorphic equivalence. One of the main things that we
shall do in this chapter is to formulate carefully and to prove Poincaré’s
theorem concerning the inequivalence of the ball and the bidisc.
The other big development which occurred near the turn of the
twentieth century is Hartogs’s theorem about analytic continuation. Al-
though we cannot treat this topic in detail here, we should like to dis-
cuss it briefly in order to further convey the flavor of several complex
variables. So that we may keep matters as simple as possible, we shall
avoid formal definitions in this discussion.
A domain U in complex space is called a domain of holomorphy
if there is a holomorphic function defined on U which cannot be ana-
lytically continued to any open set which properly contains U . It turns
out (see [KR1]) that the Weierstrass theorem of classical function the-
ory can be used to show that any open set U in the plane is a domain of
holomorphy. Briefly, let U be such an open set. Let S ⊆ U be a set in U
such that (i) S has no accumulation point in the interior of U and (ii) S
Functions of Several Complex Variables 163

Figure 1.

accumulates at every boundary point of U —see Figure 1. Then Weier-

strass’s theorem specifies the existence of a holomorphic function f on
U which vanishes precisely on S (and nowhere else). If this f were to
analytically continue past any boundary point p of U , then p would be
an interior accumulation point of the zero set of f . Thus f would be
identically zero on some connected component of U , a contradiction.
We conclude that f is a holomorphic function on U which cannot be
continued to any larger open set. See [GRK, p. 270] for the details of
this construction.
What Hartogs discovered is that, in dimensions 2 and higher, some
domains are domains of holomorphy and others are not. For example,
he proved that the domain

U = {(z 1 , z 2 ) : |z 1 | <, 2, |z 2 | < 2}

\ {(z 1 , z 2 ) : |z 1 | ≤ 1, |z 2 | ≤ 1}

has the property that any function holomorphic on U can be analyti-

cally continued to

= {(z 1 , z 2 ) : |z 1 | < 2, |z 2 | < 2}.
U

The full proof of this result may be found in [KR1]. As previously

noted, this theorem has no analogue in one complex variable and begs
the question: “Which domains in several complex variables have the
property exhibited by Hartogs’s phenomenon?” A necessary geometric
164 A Glimpse of Several Complex Variables

condition on the boundary of U was discovered by E.E. Levi, and the

problem of proving its sufficiency became known as the Levi problem.
The Levi problem was solved in considerable generality in the 1940’s
and 1950’s, but several related questions still remain open.
The point which we have tried to make in this brief essay is that
many of the most interesting questions in several complex variables
never come up in one complex variable, because they make no sense in
that context. What we can hope to accomplish in this chapter is to give
the reader a glance at the subject and suggest some of the techniques
that can be used effectively. One of our accomplishments will be to
give two fairly easy and elegant proofs of Poincaré’s theorem.

1. Basic Concepts
We introduce here only the few definitions and pieces of notation which
we shall need for our immediate purposes. For a more comprehensive
introduction to several complex variables, see [KR1, KR2, KR4]. For
simplicity of notation, we confine attention to two complex variables.
We consider domains U ⊆ C2 , where

C2 ≡ C × C

and (z 1 , z 2 ) is a typical element. Recall that C2 may be identified with

R4 by

C2 (z 1 , z 2 ) ←→ (x1 + iy1 , x2 + iy2 )

←→ (x1 , y1 , x2 , y2 ) ∈ R4 .

Thus a function of two complex variables is continuous if it is continu-

ous as a function of four real variables, it is continuously differentiable
if it is continuously differentiable as a function of four real variables,
and so forth.
Let U ⊆ C2 be a domain, that is, a connected open set. Let z and
w be complex numbers. We define

Uw = {z 1 ∈ C : (z 1 , w) ∈ U }
Basic Concepts 165

and

U z = {z 2 ∈ C : (z, z 2 ) ∈ U }.

Definition 1. A continuously differentiable function f on a domain U

is said to be holomorphic if, whenever z ∈ C satisfies U z = ∅, then
f (z, ·) is holomorphic on U z (as a function of one complex variable)
and whenever w ∈ C satisfies Uw = ∅, then f (·, w) is holomorphic on
Uw (as a function of one complex variable).

Example 1. The function

z1
f (z 1 , z 2 ) =
z2 + 1
is holomorphic on the domain

U1 = {(z 1 , z 2 ) : z 2 = −1}.

The function

g(z 1 , z 2 ) = z 1 (z 2 )2

is not holomorphic on any domain. The function


h(z 1 , z 2 ) = (z 1 ) j (z 2 ) j
j

is holomorphic on

U2 = {(z 1 , z 2 ) : |z 1 z 2 | < 1}.

In several complex variables there are two natural notions of

“neighborhood”: the ball and the bidisc. If P = (P1 , P2 ) ∈ C2 and
r > 0, then we let

B(P, r ) = {z ∈ C2 : |z 1 − P1 |2 + |z 2 − P2 |2 < r 2 }

and

D 2 (P, r ) = {z ∈ C2 : |z 1 − P1 | < r, |z 2 − P2 | < r }

166 A Glimpse of Several Complex Variables

be, respectively, the open ball and open bidisc of center P and radius r .
2
We also let B(P, r ) and D (P, r ) denote, respectively, the closed ball
and closed bidisc (defined by replacing the symbol < with the sym-
bol ≤).
A basic tool in complex analysis is the Cauchy integral formula
and its variants. We now derive one such.

2
Proposition 2. If f is holomorphic on a neighborhood of D (P, r )
(the closure of the bidisc), then for all (z 1 , z 2 ) ∈ D 2 (P, r ) we have

 
1 f (ζ1 , ζ2 )
f (z 1 , z 2 ) = dζ2 dζ1 .
(2πi)2 ∂ D(P1 ,r ) ∂ D(P2 ,r ) (ζ1 − z 1 )(ζ2 − z 2 )

Proof. The function f (·, z 2 ) is holomorphic in a neighborhood of

D(P1 , r ). Therefore the one-variable Cauchy integral formula yields,
for each fixed value of z 2 ,

1 f (ζ1 , z 2 )
f (z 1 , z 2 ) = dζ1 .
2πi ∂ D(P1 ,r ) (ζ1 − z 1 )

Now apply the one variable Cauchy integral formula to the function
f (ζ1 , z 2 ) in the integrand in the second variable: the desired formula
follows.

Corollary 2.1. A holomorphic function of two complex variables is in-

finitely differentiable as a function of four real variables.

Proof. Differentiate under the integral sign, just as in one complex vari-
able.

Corollary 2.2. A function f as in the Proposition has a power series

expansion

f (z 1 , z 2 ) = a jk (z 1 − P1 ) j (z 2 − P2 )k
Basic Concepts 167

2
which converges absolutely and uniformly on D (P, r ). The coeffi-
cients are given by
 j  k
1 ∂ ∂
a j,k = f (P).
j!k! ∂z 1 ∂z 2

Proof. The proof follows familiar lines from one complex variable.
2
Since f is holomorphic on a neighborhood of D (P, r ), we may select
2
r  > r so that f is holomorphic on a neighborhood of D (P, r  ). We
write
 
1 f (ζ1 , ζ2 )
f (z 1 , z 2 ) = dζ2 dζ1 .
(2πi) ∂ D(P1 ,r ) ∂ D(P2 ,r ) (ζ1 − z 1 ) · (ζ2 − z 2 )
2  

(∗)

However,
1 1
=
(ζ1 − z 1 ) (ζ1 − P1 ) − (z 1 − P1 )
1 1
= (∗∗)
(ζ1 − P1 ) 1 − ((z 1 − P1 )/(ζ1 − P1 ))
∞
(z 1 − P1 ) j
= .
(ζ − P1 ) j+1
j=0 1

Similarly,

1 ∞
(z 2 − P2 ) j
= . (∗ ∗ ∗)
(ζ2 − z 2 ) (ζ − P2 ) j+1
j=0 2

Convergence is absolute and uniform on {|z 1 − P1 | ≤ r } and {|z 2 −

P2 | ≤ r } respectively. Thus we may substitute (∗∗) and (∗ ∗ ∗) into (∗)
and switch the order of summation and integration to obtain

 ∞ 
∞   
1
f (z 1 , z 2 ) =
j=0 k=0
(2πi)2 ∂ D(P1 ,r ) ∂ D(P2 ,r )
168 A Glimpse of Several Complex Variables


f (ζ1 , ζ2 )dζ2 dζ1
(ζ1 − P1 ) j+1 (ζ2 − P2 )k+1
× (z 1 − P1 ) j (z 2 − P2 )k

∞
≡ a j,k · (z 1 − P1 ) j (z 2 − P2 )k .
j,k=0

Here
 
1 f (ζ1 , ζ2 )dζ2 dζ1
a j,k = .
(2πi)2 ∂ D(P1 ,r ) ∂ D(P2 ,r ) (ζ1 − P1 ) j+1 (ζ2 − P2 )k+1

The familiar Cauchy theory of one complex variable now tells us that
 j  k
1 ∂ ∂
a j,k = f (P).
j!k! ∂z 1 ∂z 2

For simplicity let us now consider power series

 j
S∼ a j,k z 1 z 2k

expanded about 0 = (0, 0). Notice that if S converges absolutely at

a point (z 1 , z 2 ) then it also converges absolutely and uniformly at the
points (µ1 z 1 , µ2 z 2 ) where |µ1 | ≤ 1, |µ2 | ≤ 1. This motivates the
following definition.

Definition 3. A domain U is called complete circular if whenever

(z 1 , z 2 ) ∈ U and |µ1 |, |µ2 | ≤ 1 then (µ1 z 1 , µ2 z 2 ) ∈ U . Note in
passing that a complete circular domain automatically contains 0.

Proposition 4. Let
 j
S∼ a j,k z 1 z 2k

be a power series. Then

C = {z = (z 1 , z 2 ) : S converges absolutely in a neighborhood of z}

Basic Concepts 169

is an open, complete circular domain. We call C the domain of conver-

gence of S.

Proof. Combine the definition of C with that of complete circular do-

main and the remark preceding it.

Although we did not assert the converse of the Proposition, it nev-

ertheless suggests that the ball and the bidisc might each be domains of
convergence for some power series. Indeed,
 j
S∼ z 1 z 2k

converges precisely on D 2 (0, 1) and on no larger open set. The power

series
 j
T ∼ (z 1 )2 + (z 2 )2
j

(when rearranged as a sum of monomials) converges absolutely on the

open ball, and on no larger open set. (Exercise: fill in the details of
these assertions.)
One of the themes of this book has been to consider conformal
self-maps of domains. It turns out that, as previously noted, “confor-
mal” is the wrong notion for functions of several complex variables.
The correct morphisms for our purposes are given in the following def-
inition:

Definition 5. Let U1 , U2 be domains in C2 . A function

F : U1 −→ U2

is called a biholomorphic mapping if

F = ( f 1 (z 1 , z 2 ), f 2 (z 1 , z 2 ))

is a pair of holomorphic functions, F is injective, F is surjective, and

F −1 is holomorphic.
170 A Glimpse of Several Complex Variables

Remark. It turns out that the condition that F −1 be holomorphic is

redundant; however this is hard to prove (see [KR1] for this result of
Osgood) so, for simplicity, we state the condition explicitly.

Example 2. Let U = D 2 (0, 1). Then

 
z 2 − 1/2
f (z 1 , z 2 ) = , i · z1
1 − (1/2)z 2
is a biholomorphic mapping of U to U .

As in the case of one variable, we let Aut(U ) denote the collection

of biholomorphic self-maps of U .

Example 3. Let U = B(0, 1). Then

√ 
z 1 − 1/2 ( 3/2)z 2
f (z 1 , z 2 ) = ,
1 − (1/2)z 1 1 − (1/2)z 1

is a biholomorphic mapping of the domain U . Of course any unitary

rotation is also a biholomorphic mapping of U .

Exercise. If U ⊆ C2 is a domain then Aut(U ) forms a group under

composition of mappings.

2. The Automorphism Groups of the Ball

and Bidisc
We shall calculate the automorphism groups of B(0, 1) and D 2 (0, 1).
We begin with some preliminary material.

Lemma 1. If f j are holomorphic functions on U and f j → f uni-

formly on compact subsets of U , then any sequence of derivatives
   
∂ ∂ k
fj
∂z 1 ∂z 2
The Automorphism Groups of the Ball and Bidisc 171

converges to
   k
∂ ∂
f,
∂z 1 ∂z 2
uniformly on compact sets.

Proof. The argument is the same as in one complex variable (Corol-

lary 5.1 of Section 0.1): Fix P ∈ U and choose r > 0 such that
2
D (P, r ) ⊆ U . Express f j on D 2 (P, r ) as a Cauchy integral of f j
on ∂ D 2 (P, r ). The result is then immediate from differentiation under
the integral sign.

Lemma 2. If f is holomorphic on a domain U ⊆ C2 and if f vanishes

on an open subset W ⊆ U then f ≡ 0 on U .

Proof. Let

C = {z ∈ U : f (z) = 0 in a neighborhood of z}.

Then C is clearly open and nonempty. On the other hand, if z ∈ ∂C ∩

U , then f and all its derivatives vanish at z. Thus the power series
expansion for f in a neighborhood of z is identically 0, and z cannot
be in ∂C. It follows that C has no boundary points in U . Therefore it
must hold that C = U .

Lemma 3. Let F be a family of holomorphic functions on a domain

U ⊆ C2 such that

| f (z)| ≤ M < ∞, ∀ f ∈ F.

Then F is a normal family.

Proof. It is enough to see that F is normal on any closed bidisc

2 2
D (P, r ) ⊆ U . Let D (P, r  ) ⊆ U be a slightly larger bidisc. Dif-
ferentiate the Cauchy integral formula on D 2 (P, r  ); the standard esti-
172 A Glimpse of Several Complex Variables

mates from one variable then give

 
 ∂f  M
  2
 ∂z (z) ≤ r  − r , ∀z ∈ D (P, r ), f ∈ F, j = 1, 2.
j

2
It follows that F is an equicontinuous family on D (P, r ). Therefore
the result follows from the Ascoli/Arzelà theorem.

Definition 4. If U1 , U2 are domains in C2 and

f : U1 −→ U2

is holomorphic then define JacC f (z), z ∈ U1 , (the Jacobian matrix of

f at z) to be the matrix
∂f ∂ f1 
1
(z) (z)
 ∂z 1 ∂z 2 
 .
 ∂ f2 ∂ f2 
(z) (z)
∂z 1 ∂z 2

Remark. Notice that JacC f (z) is distinct from the real Jacobian,
JacR f , of calculus: the latter would be a 4×4 matrix which arises from
treating f as a function from a domain in R4 to a domain in R4 .

Next we prove a Schwarz lemma in two complex variables. Al-

though it could be proved in an ad hoc manner for the ball and the
bidisc, the proof in general is not essentially more difficult; it also al-
lows us to recall a nice argument which we saw earlier in Section 3.2.

Proposition 5. (H. Cartan). Let U ⊆ C2 be a bounded domain and

P ∈ U . Let f : U → U be holomorphic and assume that f (P) = P.
If

JacC f (P) = id,

then

f (z) ≡ z.
The Automorphism Groups of the Ball and Bidisc 173

Proof. Assume without loss of generality that P = 0. Seeking a con-

tradiction, we assume that the conclusion is false.
In a neighborhood of 0 ∈ U we may expand f in a power series.
The hypotheses about f imply that this power series has the form

f (z 1 , z 2 ) = 0 + z + Am (z) + · · ·

where Am (z) represents the first nonvanishing monomial after z, Am

having degree m ≥ 2. Observe that f = ( f 1 , f 2 ) is an ordered pair and
that each summand on the right-hand side of our equation is an ordered
pair. Define

f 1 (z 1 , z 2 ) = f
f 2 (z 1 , z 2 ) = f ◦ f
···
f j (z 1 , z 2 ) = f j−1
◦ f, ∀ j ≥ 2.

Since U is bounded, the family { f j } forms a normal family (by

Lemma 3). Therefore there is a subsequence f j converging to a limit
function f . By Lemma 1, the mth derivatives of f j at P also converge
to the corresponding derivatives of 
f at P.
On the other hand, direct calculation shows that

f 2 = z + 2Am + · · ·
f 3 = z + 3Am + · · ·
etc.

and we see that in fact the mth derivatives of f j at P = 0 blow up.

The only way to resolve this contradiction is to conclude that Am ≡ 0
near P. Therefore Am ≡ 0 on U . We conclude that f (z) ≡ z, as
desired.
174 A Glimpse of Several Complex Variables

Proposition 6. Let U be a bounded complete circular domain. If

f : U −→ U

is a biholomorphic mapping and f (0) = 0 then f is linear.

Remark. This is a generalization to two variables of the fact that a

conformal map of the disc to the disc that preserves the origin is in fact
a rotation.

Proof of the Proposition. Let θ ∈ [0, 2π) and

ρθ (z 1 , z 2 ) = (eiθ z 1 , eiθ z 2 ).

Consider the map

g = ρ−θ ◦ f −1 ◦ ρθ ◦ f.

Then

JacC g(0) = ρ−θ ◦ (JacC f (0))−1 ◦ ρθ ◦ (JacC f (0)).

Here we think of ρ±θ as a linear map given by the matrix

 ±iθ 
e 0
,
0 e±iθ

hence ρ±θ is its own Jacobian. Since ρ±θ is a multiple of the identity,
it commutes with all other 2 × 2 matrices. So

JacC g(0) = id.

Now the preceding proposition applies and we may conclude that

g(z) ≡ z.

This means that

f ◦ ρθ = ρθ ◦ f. (∗)
The Automorphism Groups of the Ball and Bidisc 175

We now write f in a convergent power series in a neighborhood of 0:

 j
f (z) = a j,k z 1 z 2k .

Then (∗) says that

 j
 j
a j,k ei( j+k)θ z 1 z 2k = a j,k eiθ z 1 z 2k

for all θ (where this is an equation of ordered pairs). Equating like

monomials implies that

a j,k = 0

unless j + k = 1. It follows that f is linear near 0. By the uniqueness

of analytic continuation (Lemma 2), f is linear.

Proposition 7. Let φ ∈ Aut(D 2 (0, 1)). Then there exist a1 , a2 ∈

D(0, 1), θ1 , θ2 ∈ [0, 2π), and a permutation σ of the set {1, 2} such
that
 
z σ (1) − a1 z σ (2) − a2
φ(z) = eiθ1 · , eiθ2 · .
1 − a 1 z σ (1) 1 − a 2 z σ (2)

Proof. Let φ(0) = α = (α1 , α2 ). Define

 
z 1 − α1 z 2 − α2
ψ(z) = , .
1 − α1 z1 1 − α2 z2

Then g ≡ ψ ◦ φ ∈ Aut(D 2 (0, 1)) and g(0) = 0. It suffices to show that


g(z 1 , z 2 ) = eiθ1 z σ (1) , eiθ2 z σ (2) .

We know from Proposition 6 that g is linear. So

  
g g12 z1
g(z) = 11
g21 g22 z2

for some complex constants gi j of modulus not exceeding one. Define,

for k = 2, 3, . . . , points z 1,k , z 2,k ∈ D 2 (0, 1) by
176 A Glimpse of Several Complex Variables


z 1,k = (1 − 1/k)sgn(g11 ), (1 − 1/k)sgn(g12 ) ,

z 2,k = (1 − 1/k)sgn(g21 ), (1 − 1/k)sgn(g22 ) .

Here, for w ∈ C, we are letting


w/|w|, w = 0,
sgn w ≡
0, w = 0.

Now

g(z 1,k ) has first component(1 − 1/k)(|g11 | + |g12 |)

and

g(z 2,k ) has second component(1 − 1/k)(|g21 | + |g22 |).

Letting k → +∞ gives

|g11 | + |g12 | ≤ 1
(∗)
|g21 | + |g22 | ≤ 1.

On the other hand, apply g to α k = (1−1/k, 0) and β k = (0, 1−1/k)

and let k → +∞. We conclude that

((1 − 1/k)g11 , (1 − 1/k)g21 ) −→ ∂ D 2 (0, 1)

and

((1 − 1/k)g12 , (1 − 1/k)g22 ) −→ ∂ D 2 (0, 1).

Therefore

max{|g11 |, |g21 |} = 1
max{|g12 |, |g22 |} = 1. (∗∗)

The only way that (∗) and (∗∗) can both be true is if each column
of the matrix (gi j ) has one entry of modulus one and one entry 0. Say
The Automorphism Groups of the Ball and Bidisc 177

that |gη(k),k | = 1 for k = 1, 2, where η is a permutation of {1, 2}.

Letting σ = η−1 gives that

|g j,σ ( j) | = 1, j = 1, 2

and

g j,k = 0 if k = σ ( j).

Let g j,σ ( j) = eiθ j , j = 1, 2. Then

g(z 1 , z 2 ) = (eiθ1 · z σ (1) , eiθ2 · z σ (2) ),

as desired.

An elegant and complete treatment of the biholomorphic self-maps of

the unit ball may be found in [RU2]. We give a more elementary treat-
ment here.

Proposition 8. If a ∈ C, |a| < 1, then the map


z 1 − a (1 − |a|2 )1/2 z 2
φa (z 1 , z 2 ) = ,
1 − az 1 1 − az 1

is a biholomorphic self-map of B(0, 1) to B(0, 1).

Remark. The maps φa are natural generalizations of the Möbius trans-

formations which we used extensively in the first part of the book.

Proof of the Proposition. Now

|φa (z)| < 1

if and only if
 2  
 z1 − a   2 1/2 2
  +  (1 − |a| ) z 2  < 1
 1 − az   1 − az 
1 1
178 A Glimpse of Several Complex Variables

if and only if

|z 1 − a|2 + (1 − |a|2 )|z 2 |2 < |1 − az 1 |2

if and only if

|z 1 |2 + |a|2 + (1 − |a|2 )|z 2 |2 < 1 + |a|2 |z 1 |2

if and only if

(1 − |a|2 )|z 1 |2 + (1 − |a|2 )|z 2 |2 < 1 − |a|2

if and only if

z ∈ B(0, 1).

Exercise. The inverse of the mapping φa is φ−a .

Recall that a 2 × 2 matrix

M : C2 −→ C2

is unitary if its inverse equals its conjugate transpose. Geometrically,

M is unitary if and only if it takes any orthonormal basis of C2 (over
the base field C) to another.

Proposition 9. If g : B(0, 1) → B(0, 1) is biholomorphic and

g(0) = 0, then g is a unitary rotation (i.e., a mapping induced by
a unitary matrix).

Proof. Since B(0, 1) is complete circular, the map g must be linear (by
Proposition 6). Now g must preserve the boundary of B(0, 1). [Why?—
Think about a linear map acting on a closed, convex set; it will preserve
the extreme points.] Therefore the function g takes Euclidean unit vec-
tors to Euclidean unit vectors. But this says that g is unitary.

Now we can characterize the automorphisms of the unit ball.

The Automorphism Groups of the Ball and Bidisc 179

Proposition 10. If f : B(0, 1) → B(0, 1) is a biholomorphic map-

ping, then f is the composition of at most two unitary maps and a map
of the form φa .

Proof. Let w = f (0). There is a unitary rotation α such that

α(w) = (|w|, 0).

The map

g = φ|w| ◦ α ◦ f

is therefore a biholomorphic map of B(0, 1) to B(0, 1) which takes 0

to 0. By Proposition 9, g is a unitary map. Call it β. Then

f = α −1 ◦ (φ|w| )−1 ◦ β.

The reader may verify that (φ|w| )−1 = φ−|w| . The proof is therefore
complete.

Exercise. The group Aut(D 2 (0, 1)) acts transitively on D 2 (0, 1). The
group Aut(B(0, 1)) acts transitively on B(0, 1).

The ball and the bidisc are the two most fundamental domains in
C2 . Both are contractible, both have transitive automorphism groups,
and both are domains of convergence for power series. It came there-
fore as a great surprise when Poincaré proved that there is no biholo-
morphic mapping

φ : D 2 (0, 1) −→ B(0, 1).

He proved this result by first showing that if such a φ existed, then the
induced map

Aut(D 2 (0, 1)) h −→ φ ◦ h ◦ φ −1 ∈ Aut(B(0, 1))

would be a group isomorphism. He then analyzed Aut(D 2 (0, 1)) and

Aut(B(0, 1)) as groups and showed that they could not be isomorphic.
180 A Glimpse of Several Complex Variables

Indeed, we may assume that φ maps 0 to 0. Then φ induces an isomor-

phism of the isotropy group of 0 (the collection of self-maps that fix
0) in the connected component of the identity in Aut(D 2 (0, 1)) with
the isotropy group of 0 in the connected component of the identity in
Aut(B(0, 1)). It is easy to see that the former is abelian while the latter,
being the unitary group, is not. Thus these subgroups cannot be isomor-
phic and we have a contradiction. We have given a sufficiently detailed
description of the automorphism groups that the interested reader may
carry out Poincaré’s program as an exercise.
We shall instead approach the biholomorphic equivalence problem
using some of the geometric ideas developed in this monograph. The
next section is devoted to that topic.

3. Invariant Metrics and the Inequivalence

of the Ball and the Bidisc
We continue to use the symbol D to denote the unit disc in C. Let U ⊆
C2 be a domain and P ∈ U . Define (D, U ) P to be the holomorphic
functions f : U → D such that f (P) = 0. Define (U, D) P to be the
holomorphic functions f : D → U such that f (0) = P. We are now
going to define the Carathéodory and Kobayashi metrics. Since we are
working in a two-dimensional space, we can no longer specify a metric
as a scalar-valued function on the domain. In fact a metric will measure
the length of a vector at a point.

Definition 1. If P ∈ U and ξ ∈ C2 , then define the Carathéodory

length of ξ at P to be

FCU (P, ξ ) = sup{|JacC f (P) · ξ | : f ∈ (D, U ) P }.

Note here that JacC f (P) is a 2-tuple of complex numbers (because it

is the derivative of a mapping from two complex dimensions to one
complex dimension) and so is ξ . Thus JacC f (P) · ξ is the dot product
of two 2-tuples, and hence is a complex number.
Invariant Metrics and the Inequivalence of the Ball and the Bidisc 181

Definition 2. If P ∈ U and ξ ∈ C2 , then define the Kobayashi length

of ξ at P to be

FKU (P, ξ ) = inf{|ξ |/|g  (0)| : g ∈ (U, D) P ,

g  (0) is a scalar multiple of ξ }.

Here and throughout, | | denotes Euclidean length.

The way that we define metrics is still motivated by Riemann’s

philosophy. If γ : [0, 1] → U is a continuously differentiable curve,
we define its Kobayashi length to be
 1
K (γ ) = FKU (γ (t), γ  (t)) dt.
0

Notice that we are integrating the lengths, in the metric, of the tangent
vectors to the curve. The Carathéodory length of a curve is defined
similarly. We comment, just as in Section 3.1, that FC and FK are inte-
grable because they are semicontinuous.
One of the basic properties which we shall prove is that holomor-
phic mappings decrease distance in the Carathéodory and Kobayashi
metrics. In several variables we express this assertion as follows.

Proposition 3. Let U1 and U2 be domains and

f : U1 −→ U2

be a holomorphic mapping. If P ∈ U1 and ξ ∈ C2 , we define

f ∗ (P)ξ ≡ JacC f (P) · ξ.

Then

FCU1 (P, ξ ) ≥ FCU2 ( f (P), f ∗ (P)ξ )

and
U U
FK 1 (P, ξ ) ≥ FK 2 ( f (P), f ∗ (P)ξ ).
182 A Glimpse of Several Complex Variables

Proof. We give the proof for the Carathéodory metric. The proof for
the Kobayashi metric is similar.
Choose φ ∈ (D, U2 ) f (P) . Then φ ◦ f ∈ (D, U1 ) P . Hence

FCU1 (P, ξ ) ≥ |(JacC (φ ◦ f )(P))ξ |

= |JacC φ( f (P)) · (JacC f (P)) · ξ |
= |JacC φ( f (P)) · ( f ∗ (P))ξ |.

Taking the supremum over all φ gives

U U
FC 1 (P, ξ ) ≥ FC 2 ( f (P), f ∗ (P)ξ ).

Corollary 3.1. If f is a biholomorphic map, then f preserves both the

Carathéodory and the Kobayashi metrics; that is, the inequalities in
the Proposition become equalities.

Proof. Obvious. Apply the proposition both to f and to f −1 .

Exercise. Check that the Proposition implies that f decreases the in-
variant lengths of curves. That is, if γ is a continuously differentiable
curve in U1 and f ∗ γ ≡ f ◦ γ is the corresponding curve in U2 , then
show that C ( f ∗ γ ) ≤ C (γ ) and K ( f ∗ γ ) ≤ K (γ ).

We use Proposition 3 to define two interesting new invariants.

These invariants were trivial in one complex variable, but now they
provide crucial information.

Definition 4. Let U ⊆ C2 be a domain and P ∈ U . The Carathéodory

indicatrix of U at P is

iCP (U ) = {ξ ∈ C : FCU (P, ξ ) < 1}.

The Kobayashi indicatrix of U at P is

[i KP (U ) = {ξ ∈ C : FKU (P, ξ ) < 1}.

Invariant Metrics and the Inequivalence of the Ball and the Bidisc 183

In words, the indicatrix is the “unit ball,” in the indicated metric,

of vectors at P. Technically speaking, the indicatrix lives in the tangent
space at P.

Proposition 5. Let f : U1 → U2 be a biholomorphic mapping of

domains in C2 . Say that f (P) = Q. Then

JacC f (P) : iCP (U1 ) −→ iCQ (U2 )

and

JacC f (P) : i KP (U1 ) −→ i K

Q (U2 )

are linear isomorphisms.

Proof. Since f is distance-decreasing in the Kobayashi metric,

JacC f (P) maps i KP (U1 ) into i K
Q (U2 ). But the same observation applies
to

JacC ( f −1 )(Q) = (JacC f (P))−1 ;

it maps i K Q (U2 ) into i P (U1 ). Thus JacC f (P) is a linear isomorphism of

K

i KP to i K
Q as claimed.
The proof for iCP is identical.

Proposition 6. Let B = B(0, 1). Then we have

i0K (B) = B.

Proof. Let φ ∈ (B, D)0 . If η is any Euclidean unit vector in C2 then

consider the function

h(ζ ) ≡ φ(ζ ) · η.

Here “·” denotes the usual inner product of 2-vectors. We have that h
maps the disc to the disc and h(0) = 0. By the Schwarz lemma of one
184 A Glimpse of Several Complex Variables

variable,

|h  (0)| ≤ 1.

Since this inequality holds for any choice of η, we conclude that

|φ  (0)| ≤ 1.

Now if ξ is any vector in C2 then it follows from the preceding

calculation that

FKB (0, ξ ) = inf{|ξ |/|φ  (0)| : φ ∈ (B, D)0 }

≥ |ξ |.

On the other hand, the map

ζ
φ0 (ζ ) ≡ ξ
|ξ |

satisfies φ0 ∈ (B, D)0 and φ0 (ζ ) = ξ/|ξ | for any ζ . Thus φ0 (0) =
(1/|ξ |) · ξ is a positive multiple of ξ . Therefore

FKB (0, ξ ) ≤ |ξ |/|φ0 (0)| = |ξ |.

We conclude that

FKB (0, ξ ) = |ξ |,

hence that

i0K (B) = B.

Proposition 7. Let D 2 = D 2 (0, 1). Then

i0K (D 2 ) = D 2 .

Proof. Define the projections

π1 (z 1 , z 2 ) = z 1 and π2 (z 1 , z 2 ) = z 2 .
Invariant Metrics and the Inequivalence of the Ball and the Bidisc 185

Let η = (η1 , η2 ) ∈ C2 be any vector. By Proposition 3, we have that

2 π (D 2 )
FKD (0, η) ≥ FK 1 (π1 (0), (π1 )∗ η) = FKD (0, η1 ).

But the Schwarz lemma of one variable tells us easily that the last quan-
tity is just |η1 |. A similar argument shows that
2
FKD (0, η) ≥ |η2 |.

We conclude from these two inequalities that

2
FKD (0, η) ≥ max{|η1 |, |η2 |}.

Therefore

i0K (D 2 ) ⊆ D 2 .

For the reverse inclusion, fix η as above and consider the function
 
ζ η1 ζ η2
φ(ζ ) = , , ζ ∈ D.
max{|η1 |, |η2 |} max{|η1 |, |η2 |}

Then it is obvious that φ ∈ (D 2 , D)0 and we see that

 
 η1 η2
φ (ζ ) = , , ζ ∈ D.
max{|η1 |, |η2 |} max{|η1 |, |η2 |}

So φ  (0) is a positive multiple of η.

Therefore
2 |η|
FKD (0, η) ≤ = max{|η1 |, |η2 |}.
|φ  (0)|
The opposite inclusion now follows.

0 (B) = B and i0 (D ) = D .
Exercise. Verify that iC C 2 2

Theorem 8. (Poincaré). There is no biholomorphic mapping of the

bidisc D 2 to the ball B.
186 A Glimpse of Several Complex Variables

Proof. Suppose that

φ : D 2 −→ B

is biholomorphic. Let φ −1 (0) = α ∈ D 2 . There is an element ψ ∈

Aut(D 2 (0, 1)) such that ψ(0) = α. Consider g ≡ φ ◦ ψ. Then

g : D 2 −→ B

is biholomorphic and g(0) = 0. We will show that g cannot exist.

By Proposition 5, JacC g(0) is a linear isomorphism of i0K (D 2 ) to
i0 (B). But Propositions 6 and 7 identify these as D 2 and B respec-
K

tively. So we have that

JacC g(0) : D 2 −→ B

is a linear isomorphism. However, this is impossible. For the segment

i = {(t + i0, 1) : 0 ≤ t ≤ 1} lies in ∂ D 2 . The linear isomorphism
would map i to a nontrivial segment in ∂ B. But B is strictly convex (all
boundary points are extreme) so its boundary contains no segments.
This is the required contradiction.

Remark. It is important to appreciate the logic in this proof. The hy-

pothesized map φ (and therefore g as well) is not assumed to extend
in any way to ∂ D 2 . Indeed, given the very different natures of ∂ D 2
and ∂ B, we would expect φ to be highly pathological at the boundary.
Our geometric machinery allows us to pass to the linear map JacC g(0),
which is defined on all of space. Thus we are able to analyze the bound-
aries of the domains and arrive at a contradiction.

An important program begun by Poincaré in the early part of the

twentieth century can be described as follows: If φ : 1 → 2 is
a biholomorphic mapping of smoothly bounded domains in C2 , then
suppose that φ, φ −1 continue smoothly to the boundaries of 1 , 2 ,
respectively, in such a way that the extended functions are diffeomor-
phisms of the closures of the domains. Using just dimension theory
Invariant Metrics and the Inequivalence of the Ball and the Bidisc 187

(the number of degrees of freedom in φ versus the number of degrees

of freedom in specifying the two real (2n − 1)-dimensional bound-
aries), it can be shown that infinitely many algebraic relations must be
satisfied by the derivatives of functions parametrizing ∂1 and ∂2 .
These relations give rise to invariants which can play a vital role in the
biholomorphic classification of domains.
It is only since 1974 that a substantial beginning has been made in
fleshing out Poincaré’s program—see the discussion at the end of Sec-
tion 4.5. At about the same time, Chern and Moser [CHM] and Tanaka
[TAN] created systems for calculating the boundary differential invari-
ants. In his monumental work [FEF2], Fefferman actually gives an ef-
fective procedure for calculating the Chern/Moser/Tanaka invariants on
a large and important class of domains.
More recently, S. R. Bell [BEL] and others have come up with
simpler proofs of Fefferman’s theorem for a much larger class of do-
mains. While it is still believed that a biholomorphic mapping of any
two smoothly bounded domains must extend smoothly to their respec-
tive boundaries, we are far from being able to prove this assertion.
Epilogue 189

Epilogue
In complex analysis, geometric methods provide both a natural lan-
guage for analyzing and recasting classical problems and also a rubric
for posing new problems. The interaction between the classical and the
modern techniques is both rich and rewarding.
Many facets of this symbiosis have yet to be explored. In particu-
lar, very little is known about explicitly calculating and estimating the
differential invariants described in the present monograph. It is hoped
that this book will spark some new interest in these matters.
Appendix on the Structure Equations
and Curvature

1. Introduction
Here we give a brief presentation of the connection between the cal-
culus notion of curvature (see [THO]) and the more abstract notion of
curvature which leads to the definition of κ in Chapter 2. It is a plea-
sure to acknowledge our debt to the clear and compelling exposition in
[ONE].
First, a word about notation. We use the language of differential
forms consistently in this appendix. On the one hand, classically trained
analysts are often uncomfortable with this language. On the other hand,
the best way to learn the language is to use it. And the context of cur-
vature calculations on plane domains may in fact be the simplest non-
trivial context in which differential forms can be profitably used. In any
event, this appendix would be terribly clumsy if we did not use forms,
so the decision is essentially automatic. All necessary background on
differential forms may be found in [RU1] or in [ONE].

2. Expressing Curvature Intrinsically

First we recall the concept of curvature for a smooth, two-dimensional
surface M ⊆ R3 . All of our calculations are local, so it is convenient
to think of M as parametrized by two coordinate functions over a con-
nected open set U ⊆ R2 :

191
192 Appendix on the Structure Equations and Curvature

p
U  (u, v)−→(x1 (u, v), x2 (u, v), x3 (u, v)) ∈ M.

We require that the matrix

 
∂ x1 ∂ x2 ∂ x3
 ∂u ∂u ∂u 
 
 ∂ x1 ∂ x2 ∂ x3 
∂v ∂v ∂v
have rank 2 at each point of U . The vectors given by the rows of
this matrix span the tangent plane to M at each point. Applying the
Gram-Schmidt orthonormalization procedure to these row vectors, and
shrinking U, M if necessary, we may create vector fields

E 1 : M −→ R3
E 2 : M −→ R3

such that E 1 (x1 , x2 , x3 ) and E 2 (x1 , x2 , x3 ) are orthonormal and tan-

gent to M at each P = (x1 , x2 , x3 ) ∈ M. Denote by T P (M) the collec-
tion of linear combinations

a E 1 (x1 , x2 , x3 ) + bE 2 (x1 , x2 , x3 ), a, b ∈ R.

We call TP (M) the tangent space to M at P.

Let E 3 (P) be the unit normal to M at P given by the E 1 (P) ×
E 2 (P). The functions E 1 , E 2 , E 3 are smooth vector fields on M: they
assign to each P ∈ M a triple of orthonormal vectors.
Let δ1 , δ2 , δ3 denote the standard basis for vectors in R3 :

δ1 = (1, 0, 0),
δ2 = (0, 1, 0),
δ3 = (0, 0, 1).

(Many calculus books call these vectors i, j, and k.) Then we may
write

Ei = ai, j (x1 , x2 , x3 )δ j , i = 1, 2, 3.
j
Expressing Curvature Intrinsically 193

The matrix
 3
A ≡ ai, j i, j=1 ,

where the ai, j are functions of the space variables, is called the atti-
tude matrix of the frame (or basis) E 1 , E 2 , E 3 . Since A transforms one
orthonormal frame to another, A is an orthogonal matrix. Hence

A−1 = t A.

Definition 1. If P ∈ M, v ∈ T P (M), and f is a smooth function on

M, then define

Dv f (P) = f ◦ φ(t)

,
dt t=0

where φ is any smooth curve in M such that φ(0) = P and φ  (0) = v.

One checks that this definition is independent of the choice of φ.

Definition 2. If

α : M −→ R3

is a vector field on M,

α(P) = α1 (P)δ1 + α2 (P)δ2 + α3 (P)δ3 ,

and v ∈ TP (M) then define

v α(P) = (Dv α1 )(P)δ1 + (Dv α2 )(P)δ2 + (Dv α3 )(P)δ3 .

The operation v is called covariant differentiation of the vector

field α.

Definition 3. If P ∈ M, v ∈ TP (M), we define the shape operator (or

Weingarten map) for M at P to be

S P (v) = − v E 3 (P).
194 Appendix on the Structure Equations and Curvature

Lemma 4. We have that S P (v) ∈ T P (M).

Proof. Now E 3 · E 3 ≡ 1, hence

0 = Dv (E 3 · E 3 )

= (2 v E 3 ) · E 3

P
= −2S P (v) · E 3 (P).

Hence S P (v) ⊥ E 3 (P) so S P (v) ∈ T P (M).

Notice that the shape operator assigns to each P ∈ M a linear

operator S P on the 2-dimensional tangent space T P (M).
We can express this linear operator as a matrix with respect to the
basis E 1 (P), E 2 (P). So S assigns to each P ∈ M a 2 × 2 matrix M P .
The linear operator S P measures the rate of change of E 3 in any
tangent direction v. It can be shown that M P is diagonalizable. The
(real) eigenvectors of M P correspond to the principal curvatures of
M at P (these are the directions of greatest and least curvature) and the
corresponding eigenvalues measure the amount of curvature in those
directions.
The preceding observations about M P motivate the following
definition.

Definition 5. The Gaussian curvature κ(P) of M at a point P ∈ M is

the determinant of M P —the product of the two eigenvalues.

Our aim is to express κ(P) in terms of the intrinsic geometry of

M, without reference to E 3 —that is, without reference to the way that
M is situated in space.
To this end, we define covector fields θi which are dual to the
vector fields E i :

θi E j (P) = δi j .
Expressing Curvature Intrinsically 195

Then the θi may be expressed as linear combinations of the standard

basis covectors d x1 , d x2 , d x3 . Indeed, if A = (ai, j ) is the attitude ma-
trix then

θi = ai, j d x j

(remember that A−1 = t A). Thus, for each i, θi is a differential form;

and the standard calculus of differential forms—including exterior dif-
ferentiation—applies.
From now on, we restrict attention to 1- and 2-forms acting on
tangent vectors to M. Thus any 1-form α may be expressed as

α = α(E 1 )θ1 + α(E 2 )θ2

and any 2-form β may be expressed as

β = β(E 1 , E 2 )θ1 ∧ θ2 .

Now we define covector fields ωi, j , i, j ∈ {1, 2, 3}, by the formula

ωi, j (v) = ( v E i ) · E j (P).

Here “·” is the Euclidean dot product. We think of ωi, j as a differential

1-form. Notice that, since E i · E j ≡ δi, j , we have for v ∈ T P (M) that

0 = Dv (E i · E j )
= ( v E i ) · E j + E i · ( v E j )
= ωi, j (v) + ω j,i (v).

Thus

ωi, j = −ω j,i .

In particular,

ωi,i = 0.
196 Appendix on the Structure Equations and Curvature

If v ∈ T P (M) then it is easy to check, just using linear algebra, that


v E i = ωi, j (v)E j , 1 ≤ i ≤ 3.
j

We call the ωi, j the connection forms for M. We can now express the
shape operator in terms of these connection forms.

Proposition 6. Let P ∈ M and v ∈ TP (M). Then

S P (v) = ω1,3 (v)E 1 (P) + ω2,3 (v)E 2 (P).

Proof. We have that

S P (v) = − v E 3

3
= − ( v E 3 · E j )E j
j=1


3
= − ω3, j (v)E j
j=1

= ω1,3 (v)E 1 + ω2,3 (v)E 2 ,

since ω3,3 = 0.

Now we can express Gaussian curvature in terms of the ωi, j .

Proposition 7. We have that

ω1,3 ∧ ω2,3 = κ θ1 ∧ θ2 .

Proof. We need to calculate M P in terms of the ωi, j . Proposition 6

gives that

S P (E 1 ) = ω1,3 (E 1 )E 1 + ω2,3 (E 1 )E 2

and

S P (E 2 ) = ω1,3 (E 2 )E 1 + ω2,3 (E 2 )E 2
Expressing Curvature Intrinsically 197

so the matrix of S P , in terms of the basis E 1 , E 2 , is

 
ω1,3 (E 1 ) ω2,3 (E 1 )
MP = .
ω1,3 (E 2 ) ω2,3 (E 2 )

We know that ω1,3 ∧ ω2,3 , being a 2-form, can be written as λ · θ1 ∧ θ2 .

On the other hand,

κ = det M P
= ω1,3 (E 1 )ω2,3 (E 2 ) − ω1,3 (E 2 )ω2,3 (E 1 )
= (ω1,3 ∧ ω2,3 )(E 1 , E 2 )
= λ.

Therefore

ω1,3 ∧ ω2,3 = λ θ1 ∧ θ2 = κ θ1 ∧ θ2 .

Our goal now is to express κ using only those ωi, j with i = 3,

j = 3. For this we require a technical lemma about the attitude matrix.

Lemma 8. We have that


ωi, j = a j,k dai,k , 1 ≤ i, j ≤ 3.
k

Proof. If v ∈ TP (M) then

ωi, j (v) = v E i · E j (P).

But

Ei = ai,k δk .
k

Therefore

v E i = (Dv ai,k )δk .
k
198 Appendix on the Structure Equations and Curvature

Then

ωi, j (v) ≡ v E i · E j
 
 
= (Dv ai,k )δk · a j,k δk
k k

= (Dv ai,k )a j,k
k

= dai,k (v)a j,k .
k

As a result,

ωi, j = a j,k dai,k .
k

Now we have reached a milestone. We can derive the important

Cartan structural equations, which are the key to our intrinsic formulas
for curvature.

Theorem 9. We have

dθi = ωi, j ∧ θ j , (1)
j

dωi, j = ωi,k ∧ ωk, j . (2)
k

Proof. We have that


θi = ai, j d x j
j

hence

dθi = dai, j ∧ d x j .
j
Expressing Curvature Intrinsically 199

Since the attitude matrix A is orthogonal, we may solve the equations

in Lemma 8 for dai,k in terms of ωi, j . Thus

dai, j = ωi,k ak, j .
k

Substituting this into our last formula gives

  
 
dθi = ωi,k ak, j ∧ dx j
j k
 
 
= ωi,k ∧ ak, j d x j
k j

= ωi,k ∧ θk .
k

This is the first structural equation.

For the second equation, notice that the formula

ωi, j = a j,k dai,k
k

implies

dωi, j = − dai,k ∧ da j,k .
k

On the other hand,

 
   
ωi,k ∧ ωk, j = ak, dai, ∧ a j,m dak,m
k k  m
 
  
= ak, dai, ∧ − ak,m da j,m
k  m
  
 
=− ak, ak,m · dai, ∧ da j,m
,m k
200 Appendix on the Structure Equations and Curvature


=− dai,m ∧ da j,m
m
= dωi, j .

In the antepenultimate line we have used the fact that A−1 = t A. The
result now follows.

The following corollary will prove critical.

Corollary. We have

dω1,2 = −κθ1 ∧ θ2 .

Proof. The second structural equation gives


dω1,2 = ω1,k ∧ ωk,2
k
= ω1,1 ∧ ω1,2 + ω1,2 ∧ ω2,2 + ω1,3 ∧ ω3,2 .

Only the third summand doesn’t vanish. But we calculated in Proposi-

tion 7 that it equals −κθ1 ∧ θ2 .

The Corollary has been our main goal in this subsection. It gives
an intrinsic way to calculate Gaussian curvature in the classical setting,
hence a way to define Gaussian curvature in more abstract settings. We
now proceed to develop this more abstract point of view.

3. Curvature Calculations on
Planar Domains
Let  ⊆ C be a domain which is equipped with a metric ρ. Assume
for simplicity that ρ(z) > 0 at all points of . Define functions
(1, 0) (0, 1)
E1 ≡ and E2 ≡ .
ρ ρ
Curvature Calculations on Planar Domains 201

Then

θ1 = ρ d x and θ2 = ρ dy

are the dual covector fields. We define ωi, j according to the first struc-
ture equation:

dθ1 = ω1,2 ∧ θ2 ,
dθ2 = ω2,1 ∧ θ1 .

We define Gaussian curvature according to the Corollary to Theorem 9

in the previous section:

dω1,2 = −κθ1 ∧ θ2 .

One can check that these definitions are independent of the choice of
frame (or basis) E 1 , E 2 , but this is irrelevant for our purposes.
We conclude this Appendix by proving that the definition of cur-
vature which we just elicited from the structural equations coincides
with the one given in Section 2.1. First, by the way that we’ve defined
θ1 and θ2 , we have

dθ1 = dρ ∧ d x
= (ρx d x + ρ y dy) ∧ d x
= ρ y dy ∧ d x
ρy
= − d x ∧ ρ dy
ρ
ρy
= − d x ∧ θ2 .
ρ
Similarly,

dθ2 = dρ ∧ dy
= (ρx d x + ρ y dy) ∧ dy
= ρx d x ∧ dy
202 Appendix on the Structure Equations and Curvature

ρx
=− dy ∧ ρ d x
ρ
ρx
= − dy ∧ θ1 .
ρ

Comparison with the first structural equation gives

ρy
ω1,2 = − d x + τ dy
ρ

and
 
ρx
ω1,2 = −ω2,1 = − − dy + σ d x,
ρ

for some unknown functions τ and σ .

The only way these equations can be consistent is if

ρy ρx
ω1,2 = − dx + dy.
ρ ρ

Thus
   
∂ ρy ∂ ρx
dω1,2 = − dy ∧ d x + d x ∧ dy
∂y ρ ∂x ρ
   
ρ yy ρy ρy ρx x ρx ρx
= − + 2 dy ∧ d x + − 2 d x ∧ dy
ρ ρ ρ ρ
1  
= 2 ρρ − (ρ y )2 − (ρx )2 d x ∧ dy
ρ
1  
= 4 ρρ − (ρ y )2 − (ρx )2 θ1 ∧ θ2 .
ρ

The second structural equation now implies that

1
κ=− (ρρ − | ρ|2 ).
ρ4
Curvature Calculations on Planar Domains 203

On the other hand, in Section 2.1 we defined

 log ρ
κ =− .
ρ2
We have
∂ 1 ∂ρ
log ρ =
∂ z̄ ρ ∂ z̄
and
∂2
 log ρ = 4 log ρ
∂z∂ z̄

1 ∂ρ ∂ρ 1 ∂ 2ρ
=4 − 2 +
ρ ∂z ∂ z̄ ρ ∂z∂ z̄
1 1
=− · | ρ|2 + · ρ.
ρ 2 ρ
It follows that, according to the definition in Section 2.1,
1
κ=− (ρρ − | ρ|2 ).
ρ4
Thus, as claimed, the definition of curvature from Chapter 2 is equal to
that which arises from the structural equations.
 205

Table of Symbols

Page
Symbol Number Meaning

A 193 the attitude matrix

A2 () 138 the Bergman space
Ar,R 122 an annulus
Aut(U ) 121, 179 the automorphism group of U
[az + b]/[cz + d] 14 linear fractional transformation
B(0, r ) 48 Poincaré metric ball
B(P, R) 128 metric ball
B(P, r ) 165 a ball in C2
C 1 complex numbers


C 82 the Riemann sphere
C 62 arc of a circle
C 168 domain of convergence
C0,1 76 the domain C \ {0, 1}
C1 3 continuously differentiable
C2 105 twice continuously differentiable
Ck 105 k times continuously differentiable
C∞ 106 infinitely differentiable
C (P, Q) 37 all piecewise continuously
differentiable curves connecting
P to Q
c(P) 111 center of curvature
CR 53 points in D having Poincaré
distance R from 0
C(z, ζ ) 140 the Cauchy kernel
D 2 unit disc
d 111 internally tangent disc
 40 the Laplacian
dρ 47 the Poincaré distance
Dv 193 directional derivative
δ1 , δ2 , δ3 192 standard basis vectors
dA 138 area measure
det Jac φ 141 the Jacobian determinant
D(P, r ) 2 open disc
D(P, r ) 2 closed disc
206

Page
Symbol Number Meaning

∂ D(P, r ) 2 boundary of disc

D (0, ) 87 a punctured disc
D(C(P), r0 ) 110 externally tangent disc
D(C (P), r0 ) 110 internally tangent disc
D 2 (P, r ) 165 a bidisc in C2
2
D (P, r ) 166 closure of a bidisc
(D, U ) P 90, 180 holomorphic functions f from
U to D such that f (P) = 0
distC (z, τ p ) 120 Carathéodory distance of
z to τ p
dρ (P, Q) 37 ρ-metric distance of P to Q
dρ (P, Q) 47 Poincaré distance
dσ (z, w)  83 spherical distance
dωi, j =
 k ωi,k ∧ ωk, j 198 Cartan structural equations
dθi = j ωi, j ∧ θ j
dV 139 volume measure
E1, E2 192 a frame (basis) of vector fields
E3 192 the unit normal vector field
F 2 a holomorphic function
F 2 complex derivative of F
F 16 a normal family
FCU (P) 90, 180 Carathéodory metric on U at P
FKU (P) 94, 181 Kobayashi metric on U at P
f∗γ 43, 181 the push-forward of γ
f∗ 42 the pullback mapping
f # (z) 85 the spherical derivative
fn 58, 101 nth iterate of f
f A2 () 138 the Bergman norm
 f, g A2 () 139 the Bergman inner product
G 81 a family of holomorphic functions
γ 3, 30, 31 a curve
γ̇ 30 derivative of γ
γ P,Q (t) 50 curve of least Poincaré length
connecting P to Q
α (P) 118 Stolz region, non-tangential
approach region
 207

Page
Symbol Number Meaning

G(z, ζ ) 148 the Green’s function

id 122 the identity mapping
i(z) 127 the identity map
iCP (U ) 182 Carathéodory indicatrix
iKP (U ) 182 Kobayashi indicatrix
I (z) 83 inversion mapping
γ F(z) dz 4 complex line integral
∞ 86 point at infinity on the Riemann
sphere
Jac φ 141, 172 the Jacobian matrix
K 17, 18 a compact set
κ = κ(P) 67 curvature of the metric ρ
κU,ρ (z) = κρ (z) 67 curvature of the metric ρ
K 139, 140 the Bergman kernel for 
λ 30 a metric
 61, 62 a line in Euclidean geometry
(γ ) 30 length of γ
ρ (γ ) 34 length of curve γ in metric ρ
 K (γ ) 181 Kobayashi length of the curve γ
Mβ (P) 119 metric approach region
MP 194 matrix of the shape operator
µ(z) 76 metric of negative curvature
on C \ {0, 1}
∇v α(P) 193 covariant derivative
νP 108 unit outward normal at P
ν P 108 unit inward normal at P
ωi, j 195 connection forms
P 82 the north pole of the Riemann sphere
Pk 25 kth degree Taylor polynomial
p(x, y) 82 the stereographic projection map
p(z) 9 a polynomial
 8 a domain
(1 , ρ1 ) 43 metric pair
∂/∂z, ∂/∂z 38 complex differential operators
φa (z 1 , z 2 ) 177 biholomorphic mapping of B
φa (ζ ) 13 a Möbius transformation
{φ j } 145 an orthonormal system for A2 ()
ψ(z, w) 155 the pseudohyperbolic metric
208

Page
Symbol Number Meaning

 129 a homotopy
R 14 the real numbers
r (P) 111 radius of curvature
ρ 31 a metric (weight)
ρ 70 the Poincaré metric
ρ(z) 106 defining function
ραA (z) 72 dilated, scaled Poincaré distance
metric
ρC 103, 104 Carathéodory metric
ρ
E 103, 104 Euclidean metric
ρK 103, 104 Kobayashi metric
ρ 152 Bergman metric
ρr (z) 70 the dilated Poincaré metric
ρτ 14 a rotation
ρ(z)|dz| 34 a conformal metric
S P (v) 193 shape operator or
Weingarten map
σ 122 the reflection map
σ (z) 69 the spherical metric
σ0 20 extremal function

a (z − P) j 2 power series expansion
j j
a jk (z 1 − P1 ) j (z 2 − P2 )k 166 power series expansion
T 109 tubular coordinate mapping
T P (M) 192 the tangent space to M at P
τ 84 induced Euclidean metric on sphere
τp 120 inward normal segment at p
θi 194 covector fields
U 2, 8 a domain
(U, D) P 93, 180 holomorphic functions f from
D to U such that f (0) = P
U0 81 the slit plane
Uw 164 slice of a domain in C2
Uz 165 slice of a domain in C2
W 108 tubular neighborhood
ξ ρ,z 31 metric length of ξ
|ξ | 31 Euclidean length of ξ
References

[AHL1] L. Ahlfors, An extension of Schwarz’s lemma, Trans. Amer. Math.

Soc. 43 (1938), 359–364.
[AHL2] ——, Complex Analysis, 3rd. ed. McGraw-Hill, New York, 1979.
[APF] L. S. Apfel, thesis, Washington University, 2003.
[BEK] S. R. Bell and S. G. Krantz, Smoothness to the boundary of conformal
mappings, Rocky Mountain J. Math. 17(1987), 23–40.
¯
[BEL] S. R. Bell, Biholomorphic mappings and the ∂-problem, Ann. Math.,
114 (1981), 103–112.
[BER] S. Bergman, The Kernel Function and Conformal Mapping, 2nd ed.,
American Mathematical Society, Providence, 1970.
[BOAS] R. P. Boas, Invitation to Complex Analysis, Random House, New York,
1987.
[CHM] S. S. Chern and J. Moser, Real hypersurfaces in complex manifolds,
Acta Math. 133(1974), 219–271.
[COH] R. Courant and D. Hilbert, Methods of Mathematica Physics, Inter-
science Publishers, New York, 1953–1962.
[DFN] B. A. Dubrovin, A. T. Fomenko, and S. P. Novikov, Modern Geometry—
Methods and Applications, Spring-er-Verlag, New York and Berlin, 1984.
[EAH] C. Earle and R. Hamilton, A fixed point theorem for holomorphic map-
pings, Proc. Symp. Pure Math., Vol. XVI, 1968, 61–65.
[EPS] B. Epstein, Orthogonal Families of Analytic Functions, MacMillan, New
York, 1965.
[FK] H. Farkas and I. Kra, Riemann Surfaces, Springer-Verlag, New York and
Berlin, 1979.
[FEF1] C. Fefferman, The Bergman kernel and biholomorphic mappings, In-
vent. Math., 26 (1974), 1–65.

209
210

[FEF2] C. Fefferman, Parabolic invariant theory in complex analysis, Adv.

Math. 31(1979), 131–262.
[FUK] B. A. Fuks, Introduction to the Theory of Analytic Functions of Several
Complex Variables, Translations of Mathematical Monographs, American
Mathematical Society, Providence, R.I., 1963.
[GAR] J. Garnett, Analytic Capacity and Measure, Springer Lecture Notes
#297 Springer-Verlag, 1972.
[GRK] R. E. Greene and S. G. Krantz, Function Theory of One Complex Vari-
able, 2nd ed., American Mathematical Society, Providence, R.I., 2002.
[HAH] K. T. Hahn, Inequality between the Bergman metric and Carathéodory
differential metrics, Proc. Am. Math. Soc. 68(1978), 193–194.
[HAL] P. R. Halmos, Introduction to Hilbert Space and the Theory of Spectral
Multiplicity, 2nd ed., AMS Chelsea Publishing, Providence, R.I., 1998.
[HEI] M. Heins, On the number of 1-1 directly conformal maps which a
multiply-connected plane region of finite connectivity p(> 2) admits onto
itself, Bull. AMS 52(1946), 454–457.
[HIL] E. Hille, Analytic Function Theory, 2nd ed., Chelsea Publishing, New
York, 1982.
[KOB1] S. Kobayashi, Hyperbolic Manifolds and Holomorphic Mappings,
Marcel Dekker, New York, 1970.
[KOB2] S. Kobayashi, Hyperbolic Complex Spaces, Springer-Verlag, Berlin,
1998.
[KOK] S. Kobayashi and K.Nomizu, Foundations of Differential Geometry,
Vols. I and II, Interscience, New York, 1963, 1969.
[KR1] S. Krantz, Function Theory of Several Complex Variables, 2nd ed.,
American Mathematical Society, Providence, R.I., 2000.
[KR2] ——, What is several complex variables?, Amer. Math. Monthly, 94
(1987), 236–256.
[KR3] ——, Functions of one complex variable, The Encyclopedia of Physical
Science and Technology, Academic Press, New York, 1987.
[KR4] ——, Functions of several complex variables, The Encyclopedia of
Physical Science and Technology, Academic Press, New York, 1987.
[KR5] ——, The Elements of Advanced Mathematics, 2nd ed., CRC Press,
2002.
[KRP1] S. Krantz and H. R. Parks, A Primer of Real Analytic Functions, 2nd
ed., Birkhäuser, Boston, in press.
[KRP2] S. Krantz and H. R. Parks, The Geometry of Domains in Space,
Birkhäuser, Boston, 1999.
References 211

[KRP3] S. Krantz and H. R. Parks, The Implicit Function Theorem, Birkhäuser,

Boston, 2002.
[KRP4] S. Krantz and H. R. Parks, Distance to C k manifolds, Jour. Diff. Equa-
tions 40(1981), 116-120.
[LS] L. Loomis and S. Sternberg, Advanced Calculus, Addison-Wesley, Read-
ing, 1969.
[MIS] D. Minda and G. Schober, Another elementary approach to the theorems
of Landau, Montel, Picard and Schottky, Complex Variables, 2 (1983), 157–
164.
[MUN] J. Munkres, Elementary Differential Topology, Princeton University
Press, Princeton, 1963.
[ONE] B. O’Neill, Elementary Differential Geometry, Academic Press, New
York, 1966.
[PAI] P. Painlevé, Sur les lignes singulières des fonctions analytiques, Annales
de la Faculté des Sciences de Toulouse (1) 2 (1888), 1–130.
[PRI] I. Privalov, Randeigenschaften Analytischer Functionen, Deutsch Verlag
der Wissenschaften, Berlin, 1956.
[RU1] W. Rudin, Principles of Mathematical Analysis, 3rd ed., McGraw-Hill,
New York, 1976.
[RU2] ——, Function Theory in the Unit Ball of Cn , Springer-Verlag, Berlin,
1981.
[RU3] ——, Real and Complex Analysis, 3rd ed., McGraw-Hill, New York,
1987.
[TAN] N. Tanaka, On generalized graded Lie algebras and geometric struc-
tures, J. Math. Soc. Japan 19(1967), 215–254.
[THO] G. Thomas, and R. Finney, Calculus and Analytic Geometry, Addison-
Wesley, Reading, 1984.
[WOL] J. Wolf, Spaces of Constant Curvature, 4th ed., Publish or Perish Press,
Berkeley, 1977.
Index

abstract notion of curvature, 191 Bell, Steven R., 155, 187

Ahlfors, Lars V., 70 Bergman kernel, 138, 139
Ahlfors/Schwarz lemma, 72 constructed with an orthonormal
analytic capacity, 91 expansion, 145
analytic function, 2 constructed with differential
annulus equations, 147
automorphism group of, 123 four characterizing properties of,
compactness of automorphism 139
group of, 124 positivity of, 152
conformal mappings of, 156 real analyticity of, 157
non-transitivity of, 131 transformation under conformal
approach region, 118 maps, 141
arc length, 30, 34 Bergman metric, 151
area measure, 138 conformal invariance of, 153
argument principle, 1, 10 extremal curves in, 158
Ascoli/Arzelà theorem, 17, 85 for the disc, 153
attitude matrix, 193 geodesics in, 157
automorphism group, 121, 170 orthogonal trajectories in, 158
characterization of compactness of, real analyticity of, 157
126 Bergman space, 139
compactness of, 123 of annulus, basis for, 156
examples of, 122 Bergman, Stefan, 137
topology on, 123 bidisc, 166
of B(0, 1), 170 in C2 , 162
automorphisms of, 175
ball, 166 biholomorphic mapping, 162, 169
and bidisc, biholomorphic of the unit ball, 170
inequivalence of, 179 problem, 187
automorphisms of, 178 biholomorphisms of complete circular
in C2 , 161 domains, 174

213
214 Index

Bloch’s principle, 82 C k curve, examples, 106

Bolyai, Janos, 58 closed, k times continuously
Brouwer invariance of domain, 123 differentiable curve, 105
closed, twice continuously
calculus differentiable curve, 105
in the complex domain, 38 coincidence of different definitions of
notion of curvature, 191 curvature, 201
canonical invariant metric, 138 compactly divergent, 79, 84
Carathéodory, Constantin, 89 compactness, 16, 17
and Kobayashi metrics, comparison complete circular domain, 168
of, 94 completeness
indicatrix, 182 of a metric, 37
isometries, 93 of the Carathéodory metric, 113,
length, 180 117
length of a curve, 181 of the Kobayashi metric, 113
metric, 89, 90 complex
completeness of, 105 analysis, 29
distance-decreasing analysis of several variables, 161
property, 90, 91 calculus notation, 41
examples, 91 derivative, 2, 39
for the disc, 93 differentiable, 2
invariance of, 90 differential operators, 38
isometries of, 92 line integral, 4
non-degeneracy of, 95 plane and negative curvature, 134
upper estimate for, 116 polynomials, 6
theorem, 22, 123 conformal map, 45
Cartan structural equations, 198 conformal self-maps of the disc, 13,
Cartan’s theorem, 172 14, 45
Casorati–Weierstrass theorem, 23, 78 conformally equivalent domains, 19
Cauchy connection forms, 196
estimates, 1, 8 continuously differentiable, 2, 3
integral formula, 1, 6 contraction mappings, 57
in C2 , 166 convergence of derivatives, 8
theorem, 4 convergent power series, 6
kernel, 140 cotangent vector, 42
Cauchy–Riemann equations, 2, 39 covariant differentiation, 193
center of curvature, 111 covector fields, 194
chain rule for differentiation, 40 cross product, 192
characterization of the disc in terms of curvature, 30, 67
the Carathéodory and Kobayashi holomorphic functions, 73
metrics, 97 definition of, 67
Chern, S. S., 187 for a smooth, two-dimensional
circle of curvature, 111 surface, 191
C k boundary, 105 invariance of, 68
alternative definition, 106 of the Euclidean metric, 69
Index 215

of the Poincaré metric, 69 omitted values, 75

provenance of, 68 equibounded, 17
curve of least length, 30 equicontinuity, 17
curve of least Poincaré length, 50 equicontinuous, 17
essential singularity, 23, 24
defining function, 106 Euclid’s axioms, 60
differential Hilbert’s rendition of, 60
equations, 29 Euclidean
forms, 191, 195 dot product, 195
invariants, 187 length, 31, 33, 181
dilations, 14 notion of length, 35
direction exterior differentiation, 195
of greatest curvature, 194 externally tangent disc, 111
of least curvature, 194 extremal function, 89
directional differentiation, 193
disc factorization of polynomials, 9
as complete metric space, 49 Farkas-Ritt theorem, 56
automorphism group of, 122 Fefferman, Charles, 155, 187
automorphism group, transitivity finitely connected domain,
of, 131 automorphism group of, 125
characterization by fixed point
non-compactness of as the limit of iterates, 57
automorphism group, 125 of a holomorphic function, 57
non-compactness of automorphism function of two complex variables,
group of, 125 continuity of, 164
notation, 1 differentiability of, 164
tautness of, 135 fundamental theorem of algebra, 9, 78
distance in a metric, 37
distance-decreasing property Gauss quote about Bolyai’s work, 59
of holomorphic mappings, 181 Gaussian curvature, 69, 194, 196, 201
of the Poincaré metric, 72 in terms of ωi, j , 196
of biholomorphic mappings, 182 intrinsic calculation of, 200
domain generalized Schwarz lemma, 86
of convergence of a power series, geodesic, 35
169 arc, 51
of holomorphy, 162 in Bergman metric, 157
with C k boundary, 106 in hyperbolic geometry, 62
with infinitely many holes, 106 Gram-Schmidt orthonormalization
with transitive automorphism procedure, 192
group, 131 great circular arc, 83
dual covector fields, 201 Green’s function, 150

elliptic functions, 156 Hahn’s theorem, 157

elliptic modular function, 26 harmonic
entire function, 74 analysis, 29
216 Index

harmonic (continued) Jacobian matrix

function, 41 complex, 172
Hartogs real, 172
domain, 163
extension phenomenon, 162 k times continuously differentiable
phenomenon, 163 curve, 105
Heine–Borel theorem, 16 Kähler metric, 32, 155
Hermitian metric, 31 Kellogg, Oliver, 155
Hilbert space, 138, 143 Kobayashi, S.
holomorphic function, 2 and Poincaré metrics, comparison
differentiability of, 166 of, 97
distance-decreasing property of, 55 indicatrix, 182
examples of in two variables, 165 of the ball, 183
much like polynomials, 8 of the bidisc, 184
power series expansion of, 166 isometries, 95
of two complex variables, 165 length, 181
holomorphic logarithm, 41 of a curve, 181
Hurwitz’s theorem, 10, 84 metric, 90, 94
hyperbolic completeness of, 105
disc, 58 distance-decreasing property
domain, 133 of, 90, 95
hyperbolicity extremal property of, 133
in terms of curvature, 134 invariance of, 90
of multi-punctured plane, 134 non-degeneracy of, 95
Kobayashi/Royden metric, 90, 94
infinitely distance-decreasing property of,
connected domain with 90
non-compact automorphism invariance of, 90
group, 132
differentiable boundary, 106 Laplace operator, 40
differentiable function, 106 Laplacian, 155
inner product, 139 length
internally tangent disc, 111 in a metric, 34
intrinsic geometry, 194 of a continuously differentiable
invariance curve, 30
of the Carathéodory indicatrix, of a curve in the Poincaré metric,
183 34
of the Kobayashi indicatrix, 183 of a piecewise continuously
inverse function theorem, 110 differentiable curve, 34
inversion, 14, 156 Levi, E.E., 164
isometries, 42, 43 problem, 164
composition of, 45 Lindelöf principle, 118, 119
properties of, 43 line, 61
rigidity of, 54 linear fractional transformation, 14,
that fix a point, 100 83
Index 217

linear properties of complex orthogonal matrix, 193

derivatives, 40 osculating discs, 111
Liouville’s theorem, 1, 9, 74, 81, Osgood’s theorem, 170
162
Lobachevsky, Nicolai, 58 Painlevé, Paul, 155
Lusin area integral, 140 theorem, 22, 155
parallel, 60
parallel postulate, 61
Marty’s criterion, 84 independence of, 64
Marty’s theorem, 86 parametrization of a surface, 191
maximum principle, 1, 11, 123 Picard, Emil,
mean value property, 11 great theorem, 25, 87
meromorphic functions, 83 little theorem, 25, 78
metric, 30, 31 geometric proof of, 78
axioms, 37 proof of, 27
examples of, 32 theorems, 25, 74, 81, 133
geometry, 1 piecewise continuously differentiable
in two complex variables, 180 curve, 3
of negative curvature on plane and negative curvature, 134
twice-punctured plane, 75 Playfair’s formulation of the parallel
metrics, comparability of, 103 postulate, 61
model for geometry, 61 Poincaré, Henri
Möbius transformation, 13, 46, -Bergman metric, 151
61 distance, 47
Montel theorem, 86 metric, 32, 45, 154
Montel’s theorem, 1, 16, 17, 80 balls in, 48
morphisms, 42 characterization of, 52
Moser, Jurgen, 187 completeness of, 49
conformal invariance of, 45,
nearest boundary point, 110 52
negatively oriented curve, 3 explicit calculation of, 47
neighborhood in two complex maps which preserve, 53
variables, 165 neighborhood basis for, 48
non-Euclidean, 58 program, 187
non-Euclidean geometry, 61 theorem, 164
example of, 63 points in hyperbolic geometry, 63
nondegeneracy of the Kobayashi pole, 23, 24
metric, 133 at infinity, 24
nontangential approach, optimality of, positively oriented curve, 3
121 potential function, 155
nontangential power series
approach region, 118 coefficients, 7
limit, 118 expansion, 2
normal family, 16, 80, 83, 86, 171 principal
normal family, examples of, 80 curvatures, 194
normally convergent, 79 normal, 111
218 Index

product rule for differentiation, 40 singularities of a holomorphic

pseudohyperbolic metric, 155 function, 22
pullback, 42 special function space, 138
metric, 83 spherical
of a metric, 42 derivative, 85
of a metric under a conjugate metric, 69, 82
holomorphic function, 42 Stolz region, 118
punctured disc, 116, 132 structural equations of Cartan, 69
punctured plane sum rule for differentiation, 40
and negative curvature, 134
non-tautness of, 135 tangent
bundle, 31
radial boundary limit, 118 space, 192
removable singularity, 23, 24 vector, 31, 33, 42
at infinity, 24 taut domain, 135
reproducing tautness and hyperbolicity, 135
kernel, 137 Taylor series, 7
property, 140 topologically trivial curve, 4
Riemann mapping topology induced by invariant metrics,
function, 89 105
Riemann mapping theorem, 16, 18, transcendental entire functions, 78
19, 89, 130, 161 transitive group action, 130
outline of proof, 19 translations, 14
Riemann removable singularities tubular neighborhood, 108
theorem, 7, 23 twice continuously differentiable
Riemann’s philosophy, 181 curve, 105
Riemannian metric, 32
rigid motions of the Euclidean plane, undefinable terms of geometry, 60
44 uniform continuity, 17
rotation, 156 unit disc, 2
unit inward normal, 108
scalar multiplication rule for unit outward normal, 108
differentiation, 40 unitary
Schwarz lemma, 1, 12, 14, 55, 94 biholomorphisms, 178
Ahlfors’s version of, 70 unitary matrix, 178
in two complex variables, 172
uniqueness portion, 102 value distribution theory, 27
Schwarz–Pick lemma, 15, 55 vector
set of uniqueness for a holomorphic attached to base point, 32
function, 171 field, 192
several complex variables, 164 with position, 33
shape operator, 193
action of, 194 Warschawski, Stefan, 155
simple closed curve, 3 Weierstrass’s theorem, 163
Index 219

zeros
of a function given by a power
series, 8
of holomorphic functions, 31