Anda di halaman 1dari 589

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

This page intentionally left blank

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Harmonic Measure
During the last two decades several remarkable new results were discovered about
harmonic measure in the complex plane. This book provides a survey of these results
and an introduction to the branch of analysis that contains them. Many of these results,
due to Bishop, Carleson, Jones, Makarov, Wolff, and others, appear here in book form
for the first time.
The book is accessible to students who have completed standard graduate courses in
real and complex analysis. The first four chapters provide the needed background
material on univalent functions, potential theory, and extremal length, and each chapter
has many exercises to further inform and teach the reader.
J O H N B . G A R N E T T is Professor of Mathematics at the University of California,

Los Angeles.
D O N A L D E . M A R S H A L L is Professor of Mathematics at the University of

Washington.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

ii

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

NEW MATH MONOGRAPHS


Editorial Board
Bla Bollobs
William Fulton
Frances Kirwan
Peter Sarnak
Barry Simon
For information about Cambridge University Press mathematics publications
visit http://publishing.cambridge.org/stm/mathematics

iii

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

iv

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Harmonic Measure
JOHN B. GARNETT
University of California, Los Angeles
DONALD E. MARSHALL
University of Washington

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, So Paulo


Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521470186
Cambridge University Press 2005
This publication is in copyright. Subject to statutory exception and to the provision of
relevant collective licensing agreements, no reproduction of any part may take place
without the written permission of Cambridge University Press.
First published in print format 2005
eBook (EBL)
ISBN-13 978-0-511-34894-5
ISBN-10 0-511-34894-0
eBook (EBL)
hardback
ISBN-13 978-0-521-47018-6
hardback
ISBN-10 0-521-47018-8

Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

To Dolores and Marianne

vii

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

viii

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Contents

Preface

page xiii

I.
1.
2.
3.
4.
5.

Jordan Domains
The Half-Plane and the Disc
Fatous Theorem and Maximal Functions
Carathodorys Theorem
Distortion and the Hyperbolic Metric
The HaymanWu Theorem
Notes
Exercises and Further Results

1
1
6
13
16
23
25
26

II.
1.
2.
3.
4.

Finitely Connected Domains


The Schwarz Alternating Method
Greens Functions and Poisson Kernels
Conjugate Functions
Boundary Smoothness
Notes
Exercises and Further Results

37
37
41
50
59
66
66

III.
1.
2.
3.
4.
5.
6.
7.
8.

Potential Theory
Capacity and Greens Functions
The Logarithmic Potential
The Energy Integral
The Equilibrium Distribution
Wieners Solution to the Dirichlet Problem
Regular Points
Wiener Series
Polar Sets and Sets of Harmonic Measure Zero

ix

73
74
77
79
82
89
93
97
102

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Contents

9. Estimates for Harmonic Measure


Notes
Exercises and Further Results
IV.
1.
2.
3.
4.
5.
6.

V.
1.
2.
3.
4.
5.
6.

VI.
1.
2.
3.
4.
5.
6.

VII.
1.
2.
3.
4.
5.

104
112
112

Extremal Distance
Definitions and Examples
Uniqueness of Extremal Metrics
Four Rules for Extremal Length
Extremal Metrics for Extremal Distance
Extremal
 d x Distance and Harmonic Measure
Estimate
The (x)
Notes
Exercises and Further Results

129
129
133
134
139
143
146
149
150

Applications and Reverse Inequalities


Asymptotic Values of Entire Functions
Lower Bounds
Reduced Extremal Distance
Teichmllers Modulsatz
Boundary Conformality and Angular Derivatives
Conditions More Geometric
Notes
Exercises and Further Results

157
157
159
162
166
173
184
193
194

Simply Connected Domains, Part One


The F. and M. Riesz Theorem
Privalovs Theorem and Plessners Theorem
Accessible Points
Cone Points and McMillans Theorem
Compression and Expansion
Pommerenkes Extension
Notes
Exercises and Further Results

200
200
203
205
207
212
216
221
221

Bloch Functions and Quasicircles


Bloch Functions
Bloch Functions and Univalent Functions
Quasicircles
Chord-Arc Curves and the A Condition
BMO Domains
Notes
Exercises and Further Results

229
229
232
241
246
253
257
258

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Contents

VIII.
1.
2.
3.
4.
5.
6.
7.

xi

Simply Connected Domains, Part Two


The Law of the Iterated Logarithm for Bloch Functions
Harmonic Measure and Hausdorff Measure
The Number of Bad Discs
Brennans Conjecture and Integral Means Spectra
Numbers and Polygonal Trees
The Dandelion Construction and (c)  (a)
Baernsteins Example on the HaymanWu Theorem
Notes
Exercises and Further Results

269
269
272
281
286
289
296
302
305
307

IX.
1.
2.
3.

Infinitely Connected Domains


Cantor Sets
For Certain , dim < 1.
For All , dim 1.
Notes
Exercises and Further Results

315
315
324
331
341
342

X.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.

Rectifiability and Quadratic Expressions


The Lusin Area Function
Square Sums and Rectifiability
A Decomposition Theorem
Schwarzian Derivatives
Geometric Estimates of Schwarzian Derivatives
Schwarzian Derivatives and Rectifiable Quasicircles
1
The BishopJones H 2 Theorem
Schwarzian Derivatives and BMO Domains
Angular Derivatives
A Local F. and M. Riesz Theorem
Ahlfors Regular Sets and the HaymanWu Theorem
Notes
Exercises and Further Results

347
348
361
372
380
384
393
397
408
411
415
420
425
426

Appendices
Hardy Spaces
Mixed Boundary Value Problems
The Dirichlet Principle
Hausdorff Measure
Transfinite Diameter and Evans Functions
Martingales, Brownian Motion, and Kakutanis Theorem
Carlemans Method

435
435
441
447
456
466
470
480

A.
B.
C.
D.
E.
F.
G.

CY535/Garnett

0 521 47018 8

xii

January 27, 2005

15:41

Contents

H.
I.
J.
K.
L.
M.

Extremal Distance in Finitely Connected Domains


McMillans Twist Point Theorem
Bloch Martingales and the Law of the Iterated Logarithm
A Dichotomy Theorem
Two Estimates on Integral Means
Calderns Theorem and Chord-Arc Domains

484
497
503
512
518
520

Bibliography
Author Index
Symbol Index
Subject Index

531
555
559
561

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Preface

Several surprising new results about harmonic measure on plane domains have
been proved during the last two decades. The most famous of these results are
Makarovs theorems that harmonic measure on any simply connected domain
is singular to Hausdorff measure  for all > 1 but absolutely continuous
to  for all < 1. Also surprising was the extension by Jones and Wolff of
Makarovs > 1 theorem to all plane domains. Further important new results
include the work of Carleson, Jones and Wolff, and others on harmonic measure
for complements of Cantor sets; the work by Carleson and Makarov, Bertilsson,
Pommerenke,
others on Brennans tantalizing conjecture that for univalent
 and

2
functions | | p d xd y < if 43 < p < 4; several new geometric conditions that guarantee the existence of angular derivatives; and the Jones square
sum characterization of subsets of rectifiable curves and its applications by
Bishop and Jones to a variety of problems in function theory.
We wrote this book to explain these exciting new results and to provide
beginning students with an introduction to this part of mathematics. We have
tried to make the subject accessible to students who have completed graduate
courses in real analysis from Folland [1984] or Wheeden and Zygmund [1977],
for example, and in complex analysis from Ahlfors [1979] or Gamelin [2001],
for example.
The first four chapters, along with the appendices on Hardy spaces, Hausdorff
measures and martingales, provide a foundation that every student of function
theory will need. In Chapter I we solve the Dirichlet problem on the half-plane
and the disc and then on any simply connected Jordan domain by using the
Carathodory theorem on boundary continuity. Chapter I also includes brief
introductions to hyperbolic geometry and univalent function theory. In Chapter II we solve the Dirichlet problem on domains bounded by finitely many
Jordan curves and study the connection between the smoothness of a domains
boundary and the smoothness of its Poisson kernel. Here the main tools are
xiii

CY535/Garnett

xiv

0 521 47018 8

January 27, 2005

15:41

Preface

two classical theorems about conjugate functions. Chapter II and the discussion in Chapter III of Wieners solution of the Dirichlet problem on arbitrary
domains follow the 1985 UCLA lecture course by Carleson. The introduction
to extremal length in Chapter IV is based on the Institut MittagLeffler lectures of Beurling [1989]. Chapter V contains some applications of extremal
length, such as Teichmllers Modulsatz and some newer theorems about angular derivatives, that are not found in other books. Chapter VI is a blend of
the classical theorems of F. and M. Riesz, Privalov, and Plessner and the more
recent theorems of McMillan, Makarov, and Pommerenke on the comparison
of harmonic measure and one dimensional Hausdorff measure for simply connected domains. Chapter VII surveys the beautiful circle of ideas around Bloch
functions, univalent functions, quasicircles, and A p weights. Chapter VIII is
an exposition of Makarovs deeper results on the relations between harmonic
measure in simply connected domains and Hausdorff measures and the work of
Carleson and Makarov concerning Brennans conjecture. Chapter IX discusses
harmonic measure on infinitely connected plane domains. Chapter X begins by
introducing the Lusin area function, the Schwarzian derivative, and the Jones
square sums, and then applies these ideas to several problems about univalent
functions and harmonic measures. The thirteen appendices at the end of the text
provide further related material.
For space reasons we have not treated some important related topics. These
include the connections between Chapters VIII and IX and thermodynamical
formalism and several other connections between complex dynamics and harmonic measure. We have emphasized Wieners solution of the Dirichlet problem instead of the Perron method. The beautiful Perron method can be found
in Ahlfors [1973] and Tsuji [1959]. We also taken a few detours around the
theory of prime ends. There are excellent discussions of prime ends in Ahlfors
[1973], Pommerenke [1975], and Tsuji [1959]. Finally, the theory of harmonic
measure in higher dimensions has a different character, and we have omitted it
entirely.
At the end of each chapter there is a brief section of biographical notes
and a section called Exercises and Further Results". An exercise consisting
of a stated result without a reference is meant to be homework for the reader.
Further results" are outlines, with detailed references, of theorems not in the
text.
Results are numbered lexicographically within each chapter, so that Theorem 2.4 is the fourth item in Section 2 of the same chapter, while Theorem III.2.4
is from Section 2 of Chapter III. The same convention is used for formulas, so
that (3.2) is in the same chapter, while (IV.6.4) refers to (6.4) from Chapter IV.
Many of the results that inspired us to write this book are also covered in

CY535/Garnett

0 521 47018 8

January 27, 2005

Preface

15:41

xv

Pommerenkes excellent book [1991]. However, our emphasis differs from the
one in Pommerenke [1991] and we hope the two books will complement each
other.
Some unpublished lecture notes from a 1986 Nachdiplom Lecture course
at Eidgenssische Technische Hochschule Zurich by the first listed author and
the out-of-print monograph Garnett [1986] were preliminary versions of the
present book.
The web page
http://www.math.washington.edu/ marshall/HMcorrections.html
will list corrections to the book. Though we have tried to avoid errors, the
observant reader will no doubt find some. We would appreciate receiving email
at marshall@math.washington.edu about any errors you come across.
Many colleagues, friends, and students have helped with their comments and
suggestions. Among these, we particularly thank A. Baernstein, M. Benedicks,
D. Bertilsson, C. J. Bishop, K. Burdzy, L. Carleson, S. Choi, R. Chow, M.
Essn, R. Gundy, P. Haissinski, J. Handy, P. Jones, P. Koosis, N. Makarov, P.
Mateos, M. ONeill, K. yma, R. Prez-Marco, P. Poggi-Corridini, S. Rohde,
I. Uriarte-Tuero, J. Verdera, S. Yang and S. Yoshinobu.
We gratefully acknowledge support during the writing of this book by the
Royalty Research Fund of the University of Washington, the University of
WashingtonUniversity of Bergen Faculty Exchange Program, the Institut des
Hautes Etudes Scientifiques, the Centre de Recerca Matemtica, Barcelona, and
the National Science Foundation.

Los Angeles and Seattle


Seattle and Bergen

John B. Garnett
Donald E. Marshall

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

xvi

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

I
Jordan Domains

To begin we construct harmonic measure and solve the Dirichlet problem in


the upper half-plane and the unit disc. We next prove the Fatou theorem on
nontangential limits. Then we construct harmonic measure on domains bounded
by Jordan curves, via the Riemann mapping theorem and the Carathodory
theorem on boundary correspondence. We review two topics from classical
complex analysis, the hyperbolic metric and the elementary distortion theory for
univalent functions. We conclude the chapter with the theorem of Hayman and
Wu on lengths of level sets. Its proof is an elementary application of harmonic
measure and the hyperbolic metric.

1. The Half-Plane and the Disc


Write H = {z : Imz > 0} for the upper half-plane and R for the real line. Suppose a < b are real. Then the function




zb
zb
= Im log
= (z) = arg
za
za
is harmonic on H, and = on (a, b) and = 0 on R \ [a, b].
z

a
b
Figure I.1 The harmonic function (z).
1

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

Viewed geometrically, (z) = Re(z) where (z) is any conformal mapping


from H to the strip {0 < Rez < } which maps (a, b) onto {Rez = } and
R \ [a, b] into {Rez = 0}. Let E R be a finite union of open intervals and

write E = nj=1 (a j , b j ) with b j1 < a j < b j . Set


z bj
j = j (z) = arg
z aj
and define the harmonic measure of E at z H to be
(z, E, H) =

n

j
j=1

(1.1)

Then
(i) 0 < (z, E, H) < 1 for z H,
(ii) (z, E, H) 1 as z E, and
(iii) (z, E, H) 0 as z R \ E.
The function (z, E, H) is the unique harmonic function on H that satisfies (i),
(ii), and (iii). The uniqueness of (z, E, H) is a consequence of the following
lemma, known as Lindelfs maximum principle.
Lemma 1.1 (Lindelf). Suppose the function u(z) is harmonic and bounded
above on a region  such that  = C. Let F be a finite subset of  and
suppose
lim sup u(z) 0

(1.2)

for all  \ F. Then u(z) 0 on .


/ . Then the map 1/(z z 0 ) transforms  into a bounded
Proof. Fix z 0
region, and thus we may assume  is bounded. If (1.2) holds for all ,
then the lemma is the ordinary maximum principle. Write F = {1 , . . . , N },
let > 0, and set


N

diam()
.
log
u (z) = u(z)
|z j |
j=1

Then u is harmonic on  and lim supz u (z) 0 for all . Therefore


u 0 for all , and


N

diam()
= 0.
u(z) lim
log

0
|z j |
j=1

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Half-Plane and the Disc

Lindelf [1915] proved Lemma 1.1 under the weaker hypothesis that  is
infinite. See also Ahlfors [1973]. Exercise 3 and Exercise II.3 tell more about
Lindelfs maximum principle.
Given a domain  and a function f C(), the Dirichlet problem for
f on  is to find a function u C() such that u = 0 on  and u| = f .
Theorem 1.2 treats the Dirichlet problem on the upper half-plane H .
Theorem 1.2. Suppose f C(R {}). Then there exists a unique function
u = u f C(H {}) such that u is harmonic on H and u| H = f .
Proof. We can assume f is real valued and f () = 0. For > 0, take disjoint
open intervals I j = (t j , t j+1 ) and real constants c j , j = 1, . . . , n, so that the
step function
f (t) =

n


cj Ij

j=1

satisfies



 f f 
< .
L (R )

(1.3)

Set
u (z) =
If t R \

n


c j (z, I j , H).

j=1

I j , then
lim u (z) = f (t)

H zt

by (ii) and (iii). Therefore by (1.3) and Lemma 1.1,




supu 1 (z) u 2 (z) < 1 + 2 .
H

Consequently the limit


u(z) lim u (z)
0

exists, and the limit u(z) is harmonic on H and satisfies


sup |u(z) u (z)| 2.
H

We claim that
lim sup |u (z) f (t)|
zt

(1.4)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains


for all t R. It is clear that (1.4) holds when t
/
I j . To verify (1.4) at the
endpoint t j+1 I j I j+1 , notice that by (ii), (iii), and Lemma 1.1,






c j + c j+1
supc j (z, I j , H) + c j+1 (z, I j+1 , H)
(z, I j I j+1 , H)
2
H


 c j c j+1 
,


2
while


lim

zt j+1

c j + c j+1
2


(z, I j I j+1 , H) =

c j + c j+1
.
2

Hence all limit values of u (z) at t j+1 lie in the closed interval with endpoints
c j and c j+1 , and then (1.3) yields (1.4) for the endpoint t j+1 .
Now let t R. By (1.4)
lim sup |u(z) f (t)| sup |u(z) u (z)| + lim sup |u (z) f (t)| 3.
zt

zH

zt

The same estimate holds if t = . Therefore u extends to be continuous on H


and u| H = f . The uniqueness of u follows immediately from the maximum

principle.
For a < b, elementary calculus gives


x a

x b

1
(x + i y, (a, b), H) =
tan1
tan1

y
y
b
y
dt
.
=
2
2
a (t x) + y
If E R is measurable, we define the harmonic measure of E at z H to be

y
dt
.
(1.5)
(z, E, H) =
2 + y2
(t

x)
E
When E is a finite union of open intervals this definition (1.5) is the same as
definition (1.1). For z = x + i y H, the density
Pz (t) =

y
1
(x t)2 + y 2

is called the Poisson kernel for H . If f C(R {}), the proof of Theorem 1.2 shows that

u f (z) =
f (t)Pz (t)dt,
R

and for this reason u f is also called the Poisson integral of f .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Half-Plane and the Disc

Note that the harmonic measure (z, E, ) is a harmonic function in its first
variable z and a probability measure in its second variable E. If z 1 , z 2 H then
0 < C 1

(z 1 , E, H)
C < ,
(z 2 , E, H)

where C depends on z 1 and z 2 but not on E. This inequality, known as Harnacks inequality, is easily proved by comparing the kernels in (1.5).
Now let D be the unit disc {z : |z| < 1} and let E be a finite union of open
arcs on D. Then we define the harmonic measure of E at z in D to be
(z, E, D) ((z), (E), H),

(1.6)

where is any conformal map of D onto H. This harmonic function satisfies


conditions analogous to (i), (ii), and (iii), so that by Lemma 1.1 the definition
(1.6) does not depend on the choice of . It follows by the change of variables
(z) = i(1 + z)/(1 z) that

1 |z|2 d
.
(z, E, D) =
i
2
E |e z| 2
An equivalent way to find this function is by a construction similar to (1.1).
This construction is outlined in Exercise 1.
Theorem 1.3. Let f (ei ) be an integrable function on D and set
2
1 |z|2 d
.
f (ei ) i
u(z) = u f (z) =
|e z|2 2
0

(1.7)

Then u(z) is harmonic on D. If f is continuous at ei0 D, then


lim

D zei0

u(z) = f (ei0 ).

(1.8)

Clearly (1.8) also holds if the integrable function f is changed on a measure


zero subset of D \ {ei0 }. The function u = u f is called the Poisson integral
of f and the kernel
Pz ( ) =

1 1 |z|2
2 |ei z|2

is the Poisson kernel for the disc. If f C(D) then



u f (z), z D
U (z) =
f (z),
z D
is the solution of the Dirichlet problem for f on D.
In the special case when f (ei ) is continuous, Theorem 1.3 follows from
Theorem 1.2 and a change of variables. Conversely, Theorem 1.3 shows that

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

Theorem 1.2 can be extended to f L 1 (dt/(1 + t 2 )), again by changing variables.


Proof of Theorem 1.3. We may suppose f is real valued. From the identity
ei + z 
= 2 Pz ( ),
Re i
e z
we see that u is the real part of the analytic function
2
ei + z d
,
f (ei ) i
e z 2
0
and therefore that u is a harmonic function. One can also see u is harmonic by
differentiating the integral (1.7).
Suppose f is continuous at ei0 and let > 0. Then
| f (ei ) f (ei0 )| <
on an interval I = (1 , 2 ) containing 0 . Setting

1 |z|2
d
+ f (ei0 )(z, I, D),
u (z) =
f (ei )
i
2
2
[0,2]\I |e z|
we have





1 |z|2
i
i0 d 

( f (e ) f (e ))  (z, I, D) .
|u(z) u (z)| = 
i
2
2
I |e z|

However, lim zei0 u (z) = f (ei0 ) by the definition of u . Therefore


lim sup |u(z) f (ei0 )| < ,
zei0

and (1.8) holds when f is continuous at ei0 .

2. Fatous Theorem and Maximal Functions


When f L 1 (D) the limit (1.8) can fail to exist at every D; see Exercise 7. However, there is a substitute result known as Fatous theorem, in
which the approach z is restricted to cones. For D and > 1, we
define the cone


 ( ) = z : |z | < (1 |z|) .
The cone  ( ) is asymptotic to a sector with vertex and angle 2 sec1 () that
is symmetric about the radius [0, ]. The cones  ( ) expand as increases.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Fatous Theorem and Maximal Functions

z
0

 ( )

2 sec1

Figure I.2 The cone  ( ).


A function u(z) on D has nontangential limit A at D if
lim

 ( ) z

u(z) = A

(2.1)
z+1

for every > 1. A good example is the function u(z) = e z1 . This function u(z)
is continuous on D \ {1}, and |u( )| = 1 on D \ {1}, but u(z) has nontangential limit 0 at = 1. With fixed > 1, the nontangential maximal function
of u at is
u ( ) = sup |u(z)|.
 ( )

If u has a finite nontangential limit at , then u ( ) < for every > 1.


We write |E| for the Lebesgue measure of E D.
Theorem 2.1 (Fatous theorem). Let f (ei ) L 1 (D) and let u(z) be the
Poisson integral of f . Then at almost every = ei D,
lim

 ( ) z

u(z) = f ( )

for all > 1. Moreover, for each > 1



 3 + 6


|| f ||1 .
{ D : u ( ) > }

(2.2)

(2.3)

When u(z) is the Poisson integral of f L 1 (D) the function u = u f is also


called the solution to the Dirichlet problem for f , even though u converges
to f on D only nontangentially and only almost everywhere.
Inequality (2.3) says the operator L 1 (D) f u is weak-type 1-1. It
follows from (2.2) that u ( ) < almost everywhere, but (2.3) is a sharper,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

quantitative result. In the proof of the theorem we derive (2.2) from the estimate
(2.3).
The proof of Fatous theorem is a standard approximate identity argument from real analysis that derives almost everywhere convergence for all
f L 1 (D) from
(a) an estimate such as (2.3) for the maximal function, and
(b) the almost everywhere convergence (2.2) for all functions in a dense subset
of L 1 (D), such as C(D).
See Stein [1970]. We will use this approximate identity argument again later.
Proof. As promised, we first assume (2.3) and show (2.3) implies (2.2). Fix
temporarily. We may assume f is real valued. Set
W f ( ) = lim sup |u f (z) f ( )|.
 z

Then W f ( )

u ( ) + | f ( )|.

Chebyshevs inequality gives



{ : | f ( )| > } || f ||1 ,

so that by (2.3),


 
 
{ : W f ( ) > } { : u ( ) > /2} + { : | f ( )| > /2}

8 + 12

|| f ||1 .

(2.4)

Fix > 0 and let g C(D) be such that || f g||1 2 . Now Wg ( ) = 0 by


Theorem 1.3, and hence
W f ( ) = W f g ( ).
Applying (2.4) to f g then gives
2


{ : W f ( ) > } (8 + 12) = (8 + 12).

Therefore, for any fixed , (2.2) holds almost everywhere. Because the cones
 increase with , it follows that (2.2) holds for every > 1, except for in
a set of measure zero.
To prove (2.3) we will dominate the nontangential maximal function with a
second, simpler maximal function. Let f L 1 (D) and write

1
| f |d
M f ( ) = sup
I |I | I

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Fatous Theorem and Maximal Functions

for the maximal average of | f | over subarcs I D that contain . The function
M f is called the HardyLittlewood maximal function of f . The function M f
is simpler than u because it features characteristic functions of intervals instead
of Poisson kernels.
Lemma 2.2. Let u(z) be the Poisson integral of f L 1 (D) and let > 1.
Then
u ( ) (1 + 2)M f ( ).

(2.5)

Proof. Assume = 1. Fix z so that 0 = arg z has |0 | . Define




Pz ( ) = sup Pz () : | | ||

1 1 + |z|

,
| | |0 |
= 2 1 |z|

max(Pz ( ), Pz ( )), |0 | < | | .

Pz ( )

Pz ( )

Figure I.3 The function

Pz .

The function Pz satisfies


(i) Pz ( ) is an even function of [, ],
(ii) Pz ( ) is decreasing on [0, ], and
(iii) Pz ( ) Pz ( ).
The even function Pz is the smallest decreasing majorant of Pz on [0, ]. We
may assume f (ei ) 0, so that


f (ei )Pz ( )d
f (ei )Pz ( )d.

CY535/Garnett

0 521 47018 8

January 27, 2005

10

15:41

I. Jordan Domains

Then properties (i) and (ii) imply



f (ei )Pz ( )d ||Pz ||1 M f (1)

(2.6)

because Pz is the increasing limit of a sequence of functions of the form



  1
( j , j ) ( )
cj
2 j

with c j 0 and c j ||Pz ||1 .

Pz ( )

0 0

Figure I.4 Approximating

Pz

by a step function.

Now we claim that when z  (1),


||Pz ||1 (1 + 2).

(2.7)

Note that (iii), (2.6), and (2.7) imply (2.5). To prove (2.7) we first assume
/2 0 = arg z /2. Then by the law of sines,
|0 |
|0 |
| sin 0 |
|sin |

,
1 |z|
|1 z|
2 |1 z|
2
1
2
where = arg(z 1)/z is explained by Figure I.2. If /2 |0 | and
z  (1), then |1 z| 1 and
|0 |
|0 |

.
1 |z|
|1 z|
Hence (see Figure I.3)
Pz 1 = 2

|0 |

Pz ( )d +

2|0 | 1 + |z|
(1 + 2).
2 1 |z|

That proves (2.7) and therefore Lemma 2.2.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Fatous Theorem and Maximal Functions

11

By Lemma 2.2, the inequality (2.3) will follow from the simpler inequality
 3 f 



1
,
(2.8)
{ D : M f ( ) > }

which says that the operator L 1 f M f is also weak-type 1-1.


To prove the inequality (2.8), we use a covering lemma.
Lemma 2.3. Let be a positive Borel measure on D and let {I j } be a finite sequence of open intervals in D. Then {I j } contains a pairwise disjoint
subfamily {Jk } such that

1  
(Jk )
Ij .
(2.9)
3
Proof. Because the family {I j } is finite, we may assume that no I j is contained
in the union of the others. Writing I j = {ei : (a j , b j )}, we may also assume
that
0 a1 < a2 < . . . . < an < 2.
Then b j+1 > b j , because otherwise I j+1 I j , and b j1 < a j+1 , because otherwise I j I j1 I j+1 . If n > 1, then bn < b1 + 2 and bn1 < a1 + 2 .
Consequently, the family of even-numbered intervals I j is pairwise disjoint.
The family of odd-numbered intervals I j is almost pairwise disjoint; only the
first and last intervals can intersect. If

1  
Ij ,
(I j )
3
j even

we take the even-numbered intervals to be the subfamily {Jk }. Otherwise



2  
Ij .
(I j )
3
j odd

In that case, if
(I1 )

1
(I j ),
2
j odd

we take for {Jk } the family of odd-numbered intervals, omitting the first interval
I1 , while if
1
(I1 ) >
(I j ),
2
j odd

we take {Jk } = {I1 }. Then in each case (2.9) holds for the subfamily {Jk }.

CY535/Garnett

12

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

Lemma 2.4. The operator f M( f ) is weak-type 1-1: If f L 1 , then




{ D : M f ( ) > } 3 f 1 .
(2.8)

Proof. Let K be a compact subset of E = { D : M f ( ) > }. For each


E there is an open interval I such that I and

1
| f |d > ,
|I | I
so that
|I | <


| f |d.
I

Cover K by finitely many such intervals {I j : 1 j n}, and let {Jk } be the
pairwise disjoint subfamily given by Lemma 2.3. Then
 




3
|K |  I j  3 |Jk | 3
Jk | f |d || f ||1 ,
and letting |K | tend to |E | establishes Lemma 2.4.

By (2.5) and (2.8), inequality (2.3) holds with constant 3 + 6, and the proof

of Fatous theorem is complete.

Corollary 2.5. If u is a bounded harmonic function on D, then for every > 0


and for almost every = ei D,
f ( ) =

lim

 ( ) z

u(z)

exists, u(z) is the Poisson integral of f , and || f || = supzD |u(z)|.


Proof. Let rn 1 and let f n (ei ) = u(rn ei ). By the BanachAlaoglu theorem, the sequence { f n } has a weak-star cluster point f L (D) satisfying
|| f || lim sup || f n || supD |u(z)|. Since u(rn z) is the Poisson integral of
f n , and Poisson kernels are in L 1 , u must be the Poisson integral of f . But

then |u(z)| || f || . The corollary now follows from Fatous theorem.
In particular, the corollary implies that for any measurable E D, there
exists a unique bounded harmonic function u(z) on D such that u(z) has nontangential limit E almost everywhere. It is the function u(z) = (z, E, D).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Carathodorys Theorem

13

3. Carathodorys Theorem
Let  be a simply connected domain in the extended plane C . We say  is a
Jordan domain if  =  is a Jordan curve in C .
Theorem 3.1 (Carathodory). Let be a conformal mapping from the unit
disc D onto a Jordan domain . Then has a continuous extension to D, and
the extension is a one-to-one map from D onto .
Because maps D onto , the continuous extension (also denoted by ) must
map D onto  = , and because is one-to-one on D, (ei ) parameterizes
the Jordan curve .
Before we prove Carathodorys theorem, we use it to solve the Dirichlet
problem on a Jordan domain . Let f be Borel function on  such that f
is integrable on D . If w = 1 (z), then
2
1 |w|2 d
(3.1)
u(z) u f (z) =
f (ei ) i
|e w|2 2
0
is harmonic on , and by Theorem 3.1 and Theorem 1.3,
lim u(z) = f ( )

 z

(3.2)

whenever 1 ( ) D is a point of continuity of f . In particular, if f is


continuous on  then (3.2) holds for all  and u(z) = u f (z) solves the
Dirichlet problem for f on .
If f is a bounded Borel function on , then f is Borel and the integral
(3.1) is defined. For any Borel set E  we use (3.1) with f = E to define
the harmonic measure of E relative to  by:

1 |w|2 d
1
.
(3.3)
(z, E, ) = (w, (E), D) =
i
2
1 (E) |e w| 2
Then E (z, E) is a Borel measure on , and (3.1) can be rewritten as

f ( )d(z, ).
(3.4)
u(z) =


Equations (3.3) and (3.4) do not depend on the choice of , because for every
conformal self map T of D,
(T (w), T ( 1 (E)), D) = (w, 1 (E), D).
When f is a bounded Borel function on , (3.4) and Fatous theorem give
sup |u(z)| =  f  L () .
z

CY535/Garnett

0 521 47018 8

January 27, 2005

14

15:41

I. Jordan Domains

Moreover, Corollary 2.5 shows that every bounded harmonic function on  can
be expressed in the form (3.4).
The principal goal of this book is to find geometric properties of the harmonic
measure (z, E) more explicit than the definition (3.3). But (3.3) already points
out the key issue: for a Jordan domain questions about harmonic measure are
equivalent to questions about the boundary behavior of conformal mappings.
Proof of Theorem 3.1. We may assume  is bounded. Fix D. First we
show has a continuous extension at . Let 0 < < 1, write
B(, ) = {z : |z | < }
and set = D B(, ). Then ( ) is a Jordan arc having length

L() =
|  (z)|ds.

By the CauchySchwarz inequality

L 2 ()
so that for < 1,

0

L 2 ()
d

|  (z)|2 ds,


DB(,)

|  (z)|2 d x d y

(3.5)

= Area (D B(, )) < .

n
Un

(n )

Figure I.5 The crosscuts n and (n ).


Therefore there is a sequence n 0 such that L(n ) 0. When L(n ) < ,
the curve (n ) has endpoints n , n , and both of these endpoints must
lie on  = . Indeed, if n , then some point near n has two distinct

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Carathodorys Theorem

15

preimages in D because maps D onto , and that is impossible because is


one-to-one. Furthermore,
|n n | L(n ) 0.

(3.6)

Let n be that closed subarc of  having endpoints n and n and having smaller
diameter. Then (3.6) implies diam(n ) 0, because  is homeomorphic to
the circle. By the Jordan curve theorem the curve n (n ) divides the plane
into two regions, and one of these regions, say Un , is bounded. Then Un ,
because C \  is arcwise connected. Since

diam(Un ) = diam n (n ) 0,
we conclude that
diam(Un ) 0.

(3.7)

Set Dn = D {z : |z | < n }. We claim that for large n, (Dn ) = Un . If


not, then by connectedness (D \ D n ) = Un and

diam(Un ) diam (B(0, 1/2)) > 0,



which contradicts (3.7). Therefore diam((Dn )) 0 and (Dn ) consists
of a single point, because (Dn+1 ) (Dn ). That means has a continuous
extension to { } D. It is an exercise to show that the union over of these
extensions defines a continuous map on D .
Let also denote the extension : D . Since (D) = , maps D
onto . To show is one-to-one, suppose (1 ) = (2 ) but 1 = 2 . The
argument used to show n  also shows that (D) = , and so we can
assume j D, j = 1, 2. The Jordan curve




(r 1 ) : 0 r 1 (r 2 ) : 0 r 1
bounds a domain W , and then 1 (W ) is one of the two components of


D \ {r 1 : 0 r 1} {r 2 : 0 r 1} .
But since (D) ,
(D 1 (W )) W  = {(1 )}
and is constant on an arc of D. It follows that is constant, either by
Schwarz reflection principle or by the Jensen formula, and this contradiction

shows (1 ) = (2 ).

CY535/Garnett

0 521 47018 8

January 27, 2005

16

15:41

I. Jordan Domains

One can also prove is one-to-one by repeating for 1 the proof that is
continuous. See Exercise 13(b). Exercise 12 gives the necessary and sufficient
condition that have a continuous extension to D.
The CauchySchwarz trick used to prove (3.5) is known as a lengtharea
argument. The lengtharea method is the cornerstone of the theory of extremal
length. See Chapter IV.

4. Distortion and the Hyperbolic Metric


Let D be the open unit disc. The hyperbolic distance from z 1 D to z 2 D
is
z2
|dz|
(z 1 , z 2 ) = D (z 1 , z 2 ) = inf
,
(4.1)
2
z 1 1 |z|
where the infimum is taken over all arcs in D connecting z 1 to z 2 . Let M denote
the set of conformal self maps of D :
T (z) =

za
, a D, || = 1.
1 az

When T M, we have
|T  (z)|
1
=
,
2
1 |T (z)|
1 |z|2
and thus the hyperbolic distance is conformally invariant,
(T (z 1 ), T (z 2 )) = (z 1 , z 2 ), T M.

(4.2)

This conformal invariance is the main reason we are interested in the hyperbolic
distance.
The hyperbolic metric is the infinitesimal form |dz|/(1 |z|2 ) of the hyperbolic distance. Taking
T (z) =

z z1
1 z1 z

gives

T (z 2 )

(z 1 , z 2 ) = (0, T (z 2 )) =
0

|dz|
.
1 |z|2

Therefore the hyperbolically shortest arc from 0 to T (z 2 ) is the radius


[0, T (z 2 )], and its hyperbolic length is
1 + |T (z )| 
1
2
.
(0, T (z 2 )) = log
2
1 |T (z 2 )|

CY535/Garnett

In general

0 521 47018 8

January 27, 2005

15:41

4. Distortion and the Hyperbolic Metric

17

 z z 
1 
 2
1+

1
1 z1 z2

(z 1 , z 2 ) = log
 z z  ,
1 
 2
2
1

1 z1 z2

(4.3)

and the hyperbolically shortest curve, or the geodesic, from z 1 to z 2 is a segment


of a diameter of D or an arc of a circle in D orthogonal to D.
By (4.3) we have
 z z  e2(z 1 ,z 2 ) 1
1 
 2
= tanh (z 1 , z 2 ).
 = 2(z ,z )

1 2 +1
1 z1 z2
e
Write
e2d 1
.
e2d + 1
Then the hyperbolic ball B = {z : (z, a) < d} is the euclidean disc
 za 


{z : 
 < t},
1 az
and a calculation shows that B has euclidean radius
t = t (d) = tanh(d) =

r (a, d) =
and euclidean distance to D
dist(B, D) =

t (1 |a|2 )
1 t 2 |a|2

1t 
(1 |a|).
1 + |a|t

(4.4)

(4.5)

Therefore, if d is fixed, the euclidean distance dist(B, D) and the euclidean


diameter of B are both comparable to dist(a, D). However, for a = 0 the
euclidean center of B is not a. Figure I.6 shows two hyperbolic balls with the
same hyperbolic radius and two geodesics with the same hyperbolic length.

z1
0
a1

z2
x1
a2

Figure I.6 Hyperbolic balls and geodesics.

CY535/Garnett

0 521 47018 8

18

January 27, 2005

15:41

I. Jordan Domains

Now assume (z) is a univalent function in D, that is, assume is analytic


and one-to-one on D. After dilating, translating, and rotating the domain (D),
is normalized by (0) = 0 and  (0) = 1, so that
(z) = z + a2 z 2 + .
Important examples are the Koebe functions
z
(z) = (z) =
, || = 1.
(1 z)2

(4.6)

(4.7)

Note that
(z) =

nn1 z n

n=1

maps D to the complement of the radial slit [/4, ].


Theorem 4.1 (Koebe one-quarter theorem). Assume (z) is a univalent function on D. If (z) has the form (4.6) then
|a2 | 2

(4.8)

and
1
.
(4.9)
4
Equality holds in (4.8) or (4.9) if and only if is a Koebe function (4.7).
dist(0, (D))

Proof. First note that (4.8) implies (4.9). Indeed, suppose w


/ (D). Then
(4.8) holds for the univalent function
g(z) =

1
w(z)
= z + (a2 + )z 2 + ,
w (z)
w

so that


a2 + 1  2,
w
and together (4.10) and (4.8) give
1
.
4
To prove (4.8) we form the odd function

(z 2 )
a2
f (z) = z
= z + z3 + .
2
2
z
|w|

(4.10)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Distortion and the Hyperbolic Metric

19

Then f is univalent because is univalent, and the C = C {} valued


function

F(z) =

1
1 a2
1 
bn z n
= z + = +
f (z)
z
2
z

(4.11)

n=1

is also univalent in D. To complete the proof we use the following lemma, called
the area theorem, which we will prove after we establish (4.8).
Lemma 4.2 (area theorem). If the univalent function F(z) satisfies (4.11),
then

n|bn |2 1.

(4.12)

n=1

To establish (4.8) we apply (4.12) to F = 1/ f . Since b1 = a2 /2 and |b1 | 1,



we have |a2 | 2.
The reader can verify that equality in either of (4.8) or (4.9) implies is a
Koebe function.
Proof of Lemma 4.2. The lemma is called the area theorem because of its
proof. For r < 1, the Jordan curve r = {F(r ei ) : 0 2 } encloses an
area A(r ), and by Greens theorem,


i 2
F
i
(r ei )d.
wdw =
F(r ei )
A(r ) =
2 r
2 0

Therefore by (4.11) and Fourier series,






1
n|bn |2 r 2n
A(r ) = 2
r
n=1

and
1


n=1

n|bn |2 = lim

r 1

A(r )
0,

which yields (4.12).

Theorem 4.3 (Koebes estimate). Let (z) be a conformal mapping from the
unit disc D onto a simply connected domain . Then for all z D
1 
| (z)|(1 |z|2 ) dist((z), ) |  (z)|(1 |z|2 ).
4

(4.13)

CY535/Garnett

20

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

Proof. Fix z 0 D. Then the univalent function




z+z 0
(z 0 )
1+z
0z

(z) = 
(z 0 ) 1 |z 0 |2
satisfies (0) = 0 and  (0) = 1. Hence if w
/ (D), then


 w (z )  1
0


 
 (z 0 )(1 |z 0 |2 )  4
by (4.9), and this gives the left-hand inequality in (4.13).
To prove the right-hand inequality, fix z D, take


f (w) = 1 (z) + dist (z),  w ,
and apply the Schwarz lemma at w = 0 to the function
g(w) =

f (w) z
.
1 z f (w)

We will often use the invariant form of (4.13):


Corollary 4.4. Let be a conformal mapping from a simply connected domain
1 onto a simply connected domain 2 , and let (z 0 ) = w0 . Then
dist(w0 , 2 )
|  (z 0 )|

4|  (z 0 )|
4
dist(z 0 , 1 )

(4.14)

Proof. Applying (4.13) to (z) = (z 0 + dist(z 0 , 1 )z) gives the left-hand


inequality and the same argument with 1 gives the right-hand inequality. 
In a simply connected domain  = C, the hyperbolic distance is defined
by moving back to D via a conformal map : D . We write
 (w1 , w2 ) = D (z 1 , z 2 )
when w j = (z j ). By (4.2),  (w1 , w2 ) does not depend on the choice of the
conformal map . The quasihyperbolic distance from w1  to w2  is
w2
|dw|
,
Q  (w1 , w2 ) = inf
w1 dist(w, )
in which the infimum is taken over all arcs in  joining w1 to w2 . Since (4.13)
can be written as
4|dz|
|dw|
|dz|

,
dist(w, )
1 |z|2
1 |z|2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Distortion and the Hyperbolic Metric

21

where w = (z), we have


 (w1 , w2 ) Q  (w1 , w2 ) 4 (w1 , w2 ).

(4.15)

Consequently the geometric statement following (4.4) and (4.5) about hyperbolic distances near D remains approximately true in every simply connected
domain with nontrivial boundary.
Let  be any proper open subset of C. Then there exist closed squares {Sj },
having pairwise disjoint interiors and sides parallel to the axes, such that
(i) Sj has sidelength (Sj ) = 2n j

(ii)  = Sj , and
(iii) diam(Sj ) dist(Sj , ) < 4 diam(Sj ).
The squares {Sj } are called Whitney squares. Here is one way to construct
Whitney squares in the case diam() < : Let 2N +1 diam() < 2N +2
and partition the plane into squares having sides parallel to the axes and sidelength 2N . We call these squares 2N squares. Include in the family {Sj }
any 2N square S  which satisfies (iii), and divide each of the remaining 2N squares into four squares of side 2N 1 . Next include in {Sj } any of
these new 2N 1 squares contained in  and satisfying (iii), and continue. (See
Figure I.7). Finding Whitney squares in the case diam() = is an exercise
for the reader. Whitney squares can be viewed as replacements for hyperbolic
balls since there are universal constants c1 < c2 such that each Tj contains a
hyperbolic ball of radius c1 and is contained in a hyperbolic ball of radius c2 .

{Tk }

{Sj }
Figure I.7 Whitney squares in D and .

Whitney squares are almost conformally invariant in the following sense:


Assume  is simply connected, let : D  be a conformal mapping, let
{Sj } be the Whitney squares for  and let {Tk } be the Whitney squares for D.
Then by (4.15) there is a constant M, not depending on the map , such that for

CY535/Garnett

0 521 47018 8

22

January 27, 2005

15:41

I. Jordan Domains

each k, (Tk ) is contained in at most M squares Sj , and 1 (Sk ) is contained


in at most M squares Tj . In particular, for each d > 0, there is M(d) such
that every hyperbolic ball {z :  (z, a) < d} in  is covered by M(d) Whitney
squares.
Theorem 4.5. Let (z) be a univalent function satisfying (0) = 0 and
 (0) = 1. Then
|z|
|z|
|(z)|
,
(1 + |z|)2
(1 |z|)2

(4.16)

1 + |z|
1 |z|
|  (z)|
.
3
(1 + |z|)
(1 |z|)3

(4.17)

|  (z)|
1 + |z|
1 |z|

.
|z|(1 + |z|)
|(z)|
|z|(1 |z|)

(4.18)

and

Moreover,

Inequality (4.16) is known as the growth theorem, while (4.17) is called the
distortion theorem.
Proof. The critical inequalities are (4.17) and we prove them first. Fix z 0 D
and take


z+z 0
(z 0 )
1+z
z
0

.

f (z) =
(4.19)
 (z 0 ) 1 |z 0 |2
Then f is univalent on D, f (0) = 0 and f  (0) = 1. By (4.8),



 
1 |z 0 |2


2z 0  4,
| f (0)| =  (z 0 ) 
(z 0 )
so that when z 0 = r ei ,

 i 
 e (z 0 )
4
2|z 0 | 


.
  (z )
2
1

|z
|
1

|z 0 |2
0
0

(4.20)

Applying the general formula


Re

zg  (z)
Reg
=
|z|
r

to g = log  , we then obtain

2r + 4
2r 4
log |  (z 0 )|

.
r
1 r2
1 r2

(4.21)

Integrating (4.21) along the radius [0, z] gives both inequalities in (4.17).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. The HaymanWu Theorem

23

To prove the upper bound in (4.16), integrate the upper bound in (4.17)
along [0, z]. To prove the lower bound, we can assume that |(z)| < 41 , because
|z|
< 41 . Then by (4.9) there exists an arc D with ( ) = [0, (z)],
(1+|z|)2
and integrating |  (z)||dz| along then gives the lower bound in (4.16).
To prove (4.18), apply (4.16) at z 0 to the function f defined by (4.19). 
Again, equality any place in (4.16), (4.17), (4.18), or (4.21) implies that
is a Koebe function.

5. The HaymanWu Theorem


We give a very elementary proof, based on an idea of the late K. yma [1992], of
the theorem of Hayman and Wu. The HaymanWu theorem will be a recurrent
topic throughout this book.
Theorem 5.1 (HaymanWu). Let be a conformal mapping from D to a
simply connected domain  and let L be any line. Then


(5.1)
length 1 (L ) 4.
Hayman and Wu [1981] gave the first proof of (5.1) with 4 replaced by some
large unknown constant. yma [1992] obtained the constant 4 , Rohde [2002]
proved that the best constant in (5.1) is strictly smaller than 4 , and yma
[1993] proved that the best constant is at least 2 . The sharp constant in (5.1) is
not known. See Exercises 24 and VI.3. We present ymas elementary proof,
as modified by Rohde.
For the proof it will be convenient to replace the hyperbolic metric (z 1 , z 2 )
by the pseudohyperbolic metric, defined in D by


 z1 z2 
 = tanh(z 1 , z 2 ),

D (z 1 , z 2 ) = 
1 z1 z2 
and in  by
 (w1 , w2 ) = D ( 1 (w1 ), 1 (w2 )).
Proof. We can assume that is analytic and one-to-one in a neighborhood of
D and that L = R . Let L k denote the components of  L and let k be that
component of
 {z : z }
such that L k k . Then k is a Jordan domain symmetric about R. When
k = j, k  j R and k  j contains at most one point, because 

CY535/Garnett

0 521 47018 8

January 27, 2005

24

15:41

I. Jordan Domains

is simply connected.

k

j

Figure I.8 The domains k are shaded.


By symmetry there is a conformal mapping k : k iH such that
k (L k ) = R+ and k extends continuously to k . For 1 (k ) D,
set
= ( ), x = |k ()|, = k1 (x), and z = 1 ().
Then the composition
 1 k1 (|k |)


is a smooth map of 1 ( k  \ P) D onto 1 ( L k ) \ P  where
P and P  are finite sets.

k

k ()
Figure I.9 The map z = ( )

To prove Theorem 5.1, it suffices to show that


|| 2.

(5.2)

To prove (5.2), suppose that I = [,  ] is an open interval contained in


1 (k ) D. Set
 = (  ), x  = |k (  )|,  = k1 (x  ), and z  = 1 (  ).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. The HaymanWu Theorem

25

Then by Picks theorem (see Exercise 14(a))



 x x 
,

D (( ), ( )) =  (, ) k (, ) = i H (x, x ) = 
x + x 


and by Lindelfs maximum principle


(x, k ((I )), iH) = (z, I, 1 (k )) (( ), I, D).
Letting  we obtain the inequality
||
1 1 |( )|2

,
2 | ( )|2
1 |( )|2

(5.3)

which easily implies (5.2). This completes the proof of the HaymanWu theo
rem.
Because it does not depend on the mapping , the conclusion of Theorem 5.1
is actually stronger than (5.1). See Exercise 20.
In Chapter VII we will see that for some 1 < p < 2, independent of

p


 1 
(5.4)
( ) (w) |dw| < C p ,
L

but the largest permissible p is unknown. A slit disc shows (5.4) fails at p = 2,
and a counterexample for some p < 2, due to Baernstein, is given in Chapter VIII. In Chapter X we will determine the class of curves L for which (5.1)
holds.

Notes
This chapter is a survey of elementary material which can be found in many
other sources. We learned Lemma 2.3 from Garsia [1970], who attributed it
to W. H. Young. The proof Lemma 2.3 also shows that every open cover has

I  2. In Wheeden and Zygmund [1977] the
a subcover {Ik } such that
k
higher dimensional formulation of this statement is called the simple Vitali
lemma. The classical Vitali covering lemma is an easy consequence of this or
Lemma 2.3. See Exercise 9 or Mattilas book [1995], which has much about
covering lemmas. Steins classic book [1970] also gives an excellent discussion of the relation between covering lemmas, maximal functions, and almost
everywhere convergence and a clear introduction to Whitney squares.
Section 4 follows Duren [1983], Hayman [1958], and Goluzin [1952]. We
call (4.13) Koebes estimate and we will refer to it as such many times, but in

CY535/Garnett

26

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

[1907] Koebe only proved the left half of (4.13) with a constant 41 . Later,
in [1916] Bieberbach showed = 41 . See Duren [1983].

Exercises and Further Results


1. (a) If E is the arc (ei1 , ei2 ) of D in Figure I.10, then

 i2
e z
(2 1 ) = 2 (2 1 ), (E.1)
2 (z, E, D) = 2 arg i
e 1 z
where is the angle subtended at z by the arc (ei1 , ei2 ).
Hint: Verify (i), (ii), and (iii) from Section 1 via elementary geometry and
use Lemma 1.1. An alternative approach from the Poisson integral formula
is to write


2ei
d 2
ei + z
i

z)

1
=
log(e
ei z
ei z
d i
and integrate the real part of this expression over the interval (1 , 2 ).
(b) We also have
(z, E) =

1
(2 1 ),
2

where ei j is the other endpoint of the chord extending from ei j through z.


This is easy from (E.1) and elementary geometry.
ei1

ei2
E
ei1

ei2
Figure I.10 Harmonic measure of an arc.
(c) Using (z) = i(1 z)/(1 + z), and a change of variables, derive the
Poisson integral formula for D from the Poisson integral formula for H.
Hints:
ei + z

y
1
1 |z|2
and
=
Re
= Im
.
ei z
tz
|ei z|2
(x t)2 + y 2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

27

2. If ei D is fixed, the level curves of z Pz (ei ) are circles in D tangent


to D at ei . If E is an arc of D, the level set {(z, E) = t} is the arc of a
circle cutting D at the endpoints of E in the angle t.
3. (a) The Lindelf maximum principle, Lemma 1.1, is false if u(z) is not
assumed to be bounded above.
(b) But Lindelfs maximum principle is also true for subharmonic functions:
If u(z) is a subharmonic function on a domain  such that  = C, if u(z)
is bounded above, and if
lim sup u(z) 0,
z

for all  \ F, where F is a finite subset of , then u(z) 0 on .


See Ahlfors [1973].
(c) Show that the conclusion in 3(b) still holds if the hypothesis  = C is
replaced by the weaker hypothesis  \ F = .
4. Let u be a real valued continuous function on a region . Prove that the
following three conditions are equivalent:
(a) For each sufficiently small disc D  and each v harmonic on D, u v
and v u satisfy the maximum principle on D.
(b) u satisfies the mean value property for each sufficiently small disc D
contained in .
(c) u is harmonic on .
Hint: Use the Poisson integral on D.
5. (a) Harnacks inequality. Let p and q be points of an arbitrary plane domain
. Prove there is a constant C p,q such that whenever u(z) is positive and
harmonic on ,
1
C p,q

u( p)
C p,q .
u(q)

(b) If p and q remain in a compact subset K , then there is a constant


C K < such that C p,q C K .
(c) Harnacks principle. Let {n } be a sequence of domains and let u n be a
harmonic function on n . Assume  is a domain such that every z  has a
neighborhood Vz such that for n sufficiently large, Vz n , and u n u n+1
on Vz . Then either u n tends to uniformly on every compact subset of ,
or there is a harmonic function u on  and u n converges to u uniformly on
every compact subset of .
(d) Let n and  be as in (c), and let u n be a harmonic function on n . Now
assume there is Mz < such that for n sufficiently large, supVz |u n | Mz .
Prove {u n } has a subsequence converging uniformly on all compact subsets

CY535/Garnett

0 521 47018 8

January 27, 2005

28

15:41

I. Jordan Domains
of  to a function harmonic on .
(e) If u is harmonic on D and |u| M then
|u|

2M
.
1 |z|2


eit + z
Hint: Pz (t) = Re it
.
e z
6. The boundary Harnack inequality. Let u and v be positive harmonic functions on D with u(0) = v(0), let I D be an open arc and assume


lim u(z) = lim v(z) = 0

for all I. Prove that for every compact K D I there is a constant


C(K ) independent of u and v such that on K D
u(z)
1

C(K ).
C(K )
v(z)

7.

8.

9.

10.

11.

Hint: Using a conformal map of a subregion of D onto D, you may assume u


and v are continuous on D. Estimate the Poisson kernel on D \ I . See Wu
[1978] for a more general result.
Show there is a Borel set A D such that any function almost everywhere
equal to A is not continuous at any point D. Then show that for
f = A , (1.8) fails for all D.
Suppose G(z) is continuous on the closed disc and analytic on the open disc.
Use the Poisson kernel to show that if dG(ei )/d exists at 0 , then G  (z)
has nontangential limit at ei0 .
A Vitali cover of a measurable set E D is a family I of arcs I such
that for every E and every > 0 there exists I I such that I
and |I | < . Prove the Vitali covering lemma: If I is a Vitali cover of E
and if > 0 there exists a pairwise disjoint sequence {I j } I such that

|E \ I j | < .
Assume u(z) is the Poisson integral of a finite positive measure on D,
and let d = f ( )d + ds be the Lebesgue decomposition of . Prove
that almost everywhere d , u has nontangential limit f ( ), and that almost
everywhere ds , u has nontangential limit . Hint: Use Lemma 2.3.
Let u(z) be real and harmonic in D.
(a) If 1 < < , then
u u C, M(u ).
Moreover, if 0 < p < , then

||u || L p ( D) .
||u || L p ( D) C,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

29

(b) If 1 p < and if u is the Poisson integral of f L p , then



lim |u(r ei ) f (ei )|d = 0.
r 1

(c) If p > 1 show u is the Poisson integral of f L p (D) if and only if



d
sup |u(r ei )| p
< .
2
r <1
(d) Show u is the Poisson integral of a probability measure on D

u(z) =
Pz ( )d( ),
if and only if u > 0 and u(0) = 1.
(e) Show u is the Poisson integral of a finite (signed) Borel measure on D
if and only if

d
sup |u(r ei )|
< .
(E.2)
2
r <1
If (E.2) holds, then u has an almost everywhere nontangential limit f ( ),
but u(z) need not be the Poisson integral of f. See Exercise 10.
(f) If u is the Poisson integral of a finite signed measure on D, then the
d
converge weak-star to as r 1. It follows that u
measures u(r ei ) 2
determines the measure uniquely.
(g) Show u is the Poisson integral of f L 1 (D) if and only if the family
{u(r ei ) : 0 < r < 1} is uniformly integrable in .
(h) Let 1 p . If u L p (D), then u is the Poisson integral of some
f L p (D).
See Appendix A for the corresponding results when p < 1.
12. A compact set K is locally connected if whenever U is a relatively open
subset of K and z U K there is a relatively open subset V of K such
that V is connected and z V U. Let  be a simply connected domain
such that  contains at least two points. Prove  is locally connected if
and only if the Riemann map : D  extends continuously to D. Hint:
For one direction, use the uniform continuity of . For the other direction
repeat the proof of Carathodorys theorem with n replaced by L n .
13. (a) Let  be a simply connected domain and let  be a crosscut, that
is, a Jordan arc in  having distinct endpoints in . Prove that  \ has
two components 1 and 2 , each simply connected, and j =  j \ is
connected.
(b) Let  be simply connected, let :  D be conformal and fix z 
and . Assume and z can be separated by a sequence of crosscuts

CY535/Garnett

0 521 47018 8

January 27, 2005

30

15:41

I. Jordan Domains

n  such that length(n ) 0. Let Un be the component of  \ n such


that z
/ Un . Prove that (Un ) D.
14. Let f (z) be an analytic function on D and assume | f (z)| 1.
(a) Prove Picks theorem:
 


 f (z) f (w)   z w 



 1 f (w) f (z)   1 wz .
Equivalently, prove
( f (z), f (w)) (z, w).
In other words, analytic maps from D to D are Lipschitz with respect to the
hyperbolic metric.
Hint: Repeat the proof of the Schwarz lemma, using (S f )/T for suitable
S and T in M.
(b) Deduce that
| f  (z)|
1

.
2
1 | f (z)|
1 |z|2
(c) If equality holds in (a) at any two points z and w, or if equality holds in
(b) at any point, then f M.
15. (a) Prove (4.4) and (4.5).
(b) Find the euclidean center of the hyperbolic ball {z : (z, a) < d}.
16. Let (z) be univalent on D and assume (z) = 0 for every z.
(a) Prove
|  (0)| 4|(0)|,
using (4.9).
(b) Prove


1 |z|
1 + |z|

2

|(z)|

|(0)|

Hint: Fix z and apply (a) to


(w) =

1 + |z|
1 |z|

w+z 
1 + zw

to get
  (z) 
4


,


(z)
1 |z|2
and integrate along [0, z].

2
.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

31

17. (a) Show that if equality holds anywhere in (4.8), (4.9), (4.16), (4.17), (4.18),
or (4.21) then is a Koebe function, defined by (4.7).
(b) On the other hand, equality in (4.12) holds if and only if C \ F(D) has
area zero.
18. (a) Find an upper bound for the constant M mentioned in the remarks preceding Figure I.7.
(b) Give a construction of Whitney squares in the case diam() = .
19. A region of the form
Q = {z = r ei : 0 < < 0 + (Q), 1 (Q) r < 1}
is called a box or a Carleson box, and the arc
{ei : 0 < < 0 + (Q)}
is the base of Q. See Figure I.11. A finite measure on D is a Carleson
measure if there is a constant C such that
(Q) C(Q)
for every Carleson box.

(Q)
Q
aQ

ei0
(Q)

Figure I.11 A Carleson box.

(a) Let > 1. Prove is a Carleson measure if and only if there is C such
that


({z D : u(z) > }) C { D : u > }
whenever u is the Poisson integral of f L 1 (D). Hint: Write the set

{u > } = I j where the I j are pairwise disjoint open arcs on D, and

define Tj = D \ I
/ j  ( ). The set Tj looks like a tent with base I j .

CY535/Garnett

0 521 47018 8

January 27, 2005

32

15:41

I. Jordan Domains

Ij
Tj
Ik
Tk
Figure I.12 Tents.

Then (Tj ) C (I j ) and |u| on D \ Tj . Conversely, let Q be the
box with base I, and let u(z) = 4(z, I, D). Then u(z) > on Q while
by (2.3),


{u > } C |I |.

(b) If is a finite measure, then is a Carleson measure if and only if



sup

aD

1 |a|2
d(z) < .
|1 az|2

Hint: Let a be the center of Q. See page 239 of Garnett [1981].


(c) Let  be a countable union of rectifiable curves in D. Then
sup length(T ()) <

T M

if and only if arc length on  is a Carleson measure.


(d) A region in the upper half-plane H of the form
Q = {(x, y) : a x a + (Q), 0 < y (Q)}
is also called a box or Carleson box and a finite measure on H is called a
Carleson measure if there is a constant C such that
(Q) C(Q)
for every Carleson box in H. A box Q is called dyadic if (Q) = 2k and
a = j2k for some integers j and k. If (Q) C(Q) for all dyadic boxes,
prove that is a Carleson measure on H.
(e) If is a Carleson measure on D and is a conformal map of H onto D,
then prove is a Carleson measure on H. Conversely, if is a Carleson
measure on H supported in the unit box [0, 1] [0, 1], then 1 is a
Carleson measure on D.
Part (a) is due to E. M. Stein (unpublished).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

33

20. Let be a conformal mapping from D to a simply connected domain , let


L be any line and set  = 1 ( L). Then
(a) Arc length on  is a Carleson measure; in other words,
length( Q) C(Q)
for any box Q = {r ei : 0 < < 0 + (Q), 1 (Q) r < 1}.
(b) Moreover, for any disc B(z, r ),
length( B(z, r )) C r.

(E.3)

Hint: If r < 1 |z| or r 0.4 use a linear map and the Hayman-Wu
theorem. If 1 |z| < r < 0.4 then there is a Carleson box Q D B(z, r )
such that (Q) Cr, and part (a) gives (E.3).
21. Let (z) be univalent on D and assume (z) = 0 for every z.
(a) There is C1 < , independent of , such that
  
 (z) 
d xd y C1 .

(z)
D

Hint: Write w = u + iv = (z), w . Then


  
2 1  2
  (z) 
2
3


 (z) 2
d xd y
d xd y
 (1 |z|) 3 d xd y


(z)
1 |z|
(z)
D


C2


2 dudv

.
1 | 1 (w)| 3
|w|2

Now integrate over  {|w| > 1} and  {|w| 1} separately, and use
Exercise 16.
(b) Let f = log . There is C2 < , independent of such that

| f  |d xd y C2 (Q)
(E.4)
Q

for any Carleson box


Q = {z = r ei : 0 0 + (Q), 1 (Q) r < 1}.
In other words, | f  (z)|d xd y is a Carleson measure. For the proof we may
(Q)
assume (Q) is small. Let a = a Q = (1 2(Q))ei(0 + 2 ) and use (a) with
w = (z a)/(1 az). See Figure I.11.

CY535/Garnett

0 521 47018 8

January 27, 2005

34

15:41

I. Jordan Domains

(c) It follows from (b) that almost everywhere and in L 1 norm,


lim f (r ei ) = f (ei )

r 1

exists and f (z) is the Poisson integral of f (ei ). In fact, f (ei ) BMO
(for bounded mean oscillation), which by definition means there exists a
constant C3 < , independent of , such that for every arc I D,

1
| f |d C3 .
(E.5)
inf
C |I | I
Indeed, if Q has base I, then one can show


1
| f f (a Q )|d C
| f  |d xd y.
|I | I
Q

See Jones [1980]. Baernstein [1976] gave the first proof of (E.5). Cima and
Schober [1976] and Pommerenke [1976] have related results. When e f (z) is
not univalent, (E.4) is not equivalent to (E.5). However, in general (E.5) is
equivalent to

| f  |2 (1 |z|)d xd y C(Q).
Q

See Fefferman and Stein [1972], Baernstein [1980], Garnett [1981], or Appendix F.
22. Again let be univalent and assume (z) = 0 for all z. Prove that for r 1/2
  i 
2 
C4
 (r e ) 2
log
,
(E.6)
d

I (r ) = 

(r ei )
1r
1r
where C4 does not depend on . Hayman [1980] attributes the following
argument to P. L. Duren: By Exercise 16,
  i 
1 
 (r e ) 2
+ 8 log 2.
d
xd
y

8
log


(r ei )
1r
{|z|<r }

Take r < R < 1. Since I (r ) increases with r ,


  i 
2
 (r e ) 2
I (r ) 2

 d xd y
2
(r ei )
R r
{r <|z|<R}



1

2
+
8
log
2
.
8
log
2
1 R
R r2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

35

Now take
1
(1 r 2 ).
2
Hayman [1980] shows that the logarithm in (E.6) is necessary.
23. Let 0 < p < . An analytic function f (z) on the disc D is in the Hardy
space H p if

p
sup
| f (r ei )| p d = || f || Hp < .
1 R2 =

0<r <1 0

Let (z) be a univalent function on D, normalized by


(z) = z + a2 z 2 + . . . .
(a) For 0 < p < , verify the identity

|| p = p 2 || p2 |  |2 .
(b) If 0 < p <

1
2

then H p and
p

|||| H p

C
,
1 2p

where C is independent of p. This result is due to Prawitz [1927]. Greens


theorem and (a) give
2


d 1
p2
i p
|(r e )| d =
|(z)| p2 |  |2 d xd y
dr 2 0
2r
{|z|<r }

p2
2r

Write M(r ) = sup|z|r |(z)|. Then


2


p2
d 1
|| p d
dr 2 0
2r

|w| p2 d A(w).
({|z|<r })

|w| p2 d A(w) =

|w|<M(r )

p
p
M(r ) ,
r

while by the distortion theorem,


M(r )
Therefore
1
2

|| d p
p

0 r
0

r
.
(1 r )2

M(t)
t

p
dt

t p1 (1 t)2 p dt = O

.
1 2p

CY535/Garnett

36

0 521 47018 8

January 27, 2005

15:41

I. Jordan Domains

(c) If p > 21 , a similar argument gives


2
1
1
|(r ei )| p d C p
.
2 0
(1 r )2 p1
(d) If p > 21 , (c) and (4.18) imply
2
1
1
|  (r ei )| p d C p 
.
2 0
(1 r )3 p1
For p = 1, (c) implies (d) if is analytic but not necessarily univalent.
1
(e) The Koebe function (z) = z/(1 z)2 is not in H 2 .
In a deep paper [1974] Baernstein proved that the extremal functions for
parts (b) and (c) are the rotates of the Koebe function. See Duren [1983] for
these results and more, and see Appendix A for the basic theory of Hardy
spaces.
24. (a) Show that the best constant in the HaymanWu theorem is smaller than
4 (Rohde [2002]). Hint: By (5.3) if |( )| > 2 then ( ) lies in
a cone at with opening < C and || > 1 C. Moreover the curve
= (D) makes an angle with the radius through ( ) approximately
tan1 (1/2). Then use (4.20) to show that if makes a larger angle with the
radius at a point z then || < 2 on an interval centered at z/|z| of
length (1 |z|). Use this to show that there is a > 0, independent of so
that || < 2 on a subset of the circle of length at least .
(b) Use (5.3) to show that if we also require (0) L in the HaymanWu
theorem, then the best constant is less than 4 2.
25. (a) Let 1 2 be Jordan domains, and let E = 1 2 . Then on 1 ,
(z, E, 1 ) (z, E, 2 ).
(b) Let 0 < < 1 and let  be a Jordan domain such that
{|z| < 1 }  {|z| < 1 + }.
Let E =  {1 arg z 2 }, where 2 < 1 + 2. Show


(0, E, ) (2 1 )  C log( 1 ).
2

Hint: By (a) the worst case in (b) is for regions of the following form. Let
I = [ei1 , ei2 ]. Let  = D \ {r z : z I, r 1 } where < |I |, and let
E =  {|z| = 1 }. Show that if |z| = 1 , and dist(z, E) , then
(z, E, ) 2(z, I, D).
(c) Show the estimate in (b) is sharp, except for the value of C.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

II
Finitely Connected Domains

In this chapter we solve the Dirichlet problem on a domain bounded by a


finite number of Jordan curves. For a simply connected Jordan domain the
problem was solved in Chapter I via the theorem of Carathodory. For a multiply
connected domain the problem will be reduced to the simply connected case
using the Schwarz alternating method.
Solving the Dirichlet problem on a domain  is equivalent to constructing
harmonic measure on . In Section 2 we describe harmonic measure in terms
of the normal derivative of Greens function in the case when  consists of
analytic curves. In Section 4 we study the relation between the smoothness of
 and the smoothness of the Poisson kernel (the RadonNikodym derivative
of harmonic measure against arc length). This relation hinges on two classical
estimates for conjugate functions which we prove in Section 3.

1. The Schwarz Alternating Method


Let  be a plane domain such that  is a finite union of pairwise disjoint
Jordan curves
 = 1 2 . . .  p .
We say  is a finitely connected Jordan domain. A bounded function f on 
is piecewise continuous if there is a finite set E  such that f is continuous
on  \ E and f has left and right limits at each point of E. In this section
we solve the Dirichlet problem for piecewise continuous boundary data on a
finitely connected Jordan domain.
Theorem 1.1. Let  be a finitely connected Jordan domain, and let f be a
bounded piecewise continuous function on . Then there is a unique function
37

CY535/Garnett

0 521 47018 8

38

January 27, 2005

15:41

II. Finitely Connected Domains

u(z) = u f (z), bounded and harmonic on , such that


lim u(z) = f ( )

(1.1)

at every point of continuity of f . Moreover,


sup |u| sup | f |.


Proof. We may assume  is bounded. The uniqueness of u f is immediate from


Lindelfs maximum principle, Lemma I.1.1. The existence of u f in case p = 1
was treated in Section I.3, and so we assume p > 1. Take a Jordan arc with
endpoints a, b
/  such that 1 =  \ is simply connected and such that
 is a finite set. See Figure II.1. Then (z) = za
zb has a single valued
analytic branch defined on 1 , and we can solve the Dirichlet problem on 1
by transplanting it to the Jordan region (1 ). Take a second Jordan arc  such
that 1 =  \  is simply connected,   is a finite set, and  = .
We can also solve the Dirichlet problem on 1 .

a

b

Figure II.1 The proof of Theorem 1.1.


Let E  be a finite set and set F = E ( ) (  ). Suppose
without loss of generality that f C( \ E) is positive and bounded. To start,
let u 1 be the solution to the Dirichlet problem on 1 with boundary data

f ( ),

u 1 ( ) =
max f, .
Then u 1 is harmonic on 1 and continuous on  \ F and u 1 matches its boundary data on 1 \ F. Next let u 1 be the solution to the Dirichlet problem on 1
with boundary data u 1 ( ), 1 . Then u 1 is harmonic on 1 and continuous
on  \ F and u 1 matches its boundary data on 1 \ F. In particular, u 1 = u 1
on  \ F. By the Lindelf maximum principle we have u 1 max f = u 1
on , and therefore u 1 u 1 on  \ F. Now let u 2 be the solution to the
Dirichlet problem on 1 with boundary data u 1 ( ), 1 . On  we

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Schwarz Alternating Method

39

have u 2 = u 1 u 1 , while on  \ F we have u 2 = f = u 1 . Therefore u 2 u 1


on 1 , again by Lemma I.1.1. Consequently u 2 u 1 = u 1 on  , and by
the maximum principle, u 2 u 1 on  \ F.
Continuing in this way we obtain a decreasing sequence
u 1 u 1 u 2 u 2 u 3 . . .
of positive functions, alternately harmonic on 1 and 1 . By Harnacks principle, Exercise I.5, the limit
u(z) = lim u n (z) = lim u n (z)
n

is a bounded harmonic function on  such that u u 1 max f .


To complete the proof we show
lim u(z) = f ( )

(1.1)

whenever  is a point of continuity of f . We may assume


/ E. Take
a neighborhood V of such that W = V  is a Jordan domain and such that
W (E ) = . Let be a conformal map of D onto W. By Carathodorys
theorem (D) = W, and for w = (z) W

1 |z|2
d
f (ei )
u n (w) =
i z|2
1
2
|e
()

+

D\ 1 ()

1 |z|2
d
u n (ei ) .
2
|ei z|2

Because 1 () is a neighborhood of 1 ( ) in D, the first integral approaches f ( ) as w . Because |u n | sup | f |, the second integral tends
to zero, uniformly in n, as w . Therefore (1.1) holds and the theorem is

proved.
If  is a finitely connected Jordan domain, Theorem 1.1 shows that the map
f u f (z) is a bounded linear functional on C(). Define the harmonic
measure of a relatively open subset U  by
(z, U ) (z, U, ) = sup{u f (z) : f C(), 0 f U },
and for an arbitrary subset E of , set
(z, E) (z, E, ) = inf{(z, U ) : U open in , U E}.

CY535/Garnett

0 521 47018 8

40

January 27, 2005

15:41

II. Finitely Connected Domains

This procedure, which mimics the usual proof of the Riesz representation theorem, shows (z, E) is a Borel measure on  such that

u f (z) =
f ( )d(z, )
(1.2)


for continuous f . When  is simply connected, this definition of harmonic


measure agrees with the earlier definition (I.3.3).
For every z 1 , z 2  there exists, by virtue of Harnacks inequality, a constant c = c(z 1 , z 2 ) such that
1
(1.3)
(z 1 , E) (z 2 , E) c(z 1 , E),
c
and the constants c(z 1 , z 2 ) remain uniformly bounded if z 1 and z 2 remain in
a compact subset of . If f L 1 (, d), and in particular if f is bounded
and Borel, there is a sequence { f n } in C() such that for some fixed z 0 ,

| f n ( ) f ( )| d(z 0 , ) 0.
Write


u n (z) =

and

f n ( )d(z, )

u(z) = u f (z) =

f ( )d(z, ).

(1.4)

Then by (1.3)
u n (z) u(z)
for all z . By (1.3) we also see that the harmonic functions u n (z) are uniformly bounded on compact subsets of . Thus by Harnacks principle the
limit function u(z) is harmonic on . We call u the solution to the Dirichlet
problem for f on . If f is bounded, then we also have
sup |u f (z)| || f || L (,d) .
z

Moreover, if f is bounded and continuous at , then (1.1) also holds at


. See Theorem 2.7 or Exercise 1(b) for a proof.
Note however that the Schwarz alternating method cannot be applied directly
to a bounded Borel function because the hypothesis of Lindelfs maximum
principle holds only for piecewise continuous functions.
The next section will give a much more explicit description of the measure
(z, E) when  has some additional smoothness.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Greens Functions and Poisson Kernels

41

2. Greens Functions and Poisson Kernels


Again let  be a finitely connected Jordan domain, and assume  is bounded.
For fixed w , let h(z, w) be the solution to the Dirichlet problem for the
boundary value f ( ) = log | w| C(), and define Greens function
with pole w to be
g(z, w) = log

1
+ h(z, w).
|z w|

(2.1)

Then g(z, w) is continuous in z  \ {w}, and


g(z, w) > 0

on ,

(2.2)

g(, w) = 0

on ,

(2.3)

z g(z, w) is harmonic on  \ {w},


1
is harmonic at w.
z g(z, w) log
|w z|

(2.4)
(2.5)

These properties are easily derived from Theorem 1.1 and the definition (2.1).
By the maximum principle, properties (2.3), (2.4), and (2.5) determine g(z, w)
uniquely.
When  is unbounded we fix a
/ . For w = , we let h(z, w) solve the

Dirichlet problem on  for f ( ) = log w
a , and define


za
 + h(z, w).
g(z, w) = log 
z w
 1 
 to define h(z, ) and set
For w = , we instead use f ( ) = log a
g(z, ) = log |z a| + h(z, ). These definitions are independent of the
choice of a, and with them (2.2)(2.5) still hold and determine g(z, w).
Suppose is a conformal mapping from one finitely connected Jordan domain  onto another finitely connected Jordan domain  . Then (z) 
whenever z , because :   is a homeomorphism. It follows that
g ((z), (w)) = g (z, w)
because Greens functions are characterized by (2.3), (2.4), and (2.5).
If  is the unit disc D then clearly


 1 zw 
.

g(z, w) = log 
zw 

(2.6)

(2.7)

Consequently Greens function for any simply connected Jordan domain  can
be expressed in terms of the conformal mapping :  D.

CY535/Garnett

0 521 47018 8

January 27, 2005

42

15:41

II. Finitely Connected Domains

Theorem 2.1. Let  be a simply connected domain bounded by a Jordan curve,


let w  and let :  D be a conformal map with (w) = 0. Then
g(z, w) = log |(z)|.


Proof. Immediate from (2.6) and (2.7).

By definition, an analytic arc is the image ((1, 1)) of the open interval
(1, 1) under a map one-to-one and analytic on a neighborhood in C of (1, 1)
to C. An analytic Jordan curve is a Jordan curve that is a finite union of (open)
analytic arcs.
Lemma 2.2. Let  be a finitely connected Jordan domain. Then there exists a
finitely connected Jordan domain  such that  consists of finitely many
pairwise disjoint analytic Jordan curves and there exists a conformal map from
 onto  which extends to be a homeomorphism from  to  .

2
3 2 1

3


1
1

1 (2 )
1 (3 )

Figure II.2 The proof of Lemma 2.2.


Proof. Write  = 1 . . .  p , where each  j is a Jordan curve. Let 1 be
the component of C \ 1 containing , where C is the extended plane, and let
1 be a conformal map of 1 onto D. Let 2 be the component of C \ 1 (2 )
containing 1 (), and let 2 be a conformal map of 2 onto D. Repeating this
process for each boundary curve, we obtain a conformal map p from  to
a region  such that  consists of finitely many pairwise disjoint analytic

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Greens Functions and Poisson Kernels

43

Jordan curves. Applying Carathodorys theorem to each k , we see that p


extends to be a homeomorphism from  to  .

In Exercise 5 the proof of Lemma 2.2 is used to define Greens function in
non-Jordan, finitely connected domains.
Theorem 2.3. Let  be a finitely connected Jordan domain and let z 1 , z 2 .
Then
g(z 1 , z 2 ) = g(z 2 , z 1 ).

(2.8)

Proof. By Lemma 2.2 we may assume  consists of analytic Jordan curves.


When  consists of analytic curves, an argument with Schwarz reflection,
which we will use many times and prove in Lemma 2.4, shows there is a
neighborhood V of  to which z g(z, w) has a harmonic extension. Hence
g(z, w) is C on some neighborhood V of  and we can use Greens theorem
in the form



u
v
v
(u v v u)d xd y =
u
ds,
n
n
U
U

where n is the unit normal vector pointing out from the domain U.
Fix distinct z 1 , z 2 . We apply Greens theorem on the domain


 =  \ {|z z 1 | } {|z z 2 | } ,
when is small, with u(z) = g(z, z 1 ) and v(z) = g(z, z 2 ). Because u = v = 0
on ,


v
u
u
v
ds = 0,
n
n

and because u and v are harmonic on  , the area integral in Greens theorem
vanishes. We conclude that
2
g
d
g(z 1 + ei , z 1 ) (z 1 + ei , z 2 )

r
2
0
2
g
d
g(z 1 + ei , z 2 ) (z 1 + ei , z 1 )

r
2
0
(2.9)
2
g
d
i
i
(g(z 2 + e , z 2 ) (z 2 + e , z 1 )
=
r
2
0
2
g
d
g(z 2 + ei , z 1 ) (z 2 + ei , z 2 ) .

r
2
0

CY535/Garnett

0 521 47018 8

January 27, 2005

44

15:41

II. Finitely Connected Domains

For small, g(z j + ei , z j ) 2 log (1/) and for k = j, g(z, z j ) has bounded
derivatives near z k . That means the first and third integrals in (2.9) approach 0
as 0. By (2.1),
log
g
(z j + ei , z j ) =
+ O(1),
r

so that, as 0, the second integral tends to g(z 1 , z 2 ) and the fourth integral

tends to g(z 2 , z 1 ). Therefore (2.8) holds.

Lemma 2.4. Suppose  is a finitely connected Jordan domain and suppose


 is an analytic arc. Let u(z) be a harmonic function in . If
lim u(z) = 0

for all , then there is an open set W  such that u extends to be


harmonic on W . If also u(z) > 0 in , then
u
( ) < 0
n

(2.10)

for all .
u=0

D
V

u>0


v=0

v>0
Figure II.3 Straightening an analytic arc in .

Proof. Let . Because is an analytic arc, there is a neighborhood V of


and a conformal map : V D such that ( ) = 0,
(V ) = D = D {Imw < 0},
and
( V ) = (1, 1).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Greens Functions and Poisson Kernels

Set

u( 1 (w)),

v(w) = u( 1 (w)),

0,

45

w D = D {Imw < 0},


w D+ = D {Imw > 0},
w (1, 1).

Then v is continuous in D and v has the mean value property over sufficiently
small circles centered at any w D. Hence v is harmonic in D and U = v
defines a harmonic extension of u to V .
If U1 and U2 are extensions of u to neighborhoods V1 and V2 such that
V1 V2 is connected, then U1 = U2 in the component of V1 V2 that
contains V1 V2 . It follows that u has a harmonic extension to some open
set W .
If u > 0 in , then clearly
u/n
0 on . The strict inequality (2.10) will

hold at if and only if v/ y (0) < 0. On D, there is an analytic function


h =
v iv with Imh = v and h(0) = 0. The Taylor expansion of h at 0 is
h(w) = an w n + O(|w n+1 |),

+
with an = 0. But if n
2, then

h(D ) D = , which is a contradiction.




Hence a1 = h (0) = v/ y (0) = 0 and (2.10) holds.

When  consists of analytic curves, Greens function provides a formula


for harmonic measure that generalizes the Poisson integral formula for D.
Theorem 2.5. Assume  consists of finitely many pairwise disjoint analytic
Jordan curves, and let z . Then Greens function g(, z) extends to be
harmonic (and hence real analytic) on a neighborhood of  and
g(, z)
>0
n

(2.11)

on , where n is the unit outer normal at . If u C() is harmonic


on , then

g(, z)
ds( )
.
(2.12)
u( )
u(z) =
n
2


Because of (2.12),
Pz ( ) =
is called the Poisson kernel for .

1 g(, z)
2 n

CY535/Garnett

0 521 47018 8

46

January 27, 2005

15:41

II. Finitely Connected Domains

Proof. Fix z . By Lemma 2.4, g(, z) extends to be harmonic (and real


analytic) on some neighborhood of  and then (2.11) follows from (2.10).
To prove (2.12) we first assume that u is C on a neighborhood of . We
then apply Greens theorem on  =  \ {w : |w z| < } with small and
v(w) = g(w, z). Because w g(w, z) = u = 0 on  and because g = 0 on
, Greens theorem collapses into
2

g(, z)
ds( )
u(z + ei ) d
=

u( )
g(z + ei , z)
n
2
r
2

0
2
i
g(z + e , z) d
.

u(z + ei )
r
2
0
Since g(z + ei , z) 2 log 1/ for small , and since u is C , we have
2
u(z + ei ) d
lim
g(z + ei , z)
= 0.
0
r
2
0
As we have seen, (2.1) yields
log
g(z + ei , z)
=
+ O(1),
r

while u(z + ei ) = u(z) + O(). Hence


2
g
d
= u(z) + O(),
u(z + ei ) (z + ei , z)

r
2
0
and that gives (2.12) when u is C on a neighborhood of .
To prove (2.12) in general, set  = {w  : g(w, z) > } where is small.
By uniqueness,  has Greens function g (w, z) = g(w, z) . Therefore
g
g
(, z) =
(, z)
n
n
g is a
on  . In a neighborhood of a point 0 , the function = g + i
conformal map and


g(, z)
1
u( )ds =
u 1 ds,
2 N 
n
{Rez=}(N )
which converges to

{Rez=0}(N )

u 1 ds =

1
2

g(, z)
u( )ds.
n

Therefore


1
g(, z)
g(, z)
1
u( )ds
u( )ds
2 
n
2 
n

(2.13)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Greens Functions and Poisson Kernels

47

as 0. But (2.12) holds for u on  because u is C on a neighborhood of



 , and hence (2.13) yields (2.12) for .
Corollary 2.6. If  consists of finitely many pairwise disjoint analytic Jordan
curves and if z , then
d(z, ) =

g(z, ) ds( )
.
n
2

(2.14)

In other words, harmonic measure for z  is absolutely continuous to arc


length on , the density
1 g(z, )
d
=
= Pz ( )
ds
2 n
is real analytic on , and
d
< c2
ds

c1 <

(2.15)

for positive constants c1 and c2 .


Proof. The identity (2.14) follows from (1.2) and (2.12) and the fact that Borel
measures are determined by their actions on continuous functions. The inequal
ities (2.15) are immediate from (2.11).
One objective of this book is to compare harmonic measure for general
domains to more geometrical measures such as arc length, and Corollary 2.6 is
the first result of this kind.
Theorem 2.7. Assume  consists of finitely many pairwise disjoint analytic
Jordan curves and for  and > 1 define


 ( ) = z  : |z | < dist(z, ) .
If u(z) is a bounded harmonic function on , then for ds almost all  the
limit
lim

 ( ) z

exists,

u(z) = f ( )

(2.16)


u(z) =

Pz ( ) f ( )ds( ),

(2.17)

and
sup |u(z)| = || f || L .


(2.18)

CY535/Garnett

48

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

Conversely, if f is a bounded Borel function on , then (2.17) defines a


bounded harmonic function u(z) on  such that (2.18) holds and (2.16) holds
ds almost everywhere. Moreover, if f is continuous at 0 , then
lim u(z) = f (0 ).

z0

(2.19)

By (2.18), (2.16) and (2.17) establish an isometry between the space of


bounded harmonic functions on  and L (, ds) when  consists of analytic curves.
Proof. A simple localization argument gives the existence of the nontangential
limit f . If I is an open arc on , there exists a neighborhood V I such
that V  = I and V  is simply connected, and there exists a conformal
mapping defined on V such that (V ) = D and (I ) is an arc on D.
It follows that maps conical approach regions at V  into cones at
( ):


V  ( ) B ( ) () (( )),
where B ( ) = {z : |z | < = ( )}. Then if u(z) is a bounded harmonic
function on , we can apply Fatous theorem to u ( 1 ) to obtain (2.16) ds
almost everywhere on V .
The proof of (2.17) is exactly the same as the proof of (2.12) except
that the dominated convergence theorem is applied in
 (2.13). By (2.16) we
have || f || sup |u(z)|, and because Pz 0 and  Pz ds = 1, we have
sup |u(z)| || f || . Hence (2.16) and (2.17) imply (2.18).
To prove the converse, let f L (, ds). Then the discussion following (1.4) shows (2.17) defines a bounded harmonic function u(z) on  and
sup |u(z)| || f || . Therefore by (2.16), u has almost everywhere a nontangential limit, which we will temporarily call F, and u is the Poisson integral
of F. What we must prove is that F = f almost everywhere. Again let V be a
neighborhood of an open arc I = V  such that V  is simply connected.
For h L (, ds), define


vh (z) =
Pz ( )h( )ds( ) Pz (, V )h( )ds( ),


where Pz (, V ) is the Poisson kernel for z V . If h C() then by


(1.2), Theorem 1.1, (2.12), and (3.2) of Chapter I, lim z vh (z) = 0 for all
I , and hence by Lemma 2.4, vh extends to be harmonic in a neighborhood
W of I which does not depend on h. Thus by Exercise I.5(e) or (3.5) below, if
J is a compact subset of I and if > 0 then there is a neighborhood N of J ,
depending only on ||h|| and , so that |vh | < in N . Now take h n C()

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Greens Functions and Poisson Kernels

49

so that h n converges to f in L 1 , and ||h n || || f || . For each z V N ,


vh n (z) converges to v f (z) and so |v f (z)| < . Because > 0 was arbitrary, we
conclude that v f (z) 0 as z J . But by Theorem I.1.3,
F( ) f ( ) =

lim

( ) z

v f (z) = 0

almost everywhere in J . Consequently F = f almost everywhere and (2.16)


and (2.18) hold for all f L (, ds). 
Finally, if f is continuous at 0 I then I Pz (, V ) f ( )ds( ) is continuous

at 0 by (3.2) of Chapter I and so by the continuity of v f , (2.19) holds.
The converse can be proved another way. Using the real analyticity of
g(w, z), one can refine the proof of Lemma I.2.2 and show that
sup |u(z)| C(, )Ms f ( ),

 ( )

(2.20)

where u is the Poisson integral (2.17) and where the maximal function Ms f ( )
is the supremum of the averages of f over arcs  with :

1
Ms f ( ) = sup
| f |ds.
( )
A variation on the covering lemma shows that Ms is weak-type 1-1, and an
approximation, as in the proof of Theorem I.2.1, then yields (2.16) for the
Poisson integral of f . This is the argument that must be used in the Euclidean
spaces Rd , with d 3, and in other situations.
With some care, the conformal mapping proof of (2.16) in the text can also
be parlayed into a proof of the maximal estimate (2.20). See Exercise 9.
Let  be any finitely connected Jordan domain and let be the conformal
map, given in Lemma 2.2, of  onto a domain  , where  consists of
analytic Jordan curves. Since is a homeomorphism of  onto  , harmonic
measure can be transplanted from  to  via , just as it was in Section I.3
for simply connected Jordan domains. This gives an alternate but equivalent
definition of harmonic measure for .
In Section 4 we will consider two questions. Let  be a finitely connected
Jordan domain.
Question 1. If  has some degree of differentiability and if f C() also
has some degree of differentiability along , how smooth is the solution u f (z),
as z approaches ?
Question 2. What smoothness conditions on , weaker than real analyticity,
will ensure that g(z, )/n exists on  and that (2.14) and (2.15) still hold?

CY535/Garnett

0 521 47018 8

50

January 27, 2005

15:41

II. Finitely Connected Domains

These two questions are equivalent. Their answers will depend on Kelloggs
theorem about the boundary behavior of conformal mappings. The proof of
Kelloggs theorem in turn depends on the estimates for conjugate functions in
the next section.

3. Conjugate Functions
Let f L 1 (D) be real. For convenience we write f ( ) for f (ei ). If u(z) is
the Poisson integral of f on D, then u(z) is harmonic and real and there exists
a unique harmonic function 
u (z) such that 
u (0) = 0 and
F = u + i
u
is analytic on D. The function 
u is called the conjugate function or harmonic
conjugate of u. The nontangential limit

f ( ) =

lim

 (ei ) zei


u (z)

(3.1)

exists almost everywhere, and this has an easy proof from Fatous theorem: We
may assume f 0, so that
u (z))
G(z) = e(u(z)+i

is bounded and analytic on D. By Corollary I.2.5, G has a nontangential limit


G(ei ) almost everywhere. Since |G(ei )| = e f () and f L 1 , |G(ei )| > 0
almost everywhere. At such ei , G is continuous and non-zero on the cone
u ) has a continuous extension to
K =  (ei ). Consequently log G = (u + i
K {z : |G(z) G(ei )| < 21 |G(ei |} and the limit (3.1) exists at ei .
There is a close connection between conjugate functions and conformal
mappings. If u is harmonic and if |u| < /2, then
z
u )( )
ei(u+i
d
(z) =
0

is a conformal map from D to a simply connected domain and u = arg  .


Indeed, if a = b D, then
1
 (a + t[b a])dt = 0,
(b) (a) = (b a)
0

Re(  )

because
> 0.
When f is bounded, or even continuous, it can happen that 
f is not bounded.
For an example, let u + i
u be the conformal map of D onto the region
{0 < x < 1/(1 + |y|)}.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Conjugate Functions

51

Then u is continuous on D by Carathodorys theorem, but 


u is not bounded.
For a second example see Exercise 11. The next two theorems get around the
obstruction that 
f may be unbounded even when f is continuous.
Theorem 3.1 (Zygmund). Let f L (D) be real with || f || 1.
(a) For 0 < < /2 there is a constant C , depending only on , such that

d

C .
e| f ()|
2
(b) If f C(D), then for all <



u (r ei ) d
sup
e 
< .
2
0<r <1
Proof. Let u(z) be the Poisson integral of f in D and consider the analytic
function g(z) = 
u (z) iu(z). For r < 1, g(z) satisfies
2
i d
eg(r e ) ,
eg(0) =
2
0
and because 
u (0) = 0, that gives
2
d
u (r ei )
cos u(0) =
e
cos u(r ei ) .
2
0
But if 0 < < /2, then
0 < cos < cos u 1
because |u| 1, so that

u (r e
e

i )

d
sec .
2

Then by (3.1) and Fatous lemma


2
 d
sec .
e f ()
2
0
By repeating this argument with f ( ), we then obtain part (a) with constant
C = 2 sec .
To prove (b), fix < and take a trigonometric polynomial
p( ) =

N

(an cos n + bn sin n )
n=0

CY535/Garnett

0 521 47018 8

52

January 27, 2005

15:41

II. Finitely Connected Domains

such that  f p < /(2). Then



p (r ei ) =

N


r n (an sin n bn cos n )

n=0

is bounded, while part (a) gives


B = sup

0<r <1

u p)(r e
e|(

i )|

d
< .
2

Therefore, since |
u | |
p | + |(u
p)|, we have

u (r ei )| d
p 
B e
sup
< ,
e|
2
0<r <1


and (b) is proved.


Let 0 < < 1. The Lipschitz class C is

!
 f ( + t) f ( )
<

.
C = f L (D) : sup
t
t>0

Every f C agrees almost everywhere with a function continuous on D.


The class C is given the Lipschitz norm
 f C =  f  + sup
t>0

 f ( + t) f ( )
.
t

(3.2)

In a moment we shall prove Privalovs theorem that 


f C whenever f C .
In the Poisson integral formula
it
dt
e +z
f (t) ,
u(z) = Re
it z
e
2

the kernel (eit + z)/(eit z) is analytic in z D and real at z = 0. Then the


uniqueness of 
u shows
it
dt
e +z
f (t) .
(3.3)
u(z) + i
u (z) = F(z) =
it z
e
2

The analytic function F(z) in (3.3) is called the Herglotz integral of f . Write
u u 
u =
,
x y
when u is differentiable on an open plane set. If u is harmonic and bounded
on D then u is the Poisson integral of some f L (D) and by (3.3) and the

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Conjugate Functions

53

CauchyRiemann equations,

 

|u(z)| =  F  (z) = 

eit f (t) dt 
.
(eit z)2

Therefore (as in Exercise I.5(e))



1
dt 
|u(z)|
 f  2(1 |z|)1  f  .
it
2
|e z|

(3.4)

(3.5)

We will often use the following consequence of (3.5): If u(z) is harmonic


and |u(z)| M on B(z, R), then
sup |u|
B(z, R2 )

4M
.
R

(3.6)

To prove (3.6) simply apply (3.5) to U (w) = u(z + Rw).


The next theorem shows that f C if and only if the estimate (3.5) can be
upgraded to


|u| = O (1 |z|)1 .
Theorem 3.2. Let 0 < < 1, let f L (D) be real, and let u(z) be the
Poisson integral of f . Then the following conditions are equivalent:
(a) f C .
(b) 
f C.


(c) |u(z)| = O (1 |z|)1 .
(d) u C (D); that is, for all z 1 , z 2 D,


|u(z 1 ) u(z 2 )| = O |z 1 z 2 | .
Moreover there is a constant C1 , independent of , such that
C1
|| f ||C ,
(1 )

(3.7)

C1
(1 |z|)1 || f ||C ,
(1 )

(3.8)

|| 
f ||C
|u(z)|
and
sup

z 1 =z 2

"
#
|u(z 1 ) u(z 2 )|
C1
1
sup
(1

|z|)

|u(z)|
.
|z 1 z 2 |
|z|<1

(3.9)

The equivalence (a) (b) was first proved by Privalov [1916] who worked
directly with the imaginary part of the integral (3.3); (a) (c) was proved
by Hardy and Littlewood in [1931].

CY535/Garnett

54

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

Proof. Clearly (d)  (a) because if (d) holds, u is uniformly continuous on


D. We first show (a)  (c) and establish (3.8) and then show (c)  (d)
and establish (3.9). Then (a) (b) and inequality (3.7) will follow because
|
u | = |u| by the CauchyRiemann equations.
Assume (a) holds, that is, assume u is the Poisson integral of f C .
(3.8). By (3.5) we may assume |z| 1/2. Let eit0 = z/|z|. Because
Weitprove
it
e /(e z)2 dt = 0, (3.4) yields
 eit ( f (t) f (t )) dt 
0


|u(z)| = 
,
it z)2

(e

so that


|u(z)|
=

| f (t) f (t0 )| dt
|eit z|2

+

|tt0 |<1|z|

1|z||tt0 |<

The inequality 1 |z| |eit z|, valid for all t, gives


1|z|

2 f 
| f (t) f (t0 )| dt
dt

t
it z|2
2

|e
(1

|z|)
|tt0 |<1|z|
0
2 f 
(1 |z|)1 .
=
(1 + )
When 1 |z| |t t0 | < , the inequality |t t0 |2 c|eit z|2 ,
valid with c independent of z, gives


| f (t) f (t0 )| dt
dt

2c
f

t 2

it
2

|e z|
1|z|<|tt0 |<
1|z|
2c f 
(1 |z|)1 .

(1 )
Therefore we obtain (3.8) with
2
C1 = sup
0<<1



(1 )
2(c + 1)
c+
=
.
(1 + )

If (c) holds, we may assume (1 |z|)1 |u(z)| 1. Let z j = r j ei j D,


as in Figure II.4. We may assume |z j | 1/2 and = |z 1 z 2 | 1/2. Set
w j = (1 )z j . Then
|u(z 1 ) u(z 2 )| |u(z 1 ) u(w1 )| + |u(z 2 ) u(w2 )| + |u(w1 ) u(w2 )|.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Conjugate Functions

However,



|u(z j ) u(w j )| = 

while

rj
(1)r j

55

1
u i j 

(1 t)1 dt
(te )dt 
,
t

1
|u(w1 ) u(w2 )| |w1 w2 |Max j 1 |w j |
.

Therefore (d) and (3.9) hold.

z1
w1 = (1 )z 1
w2 = (1 )z 2

z2

ei1
ei2

Figure II.4
Now suppose (a) holds. Then (c) holds and |
u | = O((1 |z|)1 ). Therefore (d) and (3.9) hold for 
u , and 
u extends continuously to D where 
u has



boundary value f . It follows that f C because (d) implies (a). Finally since

f = f + u(0), it follows that (a) (b) and that (3.7) holds.

Let k be a non-negative integer, let 0 < 1 and let f C(D). We say
f C k+ if f is k times continuously differentiable on D and (d/d )k f C
if > 0. If F(z) is analytic on D we say F C k+ (D) if F and its first k
derivatives F  , F  , , F (k) extend continuously to D and if there is C such
that
|F (k) (z 1 ) F (k) (z 2 )| C|z 1 z 2 |
for all z 1 , z 2 D.
Corollary 3.3. Assume k is a non-negative integer and assume 0 < < 1. Let
f C(D) be real and let F(z) = u(z) + i
u (z) be the Herglotz integral of f .
k+
k+
(D) if and only if f C
(D).
Then F C
Proof. Because f = ReF, it is clear that f C k+ (D) if F C k+ (D).

CY535/Garnett

0 521 47018 8

56

January 27, 2005

15:41

II. Finitely Connected Domains

Assume f C k+ . If f ( ) has Fourier series


f ( )

an ein ,

then F has Taylor series


F(z) = a0 + 2

an z n .

n=1

Therefore d f /d , which has Fourier series


df

inan ein ,
d

has Herglotz integral i z F  (z) and the corollary follows from the special case

k = 0, which is Theorem 3.2.
Theorem 3.2 fails when = 1. The harmonic conjugate of a continuously
differentiable function on D need not have a continuous derivative, and the
conjugate of a Lipschitz function, that is, a function satisfying


 f ( + t) f ( ) M|t|,
need not be a Lipschitz function. See Exercise 13. For the same reason, Corollary 3.3 is also false when = 0 and k is a non-negative integer. However,
conjugation does preserve the Zygmund class Z of continuous f on D such
that
 f ( + t) + f ( t) 2 f ( )
< .
sup
t
t>0
The Zygmund class has norm
 f  Z =  f  + sup
t>0

When f
Define

Z ,

 f ( + t) + f ( t) 2 f ( )
.
t

we say f is a Zygmund function.


1
 


2 u(z) = u x  = |u x x (z)|2 + |u yx (z)|2 2

  2 eit f (t)


 

=  F  (z) = 
dt
,
(eit z)3

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Conjugate Functions

57

where F = u + i
u is the Herglotz integral of f L 1 (D). If f C , then
(3.8) and (3.6), applied to u x in the disc B(z, 1|z|
2 ), give us



2 u(z) = O (1 |z|)2 .
(3.10)
holds,
F  along radii shows that

Conversely,
if (3.10) 1

then integrating

u(z) = O (1 |z|)
, and f C . Thus (3.10) provides yet another
characterization of C functions, in terms of second derivatives. Zygmund functions have a similar characterization.
Theorem 3.4. Let f L (D) be real and let u(z) be the Poisson integral of
f . Then the following are equivalent:
(a) f Z .
(b) 
f Z .

(c) 2 u(z) = O (1 |z|)1 .


There is a constant C such that
 
 

(3.11)
f Z C  f Z
and
 




1
 f  || f || + sup (1 |z|)2 u(z) C  f  .
Z
Z
C
zD

(3.12)

Notice that by (3.12) and (3.10), Z C for all < 1. In particular, if


f are continuous. On the other hand, if f is Lipschitz, then
f Z , then f and 

clearly f Z . Thus we have the increasing scale of spaces


C 1 Lipschitz Z C C,
< 1, where C 1 denotes the space of continuously differentiable functions and
C is the space of continuous functions.
Proof. The logic is the same as in the proof of Theorem 3.2. First assume (a)
holds. Fix z D . We may assume z = |z| = Rez and |z| > 21 . Because

ei
d = 0
i
(e z)3
and f ( ) has Herglotz integral F(z), we have

it
e f (t) + f (t) 2 f (0)
2u
1

(z) = ReF (z) =
dt

x2
(eit |z|)3

CY535/Garnett

0 521 47018 8

January 27, 2005

58

15:41

II. Finitely Connected Domains

and
 2u  



 2 (z) = ReF  (z)
x


|t|1|z|

 f  Z |t|
c
dt +
3

(1 |z|)


1|z|<|t|

 f Z
dt
|t|2

C(1 |z|)1  f  Z .

  2 
Unfortunately this trick does not us help with ImF  (z) =  xuy . Instead,

we apply (3.6) to U = 2 u/ x 2 on B(z, (1|z|)


2 ), getting
 3u 
C f  Z


.


2
y x
(1 |z|)2

An integration then yields


|z|
 2u
 

C f  Z
 f Z


(z)  F  (0) +
.
ds C 

2
y x
1 |z|
(1 s)
0
That proves (c), and the right inequality in (3.12).
f are continuous.
Now assume (c) holds. Then by (3.10), f C and f and 
Fix and t with 0 < t and set r = 1 (t/ ). Then
f ( + t) + f ( t) 2 f ( ) = f ( + t) u(r ei(+t) )
+ f ( t) u(r ei(t) )
(3.13)

+ 2u(r ei ) 2 f ( )
+ u(r e

i(+t)

) u(r e )

+ u(r e

i(t)

) u(r ei ).

ei(+t)
ei
r ei(+t)
r ei
r ei(t)
Figure II.5

ei(t)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Boundary Smoothness

59

Because



 2
u(r ei(+t) ) + u(r ei(t) ) 2u(r ei ) |t|2 sup  u (w),
2
|w|=r
the last two terms on the right side of (3.13) are O(|t|) by (c). Then because
i
lims1 (1 s) u
s (se ) = 0 by (c), an integration by parts shows
1
2u
u
i
(1 s) 2 (sei )ds + (1 r ) (r ei ).
f () u(r e ) =
r
s
r
Therefore when (c) holds, the sum of the first three lines of the right side of
(3.13) is

t
u r (r ei(+t) ) + u r (r ei(t) ) 2u r (r ei ) + O(t)



t
t 
u r (r ei(+t) ) u r (r ei ) + u r (r ei(t) ) u r (r ei ) + O(t)


 2u
2t 2
sup 
(w) + O(t) C  t.
|w|=r r

That proves (a), and the left-hand inequality in (3.12) can be seen by following
the constants
argument.
 the previous

 in
u  = 2 u , it follows that (a) (b) and that (3.11) holds. 
Since 2

4. Boundary Smoothness
Let  be a Jordan domain with boundary  and let be a conformal mapping from D onto , so that extends to a homeomorphism from D to the
Jordan curve  = . In this section we examine the connection between the
smoothness of  and the differentiability of on D. When  has some degree of smoothness, we also study the relation between the differentiability of
f C() and the differentiability of its solution u f to the Dirichlet problem
at points of .
The results do not depend on the choice of the mapping : D  because
any other such map has the form T, with T M. We first show that the
smoothness of in a neighborhood of 1 ( ) depends only on the smoothness
of  in a neighborhood of .

CY535/Garnett

60

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

Theorem 4.1. Let 1 and 2 be Jordan domains such that 1 2 and let
1 2 be an open subarc. Let j be a conformal map of D onto  j .
Then = 21 1 has an analytic continuation across 11 ( ), and  = 0
on 11 ( ).
Proof. The analytic function = 21 1 from D into D has a continuous
and unimodular extension to the arc 11 ( ). By Schwarz reflection has an
analytic and one-to-one extension to a neighborhood of 11 ( ) in C and hence

 = 0 on 1 1 ( ).
Let  be an arc parameterized as { (t) : a < t < b}. We say  has a tangent
at 0 = (t0 ) if
(t) 0
= ei
| (t) 0 |

(4.1)

(t) 0
= ei .
| (t) 0 |

(4.2)

lim
tt0

and
lim
tt0

If (4.1) and (4.2) hold then  has unit tangent vector ei at 0 . Except for
reversals of orientation, the existence of a tangent at 0 and its value ei do
not depend on the choice of the parameterization t (t). We say  has a
continuous tangent if  has a tangent at each  and if ei ( ) is continuous
on .
Theorem 4.2. The curve  has a tangent at = (ei ) if and only if the limit


(z)
lim arg
(4.3)
z ei
D zei
exists and is finite. In that case,



(z)
= ( )
lim arg
z ei
2
D zei

(modulo 2 ).

(4.4)

The curve  has a continuous tangent if and only if arg  (z) has a continuous
extension to D. If  has a continuous tangent, then C (D) for all < 1
and  is rectifiable.
There exist Jordan domains  and conformal maps : D  such that 
has a continuous tangent but C 1 (D). An example can be built from the
connection in Section 3 between conjugate functions and conformal mappings:
If u(ei ) is continuous and |u| < 2 , then u(ei ) = arg  (ei ), where is a
conformal mapping onto a Jordan domain, but 
u (ei ) = log |  (ei )| may
not be bounded above or below.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Boundary Smoothness

Proof. Set


v(z) = arg

61


(z)
.
z ei

Then v(z) is harmonic on D and v is continuous on D \ {ei }. If  has a tangent


at = (ei ) or if the limit (4.3) exists at ei , then v(z) is bounded on D. Thus
in either case v is the Poisson integral of v(ei ). Therefore v has a continuous
extension to ei if and only if v| D has a continuous extension to ei , and if and
only if  has a tangent at . Clearly, (4.4) holds when v is continuous at ei .
Now suppose  has a continuous tangent. Then is continuous on D.
If h = 0, then
(zei h ) (z) 
Ah (z) = arg
z(ei h 1)
is continuous on D and harmonic on D. For |z| < 1,
lim Ah (z) = arg  (z).

h0

For |z| = 1 there exists k, 0 < |k| < |h|, such that

(zei h ) (z) 

= arg (zei h ) (z) = ((zeik )),
arg
h
by the proof of the mean value theorem from calculus. Consequently

lim Ah (z) = ((z)) argz ,


h0
2
with the convergence uniform on D, and hence arg(  ) is the Poisson integral
of the continuous function

((ei )) .
2
Now suppose arg  (z) has a continuous extension to D. For r < 1, the curve
i
r = {(r ei ) : 0 2 } has tangent ei ((r e )) satisfying
ei ((r e

i ))

= iei eiarg (r e ) ,
i

and ei ((z)) has a continuous extension to D \ {0}. Then for h > 0






arg (ei(+h) ) (ei ) = lim arg (r ei(+h) ) (r ei )
r 1

= lim ((r ei(+kr ) )),


r 1

where 0 < kr < h, again by the proof of the mean value theorem. Therefore


lim arg (ei(+h) ) (ei ) = lim ((z)),
h0

zei

CY535/Garnett

0 521 47018 8

62

January 27, 2005

15:41

II. Finitely Connected Domains

and a similar argument for h < 0 yields




lim arg (ei(+h) ) (ei ) = lim ((z)) + .
zei

h0

Thus  has a continuous tangent.


If  has a continuous tangent, then by part (b) of Theorem 3.1,


i

sup |  (r ei )| d = sup e(arg )(r e ) d = B <
r <1

r <1

for all < , where B depends on and . Take = 1/(1 ), where


0 < < 1. Let a < b < a + . Then for any r < 1, Hlders inequality gives
 b
1
b

i


i
| (r e )|d |b a|
| (r e )| d
|b a| B1 .
a

Therefore if a = 0 < 1 < . . . < n = b, then


b
n

|(r ei j ) (r ei j1 )|
|  (r ei )|d |b a| B 1 ,
j=1

(4.5)

and letting r tend to 1, we see that  is rectifiable and that C .

Let k be a non-negative integer and let 0 < 1. We say the curve  is of


class C k+ if  is rectifiable and if, in the arc length parameterization
 = { (s) : 0 s () = length ()},
the function is k times continuously differentiable and d k /ds k C if
> 0.
Theorem 4.3 (Kellogg). Let k 1 and 0 < < 1. Then the following conditions are equivalent:
(a)  is of class C k+ .
(b) arg  C k1+ (D).
(c) C k+ (D) and  = 0 on D.
Note that if = 0 and k 1 then (a) does not imply (c) but (a) is still equivalent
to (b). We need the elementary lemma.
Lemma 4.4. Let k be a positive integer, and let 0 < 1. Let f C 1 ([0, 1])
satisfy f  > 0 and let g f 1 . Then g C k+ if and only if f C k+ .
Proof. The case = 0 and k = 1 is clear since g  (y) = 1/ f  (g(y)) > 0. If
= 0 and k 2, the proof is by induction: if f C k and g C k1 , then
f  g C k1 and because f  > 0, g  = 1/( f  g) C k1 . Hence g C k .
Now suppose > 0 and k = 1. If f  C then

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Boundary Smoothness

63





1
1



|g (y1 ) g (y2 )| =  
f (g(y1 ))
f (g(y2 )) 


| f  (g(y2 )) f  (g(y1 ))|


min | f  |2

C|g(y2 ) g(y1 )| C  |y2 y1 | ,


and so g  C . Finally, suppose > 0 and k 2. If f C k+ , then g C k
and g (k) can be written as a sum of products of the functions
g (1) , . . . , g (k1) , f (2) g, . . . , f (k) g.
All these functions are C 1 , except perhaps f (k) g. But f (k) g C , and
thus g (k) C . The converse follows from interchanging g and f .

Proof of Theorem 4.3. If arg  C k1+ , 0 < < 1, then by Corollary 3.3,
log |  | C k1+ , and by taking exponentials  C k1+ (D) and  = 0.

Conversely, if  C k1+ (D) and  = 0, then ei arg =  /|  | C k1+ ,
and arg  C k1+ . Hence (b) and (c) are equivalent.
Assume (c) holds. If

|  (eit )|dt,
s( ) =
0

s

s  ( )

then > 0 and


4.4, and by (4.4)

arg

|  (ei )|

C k1+ . Thus  (s) C k1+ by Lemma

d
= arg  (ei(s) ) + + (s).
ds
2

(4.6)

Since arg  C k1+ and  C k1+ , we conclude from Corollary 3.3 that
d /ds C k1+ and that  is of class C k+ . Therefore (a) holds.
Now assume (a) holds, that is, assume  is of class C k+ . Then by Theorem
4.2, arg  C and by (4.6), d (s( ))/ds C. If k = 1, so that d /ds C ,
then


 d

 (s(1 )) d (s(2 )) C|s(1 ) s(2 )|
(4.7)
 ds

ds
so that by (4.5)



 d

 (s(1 )) d (s(2 )) C|1 2 |(1)
 ds

ds

for any > 0. Thus arg  C (1) and by Corollary 3.3,  C (1) .
Thus s  ( ) = |  (ei )| C (1) and |s(1 ) s(2 )| K |1 2 |. But then

CY535/Garnett

64

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

by (4.6) and (4.7), arg  C . That proves (b) when k = 1. Moreover, by Theorem 3.2 we have s  ( ) = |  (ei )| C , and s C 1+ . If k 2, we use induction. If d /ds C k1+ and if s( ) C k1+ , then by (4.6), arg  C k1+ ,
so that s  = |  | C k1+ , again by Theorem 3.2, and s C k+ . That gives

(b) in general.
Let l be a non-negative integer and let 0 < 1. If  is of class C k+ and if
f C() we say f C l+ () if f ( (s)) C l+ , when viewed as a function
of arc length on .
Corollary 4.5. Suppose  is of class C k+ , where k + > 1, and suppose
f C l+ (). Set n + = min(k + , l + ), where 0 < < 1 and n is a
non-negative integer. Let be a conformal map of D onto , let G be the
Herglotz integral of f , and let F = G 1 . Then F C n+ ().
Proof. Use Corollary 3.3, Theorem 4.3, and Lemma 4.4.

The same result holds for finitely connected Jordan domains whose boundary
curves are of class C k+ , except that conjugate functions and Herglotz integrals
cannot be defined in multiply connected domains.
Corollary 4.6. Let  be a finite union of pairwise disjoint C k+ Jordan curves,
where k + > 1, and let f C() be a C l+ function of arc length on each
component of . Set n + = min(k + , l + ), where 0 < < 1 and n
is a non-negative integer. Then u(z) = u f (z) and its first n partial derivatives
extend continuously to  and
|D n u(z 1 ) D n u(z 2 )| K |z 1 z 2 | ,
for all z 1 , z 2 , where D n denotes any n-th partial derivative.
Proof. By Theorem 4.1 it is enough to work in some neighborhood of .
Let J be the component of  such that J . Let 1 be that component of
C \ J such that  1 , and let u 1 be the solution to the Dirichlet problem
on 1 with boundary value f . Near J , u 1 has the required smoothness by
Corollary 4.5. If 1 is a conformal map of D on 1 , then v = (u u 1 ) 1
is harmonic on A = {r < |z| < 1}, for some r < 1, v is continuous on A, and
v = 0 on {|z| = 1}. By reflection v extends to be harmonic across D. Hence
v C and
u = u 1 + v 11
has as much smoothness as u 1 and 1 both have.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Boundary Smoothness

65

Corollary 4.6 answers Question 1 from the end of Section 2. The next corollary answers Question 2.
Corollary 4.7. If  consists of finitely many pairwise disjoint Jordan curves
of class C 1+ , where > 0, then
d(z, ) =

g(z, ) ds( )
.
n
2

(4.8)

In other words, harmonic measure for z  is absolutely continuous to arc


length on , and the density
1 g(, z)
d
=
ds
2 n
is of class C () and satisfies
c1 <

d
< c2
ds

(4.9)

for positive constants c1 and c2 .


Proof. Let be a conformal map from  onto , where  consists of
analytic Jordan curves. If = ( ) and z = (z ), then
g (, z)
1
g ( , z )
=
n
n
|  ( )|
and
|  ( )| =

ds( )
,
ds( )

by Corollary 4.6 and the uniqueness of Greens function. Since harmonic measure is conformally invariant, (4.8) now follows from the case when  is
analytic, as in Corollary 2.6. By Corollary 4.6, g/n C (). And (4.9)

holds because |  | > 0 on  , by Theorem 4.3.
This answers Question 2 from Section 2. When  is of class C 1 , harmonic
measure is absolutely continuous to arc length, but the density may not be
continuous, bounded or bounded below. See Exercise 14.
Exercise 20 shows how Greens theorem can be applied when  is of class
C 1+ , without first mapping to a domain with real analytic boundary.

CY535/Garnett

0 521 47018 8

66

January 27, 2005

15:41

II. Finitely Connected Domains

Notes
This chapter is based on Lennart Carlesons 1985 lectures at UCLA. The conformal maps in Figure II.2 were computed using the Zipper algorithm from
Marshall [1993]. Stein [1970] gives a good introduction to the scale of Lipschitz spaces beyond C and Z . For more complete discussions see also Adams
[1975], Adams and Hedberg [1996], Bennett and Sharpley [1988], Stein [1993],
and Triebel [1983] and [1992]. See Zygmund [1959] for the classical picture.
Theorem 4.3 and its corollaries are due to Kellogg [1929]. Other boundary
problems are discussed in Appendix B and a different approach, through the
Dirichlet principle, is presented in Appendix C.

Exercises and Further Results


1. Suppose f is real valued, bounded, and Borel measurable on , where 
is a finitely connected Jordan domain and let u = u f be the solution to the
Dirichlet problem on  for f given by (1.4).
(a) Prove
lim inf f ( ) lim inf u(z) lim sup u(z) lim sup f ( ).

 0

 z0

 z0

 0

(b) If f is continuous at , prove lim z u(z) = f ( ).


Hint: Use Theorem I.1.3, Carathodorys theorem, and Theorem 1.1 applied
to a piecewise constant function.
2. A numerical routine for solving the Dirichlet problem in a finitely connected
Jordan domain bounded by C 1+ curves can be based on the following
construction: With u n and u n+1 as the proof of Theorem 1.1, find functions
h and k, explicitly in terms of the data f and Greens functions g and g  of
the simply connected domains 1 and 1 , such that

u n+1 = h +
ku n ds on  .


It follows that u satisfies the integral equation



u=h+
kuds on .


3. (a) Let  be a finitely connected Jordan domain, let E be a finite union of


open arcs in  and let f (z) be a bounded analytic function on . If
| f (z)| M,

z ,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

67

and
lim sup | f (z)| m,

E,

then
| f (z)| m M 1 ,

z ,

where = (z, E, ). Hint: Apply Exercise I.3(b) with log | f (z)|.
(b) On the annulus {r1 < |z| < r2 } show the circle {|z| = r2 } has harmonic
measure
log |z| log r1
.
log r2 log r1
(c) (three circles theorem) Suppose that f (z) is analytic on the annulus
{r1 < |z| < r2 }, and suppose that


m j = sup lim sup | f (z)|
| |=r j

is finite for j = 1, 2. Then


(r )

sup | f (z)| m 1

|z|=r

1(r )

m2

where
(r ) =

log r2 log r
.
log r2 log r1

(d) (Lindelf) Suppose f (z) is bounded and analytic on the disc D. Let
be an arc in D terminating at a point D, and assume
lim

f (z) = a.

Then f (z) has nontangential limit a at . Hint: Take a = 0. If  is a nontangential cone terminating at , then lim inf z (z, ) > 0. Then (a) can be
used.
(e) (Lindelf) Suppose f (z) is bounded and analytic on D. If the nontangential limit f (ei ) is continuous in some deleted arc ( , ) (, + )
and if the two limits
lim f (ei(+) )
0

and
lim f (ei(+) )
0

CY535/Garnett

68

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

both exist, then these two limits are equal.


(f) Let be a univalent function on D. If is an arc in D with endpoint
D and if
lim (z) = w,

then has nontangential limit w at .


4. Suppose  is a finitely connected Jordan domain, and suppose n are finitely
connected Jordan domains bounded by disjoint C 1+ Jordan curves such that

n n+1  and n = . Fix z  and let dn denote harmonic
measure for z with respect to the region n . Prove that dn converges weakstar to d where d is harmonic measure for z with respect to . In other
words, if f C(), then prove


lim
f dn =
f d.
n

n


Hint: If u f is the harmonic extension of f | to , lim (u f f )dn = 0
and thus any weak-star cluster point d of dn satisfies  f d = u f (z).
So by the Riesz representation theorem and (1.2), d = d.
5. Let  be a finitely connected domain such that  has no isolated points.
(a) Prove there is a conformal map of  onto a domain  bounded by
finitely many pairwise disjoint analytic Jordan curves.
(b) Define g (z, w) = g ((z), (w)), and verify (2.2), (2.4), (2.5), and
lim

 z 

g (z, w) = 0.

(2.3 )

(c) Then (2.3 ), (2.4), and (2.5) uniquely determine g and this definition of
g is independent of the choice of  and .
(d) Using the preceding results and the Riemann mapping theorem, derive the
conclusion of Theorem 2.1 for any simply connected domain  not bounded
by a Jordan curve.
6. Let  be a simply connected plane domain and assume that there exists
a (Greens) function g (z, w) on   \ {z = w} satisfying (2.3 ), (2.4),
and (2.5). Let u(z) = g(z, w) + log |z w|. Prove the Riemann mapping
theorem by showing that
u (z))
(z) = (z, w) = (z w)e(u(z)+i

defines a conformal mapping from  onto D. Hint: To show is one-to-one,


compare log |T (z, w)| to g(z, w ) for some T M.
7. (a) If  is a doubly connected Jordan domain, if E is one of the boundary

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

69

components, and if = (z, E, ), show that for some choice of a > 0,
)
= ea(+i

defines a single valued conformal map of  onto an annulus.


(b) Let  be a domain such that C \  has exactly two components. Prove
there is a conformal mapping from  onto a ring {z : r1 < |z| < r2 } where
0 r1 < r2 .
(c) Let  be a finitely connected Jordan domain and suppose z 0 . Set
u (z) = log(1/)(z, B ,  \ B ),
where B = {z : |z z 0 | }. Then
|u g (z, z 0 )| C(z, B ,  \ B ),
and hence u g (z, z 0 ), uniformly on compact subsets of  \ {z 0 }.
(d) Let  be a finitely connected Jordan domain bounded by C 1+ curves.
The period of the conjugate 
g (z, z 0 ) of Greens function around a component  j of  is equal to the harmonic measure of  j at z 0 .
8. (a) Let u(z) be the Poisson integral of f L 2 . Use Greens theorem to show


2
1
1
2
|u(z)|2 log d xd y
| f f (0)| d =
2

|z|
D

2

|u(z)|2 (1 |z|)d xd y

D

4
1

|u(z)|2 log d xd y.
3
|z|
D

u(r ei ),

Hints: Replace f by
use the identity (u 2 ) = 2|u|2 and Greens
theorem, and send r 1. The final inequality is a direct calculation using
polar coordinates and Fourier series. The identity on the left can also be
proved with Fourier series.
(b) If f H 2 (see Appendix A), then by (a)


2
1
1
| f  (z)|2 log d xd y
| f f (0)|2 d =
2

|z|
D

2

| f  (z)|2 (1 |z|)d xd y

D

4
1

| f  (z)|2 log d xd y.
3
|z|

CY535/Garnett

70

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

In particular, the two area integrals are comparable.


9. (a) In Theorem 2.7, verify (2.16) when f L 1 (ds).
(b) Prove (2.20) by following the conformal mapping proof of (2.16) given
in the text.
10. (a) Prove that every analytic Jordan curve is the image of D under an analytic
map that is one-to-one on a neighborhood of D.
(b) If  is a domain in C then we can find subdomains n , each bounded
by finitely many pairwise disjoint analytic Jordan curves so that n n+1

and  = n . Hint: Use the proof of Lemma 2.2.
11. (a) If f L 1 (D), and if u is the Poisson integral of f , then

2r sin( )
d
i
f (ei ) .

u (r e ) =
2
2
1

2r
cos(

)
+
r

(b) Almost everywhere on D,




f () = lim

0 ||>


cot


f (ei )

d
.
2

(c) Define the modulus of continuity of the function f as


f () = sup{| f ( ) f ()| : | | < }.
The function f is called Dini continuous if

f (t)
dt < .
t
0
Prove that 
f is continuous if f is Dini continuous, in fact,


(t)
(t)
dt
+

()

dt.

f
t
t2
0

h(r ) exists and


(d) Let h( ) be odd and increasing on (, ]. Then limr 1 
is finite if and only if

h( )
d < .
0
(e) Use (d) to show (c) is sharp.
12. Let  be a simply connected domain. Write d(w) = dist(w, ) and let
(w, w0 ) denote the hyperbolic distance in . Let : D  be conformal.
Then C (D) if and only if


1
lim sup 2(w, w0 ) + log d(w) <

d(w)0
for any w0 . Hint: Use Theorem 3.2 and Theorem I.4.3. See Becker and
Pommerenke [1982a].

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

71

13. Corollary 3.3 is false when k = 1 and = 0. Worse yet, exists f C 1 (D)
such that 
f is not a Lipschitz function.
14. There is a simply connected domain  such that  C 1 but such that no
conformal map : D  is class C 1 on D. Worse yet, |  | can have infinite
nontangential limit at some point on D.
15. In Theorem 4.2, if  has a continuous tangent, then  H p for all finite p.
In other words,

|  (r ei )| p d < .
sup
0<r <1

See Appendix A.
16. Let  be a Jordan domain with C 1 boundary  parameterized by arc length
 = { (s) : 0 < s < ( )},

17.

18.
19.

20.

and let be a conformal map from D onto .


(a) If is Dini continuous, then  extends continuously to D and  = 0.
See Warschawski [1932a] and [1961].
(b) If C 1 and if  is Dini continuous, then  extends continuously to
D. See Warschawski [1932a].
Let  be a finitely connected Jordan domain with n + 1 boundary components 0 , 1 , . . . , n and let u(z) = (z,  j , ). Prove that u(z) has
n 1 critical points in . In other words, prove the analytic function
f (z) = u x (z) iu y (z) has n 1 zeros in , when counted with multiplicities. Hint: We may assume each k is an analytic curve. Then f = 0
on  and since u is constant on k , the vector u = f is parallel to the
normal vector on k . Hence arg f decreases by 2 as k is traversed counterclockwise, and the argument principle then shows that f has n 1 zeros
in .
Prove that conditions (a) and (b) of Theorem 4.3 are equivalent even when
= 0.
(a) Formulate a notion of nontangential convergence for a finitely connected
Jordan domain with C 1 boundary. Use it to show that (z, E, ) is the
unique bounded harmonic function on  with nontangential limits 1 on E
and 0 on  \ E, almost everywhere with respect to arc length on .
(b) If  is C 1+ for some > 0, then under the conformal mapping cones
 ( ) correspond to cones  (( )), and then (2.16) holds for f L 1 (ds).
Greens theorem can be used on a finitely connected domain bounded by
a finite number of pairwise disjoint C 1+ curves with functions u and v in
C 1 () C 2 (). Hint: Use a finite number of Riemann maps and Theorem
4.3.

CY535/Garnett

72

0 521 47018 8

January 27, 2005

15:41

II. Finitely Connected Domains

21. More generally, Greens theorem also holds for Lipschitz domains, which
will be defined in this exercise. A K -Lipschitz graph is a set of the form
 = {(x, (x)) : |x| M},
where is a real valued Lipschitz continuous function:
|(x2 ) (x1 )| K |x2 x1 |,
for some constant K . A Lipschitz graph is rectifiable, and since is absolutely continuous, arc length on  has the form
ds =

1 + (  (x))2 d x,

and almost everywhere,


(  (x), 1)
n= $
1 + (  (x))2
is a vector normal to . For B > sup[M,M] |(x)|, set
 =  = {(x, y) : |x| < M, (x) < y < B}.
(a) Prove that Greens theorem holds on  in this form: Suppose u C 1 ()
and v C 2 (), suppose u, u, v, v, and v are bounded on , and suppose u and v extend continuously to . Then


v
(E.1)
u ds.
u v + u v d xd y =

 n
Hint: Approximate  by C subdomains as follows: Let C (1, 1)
satisfy 0 and
1
d x = 1.
1

If (x) = 1 ( x ) and () (x) = (x) C , () (x) (x) uniformly, and () (x) (x) almost everywhere. Show there are n 0
such that n = n + (n ) satisfies n > n+1 > . Then derive (E.1) from
Greens theorem for n .
(b) A finitely connected Jordan domain  is a Lipschitz domain if each
p  has a neighborhood V so that V  is the image of a Lipschitz
graph under a linear mapping. Prove Greens theorem (E.1) holds for the
Lipschitz domain  under the same smoothness assumptions on u and v
that we made in (a). That is, u C 1 ();and v C 2 (); u, u, v, v, and
v are bounded on ; and u and v extend continuously to .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

III
Potential Theory

The goal of this chapter is to solve the Dirichlet problem on an arbitrary plane
domain . There are three traditional ways to solve this problem:
(i) The Wiener method is to approximate  from inside by subdomains n
of the type studied in Chapter II and to show that the harmonic measures
(z, E, n ) converge weak-star to a limit measure on . With Wieners
method one must prove that the limit measure (z, E, ) does not depend
on the approximating sequence n .
(ii) The Perron method associates to any bounded function f on  a harmonic
function P f on . The function P f is the upper envelop of a family of
subharmonic functions constrained by f on . Perrons method is elegant
and general. With Perrons method the difficulty is linearity; one must prove
that P f = P f , at least for f continuous.
(iii) The Brownian motion approach, originally from Kakutani [1944a], identifies (z, E, ) with the probability that a randomly moving particle,
starting at z, first hits  in the set E. This method has considerable intuitive appeal, but it leaves many theorems hard to reach.
We follow Wiener and use the energy integral to prove that the limit
(z, E, ) is unique. This leads to the notions of capacity, equilibrium distribution, and regular point and to the characterization of regular points by
Wiener series.
For the Perron method see Ahlfors [1979] or Tsuji [1959]. Appendix F below
includes Kakutanis theorem for the discrete version of Brownian motion.
We conclude the chapter with some potential theoretic estimates for harmonic measure.

73

CY535/Garnett

0 521 47018 8

74

January 27, 2005

15:41

III. Potential Theory

1. Capacity and Greens Functions


Let E be a compact plane set such that  = C \ E is a finitely connected
Jordan domain. By Chapter II and a conformal mapping, we see that  has
Greens function g (z, ) with pole at , and if a 
g (z, ) = log |z a| + h(z, ),
where h(z, ) is harmonic on  and continuous on  and
h(, ) = log | a|

.

(Recall that u(z) is harmonic at if u(1/z) is harmonic on a neighborhood of


0.) The quantity
= (E) = h(, ),
is called Robins constant for E, and we have
g (z, ) = log |z| + + o(1),

as z .

(1.1)

Define the logarithmic capacity of E to be


Cap(E) = e (E) .
Thus Cap(E) > 0 in the case at hand.
Suppose 1 and 2 are finitely connected Jordan domains such that  j
and set E j = C \  j . Assume there is a conformal map of 1 onto 2 , such
that for |z| large
(z) = az + b0 +

b1
+ ...
z

with a > 0. Then


g1 (z, ) = g2 ((z), ),
so that by (1.1)
(E 1 ) = (E 2 ) + log a
and
Cap(E 2 ) = aCap(E 1 ).

(1.2)

In particular, the capacity of a closed disc is the radius of the disc, because
gC \D (z, ) = log |z|.
Now let E be any compact plane set and write  for the component of C \ E
such that . Fix a sequence {n } of finitely connected domains such that
n n n+1 ,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Capacity and Greens Functions

such that
=

75

n ,

and such that n consists of C 1+ Jordan curves for some > 0. See Exercise
II.10(b). Define E n = C \ n . As we shall soon see, all our results will be
independent of the sequence {n }.
Because n n+1 , it follows from from the maximum principle that
gn+1 (z, ) > gn (z, ) on n , and hence that (E n+1 ) > (E n ) and
Cap(E n+1 ) < Cap(E n ). Now define
Cap(E) = lim Cap(E n ).
n

(1.3)

Because g (z, w) is an increasing function of , an interlacing of the domains


n shows that the definition (1.3) does not depend on the choice of the sequence
& = C \ , then by definition,
{n }. Note that if E
& = Cap( E).
&
Cap( E) = Cap(E) = Cap( E)

(1.4)

By definition, the quantity Robins constant


1 
= lim (E n )
(E) log
n
Cap(E)
is Robins constant for the arbitrary compact set E. If E F then by (1.3)
Cap(E) Cap(F) and (E) (F).

(1.5)

If Cap(E) > 0, then limn (E n ) = (E) < , and by Harnacks principle


g (z, ) = lim gn (z, )
n

defines a positive harmonic function on  having expansion


g (z, ) = log |z| + (E) + o(1)

(1.6)

at . When z = , the symmetry (2.8) of Chapter II shows that


lim gn (, w) = lim gn (w, )
n

exists for all w  \ {}. Thus by Harnacks principle the limit


g (z, w) = lim gn (z, w)
n

(1.7)

exists for all z, w  with z = w, and g (z, w) satisfies conditions (2.2), (2.4),
and (2.5) of Chapter II. The function g (z, w) is Greens function for  with
pole at w.

CY535/Garnett

0 521 47018 8

76

January 27, 2005

15:41

III. Potential Theory

Example 1.1. Suppose E is compact and connected. Let  be the component


of C \ E such that  and let :  C \ D be the conformal mapping
such that for |z| large,
b1
+ ...,
z

(z) = az + b0 +
with a > 0. For r > 1,

r = {z : |(z)| > r }
is bounded by an analytic Jordan curve and gr (z, ) = log |(z)/r |. Then
by (1.7), g (z, ) = log |(z)| and by (1.2),
1 = Cap(D) = aCap(E).
For example, the normalized conformal map = 1 of C \ D onto the interval [2, 2] is
(z) = z +

1
z

and so
Cap([2, 2]) = 1.
Consequently
Cap([, ]) =

for every interval [, ] R, by (1.2).

Example 1.2. If E D, then Cap(E) sin |E|/4 .


Proof. If E is an arc on D, then after a conformal mapping, Example 1.1 gives
Cap(E) = sin(|E|/4). Because of (1.3) we may assume E is a finite union of
arcs. Define
i
e +z
1
d,
F(z) =
4 E ei z

and let (z) = (z, E, D). Then F(1/z) = F(z) and F = 2 + i


on D.
Therefore

ReF
2
2
on  = C \ E and F(0) = F() = |E|/4. Hence H (z) = ei F(z) maps 

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. The Logarithmic Potential

77

into the right half-plane and H () = H (0). Now




 H (z) + H (0) 


G(z) = logz
H (z) H (0) 
is superharmonic on  and
lim inf G(z) 0.

z 

By the maximum principle, G(z) > 0 in  so that G(z) gn (z, ) for all n.
For |z| large
G(z) = log |z| log sin

|E|

+ ...,

so that by (1.1)
log sin

|E|

(n ) (E)

and Cap(E) sin(|E|/4).

2. The Logarithmic Potential


Let be a finite, compactly supported signed (Borel) measure. The logarithmic
potential of is the function

1
U (z) = log
d( ).
| z|
By Fubinis theorem, the integral U is absolutely convergent for area almost
every z.
Lemma 2.1. If > 0, the potential U is lower semicontinuous and superharmonic.
Proof. The lemma holds because for fixed, the function
z log

1
| z|

is lower semicontinuous and superharmonic.

The next theorem connects the notions of logarithmic potential, Greens


function, capacity, and harmonic measure. Suppose  = C \ E is bounded by
a finite family of disjoint C 1+ Jordan curves, > 0, write E for the harmonic

CY535/Garnett

0 521 47018 8

January 27, 2005

78

15:41

III. Potential Theory

measure of relative to ,
d E = d(, , ) =

1 g(, )
ds = P ( )ds.
2
n

Theorem 2.2. If  = C \ E is connected and bounded by finitely many pairwise disjoint C 1+ Jordan curves, then the integral U E (z) is absolutely convergent at every z C. The potential U E is continuous on C and
g(z, ) = (E) U E (z),
U E (z) < (E),

z ,

(2.1)

z ,

(2.2)

z E = C \ .

(2.3)

and
U E (z) = (E),

Later we shall see that E is the unique probability measure on E such that
U is constant on E. For this reason E is called the equilibrium distribution
of E.
Proof. We can assume 0
/ . Clearly, the integral U (z) is absolutely convergent at all z
/ . Since

log | | d(z, ),
g(z, ) = log |z|


we have


(E) =

log | | d E ( ).

On the other hand, for fixed z 0 ,






 z 
 

 d(z, ),



g(z, z 0 ) = log 
log 
z z0 
z0 


(2.4)

because the right side of (2.4) satisfies the conditions (2.3), (2.4), and (2.5) from
Chapter II and those conditions determine Greens function. Then sending z to
yields

1
g(, z 0 ) = (E) log
d E ( )
| z 0 |
= (E) U E (z 0 ).
For z , (2.1) is then a consequence of the symmetry of Greens function,
g(, z) = g(z, ). Then because g(z, ) > 0 on , (2.1) implies (2.2).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. The Energy Integral

79

For z
/ , v( ) = log (| |/| z|) is harmonic on a neighborhood of 
and


1
d E ( )
0 = v() = log | |d E ( ) + log
| z|
= (E) + U E (z).
Therefore (2.3) holds for z
/ .
If z , the lower semicontinuity of U gives
U (z) lim inf U (w) (E) < .
 wz

Since the integrand is bounded below, that means the integral U (z) converges
absolutely. Because  consists of C 1+ curves, Area() = 0. Then by the
superharmonicity of U and the continuity of g(z, ),
'
(

d d
d d
U (z) lim sup
U ( )
+

= .
2
2
B(z,)
B(z,)\
0
Consequently (2.1) and (2.3) hold at all z , and it follows that U is

continuous in C .
Let E be a compact set with Cap(E) > 0, and let E n = C \ n be as in
Section 1. Then by (1.7) and Theorem 2.2 any weak-star cluster point E of the
sequence { E n } satisfies both (2.1) and (2.2) on . In Section 4 we will use the
energy integral to show that there is a unique weak-star limit E independent
of the sequence E n and to establish a version of (2.3) for U E on E. A different
proof of the uniqueness of the weak-star limit { E n } for a bounded domain 
is given in Exercise 4.

3. The Energy Integral


Let be a signed measure with compact support. If




log 1  d||( ) d||(z) < ,

|z | 

(3.1)

we say has finite energy and define the energy integral I () by




1
I () =
log
d( )d(z) = U (z)d(z).
|z |
The energy integral has a very important property: It is positive definite on the
space of signed measures with finite energy and zero integral.

CY535/Garnett

80

0 521 47018 8

January 27, 2005

15:41

III. Potential Theory

Theorem 3.1. If (3.1) holds and if


I () = 0, then 0.

d = 0, then I () 0. Moreover, if

1
Proof. Write L(z) = log |z|
. By Greens theorem

1
f (z) =
L(z w) f (w)dudv
2

(3.2)

whenever f C has compact support.


First consider the special case of an absolutely continuous signed measure
d = h(z)d xd y where h C has compact support and satisfies

h(z)d xd y = 0.
(3.3)
Then the convolution U = L h is also C and by (3.3),
|U (z)|

C
|z|

(3.4)

C
|z|2

(3.5)

and
|U (z)|

for |z| large. For any f C with compact support, Greens theorem (3.2) and
Fubinis theorem give



U f d xdy = 2
h f d xdy,
U f d xdy =
because h and f have compact support. Therefore
U = 2 h.
Now by (3.4), (3.5), and Greens theorem again



1
1
I () =
U hd xdy =
U U d xd y =
|U |2 d xd y.
2
2
That shows I () 0 in this special case. Moreover, if I () = 0 then U = 0
1
U = 0.
and h = 2
To derive the full Theorem 3.1 from the special case we employ a standard
mollification argument. Let K Cc (C) be a compactly supported infinitely
differentiable function such that
(i) K is radial, K (z) = K (|z|),
(ii) K 0, and
(iii) K d xdy = 1.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. The Energy Integral

81

Set
K (z) = 2 K (z/)
and let be the absolutely continuous measure having density

h (z) = K (z) = K (z w)d(w).
Then


f d =

lim

f d,

(3.6)

for all continuous f . Furthermore, h C has compact support and satisfies


(3.3). By Fubinis theorem and (i)

K K L(z )d(z)d( ),
I ( ) =
where


K K (z) =

and

K (z w)K (w)dudv

K K L(w) =

K K (z) L(w z)d xd y.

Now because K is a radial function, K K is also a radial function, and


therefore since L(z) is superharmonic,
K K L(z) L(z).

(3.7)

Also, the continuity of L(z) viewed as a map to (, ], gives

Since

K K L(z) L(z),
d = 0,

log

1
d( )d(z) =
|z |

as 0.


log

1
d( )d(z),
|z |

for any > 0, and we can assume has support in {z : |z| < 1/2}. Therefore
by (3.1), (3.7) and dominated convergence,
lim I ( ) = I (),

and we see that I () 0.


Now assume I () = 0. Write U = U . Then

|U |2 d xd y = I ( ) 0.

(3.8)

CY535/Garnett

0 521 47018 8

January 27, 2005

82

15:41

III. Potential Theory


1
We also have U (z) = O |z|
uniformly in , because all h , < 1, satisfy
(3.3) and vanish outside a common compact set. Then by (3.8) and Lemma 3.2
below,

|U (z)|2 d xd y = 0.
lim
0

Let f
more,

have compact support. Then by (3.6) and Greens theorem once



f d = lim

1
f d = lim
0 2


f U d xd y = 0.

Therefore = 0.

C (C)

Lemma 3.2. Assume Un (z)

satisfy

|Un (z)| C/|z|,

(3.9)

and
||Un ||2 0.
Then


|Un |2 d xd y 0,

(3.10)

for every compact set K .


Proof. Assume K [L , L] [R, R]. Then
L

|Un (x, y)|2 d x 2
|Un (x, y) Un (x, L)|2 d x + 2C 2 /L
[L ,L]

L
L
L

 2

 Un 
dy d x + 2C 2 /L



R y
y

4L||Un ||22 + 2C 2 /L ,
and a second integration gives (3.10) when L is large compared to R.

4. The Equilibrium Distribution


As promised at the end of Section 2, we use the energy integral to show that the
sequence { E n } has a unique limit E and that U E = (E) on E \ A, where
Cap(A) = 0. Write P(E) for the set of all Borel probability measures on the

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. The Equilibrium Distribution

83

compact set E. Let  be the component of C \ E such that , and recall


the sequences {n } and E n = C \ n from Section 1. Set E n = C \ n and
& = C \ .
E
Theorem 4.1. Assume Cap(E) > 0. Then sequence {n } = { E n } converges
weak-star to some = E P(E), and the limit does not depend on the choice
of the sequence {E n }. Moreover,
(E) = I ( E ) = inf{I ( ) : P(E)}

(4.1)

and I ( ) > (E) for all P(E) with = E . In other words, E is the
unique probability measure on E satisfying (E) = I ( E ).
It follows that the measure E is the unique P(E) satisfying U = (E)
almost everywhere d. By definition E is the equilibrium distribution of
E, and U E is called the equilibrium potential.
Lemma 4.2. Let E be a compact set and let {n } be a sequence in P(E). If n
converges weak-star to P(E) then
U (z) lim inf Un (z),

z C,

(4.2)

and
I () lim inf I (n ).

(4.3)

Proof of Lemma 4.2. Write h(z, ) = log 1/|z | and for N > 0 define
h N (z, ) = min(h(z, ), N ). Then each h N is continuous, h = lim h N , and
h N h. Therefore


h N (z, )d( ) = lim h N (z, )dn ( )
n

lim inf h(z, )dn ( )
n

= lim inf Un .
n

Because h N increases to h, this gives (4.2). Because n n converges to



weak-star, a similar argument gives (4.3).
Proof of Theorem 4.1. We first assume that  = C \ E is bounded by finitely
many pairwise disjoint C 1+ Jordan curves and in that case we prove E is the
unique minimizer
for (4.1). For such E, U E = (E) on E by (2.3), so that

I ( E ) = U E d E = (E). Thus if P(E), then


(E) = U E d = U d E

CY535/Garnett

0 521 47018 8

January 27, 2005

84

15:41

III. Potential Theory

by Fubinis theorem. But if = E , then by Theorem 3.1,


0 < I ( E )

= I ( ) + I ( E )


U E d

U d E

= I ( ) (E).
Therefore (4.1) holds, and I ( ) = (E) if and only if = E .
Now suppose E is any compact set. By the BanachAlaoglu theorem there
is a subsequence {n j } of { E n } that converges weak-star to some P(E 1 ).
By (4.3),
I () lim inf I (n j ) = lim inf (E n j ) = (E).
Moreover, any such has support contained in

)
*

+
En

& E,
= E

(4.4)

because E n has support E n . If P(E) P(E n ) then, because (4.1) holds


for the sets E n ,
I ( ) lim (E n ) = (E).
n

Therefore I ( ) I () = (E).
. By Fubinis theorem
 Suppose  P(E) satisfies I ( ) = (E) and =
U d = U d, and by Theorem 3.1

0 < I ( ) = I ( ) + I () 2 U d.
(4.5)
Therefore


U d < (E).

(4.6)

But if
t = t + (1 t),

0 < t < 1,

then t P(E) so that

(E) I (t ) = t 2 I ( ) + (1 t)2 I () + 2t (1 t)

(E) + 2


U d (E) t + O(t 2 ).

U d
(4.7)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. The Equilibrium Distribution

85

For t small (4.6) and (4.7) are in contradiction and hence is the unique extremal
for the minimizing problem (4.1). The uniqueness implies {n } converges weakstar, because the sequence {n } has a unique weak-star accumulation point. The
uniqueness also shows that the limit cannot depend on the choice of {E n }. 
For any compact set E,
& E E E,
&
E
& By (1.4) and the uniqueness in Theoand by (4.4), E is supported on E.
rem 4.1,
E = E = E& = E&.
The capacity of any Borel set A is defined as
Cap(A) = sup{Cap(E) : E compact, E A}.

(4.8)

Corollary 4.3. If E is compact and Cap(E) = 0, then I () = + for all


P(E). Consequently, if A is Borel with Cap(A) = 0 and if is a positive
measure with I () < , then (A) = 0.
Proof. Suppose E is compact and Cap(E) = 0 and suppose P(E). Then
P(E) P(E n ) and I () (E n ) by Theorem 4.1. If A is a Borel set
with Cap(A) = 0 and if is a positive measure
with (A) > 0, then take E A
E
P(E) and I () = .
compact, such that (E) > 0. Then = (E)

Thus I ( ) = .
In particular, (4.8) and Corollary 4.3 imply that
Cap(A) = 0  Area(A) = 0.
Sets of capacity zero are also called polar sets.
Theorem 4.4. Let E be a compact set such that Cap(E) > 0 and let E be the
equilibrium distribution for E. Then
U E (z) (E),
U E (z) = (E) g (z, ) < (E),
and

zC
&
z  = C \ E,


& : U E (z) = (E)} = 0.
Cap {z E

(4.9)
(4.10)

(4.11)

Proof. Since U En (E n ) and (E n ) (E), (4.9) follows from Lemma


4.2. By (1.7), (2.1), and the weak-star convergence of E n , we conclude that

CY535/Garnett

0 521 47018 8

January 27, 2005

86

15:41

III. Potential Theory

U E (z) = (E) g (z, ) for z . If U E (z) = (E) at some z , then


by the maximum principle, U E is constant on , but
lim U E (z) = .

Therefore (4.10) holds.


& such that Cap(K ) > 0
To prove (4.11), let K be any compact subset of E
and take P(K ) with I ( ) < . For 0 < t < 1, the inequality (4.7) still
obtains, so that the coefficient of t in (4.7) is non-negative:

U E d (E) 0.
But then by (4.9)


U E d = (E)

and U E (z) = (E) almost everywhere with respect to . Thus Cap(K ) = 0


& : U E (z) = (E)}.

whenever K {z E
Theorem 4.5. Suppose is a contraction: |(z) (w)| |z w| for all
z, w E. Then
Cap((E)) Cap(E).
Consequently, if E = {|z| : z E} is the circular projection of E onto R, then
Cap(E) Cap(E )

|E |
,
4

(4.12)

where |E | denotes the length of E .

E
Figure III.1 Circular projection.
In Theorem 4.5, circular projections can also be replaced by vertical projec = {x : x + i y E}. The proof is the same.
tions: E

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. The Equilibrium Distribution

87

Proof. We may assume that Cap((E)) > 0. Let P((E)) be such that
I () < . By Lemma 4.6 below there is P(E) such that () = ,
where () is defined by
()(A) = ( 1 (A))

(4.13)

for all Borel A (E). Then



1
I () =
log
d(z)d(w) I (),
|(z)

(w)|
E E
and by Theorem 4.1, (E) ((E)) and Cap(E) Cap((E)).
In proving (4.12) we may assume that E is a finite union of smooth curves and
E is a finite union of intervals in R . Let F be the interval obtained by sliding the
intervals of E together without altering their lengths. Then F is obtained from
E by applying two contractions, and hence Cap(E) Cap(E ) Cap(F). But

then by Example 1.1, Cap(F) = |F|/4 = |E |/4.
Lemma 4.6. Let E be a compact set and let be a continuous map from E
onto (E). For P(E) define () by (4.13) above. Then maps P(E)
onto P((E)).
Proof. By its definition, the map : P(E) P((E)) is continuous, weakstar to weak-star. Given P((E)), let n be a sequence of atomic measures
n =

Nn

j=1

(n)

a j z (n) ,
j

(n)
where z denotes the point-mass at z, such that z (n)
j (E), such that a j > 0
 (n)
(n)
(n)
and a j = 1, and such that n weak-star. Pick any w j 1 (z j ),
and set

n =

Nn

j=1

a j(n) w(n) .
j

Then (n ) = n , and if a subsequence of {n } converges weak-star to , then



P(E) and () = .
In contrast to Theorem 4.5, there exist sets E R with |E| = 0 but
Cap(E) > 0.
Example 4.7. The Cantor set is obtained from the unit interval K 0 = [0, 1]
as follows. The (closed) set K n consists of 2n intervals of length 3n obtained
by removing the middle (open) one-third of each interval in K n1 . The Cantor

CY535/Garnett

0 521 47018 8

January 27, 2005

88

15:41

III. Potential Theory

set is
K =

Kn ,

n=0

and
Cap(K )

1
.
9

(4.14)

K0
K1
K2
K3
K4
K5
Figure III.2 The Cantor set.
To prove (4.14), note that by (1.3) Cap(K ) = limn Cap(K n ). The set
K n consists of two subsets K n1 and K n2 , with dist(K n1 , K n2 ) = 13 such that
Cap(K n1 ) = Cap(K n2 ) = Cap(K n1 )/3. In other words,
(K n1 ) = (K n2 ) = (K n1 ) + log 3.
j

Let = (1 + 2 )/2 where j is the equilibrium distribution for K n , j = 1, 2.


Then by Theorem 4.1,



(K n1 ) (K n2 ) 1
1 + 2
1
=
+
+
d1 ( )d2 (z),
I
log
2
4
4
2
|z |
so that by (4.1),

(K n1 ) + log 3
(K n ) 2
4


+

log 3
(K n1 ) + 2 log 3
=
.
2
2

Thus (K ) = lim (K n ) 2 log 3 and Cap(K ) 1/9.

Theorem 4.5 provides another proof of the Koebe one-quarter theorem I.4.1.
Corollary 4.8. Let (z) be a univalent function on the unit disc D. If is
normalized by (0) = 0 and  (0) = 1, then
(D) {w : |w| < 1/4}.

(4.15)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Wieners Solution to the Dirichlet Problem

89

Proof. Let a = inf{|w| : w


/ (D)}. Then
(z) =

1
( 1z )

= z + b0 +

b1
+ ...
z

is a conformal map from C \ D to a domain  and E = C \  is a connected


set with circular projection E = [0, 1/a]. By Theorem 4.5 and Example 1.1,
1 = Cap(E) Cap(E ) =

1
,
4a


which gives (4.15).

5. Wieners Solution to the Dirichlet Problem


Let  be a domain in C such that Cap(C \ ) > 0, let f C(), and let
{n } be the subdomains introduced in Section 1. Extend f to a continuous
function on C , still called f , and set

u n (z) =
f ( )d(z, , n ),
z n .
n

Then Theorem 4.1 yields the following fundamental result from Wiener
[1924a].
Theorem 5.1 (Wiener). If Cap(C \ ) > 0, the sequence {u n (z)} converges
uniformly on all compact subsets of  to a harmonic function u(z), and u(z)
does not depend on the choice of the sequence {n } nor on the choice of the
extension of f C().
We write
u(z) = u f (z)
and we call u f (z) the Wiener solution to the Dirichlet problem with boundary
values f ( ).
Proof. Suppose . As in Section 4, let n = E n = (, , n ) where
E n = C \ n . By Theorem 4.1, the sequence {n } converges weak-star to E ,
where E = C \ . Hence


lim u n () = lim
f ( )d E n ( ) =
f ( )d E ( ),
(5.1)
n

n 

and the latter integral does not depend on the sequence {n }. It also does not
depend on the extension of f because the limit measure E is supported on
.

CY535/Garnett

0 521 47018 8

January 27, 2005

90

15:41

III. Potential Theory

For z  and z = , let T be a Mbius transformation such that T z =


and note that
Cap() > 0 Cap(T ()) > 0.

(5.2)

Indeed, if Cap() > 0, then by (1.7), g (w, T 1 ()) exists and


g (w, T 1 ()) = lim gn (w, T 1 ())
n

= lim gT (n ) (T w, ) = gT () (T w, ),


n

and that implies (5.2). The argument for z = shows limn u n (z) = u(z)
exists for all z  and the limit u(z) is independent of {n } and the extension
of f . Because the functions {u n (z)} are uniformly bounded and harmonic, the
convergence u n (z) u(z) is uniform on compact subsets of  and the limit

u(z) is harmonic in .
Theorem 5.1 enables us to define harmonic measure for any domain  with
Cap(C \ ) > 0, because it says the weak-star limit
(z, , ) = lim (z, , n )
n

(5.3)

exists and is independent of the choice of {n }. Then for any f C(),

u f (z) =
f ( )d(z, , ),


where u f (z) is the Wiener solution to the Dirichlet problem for f ( ). By definition, the limit measure (z, , ) is the harmonic measure for z relative
to . By Exercise 4 of Chapter II, this definition of harmonic measure agrees
with the definition in Chapter II for finitely connected Jordan domains. As in
Chapter I, (z, A, ) is a harmonic function of z  and
0 (z, A, ) 1.
By Harnacks inequality,
(z 1 , E, ) C z 1 ,z 2 (z 2 , E, )
for every pair z 1 , z 2 of points in , where the constant C z 1 ,z 2 > 0 depends only
on z 1 , z 2 , and . Thus the measures (z, , ), z , are mutually absolutely
continuous.
If z = , then (5.1) shows that (, , ) =  . If z  and if T
is a Mbius transformation with T (z) = , then for every Borel set A,
(z, A, ) = T () (T (A)). By (5.2) and Corollary 4.3
Cap(A) = 0  (z, A, ) = 0

(5.4)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Wieners Solution to the Dirichlet Problem

91

for every  with Cap(C \ ) > 0 and every z .


If Cap(C \ ) = 0 then, as we shall see in Section 8, each bounded or positive harmonic function on  is constant. Therefore we cannot define harmonic
measure or solve the Dirichlet problem when Cap(C \ ) = 0.
A major goal of this book is to understand the relation between harmonic
measure and other measures. Every finite F  has harmonic measure zero.
(One proof: Cap(F) = 0 by (4.1), so that (z, F, ) = 0 for all z  by
(5.4).) We shall need the following quantitative form of this fact. Let {n } be
the increasing sequence of domains introduced in Section 1.
Proposition 5.2. If E = C \  satisfies Cap(E) > 0, then for any , and
any > 0,
lim

sup

0 zn \B ( )

(z, n B ( ), n ) = 0

(5.5)

uniformly in n.

Jn

1
M

M
n
n
Figure III.3 Proof of Proposition 5.2.
Proof. Suppose  and = 0. Let J = E {z : 1/M |z| M} with
M so that Cap(J ) > 0. For n as before, set Jn = {z : 1/M |z| M} \ n .
See Figure III.3. Let n be the equilibrium distribution of Jn and let n be
Robins constant for Jn . Then



z 
1
dn ( )
Un (z) =
Un (z) log
log
|z|
z 
Jn

CY535/Garnett

92

0 521 47018 8

January 27, 2005

15:41

III. Potential Theory

is harmonic on C \ (Jn {0}). For < 1/M we have




1
Un (z) log
1 , z B (0),
M
and
1
n , z C.
M
For < , these inequalities and the maximum principle give
Un (z) log

(z, n B (0), n )

Un (z) + log M + n
log( 1 M) + n

on {|z| > } n . Hence if |z| ,


log 1 + M + log M + n


(z, n B (0), n )
.
log 1 M + n

But because M is fixed and J Jn B M (0), the quantities n are bounded



above and below, and therefore (5.5) holds.
Suppose 0
/ . Then by Theorem 5.1 and the definition of g(z, ) for
finitely connected domains,
h(z, ) = g(z, ) log |z|
coincides with the solution to the Dirichlet problem in  with boundary
value log |z|. The following proposition gives a similar result when  = C .
Proposition 5.3. Let E {z : |z| < R} be a compact set such that Cap(E) > 0
and let  = C \ E. Then U E is the solution to the Dirichlet problem on
 R =  B(0, R) with boundary data

(E)
for ,
f ( ) =
for | | = R,
U E ( )
and g (, ) is the solution to the Dirichlet problem on  R with boundary
data (E) f ( ).
Proof. Because U (z) = (E) g (z, ), it is enough to show that U E is
the solution of a Dirichlet problem. We may suppose the domains {n } satisfy
{z : |z| R} 1 . Extend f to C(C) so that

on E,
(E)
on n ,
f = (E n )

on |z| = R,
U E

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Regular Points

93

and let u n be the Wiener solution of Dirichlets problem on nR = n B R (0)


for the boundary data f |nR . Then u n U En is continuous on nR and harmonic on n , and hence



U E (w) U En (w) d(z, w, nR )
u n (z) U En (z) =
|w|=R

on

nR .

Because U En U E uniformly on {|w| = R}, it follows that


lim u n (z) = lim U En (z) = U E (z)
n

R .

Therefore U E is the Wiener solution to the Dirichlet problem in


for z

 R for the boundary data f | R .
With the same proof, Proposition 5.3 is also valid if B(0, R) is replaced by
any bounded Jordan region  such that E  = .

6. Regular Points
Assume Cap(C \ ) > 0, let  and let f C(). In this section we
determine when the Wiener solution u f has boundary value f ( ) at . The
point  is a regular point for  if
lim u f (z) = f ( )

 z

(6.1)

for all f C(). Equivalently,  is a regular point if


lim (z, , ) =

 z

in the weak-star topology, where is the point mass at . A point 


not regular is called an irregular point. If  consists of finitely many Jordan
curves, then by Chapter II every boundary point is regular. However, if
 = {z : |z| < 1} \ {0},
then every bounded harmonic function on  extends to be harmonic at 0. Therefore (6.1) fails at = 0 except in the rare cases where
2
d
f (ei ) ,
f (0) =
2
0
and therefore 0 is an irregular point of .
Theorem 6.1. The point  is a regular point for  if and only if there
is > 0 such that for every > 0, there exists a function V , harmonic in

CY535/Garnett

0 521 47018 8

January 27, 2005

94

15:41

III. Potential Theory

 {z : |z | < }, satisfying
V (z) 0,

(6.2)

lim sup V (z) ,

(6.3)

V (z) > 1/2 on {z : < |z | } .

(6.4)

and
By definition a barrier is a harmonic function V on  such that
lim V (z) = 0,

and
lim inf V (z) > 0
z

for  \ { }. If V is a barrier, then there are constants C > 0 such that


the functions V = C V satisfy (6.2), (6.3), and (6.4).
Proof. Let be a regular point. Fix > 0. For 0 < < choose f C()
such that
0 f 1,
{z  : f (z) = 0} = { },
and
f (z) = 1, when |z | > .
Then clearly V (z) = u f (z) satisfies (6.2) and (6.3), and when is small enough
(6.4) also holds by Proposition 5.2.
Now suppose the functions V (z) exist. Let f C(C ) satisfy
0 f 1/2 and f 1 (0) =

(6.5)

and let n be as above. If > 0, then for sufficiently small,


f (z) V (z) +
on {z : |z | < } n . Let u n solve the Dirichlet problem on the subdomain n for the boundary value f . Then
u n (z) sup f 1/2 V (z)
on n {z : |z | = }, and by the maximum principle
u n (z) V (z) +

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Regular Points

95

on n {|z | < }. Thus


0 u f (z) V (z) +
on  {|z | < }. Because is arbitrary, (6.3) now yields
lim u f (z) = 0.

Because every real g C(C ) is a constant plus a linear combination of two



functions satisfying (6.5), it follows that is a regular point.
Theorem 6.1 implies that the regularity of is a local question; it depends
only on behavior of  in some neighborhood of .
Corollary 6.2. Let . If the connected component of  containing
consists of more than one point, then is a regular point for . In particular,
if  is simply connected, then every point of  is a regular point.
Proof. For > 0, let E be the component of  B(, /2) containing .
Then E = { }. Let be a Riemann map of C \ E onto D with () = 0.
Then | (z)| 1 as C \ E z E and




log  1(z) 


V (z)


inf log  1(z) 
<|z|<1

satisfies (6.2), (6.3), and (6.4) with = 1.

Theorem 6.3. Assume E = C \  is a compact plane set with Cap(E) > 0


and let E be the equilibrium distribution for E. Then for  the following
conditions are equivalent:
is a regular point,
lim g (z, ) = 0,

 z

U E ( ) = (E).

(6.6)
(6.7)
(6.8)

Proof. Take R > 0 such that E B(0, R). Then by Proposition 5.3, g (, )
is the solution to the Dirichlet problem in  R =  B(0, R), for the boundary
data

0
for z ,
for |z| = R.
g (z, )
If  is a regular point, then by Theorem 6.1, is also a regular point for
 R and hence (6.7) holds.

CY535/Garnett

0 521 47018 8

January 27, 2005

96

15:41

III. Potential Theory

Now suppose (6.7) holds. Then by (4.10)


U E (z) = (E) g (z, ) (E),

(6.9)

as z  E. Because log 1/|z| is superharmonic, we have



1
U E (z)d xd y,
U E ( )
r 2
B(,r )

while by Theorem 4.4,


U E (z) (E)
for all z. Therefore
0 (E) U E ( ) lim sup
r 0

1
r 2

+ lim sup
r 0

(6.10)


( (E) U E (z))d xd y
B(,r )

1
r 2

( (E) U E (z))d xd y.
B(,r )\

By (6.9) and (6.10) the first limit is 0, while by the remark after Corollary 4.3
and Theorem 4.4,
Area({z
/  : U E (z) = (E)}) = 0,
so that the second limit is also 0. Hence (6.8) holds.

To show that (6.8) implies is regular, set E = E B , 2 /4 . If and


{z : |z | = 2 /4} lie in the same component of E {z : |z | = 2 /4},

then the complement


2
in C of that component is a simply connected domain
  B , /4 , and  . By Corollary 6.2, is regular for  , and
then by Theorem 6.1, is regular
for .

Thus we may assume there is a (closed)


C 2 Jordan curve  B , 2 /4 such that is in the bounded component
W of C \ .
Let be the equilibrium distribution for E , and let U be its logarithmic
potential. We will verify conditions (6.2), (6.3), and (6.4) for the functions
V (z) = 1

U (z)
.
(E )

and this will show that is regular. First note that e (E ) = Cap(E ) > 0,
because if not, then by Corollary 4.3, E (B(, 2 /4)) = 0. But then would
be in the unbounded component of V = C \ supp( E ), where supp( E ) is
the closed support of E . Then (4.10) and the maximum principle would give
U E ( ) < (E), contrary to the assumption (6.8).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. Wiener Series

97

By Theorem 4.4, (6.2) holds for V . Note that if |z | > , then



1
2
log
U (z) =
d E (w) log .
|z

w|

E
Because (E ) log 42 , we have V (z) > 1/2 in |z | and (6.4) holds
for V .
Choose M > 0 so that when z ,

(6.11)
(E ) U (z) M (E) U E (z) .
By Proposition 5.3, each side of (6.11) is the solution to a Dirichlet problem
in W  with boundary data 0 on  W , and hence the inequality (6.11)
continues to hold in W . Sending z then yields (6.3). Then is a

regular point by Theorem 6.1.
Corollary 6.4 (Kelloggs theorem). Let  be any domain in C . Then

Cap {  : is irregular} = 0.

Proof. We may assume


/ . Then the corollary is immediate from Theorem

4.4 and Theorem 6.3.
Corollary 6.5. Let  be any domain in C with Cap(C \ ) > 0. Then
(z, {  : is irregular}, ) = 0
for all z .
Proof. By (6.8), the set of irregular points is a Borel set, so the corollary follows

from (5.4) and Corollary 6.4.

7. Wiener Series
In a second paper, [1924b], gave a geometric condition necessary and sufficient
for  to be a regular point for . Let An = An ( ) be the ring
An = {z : 2n |z | 2n+1 },
and recall the notation



(E) = log 1/Cap(E) .

Theorem 7.1 (Wiener). Let  be a domain in C and let . Then the


following are equivalent:

CY535/Garnett

0 521 47018 8

98

January 27, 2005

15:41

III. Potential Theory

is a regular point,

n=1

(An \ )


n=1

= ,

n
= .
(An )

(7.1)
(7.2)

(7.3)

Proof. We can assume that = 0, and  B(0, 1/2). It is clear that (7.3) implies (7.2) because the function E (E) is decreasing. Now assume (7.2).
By the definitions of regular point and capacity, we can replace  by a subdomain so that, for each n, An \  is bounded by finitely many C 2 Jordan curves
and so that the series (7.2) still diverges. Write n = (An \ ), let n be the
equilibrium distribution for An \ , and let u n (z) = Un (z). Fix > 0. Then
because n n log 2 there are integers N1 < N2 < . . . Nk such that
N j+1 1

 n log 2
1 + .
n

(7.4)

Nj

Set
N j+1 1

Wj =

 un
.
n
Nj

By our smoothness assumption on An \ , W j is continuous on B(0, 1), and


W j is harmonic and positive on  B(0, 1). Also, by (7.4)
N j+1 1

N j+1 1
 u n (0)
 n log 2

1.
W j (0) =
n
n
Nj

(7.5)

Nj

On the other hand, if z Ak , k 1, then


1,
if |n k| 1

u n (z) (k+1) log 2


, if n > k + 1

(n+1) log 2
, if n < k 1.
n
Hence by (7.4),
W j (z) 3 +

Nj + 1
(1 + ) 1 + 5
Nj

(7.6)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. Wiener Series

99

on B(0, 1) provided N j is sufficiently large. Also, for any > 0,


sup W j (z)

(7.7)

|z|

if N j is sufficiently large. Set Vj = 1 + 5 W j . Then by (7.5), (7.6), and (7.7)


the functions Vj satisfy the hypotheses of Theorem 6.1 (with 5), and = 0 is
a regular point of .
Now assume = 0 is a regular point of  and assume the series (7.3)
converges. Again we may replace  by a subdomain so that, for each n, An 
consists of finitely many C 2 Jordan curves and so that (7.3) still converges.
We may also assume  B(0, 1/2). Write n = (An ), let n be the
equilibrium distribution for An , and let u n = Un be its potential.
Fix , 0 < < 13 and take N 3 so that


n
< .
n

(7.8)

n=N

Let F(z) C() be harmonic on  with 0 F 1, F(0) = 1, and F = 0 on


 \ B(0, 2N +1 ). Since = 0 is a regular point there exists s < 2N such
that F(z) > 1 on  B(0, s). Let 0 < r < s be so small that if Ur (z) is
the equilibrium potential for  B(0, r ), then
Vr (z) =

Ur (z)
( B(0, r ))

on {|z| s}. Let A be the annulus {r |z| 2N +1 }. Let U denote the equilibrium potential for  A and take
V (z) =

U (z)
.
( A)

Then V + Vr F by Proposition 5.3. Consequently V 1 2 on


{|z| = s} B(0, s)  A .
Hence by the maximum principle,
V (0) 1 2.
On the other hand, V (z)
principle also gives

2
N

on |z| =

V (z)

M

u n (z)
N

Hence by (7.8) V (0) +

2
N,

(7.9)

1
M
2 , so that when 2

< r the maximum

2
.
N

which contradicts (7.9).

CY535/Garnett

0 521 47018 8

January 27, 2005

100

15:41

III. Potential Theory

Theorem 7.1 has some important applications. The first shows there exist
totally disconnected sets K such that each K is a regular boundary point
for  = C \ K .

Example 7.2. The Cantor set K =
n=1 K n of Example 4.7 has Cap(K ) > 0.
Suppose K and suppose K A j ( ). Then belongs to an interval
I K j of length 3 j and there is a second interval J K j of length 3 j with
dist(I, J ) = 3 j . See Figure III.4. If j is such that 2 j1 > 3 j+1 then one of
the three rings A j1 ( ), A j ( ), or A j+1 ( ) contains either I or J . Hence for
some n { j 1, j, j + 1}
Cap(An K ) Cap(I K ) = Cap(J K ) = 3 j Cap(K ).
Therefore
n
C >0
(An K )
for some C > 0 and for infinitely many n and by (7.3) every is a regular point
for  = C \ K .

A j+1

Aj

A j1

2 j

I
2( j1)

2( j2)

Figure III.4 Annuli and the Cantor set.


See Corollary 7.4 and Corollary D.3 of Appendix D for other proofs that
every point of the Cantor set K is regular.
Corollary 7.3 (Beurling). Let E be a compact set. Assume 0 E and set
E = {|z| : z E}. If

dr
= ,
(7.9)

E r
then 0 is a regular point for  = C \ E.
Proof. Set E n = E An and E n = {|z| : z E n }. Then Cap(E n ) |E n |/4
 n
by Theorem 4.5, and consequently we have
2 Cap(E n ) = by (7.9).
But 2n Cap(E n ) Cn/ (E n ), because Cap(E n ) 2n+1 and the function

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. Wiener Series

101

x log 1/x is increasing on [0, 1/e]. Therefore the series (7.3) also diverges
and 0 is a regular point.

Corollary 7.4. Assume E = . If

Cap E B(, r )
lim sup
> 0,
r
r 0

(7.11)

or if
lim sup
r 0

log(1/r )
> 0,
(E B(, r ))

(7.12)

then is a regular point of .


Proof. It is clear that (7.11) implies (7.12) and that (7.12) implies
n
> 0.
lim sup
n
n (E B(, 2 ))
Then by Lemma 7.5 below
lim sup
n


n

k
> 0,
(E Ak )


and (7.3) again diverges.

Lemma 7.5. Suppose E 1 , . . . , E n are disjoint Borel subsets of B(0, 1/2). Then
(

1


Ek )

n

k=1

1
.
(E k )

Proof. By the definition of capacity, we can suppose each E k is bounded by


finitely many pairwise disjoint C 2 Jordan curves. Let be the equilibrium

distribution for E = E k , let k be the equilibrium distribution for E k , let

= k / (E k ) and let U , Uk , and U be their potentials. Then because
U = (E) on E and supp( ) E, we have




1
(E) = U d = U d,
(E k )
and because Uk 0 on E and Uk = (E k ) on E k , we have U 1 on
supp(). Consequently

U d 1
and the lemma follows.

CY535/Garnett

0 521 47018 8

January 27, 2005

102

15:41

III. Potential Theory

8. Polar Sets and Sets of Harmonic Measure Zero


In this section we prove a strong form of the maximum principle. It implies that
if a compact set E is a polar set, i.e. Cap(E) = 0, and if u > 0 is a harmonic
function on C \ E, then u is constant.
Theorem 8.1. Suppose  is a domain such that Cap(C \ ) > 0, and suppose
A is a Borel subset of  such that
(z 0 , A, ) = 0,

(8.1)

holds for some z 0 . Let u(z) be a subharmonic function in . If u is bounded


above in  and if
lim sup u(z) m

(8.2)

for all  \ A, then


u(z) m
for all z .
The condition (8.1) is essential for Theorem 8.1. For example, if K is
a compact subset of  such that 0 < (z 0 , K ) < 1 for some z 0 , let
u(z) = (z, K , ). If A = ( \ K ) {regular points} then (z 0 , A, ) > 0
by Corollary 6.5. But lim z u(z) = 0 on A, while u(z 0 ) > 0.
Because a finite set has zero harmonic measure, Theorem 8.1 is a generalization of Lindelfs maximum principle, Lemma I.1.1.
Proof. On , define
V ( ) =

u( ),

,

lim sup u(z),

.

 z

Then by hypothesis, V is bounded above, and V is upper semicontinuous on 


because subharmonic functions are upper semicontinuous. By (8.1) and (8.2),

V ( )d(z, , ) m
(8.3)


for every z .
Let {n } be the subdomains from before. Then for any z ,
(z, , ) = lim (z, , n ),
n

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

8. Polar Sets and Sets of Harmonic Measure Zero

103

with respect to weak-star convergence of measures on . Let Vk ( ) be a sequence of continuous functions that decreases to V ( ) on , and let u k,n (z)
be the harmonic extension of Vk to n . Because u is upper semicontinuous on
n ,

lim sup u(z) u k,n (z) = u( ) Vk ( ) 0


n z

at every n . Hence by the ordinary maximum principle,



u(z) u k,n (z) =
Vk ( )d(z, , n )
n

for z n . Then by the weak-star convergence



Vk ( )d(z, , ),
u(z)

(8.4)

so that by monotone convergence



u(z)

(8.5)

V ( )d(z, , ),

z . Then (8.3) and (8.5) imply that u(z) m in  =

n .

Theorem 8.2. Suppose A is a Borel set. Then A is a polar set if and only if
(z, A , ) = 0

(8.6)

for every domain  such that Cap(C \ ) > 0 and for every z .
Proof. By (4.8) and the regularity of harmonic measure, we may assume
that the set A is compact. If Cap(A) = 0, then (8.6) follows from (5.4). Conversely, if Cap(A) > 0 let A be the equilibrium distribution for A and let 
be the unbounded component of C \ A. Then Cap(C \ ) = Cap(A) > 0 and

(, A , ) = A (A) = 1.
More quantitative forms of Theorem 8.2 will be given in the next section.
Because of Theorem 8.2, Nevanlinna [1953] called a polar set a set of absolute
harmonic measure 0.
Corollary 8.3. Suppose  is a domain such that Cap(C \ ) > 0, and suppose u is subharmonic and bounded above in . If
lim sup u(z) m
z

for all  \ A, where A is a Borel set and Cap(A) = 0, then


sup u(z) m.


CY535/Garnett

0 521 47018 8

January 27, 2005

104

15:41

III. Potential Theory

Corollary 8.3 is immediate from Theorems 8.1 and 8.2.


Corollary 8.4. Suppose E is a compact polar set. If u(z) is subharmonic and
bounded above in C \ E, then u is constant. If  E is a Jordan domain
then every bounded harmonic function on  E extends to be harmonic on E.
Conversely, if Cap(E) > 0, if  is the unbounded component of C \ E, and if
A is a compact subset of E such that 0 < E (A) < 1, then u(z) = (z, A, )
is a non-constant bounded harmonic function on .
If E is a compact polar set then by Corollary 8.4 every positive harmonic
function in C \ E is constant and every bounded harmonic function in C \ E
is constant.
Proof. Let z 0 C \ E, < dist(z 0 , E), and  = C \ (E B(z 0 , )). Then
Cap(C \ ) > 0 and by Corollary 8.3,
sup u(z) sup u(z).


B(z 0 ,)

Letting 0 we obtain
sup u(z) u(z 0 ).

C \E

Because z 0 is arbitrary in C \ E, u is constant. To prove the second statement,


we shrink  so that u is continous on . Let v be the solution to the Dirichlet
problem on  for the function u. Then by Corollary 8.3, u = v and v is the

harmonic extension of u to . The converse assertion is clear.

9. Estimates for Harmonic Measure


We use capacity and potentials to estimate (z, E, D \ E) where E is a closed
subset of the unit disc D. By (5.3) we may assume E D is compact and consists of finitely many analytic curves. First consider the case when dist(E, D)
is large.
Theorem 9.1. Suppose E {z : 0 < < |z| < r < 1} and set
(E) = log (Cap(E)) .
Then there are constants c1 (r ) and c2 (, r ) such that
c1 (r )
c2 (, r )
(0, E, D \ E)
.
(E)
(E)

(9.1)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

9. Estimates for Harmonic Measure

Proof. Let

105



 1 z 

g(z, ) = log 
z 

be Greens function on D and let E be the equilibrium distribution for E. Then


the Green potential

g(z, )d E ( )
V (z) =
E

is continuous on D \ E and harmonic on D \ E. Also, log r1 V (0) log 1 ,


V = 0 on D, and for z E,
(E) + log(1 r 2 ) V (z) (E) + log(1 + r 2 )
by (2.3). Hence

log r1

(E) + log(1 + r 2 )

(0, E, D \ E)

log

(E) + log(1 r 2 )

(9.2)

provided that (E) + log(1 r 2 ) > 0. This gives the left-hand inequality
in (9.1) with c1 (r ) = (log r )2 / log(r + 1/r ), because (E) log(1/r ). If
(E) 2 log 1/(1 r 2 ) the upper bound in (9.2) gives the upper bound in (9.1)
with constant 2 log(1/), and if (E) 2 log 1/(1 r 2 ), the upper bound in
(9.1) follows from the trivial estimate
(0, E, D \ E) 1

1
2 log 1r
2

(E)

The next result is called the Beurling projection theorem.


Theorem 9.2 (Beurling). If E D \ {0}, and E = {|z| : z E} is the circular projection of E, then
(z, E, D \ E) (|z|, E , D \ E ).
Proof. By Greens theorem
(z, E, D \ E) = (z) =

1
2

g(z, )

( )
ds
n
(9.3)

g(z, )d ( ),

CY535/Garnett

0 521 47018 8

January 27, 2005

106

15:41

III. Potential Theory

in which the normal n points out of D \ E, so that 0. Consider the circular


projection of ,
(A) = ({z E : |z| A}),
and its Green potential


V (z) =

g(z, | |)d .

( E)

z
E
E

|z|
0

E
E

E
E

Figure III.5 Circular Projection in D.


Because
g(|z|, | |) g(z, ) g(|z|, | |),
we have
V (|z|) (z, E, D \ E) V (|z|).
Let z n z E. By the right-hand inequality,
lim inf V (|z n |) 1,
so that by a comparison, (z, E , D \ E ) V (z). With the left-hand inequality that gives
(|z|, E , D \ E ) V (|z|) (z),
which proves the theorem.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

9. Estimates for Harmonic Measure

107

Corollary 9.3. Let  be a simply connected domain, let z  and let .


Then for r sufficiently small,
1

(z, B(, r ) , ) Cr 2 .
Proof. We may assume = 0 and z = 1. We project  \ B(0, r ) to the
positive real axis and apply Beurlings theorem in C \ B(0, r ) to get
(z, B(0, r ) , ) (1, B(0, r ), C \ (B(0, r ) [r, )))



2
4
i r
= arg
= tan1 r .

i+ r

B(0, r )

1
Figure III.6 The proof of Corollary 9.3.
The case when dist(E, D) is small leads to the connection between harmonic measure and Carleson measures. We work in the upper half-plane H and
with dyadic Carleson boxes of the form
Q = { j2k x ( j + 1)2k ; 0 < y 2k },
and write (Q) = 2k . Fix z 0 = i and for convenience assume E is compact
and


E 0 x 1, 0 < y 1/2 .
(9.4)
Let A = A E be the set of positive measures on E such that
(i) is a Carleson measure with constant 1,
(Q) (Q),

(9.5)

for every dyadic Carleson box Q (see Exercise I.19), and


(ii) the top half T (Q) = Q {y (Q)/2} of every dyadic Carleson box
Q satisfies

CY535/Garnett

0 521 47018 8

January 27, 2005

108

15:41

III. Potential Theory


(T (Q))
1

(Q)
1 + (E T (Q)) (T (Q))
(9.6)
=

1 + log

1
Cap T (Q)
Cap ET (Q)

.

Theorem 9.4. Assume E satisfies (9.4). Then there are constants C1 and C2
such that
C1 (i, E, H \ E) sup{ (E) : A E } C2 (i, E, H \ E).

(9.7)

In the special case z = 0 and E {| | > 1/2}, Beurlings projection theorem (with a larger constant) is a consequence of Theorem 9.4. Suppose
0 (A) = (0, A E , D \ E ) and let be any positive measure on E that
projects circularly onto 0 . Then an elementary comparison gives
(Q) 0 (Q ) C(Q )
for some constant C, so that C 1 satisfies (9.5). To check (9.6) for C notice
that by (4.12) the right side of (9.6) becomes smaller when E is replaced by
E and then apply the right-hand inequality in (9.2) with = 1 (Q) and
r = (1 + )/2.
When E H we still write E = {|z| : z E} for the circular projection of
E, but for z = x + i y H we define z = |x| + i y. The estimate
(z, E, H \ E)

2
(z , E , H)
3

from Hall [1937] is known as Halls lemma. With a different constant, the
most important special case, E {arg z < /4} and z = z of Halls lemma
also follows easily Theorem 9.4. See Exercise 20.
Proof of Theorem 9.4. Write = + i E. Greens function for the halfplane H is


z 


g(z, ) = log 
,
z 
and g(z, ) satisfies
g(z, ) = log

+ O(1),
|z |

if

|z |
|z |

c < 1,

(9.8)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

9. Estimates for Harmonic Measure

and



 z 2
4y


g(z, ) 
,
 1=
z 
|z |2

if |z | >

109

.
4

(9.9)

The Green potential of a measure on H is



g(z, )d( ).
G (z) =
H

Let Q denote the (logarithmic) equilibrium distribution of E T (Q). Then


for z E T (Q) we have

G Q (z) = g(z, )d Q ( ) 1 + (E T (Q)) (T (Q))
by (9.8). Thus if we define
Q =

Q
,
1 + (E T (Q)) (T (Q))

then for some constant C


C

ET (Q)

g(z, )d Q C

(9.10)

on E T (Q).
To prove the left half of (9.7) we may assume E {y 2N }. We inductively define a measure = N on E as follows: Set 0 = 0. Assume that k is
defined and has support in E { : 2N +k }. Let
Q Dk = {Q dyadic : (Q) = 2N +k+1 and Cap(E T (Q)) > 0}.
If inf ET (Q) G k < 1, take Q > 0 such that
inf (G k (z) + Q G Q (z)) = 1.

ET (Q)

(9.11)

If inf ET (Q) G k 1, take Q = 0. Then define



k+1 = k +
Q Q .
Dk+1

Finally take = N , and = . Then by construction G 1 on E and


G = 0 on R. But by (9.4) and (9.9),

G (i) C
d( ) = C (E).
E

Therefore by the maximum principle


(i, E, H \ E) G (i) C (E),

CY535/Garnett

0 521 47018 8

January 27, 2005

110

15:41

III. Potential Theory

and if we can show that c A E for some constant c, then the left-hand inequality in (9.7) will hold with C1 = (cC)1 .
To verify (9.5) for c let Q be a dyadic box of side 2N +k+1 and first assume
Q > 0. Then by (9.11) there exists z E T (Q) such that
1 = G k (z) + Q G Q (z)

g(z, )
Q

(T (Q))
C
+C
(Q)


Q\T (Q)

y
d
|z |2

(Q)
,
(Q)

and therefore (9.5) holds for Q. But if Q = 0 then (T (Q)) = 0 and



(Q) =
{ (Q  ) : Q  Q, Q  > 0, Q  maximal},
so that (9.5) also holds for Q.
To verify (9.6) for the dyadic box Q observe that
(T (Q))
(T (Q))
(Q)

Q
1 + (E T (Q)) (T (Q))

C
by (9.10) and (9.11).
To prove the right-hand part of (9.7) let A E and set


 =
(T (Q)) Q and = .

Q dyadic

Then G is harmonic on H \ E and


G = 0
on R since E is compact. By (9.4) and (9.9)

G (i)
d = (E).
E

(9.12)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

9. Estimates for Harmonic Measure

111

Now let z E and fix a dyadic box Q with z T (Q). Then by (9.5) and (9.9)


d  ( )
g(z, )d( ) C y
2
| z|(Q)/4
| z|(Q)/4 |z |

Cy


 ({| z| 2n2 (Q)})
22n (Q)2
n=0

C,
and by (9.6) and (9.8)


| z|<(Q)/4

g(z, )d( ) C

for some constant C. Therefore G (z) C on E and by (9.12)


(i, E, H \ E) C 1 G (i) C (E).


That proves the right-hand inequality in (9.7).

Every theorem in this section was proved using Green potentials and Green
potentials have a theory parallel to the potential theory in Sections 2, 3, and 4.
The Green potential of a signed measure supported on E D is the function

g(z, )d ( ).
G (z) =
E

Assume E is compact and Cap(E) > 0. Then by (9.3) and a limiting process
there is a positive measure on E such that
(z, E, D \ E) = G (z).

(9.13)

Moreover, there exists unique P(E) having minimal Green energy:



!

G d = inf
G d : P(E) G (E).
The minimum energy G (E) is called the Green capacity of E, and except for
a set of capacity zero, G = G (E) on E. Then (9.13) becomes
(z, E, D \ E) =

1
G (z).
G (E)

See Exercise 18 for the proofs of these results and for an application.

CY535/Garnett

0 521 47018 8

112

January 27, 2005

15:41

III. Potential Theory

Notes
Sections 1 through 6 follow a UCLA lecture course given by Lennart Carleson.
Most of Section 7 is in Wieners original paper [1924b]. Actually, Wiener did not
prove Theorem 7.1 in [1924b]. He proved the Rd , d 3, version of Theorem
7.1, but for C = R2 he proved  is regular if and only if


n=1

2n
= ,
(Jn \ )

(N.1)

where
Jn =

%

 
Ak : 2n1 < k 2n = z : 2n1 log2


1
2n .
|z |

In Rd Wieners proof hinged on an elegant estimate of Green potentials (see


[1924b], p. 133). In R2 the analogous estimate holds for the big rings Jn , but it is
1
decays more slowly in R2 than |X1 | does in
false for the rings An because log |z|
d
R . The first proof of Theorem 7.1 was published by de la Valle Poussin [1949].
See also Tsuji [1959]. Our proof follows Tsuji [1959] when the series diverges
and copies Wiener [1924b] when it converges. See Exercise 9 for Wieners
proof of (N.1) and for the equivalence of (N.1) and (7.2). Corollary 7.3 is from
Beurlings thesis [1933].
Theorem 9.4 is from Carleson [1982]. Potential theory is a vast subject and
we have only touched on the two-dimensional theory. For the broader picture
we refer to the books of Brelot [1965], Carleson [1967a], Landkof [1972],
Ransford [1995], Tsuji [1959], and Wermer [1974]. See Doob [1984] for the
connections with probability, and see the books of Heinonen, Kilpelinen, and
Martio [1993] and Adams and Hedberg [1996] for nonlinear potential theory.
Appendix E gives a geometric description of capacity by transfinite diameter
and constructs Evans functions for sets of capacity zero.

Exercises and Further Results


1. Suppose p(z) = ak z k + . . . + a0 is a polynomial. Set  = {| p| > } and
= (, ., ). Then 

 p(z) 
1
.

(i) g (z, ) = log 
k

a 
1
 k
(ii) () = log  .
k

1 p  (z) dz
g ds
=
.
(iii) d =
n 2
k p(z) 2i

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

113

(iv) The harmonic measure at of a component J of  is the number of


zeros of p in the bounded component of C \ J , divided by k.
(v) Formulate and prove similar results for rational functions.
2. Let  be a finitely connected Jordan domain such that  and every
component of  is a C 2 curve. Let G(z) = g (z, ) and let {z j } be the
(finite) set of critical points of G in . Prove

 G 

G
1


G(z j ),
log 
ds = () +
2  n
n
where () is Robins constant for . This identity from Ahlfors [1947]
will play a critical role in Chapter IX. Hint: Use Greens theorem and the
Taylor expansion of G x i G y near each critical point z j .
3. Let  be a bounded plane domain and let v(z) be subharmonic in a neighborhood of . Then there is a unique positive finite measure supported on
 and a unique function u(z) harmonic on  such that
v(z) = u(z) U (z)
take d = v
on . Hint: If v
2 d xd y. In the general case, replace
v by v K as in the proof of Theorem 3.1 and take a limit. This result is
known as the F. Riesz decomposition theorem for subharmonic functions.
4. (a) Let g be a compactly supported real valued C 2 function on the plane.
Then g = 1 2 where 1 and 2 are C 2 and subharmonic on an open set
containing {z : g(z) = 0}.
(b) Let  be a bounded plane domain and let n n1 . . . be an increasing

sequence of finitely connected subdomains of  such that  = n n and
n consists of C 2 Jordan curves. Fix z 1 and consider the sequence
n of probability measures on  defined by n (E) = (z, E n , n ).
Show that if is continuous on  and subharmonic on , then


dn dn+1 .
C 2 (),

Using this result, part (a), and the boundedness of {n }, prove that for all
f C(),

lim
f dn exists.
n

It follows that for bounded  the Wiener definition of harmonic measure


does not depend on the approximating sequence n .
(c) In Theorem 4.1, if = lim n = , then for any P(E) the sequence
{E n } can be chosen so that is the weak-star limit of the sequence {n }.

CY535/Garnett

0 521 47018 8

114

January 27, 2005

15:41

III. Potential Theory

5. A property is said to hold p.p. (for presque partout) on a set E if the subset
of E where the property fails has capacity zero. Suppose E is compact and
(E) > 0. Let M(E) = { : is a signed measure supported on E}. Then
1
= inf{(E) : M(E) and U 1 p.p. on E}
(E)
= sup{(E) : M(E) and U 1 p.p. on E}.
6. Let  = D\ {0} and let f C(). Prove lim z0 u f (z) = f (0) if and only
d
.
if f (0) = D f (ei ) 2
7. Let
p(z) = z d + ad1 z d1 + . . . + a0
be a monic polynomial of degree d 2 and let p (n) be the n-th iterate
p (n) (z) = p p . . . p(z)
and write

1

p (n) = p (n) .

The Julia set of p(z) is




E = z : p (n) is not a normal family in any neighborhood of z .
Here normal includes convergence to the constant f = . Thus, for example, if p(z) = z 2 then E is the unit circle {|z| = 1}. The Fatou set of p(z)
is the complement C \ E. The basin of attraction of for p(z) is


 = z : lim p (n) (z) =
n

and 
\ E. It is known that the Julia set E is a non-empty, compact, perfect set, that E is completely invariant, E = p(E) = p 1 (E),
and that E = . See Chapter III of Carleson and Gamelin [1993]. This exercise gives Brolins beautiful [1965] description of the harmonic measure
(, , ).
(a) Let U0 be a disc such that U0 p(U0 ), so that C \ U0 . The disc
U0 = {|z| < 1 +

d1


|a j |}

j=0

is an example. If Un = p (n) (U0 ) and E n = Un then p(E n+1 ) = E n ,



E n = E. Let gn (z) denote Greens function for
p 1 (E n ) = E n+1 , and
C \ E n with pole at . Then
gn (z) = log |z| + n + o(1),

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

115

where Cap(E n ) = en . Show that


gn ( p(z))
,
d
by uniqueness. Then conclude that n+1 = n /d and Cap(E) = 1.
(b) Let E be the equilibrium distribution for E. It is known that if V is open
and V E = , then for some n, p (n) (V E) = E. Again see Carleson and
Gamelin [1993]. Use this fact and Theorem 4.5 to show that E has closed
support E.

(c) If V is a neighborhood of z E then by Montels theorem n p (n) (V )
can omit at most two values 1 and 2 , and 1 and 2 do not depend on
/ {1 , 2 } and take the point mass 0 = z 0 .
the point z. Fix any point z 0
Write p (n) (z 0 ) = {z 1 , . . . , z d n }, counting points with their multiplicities,
and form the discrete measure
1 
z j .
n = n
d
gn+1 (z) =

Show n ( p(A)) = n+1 (A) for any Borel set A. Show n converges weakstar to E . Hint: Use the identity
,
|z z j |
| p (n) (z) z 0 | =
j

to estimate Un , and use Lemma 4.2 to show that any weak-star convergent
subsequence of {n } has limit E .
(d) If A E, then E ( p 1 (A)) = E (A). In other words, p is a measure
preserving transformation on the measure space (E, E ). Moreover, p is
strongly mixing: If f, g L 2 (E, E ), then



f d E gd E .
f ( p (n) (z))g(z)d E (z) =
lim
n

8. (a) Prove is a regular point for  if and only if lim z g(z, z 0 ) = 0. Give
a proof with Wiener series and give a proof without Wiener series.
(b) Prove is a regular point for  if and only if is a regular point for
 B(, r ), r > 0. Give a proof with Wiener series and give a proof without
Wiener series.
(c) Let  j = C \ E j , where E 1 and E 2 are disjoint compact sets and let
 = 1 2 . Prove  is regular for  if and only if is regular for
1 or is regular for 2 .
9. Let . (a) Let < 1 and set G n = {z : n |z | n1 }. Prove

CY535/Garnett

0 521 47018 8

January 27, 2005

116

15:41

III. Potential Theory

is a regular point if and only if


n=1

n
= ,
(G n \ )

and if and only if


n=1

n
= .
(G n )


(b) Lemma 7.5 has a converse. Let Jn = {Ak : 2n1 < k 2k }. There is
a universal constant C < so that if (Jn \ ) > C2n and if E k Ak for
2n1 < k 2n then
n

2

2n1 +1

1
4

.
(E k )
(E k )


Hint: If = k / (E k ) where k is the
 equilibrium distribution for E k ,

then |||| = 1/ (E k ). By estimating Uk dm as in the proof of Theorem 7.1, show that I () 3|||| + C1 2n ||||2 . Now apply Theorem 4.1
and the hypothesis on (Jn \ ) above.
(c) Set Br = {z : |z | r }. Prove that the following conditions are equivalent to the regularity of :
(i)

(ii)

(iii)
0

(v)

1
(B2s \ )

(iv)

2n
=
(Jn \ )




ds = .

dr
1
= .
(Br \ ) r
1
= .
(B2n \ )
n
(An \ )

= .

Hint: Note that the integrand in (ii) is decreasing. Bound the left side of each
displayed formula by a multiple of the left side of the subsequent formula.
If the first sum is infinite, then apply Theorem 7.1 to conclude is regular,
and if the first sum is finite, use part (b) to show the sum in (v) is finite.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

117

(d) Wieners paper [1924b] proved that (i) implies is a regular point. Here
is his argument. Suppose = 0 and suppose
2n

2
n=1

2n

= ,

(E.1)

where n = (Jn \ ). Set rn = 22 . Then Jn = {z : rn |z| rn1 } and


Jn Bn1 = {z : |z| rn1 }. Define



 sn z w/s
n  dn (w)
log
,
Wn (z) =
z w  n
Jn \n
n

where n is the equilibrium distribution for Jn \ n . The function Wn is a


Green potential because the kernel in its definition is Greens function for
the disc Bn . Then
Wn (z) = 0

on |z| = sn ,



2 

rn1
1


Wn (z) 1 +
log sn +
1

n
sn 
for n 2, and the critical


 sn rn rn1 /sn  C2n
1


Wn (z)
log
n
rn + rn1 
n

(E.2)
on Jn ,

(E.3)

on |z| rn .

(E.4)

Let Vn,m be the Wiener solution to the Dirichlet problem on the domain

B0 \ ( m
p=0 J2n+2 p \ ) for the boundary data


0,
if z m
p=0 J2n+2 p \ ,
Vn,m (z) =
1,
if |z| = s0 .
We show by induction on m that

m 
,
C22n+2 p
Vn,m
1
2n+2 p

on {z : |z| < r2n+2m+1 }.

(E.5)

p=0


2
To prove (E.5), we may assume m
p=0 J2n+2 p \  is bounded by disjoint C
Jordan curves. Then by the maximum principle, (E.2), (E.3), and (E.4),
Vn,0 1 W2n 1

C22n
.
2n

on B2n+1 . If (E.5) holds for m = k 1, then by the maximum principle




 k1
,
C22n+2 p
Vn,k (z) 1 W2n+2k (z)
1
2n+2 p
p=0

CY535/Garnett

0 521 47018 8

January 27, 2005

118

15:41

III. Potential Theory

on B2n+2k+1 , so that by (E.4), we obtain (E.5) when m = k. Now by (E.1)


and (E.5)
lim Vn,m (0) = 0.

For |z| = s2n and all m,


Vn,m (z) (z, B0 , B0 \ Dn ) n ,
where Dn = {z : |z| r2n1 } and it follows via Theorem 6.1 that 0 is a
 22n+1
regular point. A similar argument works if
2n+1 = .

10. (a) Let E { 21 |z| 1} be a continuum with E {z : |z| = 21 } = and


E {|z| = 1} = . Then


1
2
(0, E, D \ E) tan1 .

This follows from Beurlings projection theorem, but it also has an elementary proof, albeit with a smaller constant: We can assume 21 E. Set
E = {z : z E}. Since E is a continuum, the component U of D \ (E E)
satisfying 0 U has [ 21 , 1] U = and U D E E.

E
1
2

Figure III.7 Lower bound via reflection.


Then by the maximum principle,
(z, E, D \ E) + (z, E, D \ E) (z, (E E) U, U )
1
1

z, [ , 1], D \ [ , 1] .
2
2
Hence



1
1
2
1
1
2(0, E, D \ E) (0, [ , 1], D \ [ , 1]) = tan
.
2
2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

119

See Milloux [1924] and Garnett [1986].


(b) Let : D  be conformal, let I be an arc of D with length
|I | < 1 and center I and let z I = (1 |I |) I . Let denote the harmonic
measure on  for the
point (0). Show that if C > 1, then

B((z I ), Cdist((z I ), )) c|I |, where c is a constant depending
only on C.
11. Let K be a compact set and set  = C \ K . The set K is called uniformly
perfect if there is a constant A such that
Cap(K B(z, )) > A

(E.6)

for all z K and all < diam(K ).


(a) Condition (E.6) holds if and only if there is a constant > 0 such that
(z, B(z, 2d(z)), )
for all z , where d(z) = dist(z, K ). Exercise IX.3 will give several other
conditions equivalent to (E.6).
(b) Assume K satisfies (E.6), let U be an open subset of  and let u(z)
be continuous on  and harmonic on . If u = 0 on U and if J U is
compact, then
u(z) C(dist(z, J ))
for constants C = C(J ) and = (A) satisfying 0 < 1. See Jerison
and Kenig [1982a], Lemma 4.1. In particular, if K is uniformly perfect and
g(z) = g (z, ), then on ,
g(z) C(dist(z, K ))

(E.7)

and
|g(z 1 ) g(z 2 )| C  |z 1 z 2 | .
(c) However, (E.7) does not imply (E.6). If
K = {0}

{|z| = 2n , |argz| n }
2

n=1

for small n , then (E.7) holds but (E.6) fails.


12. ksendal [1972] showed that for any plane domain, harmonic measure is
singular to area measure (although  may have positive area). The following proof is from a lecture by Carleson: Let K be a compact subset of 
with Area(K ) > 0 and suppose is a point of area density of K :
Area(K B(, ))
>1
2

CY535/Garnett

0 521 47018 8

January 27, 2005

120

15:41

III. Potential Theory

for small . Then there is

< r () <

3
4

so that

|{|z | = r ()} \ K |
< 6,
2r ()
and by a comparison,

(z, B(, ) , ) < 42


2
on  {|z | = }. Taking 42 < 1/4 yields
22n (z, B(, 2n ) , ) 0,
and by the RadonNikodym theorem is singular to area measure.
13. Let  be a domain such that Cap() > 0 and let E  be a Borel set
such that (z, E, ) > 0 for some z . Prove
sup (z, E, ) = 1.


Find a proof without using the Perron solution.


14. Let  be a simply connected domain and let E R  have |E| = 0.
Prove (z, E, ) = 0 for all z . Hint: The result is easy if  H or
if  = {z : z }. In general, we can assume 0 =  H = . Then 0

is simply connected. Write R  = Jj where Jj is an open interval and
let  j be the component of  {z : z } with Jj  j . Then  j is also
simply connected. Now if z 0 0 , then

(z 0 , E, ) =
(, E, )d0 (z 0 , )
j

Jj


j

Jj

 j \

(z, E, )d j (, z)d0 (z 0 , ),

but by symmetry  j (,  j \ ,  j ) = 1/2 for Jj . The result now


follows from Exercise 13. See ksendal [1980].
15. (a) Let 0 , let  satisfy | | > 1, and let 0 < r < 21 . Then
(0,  B(, r ), ) C

sup
|z |< 3r
2

g (z, 0)

with constant C independent of . Hint: Assume  is a finitely connected


Jordan domain with smooth boundary. Take C (B(, 3r2 )) such that

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

121

0 1, (z) = 1 on B(, r ), and | | rC2 . Write



g(0, w)
1

ds(w)
(0,  B(, r ), )
2 
nw
and apply Greens theorem. See Jerison and Kenig [1982a].
(b) Let  be a simply connected domain containing the disc D, let z 1 
satisfy |z 1 | > 1, and set g(z) = g (z, z 1 ). Let  be the hyperbolic geodesic
of  joining 0 to z 1 and for 0 t 1 define
(t) =
Then

sup

{|z|=t}

dist(z, ).

1 1 dt 
.
g(0) C0 exp
2 0 (t)

Hint: Let (z, w) be the hyperbolic distance on  and for 0 t 1 pick


z t  with |z t | = t. Then g(z) e2(z,z 1 ) and by Koebes estimate
1

ds
dt

.
(0, z 1 )
4
dist(z,
)
4(t)

0
Carleson and Jones [1992].
(c) Let  be a simply connected domain containing the disc{|z| < 21 }, let
 satisfy | | > 1, and let
(t) = sup{dist(z, ) : z , |z | = t}.
Then for r < 14 ,

1 1 dt 
.
(0,  B(, r ), ) C exp
2 3r2 (t)

Hint: Use (a) and (b). See Jones and Makarov [1995].
(d) Part (c) is false with 23 r replaced by r . Hint: = 2,
 = 2D \ {|z 2| r, |y| }.
(e) Let  = {z : | arg z| < } and Br = {|z| r }. Use part (c) to show
(1, Br ,  \ Br ) Cr 1/(2 sin ) .
Also show by an explicit conformal map that
(1, Br ,  \ Br ) = Cr / .
Theorem IV.6.2 gives the accurate upper bound Cr / .

CY535/Garnett

122

0 521 47018 8

January 27, 2005

15:41

III. Potential Theory

For related results see also Beurling [1933], Exercise 24, Theorem G.1,
and the comments following the proof of Theorem G.1.
16. (a) The following result is from Gehring and Hayman [1962].
Let : D  be univalent and assume extends continuously to the open
arc = D {Imz > 0}. Then

 ((1, 1)) K  ( ) ,
with constant K not depending on .
Set I j = [1 2 j , 1 2 j1 ] and j = (B(1, 2 j ) \ B(1, 2 j1 ),
for j = 0, 1, . . . . Then clearly
inf (z, j , D) c1 > 0.

(E.8)

zI j

We may assume  j = (( j )) < . Suppose we have


sup |  (z)| c2 2 j  j

(E.9)

Ij

for all j. Then summing over j, and doing the same for I j we get
1
|  (x)|d x c2 (( )),
1

which gives the GehringHayman theorem.


Now by Theorem I.4.3, (E.9) holds if and only if
dist((z), ) c3  j

(E.10)

for some constant c3 and for all z I j . Let c3 be large and let j be the arc
length midpoint of the curve ( j ). Then
B j = B( j , c3  j ) ( j ).


If (E.10) fails at z j , then (z j )


/ B j . But using the inversion T z = zj j and
the continuum E = D T (), we see by (E.8) and the Beurling projection
theorem that (E.10) will hold at z j if c3 is sufficiently large.
(b) Here are two consequences. If there is an arc D, joining 0 to 1 such
that

|  (z)|ds =  < ,
(( )) =

then

 ((0, 1)) =

1
0

|  (x)|d x K .

(E.11)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

123

Write Iz = D {| z| < 2(1 |z|)}, for z D. If extends continuously to Iz then ((Iz )) K  dist((z), ). Inequality (E.11) will be used
in the next chapter.
See Heinonen and Nkki [1994], Heinonen and Rohde [1993], and Bonk,
Koskela, and Rohde [1998] for extensions.
17. Suppose E { : | | < } and suppose < r < 1. Then there are constants
c1 (, r ) and c2 (, r ) such that if |z| = r , then
c2 (, r )
c1 (, r )
(z, E, D \ E)
.
(E)
(E)
This can be proved directly by mimicking the proof of Theorem 9.1, or
derived from the statement of Theorem 9.1 by relating (E) to (T E)
where T is a Mbius
of the disc.

 transformation

z
be
Greens
function
for D and let be a compactly
18. Let g(z, ) = log  1
z 
supported signed measure on D.

(a) Prove that the Green potential G (z) = g(z, )d ( ) satisfies

J ( ) = G d 0
and J ( ) = 0 0. Hint: Follow the argument in Section 3.
(b) Let E D be compact. Prove Cap(E) > 0 if and only if there exists
P(E) such that J () < . Assuming Cap(E) > 0, prove there is a
unique P(E) such that

!

G d : P(E) G (E).
G d = inf
Also show that G (z) = G (E) for all z E except for set of capacity zero
and that on D \ E
(z, E, D \ E) =

1
G (z).
G (E)

(c) Let E [1, 0) be compact and let E = [1, a] be a closed interval


dt
dt
=
. Prove
with the same logarithmic measure as E,
E t
E t
(0, E, D \ E) (0, E , D \ E ).
This result, called the Beurling shove theorem, is also from Beurlings
thesis [1933]. See Essn and Haliste [1989], Pruss [1999], and Betsakos and
Solynin [2000] for refinements.
 = {
19. Write 
z = x + i|y| when z = x + i y and write F
z : z F}.

CY535/Garnett

124

0 521 47018 8

January 27, 2005

15:41

III. Potential Theory

(a) Let F be a compact subset of D, and let z D \ F. Prove


 D \ F).

(z, F, D \ F) (
z, F,
Hint: The left side is superharmonic in D and the right side is harmonic
in D \ F. Consequently we may assume z D R. In fact, via a Mbius
transformation of the disc, we can assume z = 0. Represent the right side as
the Green potential of a positive measure on F. Write F + = F {Imz0}
 define
and F = F {Imz<0} . When A F,
(A) = (F + A) + (F A)
 D \ F).

and let v(z) be the Green potential of . Show v(z) (z, F,
(b) Let E be a compact plane set of positive capacity, let  = C \ E, and let
 Prove
 = C \ E.
)
g (x, w) g (x, w
for x R {} and w C. Hint: We may take x = . Now use the general
and trivial fact

g (, w) = lim log R z, B(0, R), B(0, R) \ E .


R

(c) Let  be a simply connected domain with 0 , let a = inf  R,


b = sup  R, and
E = (, a)  (b, ),
 containing the lower half-plane. Then
and let 1 be the component of C \ E
1 is simply connected. Let :  D and 1 : 1 D be conformal
mappings with (0) = 1 (0) = 0. Prove
|  (x)| 4|1 (x)|,

x R.

Consequently the HaymanWu theorem, Theorem I.5.1, can be reduced


to the special case of L . That case has been proved by B. Flinn [1983].
Parts (a), (b), and (c) are attributed to Baernstein in Fernndez, Heinonen,
and Martio [1989].
20. Halls lemma asserts that if E H is compact, E = {|z| : z E}, and
z = z, then
(z, E, H \ E)

2
(z , E , H).
3

(E.12)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

z
E

125

z
E

E
E E
Figure III.8 Halls lemma.
(a) Derive the special case z = z and E {arg z < 4 } of Halls lemma
from Theorem 9.4. Hint: For 0 < x < 1, x 1/(1 + log(1/x)). Split E into
E D and E C \ D, using a reflection to estimate the harmonic measure
of the latter.
(b) The general case is also proved using Green potentials. See Hall [1937]
or Duren [1970].
(c) Contrary to intuition, (E.12) is not true with the constant 23 replaced by
1. See Hayman [1974] and Marshall and Sundberg [1989].
(d) (radial Halls lemma) If E D and if = dist(0, E) > 0 then
(0, E, D \ E) C(0, E , D),
where E is the radial projection of E onto D and C is independent of
E. Hint: Assume that E is the union of circular slits contained in {|z| > 21 }
such that no radius meets E more than once. Let V be the Green potential of
|d |/(| | log(1/| |)|. Then V (z) c(z, E, D \ E) and V (0) = |E | (ksendal, private communication).
21. Let Q = {0 < x < 1, 0 < y < 1} and let {Q j } be a sequence of pairwise
disjoint dyadic squares
Q j = {k j 2n j < x < (k j + 1)2n j , 0 < y < 2n j }

with sidelength (Q j ) = 2n j 14 . Set  = Q \ Q j and z 0 = 21 + i 34 .

Prove that given > 0 there is > 0 such that if (z 0 , Q j , ) <

then (Q j ) < .
22. The following results are from Beurlings elegant paper [1940]. Also see
Beurlings thesis [1933]. Let u(z) be the Poisson integral of a real function
f L 2 (D).
(a) If f has Fourier series

a0 
an cos n + bn sin n ,
+
2
n=1

CY535/Garnett

0 521 47018 8

January 27, 2005

126

15:41

III. Potential Theory

then


S( f ) D(u) =
D

2  u 

2
0

|u|2 d xd y =

2 r dr d =

n(an 2 + bn 2 ).

(b) Assume a0 = u(0) = 0, and assume D(u) < . Set


R
 u 
 dr.
U (Rei ) =
r
0
Then
(i) |u| U,
(ii) U is subharmonic,
(iii) F( ) = U (ei ) L 2 (D),
and

2
1 2 
1
F( ) d
n(an 2 + bn 2 ).
2 0
2
(c) Prove

2 U
0

(r e ) d r
i

2 U

2
0

2
(r ei ) d.

Then by a partial integration (with respect to


Hint: First suppose U
),
2
R 2
U
2 U
2 
2 U
U d dr =
r2
d.
r

2
r
r

0
0
0
C 2.

(d) Because
D(U ) =


U
2
1 U
2 
d,
+ 2
r
r

part (c) implies that D(U ) D(u) and by the Dirichlet principle (see Appendix C), S(F) S( f ). In fact, S(F) < S( f ) because U is not harmonic.
(e) Let be a measure on D and define

h n = cos n d( ),

kn =

sin n d( ).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

Then

U (r e ) =
i


log

127

 rn

1
d(t) =
(h n cos n + kn sin n ).
r ei eit
n
0

(f) If E D is compact, if Cap(E) < , and if = E , then



 h n 2 + kn 2
I () = lim U (r ei )d( ) =
.
r 1
n
If S( f ) < , then


$
1
i
S( f ) I ().
lim u(r e )d( ) =
(an h n + bn kn )
r 1

(g) If S( f ) , then
E = {ei : lim sup |u(r ei )| > }
r 1

satisfies
Cap(E ) e .
2

(h) Therefore, if S( f ) < , then


 1 excepti on a set of capacity zero,
limr 1 u(r ei ) exists, and in fact, 0 | u
r (r e )|dr < .

cn z n , n|cn |2 < , and
(i) Prove the theorem of Fejr that if f (z) =
0

cn ein = . Thus if S( f ) < , the Fourier
limr 1 f (r ei ) = , then

series an cos n + bn sin n converges except on a set of capacity zero.
23. (a) Let (z) be a univalent function on the unit disc. Prove that has a
1
nontangential limit and that 0 |  (r )|dr < for all D \ E, where
Cap(E) = 0. This is also from Beurling [1940].


(b) Let f (z) = cn z n be analytic on D, and assume n|cn |2 1. Then
1
E = { D :
| f  (r )|dr > }
0

1
has Cap(E ) < e . Consequently, 2
|E | < e1 . The second estimate
is from Beurling [1933]. See also Marshall [1989].
24. (a) Now let  be a simply connected domain, and let z, z 0 . Consider
the family  of rectifiable curves in  connecting z to z 0 . With Beurling,
define

(z 0 , z) = sup{ inf |  (z)|ds : univalent on , Area(()) }.


Prove
e2g(z,z 0 ) + e = 1.
2

CY535/Garnett

0 521 47018 8

January 27, 2005

128

15:41

III. Potential Theory

(b) Now let E . Consider the family  of rectifiable curves in 


having one endpoint z 0 and the other endpoint in E. With Beurling, define

L(z 0 , E) = sup{ inf |  (z)|ds : univalent on , Area(()) }.


Prove
(z 0 , E, ) e1L

2 (z

0 ,E)

This estimate, from Beurlings thesis [1933], is a precursor of extremal


length. Note that (b) follows from 23(b).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

IV
Extremal Distance

Extremal distance is a conformally invariant version of distance. As such it is


a powerful tool for estimating conformal invariants like harmonic measure in
terms of more geometric quantities. We begin the chapter with the basic properties of extremal length and of extremal distance, which is the most important
application of extremal length. We then characterize certain extremal distances
in Jordan domains by means of conformal mappings from the domains to slit
rectangles. This characterization leads to fundamental estimates for harmonic
measure by extremal distance and the famous integral

dx
.
(x)

1. Definitions and Examples


By definition, a path family in a domain  is a non-empty set  of countable
unions of rectifiable arcs in . An element  is called a curve even though
may not be connected and may have many self-intersections. The euclidean
length of the path family  is

ds,
(1.1)
inf


where ds denotes arc length. The quantity (1.1) is not conformally invariant,
but a conformally invariant version of (1.1) is



|  (z)|ds ,
(1.2)
sup inf

where the supremum


is taken over all conformal maps :  () such that

Area () = 1. See Exercise III.22 for an inequality about (1.2). However,
129

CY535/Garnett

130

0 521 47018 8

January 27, 2005

15:41

IV. Extremal Distance

the supremum in (1.2) is difficult to estimate geometrically and (1.2) becomes


more manageable if we use functions more general than |  (z)| in the arc length
formula.
By definition, a metric is a non-negative Borel measurable function on 
such that the area,

A(, )
2 d xd y,


satisfies 0 < A(, ) < . When is a metric and  is a path family, we


define the -length of  by

L(, ) = inf
|dz|,


and the extremal length of  by


 () = sup

L(, )2
.
A(, )

(1.3)

A conformal mapping transforms the metric on  into the metric


( 1 (z))|( 1 ) (z)| on (). Therefore () (()) =  () and the extremal length (1.3) is conformally invariant.
An extremal metric for  is a metric which attains the supremum (1.3).
In many important cases there exists an extremal metric of the form |  (z)|
for some conformal mapping , and in those cases (1.3) is just the square of
(1.2). The advantage of the more general expression (1.3) is that every metric
provides a lower bound for  (). Indeed, one of the main results of the theory,
Theorem 6.1 below, is proved using metrics not of the form |  (z)|.
Because of its homogeneity, the ratio (1.3) is unchanged if the metric is
multiplied by a positive constant. Thus we can normalize the metric to satisfy
A(, ) = 1 so that
 () = sup{L 2 (, ) : A(, ) = 1},
or we can normalize the metric by L(, ) = 1 so that
 ()1 = inf{A(, ) : L(, ) = 1}.
The quantity  ()1 is called the modulus or module of . Some authors
emphasize moduli instead of extremal lengths.
Extremal distance is the most useful instance of extremal length. When
E  and F  the extremal distance d (E, F) from E to F is defined by
d (E, F) =  (),

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Definitions and Examples

131

where  is the family of connected arcs in  that join E and F. The conjugate
(E, F), is defined to be the extremal length of the family
extremal distance, d
(E, F) we allow a curve

 of curves that separate E from F. To compute d

 to be any finite union (not necessarily connected) of arcs and closed


curves in  such that E and F lie in the boundaries of distinct components of
 \ . See Figure IV.1.


E

Figure IV.1 Connecting and separating curves.


Example 1.1 (The rectangle). If R = {(x, y) : 0 < x <  and 0 < y < h} is
a rectangle of length  and height h, then the extremal distance between the
vertical sides E and F is

(1.4)
dR (E, F) = .
h
The rectangle example is the building block for many results about extremal
length.
Proof. Let  be the family of connected arcs in R joining E and F and let
be a metric on R. Then
 
2

L 2 (, )
(x + i y)d x 
2 (x + i y)d x
(1.5)
0

by the CauchySchwarz inequality. Integrating (1.5) with respect to y gives


L 2 (, )h A(R, )
and hence
dR (E, F) = sup


L 2 (, )
.
A(R, )
h

Equality holds when = 1, and that proves (1.4).

(1.6)


By the condition for equality in the CauchySchwarz inequality, (1.5) is


an equality if and only if is constant almost everywhere d xd y. Therefore

CY535/Garnett

0 521 47018 8

January 27, 2005

132

15:41

IV. Extremal Distance

every extremal metric is constant almost everywhere. The conjugate extremal


distance dR (E, F) is the extremal distance between the two horizontal sides of
R, and thus dR (E, F) = h .
More generally, let  be a Jordan domain, let E = [1 , 2 ] and F = [3 , 4 ]
be two disjoint subarcs of , where 1 , 2 , 3 , and 4 are listed in the counterclockwise ordering of . There exist a unique  > 0 and a conformal mapping from  to a rectangle R = {(x, y) : 0 < x < , 0 < y < 1} so that
1 , 2 , 3 , and 4 are mapped respectively to the corners i, 0, , and  + i.
To prove the rectangle exists, map  to a half-plane and apply a Schwarz
Christoffel map or use Theorem 4.1. To prove  is unique use equality (1.4) and
the conformal invariance of extremal lengths or apply Schwarz reflection to the
map between two rectangles. By conformal invariance
 = d (E, F)
and the extremal metrics for d (E, F) are the constant multiples of |  |. Thus
(1.3) is the square of (1.2) in this case. For the rectangle we also have

(E, F) = 
d (E, F) d

E
2

1
= 1.


(E)

(1.7)

(F)

3
Figure IV.2 Extremal distance in a quadrilateral.

Recall that every metric gives a lower bound for d (E, F). Equality (1.7)
(E, F), and
is significant because every metric also gives a lower bound for d
(E, F) provides an upper bound for
when (1.7) holds every lower bound for d
d (E, F). We shall see in Theorem 4.1 and Theorem H.1 in Appendix H that
(1.7) holds in very general circumstances.
Example 1.2 (The annulus). The extremal distance between the two boundary
circles Cr and C R of the annulus A = {z : r < |z| < R} is
 
R
1
log
.
dA (Cr , C R ) =
2
r

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Four Rules for Extremal Length

133

Proof. The reasoning is similar to the rectangle case. Since (1.3) is a supremum,
we can assume 0 < r < R < . Let  be the family of connected arcs in A
joining Cr to C R and let be a metric on A. Then again by CauchySchwarz,
 R
2

R
R
2
i
L (, )
(te )dt log
(tei )2 tdt.
(1.8)
r r
r
Integrating (1.8) with respect to gives
2 L 2 (, ) log

A(A, )

and consequently
dA (Cr , C R ) = sup

1
L 2 (, )

log
.
A(A, )
2
r

(1.9)

We will get equality holds in (1.8) if and only if (z) = c/|z| a.e., where
c is a positive constant,

and for such we also get equality in (1.9). Thus
1
log Rr .

dA (Cr , C R ) = 2
The proof shows that the extremal metrics for dA (Cr , C R ) have the form
c/|z|. The conjugate extremal distance, dA (Cr , C R ), is equal to the extremal
length of the family  of closed curves in A separating the two boundary circles
Cr and C R . By a similar argument we also have
dA (Cr , C R ) =

1
2
.
R
=
d
(C
log r
A r , CR)

More generally, a ring domain is a doubly connected plane . By Exercise 7


of Chapter II every ring domain is conformally equivalent to an annulus of the
form A = {z : r < |z| < R}. Therefore the module of the family  of

curves
1
log Rr , and an
in  separating the two components of  is  ()1 = 2


(z)|
extremal metric is |
|(z)| where :  A is a conformal map. The quantity
 ()1 is also called the module of the ring domain  and denoted by

 ()1 = mod() =

1
log
.
2
r

2. Uniqueness of Extremal Metrics


It is an open problem to determine when a path family  has an extremal metric.
However, if an extremal metric does exist, then up to a multiplicative constant
it is unique area almost everywhere.

CY535/Garnett

134

0 521 47018 8

January 27, 2005

15:41

IV. Extremal Distance

Theorem 2.1. Let  be a path family on  and let 1 and 2 be metrics on 


satisfying
L 2 (, 1 )
L 2 (, 2 )
=
.
A(, 1 )
A(, 2 )
Let 3 = c1 1 + c2 2 , where c j =

1
21
.
2 A(, j )

Then

L 2 (, 3 )
L 2 (, 1 )

.
A(, 3 )
A(, 1 )

(2.1)

If equality holds in (2.1) then c1 1 = c2 2 a.e. d xd y. In particular, if 1 and 2


are extremal metrics for , normalized by A(, 1 ) = A(, 2 ), then 1 = 2
a.e. d xd y.
Proof. Without loss of generality we assume A(, 1 ) = A(, 2 ) = 1. Then
L(, 1 ) = L(, 2 ) and by the definition of L(, ),

1
L(, 1 ) + L(, 2 ) = L(, 1 ).
L(, 3 )
2
On the other hand, by the CauchySchwarz inequality,

1
1
1
A(, 3 ) = A(, 1 ) + A(, 2 ) +
1 2 d xd y 1.
(2.2)
4
4
2 
Therefore (2.1) holds. If equality occurs in (2.1), then A(, 3 ) = 1 and equality occurs in (2.2), so that 1 = 2 almost everywhere d xd y, again by the
condition for equality in the CauchySchwarz inequality. In particular, if 1
and 2 are extremal metrics, then equality holds in (2.1) and 1 = 2 almost

everywhere.

3. Four Rules for Extremal Length


Extremal lengths obey four basic rules.
1. The extension rule. Let   be domains and let  be a path family in
. Then
 () =  ().

(3.1)

Moreover, if   is a path family in  such that every    contains some


, then
 (  )  ().

(3.2)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Four Rules for Extremal Length

135

Equality (3.1) says that the extremal length depends on the path family 
and not the domain , and for this reason we often write () for  (). The
extension rule shows that d (E, F) is decreased if any of the sets E, F, or 
is increased and d (E, F) = d\(EF) (E, F) when E and F are closed.
 =  
E

F
F





Figure IV.3 Extension rule.


In Figure IV.3, two applications of the extension rule give
d (E  , F  ) d (E, F) d (E, F).
Proof. The curves in the family   are longer and fewer than the curves in the
family . Therefore  (  )  () and inequality (3.2) will follow from
equality (3.1).
Let  be any metric on  and set =  | . Then L(, ) = L(,  ) and
A(, ) A( ,  ). Taking the supremum over  yields  ()  ().
Now let be any metric on  and define the metric  on  by  =  . Then
L(, ) = L(,  ) and A(, ) = A( ,  ). Taking the supremum over ,

we get  ()  (), which proves (3.1).
2. The serial rule. Let 1 and 2 be path families contained in disjoint open
sets 1 and 2 respectively, and let  be a path family contained in a domain
 1 2 . If each  contains some 1 1 and some 2 2 , then
() (1 ) + (2 ).

CY535/Garnett

0 521 47018 8

January 27, 2005

136

15:41

IV. Extremal Distance

1

2

Figure IV.4 Serial rule.


In Figure IV.4, the crosscut F divides  into regions 1 and 2 with E 1
and G 2 , and by the serial rule,
d (E, G) d1 (E, F) + d2 (F, G).
Proof. If either (1 ) or (2 ) is 0 or , the result follows from the extension rule. Otherwise, choose metrics 1 in 1 and 2 in 2 , normalized by
the conditions L(1 , 1 ) = A(1 , 1 ) and L(2 , 2 ) = A(2 , 2 ). Then the
metric = 1 1 + 2 2 on  satisfies L(, ) L(1 , 1 ) + L(2 , 2 )
and A(, ) = A(1 , 1 ) + A(2 , 2 ) = L(1 , 1 ) + L(2 , 2 ). Thus
L 2 (, )
L 2 (1 , 1 )
L 2 (2 , 2 )
L(1 , 1 ) + L(2 , 2 ) =
+
.
A(, )
A(1 , 1 )
A(2 , 2 )
Taking the supremum over all 1 and 2 , we obtain () (1 ) + (2 ). 
3. The parallel rule. Let 1 and 2 be path families contained in disjoint open
sets 1 and 2 respectively. If  is a path family in  1 2 such that
every 1 2 contains some  , then
1
1
1

+
.
()
(1 ) (2 )

(3.3)

E2

E1
1

G2
G1
Figure IV.5 Parallel rule.

2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Four Rules for Extremal Length

137

In Figure IV.5, F is a crosscut dividing  into regions 1 and 2 and


E j = E  j and G j = G  j j = 1, 2. By the parallel rule,
1
1
1

+
.
d (E, G)
d1 (E 1 , G 1 ) d2 (E 2 , G 2 )
In situations where (1.7) holds, this inequality can also be obtained by applying
the serial rule to the conjugate extremal distances.
Proof. If () > 0, let be a metric on , normalized by L(, ) = 1. Then
L(1 , ) 1 and L(2 , ) 1 and
A(, ) A(1 , ) + A(2 , )

1
1
+
.
(1 ) (2 )

Taking the supremum (1.3) over all , we obtain (2.3).

4. The symmetry rule. Let T :   satisfy T T (z) = z and suppose that


either T (z) or T (z) is analytic. If  is a path family in  such that T () = ,
then
2
!
L (, )
: = ( T ) |JT | ,
() = sup
(3.4)
A(, )


where |JT | = |T  | when T is analytic and |JT | = |T | when T (z) is analytic.


Proof. If 1 = ( T )|JT | then


|dz| =

T 1 ( )

1 |dz|.

Because T maps  onto , L(, ) = L(, 1 ), and because T is one-to-one,


A(, ) = A(, 1 ). Then by Theorem 2.1, 2 = 21 ( + 1 ) satisfies
L 2 (, 2 )
L 2 (, )

.
A(, 2 )
A(, )
But since T T (z) = z, we have 2 T |JT | = 2 , and hence (3.4) holds.

For example, let T (z) = z, let  {z : Imz > 0}, and let 1  T ()
be a domain such that T (1 ) = 1 . Let E, F  and set E 1 = E T (E)
and F1 = F T (F). See Figure IV.6. Then
1
d (E, F)
(3.5)
2
because to compute d1 (E 1 , F1 ) it suffices by the symmetry rule to consider
only metrics satisfying (z) = (z) and = on R 1 . If is a curve
d1 (E 1 , F1 ) =

CY535/Garnett

0 521 47018 8

January 27, 2005

138

15:41

IV. Extremal Distance

 R that
connecting E 1 to F1 , then reflecting T () yields a curve 
joins E to F and


|dz| =
|dz|.

Because is a metric, we can find curves  with -length arbitrarily close


to the length of 
. Then (3.5) follows since A(1 , ) = 2 A(, ). We will use
this idea several times.

E
R
T (E)
T ()

T (F)

Figure IV.6 Symmetry rule.


There is an analogy between extremal distance and electrical resistance or
resistivity. When the opposite ends of an electrical conductor are joined to the
two terminals of a battery and the remaining sides of the conductor are insulated,
there is a drop in potential energy from one end of the conductor to the other,
and this drop is equal to the voltage of the battery. Suppose the conductor is
a uniformly thin sheet of material in the shape of a Jordan domain . Put a
copper coating on two arcs of  and attach each copper arc to a terminal of the
battery. Assume the atmosphere acts as an insulator on the rest of the surface
of the region. Then the electrical resistance of the conductor is the extremal
distance between the two intervals, the magnitude of the electric field is the
extremal metric, the field lines are the shortest curves in the extremal metric,
and the equipotential lines are the shortest curves for the conjugate extremal
distance. The extension rule says that the resistance d (E, F) increases if 
or E or F is decreased. The serial and parallel rules correspond to the familiar
rules for calculating resistance when circuits are attached in series or in parallel.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Extremal Metrics for Extremal Distance

139

4. Extremal Metrics for Extremal Distance


Let  be a Jordan domain and let E and F be disjoint closed subsets of , each
consisting of a finite union of arcs. Recall that the extremal distance between
E and F is the extremal length of the family of curves in  connecting E to
F, and the conjugate extremal distance is the extremal length of the family of
curves separating E from F.
In this section everything will hinge on the following observation: Let
R = {(x, y) : 0 < x < 1, 0 < y < h}


be a rectangle with sides parallel to the axes, let  = R \ L j where {L j } is
a finite family of horizontal line segments in R, and let E and F be the vertical
sides of the rectangle R. Then
d (E, F) = dR (E, F)
by the argument used in Example 1.1. Moreover, every extremal metric for
d (E, F) is constant almost everywhere. The same holds for the conjugate
extremal distance

(E, F) = dR (E, F)
d

because by definition the curves that separate E from F need not be connected.

L1
L2

L3
Figure IV.7 Slit rectangle.

Now suppose there exists a conformal mapping :  R \ L j , where
R and L j are as in the preceding paragraph, such that is continuous on ,
(E) is the left vertical side of R, and (F) is the right vertical side of R. Then
by the conformal invariance of extremal length and the preceding observation,

(E, F),
d (E, F) = d() ((E), (F)) = 1/ h = 1/d

CY535/Garnett

0 521 47018 8

140

January 27, 2005

15:41

IV. Extremal Distance

and an extremal metric for d (E, F) is 0 (z) = |  (z)|. In the next theorem,
we determine when  can be mapped to a slit rectangle R so that E and F
correspond to the vertical edges of R.
Theorem 4.1. Let  be a Jordan domain and let E and F be finite unions of
closed subarcs of . Assume E F = . Then there is a rectangle R having
sides parallel to the axes and a conformal map of  onto the rectangle R
with a finite number of horizontal line segments removed such that C()
and (E) and (F) are the vertical sides of the rectangle if and only if there is
an arc  such that
E

F = .

and

(4.1)

In this case, the extremal distance from E to F is the ratio of the length to height
of this rectangle, the conjugate extremal distance satisfies

(E, F) = 1/d (E, F),


d

and the extremal metrics on  for d (E, F) are the positive constant multiples
of 0 (z) = |  (z)|.
E

(F)

E
(E)
E

F
Figure IV.8 Extremal distance and slit rectangles.

Proof. Assume is a conformal map of  onto the rectangle R with a finite


number of horizontal line segments removed such that (E) and (F) are the
vertical sides of R. Let be a curve in () connecting the top and bottom
edges of the rectangle. The curve divides () into two regions U1 and U2
such that (E) U1 and (F) U2 . Then 1 ( ) is a curve in  with two
endpoints on . Because  is a Jordan curve, these endpoints divide  into
two arcs 1 and 2 with E 1 and F 2 .
Conversely, assume there is an arc such that (4.1) holds. By the conformal
invariance of extremal distance we may suppose that  is the unit disc D and
that E F D. Let d = C \ (E F) and let (z) = (z, F, d ) be the
(z) be the harmonic conjugate of in D.
harmonic measure of F in d . Let 
We claim that
= + i

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Extremal Metrics for Extremal Distance

141

is the conformal map promised by the theorem.


Because is harmonic in d , extends to be analytic across D \ (E F).
By the symmetry of d and ,


=
=0

(4.2)

on D \ (E F).
Let E , where E is the relative interior of E in D. Because = 0
on E, we can by the Schwarz reflection principle extend to be analytic in a
neighborhood W of . The extension satisfies Re > 0 on W D and Re < 0
on W \ D. This implies  ( ) = 0 and /r 0 by Lemma II.2.4. Then
because / = 0, we obtain

=
<0
(4.3)

r
at , and hence on all of E . Applying the same argument to 1 on F we
obtain


=
>0
(4.4)

r
on F .
Let be an endpoint of E F and let W be a neighborhood of so that
U = W \ is simply connected, where is the component of E F containing
. Then is continuous on W and = + i
is analytic in U . By mapping

C \ to the upper half-plane and using Schwarz reflection again, we see that
extends to be continuous on D { }. We conclude that is continuous on
D.
Now let us follow (z) as z traverses D counterclockwise. On each component of E, = 0 and 
is strictly decreasing, by (4.3). On each component of F, = 1 and 
is strictly increasing, by (4.4). On any component
is constant, by (4.2). If both endk of D \ (E F), 0 < < 1 and 
points of k are in E, then the curve (k ) traces a horizontal line segment
L k {z : 0 Rez 1}, beginning and ending on {z : Rez = 0}. If both endpoints are in F, then (k ) traces a horizontal segment L k {z : 0 Rez 1},
beginning and ending on {z : Rez = 1}. There are only two components k having endpoints on both E and F; on these 
is constant and = Re goes from
0 to 1 or from 1 to 0. Thus (D) contains the boundary of a rectangle R and the

contour (D) has winding number 1 about each point in R \ L k and winding number 0 about each point in C \ R. The argument principle then shows
that
%
(D) = R \
Lk

CY535/Garnett

0 521 47018 8

January 27, 2005

142

15:41

IV. Extremal Distance

and that is one-to-one.


The statements about the extremal distance and the extremal metrics now
follow from Theorem 2.1 and the observation made at the beginning of this
section.

By the proof of Example 1.2, the extremal distance and conjugate extremal
distance between the boundary circles of an annulus will also not change if
radial slits are removed from the annulus. Thus we can replace the rectangle R
by an annulus and obtain a ring domain version of Theorem 4.1.
Theorem 4.2. Suppose  is bounded by two disjoint Jordan curves 1 , 2 and
suppose E and F are finite unions of closed arcs with E 1 and F 2 .
Then there is a conformal map of  onto an annulus with finitely many radial
slits removed such that is continuous on  and the images of E and F are the
two boundary circles. The extremal metrics for the extremal distance between
E and F are constant multiples of |  (z)|/|(z)| and
d (E, F) =

1
1
(E, F) = 2 log R,
d

where R > 1 is the ratio of the radii of the boundary circles.

E
E

(F)

E
(E)
Figure IV.9 Extremal distance and slit annuli.
Proof. By conformal invariance, we may suppose that 2 = D and that
1 D is an analytic Jordan curve. If  \ (E F) = , take two copies
of  and form the doubled Riemann surface d by attaching them along
 \ (E F). Then d is a torus with finitely many arcs removed, where the
arcs correspond to the intervals in E F. If E F = , then set d = .
Let (z) = (z, Fd , d ) be the harmonic measure of the double Fd of F in
d . (The proof of Theorem B.4 in Appendix B explains the construction of d
and the solution of the Dirichlet problem on d .) Let 
be the (locally defined)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Extremal Distance and Harmonic Measure

143

harmonic conjugate of in . Choose a Jordan curve  homologous to


D. If



ds = 0,
(4.5)
s
then = + i
is single valued and analytic in  and by the proof of Theorem 4.1 extends to be continuous on . But then by (4.2) and (4.4), 
increases
as z traces D, in contradiction to (4.5). Thus there exists a non-zero constant
c so that



ds = 2.
c
s

Then = ec is single valued and analytic on , and is continuous on . As


decreases by 2 . Hence
z traces D, c
increases by 2 and, as z traces 1 , c
() contains the boundary of an annulus, and the remaining assertions of
the theorem follow just as in Theorem 4.1.


5. Extremal Distance and Harmonic Measure


We begin with the special case of a rectangle. Set
R L = {z : |Rez| < L and |Imz| < 1}.
Lemma 5.1. If E L = {z R L : |Rez| = L} is the union of the vertical edges
of R L , then

e 2 L (0, E L , R L )

8 L
e 2 .

(5.1)

The estimates in (5.1) are sharp, because

lim (0, E L , R L )e 2 L =

and

lim (0, E L , R L )e 2 L = 1.

L0

(5.2)

The lemma connects the harmonic measure of the two ends of R L at its center
point to the extremal distance L between the two ends. When L 1, the lower

bound in (5.1) can be improved to (.93)(8/ )e 2 L . When L < 1, better bounds


can be obtained by applying (5.1) to (0, E 1/L , R1/L ) = 1 (0, E L , R L ).
See Exercise 10.
Proof. Let SL be the infinite strip {z : Rez > L and |Imz| < 1} and let L (z)
be the harmonic measure of the left edge SL {Rez = L} in SL . Using the

CY535/Garnett

0 521 47018 8

January 27, 2005

144

15:41

IV. Extremal Distance

conformal map z e z/2 and some elementary geometry, we obtain


)
+

e 2 z (ie 2 L )
2
.
L (z) = arg

e 2 z ie 2 L
In particular, L (0) =
Thus for z R L ,

(5.3)

tan1 (e 2 L ). By (5.3) L (z) L (L) on Rez = L.

(z, E, R L ) L (z) + L (z) (1 + L (L))(z, E, R L ).


By the maximum principle, this inequality persists at z = 0, and we conclude
that
2 L (0)
(0, E, R L ) 2 L (0).
1 + L (L)
Because
L (L) = 2L (0) =

4
tan1 (e L ),

we obtain
8

1+

tan1 (e 2 L )
4

tan1 (e L )

(0, E, R L )

8
tan1 (e 2 L ).

(5.4)

The estimates (5.1) now follow from the elementary inequalities


,
t tan1 (t) min t,
4
4
valid for 0 t 1, and both equalities in (5.2) are immediate from (5.4).

Now let  be a Jordan domain, let E be an arc on  and let z 0 . Consider


all Jordan arcs  that join z 0 to  \ E, and define
(z 0 , E) = sup d\ (, E),

where the supremum is taken over all such Jordan arcs.


E

z0

Figure IV.10 Distance from a point to an arc.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Extremal Distance and Harmonic Measure

145

Both quantities (z 0 , E) and (z 0 , E, ) are conformally invariant but


(z, E, ) is strictly increasing in E, while (z 0 , E) is strictly decreasing
in E. This means there is a function connecting to , and the next result
examines that function.
Theorem 5.2. Let  be a Jordan domain, let E be a subarc of  and let
z 0 . Then
e(z 0 ,E) (z 0 , E, )

8 (z 0 ,E)
e
.

Moreover
lim (z 0 , E, )e(z 0 ,E) =

and

lim (z 0 , E, )e(z 0 ,E) = 1.

By Theorem 5.2 every choice of the arc and every choice of the metric
give an upper bound for (z 0 , E, ), because (z 0 , E) and the extremal
distance d\ (, E) are both suprema. This idea of varying is from Beurling
[1989].
Proof. By conformal invariance we may suppose that  = D, that z 0 = 0 and
that E is an arc on D. Let E 1 and E 2 be the two disjoint arcs on D given by
E 1 E 2 = {ei : e2i E}. Because the arcs E 1 and E 2 are symmetric about
0, there is a conformal map f of D onto a rectangle R such that R has center
j = f (E j ),
0 and sides parallel to the axes, such that f (0) = 0 and such that E
j = 1, 2, are the vertical sides of R. Let be any arc in D connecting 0 to

D \ E and let 
= {z R : ( f 1 (z))2 }. The map z is conformal on

D \ and the two branches of f ( z) map D \ conformally onto the two


components of R \ 
. Hence by the serial rule and by conformal invariance,
1 , E
2 ) d( E
1 , 
2 ) = 2dD (, E).
) + d(
, E
dR (E
Equality holds if 
0 is the vertical line segment in R through 0, so that
1 , E
2 ).
2(0, E) = d R ( E

E1

0 E2

Figure IV.11 Proof of Theorem 5.2.

1
E

2
E

CY535/Garnett

0 521 47018 8

146

January 27, 2005

15:41

IV. Extremal Distance

Since (z 2 , E, D) = (z, E 1 E 2 , D), we have


1 E
2 , R).
(0, E, D) = (0, E 1 E 2 , D) = (0, E
Then by Lemma 5.1,
e(0,E) (0, E, D)

8 dD (,E)
e
.

The upper estimate is also valid when E is a finite union of arcs on .


In that case we consider all Jordan arcs in  connecting z 0 to  \ E, and
define
(z 0 , E) = sup d\ (, E).

Theorem 5.3. Let  be a Jordan domain, and let E be a finite union of arcs
contained in . Then
8
(5.5)
(z 0 , E, ) e(z 0 ,E) .

Proof. We can assume  = D and z 0 = 0. Let be an arc from 0 to D \ E and


set = d\ (, E). Write 1 = {z : z 2 } and {ei : e2i E} = E 1 E 2
where 1 separates E 1 from E 2 and z 2 (E 1 ) = z 2 (E 2 ) = E. By Theorem 4.1,
there is a conformal map of D onto a rectangle R with horizontal slits removed
j = (E j ) are the vertical ends of R. Then as in Theorem 5.2 we obtain
so that E

2 ) 2dD (, E) and consequently
d R ( E1, E
8
e
.

Taking the supremum over then gives (5.5).


(0, E, D)

There is no lower bound in Theorem 5.3 because the slits can consume nearly
all of the harmonic measure. See Exercise 12.

6. The

dx
(x)

Estimate

Every lower bound for extremal distance yields an upper bound for harmonic
measure, and by (1.3) every metric yields a lower bound for extremal distance. Constructing good metrics is therefore an important method. Consider
the special case of a strip domain:
 = {(x, y) : |y m(x)| < (x)/2, a < x < b},

CY535/Garnett

0 521 47018 8

January 27, 2005

6. The

dx
(x)

15:41

Estimate

147

having varying width (x) and mid-line y = m(x).

(x)

y = m(x)

x
Figure IV.12 A strip domain.

Let a = x0 < x1 < . . . < xn = b and set j =  {z : Rez = x j }, where


j = 0, . . . , n. By the serial rule,
d (0 , n )

n


d( j1 , j ).

j=1

If x = x j x j1 is small, the region between j1 and j is approximately


a thin rectangle having x as base, (x j ) as height, and x j + im(x j ) as the
midpoint of its right vertical side. Under the linear map
z im(x j )
,
(x j )
this rectangle is sent to a rectangle centered on R with height 1 and width
x/ (x j ), so that d( j1 , j ) x/ (x j ). Consider the (non-analytic) map

 x
dt y m(x)
,
(6.1)
 : (x, y)
(x)
a (t)
b 1
from  onto a rectangle of height 1 and length a (t)
dt. The extremal metric
for d (0 , n ) is given by 0 (z) = |  (z)| = |Re(z)|, where is a conformal
map of  onto some rectangle, but the key idea here is to replace the analytic
by  and use as metric
(z) = |Re| =

1
,
(x)

z = x + i y.

If is any curve connecting 0 to n in , then



b
1
d x.
(z)|dz|

(x)
a

CY535/Garnett

0 521 47018 8

January 27, 2005

148

15:41

IV. Extremal Distance

Moreover

2 d yd x =
a

so that


d (0 , n )

1
d x,
(x)

1
d x.
(x)

More generally, let  be a finitely connected Jordan domain and assume


E {z : Rez a}  and F {z : Rez b}  are finite unions of arcs.
Suppose I x {z : Rez = x} separates E from F and let (x) be the length of
I x , a < x < b. We assume that (x) is measurable but we do not assume that
I x =  {z : Rez = x} or that I x is connected. Define the metric A by

1
(x)
if (x, y) I x ,
A (x, y) =

0
elsewhere in .
Then by the above argument
b
1
d (E, F)
d x.
(6.2)

(x)
a
Theorem 6.1. Let  be a Jordan domain and let z 0 . Let b > x0 = Rez 0
and suppose F {z  : Rez b}. Assume that for x0 < x < b, there exists
I x  {Rez = x} separating z 0 from F and set (x) (I x ). Then


b
dx
8
.
(6.3)
(z 0 , F) exp

x0 (x)
F

I x2

I x3


F

I x3


F

z0
I x1

x0


F

R
x4
x1
x2
x3
b
Figure IV.13 Crosscut estimate of harmonic measure.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Notes

149

For example, if I x = {z  : Rez = x}, then (x) is lower semicontinuous


and measurable.
 be the component of  \ {z : Rez = b} that contains z 0 and let
Proof. Let 

 is a finite union of arcs on 


F =  {z : Rez = b}. We can suppose that F
1
by replacing  with (|z| < r ) where is a conformal map of  onto D.
. By the maximum
Let {z : Rez = Rez 0 } be a curve connecting z 0 to 
principle,
 
),
(z 0 , F, ) (z 0 , F,
and then Theorem 5.3 and inequality (6.2) give (6.3).

Theorem 6.1 remains true if (x) is replaced by ({z  : Rez = x}), because that change only increases the right side of (6.3). When {z  : Rez = x}
contains several crosscuts that separate z 0 from F, sharper versions of (6.3) are
available. See Appendix G. Theorem H.8 in Appendix H extends Theorem 6.1
to finitely connected domains.
Theorem 6.1 is often used in polar coordinates, as in the following theorem.
Theorem 6.2. Let  be a Jordan domain and let E  {|z| R}. If z 0 
satisfies r0 = max(|z 0 |, dist(0, )) < R. Suppose that Jr {z  : |z| = r }
separates z 0 from E, r0 < r < R, and let r (r ) be the length of Jr . Then


R
dr
8
(z 0 , E) exp
,
(6.4)

r0 r (r )
if (r ) is measurable.
Proof. Define the metric
1
r (r )
A (z) =

in Jr ,
in  \

{Jr : r0 < r < R},

for z  and r = |z| and repeat the proof of Theorem 6.1.

Notes
Much of this chapter is based on Beurlings MittagLeffler Institut lectures,
published in [1989], and on Ahlfors [1973]. In [1973] Ahlfors says that Beurling invented extremal length in 1943 or 1944, but waited until 1946 to announce
his results. The application of the CauchySchwarz inequality in the proof of
(1.4) is often referred to as the lengtharea principle; it goes back to early

CY535/Garnett

150

0 521 47018 8

January 27, 2005

15:41

IV. Extremal Distance

work by Grtzsch [1928] and others. Both Beurling [1989] and Ahlfors [1973]
obtained the extremal metric |  (z)| = |Re(z)| of Section 4 by solving a
mixed DirichletNeumann problem for Re(z). Lemma 5.1 is from Marshall
and Sundberg [1989]. Theorem 5.3 is in Beurling [1989], with different constants. The first form of Theorem 6.1 was proved by Ahlfors [1930], but we have
followed Beurling [1989]. Fuchs [1967], Hersch [1955], and Ohtsuka [1970]
have other presentations. Lord Rayleigh [1871], [1876] also derived the estimate (6.2) in 3 dimensions for the resistance of a conductor using the serial
rule and proved that the conjugate resistance (extremal distance) is the reciprocal of the resistance for a quadrilateral.
[1988] says Carleman was
 Baernstein
dx
to measure a domain and calls it
the first to use integrals of the form (x)
surely one of the most brilliant ideas in the history of complex function theory. Carlemans original [1933] estimate of harmonic measure by the above
integral, proved with a differential inequality, will be given in Appendix G. In
Appendix H, extremal distances on finitely connected domains will be studied
by embedding the domains in their Riemann surface doubles.

Exercises and Further Results


1. Let  be a Jordan domain and let A1 , B1 , A2 , and B2 be arcs of  listed
in counterclockwise order. Assume these arcs have disjoint interiors and
assume A1 A2 B1 B2 = . Prove
Area() distC (A1 , A2 )distC (B1 , B2 ),
where distC denotes euclidean distance. Prove equality holds only if
A1 , B1 , A2 , and B2 are the four sides of a rectangle.
2. Suppose two annuli A1 and A2 are each bounded by concentric circles. Prove
that if A1 and A2 are conformally equivalent, then the ratios of the radii of
their bounding circles are the same. Find one proof with extremal lengths
and find one proof without extremal lengths.
3. Let S be a one-to-one analytic or conjugate analytic map from a domain 1
onto a domain 2 such that 1 2 = . Let 1 be a path family in 1 ,
set 2 = S(1 ) and suppose  is a path family in 1 2 .
(a) If 1 S(1 )  for every 1 1 , and if every  contains some
1 1 and some 2 2 , then () = (1 ) + (2 ) = 2(1 ). In other
words, equality holds in the serial rule. Moreover, ( ) = () where
 = {1 S(1 ) : 1 1 }. For example, if S(z) = z and S() = , set

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

151

1 =  {z : Imz > 0}. Then if E  {z : Imz > 0}, we obtain


d (E, S(E)) = 2d1 (E, R ).
(b) If every 1 1 and every 2 2 contains some  and if whenever
, (S 1 1 ) contains some 1 1 , then
1
1
2
1
=
+
=
.
()
(1 ) (2 )
(1 )
In other words, equality holds in the parallel rule. For example, if S(z) = z
and S() = , and if E 1 , F1  {z : Imz > 0}, then for E 2 = S(E 1 ),
F2 = S(F1 ), and 1 =  {z : Imz > 0},
1
1
1
2
=
+
=
.
d (E 1 E 2 , F1 F2 )
d1 (E 1 , F1 ) d2 (E 2 , F2 )
d1 (E 1 , F1 )
4. Let a < b < c be real, let  be the family of arcs in C joining [, a]
to [b, c], and let  + be the family of arcs in the upper half-plane joining
[, a] to [b, c].
(a) Show that
b a 
,
2() = ( + ) = f
cb
where f (t) is strictly increasing, f (0) = 0, and f () = .
(b) Write the function f as an elliptic integral, using the SchwarzChristoffel
formula.
(c) If C is the circle centered on R through a and c, let be the intersection
of {Rez = b} and C in {Imz > 0}. Show that
8
cb
2
+
1
e( )/2 .
(, (, a] [b, c], H) = tan

ba

5. A ring domain  separates the two sets E and F if E and F are contained
in different components of C \ .
(a) (Grtzsch [1928]) If  is a ring domain separating {0, r } from the unit
circle { : | | = 1}, then
mod() mod(D \ [0, r ]),

(E.1)

where mod() denotes the modulus of . Hint: Apply the extension rule for
the (larger) family of curves in D separating {0, r } from D. Then use the
symmetry rule, and consider curves in D {Imz > 0} connecting [1, 0] to
[r, 1].

CY535/Garnett

152

0 521 47018 8

January 27, 2005

15:41

IV. Extremal Distance

(b) (Teichmller [1938]) If z 1 , z 2 , and 0 are three distinct points and if  is


a ring domain separating {0, z 1 } from {z 2 , }, then

mod() mod C \ [|z 1 |, 0] [|z 2 |, ] .


(E.2)
Many authors call (E.2) Teichmllers module theorem, but in [1938] Teichmller called a weak
version of Theorem

4.1 of Chapter V Der Modulsatz.


(c) If  = C \ [0, 1 ] [1, ] then
mod()

C
log 1

Hint: See Exercise 4 and (1.7) or Lehto and Virtanen [1973], page 61.
(d) If  = 1 2 where 1 is a rectifiable Jordan curve with (1 ) 1
and dist(1 , 2 ) > , then
mod() C.


Hint: If E = {z : dist(z, 1 ) < } and = E , then 2 d xd y C.
(e) Given R > 0, there exists A > 0 such that every ring domain  with
mod() > A contains an annulus {r1 < |z z 0 | < r2 } with r2 /r1 > R.
6. (a) Let A and B be disjoint arcs in C. Then
2
dist(A, B)
1
.
dC (A, B)
diam(A) + dist(A, B)
Hint: Take z 0 B such that dist(z 0 , A) = dist(A, B) and define

1
if |z z 0 | < dist(A, B) + diam(A),
dist(A,B)
(z) =

0
if |z z 0 | dist(A, B) + diam(A).
(b) Suppose  is a ring domain having boundary components 1 and 2 such
that with respect to the spherical metric (z) = (1 + |z|2 )1 , diam( j ) .
Prove

mod() 2 .
4
If also dist(1 , 2 ) , prove


tan( 2 )
1
2
log
.
mod()

tan( 2 )
See Lehto and Virtanen [1973], page 34.
7. (Beurlings criterion for extremal metrics) Let  be a path family in a domain
 and
 let 0 be a metric on . Suppose  contains a subfamily 0 such that
(i) 0 |dz| = L(, 0 ) for all 0 , and

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

153


(ii) If h is real valued and Borel measurable in  and if h|dz| 0 for all

0 , then  h0 d xd y 0.
Then 0 is an extremal metric for . It follows that all other extremal metrics
are of the form c0 , where c > 0 is a constant. See Ahlfors [1973].
8. Let : D  be conformal, let I be an arc of D with length |I | < and
center I and let z I = (1 |I |) I and let be any arc in D that connects
the endpoints of I . Prove dist((z I ), ) Cdiam( ).

zI

Figure IV.14
9. Let  = (0, 1) (0, 1) be the unit square, let F = {(1, y) : 0 y 1}, and
let E = {(0, y) : 0 y 1} {(1/2, 1)}.
(a) Prove d (E, F) = 1. Hint: let
B  (z)
(z) = 1 +
,
|z | log 1/|z |
where B is a ball of radius centered at = (1/2, 1).
(b) Let E consists of the set E together with an interval centered at (1/2, 1)
of length , and let = | | be the extremal metric for d (E , F). Prove
lim0 exists and is the constant metric, but the limit is not an extremal
metric for d (E 0 , F).
(c) Prove there is no extremal metric for d (E 0 , F). See Ohtsuka [1970].
10. Let R1 = {z : |Rez| < L , |Imz| < 1}, R2 = {z : |Imz| 1}, where L > 2.
Let  be a Jordan domain with R1  R2 . Let J = {(0, y) : |y| 1},
E {z R2 : Rez L} and F {z R2 : Rez L} be such that E and
F are finite unions of closed arcs in . Then there is a constant C such that
d (E, F) d (E, J ) + d (J, F) + ce L/2 .

(E.3)

In other words, if L is large then equality almost holds in the serial rule. This
result will be used in Chapter VIII. The proof of Theorem V.5.7 will include
a simple special case of (E.3).

CY535/Garnett

0 521 47018 8

January 27, 2005

154

15:41

IV. Extremal Distance

Hints: By the extension rule, d (E, F) L. By Theorem 4.1 there exists a


conformal map from  to a rectangle R with a finite number of horizontal slits removed. We may assume R = {z : |Rez| < d (E, F), |Imz| < 1},
Im((z) z) = 0 on the top and bottom edges of R1 , Re = d (E, F)
on E, and Re = d (E, F) on F. Since |Im((z) z)| 2 on the vertical
edges of R1 , |Im((z) z)| C1 e L/2 on R1 {|Rez| 2} by Lemma
5.1. By Schwarz reflection, extends to a conformal mapping from some
domain containing the rectangle {z : |Rez| < L , |Imz| < 3} into the rectangle {z : |Rez| < d (E, F), |Imz| < 3}, and |Im((z) z)| C1 e L/2 on
D2 = {|z| < 2}. Differentiating the Herglotz integral formula on D2 shows
that there is a constant C such that
|  (z) 1| < Ce L/2 ,
for z J . It follows that
sup Re inf Re 2 sup |Im  (z)| 2Ce L/2 .
J

Therefore (E.3) holds because d (E, J ) 21 (d (E, F) + inf J Re) and
d (J, F) 21 (d (E, F) sup J Re).
11. Let R L = {|Rez| < L , |Imz| < 1} and let E L = R L {|Rez| = L} be as
in Lemma 5.1.

(a) Prove that (0, E L , R L ) (.93) 8 e 2 L when L 1. See Marshall and


Sundberg [1989].

(b) If L < 1, prove that 1 8 e 2L (0, E L , R L ) 1 (.93) 8 e 2L .


(c) Let S = {|Imz| < 1} and let FL = S {|Rez| < L}. Prove
(z, E L , R L ) =

+


2(z + (4n + 2)L , FL , S).

n=

(d) Using a conformal map, find explicit formulae for (z, FL , S), and
(0, E L , R L ).
12. Let R L and E L be as in Exercise 10.
(a) Set

 1
2
xn = sin (0, E 2n L , R2n L )
.
2
2
= 21 (xn + 1/xn ). Hint: Use the Schwarz reflection principle to
Prove xn1
relate the conformal map of D onto R L to the conformal map of D onto R2L .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

155

(b) Given n and L, set


1 L2n1
e
,
2
1
yk =
(yk1 + 1/yk1 ), 1 k n,
2


2
1
1
.
= tan

yn 1/yn

y0 =

Prove
|(0, E L , R L ) | Ce L2

n1

When n = 3 and L 1 the computation of is accurate to 16 decimal places.


In general this method gives quadratic convergence to (0, E L , R L ), which
is considerably faster than other methods such as the one given in Exercise
11.
(c) Use (a) to find a quadratically convergent algorithm to compute the conformal map of D onto R L . Hint: For large N , approximate R2n L by a semiinfinite strip. See Marshall and Sundberg [1989].
13. (a) Given > 0, construct a set E D consisting of finitely many closed
arcs so that (0, E, D) < but (0, E) = sup dD\ (, E) = 1 where the
supremum is taken over all arcs connecting 0 to D. Hint: Remove horizontal slits from a rectangle.
(b) Show that for each set E D that consists of n arcs, there are n possible
rectangles that occur in the proof of Theorem 5.3. Also show that the vertical
line through 0 can meet the excised slits in the proof of Theorem 5.3. Is there
a curve so that (z 0 , E) = dD (, E)?
14. Let  be simply connected, let z 0 , z 1 , and let j be an arc in  joining
z j to . Then

Min g(z 0 , z 1 ), 1 C exp d\1 2 (1 , 2 )


for some absolute constant C. See Beurling [1989].
15. Let L n = {r ein : 0 r 1/2, n = 2n } and
 = {z : |z| < 1, Imz > 0, |z i/4| > 1/4} \

Ln.

n=1

Let E = {ei : 0 /6} and F = {r : 0 r 1}. If  denotes the


curves in  connecting E to F and n denotes the curves in  connecting

CY535/Garnett

0 521 47018 8

January 27, 2005

156

15:41

IV. Extremal Distance

E to Fn = F B(0, 2n ), then for 1,



inf
|dz| 1
n

and


inf

3
1
|dz| =
+ + > 1.
2
8
4

The approximation does not work here because the prime end corresponding
to 0 includes the interval [0, 1/2]. See Ohtsuka [1970].
16. (a) Let n  have endpoints n (0) a  and n (1) b .
Moreover suppose that there are arcs Ik  {|z a| = k } with k 0
so that Ik separates all but finitely many {n (0)} from some fixed point
z 0 . Similarly suppose there are arcs Jk {|z b| = k } so that Jk separates all but finitely many {n (1)} from z 0 . Let be a metric on  with
A(, ) < . Prove that for all > 0 there exists a curve  connecting a to b so that


|dz| lim inf
|dz| + .

Hint: Use the lengtharea principle. See Ohtsuka [1970].


(b) Suppose  is a finitely connected Jordan domain and suppose E n and Fn
are each finite unions of pairwise disjoint closed arcs in  with E n E n+1


and Fn Fn+1 . Let E = E n and F = Fn . Use (a) to prove that
lim d (E n , Fn ) = d (E, F).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

V
Applications and Reverse Inequalities

In Section 1 we present Ahlfors solution of the Denjoy problem, historically


the first application of an estimate like (IV.6.4). In Section 2 we give Beurlings
converse to Theorem IV.6.1. It is proved by constructing a new metric. In Section
3 we discuss reduced extremal distance and derive a recent sharp theorem of
Balogh and Bonk about the lengths of conformal images of radii. It is also proved
by constructing a new metric. In Section 4 we prove Teichmllers theorem on
the addition of moduli of rings. In Sections 5 and 6 we use extremal distances to
to obtain two different, definitive results on the existence of angular derivatives.

1. Asymptotic Values of Entire Functions


An entire function f (z) has asymptotic value a if there is a Jordan arc 
tending to such that
f (z) = a.

lim

 z

For example

f (z) =

ew dw
n

has n distinct asymptotic values, one along each of the curves arg z = 2 j/n,
j = 1, . . . , n. Write
M(r ) = sup | f (z)|
|z|=r

for the maximum of | f (z)| on {|z| = r }.


Theorem 1.1. If the non-constant entire function f (z) has n distinct finite

157

CY535/Garnett

158

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

asymptotic values, then


lim inf
r

logM(r )
r n/2


> 0.

(1.1)

Proof. Assume f has n distinct asymptotic values a j taken along n distinct


curves  j . Moving each  j slightly, we may assume that  j is a simple polygonal arc whose vertices tend only to . Because the asymptotic values a j are
distinct, we may further move the  j so that  j k = {0} for j = k, and so
that 0 1 when n = 1. Then the  j divide the plane into n Jordan domains
G 1 , G 2 , . . . , G n , and when n > 1 each G j is bounded by two distinct k .
On each domain G j the function f (z) must be unbounded. When n = 1,
this holds because f is not constant. When n > 1, it follows from Lindelfs
theorem, Exercise II.3(d), applied to f where is a conformal map from
D onto G j .

We can assume | f (z)| 1 on
 j , and by the maximum principle we can
choose R0 large so that for each j = 1, . . . , n there is a z j G j with |z j | = R0
and | f (z j )| > 1. Let R > R0 . Write
j (R) = (z j , {|z| = R}, G j {|z| < R}),
and write  j (r ) for the angular measure of G j {|z| = r }. Then by the maximum principle
log| f (z j )| j (R)logM(R)
and hence by (IV.6.4),



R
dr
8
logM(R).
log | f (z j )| exp

R0 r  j (r )

(When n = 1, G 1 can be mapped to a Jordan domain using z and (IV.6.4) can


then be applied.) By our assumptions on the  j ,
n


 j (r ) = 2,

j=1

and by the CauchySchwarz inequality,

n
n


1

 j (r ) n 2 ,
 j (r )
j=1

j=1

so that
n

j=1

n2
1

.
 j (r )
2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Lower Bounds

Hence

159

 
n
R
1  R dr
n

log
,
n
2
R0
R0 r  j (r )
j=1

and there is j = j (R) such that




exp

R0

dr
r  j (r )

We conclude that
0 < min log| f (z j )|
j

8 n2
R
0

R0
R

n/2
.

logM(R)
n

R2


,


from which (1.1) follows.

2. Lower Bounds
For a domain of the form
 = {(x, y) : |y m(x)| < (x)/2, a < x < b},
the inequality (IV.6.3) of Theorem IV.6.1 is not sharp, because the mapping 
defined in (IV.6.1) is not conformal. However, minor smoothness assumptions
on m(x) and (x) will yield a lower bound for (z 0 , F) that complements
(IV.6.3). Lower bounds on harmonic measure come from upper bounds on extremal distances or, equivalently, from lower bounds on conjugate extremal
distances. In view of (IV.6.1) a natural metric for studying the conjugate extremal distance in  is

 





y

m(x)
(y m)  + m  2
1
=
+
. (2.1)
B (x, y) = 

(x)
2 (x)
2
This metric was invented by Beurling [1989].
Let E =  {x = a} and F =  {x = b}. If is any curve in  connecting the curve y = m(x) + (x)/2 to the curve y = m(x) (x)/2, then







ym

dz  = 1.
B |dz| 

CY535/Garnett

0 521 47018 8

160

January 27, 2005

15:41

V. Applications and Reverse Inequalities

Furthermore,

A(, B ) =


 (
(y m)  + m  2
1
+
d yd x
2
2
m 2
b  2
1 
m (x) + 12
(x)2
dx
+
d x.
(x)
(x)
a

m+ 2

Thus
d (E, F) =

1
(E, F)
d

'

dx
+
(x)

1 
m  (x)2 + 12
(x)2
d x.
(x)

(2.2)

Theorem 2.1. Suppose  = {(x, y) : |y m(x)| < (x)/2, a < x < b} is a


Jordan domain and suppose and m are absolutely continuous on (a, b). Suppose further that
{(x, y) : |x x0 | < , |y y0 | < } ,
and set z 0 = x0 + i y0 . If F =  {(x, y) : x = b}, then
(
'
b
1 
1 + m  (x)2 + 12
(x)2
dx ,
(z 0 , F) C() exp
(x)
x0
where

C() = exp 2

x0 +
x0

(2.3)


(x)d x

depends only on and (x) for x (x0 , x0 + ).


Proof. Define the metric
 for (x, y)  by

B (x, y)

B2 (x, y) + 2

(x, y) =

if x > x0 + ,
if x0 < x < x0 + ,
if x0 < x < x0 ,
elsewhere in ,

where B is the Beurling metric (2.1).


Let be any curve connecting z 0 to . If is any curve in  \ that sep|dz| 1. Indeed suppose {z : Rez > x0 }. Then
arates from F, then


|dz| B |dz| 1. On the other hand, if {z : Rez > x0 }, then the

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Lower Bounds

161

euclidean length of = {z : x0 < Rez < x0 + } is at least , because


separates from F. Hence
|dz| 1 |dz| 1.
Also,
b m+/2


1 x0 +

2 d xd y =
B2 (x, y)d yd x + 2
(x)d x.
x0
x0 m/2


Thus,

b
1
1
d (, F) =

dx
d (, F)
x0 (x)
b  2

1 
m (x) + 12
(x)2
1 x0 +
dx + 2
+
(x)d x
(x)
x0
x0

by (2.2). Because the crosscut is arbitrary, the result now follows from Theorem IV.5.2.

Theorem 2.1 implies that Theorem IV.6.1 is the best result possible for
domains with smooth boundaries.
Corollary 2.2. Suppose  = {(x, y) : |y m(x)| <
domain and suppose

(x)
2 ,

x > a} is a Jordan

{(x, y) : |x x0 | < , |y y0 | < } .


Set z 0 = x0 + i y0 . If

 m  (x)2

+  (x)2
(x)


d x = A < ,

then whenever b > x0 and F = {z  : Rez b},






b
b
dx
dx
8
(z 0 , F) exp
,
C exp

x0 (x)
x0 (x)
where C is a constant depending only on A and (x) for x (x0 , x0 + ).
The lower bound (2.3) and the upper bound (IV.6.3) were both obtained by
studying metrics of the form = |u|. Metrics of this form have the obvious
advantage that lengths can be estimated by





|u||dz|  u dz  ,

which is the change in u along . Other metrics of this form will appear in
Section 6.

CY535/Garnett

0 521 47018 8

162

January 27, 2005

15:41

V. Applications and Reverse Inequalities

3. Reduced Extremal Distance


Let  be a finitely connected Jordan domain, let z 0 , and let E be a finite
union of closed subarcs of . The reduced extremal distance  (z 0 , E) is a
conformally invariant version of the distance



ds : , joins z 0 to E
inf

in  from z 0 to E.
Let us try to measure the distance from z 0 to E by deleting a small disc,
B = B (z 0 ) = {z : |z z 0 | < } from  and computing the extremal distance
d\B ( B , E). If < , then by the serial rule and Example IV.1.2,
d\B ( B , E) d\B ( B , E) + d B \B ( B , B )
1
log(/).
= d\B ( B , E) +
2

(3.1)

Hence
1
log
2
is a decreasing function of . We will show in a moment that h() is bounded
above. It will then follow that the three limits
1
lim d\B ( B , E) +
log ,
0
2
h() = d\B ( B , E) +

lim d\B ( B , ) +

and

1
log ,
2


(z 0 , E) =  (z 0 , E) = lim d\B ( B , E) d\B ( B , )
0

all exist and are finite. The reduced extremal distance is defined to be the third
limit (z 0 , E).
To see that h() is bounded above, choose a component E 1 of E and let be
a conformal map of C \ E 1 to D such that (z 0 ) = 0. Then (B ) B(0, a)
for some a > 0, so that by conformal invariance and the extension rule,
h() d\B ( B , E 1 ) +

1
log
2

dD\B(0,a) ( B(0, a), D) +


and h() is bounded.

1
1
1
log =
log ,
2
2
a

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Reduced Extremal Distance

163

Lemma 3.1. The reduced extremal distance (z 0 , E) is conformally invariant.


Proof. Suppose is a conformal map defined on . We may suppose that
z 0 = 0 and (z 0 ) = 0. Because is conformal at 0, there exist 0 > 0 and
K > 0 such that for = |  (0)| K 2 and = |  (0)| + K 2 ,
B (B ) B
1
when < 0 . Because k() = d()\B ( B , (E)) + 2
log is decreasing
and bounded above and because lim0 log(/) = 0, we have

lim d()\B ( B , (E)) d()\B ( B , (E)) = lim (k() k()) = 0.


0

We conclude that


(0, E) = lim d ( B , E) d ( B , )
0

= lim d() (( B ), (E)) d() (( B ), ())


0

= lim d() ( B , (E)) d() ( B , ())


0

= (0, (E)),
and (z 0 , E) is a conformal invariant.

Recall that when E is a compact plane set, the logarithmic capacity of E is


e E , where g(z, ) = log |z| + E + o(1) (as z ), is Greens function
for C \ E, and E is Robins constant for E In [1950] and [1952] Ahlfors
and Beurling obtained the following connection between capacity and reduced
extremal distance.
Theorem 3.2 (AhlforsBeurling). If E is a finite union of closed arcs in D,
then
D (0, E) = Dc (, E) = E /,
where E is Robins constant for E.
Proof. Set d = C \ E and G(z) = g(z) + g(1/z), where g(z) is Greens
function in d with pole at . Note that g(z) g(1/z) log |z| is harmonic in d and zero on d and hence g(1/z) = g(z) log |z| in d . Thus
G(z) = 2g(z) log |z| and g(0) = lim|z| g(z) log |z| = E . Since G is
symmetric about the unit circle, G/r = 0 on D \ E and G = 0 on E. If
g ) is analytic in D

g is a harmonic conjugate of g in D, then f (z) = ze2(g+i
and log | f | = G. By repeating the proof of Theorem IV.4.1, but working
with the symmetric function G instead of with , we see that f extends to be

CY535/Garnett

164

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

continuous on D, that | f | = 1 on E, and that (arg f )/ = 0 on D \ E.


The argument principle shows that f is a conformal map of D onto the disc
with radial slits removed, because f has exactly one zero in D. Furthermore,
D = f (E) and f (0) = 0. The extremal distance from F {z : | f (z)| = } to
1
log(1/). Because g(0) = E , we have = |z|e2g(z) |z|e2 E
E is then 2
on F and hence F is close to the circle of radius e2 E . But then the proof of
Lemma 3.1 yields
D (0, E) = lim dD (F , E) dD (F , D)
0
1
1
1
1 
log
log 2
= lim
0 2

2
e E
= E /.

The proof of Theorem 3.2 was a reprise of the proof of Theorem IV.4.1, with
harmonic measure replaced by Greens function and for that reason reduced
extremal distance leads to estimates in which harmonic measures are replaced
by capacities. Theorem 3.2 includes an estimate for the harmonic measure of
E D, because by Example III.1.2,
|E|
|E|

= (0, E).
(3.2)
4
2
When E is a single arc, equality holds in the first inequality in (3.2). Thus we
have the following corollary.
e E sin

Corollary 3.3. If  is a Jordan domain and if E is a finite union of closed arcs


on , then
(z 0 , E) e(z 0 ,E) .
If E is a single arc, then
2
2
sin1 (e(z 0 ,E) ) e(z 0 ,E) .

Pfluger [1955] gives a quantitative version of Theorem 3.2.


(z 0 , E) =

Corollary 3.4 (Pfluger). If  = D, if E is a finite union of closed arcs in D,


and if B is the circle centered at 0 of radius , then
E
log 1/ log(1 )
log 1/
E
+
+
dD ( B , E)
+
.

2
Proof. By (3.1) the function
h() = dD ( B , E) dD ( B , D) = dD ( B , E) +

1
log
2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Reduced Extremal Distance

165

increases to E / as 0. This proves the right-hand inequality. By Theorem


III.2.2 , we have

1
E g(z) = U (z) =
log
d( ),
|z

|
E
g ) as
where is a probability measure and g(z) = g (z, ). If f = ze2(g+i
in the proof of Theorem 3.2, then for z B

log | f | = 2g(z) + log 2 E 2 log(1 ) + log .


Thus if log 1/r = 2 E + 2 log(1 ) log ,
dD ( B , E) d f (D) ( Br , D) = dD ( Br , D)
1
(2 E + 2 log(1 ) log ).
=
2
That proves the left-hand inequality.

See Exercise 2 for a proof of Theorem 3.2 using logarithmic potentials only.
The next result introduces a metric found in Bonk, Koskela, and Rohde [1998]
and Balogh and Bonk [1999].
Fix z 0  and let

!
dist (z 0 , w) = inf
ds : is a curve in  and z 0 , w

be the euclidean distance in  from z 0 to w.


Theorem 3.5. Let : D  be a normalized univalent function, (0) = 0
and  (0) = 1, and let E {  : dist (0, ) R}. Then
1
log R.
(3.3)
2
Equality holds in (3.3) if  is a disc of radius R, with radial slits removed.
 (0, E)

Proof. We may assume that  is an analytic curve. Since (0) = 0 and


|  (0)| = 1,

1
1
log .
 (0, E) = lim d ( B , E)
0
2

By the extension rule, we may suppose that E = {z  : dist (0, z) R}


and by another application of the extension rule, we may suppose that
dist (0, ) R for all  and E = {  : dist (0, ) = R}. Define a
metric on  by
(z) =

1
.
dist (0, z)

CY535/Garnett

0 521 47018 8

166

January 27, 2005

15:41

V. Applications and Reverse Inequalities

If : [0, L)  is a curve parameterized by arc length with (0) = 0, then


| (s)| dist  (0, (s)) s.
Any curve connecting B to E in  \ B can be extended to a curve
connecting 0 to E by adding a line segment of length . Thus


R
1
R
(z)ds =
(z)ds
ds = log

| |
s
and


2 (z)d xd y
\B

<|w|<R

1
R
d A(w) = 2 log ,

|w|2

where B = {z : |z| < }. Therefore


d ( B , E) = d\B ( B , E)
and

R
1
log
2


1
1
1
log

log R.
(0, E) = lim d ( B , E)
0
2

(3.4)

By Theorem IV.4.2, equality holds in (3.4) when E is the circle of radius R and
 is the disc of radius R with radial slits removed.

Corollary 3.6 (BaloghBonk). Let (z) be a normalized univalent function,
(0) = 0 and  (0) = 1, in D and let R > 0. Then there is a constant C,
independent of and R, such that
1
#
"
|  (r )|dr > C1/2 .
Cap D :
0

Proof. Set  = (D). By an approximation, we may suppose that  is an


analytic Jordan curve. By the GehringHayman inequality, Exercise III.16,
1
|  (r )|dr K dist (0, ( ))
0

for some constant K independent of and . Corollary


3.6 now follows from


Theorems 3.2 and 3.5 with R = /K and C = K .
For a set E D, (3.2) shows that capacity can be replaced by length in
the statement of Corollary 3.6. See Exercises III.22 and III.23 for precursors of
Corollary 3.6.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Teichmllers Modulsatz

167

4. Teichmllers Modulsatz
Let  = {R1 < |z| < R2 } and let J  be a continuum that separates  into
disjoint ring domains 1 , 2 with {|z| = R1 } 1 and {|z| = R2 } 2 .

R1

R2

1
2
Figure V.1 Teichmllers Modulsatz.
By the serial rule
mod(1 ) + mod(2 ) mod() =

R 
1
2
log
,
2
R1

(4.1)

and equality holds in (4.1) if J is a circle {|z| = r }. Theorem 4.1 says that if
equality approximately holds in (4.1) then J is approximately a circle centered
at 0.
Theorem 4.1 (Teichmller). If
mod(1 ) + mod(2 )

R 
1
2
log

2
R1

(4.2)

and if > 0 is sufficiently small, then there is a C < independent of  and


such that
sup J |z|
1
1 + C log .
(4.3)
inf J |z|

In [1938] Teichmller derived his Modulsatz, Theorem 4.1, from his special
Modulsatz, Theorem 4.2 below, and we do the same.
Let G be a simply connected domain such that 0 G. For small define
M = M (G) = mod(G \ B(0, )).

CY535/Garnett

0 521 47018 8

168

January 27, 2005

15:41

V. Applications and Reverse Inequalities

As in (3.1),
M +

log
log 
M  +
,
2
2

(4.4)

if  < . Let : D G be a conformal mapping with (0) = 0. Then for


small

M (G) = mod D \ 1 (B(0, )) ,


so that by the proof of Lemma 3.1,
lim M +

log
1
=
log |  (0)|.
2
2

log |  (0)| the reduced modulus of G. If G C is simWe call M(G) =


ply connected and G, the reduced modulus of G is defined by
1
2

M(G) =

1

1
log |
(0)|,
2

where : D G is a conformal map with (0) = . It follows from Example III.1.1 that 2 M(G) = G when G. If J is a continuum and if G 1
and G 2 are disjoint components of C \ J such that 0 G 1 and G 2 , then
by (4.1)
M(G 1 ) + M(G 2 ) 0,
and equality holds if J is a circle {|z| = r }.
Theorem 4.2. If
M(G 1 ) + M(G 2 )

(4.5)

and if > 0 is sufficiently small, then there is a C < independent of J and


such that
sup J |z|
1
1 + C log .
(4.6)
inf J |z|

By inequality (4.4), Theorem 4.1 is a corollary of Theorem 4.2.


Proof. We follow Pommerenke [1991]. We assume J is not a circle. Let
(z) =


n=1

an z n

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Teichmllers Modulsatz

169

be a conformal map of D onto G 1 . Also let

(z) =

b1 
+
bn z n
z
n=0

be a conformal map of D onto G 2 . Then by (4.5)




 b1 
 2 .
log 
a1 

(4.7)

Inequality (4.7) is the main character in the proof of Theorem 4.2.


By the proof of the area theorem, Lemma I.4.2, we have


n|an |2 = Area(G 1 ) |b1 |2
n|bn |2 .

n=1

n=1

Therefore
|b1 |2 |a1 |2 |b1 |2 +




n |bn |2 + |an |2 .
n=2

In particular, |a1 | < |b1 |, because J is not a circle. Set |z| = r = |a1 /b1 |.
Then by (4.7), e2 r < 1 and
(1 r 2 ) log(1/(1 r 2 )) C1 log(1/),
so that by the CauchySchwarz inequality
)
+

 r 2n
|(z) a1 z|2
n|an |2
n
n=2

n=1



|b1 |2 |a1 |2 log
Thus

1
1
C1 |b1 |2 log .
2

1r

)
inf |z| inf |(z)| |b1 | 1 C2

zJ

|z|=r

1
log

+
A()

(4.8)

when is small, because by (4.7)


|a1 |
1 2 .
|b1 |
Similarly, on |z| = r
+

2 )




b
1
1
1
(z)
+ b0 
n|bn |2 log
C1 |b1 |2 log .

z

1 r2
n=1

CY535/Garnett

0 521 47018 8

170

January 27, 2005

15:41

V. Applications and Reverse Inequalities

But then by (4.7),


supzJ |z b0 | sup|z|=r |(z) b0 |



|b1 | 1 + C1 log 1 B(),

(4.9)

when is small. Therefore J lies inside the circle centered at b0 of radius B()
and outside the circle centered at 0 of radius A(). Consequently
1
|b0 | B() A() = C3 |b1 | log ,

and with (4.8) and (4.9) this gives (4.6).



Corollary 4.3. Let R > 1, let  = {1 < |z| < R} be the annulus with boundary
curves 1 = {|z| = 1} and 2 = {|z| = R}, let be a curve in  joining 1 1
to 2 , and assume
sup | arg z| = .

(4.10)

If < 0 < , there is a constant C = C(0 ) such that if 1 < R e then


d\ (1 , 2 )

C2
1
log R +
,
2
log 1

(4.11)

while if R > e then


d\ (1 , 2 )

C
2
1
log R +

.
2
log R log log R + log 1

2
1

R2
2

Figure V.2 , and their reflection about |z| = R.

(4.12)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Teichmllers Modulsatz

171

By the extension rule,


1
log R
(4.13)
2
and equality holds in (4.13) if is the radial slit [1, R]. Corollary 4.3 says that
if equality almost holds in (4.13) then is almost a radial segment.
d\ (1 , 2 )

Proof. We first prove (4.11) in the case R e and then derive (4.12) for R > e
from (4.11).
Write = d\ (1 , 2 ) and assume
<

log R
+ .
2

(4.14)

 and 
Let 
be the extensions of  and by reflection through {|z| = R}, so
) = 2mod(). Then by the symmetry
that 
joins z = 1 to z = R 2 and mod(
 \
rule (see Exercise IV.3(a)) the family 
 of symmetric curves in 
joining a
 satisfies (
 ) = 2. Set
point ei 1 to R 2 ei 
B = e2

2 /log

The map
z exp

i log z 
log R

 \
sends 
to the slit ring 1 \ [1, B] bounded by  and B = {Bz : z }
and cut along the positive real axis. The map also sends 1 and {|z| = R 2 } to
the slit [1, B] from  to B and sends 
 to the family of closed curves in 1
separating  from B. Therefore for sufficiently small,
2

2
+ ...
mod(1 )
2
log R
log R
log B
C.
2
Let (z) be a conformal map from D to the domain inside of  such that
(0) = 0. Let r > 0 be so small that 2 = B({r < |z| < 1}) 1 and set
3 = 2 \ (1 ). Then
1

1
log
,
mod(2 ) =
2
r
while
1

1
log
O(r ),
mod(3 )
2
Br

CY535/Garnett

0 521 47018 8

172

January 27, 2005

15:41

V. Applications and Reverse Inequalities

because

1 (3 ) = D \ 1 B({|z| r }) D \ {|z| Br O(r 2 )}.

Therefore for r small,


mod(1 ) + mod(3 ) mod(2 ) C.

Let J = (B)1  = 1 ( B1 ). Then by (4.15) and Theorem 4.1


1
sup |z|/inf |z| 1 + C log .
J

(4.15)

(4.16)

By (4.10) we have
0

 {e log R |z| e log R },


0

B {Be log R |z| Be log R },


and
0

e log R c(0 )Be log R

with c(0 ) < 1 because R e. By Koebes theorem J {|z| c (0 ) < 1} so


that by (4.16) and the Schwarz lemma
1

sup |z|/inf |z| 1 + C log .



Therefore by definition of 

C log

and
C

2
log 1

Now assume R > e. The map z log z sends  to a region log  bounded by
{Rez = 0}, {Rez = log R}, a curve from 0 to {Rez = log R}, and + 2i. Let
n be the integer such that n 1 < log R n and let n be the region bounded
by {Rez = 0}, {Rez = log R}, , and + 2 ni. Let E n = n {Rez = 0}
and let Fn = n {Rez = log R}. Then by the parallel rule and conformal
invariance
n
n
1

=
.
dn (E n , Fn )
d1 (E 1 , F1 )
d\ (1 , 2 )

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Boundary Conformality and Angular Derivatives

173

By the R e case of (4.11) and a rescaling, we have


dn (E n , Fn )

(/n)2
1
log R + C
2 n
(log n/)3

and so
d\ (1 , 2 )

C
2
1
log R +

.
2
log R log log R + log 1

Slightly weaker forms of (4.3) and (4.11) are easier to prove and sufficient
for most applications. See Exercise 4.
Teichmllers [1938] paper actually shows that the inequality (4.3) is sharp,
except for the choice of constant. For his example Teichmller formed a Riemann surface (which is C ) by fixing q > 1, taking two domains D1 = D and
D2 = {1 < |z| }, and making the identifications:
(i) ei D1 with eiq D2 , for 0 | | < /q,
(ii) ei D1 with ei D1 , for /q | | < , and
(iii) ei/q with ei/q and with ei .
By the uniformization theorem this gives two domains G 1 and G 2 such that
J = C \ (G 1 G 2 ) is a slit cardioid. Teichmller showed (4.6) is sharp by
varying the parameter q. It follows that (4.3) is also sharp.

5. Boundary Conformality and Angular Derivatives


Let  be a simply connected domain, let , and let F :   be
a conformal mapping. In this section we study two kinds of local boundary
behavior:
(a) The conformality of F at , and
(b) The existence of a non-zero angular derivative F  ( ).
We seek geometric conditions on  that are necessary and sufficient for either
(a) or (b) to hold.
Let C. For [0, 2 ), (0, /2), and > 0, define the truncated
cone
 (, ) = {z : | arg(z ) | < , 0 < |z | < }.
Definition 5.1. We say that  has an inner tangent with inner normal ei
at  if for every (0, /2) there is an = () > 0 so that
 (, ) .

(5.1)

CY535/Garnett

0 521 47018 8

174

January 27, 2005

15:41

V. Applications and Reverse Inequalities

When = /2 we say  has [vertical] vertical inner normal.

 ( )

Figure V.3 Truncated cone at .


When  has an inner tangent, the inner normal ei may not be unique.
For example, at each point on a slit there are two inner tangents with opposite
normal directions and the Jordan domain D {0 < arg z < 3/2} has inner
tangent at 0 with inner normal ei for all [/2, ]. Nevertheless, when 
has an inner tangent, we fix one ei for which (5.1) holds and write
 ( ) =  (, ).
If  has an inner tangent at and if
lim

 ( ) z

f (z) = A,

we say the function f has nontangential limit A at .


Definition 5.2. Suppose F :   is a conformal mapping and suppose 
has an inner tangent at . We say F is conformal at if F has a nontangential limit
F( )

lim

F(z)

arg

F(z) F( )
z

 ( ) z

and if the limit


A =

lim

 ( ) z

(5.2)

exists for every (0, /2).


If F is conformal at and if | | < /2, then as z the image arc
F({arg(z ) = }) is asymptotic to the ray {arg(w F( )) = + A }. In
other words, F preserves angles between nontangential rays. If F is conformal

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Boundary Conformality and Angular Derivatives

175

at , then F() has an inner tangent at F( ) and F 1 is conformal at F( ).


If  is a Jordan domain, if F : D  is a conformal mapping, and if  has a
tangent at F( ), then by Theorem II.4.2, F is conformal at . Moreover, in this
case the convergence z need not be restricted to cones  ( ). In general,
F can be conformal at D even when  does not have a tangent at F( ).
See Exercise VII.12 or Theorem 5.5 below.
Definition 5.3. Suppose F :   is a conformal map defined on  and
suppose  has an inner tangent at . We say F has angular derivative
F  ( ) if for all (0, /2) the two limits
F( )

lim

 ( ) z

F(z)

and
F  ( )

lim

 ( ) z

F(z) F( )
z

(5.3)

exist and are finite.


Lemma 5.4. Suppose F :   is a conformal map defined on  and suppose  has an inner tangent at . Then F has an angular derivative at
if and only if F  has a finite nontangential limit at .
Proof. If F  has nontangential limit L, then by integration F has finite nontangential limit F( ) and by the fundamental theorem of calculus (5.3) holds
with F  ( ) = L . Conversely, if F has an angular derivative at and if z n
nontangentially, then uniformly in some disc {w : |w| < },
gn (w) =

F(z n + ( z n )w) F( )
(w 1)F  ( ).
zn

Therefore
lim F  (z n ) = lim gn (0) = F  ( )
and F  has nontangential limit F  ( ) at .

It is clear that if F has a non-zero angular derivative at then F is conformal at


and A = arg F  ( ). Furthermore, F 1 also has a non-zero angular derivative
at F( ). On the other hand, F can be conformal at  without having an
angular derivative at . See Exercise VI.8.
However, the conformality of F at does yield some information about the
difference quotients (5.3). Suppose  has an inner tangent with vertical inner

CY535/Garnett

176

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

normal at 0  and suppose F is conformal at 0. Form the functions


)
+
i
F(z) F(0)

f (z) = log
.
z
F(i) F(0)
Then Im f converges to 0, as 0, uniformly on compact subsets of the half
annulus
A = {z : 1/2 < |z| < 2} H.
Since Re f (i) = 0, it follows that Re f converges to 0 uniformly on compact
subsets of A. Therefore
F(z) F(0)
i
lim

= lim e f = 1
0
z
F(i) F(0) 0
uniformly on compact subsets of A. Consequently F has an angular derivative
at 0 if and only if the vertical limit
lim

F(i) F(0)
i

exists. This fact also follows from Lindelfs theorem, Exercise II.3(d).
For the rest of this section we assume
:H
is a conformal mapping from upper half-plane H onto a domain  such that
0  and we assume has nontangential limit
lim

 (0) z0

(z) = 0.

(5.4)

It will often be convenient to transform both H and  (H) to strips.


The standard strip
S = {z : |Imz| < /2}
is mapped by the function (z) = iez onto H so that (+) = 0 and
() = . Then
(z) = 1 (z) = i/2 log (iez )

(5.5)

 defined by
is a conformal map of S onto the region 
 = {i/2 log w : w }


or

}.
 = {iez : z 

With this correspondence the cones 1 (0) in H or in  are transformed to


half-strips
S = {z : |Imz| < /2 and Rez > 0},

(0, /2)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Boundary Conformality and Angular Derivatives

177

. With this transformation (5.4) holds if and only if


in S or in 
lim

S z+

Re(z) = +

for all (0, /2), and condition (5.2) holds if and only if


Im(z) Imz = A
lim
S z+

for all (0, /2). Similarly,  has an inner tangent at 0 with a vertical
normal if and only if for each (0, /2) there is x > 0 such that
.
x + S 

 (0)

i 2

(5.6)

i 2

log(w)



R x

x + S

0
Figure V.4 Transforming the half-plane version to the strip version.

i 2

The first theorem gives local geometric conditions on a region  = (H)


that are necessary and sufficient for to be conformal R. The theorem is
from Ostrowski [1937]. For convenience we treat only the special case = 0,
(0) = 0, and A = 0 in (5.2); the other cases reduce to this case by translations
and rotations.
Theorem 5.5 (Ostrowski). Suppose  is a simply connected domain in C and
suppose 0 . Let : H  be a conformal map having nontangential
limit (0) = 0. If is conformal at 0 and if (5.2) holds with A0 = 0, then 
has an inner tangent at 0 with a vertical inner normal and
lim

R x0

dist(x, )
= 0.
x

(5.7)

Conversely, if  has an inner tangent at 0 with a vertical inner normal and if


(5.7) holds, then there exists a conformal map of H onto  having nontangential limit (0) = 0 such that is conformal at 0 and the limit A0 in (5.2) is
0.

CY535/Garnett

178

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

The comb in Figure V.5 has tines [xn i, xn ] with xn = xn 0. If


is a conformal map of the upper half-plane onto the complement of the comb
(including ) and if (0) = 0, then by Ostrowskis theorem is conformal at
0 if and only if limn xn+1 /xn = 1.
0

xn

xn i
Figure V.5 A comb region.
Proof. If (0) = 0 and if the limit in (5.2) is 0, then for all (0, ) the
ray {r ei : r > 0} and its image are asymptotic and therefore  has an inner
tangent at 0 with a vertical inner normal. If (5.7) fails, there is > 0 and
a sequence xn 0 such that dist(xn , ) xn for all n. We may suppose
xn > 0 and xn  by (5.1). The sequence of harmonic functions
xn z
u n (z) = arg 1
(xn z)
is then a normal family on
A = B(1, ) {z H : 1 < |z| < 1 + },
and for 0 < < /2, u n (z) converges to 0 on A {z : < arg z < }.
Fix z = 1 i 2 A . Then u n (z) converges to 0, so that arg 1 (xn z) < 0 for
large n. Because xn z  and 1 () = H, this is a contradiction and (5.7) is
established.
Conversely, assume  has an inner tangent at 0 with a vertical inner normal
and assume (5.7) holds. Let Er be the component of  {z : |z| = r } such
that ir Er and let  be a conformal map of  onto D. By the proof of
Carathodorys Theorem I.3.1, applied to  (instead of 1 ), we conclude that
lim (z)

zEr
r 0

exists in D. Composing 1 with a linear fractional transformation, we obtain


a conformal map of H onto  such that
lim 1 (z) = 0.

zEr
r 0

(5.8)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Boundary Conformality and Angular Derivatives

179

Because conformality is a local property of , we may suppose  contains iR+ = {i y : y > 0} and lim z | 1 (z)| = +. In terms of the
 and that if is given by (5.5) then
strip regions this means that R 
1 (R) is a curve in S along which Re 1 (x) + as x + and
Re 1 (x) as x . Condition (5.7) becomes
) = 0,
lim dist(z, 

(5.9)

zS
x+

 such that
the set Er is replaced by the component Fx of {z : Rez = x} 
Fx R = , and (5.8) translates to
lim Re 1 (z) = +.

(5.10)

zFx
x+

Take S with Im = /2. For < 1, set B = {z : |z | < }, and let


 satisfying U FRe = . We claim there is
U be the component of B 
an absolute constant C so that when U =
 1

1 2
1
.
(5.11)
diam (U ) C log

U
FRe

1 (U )

Figure V.6 Distortion near the ends of vertical crosscuts.


Indeed, by the extension rule and the annulus example
d
 (U , R) dC (B , R) C +

1
1
log .
2

(5.12)

By conformal invariance and the extension rule,


1
1
1
d
 (U , R) = dS ( (U ), (R)) dS ( (U ), L),

(5.13)

where L = i/2 + R is the component of S such that


/ L. Let  be the
1
family of curves in S separating the connected set (U ) from L. Using the
spherical metric, as in Exercise IV.6, we obtain
1
= S () C(diam 1 (U ))2 .
dS ( 1 (U ), L)

(5.14)

CY535/Garnett

180

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

Now (5.11) follows from (5.12), (5.13), and (5.14). The same estimate (5.11)
also holds if S and Im = i/2.
Finally, if > 0 and if z S then dist(z, ) < 2 for Rez is sufficiently
large by (5.9). Thus by (5.11)
1 1
dist( 1 (z), S) < C(log ) 2

and
1 1
|Imz Im 1 (z)| C(log ) 2 .

(5.15)

In other words = 1 ( S ) {Rez > x()} consists of two curves situated


1
outside S , where = 2C(log 1/) 2 . Moreover, by (5.10) +. By the
maximum principle and (5.15),
1 1
|Im(w) Imw| C(log ) 2

when w M + S if M > 0 is sufficiently large. Because > 0 was arbitrary,


this proves that has nontangential limit 0 at 0 and that (5.2) holds with limit
A0 = 0.

Ostrowskis theorem gave a geometric condition at p  which is necessary and sufficient for the conformal map : H  from the disc or halfplane to  to be conformal at 1 ( p). The next theorem, due to Rodin and
Warschawski [1976] and independently to Jenkins and Oikawa [1977], gives
an extremal length condition that is necessary and sufficient for the conformal
map to have non-zero angular derivative at 1 ( p). In Section 6 we will reinterpret this extremal length condition more geometrically. Because the existence
of a non-zero angular derivative depends on the domain  and not on the choice
of the conformal map , we make the following definition:
Definition 5.6. If  is simply connected and if 0 , we say  has a positive
angular derivative at 0  if there is a conformal map of H onto  which
has nontangential limit (0) = 0 and which has angular derivative  (0) such
that 0 <  (0) < .
Theorem 5.7 (JenkinsOikawa, RodinWarschawski). Suppose  is a simply connected domain such that 0  and iR+ = {i y : y > 0} . For
r > 0, let Er denote the component of  {z : |z| = r } containing the point
ir . Then  has positive angular derivative at 0 if and only if
(a)  has an inner tangent at 0 with a vertical inner normal, and
(5.16)
(b) lims<r 0 d (Er , E s ) 1 log rs = 0.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Boundary Conformality and Angular Derivatives

181

If the Ostrowski condition (5.7) holds, then Theorem 5.7 is still true if condition (b) is replaced by the weaker
(b ) limn>m+ d (E 2m , E 2n )

log 2nm = 0.

See Exercise 5.
Er

Es
0

Figure V.7 Positive angular derivative at a tine.


Proof. First assume that  has a positive angular derivative at 0. Then (a) holds
by Theorem 5.5. To establish (b), we transform the problem using (5.5). The
 = {i/2 log w : w }
conformal map of the standard strip S onto 
satisfies
lim (z) z = C

Rez+
zS

for each (0, /2), where C is a (finite) real number; or equivalently,


lim

Rew+
wS

w 1 (w) = C

(5.17)

 {z : Rez = x} such
for each (0, /2). Let Fx denote the component of 
that x Fx . Then (b) will hold if and only if
d
 (Fs , Ft ) |s t| 0

(5.18)

, we may suppose that the constant C in (5.17)


as s, t +. By translating 
is zero. That will not affect (5.18).
With > 0 and C as in (5.11) choose so that C(log 1/)1/2 < . By
Theorem 5.5, we can choose x so that x + S  and B(, )  = if
S satisfies Re > x . Let be the component of {z : |z | = } 
such that FRe = . Then if U is the component of  \ that meets

CY535/Garnett

0 521 47018 8

182

January 27, 2005

15:41

V. Applications and Reverse Inequalities

B(, ), we have by (5.11)


1 1
diam 1 (U ) = diam 1 ( ) C(log ) 2 < .

By (5.17), we may also suppose that |Re 1 (z) x| < for all z S Fx
and all x > x . Consequently,
sup |Re 1 (z) x| < , x > x .

zFx

Thus for t > s > x ,


{z : s < Rez < t + , |Imz| < /2} 1 (Fs ) 1 (Ft ),

(5.19)

whereas
{z : s + < Rez < t , |Imz| < /2}
( 1 (Fs ) 1 (Ft )) = .

(5.20)

Then by the extension rule and conformal invariance


t s 2 < dS ( 1 (Fs ), 1 (Ft )) = d
 (Fs , Ft ) < t s + 2
and (5.18) follows.
Conversely, assume (a) and (b) hold. We first establish (5.7). If (5.7) fails,
then as in the proof of Theorem 5.5 we may suppose there is a region
A = {z : Imz > 0, 1 < |z| < 1 + } B(1, )
and a sequence xn 0, xn > 0 such that xn A . Let rn = (1 + )xn and
sn = (1 )xn . Then by the extension rule and conformal invariance,
d (Ern , E sn ) d A ({|z| = 1 + }, {|z| = 1 })
1
1+
1
rn
= C() + log
= C() + log ,

sn
for some C() > 0. Letting xn 0 then contradicts (b), and therefore (5.7)
holds.
Now choose as in Theorem 5.5, define by (5.5), and recall that Fx is
 {Rez = x} containing x. We claim that the variation of
that component of 
Rez along the image of Fx satisfies
Vx = sup |Re 1 (u) Re 1 (v)| 0

(5.21)

u,vFx

as x +. If (5.21) is true, then by a comparison with rectangles as in (5.19)


and (5.20) we obtain via (5.18) that
lim

Rew+
wS

Rew Re 1 (w)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Boundary Conformality and Angular Derivatives

183

exists for all (0, /2). Then by Theorem 5.5


lim

Rew+
wS

Imw Im 1 (w) = 0,

and (5.17) follows. But (5.17) clearly implies that has a positive angular
derivative. Thus to complete the proof of Theorem 5.7, we need only prove
(5.21).
To prove (5.21) note that by (5.18) there exists x > 0 so that for all r, s, t
with x < r < s < t < we have approximate equality in the serial rule:
d
 (Fr , Fs ) + d
 (Fs , Ft ) d
 (Fr , Ft ) .
Let f be the conformal map of the subregion of S between 1 (Fr ) and
1 (Ft ) onto the rectangle R = {z : 0 < Imz < , 0 < Rez < d
 (Fr , Ft )},
so that f 1 (Fr ) is the left vertical end Ir of R and f 1 (Ft ) is the right
vertical end It of R. Then Is = f 1 (Fs ) divides R into two regions U1 and
U2 , and
dU1 (Ir , Is ) + dU2 (Is , It ) d R (Ir , It ) .

(5.22)

Is
Ir

U1

U2

It

Figure V.8 Image of a vertical crosscut.


 with ends 
It and
Reflect R through its top edge, obtaining a rectangle R
Ir and 
2 separated by a curve 
 contains regions U
1 and U
Is which connects as to
R
as + 2i. By the symmetry rule,
Ir , 
Is ) + dU2 (
Is , 
I t ) d R (
Ir , 
It ) /2.
dU1 (

(5.23)

s connects as R to as + 2i, Theorem 4.1 can now be applied to


Because T
 to derive (5.21) from (5.23).

{e z : z R}
Exercises 4 and 5 include proofs of (5.21) without Teichmllers Modulsatz.
When  is smooth, Theorem 5.7 and the estimates in Sections 2 and IV.6
yield necessary and sufficient geometric conditions for the existence of a positive angular derivative.

CY535/Garnett

184

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

Corollary 5.8. Suppose  is a simply connected domain given by


 = {r ei : 1 (r ) < < 2 (r ); r > 0},
where j are absolutely continuous in (0, 1), 1 (r ) < /2 < 2 (r ), 1 (0) = 0,
2 (0) = , and
1
j (r )2 r dr < ,
0

for j = 1, 2. Let (r ) = 2 (r ) 1 (r ). Then  has a positive angular derivative at 0 if and only if

1
1
1

dr
(5.24)
r (r ) r
0
exists.
A related result can be found in Rodin and Warschawski [1976].
Proof. Since 1 (0) = 0 and 2 (0) = ,  has an inner tangent with a vertical
inner normal. Let
1 (ex ) + 2 (ex )

,
2
2

m(x) =

(x) = 2 (ex ) 1 (ex ),


and
 = {x + i y : m(x) (x)/2 < y < m(x) + (x)/2}.

 onto . Let Fx = 
 {z : Rez = x}.
Then (z) = iez is one-to-one from 
For s, t sufficiently large, by (IV.6.2) and (2.2) we have the estimates
t  2
t
t
1 
m (x) + 12
(x)2
1
1
d x d
dx +
d x.
 (Fr , Fs )
(x)
s (x)
s (x)
s
By our assumptions on j ,

lim

s,t+ s

1 
m  (x)2 + 12
(x)2
d x = 0.
(x)

(5.25)

Thus by Theorem 5.7,  has a positive angular derivative at 0 if and only if



t
1
1

d x = 0,
(5.26)
lim
s,t+ s
(x)
and by a change of variables, (5.26) is equivalent to (5.24).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Conditions More Geometric

185

6. Conditions More Geometric


In this section we seek more geometric conditions on a simply connected domain
 that are equivalent to the existence of a non-zero angular derivative. We
concentrate on the normalized domain
 = {i/2 log w : w }

. From this perspective,  has an inner tangent
and on the point z = + 
, and  has a positive
and vertical inner normal if and only if (5.6) holds for 
. If 
 is a simply
angular derivative if and only if (5.6) and (5.18) hold for 
 has an angular
connected domain satisfying (5.6) and (5.18), we say that 
derivative at +. We do not require that the map z ez be one-to-one on
.

We begin by observing that under mild assumptions, condition (5.26)
 {z : s < Rez < t} and
equivalent to a statement about areas. Set s,t = 
Ss,t = {z : s < Rez < t, |Imz| < /2} and note that then
t
( )d x = Area(Ss,t \ s,t ) Area(s,t \ Ss,t ).
s

Now if  has an inner tangent at 0 with a vertical inner normal, then for x
sufficiently large

(x) 2.
2
Hence if has constant sign on (s, t), then

 
 t 1
1



 s
dx
1
2



.
(6.1)
Area(Ss,t \ s,t ) Area(s,t \ Ss,t )
2 2
2
In particular, if
) + Area(
 \ S) <
Area(S \ 
then


t
 1
1 

dx = 0

s,t s  (x)

lim

and (5.26) holds.


When (5.25) fails, upper and lower estimates can still be found in terms of
 is replaced by a smaller region with a Lipschitz boundary. Let M > 0.
areas if 
Recall that an M-Lipschitz graph is the graph of a function h for which
|h(s) h(t)| M|s t|

CY535/Garnett

0 521 47018 8

186

January 27, 2005

15:41

V. Applications and Reverse Inequalities

 is a region such that R 


. Let T M denote the
for all s, t R. Suppose 
collection of isosceles triangles T such that
(i) T ,
(ii) T has base on R, and
(iii) T has sides with slope M,
and define
M =


%
{T : T T M }.

(6.2)


 M is the largest domain such that R 
M 
 and 
Then 
M consists of
two M-Lipschitz graphs.



\
M

S

M




Figure V.9 M-Lipschitz subregion.


 R is a simply connected domain such that
Theorem 6.1. Suppose 
 {z : Rez < 0} = S {z : Rez < 0}.


(6.3)

 be the region defined by (6.2) above.


M 
Let 
 has an angular derivative at +
 M ) < . Then 
(a) Suppose Area(S \ 
if and only if
 M \ S) < .
Area(
 has an angular derivative at +
 M \ S) < . Then 
(b) Suppose Area(
if and only if
 M ) < ,
Area(S \ 
provided that M > 8 .
The hypothesis (6.3) is not crucial because the existence of an angular derivative at + is a local property of  near +. See Marshall [1995] for an
example where M is small and (b) fails.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Conditions More Geometric

187

 M \ S) < , then
 M ) < and Area(
Proof. If Area(S \ 
lim dist(z, S) = 0

Rez+
M
z 

(6.4)

 has an angular derivative at +, then by


and (5.9) holds. Conversely, if 
Ostrowskis theorem, (6.4) holds. Thus we may assume (6.4) during the re {z : Rez = s} such
mainder of the proof. If Fs denotes the component of 
 has an angular derivative at + if and only if
that s Fs , then 
lim d
 (Fs , Ft )

s,t
s<t

t s
= 0.

(6.5)

To prove Theorem 6.1 we will obtain upper and lower bounds for the left side of
(6.5). The upper bound is given by Lemma 6.2, and the lower bound by Lemma
6.4.
 M is given by the two curves y = h 1 (x) + /2 and
Now suppose that 
y = h 2 (x) /2, < x < +, where h 1 > /2 and h 2 > /2. By
(6.4), lim x+ h j (x) = 0. As before, for any region U , let
Us,t = U {z : s < Rez < t}.
Lemma 6.2 (Sastry). Suppose |h j (x)| M 2 for s < x < t. Then
d
 (Fs , Ft )

 M )s,t
 M \ S)s,t
(4M 2 + 2)Area(S \ 
t s
Area(

2
(4M 2 + 2) 2

Sastry [1995] proved this lemma with different constants and used a piecewise constant metric. We will use a metric of the form = |u|, where it is
easier to estimate lengths.
Proof. To give an upper bound for extremal distance, we construct a metric that
gives a lower bound for the conjugate extremal distance. The idea is that the
 onto S is nearly the identity in S and it compresses the
conformal map of 

 S. The expansion and compression
regions near  \ S and expands near 
will be accomplished by imitating Beurlings metric from Section 2 in those
regions.
Because
d
 M )s,t (Fs , Ft ),
 (Fs , Ft ) d
 M (Fs , Ft ) = d(

CY535/Garnett

0 521 47018 8

188

January 27, 2005

15:41

V. Applications and Reverse Inequalities

without loss of generality, we may suppose (for notational convenience) that


 = (
 M )s,t . Define

 : h 1 (x) > 0 and h 1 /(2M 2 ) < y < h 1 },
U1 = {(x, y + /2) 
 : h 1 (x) < 0 and 2h 1 < y < h 1 },
U2 = {(x, y + /2) 
 : h 2 (x) > 0 and h 2 < y < h 2 /(2M 2 )},
U3 = {(x, y /2) 
 : h 2 (x) < 0 and h 2 < y < 2h 2 },
U4 = {(x, y /2) 
and
\
U = 

4

j=1 U j .



S

U1

Fs

U1
U2

U
U4

Ft
U4

U3


Figure V.10 Proof of Sastrys lemma.

 onto S which fixes


Consider the continuous map (x, y) (x, u(x, y)) of 


and is linear in y on  \ U . Thus

y
for (x, y) U ,

for (x, y) U1 ,
(y (h 1 + /2))/(2M + 1) + /2
u(x, y) = 2(y (h 1 + /2)) + /2
for (x, y) U2 ,

(y + h 2 + /2)/(2M 2 + 1) /2
for (x, y) U3 ,

for (x, y) U4 .
2(y + h 2 + /2) /2

. Note that
Set = |u| on 

1
|u|2 4(1 + M 2 )

(1 + M 2 )/(2M 2 + 1)2

on U ,
on U2 U4 ,
on U1 U3 .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Conditions More Geometric

Thus,

189




|u|2 d yd x

Ss,t

d yd x

4(1 + M 2 )Area(U2 U4 ) +

1 + M2
Area(U1 U3 )
(2M 2 + 1)2

1
 \ Ss,t ) 2Area(Ss,t \ 
)
Area(
2M 2
1
 \ Ss,t ).
)
= (4M 2 + 2)Area(Ss,t \ 
Area(
2
4M + 2

If is a curve connecting {y = h 1 (x) /2} to {y = h 2 (x) /2} and


, then
contained in 






|u||dz|  u dz  = .

Thus


d
 (Fs , Ft )

 |u|



2 d yd x

(inf |u||dz|)2

1
)
2 (t s) + (4M 2 + 2)Area(Ss,t \ 


1

Area( \ Ss,t ) .

4M 2 + 2

The next step is to derive a lower bound for the extremal distance in
 = j j where each j is an open arc with endpoints
M \ 
(6.5). Write 
l
r
 where Rez l < Rez r . For v C \ R, let Tv denote the triangle
zj , zj 
j
j
in T M with vertex v. There is a unique v j with Rez lj Rev j Rez rj so
that z lj , z rj Tv j . Then j Tv j and j consists of at most two line segments. If |Imz lj | |Imz rj |, let B j = {z : |z z lj | < Rez rj Rez lj }. Otherwise
let B j = {z : |z z rj | < Rez rj Rez lj }.
Lemma 6.3.
8
Area(B j )
M


z j




|Imz|  d x,
2

z = x + i y.

CY535/Garnett

0 521 47018 8

190

January 27, 2005

15:41

V. Applications and Reverse Inequalities

vj



z rj
S

z lj




Bj
Figure V.11 Proof of Lemma 6.3.
Proof. Suppose f is defined and continuous on [0, 1] with

1
0 x a,
f  (x) =
1 a x 1.
Then by elementary geometry or calculus
1
| f (x)|d x
0

is minimal when a =

1
2

and f (0) = 41 . Thus


1
1
| f (x)|d x .
8
0

If Imz > 0 on j , the map


)
(x, y)

x Rez lj
Rez rj Rez lj

y /2

M(Rez rj Rez lj )

transforms j into the graph of one such f . Thus





M(Rez rj Rez lj )2


M
|Imz| d x
=
Area(B j ).


2
8
8
z j
The inequality of the lemma for Imz < 0 follows by a reflection about R.
The quantity


z j




|Imz|  d x
2

is the total area between j and S.


 using the
The next lemma gives a lower bound for the extremal distance in 



geometry of  M . When  =  M , it follows immediately from the discussion
at the end of Section 5.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Conditions More Geometric

191

Lemma 6.4. If > 0 there is an s0 < so that for s0 < s < t < ,
 M )s,t
(M 8 )Area(S \ 
t s

M
 M \ S)s,t
(M + 8 )Area(

.
M 2
 M \ S)s,t 0 and
Proof. Suppose 0 t s . Then as s, t , Area(

Area(S \  M )s,t 0 as s, t by (6.4) and the inequality follows. Now
 given by
suppose that t s . Let be the metric on 


 M j Bj ,
for z 
= 1
.
0
elsewhere on 
d
 (Fs , Ft )

The metric will provide the lower bound for the extremal distance. First we
s,t . By Lemma 6.3
compute the -area of 


2 d xd y
d xd y
s,t


Ss,t

 M \ S)s,t Area(S \ 
 M )s,t +
Area(

Area(B j )

(6.6)

8
8
 M \ S)s,t (1
 M )s,t .
)Area(
)Area(S \ 
M
M
 connecting Fs to Ft , with s < t. Note that
Now suppose is a curve in 
(1 +

Rez rj Rez lj

4
max ||Imz| /2| 0
M z j

(6.7)

as j +, by (6.4). Thus by deleting an initial and terminal portion of ,



 M j B j and that
if necessary, we may suppose that begins and ends in 
if meets j , then s < Rez lj < Rez rj < t.
M




2

Fs
1

Ft
3

Figure V.12 Proof of Lemma 6.4.


Thus we can write =
either


k

k where {k } are disjoint subarcs of such that

CY535/Garnett

0 521 47018 8

192

January 27, 2005

15:41

V. Applications and Reverse Inequalities


 M j B j , or
(i) k 
(ii) k meets some j and connects FRezl to FRezrj with s < Rez lj < Rez rj < t.
j

If (i) holds, then k |dz| is at least equal to the change in Rez along k . If (ii)
holds, then
 M B j )) Rez rj Rez lj .
length(k (
Indeed, if z lj is the center of B j and if k j , then the shortest curve from
 is the straight line from z l to , since |Im | |Imz l |. Since
FRezl to in 
j
j
j
k must intersect B j , it must have length at least the radius of B j , namely
Rez rj Rez lj . A similar argument works if z rj is the center of B j .
Together with (6.7) this implies, for s, t sufficiently large, that

|dz| t s .

We conclude by (6.6) that for t s ,


d
 (Fs , Ft )

(t s )2

(t s) + (1 + 8
M )Area( M \ S)s,t (1

8

M )Area(S \  M )s,t

 M is Lipschitz and (6.4) holds, we have, as s, t , both


Because 
 M )s,t /(t s) 0. Hence for
 M \ S)s,t /(t s) 0 and Area(S \ 
Area(
s, t sufficiently large with t s ,
 M )s,t
(M 8 )Area(S \ 
t s

M 2
 M \ S)s,t
(M + 8 )Area(

.
2
M
concluding the proof of Lemma 6.4.
d
 (Fs , Ft )

M ) <
We now complete the proof of Theorem 6.1. We have Area(S \ 

if and only if Area(S \  M )s,t 0 as s, t +. Likewise, the condition
 M \ S)s,t 0 as s, t +.
 M \ S) < is equivalent to Area(
Area(
 M \ S) < , by Lemmas 6.2 and 6.4,
 M ) < and Area(
Thus if Area(S \ 
lim d (Fs ,
s,t 

Ft ) (t s)/ = 0,

and hence by the discussion at the beginning of the proof of the Theorem 6.1,
 has an angular derivative at +.


CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Conditions More Geometric

193

 has an angular derivative at +, then by


 M ) < and if 
If Area(S \ 
(6.5) and Lemma 6.2,
 M \ S)s,t 0
Area(
as s, t +. That proves part (a) of Theorem 6.1. If M > 8 and if
 M \ S) < then by (6.5) and Lemma 6.4,
Area(
 M )s,t 0
Area(S \ 
as s, t +, and that proves part (b) of Theorem 6.1.

There exist regions that do not satisfy the hypotheses of Theorem 6.1, but
have angular derivative at +. See Exercise 8.

Notes
Theorem 1.1 was conjectured by Denjoy in [1907] and proved by Ahlfors in
[1930]. Beurling [1933] and Carleman [1933] give different proofs. Ahlfors paper, which predated the theory of extremal length, uses the lengtharea method.
He obtained a version of (IV.6.4), with a slightly different (r ) and a constant
larger than 8 , but his version was good enough for (1.1). See also Warschawski
[1942] for another variant. We have followed Ahlfors original argument, but
starting from (IV.6.4). Carlemans proof is in Appendix G, and Beurlings proof
is in Exercise 1. Lord Rayleigh [1871] gave the three-dimensional version of
Beurlings upper estimate (2.2) for the resistance in a conductor using the parallel rule, assuming the conductor is a solid of revolution about an axis (m  (x) 0
in the two dimensional version) with smooth boundary. Maxwell [1891] gave
an account of Rayleighs method with a slight improvement of the upper bound.
Reduced extremal distance is close to a method used by Beurling in his thesis
and explained in Exercise III.23. Theorem 3.2 is from Ahlfors and Beurling
[1952], but Corollary 3.6 is a recent result by Balogh and Bonk [1999]. An
extension of Theorem 3.2 to finitely connected domains is given in Theorem H.9
of Appendix H. Jenkins [1970] has a slightly weaker but easier analogue of
Theorem 4.1. See Exercise 4.
Bertilsson [1999] proved (4.11) directly and derived Teichmllers Modulsatz as a corollary. He also showed that the estimate 2 /(log 1/) in (4.11) is
sharp. See Exercise 10.
The proof of Ostrowskis theorem given here is based on Warschawski
[1967]. The estimate (5.11) is due to J. Wolff [1935]. The problem of finding

CY535/Garnett

194

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

geometric conditions equivalent to the existence of a non-zero angular derivative is very old, dating back at least to Ahlfors [1930]. Warschawski [1967],
Rodin and Warschawski [1977], Baernstein [1988], and their references provide
a history of the problem. The proof of Theorem 6.1 is from Marshall [1995].
Theorem 6.1(a) was proved by Burdzy [1986], using Brownian excursions, in
the half-plane form given in Exercise 9, which is the context of the original
angular derivative problem. Half of Theorem 6.1(a), in the strip form given
in the text, was proved by Rodin-Warschawski [1986] and the other half was
proved by Sastry [1995]. For other proofs of the half-plane version of Theorem
6.1(a), see Carroll [1988] and Gardiner [1991]. Theorem 6.1(b) was proved by
Marshall [1995] and answers a question of Burdzy [1986].

Exercises and Further Results


1. In [1933] Beurling derived Theorem 1.1 from his own inequality, proved in
Exercise III.22 above. With the notation of the proof of Theorem 1.1, take
D j = G j {|z| > 1}, assume | f | 1 on D j , but take z j D j such that
log | f (z j )| e. For R > Max|z j | set
M j (R) =

sup

D j {|z|=R}

Beurling shows there is R0 such that if


log M j (R) =

| f (z)|.

R  j (R)
R0

then for all R > R0 ,


n

j=1

1
< 2.
j (R)

(E.1)

and (E.1) clearly implies (1.1).


To prove (E.1), write j (R) = (z j , |z| = R, D j {|z| < R}). Then
e j (R) log M j (R).
Now let C j be the family of curves in D j connecting z j to D j {|z| = R}
and define


L j = sup inf (( )) : is conformal on D j and Area(D j ) = .
C j

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

195

Beurlings inequality from Exercise III.24(b) gives


2

j e1L j ,
so that
2

log M j (R) e L j .
Under the logarithm D j is mapped to a region

A j 2 log R.

(E.2)
D j

with area A j and


(E.3)

Using
(z) =


1
2 log z
Aj

gives
L 2j

R 2

log

log R log
Aj
|z j |
Aj
R0

(E.4)

if R0 Max|z j |2 , and together (E.2), (E.3), and (E.4) imply (E.1).


2. This exercise outlines a proof of Theorem 3.2 that does not use Greens
functions.
(a) Let be the equilibrium distribution
for E D and let U be its

logarithmic potential. Set V (z) =
log

1
2

U (z) + U (1/z) and prove

1
1
1
1
V (z) + log
log
.
1 + |z|
2
|z|
1 |z|

(b) Use Greens theorem to prove





1
1
2
,
|V | d xd y = E + log
D (V )
2


where  = {z D : V (z) > (log )/2}.
(c) Set
u =

V + 21 log 1
E + 21 log 1

Then = |u | as the extremal metric for d({z : V = (log )/2}, E).


(d) Deduce Theorem 3.2 from (c) and (a).
3. Compare the estimates of harmonic measure given by Theorem IV.5.2,
Corollary 3.3, and Exercise III.22 when E is an arc on the boundary of
a Jordan domain.

CY535/Garnett

0 521 47018 8

196

January 27, 2005

15:41

V. Applications and Reverse Inequalities

4. The following variations on Theorem 4.1 and Corollary 4.3 are from Jenkins
[1970].
(a) Let  be the annulus bounded by 1 = {|z| = 1} and 2 = {|z| = R}
where R > 1, and let be a curve joining 1 to 2 such that
0 = inf arg z < sup arg z = .

Prove
d\ (1 , 2 )

1
3
log R
+
+ O(5 ).
2
108 2 log R

Hint: If a curve  \ joins 1 to 2 and enters { 3 < arg z <




ds
2
(log R)2 + .
9
|z|

2
3 }, then

(b) Let R be the rectangle {0 < x < L; 0 < y < } with left vertical side
E 1 and right vertical side E 2 . Suppose that T is a curve connecting the
horizontal sides of R such that
a = inf Rez < sup Rez = a + .
T

Then T divides R into two components 1 and 2 such that E j  j .


Prove
d1 (E 1 , T ) + d2 (T, E 2 )

3
L

+ O(5 ).

27 3

Hint: Bound d j (E j , T )1 from below using the metric

1
2 2 , if a + /3 < x < a + 2/3;
+ /9
(z) =
1
otherwise.
,
(c) Use (b) and the symmetry rule to prove Theorem 4.1 with the weaker
1
estimate 3 in (4.3).
5. Suppose  is a simply connected domain such that {i y : y > 0}  and
0 . For r > 0 let Er be the arc of  {|z| = r } such that ir Er .
Assume  has inner tangent at 0 with vertical inner normal and satisfies the
Ostrowski condition
dist(x, )
= 0.
lim
x
R x0

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

197

(a) Prove that


r
1
log = 0,

s
where the limit is restricted to 1 < r/s < C and r, s 0. Hint: Use estimate
(5.11) together with a normal families argument on {z : 1 < |z| < C} H}
as given in the comments after Definition 5.3.
(b) Prove
lim d (Er , E s )

lim d (Er , E s )

s<r 0

r
1
log = 0

if and only if
lim

n>m+

d (E 2m , E 2n )

(c) Use part (a) to prove (5.21).


6. (a) Show that if

x0

and

 (x)2
dx <
(x)

t
lim

s,t+ s

1
log 2nm = 0.

1
1

(x)


dx = 0

then
lim (x) = .

x+

 S then
(b) Show that if lim x+ (x) = and if 




t
2  s 1 1 d x 
= 1,
lim
s,t Area(Ss,t \ s,t )
 S using Area(s,t \ Ss,t ).
and prove a similar limit holds if 
7. Let  = {(x, y) : y > min (|x|/ log(1/|x|), 1)}. Then  =  has a tangent
at 0.
(a) If + is a conformal map of H onto  with (0) = 0, then + has infinite
angular derivative at 0.
(b) If is a conformal map of H onto C \  with (0) = 0, then has
zero angular derivative at 0.

(c) Prove that if  is replaced by 1 = {(x, y) : y > min |x|/(log |x|)2 , 1 },


then the conformal maps + and have non-zero angular derivatives at 1.

CY535/Garnett

198

0 521 47018 8

January 27, 2005

15:41

V. Applications and Reverse Inequalities

 = {(x, y) : |y m(x)| < 1 }.


8. Set 
2


(a) If |m  (x)| 1, 0 |m  (x)|2 d x < , and 0 |m(x)|d x = , then by
 has an angular derivative at +, but Theorem 6.1
(2.2) and Theorem 5.7 
does not apply.


(b) If |m  (x)| 1, 0 |m(x)|d x < , and 0 |m  (x)|2 d x = , then by
 has an angular derivative at +, but this fact can not be
Theorem 6.1 
deduced from (2.2).
9. Let  be a simply connected domain containing the positive imaginary
axis and suppose 0 . Let h M denote the smallest M-Lipschitz function whose graph is contained in .
(a) (Burdzy [1986]) Suppose
1
h (x)
h M >0 M
d x < .
(E.5)
x2
1
Then  has a positive angular derivative at 0 if and only if
1
h (x)
h M <0 M
d x > .
x2
1

(E.6)

For example if H  then  has a positive angular derivative at 0 if and


only if (E.6) holds.
(b) (Marshall [1995]) There is an M0 < so that if M > M0 and if h M
satisfies (E.6) then  has a positive angular derivative at 0 if and only if
h M satisfies (E.5). For example, if  H, then  has a positive angular
derivative at 0 if and only if (E.5) holds.
 show that Theorem 6.1 can be derived
(c) If (z) = iez is one-to-one on 
from parts (a) and (b), but with 8 replaced by a larger constant. See Marshall
[1995].
10. Let
 = {(x, y) : 0 < x < 1, h(x) < y < h(x) + 1},
where h(x) = log(1 + x/)/ log(1 + 1/) and set E = {Rez = 0} 
and F = {Rez = 1} . Let be a conformal map of  onto a rectangle
R = {(x, y) : 0 < x < 1, 0 < y < H },
with Re = 0 on E and Re = 1 on F. Then
d (E, F) =

1
1
=
.
2
H
 |Re| d A

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

Show that

 |Re|

2d A

199

1 C log1/ and hence


2

d (E, F) 1 + C

2
log 1

Hint: Write Re = x + u(x, y). By Greens theorem, it is enough to show





1
2


2

C
udy
|u|
d
xd
y
.



log 1 
Since |h  | is small, it is enough to prove this for v(x, y) u(x, h(x) + y)
defined on the unit square S. Relate v on S to |v| on S by integrating the
gradient along lines with slope 1. See Bertilsson [1999].

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

VI
Simply Connected Domains, Part One

The results of this chapter are largely elementary and independent of the three
preceding chapters. First we prove the classical [1916] F. and M. Riesz theorem.
Next we prove three almost everywhere theorems, due to Privalov, Plessner,
and McMillan. Their proofs all rely on elementary measure theory and a cone
construction (see Figure VI.2) which is a geometric variation on the proof of
Lusins theorem. We then give Makarovs elegant proof that on every simply
connected domain
 , > 1,
and the extension by Pommerenke that almost everywhere identifies a non-zero
angular derivative at with the existence of a cone at ( ).

1. The F. and M. Riesz Theorem


By definition an analytic function f (z) on D is in the Hardy space H p for
0 < p < , if

p
| f (r ei )| p d = || f || H p < .
sup
0<r <1

See Appendix A for a brief introduction to the Hardy spaces H p .


Theorem 1.1. Let  be a domain such that  =  is a Jordan curve and let
be a conformal map from D onto . Then the curve  is rectifiable if and
only if  H 1 . If  H 1 , then
||  || H 1 = () = 1 ().

(1.1)

Proof. By Carathodorys theorem, extends to a homeomorphism from D to


. First assume  H 1 . Let {0 = 0 < 1 < . . . . < n = 2 } be any partition
200

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The F. and M. Riesz Theorem

201

of [0, 2 ]. Then
n 
n 






 i j

(e ) (ei j1 ) = lim
(r ei j ) (r ei j1 )
r 1

j=1

= lim

r 1

j=1
n  j





j=1

j1



 (r ei )ir ei d 

(1.2)

||  || H 1 .
But () is the supremum, over all partitions, of the left side of (1.2). Therefore
 is rectifiable and
() ||  || H 1 .
Conversely, assume  is rectifiable. Then given r < 1, choose a partition
{0 < 1 < . . . . < n } of [0, 2 ], so that
n 




(r ei j ) (r ei j1 ) (r ) ,
j=1

where r = ({|z| = r }). Write


(z) =

n 




(zei j ) (zei j1 ).
j=1

Then is subharmonic on D and by Carathodorys theorem is continuous


on D, so that
sup (z) = sup (ei ) ().

Thus

|  (r ei )|d = (r ) (r ) + () + .

Therefore  H 1 and the equality (1.1) holds.

Theorem 1.1 is equivalent to a theorem about harmonic measure. Assume


that  =  is a rectifiable Jordan curve and let A  be a Borel set. Write
A = (E) and z 0 = (0). Then by Carathodorys theorem
1
|E|.
2
When A is an arc, Theorem A.1 from Appendix A and the proof of (1.1) yield



i
| (r e )|d =
|  (ei )|d,
(1.3)
1 (A) = 1 ((E)) = lim
(z 0 , A, ) =

r 1 E

CY535/Garnett

0 521 47018 8

202

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

because  H 1 , and if (1.3) holds for arcs then (1.3) also holds for all Borel
sets A . Consequently
(A) = 0  1 (A) = 0.
Conversely,
1 (A) = 0  (A) = 0,
because by Corollary A.2,



{ : |  ( )| = 0} = 0.

Thus when  is rectifiable, harmonic measure for  and linear measure on 


are mutually absolutely continuous,
 1  .
This argument proves the following:
Theorem 1.2 (F. and M. Riesz). Let  be a simply connected plane domain
such that  =  is a rectifiable Jordan curve and let : D  be conformal.
Then  L 1 (D). For any Borel set E D,

1 ((E)) =
|  |d,
E

and for any Borel set A ,


(A) = 0 1 (A) = 0.

(1.4)

Theorem 1.2 is equivalent to Theorem 1.1 because (1.3) holds only if  H 1 .


In [1936] Lavrentiev gave a quantitative version of (1.4). See Exercise 2 and
(5.1) below. See Exercise 4 for an example of a Jordan domain  and a set
E R  such that 1 (E) > 0 but (E) = 0.
Theorem 1.1 has another important geometric consequence. Let be a conformal map from D to a domain  such that  =  is a rectifiable Jordan
curve. Then by Fatous theorem, Theorem 1.1, and Corollary A.2,  has a
non-zero nontangential limit  ( ) at almost all D. But by calculus, has
a non-zero angular derivative at D whenever  has a non-zero nontangential limit at . Therefore has a non-zero angular derivative  ( ) at 1
almost every ( ) . In this way Theorem 1.1 is an improvement, almost
everywhere, on the pointwise theorems II.4.2 and V.5.7.
Corollary 1.3. Let  be a domain such that  =  is a rectifiable Jordan
curve and let be a conformal map from D onto . Then has a non-zero

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Privalovs Theorem and Plessners Theorem

203

angular derivative at and  has a tangent at ( ), both for Lebesgue almost


every D and for 1 almost every ( ) .
In general, can have a non-zero angular derivative at a point , so that 
then has an inner tangent at ( ) even though  does not have a tangent at
( ). Bishop [1987] exhibited a domain bounded by a non-rectifiable Jordan
curve  such that for almost every D, is conformal at but  has no
tangent at ( ). The domain is { f (x) < y < f (x), 0 < x < 1} where f (x)
is continuous, f (0) = f (1), f (x) > 0 on (0, 1), and f has no finite or infinite
derivative at any point. See Exercise 9.
If  is not a Jordan curve, then will not be one-to-one on D. However,
the nontangential limit ( ) can still exist, and in that case the existence of an
inner tangent at ( ) depends on the point D and not just on the point
( ) . But since is one-to-one on D, there can be at most two points
1 (w) at which is conformal.

2. Privalovs Theorem and Plessners Theorem


Let D, let > 1 and let 0 < h < 1. The truncated cone h is
h ( ) = {z : |z | < (1 |z|) < h}.

h
2 sec1 ()

Figure VI.1 The truncated cone h .


Let u(z) be a harmonic or meromorphic function on D.
Definition 2.1. The function u is nontangentially bounded at if there exist
M < , > 1, and 0 < h < 1 such that
|u(z)| M on h ( ).
Notice that the definition of nontangential boundedness only requires that u(z)
be bounded on a single truncated cone, whereas the definition of nontangential
convergence requires that u(z) has a limit through every cone  ( ).

CY535/Garnett

0 521 47018 8

204

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

Theorem 2.2 (Privalov). Suppose E D and suppose u(z) is nontangentially bounded at each E. Then u(z) has a nontangential limit at almost
every E.
Proof. Let > 0. By a little measure theory there exists a compact set K E
with |E \ K | < and there exist constants , h, and M such that |u| < M on
%
U=
h ( ).
K

To prove the theorem it is enough to show u has a nontangential limit almost


everywhere on K . Let Ui be one component of U and let i be a conformal
map of D onto Ui . Then u i is a bounded harmonic function on D and by
Fatous theorem, u i has a finite nontangential limit almost everywhere on
D. Because and h are fixed, Ui is a rectifiable Jordan curve. See Figure VI.2.
For Lebesgue almost every K Ui , i is conformal at i 1 ( ) and u i
has a finite nontangential limit at i 1 ( ), both by the F. and M. Riesz theorem.
But when i is conformal at i 1 ( ), u i has a nontangential limit at i 1 ( )

if and only if u has a nontangential limit at . Since K = i K Ui , we

conclude that u has a finite nontangential limit at almost every K .

i
K
Ui

Uj

K
h
Figure VI.2 Cone domains Ui .
Many proofs in this book will be simple variations on Figure VI.2 and the
previous argument. Define a cone domain to be a domain of the form
%
%
U=
 ( ) or U =
h ( ),
K

where K D is compact, > 1, and 0 < h < 1. Cone domains will be recurrent in this chapter and again in Chapter X.
Theorem 2.3. Let f (z) be a meromorphic function on D and let E D have
|E| > 0. If for each E there is = ( ) > 1 such that
lim

 ( ) z

f (z) = 0,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Accessible Points

205

then f (z) 0 in D.
Proof. Let K E be compact with |K | > 0 and take the cone domains Ui and
the maps i as in the proof of Privalovs theorem. Then f i is bounded and
analytic on D, and by the F. and M. Riesz theorem, f i has nontangential
limit 0 on a set of positive measure. Then by (1.4), f i 0 so that f 0. 
Nontangential approach regions are necessary in Theorems 2.2 and 2.3.
Bagemihl and Seidel [1954] showed that there is a non-constant analytic function on D having radial limit 0 almost everywhere on D. See Exercise 5.
Definition 2.4. Let f (z) be a meromorphic function on D.
point

D is a
A
h
Plessner point for f if for all > 1 and all 0 < h < 1, f  ( ) is dense in C.
Theorem 2.5 (Plessner). Let f (z) be a nonconstant meromorphic function on
D. Then there are pairwise disjoint Borel subsets N , G, and P of D such that
D = N G P and
(a) |N | = 0,
(b) At each G, f has a finite nontangential limit f ( ) and f ( ) = 0, and
(c) Each P is a Plessner point for f .
Proof. Let P be the set of Plessner points for f and let E = D \ P. Then
P and E are Borel sets, and the theorem is equivalent to the assertion that f
has a finite non-zero nontangential limit at almost every E. Let {wn } be a
countable dense subset of C and set
"
E n = : there exist > 1, 0 < h < 1,
#

and > 0 such that  f wn | on  h ( ) .

Then E = n E n . By Privalovs theorem 1/( f (z) wn ) has a nontangential


limit almost everywhere on E n and hence f has a nontangential limit at almost
every E. By Theorem 2.3, applied to f and to 1/ f , the limit is finite and

non-zero almost everywhere.

3. Accessible Points
Now let  be a simply connected domain and let : D  be a conformal mapping. We do not assume  is a Jordan curve. Nevertheless, has a
nontangential limit ( ) at almost every D, because the function
z
(z) (0)

CY535/Garnett

0 521 47018 8

206

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

is bounded and analytic on D. When has a finite nontangential limit at D


we define ( ) to be that nontangential limit. Thus we have defined ( ) at
almost every D. Better yet, by Corollary V.3.6,
1

 
 (r )dr <
(3.1)
0

except on a capacity zero subset of D. If (3.1) holds then has a radial limit
at , and it follows from Lindelfs theorem, Exercise 3(d) of Chapter II, that
has a nontangential limit at . Therefore the set where ( ) is not defined is
smaller than many sets of linear measure zero; it has capacity zero and Hausdorff
dimension zero. We say w  is an accessible point of  if w is the endpoint
of an open arc .
Lemma 3.1. If has a nontangential limit ( ) at D, then ( ) 
and ( ) is an accessible point of . Conversely, every accessible w  is
a nontangential limit w = ( ).
Proof. Suppose has nontangential limit at D. Then clearly ( ) .
But if ( )  then there is z D with ( ) = (z), and that means that
1 is not single valued in some neighborhood of (z). Therefore ( ) 
and clearly ( ) is an accessible point.
For the converse, suppose w  is the endpoint of an arc . The preceding argument above shows that lim zw | 1 (z)| = 1. Because has nontangential limits a.e. and 1 ( ) is an arc, Theorem 2.3, applied to (z) w,
shows that lim zw 1 (z) = D exists. It then follows from Lindelfs

theorem that has nontangential limit w at .
Write
A = {accessible points of }.
Then A is the range of the boundary function ( ), defined except on a set of
capacity zero, and A is dense in .
Theorem 3.2. Let  be a simply connected domain. Then the set A of accessible
points is a Borel subset of  and for all z ,

If E A, then

(z, A, ) = 1.

(3.2)

z, E,  = 1 (z), 1 (E), D .

(3.3)

In particular,
(z, E, ) = 0 | 1 (E)| = 0.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Cone Points and McMillans Theorem

207

Proof. Clearly, A is a Borel set. We may assume z = (0). Let rn 1, and set
n = ({z : |z| < rn }), and let f C(). Then by the definition of harmonic
measure, by Lemma 3.1, and by dominated convergence,



1
1
f ( )d( ) = lim
f ((rn ei ))d =
f ((ei ))d. (3.4)
n 2
2

The case f = 1 of (3.4) gives (3.2). Each side of (3.3) defines a Borel measure and since Borel measures are determined by their actions on continuous

functions, (3.4) also implies (3.3).

4. Cone Points and McMillans Theorem


Recall from Chapter V that if has a non-zero angular derivative at , then
is conformal at . Using conjugate functions or Ostrowskis theorem and
Corollary V.5.8 it is easy to make an example where has no non-zero angular
derivative at but is conformal at . See Exercise 8 and Exercise V.7. On
the other hand, Theorem 6.1 below will show that if is conformal at all
E D, then has a non-zero angular derivative at almost every E.
We permanently write
"
#
G = D : has non-zero angular derivative at
and

"
B = D : has a nontangential limit at , and
#
lim inf |  (z)| = 0, for all > 1 .

(4.1)

 ( ) z

Theorem 4.1. Let be a univalent function on D. Then G B = and


|G B| = 2.
Proof. Because has a nontangential limit almost everywhere, this theorem is
nothing but Theorem 2.3 and Plessners theorem, applied to the function  . 
We call a point w  a cone point of  if w is the vertex of an open
isosceles triangle T . Every cone point is an accessible boundary point. In

CY535/Garnett

208

0 521 47018 8

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

review, we know that the following implications hold pointwise:


has non-zero angular derivative at

is conformal at

 has an inner tangent at ( )

( ) is a cone point of ,
but that all three converse implications fail pointwise. However, Theorem 6.1
will show that these four conditions are almost everywhere equivalent. Write


K = K () = cone points for  .
Then it is clear that (G) K .
Theorem 4.2 (McMillan). Let  be a bounded simply connected domain. Then
K is a Borel set, with -finite 1 measure, and when E K ,
(E) = 0 1 (E) = 0.

(4.2)

Moreover, at almost every w K ,  has an inner tangent.


Proof. Let {L n } be the countable set of lines having rational slope and rational
y intercept. For each n, put w K n if w  \ L n and w is the vertex of an
open isosceles triangle Tn (w)  such that
(a) Tn (w) has base on the line segment L n {|z| n},
(b) Tn (w) has vertex angle n , and
(c) Tn (w) has height h n (w) satisfying n1 h n (w) n.

Then K = n K n , where K n is compact, so that K is an F set. Note that

dist(K n , L n ) n1 . Let n = K n Tn (w). Then n , and by (a), (b), and
(c), n has finitely many components n, j . Each component n, j is bounded
by a Jordan curve n, j , and except for endpoints on L n , these Jordan curves are

pairwise disjoint. We have K (n, j \ L n ) and K has -finite 1 measure.
If we assume the line L n is horizontal, then n, j is the union of an arc of L n
and the graph of a Lipschitz function defined on that arc. In particular n, j is
rectifiable, and  has an inner tangent at 1 almost every point of n, j , and
thus at 1 almost every point of K .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Cone Points and McMillans Theorem

209

w
U

Tn (w)
Ln
Figure VI.3 n-Lipschitz subregion.
Then to prove (4.2) we can assume E n, j \ L n . Suppose 1 (E) > 0.
Then by the F. and M. Riesz theorem, (z, E, n, j ) > 0 for all z n, j , and
therefore (z, E, ) > 0.
Conversely, suppose E n, j \ L n is a compact set such that 1 (E) = 0.
We may suppose L n is the real axis. Set
%
Tn (w).
U=
E
1
units in the direction of n, j to a parallel line L n . When
Translate L n by 10n

w E, let Tn (w) Tn (w) be the isosceles triangle having vertex w, base on

. Set
L n , and vertex angle n+1
%
Tn (w).
V =
E

units in the direction of n, j to a parallel line L n . Repeat


Translate L n by
this construction with L n and with triangles Tn (w) now having bases on L n

and vertex angle n+2


. Set
%
Tn (w).
W =
1
5n

Then W has finitely many components. We may assume E has no isolated


points so that W is not a triangle. We can also assume W , and hence U and V ,
is connected by replacing E by a subset if necessary. Then
W V U n, j 
and like n, j , U, V , and W consist of a segment of L n , L n , L n , respectively,
and a Lipschitz graph over that segment. Each component k of V \ E is an
arc having endpoints in E, and these endpoints are also the endpoints of a
component k of W \ E and a component k of U \ E. Each of the arcs

CY535/Garnett

0 521 47018 8

210

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

k , k , k is a polygonal arc having at most three sides and angles bounded


below.


k
C \ Uk

k V \ E
k W \ E

L n

L n L
n
Figure VI.4 Nested regions in McMillans theorem.
Let Uk be the unbounded component of C \ (k k ). Only one k satisfies
k L n = . Then for each E
inf (z, k , Uk ) = > 0,
k

uniformly in k, by a normal families argument for instance. Therefore


inf (z,  \ E,  \ V ) > 0

U \E

(4.3)

and by the maximum principle


inf (z,  \ E, ) > 0.

U \E

But U is rectifiable and 1 (E) = 0, so by the F. and M. Riesz theorem,


(z, E, U ) = 0. Thus (z,  \ E, ) on U V \ E and
= inf (z,  \ E, ) .
V \E

Therefore

(z,  \ E, ) = (z,  \ E,  \ V ) +

(,  \ E, )d\V (z, )

(z,  \ E,  \ V ) + (z, V,  \ V )
= (1 )(z,  \ E,  \ V ) + ,
for all z  \ V . But taking z U \ E then gives
(1 ) + ,
so that = 1 and (E) = 0.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Cone Points and McMillans Theorem

211

In [1969] McMillan proved Theorem 4.2 and also the following striking result, known as the McMillan twist point theorem: At almost every
D \ G, arg((z) ( )) is unbounded above and unbounded below on
every curve D having one endpoint . When arg((z) ( )) is so unbounded, ( ) is called a twist point of . In particular, for almost every
D \ G, the radial image {(r ) : 0 < r < 1} is a rectifiable arc in  with
endpoint ( ), but arg(w ( )) has no upper or lower bound along this arc.
McMillans twist point theorem will be proved in Appendix I. Example 4.3
below will show that we can have |D \ G| = 2 .
When  is a Jordan curve, is one-to-one and the notions of twist point or
inner tangent depend only on  near the point w = ( ) . But suppose
 is not a Jordan curve and let w . If there exists 0 1 (w) such that
on some curve D with endpoint 0 ,
arg((z) (0 )) is unbounded above and below,

(4.4)

then (4.4) holds for every 1 (w) and for every curve D ending at 0 .
Now assume  is a Jordan curve and let 1 and 2 be the two simply
connected components of C \ . Fix p j  j , let j : D  j be a conformal
mapping with j (0) = p j and write j = ( p j , ,  j ). Set


Tn() = w  :  has a tangent at w .
If w Tn() then by Theorem II.4.2, j is conformal at j 1 (w) for j = 1, 2.
Consequently, by Theorem 4.2 the three measures 1 , 2 , and 1 are in the
same measure class on Tn = Tn(),
Tn 1  Tn 1  Tn 2  Tn 1 .
Theorem 6.3 below will show, without using the twist point theorem, that on
 \ Tn(), 1 2 . Write Twi() for the set of twist points of . In [1987],
C. J. Bishop also made an example of a Jordan curve for which
1 (Twi()) = 1,
but

2 Tn() Twi() = 0.

Bishops construction is outlined in Exercise 10.


Example 4.3 (The von Koch snowflake). Let 0 1 . . . be the increasing
sequence of Jordan domains defined
as follows: 0 = 0 is the equilateral

triangle with vertices {0, 1, 21 + i 23 }; n = n consists of 3 4n segments


Jn, j of length 3n ; and n+1 is obtained from n by replacing the middle third

CY535/Garnett

0 521 47018 8

212

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

In, j of each Jn, j by the two remaining sides of the unique equilateral triangle
Tn, j satisfying In, j = Jn, j Tn, j and Tn, j n = . Clearly n+1 n .

0

1
2
Figure VI.5 The von Koch snowflake.

5


Then  = n is a Jordan domain and  =  = lim n , where the limit is
taken in the Hausdorff metric. The Jordan curve  =  is called the

von Koch
4
and
0
<

snowflake. The Hausdorff dimension of  is log
log 4  < . See
log 3
log 3

Appendix D. Let be the conformal map of D onto . If  has an inner tangent


at ( ), then for any > 0,  contains a truncated cone with vertex ( ) and


with aperture . Consequently ( ) n ( n ). But E = n ( n )
2
is the union of a sequence of linear images of the Cantor set and dimE = log
log 3 .
Therefore the set of points of  with an inner tangent in  has linear measure
0 and by Theorem 4.2, harmonic measure 0 with respect to . The exterior
 = C \  has inner tangents only on a countable subset of , and
domain 
. Thus the conformal
hence on a set of harmonic measure 0 with respect to 
 respectively have angular derivatives only on
maps and 
of D onto  and 
a of measure 0 by Theorem 4.2. The theorems in Section 6 below imply that the
three measures 1 , 1 =  , and 2 = 
 are pairwise mutually singular
on .

5. Compression and Expansion


Let  be a simply connected domain, let be harmonic measure for some
fixed z 0  and let : D  be a conformal mapping. If  () > 0 and
 , then the conformal map compresses a set E D having full
harmonic measure into a set (E) having  ((E)) = 0, and expands a
set E  with (E) = 0 into a set (E) with  ((E)) > 0. On the other
hand, if   , then cannot compress a set E with |E| > 0 into a set with
 ((E)) = 0.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Compression and Expansion

213

If  is a rectifiable Jordan curve then by the F. and M. Riesz theorem,


 1  . In [1936] Lavrentiev made the more precise lower estimate
(E)

C log 1 ()
,
1 + | log 1 (E)|

(5.1)

when  is rectifiable, E , and dist(z 0 , ) 1. See Exercise 2 for the


proof.
If  is not rectifiable, no inequality like (5.1) can hold. Lavrentiev [1936]
exhibited a Jordan domain for which  1 . McMillan and Piranian [1973]
constructed a Jordan domain such that
1 .
Then Kaufman and Wu [1982] built, for any measure function of the form
1
,
2
a Jordan domain for which h . An even sharper result will be proved in
Chapter VIII. These examples show can compress a set E with |E| = 2 into
(E) where h ((E)) = 0.
On the other hand, cannot compress harmonic measure very much. By
Beurlings projection theorem (see Corollary III.9.3)   1 . In the very in2
fluential paper [1973], Carleson improved Beurlings estimate to   for
some unknown , 21 < < 1. Then, in the paper that inspired much of this
book, Makarov [1985] proved the following theorem.
h(t) = t exp(| log t| ), 0 < <

Theorem 5.1 (Makarov). Let be harmonic measure for a simply connected


domain  and let 0 < < 1. Then
  .
However, cannot expand harmonic measure very much either. In [1972]
and [1981] ksendal proved 2 for plane domains of any connectivity,
and he conjectured that  for all > 1 and all domains. See Exercise
III.12 for a proof of ksendals result. Makarov, also in [1985], established the
ksendal conjecture for simply connected domains:
Theorem 5.2 (Makarov). Let be harmonic measure for a simply connected
domain  and let h(t) be a measure function such that
h(t)
= 0.
t0 t
lim

Then is singular to h , h .

CY535/Garnett

214

0 521 47018 8

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

In particular,  if > 1. Together Theorems 5.1 and 5.2 say that on


a simply connected domain, harmonic measure has dimension 1 and can
neither expand nor compress harmonic measure very much. Later Jones and
Wolff [1988] proved that for every plane domain,  for all > 1. The
JonesWolff theorem will be proved in Chapter IX.
Theorem 5.1 depends on Makarovs [1985] law of the iterated logarithm.
It will be proved in Chapter VIII. Here we give Makarovs elegant proof of
Theorem 5.2. We need a simple lemma.
Let E D and let S = {z k } be a sequence in D. We say S is nontangentially dense a.e. on E if there is > 1 such that for almost every E,
S  ( ).
Lemma 5.3. Assume {z k } is nontangentially dense a.e. on E D and let
be the conformal map from D onto a simply connected domain . Then set
wk = (z k ), rk = dist(wk , ), Bk = B(wk , 2rk ), and
%

V = 
Bk A,
where A is the set of accessible boundary points. Then
|E \ 1 (V )| = 0.

(5.2)

Proof. Suppose (5.2) is false and let Wk be the component of Bk  such that
wk W k .


Wk

wk

Bk

Figure VI.6 Proof of Lemma 5.3.


Since  is connected,
(wk , V, ) (wk ,  Bk , Wk ) c,
by the Beurling projection theorem; see Exercise 10 of Chapter III. Consequently u(z) = ((z), V, ) satisfies u(z k ) c. But u has nontangential limits a.e. and almost every E is the nontangential limit of a subsequence of

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Compression and Expansion

215

{z k }, so that
lim

 ( ) z

u(z) c

almost everywhere on E. On the other hand, by Theorem 3.2


u(z) = (z, 1 (V ), D),
so that u(z) has nontangential limit 1 (V ) almost everywhere on D. Consequently
1 (V ) c


almost everywhere on E, and (5.2) follows.


Proof of Theorem 5.2. Fix > 0. When z D and |z| > 21 , define


I (z) = D : |z | < 2(1 |z|) .
Then 1 |z|2 c|I (z)| and I (z) if and only if z 2 ( ).

I (z)

Figure VI.7 I (z) consists of the base points of all cones 2 containing z.
Let be the conformal mapping from D to , and set = 2. By Theorem 2.3,


lim inf   (z) <
 ( ) z

for almost every D, and the sets






E n = D : lim inf   (z) < n
 ( ) z

 
satisfy E n E n+1 and  E n  = 2. Every E n is covered by arbitrarily
small arcs I (z) such that
  
 (z) < n,
(5.3)
and
1 |z|2 < n ,

(5.4)

CY535/Garnett

216

0 521 47018 8

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

where n < /n is so small that


h(t)

< n+2
(5.5)
t
n2
whenever t < 4nn . By the Vitali covering lemma, Exercise I.9, there is a
sequence {z n, j } satisfying (5.3) and (5.4) such that
%


 En \
(5.6)
I (z n, j )= 0
j

and

|I (z n, j )| 2.

Then by (5.3), (5.4), and (5.5),




C
h 2|  (z n, j )|(1 |z n, j |2 ) n .
2
j

With fixed, we take {z k } =


Bk = B(wk , 2rk ), and

n {z n, j },

(5.7)

wk = (z k ), rk = dist(wk , ),

V =  (Bk )).
Then by (5.4) and Theorem I.4.3,
2rk 2|  (z k )|(1 |z k |2 ) 4,
so that (5.7) yields

h(2rk ) C.


Consequently, V =
m V m1 satisfies h (V ) = 0. But by (5.6), {z k } is nontangentially dense a.e. in D. Therefore
( , V 1 , ) = 1
m

by Lemma 5.3 and ( , V, ) = 1.

6. Pommerenkes Extension
In [1986a], Pommerenke extended Makarovs argument to obtain the following
improvement of Theorem 5.2. We recall the sets
"
#
G = D : has non zero angular derivative at ,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Pommerenkes Extension

217

"
B = D : has a nontangential limit at , but
#
lim inf |  (z)| = 0, for all > 1 ,
 ( ) z

and

"
#
K = cone points for  .

Then we know G B = , |G B| = 2 and (G) K .


Theorem 6.1 (Pommerenke). Let  be a simply connected domain and let
: D . Then there is a subset S (B) \ K such that
1 (S) = 0.

(6.1)

(S (G)) = 1.

(6.2)

and

In particular, the set P = (G) S has finite 1 measure and (P) = 1.


Consequently,
1 (K \ (G)) = (K \ (G)) = 0.
Proof. We follow the proof of Theorem 5.2. Again by the Vitali covering
lemma, there are {z n, j } such that
|  (z n, j )| < 2n3 ,

(6.3)

%


B \
I (z n, j ) = 0,
j

and

|I (z n, j )| 2.

Now take wn, j = (z n, j ), rn, j = dist(wn, j , ), Bn, j = B(wn, j , 2rn, j ), and
%
Vn =  ( Bn, j ).
j

Then as before, (6.3) yields




2rn, j
2|  (z n, j )|(1 |z n, j |2 ) c 2n ,
j

CY535/Garnett

0 521 47018 8

218

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

and
V =

*%

Vn

k nk

has 1 (V ) = 0. Set
S = V (B).
Then (6.1) holds for S. By Beurlings projection theorem, (wn, j , Vn , ) c,
and
%
{z n, j }
nN

is nontangentially dense on B. Therefore by Lemma 5.3


%


B \
1 (A Vn ) = 0
nk

for any k, where A is the set of accessible points. Hence




 B \ 1 (S) = 0,
and (6.2) then follows from Theorem 4.1. Since 1 (S) = 0, if we replace S by
S \ K then (6.1) and (6.2) still hold by Theorem 4.2 and (K \ (G)) = 0. By

Theorem 4.2, 1 (K \ (G)) = 0.
By McMillians twist point theorem, Appendix I, almost every w S is a
twist point of . From Theorems 6.1 and 4.2 we have the following corollary.
Corollary 6.2. Let  be a simply connected domain and let S  be the set
given in Theorem 6.1. Then
 1 (S) = 0 |G| = 2,
and
1 |G| = 0.
Theorem 6.3. Suppose  is a Jordan curve, and let 1 and 2 be the two
components of the complement C \ . Let E be a Borel subset of  such
that (z j , E, ) > 0 for z j  j , j = 1, 2, and let j | E be the restriction of
j = (z j , ,  j ) to E. Then
1 | E 2 | E
if and only if
1 (Tn() E) = 0,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Pommerenkes Extension

219

where Tn() is the set of tangent points of .


Browder and Wermer [1963] gave the first example of a Jordan domain for
which 1 (Tn()) = 0. See also Gamelin and Garnett [1971] and McMillan
and Piranian [1973].
Proof. Assume 1 (Tn() E) > 0. On Tn (), j  1  j by Theorem
4.2, so that for j = 1, 2, j | E (Tn()) > 0 and
1 | E 2 | E .
Now assume 1 | E 2 | E . The argument is exactly like the proof of Theorem V.1.1. We may assume dist(z j , ) 1. Let w . For 0 < t < 1, let Jj (t)
be any arc in  j {z : |z w| = t} such that Jj separates z j from w, and write
t j (t) = (Jj (t)).
Then by Theorem IV.6.2,

Because
have



1
dt
8
.
j (B(w, r )) exp

r t j (t)

1
2

(1 + 2 ) 4, (by CauchySchwarz) and 1 + 2 2, we


1
2
1
+
.
1
2

Therefore

 1
64
2dt
64
1 (B(w, r ))2 (B(w, r )) 2 exp
= 2 r 2.
t

(6.4)

By assumption, there is a compact set E N E such that j (E N ) > 0 and


1 (S)
2 (S) N 1 (S)
N
for all Borel S E N . Cover S E N by balls B(wk , rk ), wk , and rk < 1.
Then by (6.4),


rk j (S),
8 N
and hence

1 (S) j (S).
4 N

(6.5)

CY535/Garnett

220

0 521 47018 8

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

Let j be a conformal map of D onto  j , let G j D be the set where j


has non-zero angular derivative and let S = Sj be the set given by Theorem 6.1
that satisfies (6.1) and (6.2) for  j . Then 1 (S E N ) = 0 by (6.1) and (6.5).
Therefore

1 1 (G 1 ) 2 (G 2 ) E N > 0
so that by Theorem 4.2,

1 1 (G 1 ) 2 (G 2 ) E N > 0,

which proves the theorem because


1 (G 1 ) 2 (G 2 ) Tn().

The proof of Theorem 6.3 yields a purely geometric result about the tangent
points of a Jordan curve . Let w , take j (t) as in that proof, and set
(w, t) = max{| j (t)| : j = 1, 2}.
Then by Taylor series
2 (w, t) 2
1
1
2
+
+
,
1
2

and the proof of (6.4) yields


2
64
1 (B(w, r )) 2 (B(w, r ))
2 exp 2
r
r

2 (w, t)

dt 
.
t

(6.6)

But by Theorem 4.2 the left side of (6.6) is bounded below at 1 (or j ) almost
every point of Tn(). Consequently we have the following corollary.
Corollary 6.4. If  is a Jordan curve, than at 1 almost every tangent point
w of ,
1
dt
(6.7)
2 (w, t) < .
t
0
Carleson has conjectured that conversely,  has a tangent at almost every point
where (6.7) holds. Bishop and Jones [1994] has a result slightly weaker than
this conjecture. See the remarks following Theorem X.2.5 below.
We continue to assume 1 and 2 are the two complementary components
of a Jordan curve . Let K j be the cone points for  j , let G j be the set where
the conformal map j of D onto  j has non-zero angular derivative, and let
Sj  j \ K j be a set satisfying (6.1) and (6.2) with respect to  j . Note that

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

221

(G j ) Sj = . Consider the set

E = (G 1 ) S1 (G 2 ) S2 = (G 1 ) (G 2 )


S1 S2




(G 1 ) \ (G 2 ) S2 (G 2 ) \ (G 1 ) S1


S1 \ S2 ) S2 \ S1 ).

By (6.2), 1 (E) = 2 (E) = 1. By Theorems 6.3 and 6.1 we have the following
table.
Subsets of 

Measures

(G 1 ) (G 1 ) Tn()

1  1  2  1

S1 S2

1 2 , while 1 = 0


(G j ) \ (G k ) Sk
Sj \ Sk

k = 0, but j  1  j

K 1 K 2 \ (G 1 ) (G 2 )

k = 0 and 1 = 0
1 = 2 = 1 = 0

Examples that limit these results can be found in Exercises 8, 9 and 10.

Notes
Sections 1 through 4 are classical. Sections 5 is from Makarov [1985]. See
Pommerenke [1986a] and Bishop, Carleson, Garnett and Jones [1989] for the
results of Section 6.

Exercises and Further Results


1. (a) Let be the conformal mapping from D to a simple connected domain
, and assume  H 1 . Prove that  is a rectifiable curve, although not
necessarily a Jordan curve.
(b) Let K be a compact connected subset of C such that 1 (K ) < . Let
 j be the components of C \ K , and let j be a conformal mapping from D
onto  j . If  j is bounded, then j H 1 , while if  j and (0) = ,

CY535/Garnett

0 521 47018 8

222

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

then z 2 j (z) H 1 . Furthermore,


 2  
 j  d = 21 (K ).
j

See Alexander [1989].


2. Suppose  is a bounded Jordan domain with rectifiable boundary, and suppose z 0  is such that dist(z 0 , ) 1. Then for E ,
(z 0 , E)

C log 1 ()

.
1 + log 1 (E)

See Lavrentiev [1936]. Hint: Take z 0 = (0). Then () =


|  (0)| 1. By subharmonicity,

log |  |d 0

|  |ds and

so that by Jensens inequality






log |  | 2log+ |  |
1
  1
4 log  
1

= 4 log 1 ().
Write E = (S). We may assume 2 (E) = |S| and 1 (E) =
are small. Again by Jensens inequality,

1 (E)
1
,
log |  |d log
|S| S
|S|
so that
|S|

|S| log
1 (E)


D\S


S

|  |d

log |  |d 4 log 1 ().

3. We recall the HaymanWu theorem: There exists a constant C such that if


: D  is univalent and if L is a line, then


length 1 (L) C.
In Section 5 of Chapter I, we showed that C 4. Here we give in outline
ymas [1993] example that C 2 .
(a) If I is an interval in R and if 0 < < 1, let S I, be the longer circular
arc in H having the same endpoints as I and meeting
R at angles
and

. Let  I, be the unbounded component of C \ S I, S I,/2 . Then

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

223

for z I
(z, S I, ,  I, ) =

1
( + /2)
< + .
(2 /2)
2

(b) Given > 0, there is b > 0 and > 0 such that if I is a subarc of D
with |I | < b and J is a crosscut in D joining the endpoints of I with
sup (z, I, D) <
J

1
+ ,
2

then

length(J ) > (1 ) |I |.
2
Hint: The level set  = {z : (z, I ) = 21 + } is a circular arc, and if b and
are small enough, then length(J ) length() > (1 ) 2 |I |.
(c) Fix > 0, and let and b be from part (b). Construct open intervals
In R of bounded length such that for n = m,
In Im = ,
S In ,/2 S Im ,/2 = ,
and


% 

In  = 0.
R \

Write Sn = S In , , let Vn be the Jordan domain with Vn = In Sn , and set


%
 = {y < 0}
(Vn In ) .
Then  is locally rectifiable, so that if : D  is conformal, then by

the F. and M. Riesz theorem,  1 (R \ In ) = 0. Hence

 

 1 (Sn ) = 2.
Taking with |Im(0)| large gives | 1 (Sn )| < b for all n. Then by (a) and
(b),


length 1 (In ) > (1 ) | 1 (Sn )|


2
and thus

 


length 1 (R) =
length 1 (In ) (1 ) 2 .
4. There exists a simply connected Jordan domain  and a set E R 
such that 1 (E) = |E| > 0 but (z, E, ) = 0. Let 0 = (0, 1) (0, 1),

CY535/Garnett

0 521 47018 8

224

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

and E 0 = [0, 1] and form closed sets E n+1 E n and domains n+1 n
thusly: E n consists of 2n intervals In, j with |In, j | = an , 2n+1 an+1 < 2n a n ,
but lim 2n an > 0. Let In, j \ E n+1 be the middle open subinterval Jn, j In, j
bn

with |Jn, j | = an 2an+1 . Let bn 0 satisfy lim 2n e an = 0, and take


%

J n, j [0, bn+1 ] .
n+1 = n \
j

Then E = E n has |E| > 0 and  =


E , and for z ,

n is a Jordan domain such that

(z, E, ) lim (z, E n , n ) = 0


n

by Theorem IV.6.1 or Theorem VI.6.1. See Lohwater and Seidel [1948].


5. (a) Let D be an arc with non-empty interior and let f (z) be a meromorphic function on D such that
lim f (r ) = 0

r 1

for all . Prove f 0. Hint: Use Baire category and the proof of Privalovs theorem.
(b) If E D is an F set of the first Baire category in D, and if g(z) is
any continuous function on the open disc D, then there is a function f (z)
analytic on D such that

lim f (r ) g(r ) = 0
r 1

for all E.
(c) There is a non-constant analytic function f (z) on D having radial limit
0 almost everywhere on D. See Bagemihl and Seidel [1954].
6. Let f (z) be an analytic function on D and let E D. If Re f is nontangentially bounded on E, then f has a nontangential limit almost everywhere
on E.
7. Let  be a simply connected domain, let  be a compact set, let be
an arc in  connecting z 0  to some  \ , and let  be the set of
curves in  joining to . Prove
(z 0 , , )
Hint: Use Theorems 3.2 and IV.5.3.

8 ()
e
.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

225

z0

Figure VI.8 Extremal distance estimate of harmonic measure.

8. (a) Let u(z) be real valued and continuous on D, and harmonic on D, and
u (r ) = . Then
assume supzD |u(z)| < 1 but limr 1 
z
u )( )
ei(u+i
d
(z) =
0

is a conformal mapping from D to a simply connected Jordan domain, is


conformal at = 1, but does not have a non-zero angular derivative at
= 1.
(b) Find a simply connected domain  such that for some D, the conformal map : D  is not conformal at even though  has an inner
tangent at ( ).
9. Suppose the Jordan curve  is a finite union of linear images w = Az + B
of graphs
k = {y = f k (x)}.
Let 1 and 2 be the components of C \  and let j be harmonic measure
for some point in  j .
(a) Prove that j  1 on . Hint: Use the twist point theorem from
Appendix I or use Lemmas I.2 and I.3 directly.
(b) Suppose f k has no finite or infinite derivative at any point. For examples,
see Hobson [1926]. Then j  1 but 1 2 . Consequently if j is the
conformal map from D to  j then at almost every D, is conformal
but  has no tangent at ( ).

(c) Suppose f k (x) = 2n cos 2n x. Then f k is nowhere differentiable, but
f k Z . See Zygmund [1959], page 47. It follows that  has tangents j
almost everywhere.
(d) If f k is nowhere differentiable, but f k Z , then the graph {y = f k (x)}
has a vertical tangent on a set of positive linear measure.
See Bishop [1987] for (a)(d).

CY535/Garnett

226

0 521 47018 8

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

10. From the interval E 0 = [0, 1] remove the middle open interval I1,1 of length
|I1,1 | = 1/8, obtaining a set E 1 . Inductively, obtain E n by removing the middle interval In, j of length |In, j | = 22n1 from each of the 2n1 components

of E n1 . That leaves E = E n with |E| = 3/4. Each deleted interval In, j
is the base of a rectangle Rn, j having dimensions 22n1 2n2 . Let the
tower Tn, j consist of the three sides of Rn, j \ In, j . Then
%
Tn, j
1 = E
n, j

is a Jordan arc connection 0 to 1. Note that 1 falls inside the triangle with
vertices {0, 1, 1/2 + i/6}. On each side of the unit square [0, 1] [0, 1] put
an isometric copy of , with the towers pointing out from [0, 1] [0, 1], to
obtain a Jordan curve 1 . Let 1 denote the bounded domain with 1 = 1 .

Tj3
Jj3
I j3
Jj3
Figure VI.9 First generation towers.
Partition each tower Tn, j into 2n + 1 segments of length |In, j | = 22n1 ,
and repeat the tower construction on each segment using the same length
ratios as before. That gives a second Jordan curve 2 bounding a domain
2 1 . Continuing, we get a Jordan curve  = lim n bounding a domain

 be the exterior C \ ( ). Write (E) = (E, )
 = n n . Let 
). Prove  1  but 
1 (Bishop [1987]).
and 
= (E, 

Figure VI.10 Second generation towers.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

227

11. Let  be the von Koch snowflake curve. Prove


0 <  log 4 () < ,
log 3

4
and consequently that dim() = log
log 3 . (See Appendix D for the definition
of  .)
12. Let S = {z k } be a discrete sequence in D. Then the following are equivalent:
(i) S is a nontangentially dense a.e. on D.
(ii) For a.e. D there is > 0 such that S  ( ).
(iii) For every bounded harmonic function u(z) on D,

sup |u(z)| = sup |u(z k )|.


D

(iv) There is M < such that for every bounded analytic function f (z)
on D,
sup | f (z)| M sup | f (z k )|.
D

(v) For each z D, there exist positive weights k such that for every
bounded harmonic function u on D,

u(z) =
k u(z k ).

(vi) For some z 0 D \ S, there are complex weights k with |k | <
such that for every bounded analytic function f on D,

f (z) =
k f (z k ).
See Brown, Shields, and Zeller [1960], and Hoffman and Rossi [1967].
13. Let  be a simply connected domain and fix a point z 0 . For w , set
() = (w, ) = (z 0 , B(w, ) , ). Then at almost every cone
point, the limit
lim

()
=L

exists and 0 < L < . Let S be the set given by Theorem 6.1. Then
almost everywhere on S
lim sup
0

()
= .

Hint: Use the results or proofs in Section 5. Bishop [1994] conjectured that
lim inf
0

()
=0

almost everywhere on S. This has been proved recently by S. Choi [2004].

CY535/Garnett

228

0 521 47018 8

January 27, 2005

15:41

VI. Simply Connected Domains, Part One

14. Let  be a simply connected domain, and let : D  be conformal. Fix


w0 . For accessible D and a = ( ), let (a, r ) be an arc of the
circle {|w a| = r } such that (a, r ) is a crosscut of  separating w0 from
{(r ) : r0 < r < 1} for some r0 < 1. (Thus (a, r ) actually depends on
.) Write
L(a, r ) = ( (a, r )),
and

A(a, r ) =

L(a, s)ds.
0

McMillan [1970] proved these results:


(a) Almost everywhere with respect to ,
lim sup
r 0

A(a, r )
1
.
2
r 2

(b) Consequently, almost everywhere,


lim sup
r 0

1
L(a, r )
.
2r
2

(c) There is an example for which, almost everywhere,


lim inf
r 0

A(a, r )
A(a, r )
= 0, and lim sup
= 1.
r 2
r 2
r 0

See McMillan [1970] and [1969] or ONeill [1999] for the proof of (a),
and note that (b) also follows from Theorems 5.2 and IV.6.2. In [1970] had
McMillan conjectured that almost everywhere,
lim inf

A(a, r )
1

2
2
r

lim inf

1
L(a, r )
.
2r
2

r 0

and
r 0

These conjectures were recently established by ONeill and Thurman [2000]


and [2001].

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

VII
Bloch Functions and Quasicircles

1. Bloch Functions
An analytic function g(z) on D is a Bloch function if
||g||B = sup |g  (z)|(1 |z|2 ) < .
zD

(1.1)

We write B for the space of Bloch functions and we call ||g||B the Bloch norm1
of g. Because

d z + z 0 
g
(1.2)
 = g  (z 0 )(1 |z 0 |2 ),
dz 1 + z 0 z 
z=0

we have
||g||B = sup |(g T ) (0)|,
T M

(1.3)

where M is the set of conformal self maps of D:


T (z) =

z+a
,
1 + az

with a D and || = 1. Thus the Bloch space B is invariant under M:


||g T ||B = ||g||B , T M.

(1.4)

If g is bounded, then g B by (1.2). In fact by (II.3.4) and (II.3.5), g B if


only Reg is bounded. The function
g(z) =

z2

n=1
1 Actually ||g|| is only a semi-norm, but ||g|| = 0 only if g is constant.
B
B

229

(1.5)

CY535/Garnett

0 521 47018 8

230

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

is a good example of a Bloch function in no H p space. See Theorem 1.3 below.


A function g(z) on D is normal if {g T : T M} is a normal family, in the
extended sense that f is a permitted limit. By Martys theorem (Ahlfors
[1979]) g(z) is normal if and only if
|(g T ) (z)|
2
T M 1 + |g T (z)|
sup

is bounded on each compact subset of D. Since M is transitive, it follows from


(1.2) that g is normal if and only if
|g  (z)|(1 |z|2 )
< .
1 + |g(z)|2
D

sup
Hence (1.1) yields:

Theorem 1.1. Every Bloch function is a normal function.


Recall that the hyperbolic distance from z 1 D to z 2 D is
z2
dz
(z 1 , z 2 ) = inf
,
2
z 1 1 |z|
where the infimum is taken over all arcs in D connecting z 1 to z 2 .
Theorem 1.2. Let g(z) be analytic on D. Then g B if and only if g is Lipschitz
continuous as a map from the hyperbolic metric on D to the euclidian metric
on C. Furthermore,
|g(z) g(w)|
.
(z, w)
z,wD

||g||B = sup
Proof. If


|g(z 1 ) g(z 2 )| B

z2
z1

1
|dz|,
1 |z|2

then a differentiation shows ||g||B B.


Conversely, if g B then we get
|g(z 1 ) g(z 2 )| ||g||B (z 1 , z 2 )
by integrating along a hyperbolic geodesic.
Theorem 1.3. Let g(z) be analytic on D and let G(z) be the primitive
z
g(w)dw.
G(z) =
0

(1.6)


CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Bloch Functions

231

Then g B if and only if G is continuous on D and G(ei ) is a Zygmund


function, G(ei ) Z . When g B, there is a constant C independent of g
such that
1
(gB + |g(0)|) G Z C(gB + |g(0)|).
(1.7)
C
The Bloch function g has nontangential limit at ei0 D if and only if its
primitive G(ei ) is differentiable at 0 .
Proof. Suppose g B. Then (1.6) gives, for 0 < r < s < 1,


  s


G(sei ) G(r ei ) = 
g(tei )ei dt  |g(r ei )|(s r )
r

1 + t 
gB s
dt,
+
log
2
1t
r

(1.8)

and hence G extends continuously to D. Then (1.1) and Theorem II.3.4


give (1.7).
The proof of (1.8) also gives the estimate
1
1
i
i
|g(se ) g(r e )|ds gB
(r, s)ds
r

2 
1+r
gB (1 r ).

= gB log

(1.9)

Now suppose gB 1 and suppose g has nontangential limit 0 at ei . Then


(1.9) implies that dG
d = 0. Indeed, let be small and take > 1 so that
(1 t)ei(+t)  (ei )
for all small t.

ei(+t)
(1 t)ei(+t)
ei
 ( )

Figure VII.1

(1.10)

CY535/Garnett

0 521 47018 8

232

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Set r = 1 t. Then

G(ei(+t) ) G(ei ) = lim


=

sei(+t)

s1 sei
r ei(+t)
r ei

g(z)dz

g(z)dz e

g(sei )ds

(1.11)

r
i(+t)

g(r e
)
+ (1 r )e
1

g(sei(+t) ) g(r ei(+t) ) ds.
+ ei(+t)
i(+t)

By (1.10) the first three terms on the far right-hand side of (1.11) are each o(t),
because g(z) 0 when  (ei ) z ei , and by (1.9) the last term in (1.11)
is bounded by
2 
= O(t).
gB log
1+r
Hence G(ei(+t) ) G(ei ) = o(t) and G  (ei ) = 0.
The converse assertion, that G  has a nontangential limit wherever
is a simple property of the Poisson kernel. See Exercise I.8.

dG
d

exists,


Theorem 1.3 shows that the Bloch function (1.5) has a nontangential limit
at no point, because the Weierstrass function

Re G(z) =
2n cos(2n )
is differentiable at no point. See Katznelson [1968], page 106.

2. Bloch Functions and Univalent Functions


We fix a pair of functions (z) and g(z) analytic on D such that


g(z) = log  (z) , (0) = 0,
or equivalently


(z) =

(2.1)

e g(w) dw.

Theorem 2.1. If is univalent on D, then g B and ||g||B 6. Conversely,


there is < 1 such that if ||g||B < , then is univalent and (D) is bounded
by a Jordan curve.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Bloch Functions and Univalent Functions


Proof. Fix z D and write
(w) =

w+z
1+zw

(z)

 (z)(1 |z|2 )

233

Then is a normalized univalent function, (0) = 0 and  (0) = 1, and by


Theorem I.4.1, |  (0)| 4. Then because
 (0) =

 (z)
(1 |z|2 ) 2z = g  (z)(1 |z|2 ) 2z,
 (z)

it follows that
|g  (z)|(1 |z|2 ) 6.
To prove the converse, let I be an arc on D with |I | . Let C I be the
circle that intersects D orthogonally at the endpoints of I , and let z I be the
point on C I closest to 0.
Lemma 2.2. If gB < 2, then at each D, ( ) = limD z (z)
exists and




( ) (z I ) ( z I )  (z I ) 2   (z I )|I |
(2.2)
2
for all I. Furthermore, extends continuously to D and C 1/2 (D).
Proof. Let J = D {Rew 0} and let T M satisfy T (J ) = I. Then
T (w) =

z I w + |z I |
.
|z I | 1 + |z I |w

Note that when Rew 0,


|T  (w)| 1 |z I |2 |I |

(2.3)

I
zI

Figure VII.2

CY535/Garnett

0 521 47018 8

234

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Suppose for a moment that the limit ( ) exists. Then


g(z)g(z )
I 1 dz
( ) (z I )  (z I )( z I )=  (z I )
e


= (z I )

zI
T 1 ( )

e g(T (w))g(T (0)) 1 T  (w)dw.

e|z|

1|
1 is trivial by Taylor series. With (2.3) and the
hypothesis gB , it yields

T 1 ( )

 


1 + |w|  2
( ) (z I )  (z I )( z I )   (z I )|I |
1 d|w|
1 |w|
0

2  
(z I )|I |,

2
The inequality |e z

which is (2.2). Since (1.6) and (2.1) give



|  (z I )| e|g(0)|

2
1 |z I |

/2

Theorem II.3.2 shows that extends continuously to D, C 1/2 (D), and


|(1 ) (2 )|

C
|z 1 z 2 |1/2 .
2

To conclude the proof of Theorem 2.1, let 1 , 2 D and let I be the shorter
arc of D having endpoints 1 and 2 . Then by (2.2) and triangle inequalities,
|(1 ) (2 )| |  (z I )||1 2 |

4
|  (z I )||I |
2


2  
1
| (z I )||1 2 |,
2

(2.4)

2
because |I | 2 |1 2 |. Thus when 2
< 1, is a one-to-one map from
D onto a Jordan curve, and it then follows by the argument principle that is

univalent on D.

Lemma 2.2 proves more than it says. A Jordan curve  is a quasicircle if


there is a constant A such that
|w1 w| + |w w2 |
A,
(2.5)
|w1 w2 |
whenever w  lies on that subarc of  having endpoints w1 and w2 and
smaller diameter. See Ahlfors [1963] and [1966]. We call (2.5) the Ahlfors
condition. Any C 1 curve or Lipschitz curve is a quasicircle and the von Koch

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Bloch Functions and Univalent Functions

235

snowflake is a nonrectifiable quasicircle. A quasidisc is a domain bounded by


a quasicircle.
w1

w
w1
w2

w2

Quasicircle

Not Quasicircle
Figure VII.3

Theorem 2.3. There is 0 < 1 such that if gB < < 0 , then (D) is a
quasicircle with constant

(2.6)
A A() 2(1 + C1 ).
Proof. Take w j = ( j ) and let w = ( ) lie on the smaller diameter subarc
(I ) of D with endpoints 1 and 2 . Replacing by T for some T M
if necessary, we can assume that |I | . Then by (2.2),




( j ) ( ) ( j )  (z I ) 4   (z I )|I |,
2
so that


8
|w1 w| + |w w2 |   (z I )|(|1 | + | 2 | +
|  (z I )||I |
2

4  
| (z I )| |1 | + | 2 | .
1+
2
Then by (2.4) we obtain (2.5) with the estimate
2 + (4 1) 
2.
A A() =
2 (2 + 1)
Therefore (D) is a quasicircle and (2.6) holds provided
<

2
.
(2 + 1)

The constants and 0 in Theorems 2.1 and 2.3 are not sharp. Becker
[1972] showed that is univalent if gB 1, and that (D) is a quasicircle

CY535/Garnett

236

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

if gB < 1. The constant 1 is sharp in both of Beckers results, but when
gB = 1, (D)
and Pommerenke [1984].
is still a Jordan curve. See Becker
The constant 2 in (2.6) is needed because A = 2 for a circle.
Later we will need the following geometric result, which can be viewed as
a converse of Theorem 2.3. For > 0 write
1
and Imz > },
D = {z : |z| <

and
L = {x R : x + i D }.
When w  choose w0  to satisfy |w w0 | = dist(w, ), and write
(z) = w (z) = w0 i(w w0 )z, D (w) = (D ), and L (w) = (L ).
Theorem 2.4. Let be a conformal map from D onto a simply connected
domain , and let g(z) = log(  (z)). Then the following conditions are equivalent:
(a) There are > 0 and M < such that



(2.7)
1 |z|2 g  (z) ,
sup
{z:(z,z 0 )<M}

for all z 0 D.
(b) The family of maps
(z) =

z+
1 z

 ( )

( )

, D,

1 | |2

contains no sequence converging, uniformly on compact subsets of D, to a map


of the form
z
,
|| = 1.
() (z) =
1 + z
(c) There is > 0 such that for all w , either
D (w)  = ,
or there exists w1 L (w) so that
B(w1 , |w w0 |)  = .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Bloch Functions and Univalent Functions

D (w)
w
w0

237

w
w

0|

D (w)
w
w0 w1

L (w)

w
w

0|

L (w)

Figure VII.4 The two possibilities in Condition (c).


Proof. Suppose (a) is false. Then there exists {z n } D, with z n D
such that
lim

sup

n {z:(z,z )M}
n

(1 |z|2 )|g  (z)| = 0,

for all M < . Then, uniformly on compact subsets of D,


z+z 
n
g(z n ) 0,
g
1 + zn z
and


z+z n
1+z n z
 (z n )

so that
z n (z) =


 (z

z+z n
1+z n z

1,


n )(1 + z n

z)2

1
(1 + z)2

(2.8)

Because z n (0) = () (0) = 0, it follows that


z n (z) () (z)
and that (b) is false.
Now suppose (b) fails. Then there is a sequence {z n } D such that
z n (z) () (z), uniformly on compact subsets of D. If = , then we can
reverse the previous argument to conclude that (a) does not hold. If = then
by (2.8)


z+z n
1 + z 2
 1+z
z
n

.
(2.9)
 (z n )
1 + z

CY535/Garnett

0 521 47018 8

238

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Fix | | < 1 and |z| < 1. Set


n =

+ zn
,
1 + zn

n =

1 + zn
1 + zn

1 +
1 +

, and wn =

n z +
1 + n z

Then n and
z + n
wn + z n
=
.
1 + z n wn
1 + n z
Because wn (z + )/(1 + z) D and (2.9) holds uniformly on compact
subsets of D
z+n

 wn +z n

 1+
 (z n )
z
1+z
w
n
n n
=

+z n

 (n )
 (z n )
 1+z
n
z+
+2 )
)
+2
1 + 1+ z
1 +

z+

1 +
1 + 1+
z
)
+2
1 + z
=
,
1 + z
where
=

1 + 1 +

.
1 + 1 +

(2.10)

When | | = 1 but = , we can solve (2.10) for with | | < 1. Take k


and n,k so that limn n,k = and limn n,k = (k ) . Then a diagonalization argument will give (2.9) with = . Hence (a) implies (b).
The equivalence of (b) and (c) is nothing but the Carathodory convergence

theorem; see Tsuji [1959], page 381.
We remark that if (2.7) fails for M = 1 then it fails for all M < by
Theorems 1.6 and 2.1.
Theorem 2.4 is from Jones [1989]. Exercise 2 gives a result from Pommerenke [1978] that complements Theorems 2.3 and 2.4.
Together Theorems 1.3 and 2.1 set up a one-to-one correspondence
u h ) ,
log  = (u h + i

(2.11)

between univalent functions and small norm Zygmund functions h. This


correspondence was first noticed by Duren, Shapiro, and Shields [1966]. We
give two applications.
Example 2.5. There exists a Jordan domain  such that the conformal mapping
: D  has a non-zero angular derivative at no point.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Bloch Functions and Univalent Functions

239

Proof. Actually we will give two examples. For the first example, let
 n
log(  ) =
z2 .
If is small, then maps D to a quasidisc; and by the remark following Theorem 1.3, has an angular derivative at no D.
The second example is almost the von Koch snowflake, Example VI.4.3.
The domain bounded by the snowflake  has an inner tangent only on a set
2
1 (E) where dim(E) = log
log 3 . On the other hand, the conformal map from
D to the exterior  of  has a non-zero angular derivative at no point of D.
Indeed, the exterior domain  has an inner tangent only at the vertices of ; and
at each such vertex ( ),  contains a truncated cone of aperture 7
6 , which

means that cannot be conformal at .
7
6

Figure VII.5 Exterior angle at a vertex.


Duren, Shapiro, and Shields [1966] introduced (2.11) in order to construct
a Jordan domain that is not a Smirnov domain. If  is a Jordan domain with
rectifiable boundary  the conformal map : D  satisfies  H 1 and



i
 i d
log | (z)| =

Pz (e ) log | (e )|
Pz (ei )ds ( ),
2
D
D
where the singular measure s is the weak-star limit

d
d 
s = lim log |  (ei )|
log |  (r ei )|
.
r 1
2
2
Because  H 1 ,
log |  (z)|

Pz (ei ) log |  (ei )|

d
2

by Theorem A.1 and it follows that s 0. The domain  is called a Smirnov


domain if s = 0, or equivalently if

d
log |  (z)| = log |  (ei )|Pz (ei )
(2.12)
2

CY535/Garnett

0 521 47018 8

240

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

for all (or one) z D. Smirnov domains are important in approximation theory
because  is a Smirnov domain if and only if, for 1 p < , the analytic
polynomials
a0 + a1 z + a2 x 2 + . . .
are L p (, ds) dense in the space, called H p (), of analytic functions satisfying
( f )(  )1/ p H p (D). See Duren [1970], p. 173.
Example 2.6. There exists a Jordan domain  with rectifiable boundary such
that  is not a Smirnov domain.
Proof. Kahane [1969] exhibited an increasing function h Z such that h  = 0
almost everywhere. A different construction appears in Piranian [1966]. We
may assume h Z is small. Then the map defined by (2.11) has image (D)
bounded by a rectifiable Jordan curve and


Pz ( )d( ),
log | (z)| =
D

where is the singular measure having primitive h. Hence (D) is not a Smirnov
domain.
To construct Kahanes function we partition the unit interval into 4n subintervals,
In, j = [ j4n , ( j + 1)4n ),
and define a sequence n of probability measures. Let d0 = d x. When n 0,
assume dn is absolutely continuous with constant density dn, j 0 on each
In, j . Write
In, j = In+1,4 j In+1,4 j+1 In+1,4 j+2 In+1,4 j+3 .
If dn, j > 0, define
dn+1,k =

dn, j 1,
dn, j + 1,

if k = 4 j, 4 j + 3;
if k = 4 j + 1, 4 j + 2.

If dn, j = 0, define dn+1,k = 0 for all In+1,k In, j . Then the sequence n
converges weak-star, because n+1 (In, j ) = n (In ). Take = lim n and
h( ) = ([0,

)).
2

 
If Sn is the closed support of n , then Sn+1 Sn and  Sn  = 0 because an
unbiased random walk on the integers which begins at 1 almost surely hits 0.
See for example Feller [1968]. Therefore h is singular.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Quasicircles

241

To see that h Z , note that whenever I and J are adjacent intervals of the
same length,
|(I ) (J )| C|I |.

(2.13)

By construction, (2.13) holds whenever I = In, j and J = In, j+1 , even though
In, j and In, j+1 may fall in different In1,k . In general, let 4n |I |/2 < 4n+1 ,
and take In, j I and In,k J . Then by the 4-adic case of (2.13),
|(In, j ) (In,k )| 2C4n ,
(I ) (In, j ) 4C4n , and (J ) (In,k ) 4C4n . That establishes

(2.13) for general I and J and hence h Z .

3. Quasicircles
We first describe without proof some of the magical properties of quasiconformal mappings and quasicircles. The proofs can be found in the delightful books
of Ahlfors [1966] and Lehto and Virtanen [1973], and some are outlined in the
exercises. We end the section with the Jerison and Kenig characterization of
quasicircles in terms of harmonic measure.
Let  and  be domains in the extended plane C , let f :   be an
orientation preserving homeomorphism, and let K 1. Then we say f is a
K-quasiconformal mapping if
(a) f is absolutely continuous on almost every horizontal or vertical line segment in , and
(b) the (area almost everywhere defined) derivatives
fz =

fx i f y
2

and

fz =

fx + i f y
2

satisfy
| fz |

K 1
| fz |
K +1

(3.1)

almost everywhere on .
When condition (a) holds we say f is AC L, for absolutely continuous on
lines. The smallest constant K for which (3.1) holds is called the dilatation of
f on .
Let f :   be K -quasiconformal, let z  \ {, f 1 ()}, and sup-

CY535/Garnett

0 521 47018 8

242

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

pose B(z, r ) . Define


m(z, r ) =
M(z, r ) =



inf  f (w) f (z),

|wz|=r



sup  f (w) f (z),

|wz|=r

and
H (z) = lim sup

r 0

M(z, r )
.
m(z, r )

f (z)
M

Figure VII.6 The quasiconformal image of a small circle.


When f is AC L it follows from (3.1) that
H (z) K

(3.2)

almost everywhere on . Conversely, f is K -quasiconformal if f is an orientation preserving homeomorphism for which (3.2) holds almost everywhere. The
equivalence of (3.1) and (3.2) is easy to prove when f is an orientation preserving diffeomorphism, but the general case requires the differentiation theorem of
Gehring and Lehto [1959]. Surprisingly, Heinonen and Koskela [1995] derived
(3.1) from a weaker form of (3.2).
If f is a K -quasiconformal, then f 1 is also K -quasiconformal, and f is a
1-quasiconformal mapping if and only if f is a conformal mapping.
Let f be a K -quasiconformal on a domain , let  be a path family in ,
and write f () = { f ( ) : }. Then (3.1) implies
1
 () f () ( f ()) K  ().
K

(3.3)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Quasicircles

243

Therefore a K -quasiconformal mapping can only increase or decrease an extremal distance by the factor K ,
1
d (E, F) d f () ( f (E), f (F)) K d (E, F),
K
whenever f is K -quasiconformal on a neighborhood of . It follows that a
K -quasiconformal mapping can only increase or decrease the module of a ring
domain W  by the same factor K ,
1
mod( f (W )) mod(W ) K mod( f (W )).
(3.4)
K
Conversely, if f is an orientation preserving homeomorphism defined on an
open set U and if (3.4) holds for every ring W U , then f is K -quasiconformal
on U . Again it is not hard to prove that an orientation preserving diffeomorphism
is quasiconformal if and only if it satisfies (3.4), but in the general case the
argument is much deeper.
One consequence of (3.3) is the large-scale version of (3.2):
Proposition 3.1. If f : C C is K -quasiconformal, then for all z C and
all r > 0,
M(z, r ) e8K m(z, r ).
In other words, if |z 1 z| |z 2 z|, then
| f (z 1 ) f (z)| e8K | f (z 2 ) f (z)|.
The proof of Proposition 3.1 is outlined in Exercise 3.
If f : C C is K -quasiconformal and if f (D) = D, then (3.4) implies
there is a constant M = M(K ) such that
1
| f (ei(+t) ) f (ei )|

M,
M
| f (ei ) f (ei(t) )|

(3.5)

for all and all t < . The converse is also true. See Beurling and Ahlfors
[1956], Ahlfors [1966], or the sketch in Exercise 4. Homeomorphisms of D
satisfying (3.5) are called quasisymmetric functions.
Theorem 3.2. Let f be an orientation preserving homeomorphism of D to D.
Then f is the restriction to D of a quasiconformal homeomorphism of C to C if
and only if (3.5) holds. If (3.5) holds with constant M then the extension satisfies
(3.1) with K K (M), and if f is K -quasiconformal and f (D) = D, then
(3.5) holds with M M(K ).
Let f : C C be K -quasiconformal. It then follows easily from (3.4) that
 = f (D) is a quasicircle with constant A A(K ). The converse holds also.

CY535/Garnett

244

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Theorem 3.3. Let  be a Jordan curve in the plane. Then  satisfies the Ahlfors
condition (2.5) if and only if
 = f (D), where f : C C is K -quasiconformal .

(3.6)

If (2.5) holds with constant A, then (3.6) holds with K K (A), and if (3.6)
holds with constant K , then (2.5) holds with A A(K ).
See Exercise 5 for a proof of Theorem 3.3.
Corollary 3.4. Let  be a quasidisc and let : D  be a conformal mapping. Then has an extension to a quasiconformal mapping from C onto C.
Proof. Let f be a quasiconformal mapping with  = f (D) and f (0) = (0).
Set F = f 1 . Then F is a homeomorphism of D and F is quasiconformal
on D. Extend F to C \ D by reflection
F(z) =

1
F( 1z )

Then F is a quasiconformal self map of C and f F is the desired quasiconformal extension of .



A positive measure on a Jordan curve  is a doubling measure if there is
a constant C such that
(I ) C(J )
whenever I and J are adjacent subarcs of  with diam(I ) diam(J ). Interchanging I and J , we see that is doubling if and only if
diam(I ) C  diam(J )
whenever I and J are adjacent arcs with (I ) (J ). If f : D D is a
homeomorphism, then f satisfies the BeurlingAhlfors condition (3.5) if and
only if (E) = | f (E)| is a doubling measure on D. Kahanes measure in
Example 2.6 is a doubling measure on R singular to Lebesgue measure. From
Jerison and Kenig [1982a] we have the following theorem.
Theorem 3.5. Suppose  is a Jordan curve in the plane and suppose 1 and
2 are the two components of the complement C \ . Let z j  j and let
j (E) = (z j , E,  j ). Then  is a quasicircle if and only if both 1 and 2
are doubling measures on .
Proof. Assume 1 and 2 are doubling measures and assume z 2 = 2 .
Let 1 : D 1 and 2 : C \ D 2 be conformal mappings such that

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. Quasicircles

245

2 () = and 1 (0) = z 1 , and define the welding map h : D D by


h = 21 1 .
Let I and J be adjacent subarcs of D such that |I | = |J | . Because 1 is
doubling,
diam(1 (J )) C1 diam(1 (I )).
Then because 2 is doubling,
| 1 (1 (I ))|
|21 (1 (J ))|
= 2 (1 (J )) C2 C1 2 (1 (I )) = 2
.
2
2
Consequently
|h(J )|
C1 C2 ,
|h(I )|
and after interchanging I and J we see that h satisfies (3.5). Now let H : C C
denote the quasiconformal extension of h given by Corollary 3.4 and define

if |z| 1,
1 (z),
1 (z) =
2 (H (z)), if |z| > 1.
Then 1 is a quasiconformal map such that 1 (D) = , and  is a quasicircle.
Conversely, assume  is a quasicircle and let : D 1 . Let I1 and I2 be
adjacent arcs of  with diam(I1 ) = diam(I2 ) diam()/4. Let I3 be an arc
adjacent to I2 with diam(I3 ) = diam(I2 ) and I3 I1 = , and let w j , w j+1 be
the endpoints of I j , j = 1, 2, 3. Then by the Ahlfors condition (2.5),
dist(I1 , I3 ) |w2 w3 | Adiam(I2 ) = Adiam(I1 ) = Adiam(I3 ),
so that by Exercise 6(b) of Chapter IV and a rescaling,
dC (I1 , I3 ) A .
Set Jk = 1 (Ik ). We may suppose that (0, k Jk , D) 1/2. Then by Corollary 3.4 and (3.4), dC (J1 , J3 ) K A . Hence diam(J2 ) C diam(J1 ), for otherwise an annulus of large modulus separates J1 and J3 . Thus 1 is doubling.

The proof for 2 is of course the same.
The proof of doubling in Jerison and Kenig [1982a] is a different argument
that applies to higher dimensional generalizations of quasidiscs known as nontangentially accessible domains. See Exercise 9 for a Jordan curve on which
2 is doubling but 1 is not doubling. A one-sided version of the definition of
quasicircle yields domains called John domains, which are discussed in Exercise 13.

CY535/Garnett

246

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

4. Chord-Arc Curves and the A Condition


A compact set K is called Ahlfors regular if there is a constant A such that
1 (K B(z, r )) Ar
for all z K . A Jordan curve  is called a chord-arc curve or a Lavrentiev
curve if  is an Ahlfors regular quasicircle. Thus  is a chord-arc curve if and
only if  is rectifiable and there is a constant A such that for all w1 , w2 ,
(w1 , w2 )
A ,
|w1 w2 |

(4.1)

where (w1 , w2 ) is the length of the shorter arc of  joining w1 to w2 .

Figure VII.7 Chord-arc curve.


Note how (4.1) is the rectifiable analogue of (2.5). Curves satisfying (4.1)
were introduced by Lavrentiev in [1936]. A Jordan curve  is called a Lipschitz
curve if
 = {e F()+i : 0 2 },
where F( + 2 ) = F( ) and |F( + t) F( )| A|t| for all t. The smallest
constant A is called the Lipschitz constant for . It is clear that every Lipschitz
curve is a chord-arc curve. The object of this section is to prove a result, parallel
to Theorem 3.5, that characterizes chord-arc curves via harmonic measure.
Let and be continuous positive measures on a curve . We say and
are A -equivalent if there exists < 1 and < 1 such that whenever I is a
subarc of  and E I ,
(E)
(E)

.
(I )
(I )

(4.2)

To see that (4.2) defines an equivalence relation, simply replace E by I \ E.


It is not hard to show that   if and are A -equivalent. If 
is a chord-arc curve and if is A -equivalent to arc length on , then is
a doubling measure. However, not every doubling measure on the unit circle

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Chord-Arc Curves and the A Condition

247

is A -equivalent to arc length and Example 2.6 exhibits a doubling measure


singular to Lebesgue measure. A stopping time argument given in Exercise
14 shows that A is equivalent to this stronger looking condition: There exist
> 0 and C > 0 such that for all arcs I  and all E I ,
(E) 
(E)  1
(E)

.
C 1
(I )
(I )
(I )
When = k( )d and = d are A -equivalent, we say k is an A -weight
and write k( ) A .
Now let  be a rectifiable curve and let = k(z)ds be a positive finite
measure on . We say k(z) satisfies a reverse Hlder condition if there exist
> 0 and C > 0 such that whenever I  is a subarc having length (I ),

 1
1
1
1+
1+
k(z) ds
C
k(z)ds.
(I ) I
(I ) I
When p < we say k(z) is an A p -weight if there is C > 0 such that for all
arcs I ,
)
+ p1


1
1
1  p1
1
k(z)ds
ds
C.
(4.3)
(I ) I
(I ) I k(z)
The A p -condition (4.3) is important because of the famous theorems of Muckenhoupt and of Hunt, Muckenhoupt, and Wheeden telling us that when is
a measure on the unit circle and 1 < p < , the HardyLittlewood maximal
operator f M f or the conjugation operator f 
f is bounded on L p () if
and only if = kd x where k is an A p -weight. See Muckenhoupt [1972], Hunt,
Muckenhoupt and Wheeden [1973], or Garnett [1981].
Proposition 4.1. Let  be a rectifiable curve and let = k(z)ds be a finite
positive measure on . Then the following conditions are equivalent:
(a) and arc length measure ds are A -equivalent,
(b) k(z) satisfies a reverse Hlder condition, and
(c) There is p, 1 < p < , such that k(z) is an A p -weight.
There exist (, ), p(, ), and C = C(, ) such that when (a) holds with
constants and , (b) holds for C and all < (, ) and (c) holds for C and
all p p(, ).
Some parts of the proof of Proposition 4.1 are easy applications of Hlders
inequality but at one point a stopping time is needed. See Exercise 14, Coifman
and Fefferman [1974], Garnett [1981], or Torchinsky [1986].
Theorem 4.3 will show that every chord-arc curve satisfies the conditions of
Proposition 4.1. The easier Theorem 4.2 shows that Lipschitz curves satisfy a

CY535/Garnett

0 521 47018 8

248

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

strong form of condition (c).


Theorem 4.2. Let  be a domain bounded by a Lipschitz curve  (as above), let
: D  be a conformal mapping, and let (E) = (0, E, ). Then there
is a p < 2, depending only on the Lipschitz constant of , such that |  | and
1
d
1/|  | are A p -weights on D and k =
is an A 2 p -weight on .
ds
Proof. Let : D  with (0) = 0. Then almost everywhere the magnitude
of the angle between the tangent vector to  and the radius vector is
  ei  (ei )  


 

 arg(1 + i A) = < ,
arg
i
(e )
2
and so by Zygmunds theorem, Theorem II.3.1,
+ 1 ) 
+
) 



 z  
 




sup
  Pa d
 z   Pa d
aD
) 
 
 + 1 ) 

 z 


dt < ,


 dt
= sup

T


 z 

T M
for all < /(2). If I is an interval in D, take a D such that a/|a| is the
center of I and 1 |a| = |I |/2. Then Pa C/|I | on I and since |(z)/z| is
bounded above and below on D,
+1 )
)

  +

1
1
1

 dt < .
sup
| | dt
|I | I
|I | I   
I
This shows that |1/  | A p with p = 1 + 1/ > 1 + 2/ and a similar argument shows that |  | A p . Because k = 1/|  1 |, we obtain for J = (I )
1


q1

  q1
1
1
1
kds
ds
(J ) J
(J ) J k
+q1
)
q

1
 q1 d
|I | I | |
=

)
1
|I |

.)

1
|I |

+q


I

|  |d
1
d
|  |

+

and k Aq for q > /( 2).

1
|I |

| |

q
q1

q
q1

/q
< ,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Chord-Arc Curves and the A Condition

249

Theorem 4.3 (Lavrentiev). Let  be the bounded domain bounded by a chordarc curve , let z 0 , and let (E) = (z 0 , E, ).
(a) Then and arc length on  are A -equivalent, with constants depending
only on the chord-arc constant of  and dist(z 0 , )/diam().
(b) Moreover, if : D  is a conformal mapping with (0) = z 0 , then arc
length and harmonic measure for z 0 on the curves ({|z| = r }), 0 < r < 1
are A -equivalent with constants independent of r .
Proof. We follow Jerison and Kenig [1982b]. To prove (a), let I be a small arc
on D, form the box
Q = {z = t : 1 |I | < t < 1; I },
and let z Q be its center. Given that has a quasiconformal extension to C,
Proposition 3.1 tells us there is r > 0 such that
B((z Q ), r ) (Q) B((z Q ), Cr ),

(4.4)

where C depends only on the quasiconformality of the extension of . Since 


is a chord-arc curve,
((I )) Adiam((I )) Ar.
Let J = (1 |I |)I be the top edge of Q. By the Koebe estimates (I.4.13),

((J )) =
|  (z)|ds 4dist((J ), ) Ar.
J

Now let L = (, 0 ), 0 = (1 |I |), be one of the edges in Q \ (I J ) of


Q. Let c = e8K , where K is the dilatation of the quasiconformal extension of
from C to C, and define k L to be the point of smallest modulus such that
|(k ) ( )| = ck |(0 ) ( )|.
If 2 N c, then
|(k ) ( )| 2 N |(k+1 ) ( )|,
and because 1 also has a K -quasiconformal extension, Proposition 3.1 and
induction yield
|k | (2c) N |k+1 |.

(4.5)

Write L k = [k+1 , k ]. Then by Proposition 3.1 and the definition of k+1 ,


dist((L k ), ) c|(k ) ( )| = ck+1 |(0 ) ( )|,

CY535/Garnett

250

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

so that by the Koebe estimates


((L k )) 4ck+1 |(0 ) ( )| log
With (4.5) we then obtain
((L)) |(0 ) ( )|

| | 
k
.
|k+1 |

4ck+1 N log(2c) Cr.

Thus we have established


((Q)) C r.

(4.6)

With (4.6) and (4.4) we can apply Lavrentievs estimate (VI.5.1) to the
domain (Q) and the point (z Q ), and the bounds will not depend on Q.
(E)
Let E (I ). Given > 0 there exists > 0 such that if ((I
)) < then by
(VI.5.1)
((z Q ), E, (Q))
(z Q , 1 (E), Q)
=
< .
(z Q , I, Q)
((z Q ), (I ), (Q))

(4.7)

But if (4.7) holds with suitably small , then


| 1 (E)|
(z 0 , E, )
=
< 1/2,
(z 0 , (I ), )
|I |
and we conclude that and ds are A -equivalent on .
(b) By Lemma 4.4 below, the function

|  (zei )|d + log
I (z) = log

|z|/A
|(zei ) (zei )|

is subharmonic on D for A > 0. Because  is chord-arc, we can choose A


so that I 0 on D. Then by the maximum principle, I (r ) 0, and therefore

({|z| = r }) is a chord-arc curve with a constant independent of r .
Lemma 4.4. If f (z) is analytic and nowhere zero on D and if 0 < < 2
then

| f (zei )|d
J (z) = log

is subharmonic on D.
Proof. Write f = g 2 . Uniformly on compact sets, the function J (z) is a limit
of functions of the form
n

K (z) = log
|g j |2 ,
j=1

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Chord-Arc Curves and the A Condition

251

and a calculation with CauchySchwarz shows that


2
K (z) 0.
zz


Lemma 4.4 is also true without the assumption that f is nowhere zero, but
the calculation is a bit more tedious.
Corollary 4.5. Let  be a bounded domain bounded by a chord-arc curve and
let : D  be a conformal map. Then
(a)  is a Smirnov domain, and
(b) There is q > 0 such that 1 H q .
The index q in (b) depends only on the chord-arc constant of  and on the
dilatation of the quasiconformal extension of .
Proof. We first prove (b). By Theorem 4.3, the measures d and |  (r ei )|d
are A -equivalent, with constants independent of r . Thus by part (c) of Propo1
, depending only on the constants of , such that
sition 4.1, there is q = p1

q d
1


sup   i 
<

(r
e
)
2
r
and (b) holds.
Recall that  is a Smirnov domain if and only if

d
log |  (0)| =
log |  (ei )| .
2
D
Because  H 1 , we have
log |  (0)|

log |  (ei )|

d
.
2


q
But 
H 1 by (b), so that we also have

d
log |  (0)|
log |  (ei )| .
2
D
Therefore  is a Smirnov domain and (a) holds.

Theorem 4.6. Let  be a Jordan curve and let  be the bounded component of
C \ . Then  is a chord-arc curve if and only if the following three conditions
all hold:
(i)  is a quasicircle,

CY535/Garnett

252

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

(ii) Arc length on  is A -equivalent to harmonic measure for any z ,


and
(iii)  is a Smirnov domain.
Examples showing that Theorem 4.6 is sharp are given in Exercise 16.
Proof. We have already seen that (i), (ii), and (iii) hold when  is a chord-arc
curve. Now assume (i), (ii), and (iii) hold. Let : D , let I D be an
arc such that diam((I )) diam()/2, let
Q = {z = t : 1 |I | < t < 1; I },
and let z Q be its center. Because Q is a quasicircle and has a quasiconformal
extension, there exists r > 0 and C > 0 such that
B((z Q ), r ) (Q) B((z Q ), Cr )
and the endpoints w1 and w2 of (I ) satisfy
|w1 w2 |

r
.
C

By (ii) and Proposition 4.1, |  | satisfies the A p -condition



1 1 1
1 p
1
  p1 d
|  |d C
|I | I
|I | I 
for some p > 1. But by Jensens inequality,
1 1 1
1 p
1

  p1
exp
log |  |d
   d
|I | I
|I | I
so that
1
|I |

1

|  |d C exp
log |  |d
|I | I
I

d 
.
log |  (ei )|Pz Q ( )
C exp
2
D

But by (iii) and (2.12),


log |  (z Q )| =
so that
1
|I |

log |  (ei )|Pz Q ( )

|  |d C|  (z Q )|
I

d
,
2

2Cr
,
|I |

by Koebes estimate. Therefore



((I )) = |  |d 2C 2 |w1 w2 |
I

(4.8)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. BMO Domains
and  is a chord-arc curve.

253

The left inequality in (4.8) is a reverse Jensens inequality for A p -weights.


Pommerenke [1991] gives a more geometric proof of Theorem 4.6.
Theorem 4.7. Let  be a quasicircle, let  be the bounded component of C \ ,
and let : D  be a conformal mapping. Then  is a chord-arc curve if
and only if arc length and harmonic measure for (0) are A -equivalent on
({|z| = r }), 0 < r < 1, with constants that are independent of r .
Proof. If  is a chord-arc curve then by Theorem 4.3 arc length and harmonic
measure on ({|z| = r }) are A -equivalent uniformly in r . Conversely, the
A -equivalence implies conditions (i), (ii), and (iii) of Theorem 4.6 hold on
({|z| = r }), uniformly in r . Consequently the curves ({|z| = r }) are chordarc uniformly in r < 1, and  is chord-arc.


5. BMO Domains
Let f L 1 (D). Recall from Appendix F that f has bounded mean oscillation,
f BMO, if

1
| f f I |d = || f ||BMO < ,
(5.1)
sup
I |I | I
where I is a subarc of D and
fI =

1
|I |


f d
I

is the average of f over I . Condition (5.1) is very strong. By the JohnNirenberg


theorem (see Exercise F.2), it implies that f L p (D) for all p < . Condition (5.1) also implies that if the Poisson integral f (z) is analytic in D, then
| f  (z)|(1 |z|2 ) C|| f ||BMO .
See Exercise F.3. In other words,
|| f ||B C|| f ||BMO
if f (z) is analytic and f (ei ) BMO. In this section we describe the simply
connected domains  for which the Bloch function g = log(  (z)) BMO. A
second characterization will be given in Chapter X.
Let k( ) 0 be a function on the unit circle D and write g = log k.

CY535/Garnett

0 521 47018 8

254

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Lemma 5.1. Suppose k( )d is A -equivalent to d . Then g = log k BMO.


Conversely, if g BMO then there is a constant b > 0 such that ebg d and d
are A -equivalent.
Proof. Suppose k( )d and d are A -equivalent. Let I D be a subarc.
By Proposition 4.1, k satisfies the A p -condition (4.3),
1
 1
g I g
 p1
d
exp(g g I )d
exp
C
(5.2)
|I | I
|I | I
p1
By Jensens inequality

1
g I g
 p1
1
d
exp(g g I )d 1 and
exp
1.
|I | I
|I | I
p1
Thus (5.2) holds if and only if each factor is bounded and we have

exp
and
exp


1

1
(g g I )d
exp(g g I )d C 
|I | I
|I | I

1 g g  1
g g I
 p1
I
d
d
exp
C .
|I | I p 1
|I | I
p1

Therefore
sup
I

1
|I |



g g I d <
I

and g BMO.
Conversely if g BMO then by the JohnNirenberg theorem (Exercise F.2)
there is b > 0 such that

1
sup
eb|gg I | d < .
|I
|
I
I
Then ebg A2 and ebg d and d are A -equivalent by Proposition 4.1.

Now let : D  be a conformal mapping and write g = log(  ). We


say  is a BMO domain if g = log(  ) BMO. We say  is a chord-arc
domain if  is bounded by a chord-arc curve. If  is a chord-arc domain,
then by Lemma 5.1 and Theorem 4.3, Reg(ei ) = log |  (ei )| BMO, so
that g BMO by Exercise F.3. More generally, Zinsmeister [1984] proved that
every domain bounded by an Ahlfors regular Jordan curve is a BMO domain.
However, not every Jordan BMO domain has an Ahlfors regular boundary. See
Exercise 17.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. BMO Domains

255

Theorem 5.2. Every chord-arc domain is a BMO domain. Conversely, there


is > 0 such that if g BMO and ||g||BMO < , then  is bounded by a
chord-arc curve.
Proof. We have just seen that g = log(  ) BMO when  is a chord-arc
domain. Conversely, since ||g||B C||g||BMO , Theorem 2.3 gives us a 0 such
that  is a quasicircle if ||g||BMO < 0 . By Lemma 5.1 there is 1 such that
if ||g||BMO < 1 , then |  (r ei )| A , uniformly in r < 1. Thus by Theorem

4.7,  is a chord-arc domain if ||g||BMO min(0 , 1 ).
Theorem 5.3. The following are equivalent.
(a)  is a BMO domain.
(b) There exist > 0 and C > 0 such that if z 0  there is a subdomain
U  such that
(i) z 0 U,
(ii) U is rectifiable and (U) C dist(z 0 , ), and
(iii) (z 0 ,  U, U) .
(c) There exist > 0 and C > 0 such that if z 0  there is a subdomain
U  such that
(i) z 0 U and dist(z 0 , ) C dist(z 0 , U),
(ii) U is chord-arc with constant at most C and (U) C dist(z 0 , ),
and
(iii) ( U) dist(z 0 , ).

z0
U

Figure VII.8 Chord-arc subdomain.


A map G : C C is bilipschitz if there is a constant A such that
A1 |z 1 z 2 | |G(z 1 ) G(z 2 )| A|z 1 z 2 |
whenever z 1 = z 2 .

CY535/Garnett

256

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Corollary 5.4. The family of BMO domains is invariant under bilipschitz


homeomorphisms of the plane.
Proof. Condition (c) is invariant under bilipschitz mappings.

Corollary 5.5. Let  be a quasicircle, and let 1 and 2 be the components


of C \ . Then 1 is a BMO domain if and only if 2 is a BMO domain.
Proof. Let F : C C be a quasiconformal map such that 1 = F(D). The
quasiconformal reflection G : 1 2 is defined by
)
+
1
G(z) = F

,
F 1 (z)
and Ahlfors [1963] shows that G can be chosen to be bilipschitz.

Proof of Theorem 5.3. We show (a)  (b)  (c)  (a).


(a)  (b): By (a) we may assume 1 (z 0 ) = 0. Indeed, if T M has
T (0) = 1 (z 0 ) then || log T  ||BMO < Const, so that by the conformal invariance of BMO (see Exercise F.3),

gT (z) = log ( T ) (z) = log  (T (z)) + log T  (z),


has ||gT ||BMO ||g||BMO + Const.
Fix > 1 and recall the nontangential maximal function
G( ) = (g g(0)) ( ) = sup |g(z) g(0)|.
 ( )


Let > 0 be large and form E = { : G( ) } and V = E  ( ). If
is sufficiently large, then (z, E, D) 1/2 on V \ E. It follows that
(0, E, V) |E|
1 if is sufficiently large. On V,
|  (z)| e |  (0))| 4e dist(z 0 , ),
and thus (b) holds for U = (V).
(b)  (c): This is another cone domain construction. See Exercise 19.
(c)  (a): Set (0) = z 0 . Since the domain U in (c) has chord-arc boundary,
has a non-zero angular derivative at almost all points of 1 (U ) and
by Theorem 4.3,
(z 0 , U , ) (z 0 , U , U) > = (C, ) > 0.
Consequently we can construct a cone domain
%
 ( ) 1 (U)
V=
E

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. BMO Domains

257

such that |g(w) g(0)| M on V and (0, E, D) , where E D is compact. In particular, |g( ) g(0)| M on E and |E| 2 . Let I1 be a component of D \ E with center c I and take w1 = (1 |I |)c I and z 1 = (w1 ).
Then w1 has bounded hyperbolic distance to V, so that |g(w1 ) g(0)| C M
since ||g||B 6. Repeating this construction with z 1 in place of z 0 , we obtain
E 1 I1 with (w1 , E 1 , D) and |g( ) g(w0 )| 2C M on E 1 . Since
(w1 , I1 E 1 , D) (w1 , D \ I1 , D)

we obtain |I1 E 1 | c2 2 |I1 |. We repeat this construction on all complementary intervals of E and of the newly constructed sets E j I j . After n generations we arrive at a set Fn such that |g( ) g(0)| nC M on Fn and such that
|D \ Fn | |D \ E| (1 c2 2 )n . Thus
(0, { : |g( ) g(0)| > }, D) c3 ec4 ,
where c3 and c4 depend only on the constants in (c). Applying this argument to
z+w0
), we obtain
(z) = ( 1+w
0z
(w0 , { D : |g( ) g(w0 )| > and | w0 | 2(1 |w0 |)}, D)
c5 ec4 .
That means that g BMO.

Notes
Section 1 follows Makarov [1990a]. Anderson, Clunie, and Pommerenke [1974]
and Anderson and Pitt [1988] and [1989] have more on Bloch functions. The
first Jordan domain not a Smirnov domain was constructed by Keldysh and
Lavrentiev [1937]. We have only touched the beautiful theory of quasicircles. Gehring [1982] has much more, including seventeen different characterizations of quasicircles. Chord-arc curves originated in Lavrentiev [1936]
and resurfaced in Pommerenke [1977], Coifman and Meyer [1979], and Jerison and Kenig [1982a] and [1982b]. It is an open problem to show that
{log  : : D  is conformal and  is chord-arc} is a connected subset
of BMO. See Semmes [1986b], [1988a] and [1988b], Zinsmeister [1984],
[1985], and [1989], Astala and Zinsmeister [1991], MacManus [1994], and
Bishop and Jones [1994] for more about chord-arc curves and BMO domains.
Theorem 5.3 is from Bishop and Jones [1994].

CY535/Garnett

0 521 47018 8

258

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Exercises and Further Results


1. Let be a univalent function in D and suppose 0 < r < R < 1.
(a) Use Theorem 2.1 and Exercise I.23(d) to show there is C independent of
r and R such that

(1 r )3  it
|  (Rei(+t) )|d C
| (r e )|.
(1 R)3
||<1r
(b) Use (a) to show


(1 r )2

it
|  (r eit )|dt.
| (Re )|dt C
(1 R)2
2. Let g B. We say g B0 the little Bloch space if
lim |g  (z)|(1 |z|2 ) = 0.

|z|1

A Jordan curve  is asymptotically conformal if


|w w1 | + |w w2 |
0 as |w1 w2 | 0,
|w1 w2 |
w(w1 ,w2 )
sup

where (w1 , w2 ) is the subarc of  having endpoints w1 and w2 and smaller


diameter. Let g = log(  ) where is the conformal map from D to the
domain bounded by a Jordan curve . Prove  is asymptotically conformal
if and only if g B0 . (Pommerenke [1978]).
3. (a) The spherical distance between z C and w C is

2ds
S(z, w) = inf

1
+
|z|2

with the infimum taken over all rectifiable arcs joining z to w. If |z| = 1 then
S(0, z) = S(z, ) = 2 .
(b) Let |z 1 | = |z 2 | = 1 and let  be the family of curves that separate {0, z 1 }
1
from {z 2 , }. Using the metric (z) = 2 1+|z|
2 , prove

.
4
(c) Use (b) to prove Proposition 3.1. Hint: We may assume z = f (z) = 0,
|z 1 | = |z 2 | = 1, and | f (z 1 )| < | f (z 2 )|. Let  be the set of circles of radius
| f (z 1 | < r < | f (z 2 )|. Then
()

()

1
| f (z 2 )|
log
,
2
| f (z 1 )|

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

259

while by (b)
( f 1 ())

.
4

This proof is from Gehring [1982].


4. (a) Let f : C C be quasiconformal and suppose f (D) = D. Use (3.4)
and Exercise IV.5 to show that (3.5) holds for f .
(b) The converse, that f has a quasiconformal extension whenever (3.5)
holds, is easier to prove on the line. Let f : R R be an increasing homeomorphism such that for all x R and all h > 0,
f (x + h) f (x)
1

M.
M
f (x) f (x h)
Extend F to the plane by defining

1 1
F(x + i y) =
( f (x + t y) + f (x t y))dt
2 0

i 1
+
( f (x + t y) f (x t y))dt
2 0
when y = 0 and F(x) = f (x). Prove F : C C is a homeomorphism.
(c) Prove F is ACL and satisfies (3.1) almost everywhere.
(d) Prove F is bilipschitz,
C 1 |z w| |F(z) F(w)| C|z w|,
where C = C(M).
See Beurling and Ahlfors [1956] or Ahlfors [1966].
5. Let  be a Jordan curve that satisfies the quasicircle condition (2.5) and let I
and J be disjoint subarcs of . Let 1 and 2 be the components of C \ 
with 2 .
(a) Prove
C 1 d1 (I, J ) d2 (I, J ) Cd1 (I, J ),
where the constant C depends only on the constant in (2.5). Elementary
proofs can be found in Ahlfors [1963] and Gehring [1982].
(b) Let 1 : D 1 and 2 : C \ D 2 be conformal mappings such
that 2 () = , and let h : D D be the welding map h = 21 1 .
Using part (a) and Exercise IV.5 shows that h satisfies the BeurlingAhlfors
M-condition (3.5).
(c) Let H : C C be the quasiconformal extension of h guaranteed by (b)

CY535/Garnett

260

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

and Theorem 3.2. Define

1 (z) =

if |z| 1;
1 (z),
2 (H (z)), if |z| > 1.

Then 1 is a quasiconformal map such that 1 (D) = , and  is a quasicircle.


(d) Let 1 < < 2. Construct a quasicircle  such that
0 <  ( ) < .
See Gehring and Visl [1973].
(e) Exhibit a Jordan curve of positive area. It is known that every quasicircle
has area zero.
6. (a) A domain  is an M-extremal distance domain (or M-QED domain)
if whenever E and F are disjoint compact sets in ,
d (E, F) MdC (E, F).
Note that the reverse inequality with M = 1 is trivial. Show the unit disc D
is a 2-QED domain and show the constant 2 is sharp. It then follows by (3.3)
that a K quasidisc is an M(K )-QED domain. Now suppose  is a Jordan
curve and define 1 and 2 as in Exercise 5. Assume that 1 is a QED
domain, and let I and J be disjoint arcs on . Then
M 1 d1 (I, J ) d2 (I, J ) Md1 (I, J ),
in which the second inequality is obtained from taking conjugate extremal
distances. From this it follows as in Exercise 5 that  is a quasicircle. See
Gehring and Martio [1985]and Yang [1992] and [1994] for further results.
(b) Let be a conformal mapping from a quasidisc  to a simply connected
domain W and let U W be a quasidisc. Then 1 (U )  is also a quasidisc. This is easy from (a). See Fernndez, Heinonen, and Martio [1989].
7. Let : D  be conformal and let L be a line. Then there are p > 1 and
C > 0, independent of  and , such that

|( 1 ) (w)| p |dw| C.
L

This result is due to Fernndez, Heinonen, and Martio [1989]. By Exercise


III.16 we may assume L = U where U  is a closed disc. Then by
Exercise 6, 1 (L) is a quasicircle. On the other hand, by (I.4.15) 1 (L) is
an Ahlfors regular curve. Therefore 1 (L) is a chord-arc curve and |( 1 ) |
is an A -weight, which means that ( 1 ) L p (U ). Moreover, a check
of the constants will show p and C are independent of  and .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

261

8. Let  be a simply connected domain and let : D  be conformal. For


z D and |z| > 1/2, write I (z) = { D : | z| < 2(1 |z|)}.
(a) If  is a quasidisc, then
C(K )(1 |z|)|  (z)| diam(I (z)) C(K )1 (1 |z|)|  (z)|
and

(E.1)

dist (z), (I (z)) C(K )(1 |z|)|  (z)|,

where K is the smallest dilatation of a quasiconformal extension of to a


quasiconformal  : C C with () = . Hint:

Use (3.4) and estimate the modulus of C \ I (z) B(z, (1 |z|)/2) .


(b) Conversely, if (E.1) holds whenever |z| < 1/2, then  is a quasidisc.
(c) For z  write d(z) = dist(z, ) and for > 1 and w  write
 (w) = {z  : |z w| < d(z)}.
Then there are = (, K ) > 1 and (, K ) > 1 such that when D,
() ( )  (( )) ( ( )).
Hint: Use (a).
(d) If  is a chord-arc curve, then (E.1) can be improved to

|  ( )|ds C(, K )1 (1 |z|)|  (z)|.
C(, K )(1 |z|)|  (z)|
I (z)

9. Let  be a Jordan curve in the plane, let 1 and 2 be the two components of C \ , so that 2 , let 1 : D 1 and 2 : C \ D 2
be conformal mappings, and let h = 21 1 be the welding map.
(a) Let j be harmonic measure for some point in  j . Prove h  = 0 if and
only if 1 2 .
(b) Prove there exists  such that 1 is doubling but 2 is not doubling. Hint:
See Bishops towers, Figure VI.10.
10. A Jordan curve in  C such that  is called a quasicircle if  is
the quasiconformal image of R, or equivalently,  is a Mbius image of
a bounded quasicircle. Prove that a graph {y = f (x) : < x < } is
a quasicircle if and only if f is a Zygmund function, f Z (Jerison and
Kenig [1982a]).
11. Let  be a quasicircle in C and assume  and write 1 and 2 for the
complementary components of . Ahlfors [1963] proved that in this case
there exists a quasiconformal G : 1 2 which is bilipschitz,
C 1 |z w| |G(z) G(w)| C|z w|.

CY535/Garnett

262

0 521 47018 8

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

(a) If 1 1 is a rectifiable Jordan curve, then 2 = G(1 ) 2 is rectifiable and


1 2 = 1 .
(b) Let j be harmonic measure for some point z j  j , and let K j be the
set of cone points for  j . If 1 (K 1 ) > 0, then 1 (K 1 K 2 ) = 1 (K 1 ) and
K

1 K 2

1  K 1 K 2 1  K 1 K 2 2 .

These results are from Bishops thesis [1987].


12. Suppose the Jordan curve  is a finite union of linear images w = Az + B
of graphs
k = {y = f k (x)}.
Let 1 and 2 be the components of C \  and let j be harmonic measure for some point in  j . If f k Z , then by Exercises 10 and 11  is
a quasicircle and 1  Tn 1  2  1 . See Exercise VI.9 and Bishop
[1987].
13. Let  be a simply connected domain such that
/ . Then  is called a
John domain if there is z 0  and c > 0 such that for every z 1  there
is an arc  joining z 0 to z 1 such that
dist(z, ) c|z z 1 |
for all z . Thus in Figure VII.9 a slit disc is a John domain but the
teardrop is not a John domain. This exercise will survey some properties of
John domains. See John [1961], Martio and Sarvas [1979] and especially
Nkki and Visl [1991] for much more.

z0

John

Not John
Figure VII.9

(a) Recall that a crosscut of  is a Jordan arc  having both its endpoints
in . Every crosscut of  divides  \ into two components 1 and
2 . Prove  is a John domain if and only if there is a constant C such that

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

263

whenever is a crosscut of ,
min diam( j ) C diam( ).
j

Figure VII.10 Both sides John.


(b) The boundary of a John domain  is locally connected. Consequently, if
is a conformal map from D onto a John domain, then has a continuous
extension to D. However, a John domain need not be a Jordan domain.
(c) A Jordan curve is a quasicircle if and only if both of its complementary
components are John domains.
(d) Let  be a bounded simply connected domain and let : D  be
a conformal mapping. Recall I (z) = { D : | z| < 2(1 |z|)} and
write Bz = D B(z, 2(1 |z|)). The next theorem is from Pommerenke
[1982a]. See also Pommerenke [1964] and [1991].
Theorem. The following are equivalent:
(1)  is a John domain.
(2) There is C1 such that for whenever 1/2 < |z| < 1,
diam(Bz ) C1 dist((z), ).
(3) There exist , 0 < 1, and C2 such that if 0 < r < s < 1 and D,
1 s 1
|  (s )|

C
.
2
|  (r )|
1r
(4) There exist , 0 < 1, and C3 such that if I J are subarcs of D
|I | 
diam(I )
C3
.
diam(J )
|J |
To prove (1)  (2), use Beurlings projection theorem to show that
dist((z), (Jj )) K dist((z), ),

CY535/Garnett

0 521 47018 8

264

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

where Jj , j = 1, 2, are the two components of 2I (z) \ 23 I (z). Then by Exercise III.16(b) obtain two geodesic arcs j with initial point z and terminal
points in Jj , j = 1, 2, such that
(( j )) Cdist((z), (Jj )).
Write Uz for the domain bounded by 1 , 2 and the arc of D containing
I (z). When  is a John domain (a) shows
diam((Uz )) Cdist((z), ).
But because Bz \ Uz has bounded hyperbolic diameter, Koebes theorem
gives
diam(Bz \ Uz ) Cdist((z), )
even if  is not a John domain.
To show (2)  (3), we suppose = 1 and define
1
g(s) =
|  (x + i y)|2 d xd y.
s

|y|1x

By (2)
g(s) Area((Bs )) C12 (1 s)2 |  (s)|2 .
By (I.4.14) there is an absolute constant a > 0 such that
g(s) a(1 s)2 |  (s)|2
and
g  (s) a(1 s)|  (s)|2 ,
since the sets {x + i y : |z| 1 x} have bounded hyperbolic diameter.
Therefore
d
(1 s) g(s) 0,
ds
with = a/2C12 and we obtain (3) with this and C2 = a 1/2 C1 . Note that
when (2) holds, C .
If (3) holds then by Koebes estimate (1) holds with z 0 = (0) and with
([0, 1 (z)]). Thus (3)  (1) and a hyperbolic geodesic can be taken
for the curve .
To show (3)  (4), let I J be subarcs of D. We can assume J = I (z)
and I = I (w). Then by (2)
diam(I ) C1 (1 |w|)|  (w)|.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

Let z  =

|z|
|w| w.

265

Then by (3)
(1 |w|)|  (w)| C2

1 |w| 
1 |z  |

|  (z  )|.

But |  (z  )| a|  (z)| again via (I.4.14) and for arbitrary simply connected
domains
1 |z|

diam(I (z)) a  diam B z,
.
2
Therefore (4) holds with the exponent from (3).
Now assume (4) and let 1/2 < |z| < 1. Given > 0 there is A > 0 such
that by Theorem V.3.5
|( ) (z)| A dist((z), )

on 2I (z) \ I j where | I j |/|2I (z)| . Then if is sufficiently small
(4) implies that max j diam((I j )) 21 diam(2I (z)) and hence


1
diam(2I (z)).
2
It then follows from V.3.5 exactly as in the proof of (1)  (2) that
diam(Bz ) C dist((z), ). Therefore (2) holds.
(e) If  is a John domain then harmonic measure is doubling.
14. (a) If two measures and are A -equivalent, then   .
(b) If and are A -equivalent and if is a doubling measure, then is a
doubling measure.
(c) There is a doubling measure on the unit interval singular to Lebesgue
measure. Hint: Start with a quasicircle for which 2 1 .
(d) Assume (4.2) holds for = d x
and on [0, 1]. Let N 1 and suppose
N
E I [0, 1] satisfy (E) < 2 . We may suppose I is a dyadic interval.
Let {I j } be the maximal dyadic subintervals of I such that
diam(2I (z)) 2 A dist((z), ) +

(E I j )
.
(I j )


By Lebesgues theorem,
I j \ E = 0. Let I j denote the dyadic interval
containing I j with |I j | = 2|I j |. Since I j is maximal, (E I j ) (I j )

by (4.2). Set E 1 = I j . Then E E 1 , (E) (E 1 ), and
 (I j )

N 1
2
(E 1 )
2
(E) <
.
(I )
(I )

CY535/Garnett

0 521 47018 8

266

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

Continuing by induction we obtain


(E)
N
(E)

N ,

(I )
2
(I )
which means
(E)
1 (E) log 1 /log 2

.
(I )
(I )
(e) Adapt the proof outlined in (d) to the case when and are continuous
positive measures.
(f) If k is an A -weight, then k satisfies the reverse Hlder condition. Fix I
and large, and assume I kd x = 1. Let {I j,1 } be the maximal intervals in
the dyadic decomposition of I such that

and set E 1 =

< f I j,1 2,
I j,1 . Then k on I \ E 1 and
|E 1 | C  1

.
|I |
1

Now let {I j,2 } be the maximal intervals in the dyadic decomposition of I


such that

and set E 2 =

2 < f I j,1 22 ,
I j,2 E 1 . Then
1
|E 2 | 2n1 C n  1
.

|E 1 |
n

Continuing, we obtain the reverse Hlder condition



1
 1
1
1+
k 1+ d x
C
kd x
|I | I
|I | I
1
whenever 1 + < 1
.
(g) Conversely, the reverse Hlder condition and Hlders inequality yield

.
A for = 1+

(h) By Hlders inequality, A p implies A with = p1


p . Conversely, A
and its symmetry imply
 1
1 1

C
1+


|I |,
dx

k
kd
x
I
I
I kd x

from which A p follows with p = 1 + 1 .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

267

15. Let  be a Jordan domain, let 1 : D  and 2 : C \ D C \  be


conformal mappings.
(a) Prove that the welding map 2 1 1 is in A if and only if  is a BMO
domain and a quasidisc. (Astala and Zinsmeister [1991]).
(b) Prove |  | A and  is a Smirnov domain if and only if there is a
constant K such that

|  |ds K dist((z), )
Iz

|ei

z
|z| |

where Iz = { :

< 1 |z|}.
(c) Prove that  is an Ahlfors regular John domain if and only if  is a
Smirnov domain and |  | is an A -weight. See Pommerenke [1982a] and
Zinsmeister [1984].
16. (a) Example 2.6 is a quasicircle for which |  | is an A -weight, but it is not
a Smirnov domain.
(b) Construct a Zygmund function f ( ) such that f 1, f  L 1 but
/ BMO. Then  = {r ei : 0 r < f ( )} is a quasidisc and a Smirnov
f
domain, but |  | is not in A .
(c) Construct a Jordan Smirnov domain  such that |  | A but  is not
a quasidisc.
17. (a) Let map D to a domain bounded by an Ahlfors regular Jordan curve.
Then for p < 1/5 there is C such that

|  | p ds C(1 |z|)|  (z)| p .
I (z)

Hint: By Hlder,


|  | p ds
I (z)

I (z)

 p
1 p
2p
|  ( )|
1 p ds
ds
|(
)

(z)|
.
|( ) (z)|2
I (z)

The first factor can be estimated directly, using the regularity of D, and the
second factor can be handled with Prawitzs theorem.
(b) Under the assumptions in (a), g = log(  ) BMO. Hint: If p is small,
then (  ) p =  where is a conformal map of D to a quasidisc. But by
(a), |  | A , so that log(  ) BMO.
(c) Let be a conformal map from the half-plane onto the region above the
curve  = {y = sin(x 2 ), x R}. Then log(  ) BMO, but  is not Ahlfors
regular. See Zinsmeister [1984] and [1985].
18. The space VMO, for vanishing mean oscillation, consists of all g BMO

CY535/Garnett

0 521 47018 8

268

January 27, 2005

15:41

VII. Bloch Functions and Quasicircles

such that
lim sup

0 |I |<

1
|I |


|g g I |ds = 0,
I

and VMOA = VMO BMOA. See Sarason [1975] or Garnett [1981]. A


Jordan curve  is said to be asymptotically smooth if  is rectifiable and if
lim

sup

0 |w1 w2 |<

(w1 , w2 )
= 1,
|w1 w2 |

whenever w1 , w2  and (w1 , w2 ) is the length of the shorter  subarc


joining w1 to w2 . Let g = log(  ) where maps D to a Jordan domain. Prove
g VMOA if and only if  is asymptotically smooth. See Pommerenke
[1978].
19. Let  be a Jordan domain such that B(0, 1) . Assume () M and
let E  satisfy (0, E, ) . Prove there is = (, M) such that
there is a chord-arc domain U  with B(0, 1) U, (0, E U, U)
and (E U) . Hint: Take a union of fixed angle cones over a subset
of E.
20. Let  be a Jordan curve bounding domains 1 and 2 and let j : D  j be
conformal. Suppose, as happens in Example 2.6, that log |1 | is the Poisson
integral of a negative singular measure on D, so that 1 is badly not a
Smirnov domain. Prove there is c > 0 such that |2 | c on D, and that
consequently 2 is a Smirnov domain. Hint: Use the estimate (6.4). This
result is due to Jones and Smirnov [1999].

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

VIII
Simply Connected Domains, Part Two

We begin the chapter with the famous theorems from Makarov [1985], [1987],
and [1990a] on Hausdorff measures and harmonic measure in simply connected domains. Next we discuss the Brennan conjecture on the L p integrability of derivatives of univalent functions and prove some related theorems from
Carleson and Makarov [1994]. Then we give Baernsteins counterexample to
the L p version of the HaymanWu theorem.

1. The Law of the Iterated Logarithm for Bloch Functions


Theorem 1.1 (Makarov). There is a constant C > 0 such that whenever g(z)
is a Bloch function on D,
|g(r )|

lim sup
r 1

log

1
1r

1
log log log 1r

CgB .

(1.1)

Inequality (1.1) is called Makarovs law of the iterated logarithm because


it was first proved by Makarov in [1985] and because it resembles the classical Khinchin law of the iterated logarithm in Feller [1968] or Chung [1974].
Consider for example the bad Bloch function
g(z) =

z2 .

(1.2)

n=1

The series (1.2) behaves much like a sum of independent identically distributed
random variables, and Salem and Zygmund proved in [1950] that its partial

269

CY535/Garnett

0 521 47018 8

270

sums Sn (z) =

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two


n
k=0

z 2 satisfy
|Sn ( )|
lim sup $
=1
n log log n

(1.3)

for almost every D. But if g(z) is defined by (1.2), then (1.3) is equivalent
to
|g(r )|
= 1,
lim sup
1
1
r 1
log 1r
log log log 1r
which is an instance of (1.1).
Theorem 1.1 has a converse, proved by Jones in [1989].
Theorem 1.2. If ||g||B 1 and if there exist > 0 and M < such that for
all z 0 D,



(1.4)
1 |z|2 g  (z) ,
sup
{z:(z,z 0 )<M}

then almost everywhere on D,


Reg(r )

lim sup
r 1

log

1
1r

1
log log log 1r

c(, M) > 0.

(1.5)

When is a conformal mapping from D onto a domain , Theorem VII.2.4


gave a geometric condition on  that is necessary and sufficient for the Bloch
function g = log(  ) to satisfy condition (1.4). For example, (1.4) holds when
maps D to the domain inside the snowflake curve.
Proof of Theorem 1.1. This proof is due to Carleson (unpublished) and independently to Pommerenke [1986b]. We give Pommerenkes version, which
yields the best known constant C = 1 in (1.1). A different proof, from Makarov
[1990a], is given in Appendix J.
We may assume g(0) = 0 and ||g||B = 1. Let p 0 be an integer and consider the integral means
2
1
|g(r ei )|2 p d.
I p (r ) =
2 0
The proof begins with the identity

d
4 p 2 r 2

(r I p (r )) =
|g(r ei )|2 p2 |g  (r ei )|2 d,
dr
2 0

(1.6)

which is known as Hardys identity. Hardys identity can be proved by examining the Fourier series of |g| p or by differentiating I p (r ) directly. A corollary

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Law of the Iterated Logarithm for Bloch Functions

of Hardys identity is the inequality




1 p
1 p
I p (r ) p! log

p!
log
,
1r
1 r2

271

(1.7)

valid for ||g||B = 1 and for integral p 0. Indeed, (1.7) is trivial when p = 0,
and to prove (1.7) when p 1 we use (1.6) and induction to get
4 p2r
d
I p1 (r )
(r I p (r ))
dr
(1 r 2 )2
1  p1
4 pp!r
log

2
2
(1 r )
1 r2


d
1 p
d
r
log
,
p!
dr dr
1 r2
and since I p (0) = 0, an integration then yields (1.7).
Now let us apply the HardyLittlewood maximal theorem to the function
|g(r ei )| p L 2 . By inequality (1.7), gr (ei ) = sup<r |g(ei )| satisfies


1
1 p
(1.8)
|gr (ei )|2 p d C p! log
2
1r
with C independent of p. Next fix > 1 and set
A p (r ) =
Then

1
1
1
 .

 p+1
1
1r
1
log log 1r
log 1r

A p (s)ds

while by (1.8)
1

|gs (ei )|2 p d ds

A p (s)


C p!
r

1
1
C
 ,

p
1
p log 1
log
log
1r
1r

ds
1
1



C p!.
1
1
1
s
log 1s
log log 1s

It follows, via Fubinis theorem and Chebychevs inequality, that the set
'
(
1
i 2 p
2
A p (s)|gs (e )| ds > C p p!
Ep = :
r

(1.9)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

272

VIII. Simply Connected Domains, Part Two

satisfies |E p |
of gs ,

1
. Therefore if
p2

|g(r ei )|

-

1
1
log 1r
log log log 1r


p> p0

E p , then by (1.9) and the definition


1
1
3

1
2p
1
C 2 p C2 p p 2 p p! 2 p log log 1r
1
log log log 1r

. (1.10)

1
Finally setting p = log log log 1r
in (1.10) and using Stirlings formula, we

obtain (1.1) almost everywhere with constant C = 1.

Jones [1989] proof of Theorem 1.2 is elementary and similar to the classical
proof of the law of the iterated logarithm for Bernoulli trials in Feller [1968].
We give his proof in martingale language in Appendix J.

2. Harmonic Measure and Hausdorff Measure


A real valued function h(t) on an interval (0, t0 ) is called a logarithmicoexponential function or an L-function if h(t) is defined by a finite algebraic
combination of exponential functions and logarithm functions. The main theorem about L-functions is from Hardy [1954]: If h and g are L-functions, then
the limit
h(t)
lim
t0 g(t)
exists in the extended interval [0, ]. Because the derivative of an L-function
is again an L-function, it follows that every L-function is monotone on some
interval (0, t1 ). Every measure function in this section will be assumed to be a
positive, increasing and continuous L-function. When we are comparing harmonic measure to a Hausdorff measure h , we may also assume its measure
function h(t) satisfies
lim

t0

h(t)
= .
t

(2.1)

Indeed if limt0 h(t)


t = 0 then we have h by Theorem VI.5.2, and if
<
then the situation is explained by Corollary VI.6.2. We
0 < limt0 h(t)
t
1
write h for the inverse function of h, so that h 1 (h(t)) = t.
Theorem 2.1 (Makarov). Let h be a positive increasing L-function satisfying
(2.1). Let be a conformal mapping from D onto a simply connected domain
 and let denote harmonic measure for some point w0 . Then

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Harmonic Measure and Hausdorff Measure

(a)  h lim inf

(1 r 2 )|  (r )|
> 0 a.e. on D,
h 1 (1 r )

(b) h lim inf

(1 r 2 )|  (r )|
= 0 a.e. on D, and
h 1 (1 r )

r 1

r 1

273

(c) there is a set A  of -finite h measure such that (A) = 1 if and


only if
lim inf
r 1

(1 r 2 )|  (r )|
< a.e. on D.
h 1 (1 r )

Theorem 2.2 (Makarov). There is an absolute constant C > 0 such that if


h(t) = te

C log

1
t

log log log

1
t

then for every simply connected domain ,


 h .

(2.2)

Conversely, there is c < C such that if


h(t) = te

c log

1
t

log log log

1
t

(2.3)

then there exists a Jordan domain  for which


h .

(2.4)

Note that Theorem 2.2 contains Theorem VI.5.1: If < 1, then w   . Let
us derive Theorem 2.2 from Theorem 2.1 before we prove Theorem 2.1.
Proof of Theorem 2.2. Let  be any simply connected domain, let be a
conformal map from D onto  and let g = log(  ). Then by Theorem VII.2.1,
||g||B 6, and by Theorem 1.1 there is a constant C > 0 such that
lim sup
r 1

|Reg(r )|
1
C
1

log 1r log log log 1r

almost everywhere on D. Take


h 1
1 (t) = te

C log

1
t

log log log

1
t

Then by (2.5),
lim inf
r 1

(1 r )|  (r )|
h 1
1 (1 r )

(2.5)

CY535/Garnett

274

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

almost everywhere, and by Theorem 2.1(a),  h 1 . But then a little calculus


shows that


C log 1t log log log 1t
h 1 (t) = o te
as t 0, and (2.2) follows.
Conversely, let  be any Jordan domain that satisfies condition (C) of
Theorem VII.2.4. The snowflake is one example. Let be the conformal
mapping from D to  and let g = log(  ). Then by Theorem VII.2.4 and
Theorem 1.2, there is c3 > 0 such that
lim sup r 1

Reg(r )


 c3 > 0,
1
1
log 1r log log log 1r

almost everywhere on D. If c < c3 , take c with c < c < c3 and set


h 1
2 (t) = te

c log

1
t

log log log

1
t

Then
lim inf
r 1

(1 r )|  (r )|
h 1
2 (1 r )

=0

almost everywhere, so that by (b) of Theorem 2.1,


h 2 .
If (2.3) holds, then for small t
h(t) h 2 (t)
and thus (2.4) holds for c < c3 .

Appendix K gives more examples where (2.4) holds.


The extremal constant
C1 = inf{C : (2.2) holds}
is not known. However the above proof of (2.2) shows C1 < 6C2 where
C2 = inf{C : (1.1) holds}
and the proof of Theorem 1.1 yields C2 1, but the value of C2 is not known.
See Bauelos and Moore [1999]. Bauelos [1986] and Lyons [1990] have other
proofs that C2 1. The extremal
sup{c : (2.4) holds for some domain}

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Harmonic Measure and Hausdorff Measure

275

is also not known.


Theorem 2.1 itself is a consequence of the next lemma.
Lemma 2.3. Let be the conformal mapping from D to a simply connected
domain  and let E D be a Borel set.
(a) If
lim inf

(1 r )|  (r )|
A
h 1 (1 r )

almost everywhere on E, then there exists E 1 E such that


|E \ E 1 | = 0
and
h ((E 1 )) 4 A|E|.
(b) There are constants c1 > 0 and q > 0, not depending on E, such that if
lim inf

(1 r )|  (r )|
>B>0
h 1 (1 r )

(2.6)

almost everywhere on E, then


h ((E)) c1

B
1

(1 + B) 2

|E|q > 0.

Each of the six implications of Theorem 2.1 follows immediately from either
part (a) or part (b) of Lemma 2.3, and the remaining details of the proof of
Theorem 2.1 are left to the reader.
Proof of Lemma 2.3(a). The proof of (a) is almost the same as the proof of
Theorem VI.5.2. We may assume A > 0. Fix = max{2, 21A } and define
I (z) = { D : |z | < (1 |z|)}.
Let {z n } be a sequence in D such that
(1 |z n |)|  (z n )| Ah 1 (1 |z n |),
and such that {z n } is nontangentially dense on E. Fix > 0. By the Vitali
covering lemma there is a subsequence {z k } of {z n } so that the intervals I (z k )
are pairwise disjoint,
(1 |z k |)|  (z k )| <
and
|E \

,
2

I (z k )| = 0.

CY535/Garnett

0 521 47018 8

276

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

Take wk = (z k ), rk = dist(wk , ), Bk = B(wk , rk ), and


%

V = 
Bk .
Then rk 2(1 |z k |)|  (z k )| < and


h(rk ) h 2 Ah 1 (1 |z k |) .
If A 41 , then
h(rk ) (1 |z k |),
and if A > 14 , then since

h(t)
t

is decreasing by (2.1),

h(rk ) 4 A(1 |z k |).


In either case we obtain



h(rk ) 4 A

|I (z k )| 4 A|E|.

Therefore V = V1/m satisfies h (V ) 4 A|E|. On the other hand, because


2, Lemma VI.5.3 implies that |E \ 1 (V )| = 0, and thus (a) holds for

E 1 = E 1 (V ).
For the proof of (b) we need three additional lemmas.
Lemma 2.4. Let be the conformal mapping from D onto a simply connected
domain , and let be a crosscut of  with endpoints w1 , w2 . Set
j = 1 (w j ), j = 1, 2. If is the geodesic in D connecting 1 to 2 , let z
satisfy |z | = inf |z|. Then
diam( ) c(1 |z |)|  (z )|,

(2.7)

for some absolute constant c > 0.


Proof. Applying a Mbius transformation to the disc and linear map to , we
may suppose that = (1, 1), z = 0, (0) = 0 and  (0) = 1. We need to
prove diam( ) c > 0. Let Br = {|z| r }. By the Koebe one-quarter theo1
. On the other
rem, B 1 . If B(1 , diam( )) B 1 = , then diam( ) 20
4

hand, if B(1 , diam( )) B 1 = , then 1 ( ) separates B 1 1 (B 1 )


24

from a semicircle, say T + = D {Imz > 0}. Thus


dD (B 1 , T + ) dC (B 1 , B(1 , diam( )) log
24

and (2.7) holds.


C
,
diam( )


CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Harmonic Measure and Hausdorff Measure

277

The second lemma is from Carlesons fundamental paper [1973]. It is the


key to many of the deeper results in this chapter.
Lemma 2.5. Let  be a simply connected domain. Fix w0  and let 0 .
For
|w0 0 |
,
0<r <
2
set
 = B(0 , 2r ),
D = B(0 , r ), and D
and for any M > 0 take
0
M1
.
k0 = k0 (M) = 1 +

 contains
Then there exist r0 = r0 (M) > 0 such that if r < r0 , then  D
N

2
1
k0 log
log 2
r

crosscuts
1 , . . . , N
of , such that each j separates w0 from a continuum j  and
(D  \

N
%

j ) < r M ,

(2.8)

j=1

where (E) = (w0 , E, ).


D
D
1,2

2
1,1

U1
1

w0

Figure VIII.1 Proof of Lemma 2.5.

CY535/Garnett

278

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

 connecting w0 to . Let { j } be the set


Proof. Fix one curve  \ D

of component arcs of  D having the property that there exists a curve
 through
P  connecting w0 to w j  D such that P first crosses  D
j . By Exercise I.13,  \ j consists of two simply connected components;
/ U j . Because  is simply connected
take U j to be that component with w0
 = , U j Uk = whenever j = k . Also by Exercise I.13,
and  \ D
j = (U j ) \ j  is a continuum and j separates j from w0 .

Now U j D is a union of arcs j,k , and by Exercise I.13, each j,k separates

w0 from a continuum j,k j . Let j = k j,k and let  j be the family of
paths in  joining to j . Then by Theorem IV 5.3 and Exercise VI.7,
j = ( j )

8 ( j )
,
e

(2.9)

even if  is not a Jordan domain or j is not connected. Let  j be the set of


 to k j,k . Then every path in  j contains a path from
paths in U j joining D

 j , so that by the extension rule,
( j ) ( j ).
By the parallel rule,

j

1
1
2

.


( j )
log 2
D, D
d D\D


Hence for k = 1, 2, . . . ,
"
# j:

1
2
1 #
1

k log

1
( j )
log 2
r
k log r

and

#
"
1
2
8
k log .
# j : j r k

log 2
r
0
1
Then since k0 = 1 + M
,

j 8 r k0



1 8
2
log
(k + 1)r k r M
log 2
r

k=k0

if r < r0 and if r0 is small. Let

#
"
8
N = # j : j r k0 .

(2.10)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Harmonic Measure and Hausdorff Measure

279

1
2
Then N k0 log
2 log r by (2.10). Select 1 , . . . , N to have the largest j .
Then

(D  \

N
%

j )

j r M .

j 8 r k0

Lemma 2.6. Assume the L-function h(t) satisfies (2.1) and


2/3
Ct h(t) ct exp log(1/t)
.

(2.11)

Then there is C2 > 0 such that if 0 < t < C2 then


B

h(t/B) h(t),

(2.12)


t
C3 h(t).
log(1/t)

(2.13)

(1 + B) 2
and there is C3 < such that

log(1/t)h

Proof. By (2.1), h(t)/t is decreasing

and (2.12) holds if B 1. Assume B > 1.



Set x = 1/t and g(x) = log h(t)
t . Then by (2.11)
C  g(x) (log x)2/3
for large x. Then since g(x) and (log x)2/3 are both L-functions, Theorem 19
of Hardy [1954] gives
2
(log x)1/3
3x
for x large and integration then yields (2.12) and (2.13).
g  (x) c

Proof of Lemma 2.3(b). To prove (b) we may assume that h(t) satisfies (2.11).
Indeed, since (2.11) holds for every measure function of the form
h 1 (t) = te

C log

1
t

log log log

1
t

we will have established Theorem 2.2 if we prove (b) under the additional
assumption (2.11). On the other hand, if (2.11) fails, then by Hardys theorem
on L-functions,
h 1 (t) = o(h(t)).
But then we can use Theorem 2.2 (which would have been proved for the
measure function h 1 ) to obtain h 1 ((E)) > 0 and h ((E)) = whenever
E D has |E| > 0. Therefore part (b) of Lemma 2.3 will also hold when
(2.11) fails.

CY535/Garnett

0 521 47018 8

280

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

Now assume (2.6) holds for E with |E| > 0. Fix > 2. By (VII.1.6) and
Theorem VII.2.1 there is a constant c > 0 so that for E
(1 |z|)|  (z)|
> cB > 0.
h 1 (1 |z|)

lim inf

z ( )

1
n

Let  ( ) =  ( ) {z : |z| > 1 n1 }. For sufficiently large n, the closed set


E n = { E :

inf
1

zn ( )

1
(1 |z|)|  (z)|
(cB + )} E
1
n
h (1 |z|)

satisfies |E| < 2|E n |, and we may replace E by E n and B by B  = cB + n1 .


Cover (E) by discs D of radius r < r0 such that

h(r ) 2h ((E)).
(2.14)
 ) < 1 ,
By choosing r0 sufficiently small we can assume that 2 ( D
n
dist((0), (E)) > 4r0 , and h(r0 )/r0 > B, since h(t)/t is decreasing. Then
()
 for
by the M = 1 case of Lemma 2.5 there are subarcs j D
1 j N ()
()

such that j
and

2
1
log( ),
log 2
r
()

is a crosscut of  separating w0 from a continuum j

(D  \

N%
()

()

j ) r

h(r )
.
B

(2.15)
()

On the other hand, if (1 ) and (2 ) are the endpoints of an arc j , then


 ) 1 . By discarding some () if necessary, we will
|1 2 | < 2 ( D
j
n
()

suppose that j E = . Thus z E, since > 2, where z is the point


closest to the origin on the geodesic connecting 1 and 2 . By (2.6) and Lemma
2.4
c2
()
()
diam j ,
h 1 (( j )) h 1 (1 |z |)
B
so that
N
()

j=1

()

h 1 (( j ))

2c2 r
.
B

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. The Number of Bad Discs


()

Fix , set t j = h 1 (( j )), t =


N
()


()

( j ) =

j=1

Because N ()


2
log 2

t j s

t j 2c2 r /B, and s = t/ log 1/t. Then

h(t j ) =

h(t)
t

h(t j ) +

h(t j ).

t j >s

log(1/t), (2.13) gives


2c r 

t
2
2
log(1/t)h
C4 h
.
log 2
log(1/t)
B

is decreasing, (2.13) also gives

h(t j ) =

t j s


t j s

j=1

h(t j )

Moreover, since

N
()


281

2c r 
 h(t j )  h(s)
2
C3 h
.
tj

tj
t
s
B
j
t s
t s
j

Therefore by (2.12)
N
()

j=1

( j() )

(1 + B) 2
h(r ).
C4
B

(2.16)

Now (2.14), (2.15), and (2.16) yield


1

(1 + B) 2
|E| = ((E)) C1
h ((E)),
B
which is assertion (b).

3. The Number of Bad Discs


Let  be a simply connected domain, normalized so that  and
diam() = 2, and write (E) = (, E, ). It follows from Beurlings
projection theorem that
(B(, )) 1/2 ,

(3.1)

for all  and all . The power of in inequality (3.1) is sharp if  is a


line segment with one end at , because then (B(, )) c 1/2 for 1. In
their [1994] paper, Carleson and Makarov estimated the number of times (3.1)
can almost fail. For > 0 define N (, , ) to be the maximum cardinality of
sets of the form


1
k  : B(k , ) are pairwise disjoint and (B(k , )) 2 + .

CY535/Garnett

0 521 47018 8

282

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

Theorem 3.1. There are constants A and K such that for all 
N (, , ) A K .

(3.2)

The constant
K = inf{K : (3.2) holds for some constant A}
will play a leading role in the later sections of this chapter and determining K
is an important open problem.
Proof. Fix small and consider the annuli
A j ( ) = {z : j < |z | < j1 },
having moduli m =
and write

1
2 log 1/. Let

mj =

inf

E j denote any subarc of  {|z | = j }

{E j ,E j1 }

dA j (E j , E j1 )

and
X j ( ) = m j m.
Then X j 0 and X j = 0 if and only if A j  lies on a line segment with

endpoint .
Lemma 3.2. Given > 0, there is n() > 0 such that if n > n(), = n and

1
B(, ) 2 + ,
then
n1


X j ( )

j=1

1
3
n log .

Proof. We use Lemma 2.5 with w0 = and M =


1 +

(3.3)

1
2

+ 2. Take n so large that

1
2
log .
log 2

Then by Lemma 2.5 there is an arc E {|z | = 2}  that separates


from a continuum  for which
1

() = ( B(, 2))

2 + 2 +2
2
log 2

log

2 +2 .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. The Number of Bad Discs

283

Take J = {|z | = 1}. Then J meets  so that by Theorem IV.5.3,



1 1
1
dist (E, J ) 1 +
+ 2 log .
2

Set E n = E. Because  is simply connected we can by reverse induction (i.e.,


for j = n 1, n 2, . . . , 2, 1) find arcs E j  {|z | = j } such that E j
separates E j+1 from J . Then by the serial rule,
n1


distA j (E j , E j1 ) dist  (E, J ).

j=1

Hence
n1

1 
1
(n 1)
1 1
log +
+ 2 log ,
Xj 1 +
2

j=1

and (3.3) holds when n

1
2

+ 1.

Lemma 3.3. There are absolute constants > 0 and > 0 such that if ,
k 2, and
X 1 ( ) k ,
then for all   A1 ( ) and all 2 j k,
1
X j (  ) log .

(3.4)

A j (  )



A1 ( )

Figure VIII.2 Proof Lemma 3.3.


Proof. Since diam() = 2 the continuum  meets both boundary components of A1 ( ). Let be the smallest angle of a sector of A1 ( ) that contains

CY535/Garnett

284

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

A1 ( ) . Then by Corollary V.4.3 or Exercise V.4,


X 1 ( ) c

3
(log 1 )2

where c is an absolute constant. Consequently


<

k
2

(3.5)

if > 3 and is sufficiently small. But now if  A1 ( )  and 2 j k,


then the annulus A j (  )  contains two crosscuts of A j (  ) separated by
angle 3 . Note this is even true for |  | > 1 because diam() = 2.
1
.

That gives (3.4) with = 10
To complete the proof of Theorem 3.1, we let D be any closed disc of radius
1 and let
N  = N  (, n, , )
be the maximum number of points p D  such that
| p q | n , p = q

(3.6)

and
n1


X j ( p ) .

(3.7)

j=1

Also let
N  (n, , ) = sup N  (, n, , ).
{,D}

Fix small. We claim there exist constants C1 and C2 such that


N  (n, ) N  (n, , ) C1 eC2 .

(3.8)

By Lemma 3.2, (3.8) implies (3.2) with K = 3 C2 .


We prove (3.8) by induction on n. Because diam() = 2, we clearly have
N  C 2n and we obtain (3.8) for n 2 by taking C1 sufficiently large.
Now assume n 3 and assume (3.8) holds for all k < n. Take 1 , . . . n 
satisfying (3.6) and (3.7). Let and be the constants in Lemma 3.3. There
are three cases.
Case 1: min D X 1 ( ) > 2 .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. The Number of Bad Discs

285

In this case
n1


X j ( p ) 2 .

j=2

1
2

Cover D  by c discs of radius . Consider one such disc and suppose


it contains points 1 , , m that satisfy (3.6) and (3.7). Rescaling to a disc of
radius 21 , we see by induction that
m N  (n 1, 2 ) C1 eC2 (

2 )

Hence
N  (, n, ) c

1
2

eC2 C1 eC2 C1 eC2 ,


2

if C2 C2 ().
Case 2: min D X 1 ( ) n .
Fix D  with X 1 ( ) n . Then by Lemma 3.3, X j ( ) log 1
for 2 j n and all A1 (z) . Thus if < (n 1) log 1 , then all
p D  with (3.6) and (3.7) satisfy p B(, ), and by a rescaling
N  (, n, ) N  (n 1, ) C1 ec2 .
On the other hand, if (n 1) log 1 , then by (3.6)
N  (, n, ) c

1
2n

2n

ce (n1) C1 eC2 ,

if C1 and C2 are sufficiently large.


Case 3: n < min D X 1 ( ) 2 .
Let the minimum be
1 attained at D , and choose k 2 so that
X 1 ( ) (k+1) , k . Then by Lemma 3.3,
X 2 , . . . , X k log


k
on A1 ( ) . By (3.5) we can cover D  \ B(, ) with c 1 discs of
radius k , each centered in A1 ( ). Then by rescaling and induction,
N  (, n, ) N  (n 1, k ) + c
C1 eC2 (
C 1 eC2
provided C2 C2 ().

k )

+c

1
k

1
k

1
N (n k, (k 1) log )

C1 eC2 ((k1) log )

CY535/Garnett

286

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

4. Brennans Conjecture and Integral Means Spectra


Let be a conformal mapping from the disc D onto a simply connected domain
. Brennans conjecture is that


|  |2 p d xd y =
|( 1 ) | p d xd y <
(4.1)


/ ( 43 , 4), the integral (4.1) diverges for the


whenever 43 < p < 4. When p
z
Koebe function (z) = (1z)2 . When p = 2 the integral has value . When

< p < 3, (4.1) is a consequence of the distortion estimate (I.4.17). See Metzger [1973]. When 43 < p < 2, (4.1) follows from the theorem of Prawitz, Exercise I.23(d), and a simple integration. Using the methods in Carleson [1973],
Brennan [1978] established (4.1) in the range 3 < p < 3 + for some unknown
> 0. In Appendix L we give Pommerenkes [1985a] elementary proof of (4.1)
for p < 3.399. Brennans conjecture is true for many special types of domains;
see Exercise 8 and the paper [1998] of Baraski, Volberg, and Zdunik.
Brennans conjecture is a point on a spectrum of questions about integral
means of derivatives of univalent functions. Let be univalent on D, let t R,
and define
"
1
#
.
(4.2)
(t) = inf : |  (r ei )|t d = O
1r
5
3

The function (t) is the integral means spectrum of . Clearly (t) 0. By


Hlders inequality, (t) is a convex and continuous function. By the distortion
theorem (I.4.17),

if s < 0,
(t) s,
(t + s)
(4.3)
(t) + 3s, if s > 0.
The Koebe function k(z) =

z
(1z)2

has integral means spectrum

3t 1,
k (t) = 0,

|t| 1,

if t 13 ,
if 1 t < 13 ,
if t < 1.

(4.4)

Define the (unbounded) universal integral means spectrum to be




B(t) = sup (t) : is univalent .
Thus B(t) is the smallest number such that for all > 0 and all univalent
there is constant C(, ) such that

1
B(t)+
.
|  (r ei )|t d C(, )
1r

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Brennans Conjecture and Integral Means Spectra

Theorem I.4.5 yields the upper bound



3t, if t 0,
B(t)
|t|, if t < 0,
and (4.4) shows that

3t 1,
B(t) 0,

|t| 1,

287

(4.5)

if t 13 ,
if 1 t < 13 ,
if t < 1.

(4.6)

Feng and MacGregor [1976] proved B(t) = 3t 1 for t 2/5. See Exercise 7.
Theorem 4.2 below shows that B(t) = |t| 1 for t < K /2, and Appendix L
includes Pommerenkes [1985a] proof that
1/2
1
1
(t) 3t 2 + 7|t|3
(4.7)
t + 4t 2
B(t) + t+
2
4
for all t.
3t 1
|t|
|t| 1

3t

K2
1
0 13 25
Figure VIII.3 Known bounds for B(t).
Lemma 4.1. Brennans conjecture holds if and only if
B(2) = 1.
Proof. Assume B(2) = 1 and let 2 < p < 4. Then B(2 p) < < 1 by
convexity, so that for all

1 
.
|  (r ei )|2 p d C
1r
Then Fubinis theorem gives (4.1). Conversely, if (4.1) holds for p < 4 then
because the means

|  (r ei )|2 p d

CY535/Garnett

0 521 47018 8

288

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

are increasing in r ,

|  (r ei )|2 p d

1
1r

|  (z)|2 p d xd y

{r <|z|<1}

and B(2 p) 1. Therefore by continuity B(2) 1, while (4.4) implies


B(2) 1.

The proof of Lemma 4.1 also shows
sup{ p : (4.1) holds} = inf{ p : B(2 p) 1}.
Theorem 4.2. Let K be a constant for which the conclusion (3.2) of Theorem
3.1 holds. Then for t < K /2,
B(t) = |t| 1.
Define
K = inf{K : (3.2) holds for some constant A}.

(4.8)

By Theorem 4.2, Brennans conjecture will be true if K 4. We will see


conversely in Section 5 that Brennans conjecture is false if K > 4. Section 5
also contains an example that shows B(t) > |t| 1 for t (2, 1). Therefore
K 4 and Brennans conjecture is equivalent to the conjecture that K = 4.
Proof. Fix t < K /2 and assume  and diam() = 2. Let : D 
be univalent and satisfy (0) = . For 1 r small we partition the circle
C
1r
arcs I j of equal length , with 1r
{|z| = r } into 1r
16 < 8 and write z j

for the midpoint of I j . Because || log( )||B 6, we have

|  |t d |  (z j )|t .
(4.9)
Ij

Let n and k be positive integers with n large and set


A(k) = { j :

(k+1)
n

< |  (z j )| n }

and
N (k) = card A(k).
By the distortion theorem, |  (z j )| 1r
8 > , so N (k) = 0 for k n. By
Koebes estimate if j A(k) then





B(z j , ) B (z j ), |  (z j )|
2
8
k+1
1


1+ n 
1+ n 
B(z j ,
) .
B (z j ),
8
32

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Numbers and Polygonal Trees

289

1 (1+ n )
k
Set = 10

and = 21 n+k
. Since n 1 for large n, the discs
{B((z j ), )} jA(k) are pairwise disjoint and
1


1+ n
), {|z| < 1 r }
, B((z j ), ),  0, B(z j ,
40

C = C 2 + .
By Theorem 3.1

N ( j) A

K
2

(1 nk )

jk

Therefore by (4.9) and a summation by parts,



n1

k
n t+1 N (k)
|  (r ei )|t d 2 + C
k=0
n1

K
k
t
K k

2 + C A 1 2 1 +
n t (1 n ) 2 n


C

t
t + K /2



k=0
|t|1

1
1r

since t < K /2.

5. Numbers and Polygonal Trees


Let be the conformal map from D onto a simply connected domain  and let
limits
1 = 2 be points of D at which has nontangential

(1 ) and (2 ).
For > 0 let B ( j ) be the component of  B ( j ), that contains some
tail segment
{(r j ) : r0 < r < 1}
of the image arc {(r j ) : 0 r 1}. (If ( j ) = , we let B ( j ) be the
corresponding component of  {|z| > 1/}.) Let  =  (, 1 , 2 ) be the
family of paths in  joining B (1 ) to B (2 ), and let  be the set of paths in
C joining B((1 ), ) to B((2 ), ). If < let
 j =  j (, ) = { B ( j ) :  },
j = 1, 2. Then by the basic rules for extremal lengths we have
( ) ( ) + ( 1 ) + ( 2 ),

CY535/Garnett

290

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

and
( j )

1
log / , j = 1, 2.
2

Then
0 ( ) ( )

1
+ ( 1 )
log /
2

1
2
log /
+ ( )
2
( ) ( ) + O( 2 + 2 ),

(5.1)

where we get O( 2 + 2 ) by comparing C \ {B((1 ), ) B((2 ), )} to an


annulus bounded by concentric circles. Therefore the limit

lim ( ) ( )
0

exists in [0, ]. Define

= (, 1 , 2 ) = lim e2(( )( )) .


0

(5.2)

Then
0 (, 1 , 2 ) 1.

If > 0 then for < small the two middle


terms in (5.1) are small,

and if
is fixed, then by Corollary V.4.3, E j B(( j ), ) \ B(( j ), ) is confined
to a narrow sector with vertex ( j ), where E j is the connected component
of  B(( j ), ) which contains ( j ). Consequently (, 1 , 2 ) > 0 can
hold for at most one point j 1 (( j )) and we can unambiguously define

(, (1 ), (2 )) = (, 1 , 2 ).

( j )

(r j )

Figure VIII.4

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Numbers and Polygonal Trees

291

It is clear that (, 1 , 2 ) is invariant under Mbius transformations, and we


will always take 2 = .
Lemma 5.1. Suppose : H  is a conformal mapping such that
(z)
1,
z2

(z ).

(5.3)

Let x R and assume that (x) = a,  (x) = 0, and  (x) exists. Then
(, a, ) =

2
|  (x)|

(5.4)

Proof. If is small, 1 (B (x)) is approximately a half-disc of radius

2
|  (x)|

while 1 (|z| = 1 ) is approximately the semicircle of radius 2 . Then


)
+
1 |  (x)|  21
1
( ) log

2
while
( )

1
1
log 2 ,
2

and therefore (5.4) holds.

For example, suppose  is a finite connected union of line segments with


extreme points a1 , . . . , an and . In this case we say  is a polygonal tree.

a3

a2
a1

a5

a4

a7

a6
Figure VIII.5 Polygonal tree.

Let : H  satisfy (5.3) and let a j = (x j ). Then


(, a j , ) =

2
|  (x

On the other hand, if a


/ {a j },
(, a, ) = 0.

j )|

CY535/Garnett

0 521 47018 8

292

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

When  is a polygonal tree we write


j = (, a j , ).
Example 5.2. Let 0 < t < 1 and set
$
(z) = (z 2 1)(z 2 t 2 ).
Then is the conformal map from H to a domain  bounded by the half line
[t, ] and the segment on the imaginary axis [ib, ib] where b = (1 t 2 )/2,
and satisfies (5.3).
ib
t
ib
Figure VIII.6
This example has three non-zero numbers, and by (5.4) they are
1 = (, t, ) =

2t
1 + t2

and
2 = 3 = (, ib, ) =

1 1 t2
.
2 1 + t2

Note that


j2 =

1
2t 2
<1
+
2 (1 + t 2 )2

since 0 < t < 1, but that for p < 2,




j =

2 p t p + 21 p (1 t 2 ) p
(1 + t 2 ) p

(5.5)

= 1 + 21 p (1 t) p + o((1 t) p ) > 1
if 1 t is small.
Theorem 5.3. Let K be defined by (4.8) and let p 1. Then the following are
equivalent:
 p
(a)
j 1 for every polygonal tree.
(b) K 2 p.
(c) B( p) = p 1.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Numbers and Polygonal Trees

293

From (5.5) and the theorem it follows that K 4. Consequently we have


the following corollary.
Corollary 5.4. The following are equivalent:
 2
(a )
j 1 for every polygonal tree.
(b ) K = 4.
(c ) Brennans conjecture is true.
Proof of Theorem 5.3. The implication (b)  (c) is Theorem 4.2, and we
will prove (c)  (a) in Section 6.
(a)  (b): Assume (b) is false and let p < q < K /2. By the definition of

K there exists, for any A > 0, C > 0, > 0, and > 0, a simply connected
domain  such that  and diam() = 2 and there exist

1
2q+1
(5.6)
N A 2q log

points j  such that


| j k | 2
and
1

(, B( j , ) , ) 2 + .
By Lemma 2.5 there is an arc
j  B( j , 2)
having endpoints in  such that j separates from a continuum j 
and
1

(, j , ) c

2 +
log 1

Here c denotes a constant having possibly different values at different occurrences.


Let : D  be conformal with (0) = . Then 1 ( j ) is an arc in D
with endpoints on D. Let j be the hyperbolic geodesic in D having the same
endpoints as 1 ( j ) and let z j be the midpoint of j . Then
1

1 |z j | 1 ( j ) (, j , ) c

2 +
log 1

(5.7)

Set r j = (1 |z j |)/2 and B j = B(z j , r j ). By Koebes estimate and the


GehringHayman Theorem, Exercise III.15,
diam(B j ) dist((z j ), ) c diam(( j )) c diam( j ).

CY535/Garnett

0 521 47018 8

294

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

Therefore
(B j ) B( j , c),

(5.8)

and hence no point lies in more than C1 balls B j . Also, by (I.4.17)


1 |z j | c|  (z j )|

c
,
1 |z j |

(5.9)

and by (5.7) and (5.9) we have


B j {z : 1 c 1/2 |z| 1

}.
2

(5.10)

Moreover, (5.9) gives


2+q

 q

| |

d xd y

Bj

cr j2 |  (z j )|q

rj

so that by (5.10), (5.8), and (5.6)





1 2q+1 2+q
N 2+q

q
2q
log
| (z)| d xd y
c A
.
1
C1 q

q
1c 2 |z|1 2
Consequently there exists R,
1

1 c 2 < R < 1

(5.11)

such that



1 q1
1 2q+1 2+q

i q
12
2q
| (Re )| d c A
log

c
A

q
1 R
(5.12)
if is small.
From (5.12) it follows easily that (c)  (b); i.e., that B( p) > p 1 if
2 p < K ; but let us continue and prove (a)  (b).
The left-hand side of (5.12) increases with R and 1 1+R
2 is comparable to
1

1 R. Thus we may suppose that R = e n 1 n1 and that is analytic in


a neighborhood of D with  = 0 in D. Replacing with (z) if necessary,
we may suppose that
 q1 3
2
1
1
d
c A.
(5.13)
|  (e n ei )|q

n
2
2
Let 1 be the conformal map of D onto D \ [ 1e , 1] such that 1 (0) = 0 and

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Numbers and Polygonal Trees

295

(1) = 1e . Then


1 (z n )
n (z) = 1 (z ) z
zn
n

1
n

1
n

is a conformal map of D onto D \ k where {k } are n equally spaced radial


slits. Set k = e2ik/n , k = 1, . . . , n. Then n = 0 at {k } and n (k ) = Rk .
Moreover
1

|n (k )| = n|1  (1)||1 (1)|1 n cn.


Let 0 be the conformal map of H onto D \ [0, 1) given by
0 (z) =

(z +

1
z 2 1)2

and set
= C0 n 0 ,
where
1

41 (0) n
4e n
C0 =

.
lim z z(z)
diam()
1

n 0 (H)
Figure VIII.7
Then lim z (z)/z 2 = 1 and
k =

2
2
= 
,
|  (xk )|
| (Rk )||n (k )||0 (xk )|

where xk = 01 (k ). Since |0 ( 1 (ei ))| c1 < on { 2

3
2 },

CY535/Garnett

0 521 47018 8

296

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

c
n|  (Rk )|

provided 2 arg k 3
2 .
By (5.13) and Koebes estimate
 q1 
 q
1
1
1
c A,
k c
n
|  (Rk )|q 2n
where the sums are taken over those k for which 2 arg k 3
2 . Approximate
(R) by a polygonal tree. Then (a) is violated because A is large, c is a constant
and p < q.


6. The Dandelion Construction and (c)  (a)


Let T be a polygonal tree with extreme points and a1 , a2 , . . . , a N . Let a0 T
be such that the infinite segment (a0 , ] is a relatively open subset of T in the
direction ei0 , and set T0 = T \ (a0 , ]. The numbers are dilation invariant,
and we may suppose that T0 B(a0 , 1). Choose 0 > 0 so that for each j 1
T B(a j , 0 ) = [a j , a j 0 ei j )
is a line segment and B(a j , 0 ) B(ak , 0 ) = if j = k. Set T1 = T and for
< 21 0 set
%

a j + ei( j +0 ) (T0 a0 ) .
T2 = T1
B(a1,2 , 0 )
a1

a1

T
a0

a0
a2,2,2

a2

a2,1
a2
B(a2 , 0 )

Figure VIII.8 Dandelion construction.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. The Dandelion Construction and (c)  (a)

297

Then T2 is a new polygonal tree with extreme points and the N 2 finite points
a j1 , j2 = a j1 + ei( j1 +0 ) (a j2 a0 ).
For each pair j1 , j2 ,
T2 B(a j1 , j2 , 0 ) = [a j1 , j2 , a j1 , j2 0 ei j1 , j2 )
is again a line segment, and the balls {B(a j1 , j2 , 0 )} are pairwise disjoint. We
continue, attaching a translation and rotation of 2 T0 at the N 2 finite extreme
points of T2 to obtain another polygonal tree T3 T2 with N 3 finite extreme
points, which we label a j1 , j2 , j3 . This construction can be iterated indefinitely,
yielding at stage n a tree Tn Tn1 having N n finite extreme points a j1 , j2 ,..., jn ,
and Tn+1 is constructed from Tn by attaching rotations and translations of T0
dilated by the factor n . Note that B(a j1 ,..., jn , n1 0 ) B(a j1 ,..., jn1 , n2 0 )
since n1 0 + n1 < n2 0 and so for each n the balls {B(a j1 ,..., jn , n1 0 )}
are pairwise disjoint.
Set
n = C \ Tn ,
T =

Tn ,

and
=

n

= C \ T .

We call T the dandelion generated by T . Note, however, that T also depends


on the choice of the dilation factor .
Write
J = ( j1 , j2 , . . . jn )
for a multi-index of length |J | = n in {1, 2, . . . N }n , and define
a J = a j1 , j2 ,..., jn ,

J = (n , a J , ),

nJ = n

%
|I |=n
I = J

B(a I , 2 n ),

CY535/Garnett

298

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

and
J  = ( j1 , j2 , . . . , jn1 ).
By the extension rule
2J (nJ , a J , ) J ,

and by the proof of (5.1)


J J  jn (1 + O()).

(6.1)

The next lemma shows that the reverse inequality to (6.1) almost holds if the
dilation factors are small enough.
Lemma 6.1. Given > 0 there exists = () such that for every multi-index
J of length n,

2
2J (1 )2
(6.2)

jn
J
2
and
2J (1 )n j1 j2 . . . jn .
J

(6.3)

Proof. We may choose small enough that

) j ,
2
for j = 1, . . . , N . By the extension rule and induction, (6.2) implies (6.3) and
we only prove (6.2). Because numbers are translation invariant, we may
assume a0 = 0. Choose so small that

4
1
1
0
log
,
2 c
1 2
j
j (1 , a j , ) (1

where c is the constant given in Exercise IV.10. By construction,


nJ {z : n1 < |z a J  | < n2 0 }
is a radial line segment and


J
\ B(a J  , 2 n1 ).
nJ \ B(a J  , 2 n ) n1
3

Set  = {z : |z a J  | = n 2

0 }. If is small then by Exercise IV.10 and

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. The Dandelion Construction and (c)  (a)

299

the extension rule,


1
( ) dnJ ( B(a J , ), B(0, ))

 1
1
4
dnJ ( B(a J , ), ) + dnJ (, B(0, )) + c

0
1
1
1
log
dnJ ( B(a J , ), ) + d J  (, B(0, )) +
n1

2
1

Removing a radial slit from an annulus does not affect the extremal distance
between its bounding circles, so that as in the proof of (5.1)


1 
1 
d J  , B(0, ) dC , B(0, )
n1



1 
1 
d J  B(a J  , ), B(0, ) dC B(a J  , ), B(0, ) + O(),
n1

and

dnJ B(a J , ),  dC B(a J , ), 




0

= d jn B(a jn , n1 ), B(a0 ,
) dC B(a jn , n1 ), B(a0 ,
)
1

n1

)
d jn B(a jn , n1 ), B(a0 ,
1

n1

) + O()
dC B(a jn , n1 ), B(a0 ,

and


dC B(a J , ),  +dC , B(0, ) = dC B(a J , ), B(0, ) + O().

Thus

2
2J = lim e2(( )( )) 2

jn
J  (1 ),
0
2
for sufficiently small, which implies (6.2).

Proof of (c)  (a): Fix p > 0 and assume (a) is false. Then there is a polygonal
tree T with finite extreme points a1 , a2 , . . . , a N such that
N

j=1

j = A > 1.

CY535/Garnett

0 521 47018 8

300

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

We make the dandelion construction with dilation parameter so small, for


n 2, that Lemma 6.1 holds with satisfying
(1 ) p A > 1.

(6.4)

Let : H  be a conformal map satisfying (5.3) and set


1 + z

: D .
(z) = i
1z
For a multi-index J = ( j1 , . . . , jn ), let
J : D nJ ,
with J (0) = (0) and J (1) = . If f = 1 , f J = 1
J , and J = f J (a J )
then
1
8
C

.
(6.5)
| J ( J )| =
2J
2
|1 J |4
J
By construction B(a J , n1 0 ) Tn is a line segment terminating at a J , so
that by the Schwarz reflection principle,
 n1  1
2 0 2
n1
2
z + ...
(6.6)
f J (a J + 0 z ) = J +
J ( J )
is univalent on D. By the growth theorem (I.4.16)
 n1  1
 2 0  2

n1
2
 |z| 1 + O(|z|) ,
| f J (a J + 0 z ) J | =  

J ( J )

(6.7)

uniformly in J . Set
U J =  J \ B(a J , n ) ,
and
V J = f J (U J ) D.
Tn
J

n
aJ
AJ

UJ

fJ
f J (A J )

Figure VIII.9 The map f J near a J .

D \ VJ

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. The Dandelion Construction and (c)  (a)

301

If J : V J D is conformal with J (0) = 0 and J (1) = 1, then by


Schwarzs lemma
1 | f (z)| 1 | J f J (z)|

(6.8)

for z U J . Set
A J = {z : 2 n < |z a J | < 3 n and dist(z, T ) n }.
Then for z A J ,
dist(z, U J ) c dist(z, T ),
so that by (6.8) and Koebes estimate, Theorem I.4.3,
| f  (z)| c|( J f J ) (z)|.

(6.9)

By (6.6) and the distortion theorem (I.4.17),


 n1  1
2
d

 f J (a J + n1 0 z 2 ) 1 2 0

dz
7 | ( J )|
for |z| 1/2, and hence by (6.5)
| f J (z)|

2J

C n

1
2

(6.10)

for z A J . It follows from (6.7) that V J D is approximately a semi-circle


centered at J , so that by the distortion theorem there is a constant c1 independent of J so that
| J (w)| c1
for w f J (A J ), and so by (6.9), (6.10), and (6.11)
 1
2J 2


| f (z)| C n
,

for z A J .
Set
W = f (|J |=n A J ) D.

(6.11)

(6.12)

CY535/Garnett

0 521 47018 8

302

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

Then by the distortion theorem, (6.12), (6.3), and the binomial theorem,

(1 |w|) p2+2 |  (w)| p d xd y
W

(1 | f (z)|) p2+2 | f  (z)| p+2 d xd y
=
|J |=n

AJ

|J |=n

C n

( n ) p2+2 | f  (z)|2 p+2 d xd y


AJ

2J p+

|J |=n
N

0
p+ 1n
C (1 ) p+
j
.
j=1

By (6.4)
(1 ) p+

N


p+

>1

j=1

for sufficiently small, so that



(1 |w|) p2+2 |  (w)| p d xd y = +.
D

(6.13)

1
Because 0 (1 r )1+ dr < for > 0, (6.13) implies B( p) > p 1 and

(c)  (a).

7. Baernsteins Example on the HaymanWu Theorem


Exercise VII.5 described the improvement, by Fernndez, Heinonen and Martio [1989], of the HaymanWu theorem: There is p > 1 and C > 0 such that
whenever  is a simply connected domain, f :  D is conformal, and L is
a line,

| f  | p d x C.
(7.1)
L

The example  = D \ [0, 1) shows that (7.1) fails when p = 2. A natural parallel exists between (7.1) and the Brennan problem; and in Havin, Hruv, and
Nikolskii [1984], Baernstein conjectured that (7.1) should hold for all p < 2.
However, in [1989] Baernstein employed a dandelion construction to show that
(7.1) fails for some p < 2.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. Baernsteins Example on the HaymanWu Theorem

303

Consider the tree


1

T = {(1 + 2 2 )ei : 0 } {i y : y 1 + 2 2 }
and its truncation
1

T0 = {(1 + 2 2 )ei : 0 }
1

at the fork point a0 = i(1 + 2 2 ). Let > 0 be a constant to be determined


later. Let T1 = T and by induction assume Tk has 2k finite extreme points a J ,
all of which fall on the line
"
k #
1

.
Imz = (1 + 2 2 )
1
Define
Tk+1 = Tk

a J + k (T0 a0 ) ,


and note that the induction hypothesis then holds for Tk+1 . Set T = Tk ,
1

 = C \ T , k = C \ Tk , and L = {Imz = (1 + 2 2 ) 1
}. Take
conformal maps f :  D and f k : k D such that f (i) = f k (i) = 0.
This construction differs from the dandelion construction in Section 6 because
T is not polygonal, but it is not hard to see that all of the estimates from Section
6 still hold. Alternatively, approximate T0 by a polygonal tree such that the final
1
line segments ending at z = (1 + 2 2 ) are vertical and such that inequality
(7.3) below holds.

a0

R
L
Figure VIII.10

CY535/Garnett

0 521 47018 8

304

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

Theorem 7.1. If is sufficiently small there exists p, 0 < p < 2 such that

| f  | p d x = .
(7.2)
L
1

Proof. The domain 1 = C \ T has two numbers, 1 = (1 , 1 + 2 2 , )


1
and 2 = (1 , (1 + 2 2 ), ). Set



$
2z 1
1+z
2
21
2
, 3 (z) = z 1, 4 (z) = i(1 + 2 )
,
1 = z , 2 =
z+1
1z
and
= 4 3 2 1 .
Then is a conformal map of H onto C \ T and
(z)
=1
z z 2
lim

and (2 4 ) = (1 + 2 2 ). By explicit calculation,

1
1
|  (2 4 )| = |  (2 4 )| = 8( 2 1),
so that

1+ 2
> 1.
(7.3)
2
We will use the notation from Section 6. Choose > 0 so that (1 )A > 1.
By Lemma 6.1 with p = 1, and the binomial theorem,

2J [(1 )A]n ,

(7.4)
A 1 + 2 =

|J |=n

for sufficiently small. Take such , fix n, and for |J | = n set


IJ = L AJ .
Then |I J |

C n and the intervals {I

: |J | = n} are pairwise disjoint. By (6.12)


 p

2J 2 n

 p
| f | dx c n
,

IJ
J

and by (7.4) if p 2


0
1n
p
| f | p d x =
| f  | p d x C 1 2 (1 )A .
I J

|J |=n

IJ

But there exists p0 = p0 () < 2 such that if p > p0


p

1 2 (1 )A > 1

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. Baernsteins Example on the HaymanWu Theorem


and therefore f 
/ L p (L).

305

Notes
See Zygmund [1959] and M. Weiss [1959] for more about lacunary series
similar to (1.2). Makarovs original [1985] proof of Theorem 1.1 used Fourier
series instead of Hardys inequality. For Theorem 2.1 see Makarov [1987] and
[1990]. Theorem 2.2 is from Makarov [1985], but we have followed the proof
in Makarov [1987]. A different proof by Rohde [1988] can also be found in
Pommerenke [1991].
In [1987] Makarov extended his compression theorem VI.5.1 to other Hausdorff dimensions. If is univalent on D and if E D, then for all p > 0,
dim(E)

p dimE
.
( p) + p + 1 dimE

See Exercise 5 for the proof.


Pommerenkes proof of (4.1) for p < 3.399 given Appendix L below has
been improved by Bertilsson [1998] and [1999] to yield p < 3.421. Shimorins
recent paper [2003] establishes Brennans conjecture up to p < 3.7858; Hedenmalm and Shimorin [2004] have improved Shimorins estimate further.
Define the bounded universal integral means spectrum to be


Bb (t) = sup (t) : is univalent and (D) is bounded .
Thus Bb (t) B(t). Makarov [1998] proves


B(t) = max Bb (t), 3t 1 .
See Exercise 9. Consequently there is a transition point t such that

3t 1, if t t ,
B(t) =
Bb (t), if t < t ,
and by Exercise 7, 13 < t < 25 . By Exercise 9(a), Bb (t) = t 1 for t 2.
Much deeper is the Jones and Makarov [1995] result that
Bb (t) = t 1 + O(t 2)2 , as t 2,
which implies Bb (2) = 1 and Bb (2) = 1. Near t = 0 the lower bound
Bb (t) = B(t) 0.117t 2 ,
from Makarov [1986a] and Rohde [1989], complements the upper bound (4.7),
and very recently the upper bound near t = 0 has been improved by Hedenmalm

CY535/Garnett

306

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

and Shimorin [2004] to


Bb (t) 0.437t 2 .
Carleson and Jones [1992] proved that
"
#

log |nan |
Bb (1) = sup lim sup
:=
an z n is bounded and univalent ,
log n
n
and conjectured that Bb (1) = 41 . Using the computer program Zipper, see
Marshall [1993], they obtained the estimate Bb (1) 0.21. Computer experiments by Kraetzer [1996] for polynomial Julia sets also support the Carleson
Jones conjecture, but the best results known to date without computer assistance
are
0.17 Bb (1) 0.4884
from Pommerenke [1975], Duren [1983] and Grinshpan and Pommerenke
[1997], though Hedenmalm and Shimorin [2004] have improved the upper
bound to 0.46 with the assistance of computer calculations. Brennans conjecture and the CarlesonJones conjecture would both follow from the beautiful
and general conjecture in Kraetzer [1996]:
t2
, for |t| 2.
(BCJK)
4
Pommerenke [1999] calls (BCJK) the BrennanCarlesonJonesKraetzer conjecture. Note that g(t) = t 2 /4 is the only quadratic polynomial that satisfies both
g(0) = g  (0) = 0 and g(2) = g  (2) = 1. It is not yet known if Bb (t) = Bb (t).
Binder [1997], [1998a] studied the counterpart of (BCJK) for complex parameters t: If is bounded and univalent, is

1  |t|2 +
4

i t
| (r e ) |d C(, )
1r
Bb (t) =

for all > 0 and all t C with |t| < 2? Makarov [1998] characterized those
functions (t) that have the form (t) for some bounded univalent function .
See Bertilsson [1999], Hedenmalm and Shimorin [2004], Makarov [1998], and
Pommerenke [1999] for more complete discussions of Brennans conjecture,
integral means spectra, and related topics.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

307

Exercises and Further Results


1. Let g(z) =

k=0

z 2 , z D and Sn ( ) =

n

k=0

2k ,

D. Prove

|Sn ( )|
lim sup $
=1
n log log n

(1.3)

holds if and only if


|g(r )|

lim sup
r 1

log

1
1r

1
log log log 1r

= 1.

2. Prove Hardys identity (1.6).


3. Pommerenke [1986b] shows that (1.1) fails with constant C = 0.685 by
studying
g(z) = C1

z 15

n=1

with C1 chosen so that ||g||B = 1.


4. Let  be a simply connected domain, let 0 , let K  be a
continuum, and let > 0 and > 0. There is 0 > 0 so that if 0 < < 0
and (, B(0 , ), ) , then there is a crosscut B(0 , 2) of 
such that
1
+
log .
dist (K , )

Hint: Use the proof of Lemma 2.5.


5. Let be a conformal map from D to a simply connected domain . The
following results are in Makarov [1987]. See also Pommerenke [1991].
(a) Let 0 < t 1. Show that Lemma 2.5 is still true when the conclusion
(2.8) is replaced with
Mr t ( 1 (D  \

N
%

j )) < Ct r M .

j=1

Hint: Follow the proof of Lemma 2.5 but use Pflugers theorem V.3.4. and
Corollary D.3 from Appendix D to replace (2.9) with
Mr t ( j ) Ct e( j ) .
(b) For all Borel E D and all p > 0,
dim(E)

p dimE
.
( p) + p + 1 dimE

CY535/Garnett

308

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

We may assume  is a Jordan domain. Take t < 1, set


s=

pt
>t
( p) + p + 1 t

and
=

ps
<s<1
p+s

and assume Mr s ((E)) = 0. Cover (E) by discs D = B( , r ) so that


 s
2
1
r < . For M = s let j and j , 1 j N () log
2 log( r ), be the
crosscuts and continua given by the version of Lemma 2.5 proved in (a).
Then by (a)


Mr t ( 1 (D  \

N
%

j )) < Ct .

j=1

On the other hand, let I j = 1 ( j ) and z j = (1 |I j |)cj where cj is


the center of I j . Then by Hlders inequality and Exercise IV.8,
N
N




1
1 
diam j
|I j || (z j )| C log( )
r
j=1

j=1


1
1
C log( )
r
r
because the j are disjoint subarcs of B( , 2r ). But taking r = 1 |I j |
we also have

|I j | p |  (z j )| p C|I j | p1 |(r ei )| p d C|I j |(( p) p1)
for any > 0. Together, these inequalities yield

|I j |t C,

which proves the assertion.


(c). Use (b) and (4.7) to prove dim(E) 1 whenever dimE = 1 and show
this result implies Theorem VI.5.1.
6. Prove (4.5) and (4.6).
7. Let be a univalent function on D and let t > 2/5. Then

1
.
|  (r ei )|t d C(t, )
(1 r )3t1
In other words, B(t) = 3t 1. For t > 1/2, use Exercise I.23(d). For
2/5 < t 1/2, set p = 2/(2 t), q = 2/t, and fix an integer k 2 such

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

309

1/k

and use Hlders inthat 2 pt (1 1/k) > 1. Then write g(z) = (z k)
i
pt
(k1)

d and |g (r ei )|2 d with care.
equality, but estimate |g(r e )|
See Feng and MacGregor [1976] for the details.
8. (a) Let be a conformal mapping from D onto a simply connected domain
 such that (0) = 0. The domain  and the map are said to be starlike
if the segment [0, w]  whenever w . This holds if and only if
z  (z) 
>0
Re
(z)
on D. More generally, and  are called close-to-convex if there exists
starlike such that
z  (z) 
Re
>0
(E.1)
(z)
on D, and is close-to-convex if and only if C \  is a union of closed halflines meeting only at their endpoints. See Duren [1983] or Pommerenke
[1991] for the proofs of these two geometric results.
(b) Let be starlike. Then for 0 < q < 2,

|  (z)|q d xd y < .
D

Thus Brennans conjecture is true for starlike univalent functions. For the
proof, use (E.1) and the facts that if F(z) is analytic on D and ReF(z) > 0,
then F H p for all p < 1 by Exercise A.4 and |F(z)| C/(1 |z|) by the
Schwarz lemma.
(c) If is close-to-convex and 43 < q < 4, then

|  (z)|q d xd y < .
D

Hint: Extend the proof of (b). Brennan [1978] attributes this result to B.
Dahlberg and J. Lewis.
(d) If is close-to-convex and E D is a Borel set, then
dim(E)

dimE
.
2 dimE

(e) If 0 < < 1, there is a starlike domain  bounded by a rectifiable Jordan


curve and E D such that  (E) > 0 and

.
dim(D)
2
See Makarov [1987] for (d) and (e).

CY535/Garnett

0 521 47018 8

310

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

9. The bounded universal integral means spectrum is


"
#
Bb (t) = sup (t) : is univalent and (D) is bounded .
Thus Bb (t) B(t).
(a) Suppose  = (D) is bounded. Then

4 dist (z), 

| (z)|
= o
2
1 |z|
1 |z|
and
1
|  (z)|

.
1 |z|


= O

Consequently when is bounded,


(t h) (t) + s, s > 0.
On the other hand, for r 21 ,


(1 r ) |  (r ei )|2 d 2
r

1 2

|  (tei )|2 d tdt 2 Area(),

so that (2) 1. It follows that


Bb (t) t 1
for t 2. Pommerenke [1999].
(b) Let  C be a simply connected domain such that  is bounded and
let : D  be conformal. Then
(t) Bb (t),
for all real t. Hint: Consider the restrictions of to two subarcs of the circle.
(c) Let be bounded and univalent and let = 1 r be small. Partition

and let z k be the center of Ik . Then


{|z| = r } into arcs Ik with |Ik | = 2r

 1
a t1
j
|  (r ei )|d
N (a j ),
(E.2)

where
N (a) = #{Ik : |  (z k )| a },
and
aj =

j
log 1

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

311

Because the sum (E.2) has O log 1 terms, it compares to its largest term
and we obtain

log N (a j )

(t) = lim sup sup a j t 1 +


.
(E.3)
log 1
j
0
Moreover, let  = (D), let 21 , and define
f  () = lim lim sup
0

log N (, , )

log 1

where
N (, , ) = #{disjoint B = B(, ), , + (B) }.
Then for =

1
1a ,

f  ()
log N (a)

log 1

(E.4)

Hence (E.3) gives


(t) = sup

and


f () t

1+t


F() t
Bb (t) = sup
1+t

(E.5)

where
F() =

sup

 bounded

f  ().

It is clear that f  () and F() are nondecreasing. See Makarov [1998] for
(E.5) and several related results.
(d) Makarov [1998] also showed


(E.6)
B(t) = max Bb (t), 3t 1 .
Therefore there is a transition point t such that

3t 1, if t t ,
B(t) =
Bb (t), if t < t ,
and by Exercise 7, 13 < t < 25 .
To prove (E.6), we can assume by (b) that = 1 where is a bounded
univalent function such that (0) = 1. We consider the arcs Ik and their

CY535/Garnett

0 521 47018 8

312

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

centers z k from (c). For 0 < b 2 and 1 < a < 1 be of the form
define

j
log

and

N (a, b) = #{(a, b) : |(z k )| b , |  (z k )| a }.


Then

|  (r ei )|t d

 1
(a+2b)t1
(a,b)

N (a, b).

In this sum the number of terms is O (log 1 )2 , and therefore we have



log N (a, b)

(t) = lim sup sup (a + 2b)t 1 +


.
log 1
0 (a,b)

(E.7)

To estimate N (a, b), fix a, b, and . By Lemma 2.5 we can assume


{z : |z| = 1 , |(z)| b , |  (z)| a }
is an arc I of length |I | = N (a, b) = and center c I . Let = 1ab ,
let z I = (1 |I )c I , and define
T (z) =

z + zI
1 + zI z

and
(z) =

(T (z))
.
b

| 1 and |
 | ab . Partition T 1 (I ) into N (a, b)
Then on T 1 (I ), |
1
 these
arcs of diameter (and harmonic measure) 1 = 1ab . Under
1ab
arcs have images of diameter
= . It follows that
log N (a, b)
log

(1 a b)F

.
1ab

(E.8)

By the Beurling projection theorem the worst case of (E.8) is given by = b2 ,


and then (E.7) and (E.5) yield (E.6).
10. (a) Let T be a polygonal tree with finite extreme points a1 , a2 , . . . , a N .
Suppose that f : D  = C \ T is a conformal map such that f (k ) = ak .
Prove that
2
N

1
d
1
2
=
.
lim (1 r )
 (r ei )|2 2
 ( )|2
r 1
|
f
|
f
k
0
k=1

Prove an analogous statement with 2 replaced by p.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

313

(b) Set k = (, ak , ), where  = C \ T and let : H  be a conformal map such that (z)/z 2 1 as z . For fixed t > 0, define


1+w
f (w) = f t (w) = it
: D .
1w
Prove

t r 1

2
 
N


 f (0)  d = 64
k2 .
 f  (r ei )  2

2 

lim lim (1 r 2 )
0

Prove

k=1


lim lim (1 r )

r 1 t

2
 

 f (0)  d = 64.
 f  (r ei )  2

2 

f (0))/ f  (0)

Note that ( f (w)


tends to the Koebe function uniformly on
compact subsets of the disc as t . Consequently, if the Koebe function
is a local maximum for the mean of 1/| f  |2 on each circle of radius r < 1,
then Brennans conjecture is true. See Bertilsson [1999], p. 85, for the best
result in this direction.
11. Let : H  be a conformal mapping.
(a) For every x R and every > 0, the nontangential limit
z x 


lim   
J (x) =
 (x) zx (z)
exists and is finite. Write  () = {x + i y H : |x| y}. Then also
  (z) 


lim
J () =


 () z
z
exists and is finite.
(b) If J (x) > 0 and J () > 0, then lim (x) zx (z) = a exists and
is finite, and lim () z (z) = , and
(, a, ) = J (x)J ().
(c) Conversely, if a  and (, a, ) > 0, then there exists : H 
such that lim (0) z0 (z) = a, J (0) > 0, and J () > 0.
See Bertilsson [1999] for (a), (b), and (c).
12. For p > 0, B( p) = p 1 if and only if there is a constant C such that for
all normalized univalent functions (z) = z + a2 z 2 + . . . and all r < 1,

1  p1
.
|  (r ei )| p d C
1r

CY535/Garnett

314

0 521 47018 8

January 27, 2005

15:41

VIII. Simply Connected Domains, Part Two

This observation is due to Bertilsson [1999]. Its proof is included in the


argument following (5.12) together with (c) implies (a) of Theorem 5.3.
13. (a) It is a deep and difficult theorem of Astala [1994] that if F : D D is
p
K -quasiconformal, then for p = K2K
1 , the Jacobian J F is weak L :
C
.
p
Eremenko and Hamilton [1995] have an alternate proof, also quite deep.
(b) Let  be a simply connected domain. Assume there exists G :  D
such that G is K -quasiconformal and locally Lipschitz,
Area{z D : J F (z) > }

lim sup
wz

|G(z) G(w)|
M < ,
|z w|

for all z . Then whenever :  D is conformal map and p <



|  (z)| p d xd y < .

2K
K 1 ,

G 1 , change variables and use (a). In particular, Bren-

Hint: Write F =
nans conjecture is true if, for every > 0 and every simply connected domain , there is a locally Lipschitz 2 + quasiconformal map of  onto D.
See Bishop [2002a] and [2002b].
(c) Suppose : D  is a conformal map, and suppose there exists a
K -quasiconformal h : D D such that h is bilipschitz with respect to the
hyperbolic metric on D and such that for some C < ,
1 |h(z)| C|  (z)|(1 |z|)
for all z D. Then G = h 1 satisfies the hypotheses in part (b). See
Bishop [2002a] and Exercise 2, Appendix J.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

IX
Infinitely Connected Domains

Let  be a domain such that  and Cap() > 0. In this chapter we study
the relation between Hausdorff measures and harmonic measure on . When
 is finitely connected, the problem is well understood by Theorem VII.2.2
and by simple comparisons, and thus we assume that  has infinitely many
components.
In Section 1 we give the proof by Batakis of the MakarovVolberg theorem
that if  = C \ K and K is a certain type of Cantor set then there exists F K
such that (F, ) = 1 but dimHaus (F) < dimHaus (K ). In Section 2 we prove
the result of Carleson [1985] and Jones and Wolff [1986] that if K = C \  is
totally disconnected and satisfies some reasonable geometric hypotheses then
there is F K such that (F, ) = 1 but dimHaus (F) < 1. Then in Section
3 we discuss the [1988] JonesWolff theorem that for every  there exists
F  with (F) = 1 but dimHaus (F) 1.

1. Cantor Sets
Fix numbers 0 < a < b < 1/2. For every sequence {an } with
a an b

(1.1)

form the Cantor set K = K ({an }) as follows:


*
K =
Kn ,

where K 0 = [0, 1] [0, 1], K 1 = 4j=1 Q 1j K 0 , and each Q 1j is a square of
side 1 = a1 containing a corner of Q 0 = K 0 . At stage n
%
Q nJ ,
Kn =

315

CY535/Garnett

0 521 47018 8

316

January 27, 2005

15:41

IX. Infinitely Connected Domains

where Q nJ is a square of side n = a1 a2 an having its sides parallel to


the axes, and J = ( j1 , j2 , . . . , jn ) is a multi-index of length |J | = n with

n+1
jk {1, 2, 3, 4}. Then Q nJ {Q n+1
J, j : j = 1, . . . , 4}, and the Q J, j are components of K n+1 containing the four corners of Q nJ . See Figure IX.1.
n
V(J,3)
n
W(J,3)

Q n(J,1)

Q n1
J

Q n(J,4)
Figure IX.1

Q n+1
(J,2,3)

Q n(J,2)

It is clear from (1.1) and Chapter III that Cap(K ) > 0. It is also clear that
K ({an }) has Hausdorff dimension


({an }) = sup > 0 : lim inf 4n n =
n


(1.2)
= inf > 0 : lim sup 4n n = 0
n

because the Hausdorff measure  (K ) can be estimated using coverings by


the squares Q nJ .
Set  = C \ K and write (F) = (, F, ).
Theorem 1.1. There exists F K such that
(F) = 1

(1.3)

dimHaus (F) < dimHaus (K ).

(1.4)

but

The proof of Theorem 1.1 depends on four lemmas. Write K Jn = K Q nJ


and take the squares
W Jn = (1 + c1 )Q nJ
V Jn = (1 + c2 )Q nJ

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Cantor Sets

317

concentric with Q nJ , where 0 < c1 < c2 are constants that depend only on a
and b and c2 is so small that for some constant c3 > 0
dist(V Jn1 , V Jn2 ) c3 n > 0
whenever J1 = J2 .
Lemma 1.2. There exist constants 1 > 0 and 2 > 0, depending only on a
and b, such that for all z W Jn
(z, K Jn , ) (z, K Jn , V Jn \ K Jn ) 1 ,
and for all compact E

(1.5)

K Jn

(z, E, ) (z, E, V Jn \ K Jn ) 2 (z, E, ).

(1.6)

Proof. Let g(z, w) be Greens function for V Jn . Then by a change of scale there
are constants C1 and C2 such that
(1.7)

g(z, ) > C1
whenever z W Jn and K Jn , and
g(z, ) log

1
+ log n + C2
|z |

(1.8)

whenever z, K Jn . Let be the equilibrium distribution for K Jn and consider


the Green potential

U (z) = g(z, )d( ).
Then U (z) is a positive harmonic function on V Jn \ K Jn . Since Cap(K Jn ) C3 n
by Chapter III, (1.8) gives 0 < U (z) C4 on K Jn . On the other hand, U (z) = 0
on V Jn . Therefore on V Jn \ K Jn we have
(z, K Jn , ) (z, K Jn , V Jn \ K Jn ) U (z)/C4

C1
C4

C1
by (1.7), and (1.5) holds with 1 = C
. A very similar argument was used in
4
Chapter III.
In (1.6) the leftmost inequality is obvious. To prove the rightmost inequality
take z 0 W Jn so that (z 0 , E, ) = sup W n (z, E, ). Then by the maxiJ

CY535/Garnett

318

0 521 47018 8

January 27, 2005

15:41

IX. Infinitely Connected Domains

mum principle and (1.5)

(z 0 , E, V Jn \ K Jn ) = (z 0 , E, )

V Jn

(, E, )d(z, , V Jn \ K Jn )

(z 0 , E, ) (1 1 )(z 0 , E, )
= 1 (z 0 , E, )


and (1.6) follows by Harnacks inequality.

Choose the index jn+1 so that for every J = ( j1 , j2 , . . . , jn ), Q n+1


(J,1) is the

upper right corner of Q nJ and Q n+1


(J,4) is the lower left corner.
Lemma 1.3. There is > 0 such that for all n 2,

n
n
(, K 1,1,...,1,1
, ) (1 + )(, K 1,1,...,1,4
, ).

Proof. For j = 1, 4 write


n
u j (z) = (z, K 1,1,...,1,
j , )

and
n1

n
U j (z) = (z, K 1,1,...,1,
j , C \ K 1,1,...,1 ).
n1
n
n
Let L be the diagonal of K 1,1,...,1
that separates K 1,1,...,1,1
from K 1,1,...,1,4
n
and let H be the half-plane bounded by L and containing K 1,1,...,1,4 . Then on
n1
W = H \ K 1,1,...,1
n
U4 (z) U1 (z) = (z, K 1,1,...,1,4
, W ).

L
H
n
K (1,...,1)

n
K (1,...,4)

n
W(1,...,4)

Figure IX.2 The proof of Lemma 1.3.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Cantor Sets

319

U4 U1 > 0

(1.9)

U4 U1 > c4 > 0

(1.10)

Therefore

n1
on W , while on K 1,1,...,1,4

with constant c4 independent of n, by Lemma 1.2 and Harnacks inequality.


Now

u 1 (z) u 4 (z) = U1 (z) U4 (z)
(U1 ( ) U4 ( ))d(z, , ).
n1
K \K 1,1,...,1

Hence by (1.9) and (1.10)


n1
u 1 () u 4 () c4 (, K 1,1,...,1,4
, ).

On the other hand, by Lemma 1.2 and Harnacks inequality


n1
, ) c5
(z, K 1,1,...,1,4
n
on W1,1,...,1,4
. Therefore u 1 () (1 + c4 c5 )u 4 () by the maximum prin
ciple and the lemma holds with = c4 c5 .

The next lemma is from Carleson [1985] and Makarov and Volberg [1986].
Lemma 1.4. Let  be a domain such that , and let
A1 B1 A2 B2 . . . An Bn
be simply connected Jordan domains such that for each j
j = Bj \ A j 
and
mod( j ) > 0.

(1.11)

Let u and v be positive harmonic functions on  such that


u = v = 0,

on

 \ A1 .

Then there exists C = C() > 0 and q = q() < 1 such that on  \ Bn ,


 u(z)3v(z)



3
(1.12)
1 Cq n .

 u() v()


CY535/Garnett

0 521 47018 8

320

January 27, 2005

15:41

IX. Infinitely Connected Domains

B3

1

3

2

A3

Figure IX.3
Proof. By (1.11) there is a conformal mapping
j :  j {1 < |z| < 1 + j }
with ( A j ) = {|z| = 1} and j = e2 1. For k = 1, 2 set

k j

}
j(k) = j1 {|z| = 1 +
3
and

3
(k)
max{u(z) v(z) : z j }
.
=
3
(k)
min{u(z) v(z) : z j }

a j(k)

By Harnacks inequality there is C = C() such that


(k)

C.

1 aj
We claim there is 3 > 0 such that
(2)

aj

(1)

aj

(1)

1 3 (a j

1) .

(1.13)

Let us accept (1.13) for the moment and complete the proof of (1.12). By
the maximum principle,
a1(1) a1(2) a2(1) . . . an(2)
(2)

because u = v = 0 on  \ A1 . Write a j

= 1 + b j . Then by (1.13)

1 + b j+1 (1 + b j )(1 b j ) 1 + (1 )b j

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Cantor Sets

321

and bn C(1 )n1 . Since on  \ Bn


4
u(z)  u() 
(1 + bn ),
(1 + bn )1
v(z)
v()
we therefore obtain (1.12).
To prove (1.13) we can assume min (1)
j

u(z)
v(z)

= 1. Then on j(1) ,

(1)

v(z) u(z) a j v(z) = (1 + b)v(z)


where b > 0 by Harnacks inequality. Set
b
{u (1 + )v(z)}
2
b
(1)
E 2 = j {u > (1 + )v(z)}
2
(1)

E1 = j

(1)

and let z denote harmonic measure at z for the component of  \ j


taining

j(2) .

con-

j(2) ,

Since u = v = 0 on  \ A1 we have, for z




b
u(z) (1 + )
v( )dz ( ) + (1 + b)
v( )dz ( )
2 E1
E2

and


v(z) =

E 1 E 2

v( )dz ( ).

Hence
b
u(z) (1 + b)v(z) z (E 1 ) min v( )
E1
2
and since by Harnacks inequality v(z) C1 min E 1 v( ), with C1 = C1 (), we
obtain

C1 bz (E 1 ) 
(1)
.
u(z) v(z)a j 1
2(1 + b)
Similarly,
b
u(z) v(z) + z (E 2 ) min v( )
E2
2


v(z) 1 + C2 ()z (E 2 ) ,
and we conclude that
(2)

aj

(1)

(1)

1 C3 ()bz ( j )


(1)
a j 1 C4 ()b .

aj




CY535/Garnett

322

0 521 47018 8

January 27, 2005

15:41

IX. Infinitely Connected Domains

Lemma 1.5. There is N = N (a, b) such that for every n and J there exists
J  = J  (J ) = (J, jn+1 , . . . , jn+N ) such that
1 (K Jn )
.
3 4N
Proof. Fix J and N and for j = 1, 4 write
(K Jn+N
)


(1.14)

n+N
v j (z) = (z, K J,1,...,1,
j , )

and
n+N
n
n
Vj (z) = (z, K J,1,...,1,
j , V J \ K J ).

Let z W Jn . By Lemma 1.4,



 

 V (z)3V (z)
  V (z)3V (z)

4
4
 1
  1

3
3
1 
1 + Cq N .

 v1 () v4 ()
  v1 (z) v4 (z)

But




 V (z)3V (z)

v4 (z)  V1 (z) V4 (z) 
4
 1

3

1 =



 v1 (z) v4 (z)
 V4 (z)  v1 (z)
v4 (z) 



v4 (z)
 v4 ( ) v1 ( ) 
=


d(z, , V Jn \ K Jn )
V4 (z) V Jn  v4 (z)
v1 (z) 


3


v4 (z)
v4 ( )  v1 ( ) v1 (z)

3
=
1d(z, ,V Jn \ K Jn )


V4 (z) V n v4 (z)  v4 ( ) v4 (z)
J

1
(1 + Cq N )Cq N ,
2
by Lemmas 1.2 and 1.4.
If we take J = (1, 1, . . . , 1) we get
v1 ()
1 + ,
v4 ()
by Lemma 1.3 and if N is large,
V1 (z)

1+
V4 (z)
2
by the two preceding estimates. However, VV41 (z)
(z) does not depend on J . Therefore
we have
v1 ()

>1+
v4 ()
4

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. Cantor Sets

323

for all J if N is large. Consequently


inf
J

(K Jn+N
) 4 N +1  N

N +1
4
n
(K J )
4
+

and (1.14) holds if N is replaced by N p , where p = p(, N ).

Conclusion of Proof of Theorem 1.1. The proof uses the central limit theorem
exactly as in the hypothesis testing section of an elementary statistics course.
1 N
On any K Jn take two probabilities P1 and P2 such that P1 (K Jn+N
 (J ) ) = 3 4

N . Write J = J (z) when z K n . Let n be large, take


and P2 (K Jn+N
 (J ) ) = 4
n
1
J
n j+1 = n j + j and set
(n +m)N +N

"

Ej = z K :

j
#{m j : z K J  (J
(z))

2 N #
.
4
3

By the central limit theorem for the binomial distribution, there are constants
C1 and C2 > 0 (depending on N ) such that if n 1 is large
P1 (E j ) eC1 j

(1.15)

P2 (E j ) 1 eC2 j .

(1.16)

and

Set F =

k=1

Ak , where
Ak =

K \ Ej .

(1.17)

j=k

Then
K\F=

Ej ,

k=1 j=k

and by (1.14) and (1.15)


(K \ F) lim

eC1 j = 0.

j=k

On the other hand, since the E j are independent it follows from (1.16) and
(1.17) that for j > k, Ak can be covered by eC2 (n k n j ) 4n j N squares of side
n j N . Let = dimHaus (K ), let a be as in (1.1), and let 0 < < satisfy
eC2

1
N ( )
a

<1

CY535/Garnett

324

0 521 47018 8

January 27, 2005

15:41

IX. Infinitely Connected Domains

for some > . Then by (1.1) and (1.2),

 (Ak ) lim eC2 (n k n j ) 4n j N n j N = 0.


j

Therefore  (F) = 0.

2. For Certain , dim < 1


Let K C be compact and let  = C \ K . We assume there are constants ,
C1 , and C2 such that for all z 0 K and all 0 < r < diam(K )
Cap(B(z 0 , r ) K ) C1r,

(2.1)

and assume there is a ring domain Ar = A(z 0 , r )  {r < |z z 0 | < r }


bounded by rectifiable Jordan curves 1 and 2 such that
( j ) C2 r

(2.2)

dist(1 , 2 ) r.

(2.3)

and

Ar

r/

Figure IX.4
Theorem 2.1. Assume K satisfies (2.1), (2.2), and (2.3). Then there exists
= (, C1 , C2 ) > 0 and there exists F K such that (, F, ) = 1 but
dimHaus (F) 1 .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. For Certain , dim < 1

325

The proof of Theorem 2.1 consists of three steps. We first construct a dyadic
approximation of K by sets Q n, j called cubes. Next we employ a stopping
time to select certain of the cubes Q n, j . Then a lower bound for the harmonic
measure of the stopped Q n, j is obtained from an analysis of the critical points
of Greens function. Assume diam(K ) = 1 and set (E) = (, E K , ).
Throughout the proof c and C denote constants which may change with each
occurrence but which depend only on , C1 , and C2 .
Lemma 2.2. For n = 1, 2, . . . there exist simply connected Jordan domains
Q n, j such that
%
K
Q n, j ,
(2.4)
j

Q n, j {z : dist(z, K ) < 4n },

(2.5)

dist( Q n, j , K ) c4n ,

(2.6)

diam(Q n, j ) c

4n
,

(2.7)

dist(Q n, j , Q n,k ) c4n ,

(2.8)

dist( Q n, j , Q n+1,k ) c4n .

(2.9)

and

Moreover, Q n, j is a rectifiable Jordan curve of class C 3 with arc length parameterization z = z(s) satisfying
 d3z 
 d2z 




(2.10)
 2  C4n and  3  C42n .
ds
ds

Q n+1,k

Figure IX.5

Q n, j

CY535/Garnett

326

0 521 47018 8

January 27, 2005

15:41

IX. Infinitely Connected Domains

Proof. Let {Sn, j } denote the components of {z : dist(z, K ) < 4n } and let
z 0 K Sn, j . Then by (2.3) Sn, j is contained in the bounded component of
n
C \ A(z 0 , 4 ), and for k = j,
Sn,k A(z 0 ,

4n
) = .

By (2.2), (2.3), and Exercise IV.5(d),


mod(A(z 0 ,

log Rn, j
4n
)) =
C.

2
n

Let f n, j be a conformal mapping from {1 < |z| < Rn, j } onto A(z 0 , 4 ) and let
$
n
n, j be the core curve f n, j ({z : |z| = Rn, j }) of A(z 0 , 4 ). Then by Koebes
theorem there is c = c() such that
dist(n, j , K ) c4n

(2.11)

and n, j satisfies the smoothness condition (2.10). Let {Un, j } be the bounded

components of the complements of the components of n, j . Then each Un, j
is simply connected. Let {Vn, j } be the set of Un, j such that K Un, j =
and Un, j is maximal with respect to set inclusion. Then by (2.11) and the
smoothness of n, j there exists Q n, j Vn, j satisfying (2.4)(2.10) such that

K Vn, j Q n, j .
Like dyadic squares, the cubes Q n, j have the property
Q m,k Q n, j or Q m,k Q n, j =
if m > n. Furthermore
K =  =

*%
n

Q n,k .

Now fix integers M and m and set




B M = Q n,k : n m, (Q n,k ) M4n and Q n,k is maximal ,


G M = Q m, j : Q m, j Q n,k for all Q n,k B M
and
 = C \


%

%
(Q n,k ) (Q m, j ) .
BM

GM

 = g
).
(E) = (, E, 
Write G(z)
 (z, ) and 

(2.12)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. For Certain , dim < 1

327

Lemma 2.3. There is a constant c = c() such that if Q n, j B M G M , then



(Q n, j ) c(Q n, j ),

(2.13)

,
and on 

G
cM.
(2.14)
n
n, j Q n, j and
n, j Q n, j with K Q
Proof. By Lemma 2.2 there is open Q
n, j by
n, j , Q n, j ) c4n . Then (z, K Q n, j , ) c at all z Q
dist( Q
(2.1) and the proof of Lemma 1.2. Then (2.13) follows by the maximum prinn, j .
) on 
\Q
ciple, since (z, Q n, j , ) 0 = (z, Q n, j , 
n+1
for all Q n,k B M G M and by (2.10)
By construction (Q n,k ) M4
sup

Q n,k



G
G
c inf
.
Q n,k n
n

Hence by (2.13) and Lemma 2.2,


sup

Q n,k


2 c
G


( Q n,k ) cM,
n
( Q n,k )


so that (2.14) holds.


The heart of the proof of Theorem 2.1 is the identity




G
G
1
 j ),
log
ds = +
G(z
2 
n
 n

(2.15)

 and = log Cap( 


) is Robins
in which {z j } is the set of critical points of G
. See Exercise III.2 for the proof. We assume Cap(K ) < 1 and
constant for 
take M so large that > 0 in (2.15). Then (2.14) and (2.15) yield

 j ) log(cM).
(2.16)
G(z
j

Now let N = N () be a positive integer to be determined later and define




U N = Q n, j : Q p,k Q n, j and Q p,k B M G M p > n + N .
If N () is sufficiently large, it follows from (2.1), (2.6), and (2.7) that whenever
/ B M G M such
Q n, j U N there exist two disjoint cubes Q n+N ,k1 , Q n+N ,k2
that
Q n+N ,k1 Q n+N ,k2 Q n, j .
Lemma 2.4. There exist constants c and N such that

(2.17)

CY535/Garnett

0 521 47018 8

328

January 27, 2005

15:41

IX. Infinitely Connected Domains

(a) If Q n, j U N then
 c
inf G(z)
(Q n, j ).

(2.18)

Q n, j

(b) If Q n+N ,k Q n, j , then



c inf G
Qn, j

 1 inf G.

sup G
2 Q n, j
Q n+N ,k

(2.19)

n, j Q n, j such that Q


n, j \ Q n, j 
 and
Proof. (a) There is Q
n, j , Q n, j ) c dist( Q n, j , 
).
dist( Q
n, j and by Harnacks in) c on Q
By (2.1) and Lemma 2.2, 
(z, Q n, j 
n, j . Hence

equality and a comparison G(z, ) > C for all z Q n, j and Q
for z Q n, j ,


G(z)
G(z, )d
(Q n, j ).
n, j (, ) c
\ Q
n, j
Q

(b) Write
Q n, j = Q 0 Q 1 Q 2 . . . Q N = Q n+N ,k
where Q p = Q n+ p,k for some k. By (2.1) and Lemma 2.2 there is > 0 such
that for all p and all z Q p ,
) 1 .
< (z, Q p1 , Q p1 
Hence by induction
) (1 ) N .
N < (z, Q 0 , Q 0 
Also, by Lemma 2.2 and Harnacks inequality,
 C inf G(z),

sup G(z)
Q0

Q0

 = 0 on Q 0 
,
so that since G
 sup G
 C(1 ) N inf G,

c N inf G
Q0

Q0

QN

and if
C(1 ) N <
then (2.19) holds.

1
,
2

(2.20)


CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. For Certain , dim < 1

329

Fix N so that (2.17) and (2.20) both hold, and note that the lower bound in
(2.19) holds with c = c(, N ).
 such that
Lemma 2.5. If Q n, j U3N , there is a critical point z n, j of G
%
z n, j Q n, j \
Q n+3N ,k
and
 n, j ) c
(Q n, j ).
G(z
Proof. There are 4 cubes Q n, j Q n+N , j1 Q n+2N , j2 Q n+3N , j3 , which
/ B M G M and by Lemma 2.4
we relabel as Q 1 Q 2 Q 3 Q 4 , so that Q k
 < 1 inf G.

sup G
2 Qk
Q k+1
 < b} 


.
Take a = sup Q 1 G(z),
b = inf Q 4 G(z),
and U = Q 1 {a < G
2
3

By (2.17) and (2.19) Q \ Q contains a component of U {G = a} and
 in U and by (2.18)
hence by Exercise II.17 there is a critical point z n, j of G
and (2.19)
 c
 n, j ) c inf G
(Q n, j ).
G(z
Q n, j

Proof of Theorem 2.1. By (2.8) and (2.12) each Q p, j B M G M falls in


p N cubes Q n,k U M . Therefore by (2.16) and (2.18),


n
(Q n,k ) + c
m
(Q m,k ) log(cM),
c
BM

GM

and we can take m large and get





(Q m,k )

GM

Set

so that

"
#
B M = Q n,k B M : (Q n,k ) 4
(Q n,k )



(Q n,k )

B M \B M

Then

1
.
4


B M


(Q n,k )

1
.
4

1
2

CY535/Garnett

0 521 47018 8

January 27, 2005

330

15:41

IX. Infinitely Connected Domains

and on B M

(Q n,k )
1

c
4
(Q n,k )
by (2.13). Therefore


(Q n,k )

B M

1
.
2c

Each Q n, j B M can be covered by a ball of radius


c4n rn, j C4n
by Lemma 2.2 and
M4n (Q n, j ) M4n+1
by the definition of B M . Therefore if M is large,

1
C log M

rn, j log
rn, j
M
and


rn, j

 (Q n, j )
M

1
.
2M

Choose so that


rn, j

rn, j

1
.
4M

Then
M c

because by (2.21)
log 1
2M

rn, j log

rn, j <

1
C log M
.

rn, j
M

Consequently


(rn, j )1

rn, j

while


rn, j

c 

(Q n, j ) C M 1c ,
M

(Q n, j ) c.

(2.21)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. For All , dim 1

331

Taking > 0 small, we see there exists, for n = 1, 2, . . . a sequence Bn, j of


balls of radii rn, j such that
%
( Bn, j ) c
j

and

rn, j 2n .

Then
E=

%
%
*


Bn, j

m=1 n=m j

satisfies
(E) c
and
dimHaus (E) 1 .
By Harnacks inequality there is c such that
(z, E, ) c
at all z j Q 1, j . Now for every Q n, j given by Lemma 2.2 this construction
yields a set E n, j K Q n, j such that
inf (z, E n, j , ) c

Q n, j

and
dimHaus (E n, j ) 1 .
Now set
F=

E n, j .

n, j

Then dimHaus (E) 1 , and by the maximum principle (z, F, ) c for



all z , which implies that (F) = 1.

3. For All , dim 1


Let K C be a compact set of positive capacity and set  = C \ K .

CY535/Garnett

0 521 47018 8

332

January 27, 2005

15:41

IX. Infinitely Connected Domains

Theorem 3.1. There exists F  such that (F) = 1 and dimHaus (F) 1.
The proof consists of two lemmas and a domain modification construction.
 be closed squares with the same center 0 , with sides
Lemma 3.2. Let Q Q
 = R(Q). Assume
parallel to the axes, and with sidelengths (Q) and ( Q)
\ Q 
Q
and
E = Q\
has Cap(E) > 0. Fix > 0 and let B be a closed disc with center 0 such that

1+
e B = Cap(B) = ((Q)) Cap(E)
= ((Q)) e(1+) E . (3.1)
Set
 = ( Q) \ B

and
).

= (, , 
Then if R = R() is sufficiently large,

(B) C()(E),

(3.2)


(A) (A).

(3.3)


and for all A  \ Q,

A

Q

A

Q





Figure IX.6

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. For All , dim 1

333

 \ Q . We may assume
 are open because Q
Proof. The sets  Q and 
 > 0, and after a change of scale we may assume (Q) = 1. Let
Cap( \ Q)
g(z, ) = log

1
+ h(z, )
|z |

be Greens function for  E. Then



R 1

h(z, 0 ) =
,
log | 0 |d\ Q(z, ) log
2

\ Q

(3.4)

and
1

| h(z, )| = O
R
when B E. Let E and B be the equilibrium distributions for E and B
and fix z 0 Q. Then on B,

1

,
u(z) =
g(z, )d B ( ) = B + h(z 0 , 0 ) + O
R
B
and on E,

.
g(z, )d E ( ) = E + h(z 0 , 0 ) + O
R
E

 and u(z) v(z) + O 1 on  \ Q


 (beFurthermore u = v = 0 on  \ Q
R

cause E > 0). Hence on  \ Q,
v(z) =

)
(z, B, 

u(z)

B + h(z 0 , 0 ) + O R1

v(z) + O R1

B + h(z 0 , 0 ) + O R1

E + h(z 0 , 0 ) + O R1


(z, E, ),
(1 + ) E + h(z 0 , 0 ) + O R1

by (3.4), and with (3.4) that implies (3.2).


To prove (3.3) we assume 0 = 0. Write U = {|z| <
1 z 


g(z, ) = log z R R 
R R

R
2 },

let

CY535/Garnett

0 521 47018 8

334

January 27, 2005

15:41

IX. Infinitely Connected Domains

be Greens function for U , let g B be Greens function for U \ B, and let g E be


Greens function for U \ E. Then on U \ B

g B (z, ) = g(z, )
g(w, )dU \B (z, w)
B


C2 
(z, B, U \ B),
g(z, ) log 2 +
R
and on U \ E

g E (z, ) = g(z, )

g(w, )dU \E (z, w)


C1 
(z, E, U \ E).
g(z, ) log 2 +
R
If |z| = R/2 and if R is large, then by hypothesis (3.1)
(z, B, U \ B) (z, E, U \ E),
and hence
g B
g E

n
n

(3.5)

at bU.
Now suppose
(z, A)
(z 0 , A)
= sup
= > 1.

(z 0 , A) |z|=R/2 
(z, A)
Then by the maximum principle

(z, A) (z, A) > 0
on U, because 
> on A. However, using (3.5) at z 0 then yields
0 = 
(z 0 , A) (z 0 , A)


1
1
g B (z 0 , )
g E (z 0 , )

(, A)ds( )
(, A)ds( )
=
2 U
n
2 U
n

1
g B (z 0 , )
>

(, A) (, A) ds( )
2
n
0,
which is a contradiction, and therefore (3.3) follows from the maximum principle.

Lemma 3.3. Write g(z) = g (z, ). Suppose  {|z| < 1} is a finite union
of Jordan curves that separate K = C \  from and suppose there are con-

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. For All , dim 1

stants c1 , . . . c N , possibly not distinct, such that  =

335
N

j=1  j

where

 j {g(z) = c j }
is a Jordan curve. Then
I () =

1
2

g
log |g|ds log 2.
n

(3.6)

2

1
K

3

Figure IX.7
Proof. Replacing K by {g } for small , we can assume  is a finitely
connected domain with smooth boundary. Let k be the critical points of g that
lie outside . Then by Greens theorem, almost exactly as in Exercise III.2,
 1


g(k ) +
g log |g|ds +
I () =
2  j n
k
j

 cj
=
log |g|ds + ,
g(k ) +
2  j n
k

where = () (K ) is Robins constant for . Let j be the (finite) number


of critical points of g inside  j and let j be the (finite) number of components
of K inside  j . Then j = j 1, by Exercise II.17 and

1
log |g|dz = j j = 1
2  j n
by Greens theorem. Therefore
I () =


k

g(k )


j

cj + .

CY535/Garnett

0 521 47018 8

336

January 27, 2005

15:41

IX. Infinitely Connected Domains

Now write n(c) for the number of components of {g = c} outside . Then




g(k ) =
(n(c) 1)dc
0

and

j

cj =
0

{cc j } (c)dc

c0

n(c)dc,
0

where c0 = sup{c : n(c) > 1}. Hence


I () c0 .
But since  {|z| < 1}, all critical points fall into {|z| < 1}, and if |z| < 1 then

g(z) = +

log |z |d( ) + log 2.


K

Hence c0 + log 2 and (3.6) follows.

Proof of Theorem 3.1. We now modify the domain  = C \ K . We assume


K {|z| < 1/2}. Fix > 0 and fix an integer R so that R > 2 + R() where
R() is given by Lemma 3.2. Choose a large constant M and a small constant
> 0 so that M log 1 . Consider the grid G of dyadic squares having side
and lower left corners (m + ni). Partition G into R 2 subfamilies G p,q , where
1 p, q R, by the rule
G Q G p,q (m, n) ( p, q) mod (R R).
Write
K p,q =

K Q,

G p,q

 p,q = C \ K p,q ,
and
p,q (E) = (, E,  p,q ).
Then
(E)

i, j (E K p,q ),

and to prove Theorem 3.1 we may assume at the start that  =  p,q for some
p, q. To modify  =  p,q we alternate two constructions:

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. For All , dim 1

337

disc construction: Let Q be a square of side . Replace E = K Q by a closed


ball B with the same center as Q and with capacity

1+
Cap(B) =  CapE
.
 with a new harmonic measure 
Replacing E by B yields a new domain 
and
(3.2) and (3.3) hold for 
if  (R Q \ Q) = .
annulus construction: Given a square Q of side (Q) with sides parallel to
 =  (R Q \ Q) and K by K \ (R Q \ Q) . This
the axes, replace  by 
) which satisfies 
yields a new harmonic measure 
= (, , 
(A) (A)
whenever A (R Q \ Q) = .
To begin, take  =  p,q and perform the disc construction on every square
Q j G p,q . This yields a new domain 1 , whose complement is a union of discs
B j , and a new harmonic measure 1 .
Now choose a maximal dyadic square Q 1 such that (Q 1 ) and
1 (Q 1 ) M(Q 1 ).
If no such Q 1 exists, the domain modification stops. If Q 1 exists, perform the
annulus construction on Q 1 , obtaining a new 1 , and then perform the disc
construction on Q 1 to 1 , replacing K Q 1 by a disc B 1 . Denote by 2 and
2 the new domain and harmonic measure thusly obtained.
Next choose a maximal dyadic square Q 2 such that (Q 2 ) ,
Q 2 Q 1 ,
and
2 (Q 2 ) M(Q 2 ).
If no Q 2 exists, stop. If Q 2 exists, perform the annulus construction on Q 2 .
However, if B 1 (R Q 2 \ Q 2 ) = , do not remove B 1 R Q 2 \ Q 2 from 2 .
Then perform the disc construction on Q 2 to 2 , replacing K Q 2 by a disc
B 2 , and getting a new domain 3 with harmonic measure 3 .
By induction assume we have replaced squares Q 1 , . . . , Q n1 by discs
1
B , . . . , B n1 and formed new domains 1 , . . . , n . Choose a maximal dyadic
square Q n such that (Q n ) ,
Q n Q j , for all j n 1,
and
n (Q n ) M(Q n ).

CY535/Garnett

0 521 47018 8

338

January 27, 2005

15:41

IX. Infinitely Connected Domains

If no Q n exists, stop. If Q n exists, perform the annulus construction on Q n


to n . However, if j n 1 and B j (R Q n \ Q n ) = , do not remove
B j R Q n \ Q n from n . Then perform the disc construction on Q n to n ,
replacing K Q n by a disc B n , and getting a new domain n+1 with harmonic
measure n+1 .
Because there are only finitely many candidate squares and no square Q j is
repeated, the modification stops after a finite number of steps at a final domain
 = C \ K with harmonic measure . The complement K is the union of
finitely many disjoint nonremoved discs
{B k : k S} {B j : j T }.
If k > j and B j (R Q k \ Q k ) = , then B j 2Q j and (Q k ) (Q j ).
Hence B j R()Q k = , and (3.2) and (3.3) hold for each stage of the construction. Consequently
(B k ) C()(Q k )
for all k S, and
(Q j ) C()(Q j )
for all j T. By the construction
(Q k ) cM(Q k )

(3.7)

for all k S, and


({|z z 0 | < r }) C Mr
if z 0 Q = Q j or Q j and r (Q).
Lemma 3.4.
K

%
S

2R Q k

Qj.

(3.8)

Proof. Let Q G p,q and set E = K Q. If Q = Q j , j T then (3.8) is


obvious. If not there is a first index j1 such that Q 0 R Q j1 \ Q j1 and Q 0
gets removed. If j1 S, then (3.8) is true, and if j1
/ S there is a first index
j2 > j1 such that Q j1 R Q j2 \ Q j2 . However, in that case (Q j2 ) 2(Q j1 ),
because if (Q k ) = (Q j ) then
Q j R Q k Q k R Q j .
Since there exists an increasing chain j1 < j2 < . . . < jn with jn S and

Q jk R Q jk+1 \ Q jk+1 , we conclude that Q 2R Q jn .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. For All , dim 1

339

Now let Q = Q k , k S or Q = Q j , j T, let B be the corresponding disc


and assume z 2Q \ B. Then Greens function g(z) = g (z, ) has the form

g(z) =

log |z |d ( ) +

K \B

log |z |d + (K )

= u(z) + v(z) + (K ).
By (3.7) and a partial integration

d ( )
1
|v(z)|
C M log
.
(Q)
K \B |z |
Let r denote the distance to the center z 0 of Q, and let r (B) be the radius of B.
Then simple calculations give

d ( )
|u(z)|
B |z |
and
u
=


Re
B

(z )z 
d ( )
|z |2 |z|

(B)
r (B) (B)

.
|z z 0 |
|z z 0 |2

Set
= (B)Max

(B)
1
M 2 log (Q)


, 2r (B) .

If > 2r (B), then | u


r | > |v| on the circle = {|z z 0 | = } and the level
set g = inf g(z) is a closed curve separating B from . On this level curve
|g(z)|

(B)
1
C M 2 log
.

(Q)

(3.9)

If = 2r (B) we take = B {g = 0} for the level curve. Then because


R()Q \ B  , sup B |g| C inf B |g| by reflection, and (3.9) also
holds for .

Now  = ST separates K from and



g
1 
1
log+ |g|ds C
(B) log+ C M 2 log
2  n
(Q)
ST

C  log log

CY535/Garnett

0 521 47018 8

January 27, 2005

340

15:41

IX. Infinitely Connected Domains

because M log 1 . Therefore by Lemma 3.3,




1
1
g 
log |g|ds C log log .
2  n

Consider the measure function h(r ) = r 1+ and the Hausdorff content Mh


defined in Appendix D. By (3.7)

%
2R Q k ) (2R)1+
((Q k ))1+
Mh (K
S

C 1+

(Q k )

C R 1+
.
M

Let r j be the radius of B j , j T, and set


T1 = { j T : (B j ) /2 r j }.
By (3.1) and (D.11) from Appendix D
Cap(K Q ) 
j
Mh (K Q j ) r j1+ log
rj
for any j T. Therefore

%

(K Q j ) C(1 + ) /2
(B j ) C /2 .
Mh
T1

T1

Finally, we have

%
(K Q j )
T2

1 
(B j ),
C()
T2

and by (3.9)|g| on B j . Consequently



 1
g
ds
(B j ) =
2 B j n
T2
T2


C 
g 

log |g|ds
1
log
B j n
T2

C log log 1
log 1

Therefore, with fixed and > 0 given, we can choose M and so that there
exists A K such that Mh (A) < while (K \ A) < . That means h ,

and the Theorem follows.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Notes

341

Notes
Section 1 follows the elementary arguments in Batakis [1996]. Earlier Makarov
and Volberg [1986] had given an explicit formula for
dim = inf{ :  }

(N.1)

in terms of the critical points of Greens function. See also Volberg [1992],
[1993] and Lyubich and Volberg [1995]. These papers rely on a beautiful connection between information theory and harmonic measure for self-similar sets
originally due to Carleson [1985] and to Manning [1984] in the case of Julia
sets that we regretfully must omit from this book. See Makarov [1990a] and
[1998] for complete details and several further related results.
Theorem 2.1 is an unpublished result of Jones and Wolff [1986]. Carleson
[1985] had obtained a similar theorem for certain Cantor sets using information
theory and the identity (2.15) on Greens function critical points. Jones and
Wolff later made a direct proof of the more general Theorem 2.1, still using
(2.15). The Greens function identity (2.15) itself first appeared in Ahlfors
[1947] paper on analytic capacity. See Widom [1969], [1971a], and [1971b]
and Jones and Marshall [1985] for other applications.
In the substantial paper [1993], Wolff extended Theorem 3.1 to show that
for all  there is F  having finite linear measure and (F) = 1.
Makarovs paper [1998] and the recent work [2003] by Binder, Makarov,
and Smirnov explore the connection between other Hausdorff measures and
harmonic measure in general domains. Let be harmonic measure for some
domain  and define the lower pointwise dimension of at z  as
(z) = lim inf
0

log (B(z, ))
.
log

Set
f () = dim{z  : (z) }

(N.2)

and define the universal dimension spectra to be


() = sup f ()


and
SC () =

sup

 simply connected

f ().

Makarov [1998] conjectures that for all 1,


() = SC ().

(N.3)

CY535/Garnett

342

0 521 47018 8

January 27, 2005

15:41

IX. Infinitely Connected Domains

Conjecture (N.3) is connected to the Brennan conjecture because Makarov


[1998] proves that


SC () = inf (1 + Bb (t)) + (1 )t , ( 1),
t0

where Bb (t) is the bounded universal integral means spectrum defined in Exercise VIII.9.
To prove (N.3) it is enough to work with certain Cantor sets. Let D be a
simply connected domain, let {D j } be a finite collection of pairwise disjoint

Jordan domains such that D j D, and suppose
%
F:
Dj D

is an analytic function such that each restriction F  D j is univalent. Write
F (n) = F F . . . F
for the n-th iterate of F. The set
K =

Fn

1 %


Dj

n=1

is called a conformal Cantor set. Carleson and Jones [1992] and Makarov
[1998] show that (N.3) holds if and only if
f () SC ()

(N.4)

whenever C \  is a conformal Cantor set. In [2002] Binder, Makarov, and


Smirnov established (N.4) for those conformal Cantor sets for which F is a
polynomial.

Exercises and Further Results


1. (a) Let K be a Cantor set satisfying the hypotheses of Theorem 1.1. Then
there exists < dimHaus (K ) such that there is F K with dimHaus (F) =
and (F) = 1, but such that if E K has dimHaus (E) < then (E) = 0.
(b) However, if the definition of the Cantor set K at the beginning of Section
1 is generalized so that each square Q nJ is replaced by four corner squares
Q nJ, j , j = 1, 2, 3, 4, having sidelengths
(Q nJ, j ) = a nJ (Q nJ )
that depend on J but satisfy
0 < a a nJ b < 1/2,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

343

then there is an example such that


dimHaus () = dimHaus (K ).
This is due to Bishop and independently to Ancona and Batakas.
(c) If the Cantor set K is constructed with an = , 0 < 1/4, harmonic
measure on K had dimension greater than 1 C for some constant C. This
result is due to Ancona.
See Batakas [1996] for the proofs of all these results.
2. Let  be a domain such that for some > 0
(w, B(w, 2dist(w, )), )

(E.1)

for all w . Then harmonic measure for  satisfies


h

(E.2)

for every measure function h(t) such that


h(t)
= 0.
t0 t
lim

The proof outlined below is a simple variation of the original Makarov argument given in Section VI.5.
(a) Let : D  be a universal covering map (see Ahlfors [1973], Fisher
[1983], or Gamelin [2001].) Let {z n } D be nontangentially dense on D
and let wn = (z n ). Then

B(wn , 2dist(wn , )) = 1.


Hint: Consider u when 1 u is the harmonic measure above and use
Exercises III.12 and VI.11.
(b) Let : D  be a universal cover. If w0 = (z 0 ) then
dist(w0 , ) |  (z 0 )|(1 |z 0 |2 ).
Hint: Since  has no zeros the monodromy theorem yields an inverse map
: B(w, dist(w, )) D such that (w) = w and (w0 )) = z 0 .
Now apply the Schwarz lemma. Also see Lemma X.5.5.
(c) Use parts (a) and (b) and the proof of Theorem VI.5.2 to show that (E.1)
implies (E.2). We do not know who first discovered this argument, which is
mentioned in the introduction of Jones and Wolff [1988].
3. Condition (E.1) has many equivalents and a compact set K C is called
uniformly perfect if  = C \ K satisfies any one of them. Exercise III.11
gave a capacity condition equivalent to (E.1). Thirteen more equivalent conditions are listed below.

CY535/Garnett

344

0 521 47018 8

January 27, 2005

15:41

IX. Infinitely Connected Domains


(i) There is a constant c > 0 such that if : D  is a universal cover
and if w0 = (z 0 ) then
dist(w0 , ) c|  (z 0 )|(1 |z 0 |2 ).
In other words, the inequality in Exercise 2(b) can be reversed. See
Pommerenke [1979].
(ii) Let : D  be a universal cover map. The expression
d(w) =

|  (z)||dz|
, (z) = w
1 |z|2

which is independent of the choices of z or , is the infinitesimal form


of the hyperbolic metric on . In other word, the hyperbolic distance
between z 1 and z 2 is

(z 1 , z 2 ) = inf d(w)

where the infimum is taken over all arcs in  that connect z 1 to z 2 .


Now K is uniformly perfect if and only if there is a constant c > 0
such that

|dw|
.
(z 1 , z 2 ) c inf

dist(w, )
(iii) There is a constant C > 0 such that every closed curve in  not homotopic to a point in  has hyperbolic length a least C. See Pommerenke
[1979].
(iv) There is a constant C such that if R  is a ring and each component
of C \ R meets K , then mod(R) C. See Beardon and Pommerenke
[1979].
(v) There is a constant C  such that if A  is an annulus and each
component of C \ A meets K , then mod(A) C. Use (iii) and
Exercise IV.5.
(vi) If : D  is a universal cover, then g = log(  ) is a Bloch function. Pommerenke [1979].
(vii) There is a constant c > 0 such that if : D  is a universal cover
and if (z 1 ) = (z 2 ) then (z 1 , z 2 ) c. Pommerenke [1979].
(viii) There is a constant > 0 such that if : D  is a universal cover
and if 1 (w) = {z j }, then
,  z j z k 
inf

 .
k
1 zk z j
j; j =k

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

345

Pommerenke [1984]. This the famous Carleson [1958] condition that


{z j } be an interpolating sequence for H . It also has many equivalents.
See Garnett [1981], for example.
(ix) The domain  has a strong barrier; i.e., there is a nonconstant positive
superharmonic function U (z) on  and a constant c > 0 such that in
the distribution sense
cU (z)
0.
U (z) +
dist(z, )2
See Ancona [1986].
(x) There is a constant C such that Hardys inequality holds: If u(z) C 1
has compact support contained in , then




|u(z)| 2
u(z)2 d xd y.
d xd y C
dist(z, )


Ancona [1986].
(xi) Let : D  be a universal cover map. Then K is uniformly perfect
if and only if there is a constant C such that whenever D  is a
disc, every component of 1 (D) is a C-quasidisc. See Heinonen and
Rohde [1995].
(xii) Let g(z, ) be Greens function for . There is > 0 and C <
such that for all 

g(z, )1+ < C.
z g(z, )=0

See Gonzlez [1992].


(xiii) Again let : D  be a universal covering map. A fundamental
domain for is a set F D such that (F) =  and is one-toone on F. Then K is uniformly perfect if and only if there exists a
fundamental domain F for such that F is a quasicircle. Gonzlez
[1992].
The equivalence of (i), (ii), (v), (vi), and (viii) is independently due to Behrens
[1976].
4. For 1 j, k N let B j,k be a closed disc with center ( Nj , Nk ) and radius

1
, and set K = B j,k and  = C \ K . Assume all discs B j,k have
r j,k < 2N
the same harmonic measure at ,
(, B j,k , ) =

1
.
N2

CY535/Garnett

0 521 47018 8

346

January 27, 2005

15:41

IX. Infinitely Connected Domains

Prove there is C independent of N so that



r j,k C.
Hint: Use Chapter III.
5. (a) Let be harmonic measure for some domain and define dim by (N.1).
Prove that if = dim, then f () = , where f is defined by (N.2).
(b) Using the estimate Bb (t) Ct 2 from Chapter VIII, prove that Conjecture
(N.3) implies
() c( 1)2
for 1 2.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

X
Rectifiability and Quadratic Expressions

First we introduce three topics, two old and one new.


(i) The first topic is the classical Lusin area function. The Lusin area function
gives another description of H p functions and another almost everywhere
necessary and sufficient condition for the existence of nontangential limits.
The area function is discussed in Section 1. In Appendix M we prove the
JerisonKenig theorem that the area function determines the H p class of
an analytic function on a chord-arc domain.
(ii) The second topic is the characterizations of subsets of rectifiable curves
in terms of certain square sums. These theorems, from Jones [1990], are
proved in Sections 2 and 3.
(iii) The third topic is the Schwarzian derivative, which measures how much
an analytic function deviates from a Mbius transformation. Section 4 is
a brief introduction to the Schwarzian derivative.
Then we turn to the chapters main goal, an exposition of the two papers
[1990] and [1994] by Bishop and Jones. In Section 5 the Schwarzian derivative
is estimated by the Jones square sums and by a second related quantity. In
Section 6 rectifiable quasicircles are characterized by a quadratic integral akin
to the Lusin function but featuring the Schwarzian derivative. In Sections 7, 8,
and 9 the same quadratic integral gives new criteria for the existence of angular
derivatives and further characterizations of BMO domains. Section 10 brings
together most of the ideas from the chapter to prove a local version of the F.
and M. Riesz theorem, and in Section 11 this F. and M. Riesz theorem leads us
to the most general form of the HaymanWu theorem.

347

CY535/Garnett

0 521 47018 8

348

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

1. The Lusin Area Function


Let F(z) be an analytic function on D. The area function of F is

1

 F  (z)2 d xd y 2 ,
AF( ) = A F( ) =
 ( )

where > 1 and



 ( ) = z D : |z | < (1 |z|) .

2
AF( ) is called the area function because AF( ) is the area of F( ( ))
when values are counted with their multiplicities. We will need a famous theorem of Caldern [1965].
Theorem 1.1. Let F(z) be analytic in D and let 0 < p < . Then F H p if
and only if AF L p (D), and there is a constant c = c( p, ) such that
c||AF|| p ||F F(0)|| H p c1 ||AF|| p .
p

(1.1)

A complete proof of Calderns theorem will be given in Appendix M. See


also Caldern [1965] or Stein [1970] and [1993]. In this section we prove
Calderns theorem for p = 2 on certain cone domains and we prove the local
theorem that modulo sets of measure zero, AF( ) < if and only if F has a
nontangential limit at .
On D the p = 2 case of Calderns theorem has a simple Fourier series proof.
By Fubinis theorem





2
AF( ) ds =
|F  (z)|2 { : z  ( )}d xd y,
D

and by the definition of 




1
(1 |z|2 ) { : z  ( )} c()(1 |z|2 ).
c()
Hence

AF( ) ds c()

|F  (z)|2 (1 |z|2 )d xd y


c2 ()

(1.2)

2
AF( ) ds.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Lusin Area Function


Now write F(z) =

349

an z n . Then
||F F(0)||2H 2 = 2

|an |2

n=1

and


  2
 F (z) (1 |z|2 )d xd y =

n=1

n
|an |2 .
n+1

Therefore (1.1) holds at p = 2 and



  2
1
 F (z) (1 |z|2 )d xd y 1 ||F F(0)||2 2 . (1.3)
||F F(0)||2H 2
H
4
2
D

The case p = 2 also has a simple geometric proof using (1.2) and Greens
theorem. See Exercise 8 of Chapter II.
Let U be a simply connected domain such that U is a rectifiable curve,
let F(z) be analytic on U , and let be a conformal map from D onto U . By
definition we say F(z) is in H p (U ), 0 < p < , if
1

F (z) (  (z)) p H p (D).


When F H p (U ) we take
p

||F|| H p (U ) =

 i
 p  i
 F (e )  | (e )|d.

By the F. and M. Riesz theorem,  has non-zero nontangential limit almost


everywhere on D, and it follows that every F(z) H p (U ) has nontangential
limit F( ) arc length almost everywhere on U and



 i
 p  i

p
 F (e )  | (e )|d =
 F( ) p ds.
||F|| H p (U ) =
D

p
||F|| H p (U )

Consequently the norm


does not depend on the choice of .
We define a special cone domain to be a cone domain having either the
form
%
 ( ),
(1.4)
U=
E

where E is a compact subset of D, or the form


%
U=
h ( ),
E

(1.5)

CY535/Garnett

350

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

where E is a compact subset of an arc I D such that |I | h < 1 and where


h ( ) =  ( ) {z : 1 |z| < h}
with > 2. See Figure X.1. When U is a special cone domain, we define the
central point of U in case (1.4) by cU = 0 and in case (1.5) by
3h
)c I
4
where c I is the center of the arc I. Every special cone domain U is connected
and U cU if U = because > 2 and h |I |.
cU = (1

E
h

cU

Figure X.1 The special cone domain U =


E

h ( ).

When U is a domain and z U, we write d(z) = dist(z, U ) and we define


the cone (relative to U ) at U to be
 (, U ) = {z U : |z | < d(z)}.

 (, U )


Figure X.2  (, U ) and (  , U ).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Lusin Area Function

351

If F(z) is analytic on U we also define the area function



1

 F  (z)2 d xd y 2 .
AF( ) = A F(, U ) =
 (,U )

Theorem 1.2. Let U be a special cone domain and let F(z) be analytic on U .
Then F H 2 (U ) if and only if
AF L 2 (U , ds)
and if and only if

  2
 F (z) d(z)d xd y < .

(1.6)

There exist constants C1 = C1 () > 0 and C2 > 0 such that




2
  2
 F (z) d(z)d xd y C2 ||F||2 2 . (1.7)
C1 ||F||2H 2 (U )  F(cU ) +
H (U )
U

Furthermore, F H 2 (U ) if and only if



  2
 F (z) d(z)3 d xd y <
U

and there is C3 = C3 () > 0 such that




  2
  2
1

2


 F (z) d(z)3 d xd y
F (z) d(z)d xd y |F (cU )| +
C3
U


C3

  2
 F (z) d(z)d xd y.

(1.8)

Of course when U = D (1.6) is nothing but the case p = 2 of Calderns theorem 1.1 above. Theorem 1.2 is a special case of a result from Kenigs thesis [1980], which generalized Calderns theorem for 0 < p < to the more
general setting of Lipschitz domains. Later Jerison and Kenig [1982a] extended
Calderns theorem to domains bounded by chord-arc curves. See Appendix
M.
The proof of Theorem 1.2 hinges on three geometric properties of special
cone domains:
(i) For all > 1 there are constants c1 = c1 (, , h) and c2 = c2 (, , h)
such that
c1 d(z) 1 ({ U : z  (, U )}) c2 d(z).

(1.9)

CY535/Garnett

352

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

d(z)

Figure X.3  =  (, U ) is shaded.


(ii) There exists = (, h) such that
B(cU , ) U.

(1.10)

(iii) Let : D U be a conformal mapping satisfying (0) = cU . Then



 

  (ei )ei  
i


d


 
arg (e ) cU 
arg
 = arg


(ei ) cU  
d
2
(1.11)

(, h) < .
2
Inequality (1.11) gives control almost everywhere on U on the angle between
the normal vector and the position vector emanating from cU .

cU

Figure X.4 The angle in (1.11).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Lusin Area Function

353

Proof of Theorem 1.2. Again Fubinis theorem gives




(AF( ))2 ds =
|F  (z)|2 1 { U : z  (, U )} d xd y,
U

and then (1.9) yields




2
AF( ) ds c(, U )
|F  (z)|2 d(z)d xd y
U

(1.12)

c (, U )
Therefore AF
of (1.2) for the special cone domain U.
To prove (1.7) define


|F( )|2 ds =
A=

2
AF( ) ds.

L 2 (U ) if and only if (1.6) holds. Notice (1.12) is the analogue



 F 2 |  |d,


 F  (z)2 |  (z)|(1 |z|2 )d xd y,

B=
D

and

D=
D


2
   (z) 

 F 2 
 (1 |z|2 )d xd y.
  (z) 

By the Koebe theorem,

B8

  2
 F (z) d(z)d xd y 8B.

We may assume F(cU ) = 0. By (1.3) we have



$

2
d
F((z))  (z)  (1 |z|2 )d xd y
A2

dz
D

2
1



(F ) (z) |  (z)|(1 |z|2 )d xd y


2
D

   (z) 2

 F 2 

 (1 |z|2 )d xd y
 (z)
D

=B

D
.
2

CY535/Garnett

0 521 47018 8

354

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Write g = log(  ). By (1.11), |Img| = | arg(  )| < 2, so that by Exercise 3


of Appendix F, g BMO and
  2
  (1 |z|)d xd y = |g  (z)|2 (1 |z|)d xd y

is a Carleson measure with norm bounded by a fixed constant C. Therefore



(F )(z)2 |  (z)||g  (z)|2 (1 |z|2 )d xd y C A
D=
D

and
B (1 + 2C)A
with C independent of and h.
The proof that A C()B will exploit the inequality (, h) <
Because (, h) < 2 and |(ei )| 1, we have



2   (ei )ei 

 F  
A
d
 (ei ) cU 
D



2 ei  (ei )  




sec 
F
d .
(ei ) c
D

in (1.11).

Hence by Greens theorem (see Exercise 8 of Chapter II),


 



 (z) 
1
z


A sec 
log d xd y .
|F |2


(z) cU
|z|
D

Write h(z) =

z  (z)

. Then

1
A 4 sec
|(F ) |2 |h| log d xd y
|z|
D


1


+ 4 sec
(F )(F ) h   log d xd y,
|z|


and because 

(z)cU

z
(z)cU


 c() by the maximum principle,

A 4c()(sec )B + 4 sec



1


(F )(F ) h   log d xd y,
|z|
D

again by Exercise II.8. Now by (1.11), h(z) = e V (z) where |ImV | <

2,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Lusin Area Function

355

and h  = V  h. Therefore by CauchySchwarz




1


(F )(F ) h   log d xd y
|z|
D



1



=
(F ) (F )V  |h| log d xd y
|z|
D


!1
2
2
1



(F )  |h| log d xd y


|z|
D


!1
2
2
1


 2

F  |V | |h| log d xd y
|z|
D

c()B

1
2


!1
2
2  2
1

(F ) h  |V | log d xd y .
|z|
D

But since |ImV |


|V  |2 log

1
d xd y
|z|

is a Carleson measure with constant bounded by c(), and we conclude that


1

A C()(B + B 2 A 2 ),
so that (1.7) holds.
To prove (1.8) note that because |F  (z)|2 is subharmonic we have

 

  2
 F (cU )2 c1 (, h)
 F (z) d(z)d xd y,
U

by property (ii) and


  2
 F (z) 2
4
for z U and < d(z). Hence
  2
 F (z) d(z)3

Now if |w z| <

d(z)
2


 
 F (w)2 dudv

B(z,)

32
d(z)


 
 F (w)2 dudv.

B(z, d(z)
2 )

then
3
d(z)
d(w) d(z)
2
2

CY535/Garnett

0 521 47018 8

356

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

so that by Fubinis theorem,






  2
 
3
 F (z) d(z)3 d xd y 32
 F (w)2
d xd ydudv

2d(w)
U
U
B(w,d(w))

2
 
 F (w) d(w)dudv,
48
U

which establishes the right-hand inequality in (1.8).


To prove the other inequality in (1.8) we again use property (ii) to get

  2
 F (z) d(z)3 d xd y
|F  (z) F  (cU )|2 d(z) C()
U

B(cU , 2 ).

for z

Therefore


|F (z) F (cU )| d(z)d xd y c ()


2

B(cU , 2 )

|F  (z)|2 d(z)3 d xd y.

zcU
When z U \ B(cU , 2 ) write z 0 = cU + 2 |zc
. Then
U|

z

 2
 
 F  (w)ds
 F (z) F  (z 0 )2
z
z0
z
2
 
3
3


2
F (w) d(w) ds

d(w) 2 ds
z0
z0
z


1
 F  (w)2 d(w) 23 ds
C  ()d(z) 2
z0

because by trigonometry d(w) d(z) + c2 ()|z w| if w [z 0 , z].

z
w
z0
cU

Figure X.5

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Lusin Area Function

357

w
Write w for the projection of w U to U along the radius [0, |w|
]. Then
again by trigonometry

d(z) c3 ()|z w |
when z [w, w ]. Using polar coordinates we then obtain

|F  (z) F  (cU )|2 d(z)d xd y
U


2

 
C()
F (w) d(w)3 d xd y
U

+ C  ()


 
 F (w)2 d(w) 23

U {|zcU |> 2 }



C ()

w
w

|z w | 2 dr dudv


 
 F (w)2 d(w)3 dudv,

which proves the other half of (1.8).

We say two statements P and Q about points on D are almost everywhere


equivalent if the symmetric difference
{P true} {Q true}
has measure zero. For example, if u(z) is a harmonic function on D and if > 1,
then by Privalovs theorem, u ( ) < is almost everywhere equivalent to
u has nontangential limit at . The next theorem is the parallel result for area
functions.
Theorem 1.3. Let F(z) be an analytic function on D and let > 1. Then the
statements
AF( ) <
and
F has nontangential limit at
are almost everywhere equivalent.
A simple corollary of Theorem 1.3 and Plessners theorem is that Theorem 1.3 is also true when F(z) is a harmonic function. The martingale analogue of Theorem 1.3 is the Lvy theorem J.3 in Appendix J. One half of
Theorem 1.3 is due to Marcinkiewicz and Zygmund [1938] and the other half
to Spencer [1943].

CY535/Garnett

358

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

The proof of Theorem 1.3 depends on a fourth geometric property of the


special cone domains having the form (1.4) and of their cones  (, U ) : If
E = U D and if z U , then d(z) 1 |z| and hence
 (, U )  ( ).
Conversely, suppose z  ( ) U and z  ( ) satisfy |z z | d(z),
with < . Then
1 |z| d(z) + 1 |z |
and
|z | |z | + d(z).
Hence
|z |
|z |
|z | + d(z)
d(z) +
d(z) +
,

so that
 ( )  (, U ) U

(1.13)

when
=

1+
1

(1.14)

 ( )
cU = 0
 (, U )

Figure X.6 Cones satisfying (1.13).


Proof of Theorem 1.3. First we prove the MarcinkiewiczZygmund half. Let
K D be compact and assume F has nontangential limit at every K . Fix

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

1. The Lusin Area Function

359

> 1. Take > and define by (1.14). Taking K smaller, we can assume
F is bounded on the cone domain
%
U=
 ( ).
K

Thus F H 2 (U ) and by Theorem 1.2


A F(, U ) <
almost everywhere on K . Therefore by (1.13)
A F( ) <
almost everywhere on K .
To prove Spencers half of the theorem, we can assume that for all ei E,

|F  (z)|2 d xd y N
 (ei )

for some constant N . Let K E be a compact set such that for fixed ,
|E ( h, + h)|
3/4
(1.15)
2h
whenever 0 < h < . By the theorem on points of density, it is enough to prove
F has nontangential limit almost everywhere on K . Set
%
 (ei ).
U=
ei K

Then by (1.15) there is c() > 0 such that



2 N
|F  (z)|2 d xd yd
K  (ei )


=
U



|F  (z)|2  K { : z  ( )}d xd y

c()

|F  (z)|2 (1 |z|2 )d xd y.

c ()

|F  (z)|2 dist(z, U )d xd y.

Therefore F
by Theorem 1.2, and F has nontangential limits almost

everywhere on K by the F. and M. Riesz theorem.
H 2 (U )

CY535/Garnett

0 521 47018 8

360

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Theorem 1.4. Let F(z) be analytic in D and let > 1. Then



  2
 F (z) (1 |z|2 )2 d xd y <

(1.16)

 ( )

is almost everywhere equivalent to


F has nontangential limit at .
Proof. We outline two proofs. The first proof uses Theorem 1.2. Suppose

  2
 F (z) (1 |z|2 )2 d xd y M
 ( )

for all E. Given > 0 and c < 1, there exist a compact set E 0 E and a
0 > 0 such that |E \ E 0 | < and
|B(, ) E| c
for all E 0 and all < 0 . Set
U=

(1.17)

 ( ).

E0

By (1.17),





 E {| z| < (1 |z|)} c (1 |z|2 )

when z U and 1 |z| . Therefore




  2
  2
 F (z) dist(z, U )3 d xd y
 F (z) (1 |z|2 )3 d xd y
U

  2
 F (z) (1 |z|2 )2 d xd y

E  ( )

C M.
Then by Theorem 1.2, F H 2 (U ) and F has nontangential limits almost everywhere on E 0 .
Conversely, assume F has nontangential limit at every E. Fix > .
Given > 0 there is a compact set E 0 E such that |E \ E 0 | < and F is
bounded on
%
 ( ).
U=
E0

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Squares Sums and Rectifiability

361

Thus F H 2 (U ) and by Theorem 1.2



|F  (z)|2 d(z)3 d xd y < .
U

Let
U0 =

 ( ).

E0

If z U0 , then 1 |z|2 c(, )d(z). Consequently by Fubinis theorem



|F  (z)|2 (1 |z|2 )2 d xd y
E 0  ( )



|F  (z)|2 (1 |z|2 )2 { E 0 : z  ( )}d xd y

U0

|F  (z)|2 d(z)3 d xd y,

and (1.16) holds almost everywhere on E 0 .

The second proof combines Theorem 1.3 with the pointwise result below.
Lemma 1.5. Let > 1 and let F(z) be analytic on  ( ).
(a) If 1 < < there is C(, ) such that for all D,


  2
  2
 F (z) (1 |z|2 )2 d xd y C(, )
 F (z) d xd y.
 ( )

 ( )

(b) There is C such that for all D,





  2

  2
 F (z) d xd y C  F  (0)2 + C
 F (z) (1 |z|2 )2 d xd y.
 ( )

 ( )

The proof of Lemma 1.5, which is very like the proof of (1.8), is an exercise
for the reader.

2. Squares Sums and Rectifiability


Let E be a bounded set. We seek geometric conditions that are necessary and
sufficient for E to lie on a rectifiable curve. Clearly one necessary condition
is 1 (E) < , and if E is connected then 1 (E) < if and only if E is a

CY535/Garnett

0 521 47018 8

362

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

subset of a rectifiable curve. In fact, a connected set E is the trace of a (selfintersecting) rectifiable closed curve of length  if and only if 1 (E) /4;
see Falconer [1985] or David and Semmes [1991]. However, there exist sets
E such that 1 (E) = 0 but E is not contained in any rectifiable curve; see
Exercise 3.
For z E and t > 0, define

!
1
E (z, t) = inf
sup dist(w, L) : L is a line through z .
t
EB(z,t)
Thus E (z, t) is a scale invariant measure of the least deviation E B(z, t)
has from lines through the point z. Because z E, the width of the narrowest
strip containing E B(z, t) is comparable to t E (z, t).
t
2t E (z, t)

Figure X.7
Theorem 2.1. There is a constant C such that if E is a bounded set and
diam(E)
dt
(2.1)
E2 (z, t) K
t
0
for all z E, then there is a rectifiable curve  such that E  and
() CeC K diam(E).
Theorem 2.5 will provide a converse of Theorem 2.1: If E  and  is a
rectifiable curve, then (2.1) holds at 1 almost every point of E. See Exercise 7
for a refinement of Theorem 2.1.
To simplify the proof of Theorem 2.1 we introduce a dyadic variant of
E (z, t). When Q is a dyadic square of side (Q), write a Q for the concentric
square with side (a Q) = a(Q) and define

!
1
inf sup dist(z, L) : L is any line .
(2.2)
E (a Q) =
(a Q)
Ea Q

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Squares Sums and Rectifiability

363

Then 2 E (a Q)(a Q) is the width of the narrowest strip containing E a Q.


Lemma 2.2. There are constants c1 , c2 , and c3 such that if z E, then for all
d > 0,
d


dt
E2 (3Q) : (Q) d, z Q
E2 (z, t)
c3
t
0
(2.3)
d
dt
2
E (z, t) .
c1 + c2
t
0
Proof. By a change of scale we may assume d = 1. Note that if Q is a dyadic
square containing z, then

3Q B(z, 2 2(Q))
and

2 2
E (3Q)
E (z, 2 2(Q)).
3
Furthermore, if t s then
s
E (z, t) E (z, s).
t
Write Q n = Q n (z) for the unique dyadic square
#
"
Q n = j2n x < ( j + 1)2n , k2n y < (k + 1)2n

(2.4)

(2.5)

such that z Q n . Then by (2.4) and (2.5)


n=0

E2 (3Q n )


8
n=0

E2 (z, 2n+ 2 )

5
n+
32  2 2 2
dt

E (z, t)
n+ 23
9 log 2
t
n=0 2
+
)
1
5
dt
32
2
E (z, t) + log 2 .

9 log 2
t
2
0

To prove the lower estimate, note that when we reflect a strip S containing
E 3Q about a line through z E Q parallel to the sides of S we obtain a
strip having centerline through z and not more than twice as wide as S. Since
B(z, (Q)) 3Q, we therefore have
1
E (z, (Q)),
6
and it follows from (2.5) and (2.6) that
E (3Q)

(2.6)

CY535/Garnett

364

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

1  2
E (z, 2n )
36
n=0
2n

1
dt

E2 (z, t)
n1
144 log 2
t
n=0 2
1
dt
1
E2 (z, t) .

144 log 2 0
t

E2 (3Q n )

n=0

Proof of Theorem 2.1. In the proof we assume diam(E) 1. Then by (2.1)


and Lemma 2.2

E2 (3Q) c1 + c2 K
(2.7)
(Q)1,zQ

for all z E. We first describe the basic construction underlying the proof. Let
S be a closed rectangle with side lengths L and L, where 1, and assume
E meets each side of S. Partition each L side of S into three segments of length
L/3, and define the middle third M(S) as that subrectangle with each L side
replaced by its middle third. There are two cases.
Case 1: There exists a point p E M(S). See Figure X.8. Cut S into two
rectangles A0 and A1 by the line through p orthogonal to the L sides of S,
and let j L be the length of the projection of A j onto an L side of S. Then
1/3 j 2/3. Cover A j E by a rectangle Sj such that A j E meets each
side of Sj and such that the shortest side of Sj is as small as possible. Then
S0 S1 E = since p S0 S1 . Let L j be the length of the longest side of
Sj . By the Pythagorean theorem,
L j ( 2j + 2 )1/2 L (1 + 5 2 ) j L .
A1

M(S)

E
L

S
S1

A0

(2.8)

p
S0
L
Figure X.8 Case 1.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Squares Sums and Rectifiability

365

Case 2: E M(S) = . Let A0 and A1 be the two components of S \ M(S).


See Figure X.9. Cover A j E by a rectangle Sj such that each side of Sj meets
A j E and Sj has shortest side as small as possible. Join S0 to S1 by the shortest
segment T having endpoints in E Sj . Write 0 = 1 = 1/2, write L j for the
length of the longest side of Sj , and write |T | for the length of T . Then (2.8)
still holds and
|T | (1 + 2 )L .

(2.9)

A1
M(S)
S1

A0
T

L1
L

S0

Figure X.9 Case 2.


In both cases
0 + 1 = 1,

(2.10)

and there is a constant C such that Sj has shortest side of length j L j and
j C E (3Q)

(2.11)

for any dyadic Q such that Q Sj = and diam(Sj ) (Q) < 2 diam(Sj ),
because in that case Sj 3Q.
To start, assume E Q 0 = [0, 1] [0, 1], but assume no smaller dyadic
Q contains E. Let S0 be a rectangle such that E S0 , E meets all four sides
of S0 , and S0 has smallest shortest side. Write its side lengths as L 0 and L 0 ,
where 1 and C(3Q 0 ).
Now perform the basic construction on S0 , getting two rectangles S0,0 and
S0,1 , and (in Case 2) a line segment T0 . Continue. After n steps we have
rectangles S I , I = (i 1 , i 2 , . . . , i n ), i j = 0, 1, with longsides L I , and we have
weights I [1/3, 2/3] and factors I 1. For these rectangles, conditions
(2.8), (2.10), and (2.11) hold. If Case 2 is applied to S I we also get a segment
TI of length |TI |, for which (2.9) holds.
After N = 25 steps the diameter of a rectangle S I drops by at least 1/2.
Indeed, if we begin with a rectangle that has side lengths 1 and 1 and

CY535/Garnett

0 521 47018 8

366

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

diameter d0 = (1 + 2 )1/2 , then in the worst case the construction yields a


rectangle having sides and ( 2 + 2 )1/2 and diameter d1 = (2 2 + 2 )1/2 .
Since 2/3, we have d1 (17/18)1/2 d0 , and since (17/18)25/2 1/2, the
claim holds. Consequently each dyadic square Q can occur at most N times
during iterated applications of the basic construction.
For a multi-index I = (i 1 , i 2 , . . . , i n ) of size |I | = n and for m < n, write
Im = (i 1 , i 2 , ...., i m ). Let {I (1) , . . . , I ( p) } be a set of multi-indices having size
|I ( j) | n such that I ( j) I (k) for j = k. Then by (2.8) and (2.10),

 ,

,
(1 + 5 I2m )
Im .
L I ( j) L 0
|I |=n 1mn

1mn

By (2.11) and the hypothesis (2.7),


,
(1 + 5 I2m ) Ce5c2 N K ,
1mn

and by (2.10)

 ,

Im = 1.

|I |=n 1mn

Therefore

L I ( j) CeC K

(2.12)

for some new constant C, and in particular



L I CeC K .


|I |=n

Next we estimate |TI |. Write



{|TI | : Case 2 has been used k times on I1 , I2 , . . . , In = I }.
Mk =

To estimate Mk note that there is a constant 0 < 1 such that if we use Case 2
and L 0 + L 1 0.9L , then 0 and L 0 + L 1 C L . Furthermore, by (2.7)
the alternative L 0 + L 1 0.9L can only hold for at most N  subindices Im .
Therefore by (2.9) and (2.12),


Mk (1 + 02 )C N (0.9)kN CeC K
and

Mk C  e C K .

(2.13)

At the nth stage of the construction we have a connected union n of rectangles and segments such that E n . By (2.12) and (2.13) there is a contin-

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Squares Sums and Rectifiability

367

uous map n : [0, 1] n such that n ([0, 1]) has bounded length, such that
n ([0, 1]) meets all four sides of each rectangle S I , |I | = n, and such that n+1
is built by modifying n on each n1 (S I ), with |I | = n. Then lim n defines a

rectifiable curve containing E.
An earlier characterization of subsets of rectifiable curves was also given by
Jones in [1990]. For any set E define

{ E2 (3Q)(Q) : Q dyadic, (Q) diam(E)}.
2 (E) = diam(E) +
Theorem 2.3. Suppose E is a bounded set. Then E is contained in a rectifiable
curve if and only if 2 (E) is finite. Moreover, there are constants c1 , c2 so that
c1 2 (E) inf 1 () c2 2 (E).
E

(2.14)

Theorem 2.3 is sometimes called the traveling salesman theorem because


it bounds the total distance the salesman must travel to visit a finite set E of
cities when each pair of cities in E is connected by a straight road. Exercise 6
describes the minimal spanning tree problem and the simple greedy algorithm
which gives a shortest spanning tree. However the greedy algorithm gives no a
priori bound like (2.14) for arbitrarily large sets. See Garey and Johnson [1979]
and Lawler, Lenstra, Rinnooy Kan and Shmoys [1990] for more information
about spanning trees and the classical traveling salesman problem.
We first prove the right-hand inequality of (2.14), that there exists a curve
 with 1 () c2 2 (E). The argument is like the proof of Theorem 2.1. Let
Rn be the sum of the diameters of the rectangles at stage n. From (2.8) we have
L 0 + L 1 L + C E2 (3Q)(Q)

(2.15)

for any Q such that Q E S = and diam(S) (Q) 2 diam(S). As


noted before, each such Q can occur at most N = 25 times during iterations of
the construction, so that
Rn C 2 (E).
To estimate the lengths of the segments created by applications of Case 2,
note again that if Case 2 is used and if L 0 + L 1 0.9L, then 0 and
|T | C E2 (3Q)(Q). Thus the sum of the lengths of the middle segments
created from applications of Case 2 with 0 is at most C 2 (E).
Now write
Rn = In + I In
where In is the sum of the lengths of the rectangles at stage n to which Case 1
will be applied or for which 0 , and where I In is the sum of the lengths of

CY535/Garnett

368

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

the rectangles to which Case 2 will be applied and for which < 0 . Let Tn+1
denote the sum of the lengths of the middle segments created in Case 2 with
< 0 . Then

2 (3Q)(Q) + 0.9I In ,
Rn+1 In + C
where the sum is taken over all Q which intersect a rectangle R, created at stage
n and such that diam(R) (Q) < 2 diam(R). Moreover by (2.9),

2 (3Q)(Q),
(0.1)Tn+1 (0.1)I In + C
and so
Rn+1 + (0.1)Tn+1 Rn + C

Thus
(0.1)Tn+1 (Rn Rn+1 ) + C

2 (3Q)(Q).

2 (3Q)(Q),

and because Rn is bounded,




Tn C diam(E) +
2 (3Q)(Q).
n

Q dyadic

The left-hand inequality in (2.14) will be proved in the next section. Here
we only treat the special case of a Lipschitz graph.
Lemma 2.4. Let  be the graph of a Lipschitz function. Then (2.14) holds for
all E .
Proof. We use an elegant argument from Jones [1990]. Assume that, after a
translation and a dilation,


 = 0 x 1, y = f (x) ,
where f (0) = f (1) and


 f (x2 ) f (x1 ) M|x2 x1 |.
It is enough to show the left side of (2.14) holds for E = . Set
5 j ( j + 1) 6
,
I jn = n ,
2
2n
and
 nj = f (I jn ),
and let Jjn be the segment
5
j
( j + 1) 6
) .
Jjn = f ( n ), f (
2
2n

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Squares Sums and Rectifiability

369


J2n+1
j+1

J2n+1
j
Jjn
I jn

Figure X.10
n+1
n
Then J2n+1
j , J2 j+1 , and Jj are the three sides of a triangle and
#
"
n+1
n, j = 2n sup dist(z, Jjn ) : z J2n+1
j J2 j+1

satisfies
n+1
n
n 2
(J2n+1
j ) + (J2 j+1 ) (Jj ) c2 n, j ,

with c = c(M) > 0, again by the Pythagorean theorem. Therefore



2
c2m m,k
2().

(2.16)

m,k

Now set

 



n, j = 2n sup dist(z, Jjn ) : z  nj
2nm sup m,k : Ikm I jn .
m=n

Write p = m n and



sup m,k : Ikm I jn = n+ p,k (n, j).

Then (2.16) and Minkowskis inequality give


"

2
2n n,
j

#1
2

n, j

"

n, j

2 p n+ p,k (n, j)

p=0

"


2
2n 22 p n+
p,k (n, j)

p=0 n, j


p=0

2 # 21

2 2

"

2
2m m,k

#1
2

(2.17)

#1
2

m,k

1
2c () 2 .
Extend f to [1, 2] by defining f (x 1) = f (x), and let (t) be the translate {y = f (x + t)}. We get numbers n, j (t) that also satisfy (2.17). But then

CY535/Garnett

0 521 47018 8

370

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

for fixed n,



 (3Q) C
2

(Q)=2n2

n, j (t)2 dt,

because with probability 1/2 there exists j such that {Rez t : z Q} In, j .
That gives (2.14) for .

We make two observations about the construction. First, if E is a finite set,
then the curve constructed consists of line segments joining the points of E,
and these segments are the TI from Case 2. Second, if E is connected, then by
(2.14)
C1 2 (E) 1 (E) C2 2 (E).
If E is connected it is not necessary to construct the middle segments TI in
Case 2. Without the TI , every stage of the construction yields a union of rectangles, each meeting E, and this union is connected because E is connected.
Because the rectangles are shrinking, the constructed set is exactly E, although
the parameterized curve may trace portions of E several times.
The next result, from Bishop and Jones [1994], is a converse to Theorem
2.1.
Theorem 2.5. Let  be a Jordan curve. Then, except for sets of zero 1 measure,
z is a tangent point of  if and only if
1
dt
2 (z, t) < .
(2.18)
t
0
Proof. Write Tn() for the set of tangent points of . In Chapter VI we proved
that
%
Kn ,
Tn() =
n

where, after an affine transformation,




K n = 0 x 1; y = f (x) = g(x) ,
where f (x) and g(x) are Lipschitz functions, and f (x) g(x), 0 < x < 1.
Moreover, there is a function h(x) such that
f (x) h(x) g(x), on 0 < x < 1,
and there is a neighborhood Vn of K n such that


Vn  = y = h(x) : 0 < x < 1 .

(2.19)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

2. Squares Sums and Rectifiability

By Lemma 2.4,


2f (3Q)(Q) < and
g2 (3Q)(Q) < .

371

(2.20)

But whenever K n Q = ,

2 (3Q) C 2f (3Q) + g2 (3Q) .

(2.21)

Indeed, we may assume that g (3Q) is small and that f (3Q) g (3Q) by
symmetry. Let S be a strip of width g (3Q)(Q) containing 3Q {y = g(x)}
and let T be a parallel strip having width Mg (3Q)(Q) and having the same
top edge as S. When M is large, 3Q {y = f (x)} T (for otherwise by (2.19)
f (3Q) > g (3Q)) and consequently 3Q {y = h(x)} T and (2.21) holds
with C = C(M).
Now by (2.20) and (2.21)

2
V
(3Q)(Q) < .
n
(Q)<1

Then (2.3) and integrating the left side of (2.18) against arc length on K n yields
inequality (2.18) at almost every point of K n .
Conversely, let E = {z  : (2.18) holds}. Then by Theorem 2.1 and the
construction in Chapter VI,
%
En ,
E = E0
n

where 1 (E 0 ) = 0 and where, after an affine transformation,


E n {(x, f n (x)) : 0 < x < 1},
and f n is a Lipschitz function. We may suppose that z = (0, f n (0)), that f n is
differentiable at x = 0, that f n (0) = 0, and that (0, f n (0)) is a point of density
of E n . See Figure X.11. For fixed and for r small,




{x (r, r ) : | f n (x) | < |x|, (x, f n (x)) }
(1 )
2r
and
r
dt
2 (0, t) < .
t
0
But then

ir r

ir r

,
B
,
=
 B
2 4
2 4

because otherwise  (z, t) > 14 for all t ( r2 , r ). Therefore z is a cone point



for both components of C \ , and almost all such points lie in Tn().

CY535/Garnett

0 521 47018 8

372

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

ir/2

y = x

Figure X.11
Theorem 2.5 is an improvement of Corollary VI.6.4. Let 1 and 2 be the
components of C \ , let t j (t) be the length of an arc of  j {z : |z w| = t}
that separates w from a point z j  j and let
(w, t) = max{| j (t)| : j = 1, 2}.
Then clearly
(w, t) C(w, t).
Therefore Theorem 2.5 implies that
1
dt
2 (z, t) <
t
0

(2.22)

at 1 almost every tangent point of , which is the assertion of Corollary VI.6.4.


To our knowledge, it has not been determined if  has a tangent at almost every
point where (2.22) holds.

3. A Decomposition Theorem
To complete the proof of the left-hand inequality in (2.14) we use the following
decomposition theorem, also from Jones [1990]. We call a simply connected
domain  an M-Lipschitz domain if, after a translation and a dilation, 0 
and
 = {r ( )ei : 0 2 },
where
|r (1 ) r (2 )| M|1 2 |

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. A Decomposition Theorem

373

and
1
r ( ) 1.
M +1
Theorem 3.1. There is a constant M such that whenever  is a simply connected domain with 1 () < there exists a rectifiable curve  such that
\ =

j ,

(3.1)

j=0

such that each  j is an M-Lipschitz domain, and such that



1 ( j ) M1 ().

(3.2)

Proof. The proof is in two steps. The first step is to use a corona construction as
in Garnett [1981] or David and Semmes [1991] to obtain (3.1) and (3.2) with the
Lipschitz domains  j replaced by domains U j bounded by uniformly chord-arc
curves. The second step is to further decompose each chord-arc domain U j into
M-Lipschitz domains.

Let be a conformal map from D onto . Write F =  and g = log(  ).


The proof will use the familiar inequality
||g||B 6

(3.3)

many times. By the F. and M. Riesz theorem, F H 2 (see Exercise VI.1) and
by (1.3) or Exercise II.8,


1
1


2
| (z)||g (z)| log d xd y = 4
|F  (z)|2 log d xd y
|z|
|z|
D
D
(3.4)
2||F  ||2H 2
21 ().
Set D0 = {|z| < 21 } and U0 = (D0 ). Then by Theorem I.4.5, there is a
constant M independent of  such that U0 is an M-Lipschitz domain. Since
 H 1 , we also have
1 (U0 ) 1 ().

(3.5)

Next form the dyadic Carleson boxes


Q = {r ei : 1 2n r < 1, j2(n+1) < ( j + 1)2(n+1) },

CY535/Garnett

0 521 47018 8

374

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

0 j < 2n+1 , of sidelength (Q) = 2n and their top halves


T (Q) = Q {1 2n r < 1 2(n+1) }
%
= Q \ {Q  : Q  Q, Q  = Q}.
Write z Q for the center of T (Q).

zQ

Figure X.12
Fix > 0 to be determined later and consider a Carleson box Q. If
sup |g(z) g(z Q )| ,

T (Q)

(3.6)

we say Q is a type 0 box and define D(Q) = T (Q). If Q is of type 0, then again
by (3.3) U Q = (D Q ) is a chord-arc curve with chord-arc constant bounded
by a fixed constant C0 .
If (3.6) fails for Q, define G(Q) to be the set of maximal Carleson boxes
Q  Q for which
sup |g(z) g(z Q )| ,

T (Q  )

and define
D(Q) = Q \

(3.7)

Q.

G(Q)

Then D(Q) is a chord-arc curve with chord-arc constant at most 4 and


sup |g(z) g(z Q )| .

D(Q)

(3.8)

It follows that the domain U Q = (D(Q)) is bounded by a chord-arc curve


with constant at most 5 if is small.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. A Decomposition Theorem

375

Q
Q
D(Q)

Q

Figure X.13
Write {D j , j 1} for the set of maximal (with respect to set inclusion)
regions D(Q). Then the family {D0 } {D j , j 1} is pairwise disjoint. Write
U j = (D j ), j = 0, 1, . . . . To prove (3.2) we first show that

1 (U j ) M1 ().
(3.9)
j0

If D j = D(Q) for Q of type 0, then by (3.3)



1 (U j ) =
|  (z)|ds
T (Q)

c(Q)|  (z Q )|

1
c
|  (z)||g  (z)|2 log d xd y.
|z|
T (Q)

Since 1 almost every point lies in at most 2 sets U j , (3.4) then gives

1 (U j ) 21 ().
(3.10)
type 0

If Q is not of type 0, we say Q has type 1 if


1 (D D Q )

(Q)
,
2

and we say Q has type 2 if (3.11) fails.


When D j = D(Q) for Q of type 1, (3.8), (3.3), and (3.11) give


1 (U j ) =
|  (z)|ds C
|  (z)|ds.
D j

bDD(Q)

(3.11)

CY535/Garnett

0 521 47018 8

376

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Therefore


1 (U j ) C

type 1

|  (z)|ds 21 ().

(3.12)

Now assume D j = D(Q) where Q is a type 2 box. Then D j is a chord-arc


domain and D j is an Ahlfors regular curve, with constant independent of j.
Let {Jk } denote the top edges of the boxes in G(Q). Then

1 (D(Q))
1 (Jk )
12
because Q is of type 2. By (3.3), (3.7), and equicontinuity there are > 0 and
E k Jk such that |g(z) g(z Q )| on E k and 1 (E k ) 1 (Jk ). Consequently

|F(z) F(z Q )|2
ds C()(Q),
|F(z Q )|2
D(Q)
and


1 (U j ) =

D(Q)

|F(z)|2 ds C  ()


D(Q)

|F(z) F(z Q )|2 ds.

We would like to estimate the latter integral via (1.7), but unfortunately
D(Q) is not a Lipschitz domain. However D(Q) is a chord-arc domain, and
hence Theorem M.1 of Appendix M gives1

|F(z) F(z Q )|2 ds
D(Q)

(3.13)

C
|F  (z)|2 1 B z, 2dist(z, D(Q)) d xd y
D(Q)

Now because D(Q) is Ahlfors regular, (3.13) gives




|F(z) F(z Q )|2 ds C
|F  (z)|2 dist(z, D(Q))d xd y
D(Q)

D(Q)

|F  (z)|2 log

C
D(Q)

Therefore

type2


1 (U j ) C
D

|F  (z)|2 log

1
d xd y.
|z|

1
d xd y C1 (b).
|z|

1 For a proof of (3.13) using Theorem 1.2. only, see Exercise 9.

(3.14)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. A Decomposition Theorem

377

Together (3.5), (3.10), (3.12), and (3.14) yield (3.9).


We have already showed that U0 = (D0 ) is an M-Lipschitz domain. If
U j = (T (Q)) for a type 0 box Q, partition T (Q) into eight polar rectangles
Sk by dividing the top edge of T (Q) into four equal intervals and halving its
radial side. Then by (3.3) (Sk ) is an M-Lipschitz domain and

1 ((Sj )) C1 (U j ).
j

Finally, suppose U j = (D(Q)) for a box Q not of type 0. The construction


is easier to explain in the upper half-plane H and we assume Q is a square in
H with base and interval I R. See Figure X.14. Whenever J = Jk is the top
edge of a box Q k G(Q) we construct a tree T = T (Q k ). Let S 1 be a segment
of slope 1 joining the left endpoint of J to R and let S 2 be a segment of slope 1
joining the right endpoint of J to R, and define
T 1 = J S1 S2.
Assume by induction that T n T n1 has been defined and that
T n {Imz = 4n (J )}
1 T of slope 1. Let S 2
consists of 2n points z n, j situated on segments Sn,
n
j
n, j
n
1 that joins z
be the segment of length 2 4 (J ) orthogonal to Sn,
to
R
n, j
 2
 jn
n+1
n
= T j Sn, j . Then T = T (Q k ) = n T satisfies
and set T


1 (T ) = 1 + 2 2 +
2n 4n (Q k ).
n=1

Q
Jk
Qk
I
Figure X.14

(3.15)

CY535/Garnett

378

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Write
Dj \

T (Q k ) =

U j,k .

G(Q)

Then by (3.8) each (U j,k ) is an M-Lipschitz domain, and by (3.15) and (3.8),

1 ((U j,k )) C|  (z Q )|(Q),
k

which establishes (3.1) and (3.2).

Corollary 3.2. There exits a constant M < such that if  is a connected


  such
plane set with 1 () < , then there exists a connected plane set 
 ) M1 (), the bounded components D j of C \ 
 are M-Lipschitz
that 1 (

, and the boundary of the unbounded component of
domains with  D j
C\
 is a circle at least 3 21 () units from .
For the proof apply the theorem to each
bounded component of the union of
, a line segment and a circle at least 3 21 () units from .
Now let  be connected with 1 () < , let {D j } be the Lipschitz domains
given by Corollary 3.2 and write  j = D j and j = diamD j . Let Q be any
dyadic square and define
F(Q) = { j :  j 3Q = , j (Q)},
and
G(Q) = { j :  j 3Q = , j < (Q)}.
Then we have the following lemma.
Lemma 3.3. There is a constant C1 such that if (Q) diam(),

1 
2 (3Q) C1
2 j (3Q) + C1 2
Area(D j ).
 (Q)
F(Q)

(3.16)

G(Q)

Proof. We may assume (Q) = 1 and  (3Q) > 0, so that 3Q  j = for




2
some  j and 3Q D j . If F(Q) = then
G(Q) Area(D j ) 9 (Q).
Thus we can assume there exists 1 F(Q). Let L be a line such that
d = sup dist(z, L) 1 (3Q)(3Q),
1 3Q

and let z 0  3Q have maximal distance d0 = dist(z 0 , 1 ). Let z 1 1 have


1
minimal distance to z 0 and let z 2 = z 0 +z
2 . There are three cases, as shown in
Figure X.15.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

3. A Decomposition Theorem

1

1

1

z1

j

2

d1

z0

379

z3

Dj
Case 1

Case 2

Case 3

Figure X.15
Case 1: F(Q) = {1 }.
Then
B = B(z 2 ,

%
d0
)
Dj
2
G(Q)

and hence
2 (Q) (d + d0 )2 82 1 (2Q) + 2d02 82 1 (2Q) +

8 
Area(D j ).

G(Q)

Case 2: F(Q) = {1 , 2 } for disjoint D1 and D2 .


In this case we may assume 2 j (3Q) < 0 , j = 1, 2, since otherwise (3.16)

holds with C1 02 . Let d1 = sup2 3Q dist(z, 1 ). Then if 0 is sufficiently


small
 (3Q) 1 (3Q) + 2 (3Q) + d1
because for j = 1, 2, D j is an M-Lipschitz domain with diameter j (Q).

Also because D2 is an M-Lipschitz domain there exists z 3 (3Q) \ D1 D2


such that dist(z 3 ,  j ) cd1 , j = 1, 2. Consequently there is constant C2 such
that

2 (3Q) C2 2 1 (3Q) + 2 2 (3Q) + C2 d12




C2 2 1 (3Q) + 2 2 (3Q) + C1
Area(D j ).
G(Q)

Case 3: F(Q) contains at least three distinct  j .


Then because each D j is an M-Lipschitz domain there exists at least one

 j F(Q) such that  j (3Q) C1 , and (3.16) is clear in this case.

CY535/Garnett

0 521 47018 8

380

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

To finish the proof of Theorem 2.3, let  be a rectifiable curve and let { j }
be as in Corollary 3.2. By Lemma 2.4

2 j (3Q)(Q) c2 ( j ).
Q

2n

If j <
there are at most four dyadic squares Q such that (Q) = 2n and
D j G(Q). Hence


 1 
1
Area(D j ) =
Area(D j )
(Q)
(Q)
Q

G(Q)

D j G(Q)

Area(D j )

(2m j )1

m=0

Area(D j )( j )1

( j ) C M().

Therefore by Lemma 3.3


2 () C  1 ().

4. Schwarzian Derivatives
The Schwarzian derivative of a locally univalent analytic function is


  
1  (z) 2

(z)

S(z) =

2  (z)


  
3  (z) 2
(z)

=
.
 (z)
2  (z)

(4.1)

A calculation with the first expression in (4.1) shows that if


T (z) =

az + b
,
cz + d

where ad bc = 0, is a Mbius transformation, then


ST (z) = 0.

(4.2)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Schwarzian Derivatives

381

Conversely, if S = 0 on some disc, then for constants a and A = 0,


 (z)
2
=
,
 (z)
az

 (z) =

A
,
(a z)2

and (z) is a Mbius transformation. A calculation with the second expression


in (4.1) gives

2

(4.3)
S( )(z) = S((z))  (z) + S(z),
and therefore
S(T )(z) = S(z)

(4.4)

whenever T is a Mbius transformation. If is a locally univalent meromorphic


function we define

S = S 1/
in a neighborhood of a pole of . By (4.4), this definition is consistent with the
original definition (4.1) of S.
If is defined on the disc D and if T M is a self map of D, then by (4.2)
and (4.3)


2

2 

2 



1 |z|2 S( T )(z) = 1 |z|2 S(T (z))T  (z)


2 


= 1 |T (z)|2 S(T (z)).
Consequently the expression

2 

1 |z|2 S(z)

(4.5)

is a conformal invariant on D.
Theorem 4.1. If is univalent on D, then


2 
1 |z|2 S(z) 6.

(4.6)

Conversely, if is analytic on D and if




2 
sup 1 |z|2 S(z) < 2,

(4.7)

then is univalent and (D) is a quasicircle.


Proof. Suppose is univalent on D. By (4.4) we may assume
(z) = z + a2 z 2 + a3 z 3 + . . . .

CY535/Garnett

382

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

By the conformal invariance of (4.5) we only need to prove (4.6) at z = 0, where


it becomes


3
 

 (0) (  (0))2  = |6a3 6a22 | 6.
2

1
n
But since the function F(z) = (z)
= 1z +
n=0 bn z is univalent, the area
theorem, Lemma I.4.2, gives
|a3 a22 | = |b1 | 1,
which proves (4.6).
We omit the proof of the converse (4.7). Nehari first proved is univalent in
[1949], and it is the famous theorem of Ahlfors and Weill [1962] that (D) is
a quasicircle. See also Ahlfors [1963] and Lehto [1987]. The constants
 6 and 2
in (4.6) and (4.7) are both sharp. However, if supD (1 |z|2 )2 S(z) 2, then
is univalent (Nehari [1949]), and if (1 |z|2 )2 |S(z)| < 2 for all z D, then

(D) is a Jordan curve (Gehring and Pommerenke [1984]).
We will also need the local version of Theorem 4.1.
Theorem 4.2 (Ahlfors). Let U be a K -quasidisc and let : U D be conformal. Then there is = (K ) > 0 such that if is meromorphic on U and
if
1 |(z)|2 2 

S(z) < ,
sup

| (z)|
U
then extends to a quasiconformal mapping of C . In particular, is one-to-one
on U and (U ) is a quasicircle.
We will not prove Theorem 4.2 either. See Ahlfors [1963] or Lehto [1987].
Lehto [1987] has a much fuller discussion of Theorems 4.1 and 4.2, which lead
to the Bers identification of universal Teichmller space with an open connected
subset of the space of Schwarzians of univalent functions under the norm


2 
||||S = sup 1 |z|2 S(z).
D

Gehring [1977] proved a converse of Theorem 4.2: If U is a simply connected


domain for which the conclusion of Theorem 4.2 holds for some > 0, then U
is a quasidisc.
We always write

g(z) = log  (z)


(4.8)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

4. Schwarzian Derivatives

383

when is univalent on D. Theorem 4.1 is reminiscent of Theorem VII.2.1,


which said that for univalent
  (z) 


(4.9)
(1 |z|2 )|g  (z)| = (1 |z|2 )   6,
(z)
and (4.7) resembles Beckers [1972] converse: If
||g||B = sup(1 |z|2 )|g  (z)| < 1,
D

then is univalent. In fact, the interplay between the two conformal invariants
(1 |z|2 )|g  (z)|
and

2 
1 |z|2 S(z)

will be crucial throughout this chapter.


Note that if (1 |z|2 )|g  (z)| is small on


2
a hyperbolic ball B, then 1 |z|2 S(z) is also small on compact subsets
zz 1
of B. However, if (z) = zz
, then S = 0, but
2
g  (z) =

2
 (z)
=
 (z)
z2 z

(4.10)

is not small.
Lemma 4.3. Let > 0, a > 0, b > 0, and c > max(a, b) be given. Then
there is = (, a, b, c) such that if is univalent on the hyperbolic ball
{z : (z, z 0 ) < c}, and if


2 

(4.11)
1 |z|2 S(z) <
on the hyperbolic ball {z : (z, z 0 ) < a}, then for some Mbius transformation T ,

z z 
z z 0  
0 

(4.12)
 + (z) T 1
<
T ((z))
1 z0 z
1 z0 z
on the hyperbolic ball {z : (z, z 0 ) < b}.
Proof. We may assume z 0 = 0. There is a unique Mbius transformation T
so that
T (0) = 0, (T ) (0) = 1 and (T ) (0) = 0,

(4.13)

and we replace by T . By (4.8)


1
S(z) = g  (z) (g  (z))2 ,
2

(4.14)

CY535/Garnett

0 521 47018 8

384

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

and under the substitution


g  = 2

h
,
h

(4.14) becomes
S
h = 0.
2
This differential equation has a unique solution satisfying
h  +

h(0) = 1 and h  (0) = 0,


and the solution is obtained by power series: If
S(z) =

an z n ,

n=0

and
h(z) = 1 +

bn z n ,

n=2

then for n 2,
an2  bk an2k

.
2
2
n2

n(n 1)bn =

(4.15)

k=2

If is sufficiently small, then by (4.11) all an are small for small n. Thus for
any > 0 there is > 0 so that (4.11) and (4.15) give
sup |h(z) 1| < .

(z,0)<b

(4.16)

Unwinding from h now gives (4.12) for , if (4.16) holds for = ().

5. Geometric Estimates of Schwarzian Derivatives


Let be the conformal map of D onto a bounded simply connected domain
. In this section
we
give two different geometric estimates for the invariant

2
Schwarzian 1 |w|2 |S(w)|. The two estimates are not equivalent, but both
estimates measure how much  deviates from a half-plane or disc when  is
observed from the vantage point (w). The second estimate uses the numbers
E from Section 1, while the first estimate uses a more function theoretic
quantity  (z , t) defined immediately below.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Geometric Estimates of Schwarzian Derivatives

385

Fix ei D such that the nontangential limit z = (ei ) exists. For t > 0,
let L be the set of lines L such that
(a) L B(z , 2t ) = ,
and such that one component Wt ( ) of B(z , 2t) \ L satisfies both
(b) Wt ( ) , and
(c) when {(r ei : 0 < r < 1} is oriented by increasing r , Wt ( ) contains the
first arc of ({r ei : 0 < r < 1}) B(z , 2t) \ L ,
If L = , define
 (z , t) =

"
#
1
inf sup dist(z, ) : z L B(z , 2t) ,
2t LL

and if L = , put  (z , t) = 1.
(r ei )

(r ei )


Wt ( )

Figure X.16 On the left ; on the right = 1 while is small.


See also Exercise 12 for some remarks about the relation between  (z, t) and
 (z, t).
Theorem 5.1. Let z 0 = (w0 )  and let z = (ei )  satisfy
d = |z z 0 | 2dist(z 0 , ).
Then for s, 0 < s < 1, there is C = C(s), depending only on s, such that




2 

1 |w0 |2 S(w0 ) C
 (z , 2k d)2ks .

(5.1)

k=0

Proof. We may assume z 0 = w0 = 0 and d = 1. Write k =  (z , 2k d).


Fix = (s), 0 < < 0 where 0 will be determined later. We can suppose
0 , because otherwise (4.6) gives (5.1) with constant C = 6 . Since  is
bounded, limk k = 1 and there exists a first integer N such that N +1 > .
Let L 0 be a line with L 0 B(z , 21 ) = such that L 0 gives the minimum
0 . The line L 0 exists because 0 . Let H0 be the component of C \ L 0 such

CY535/Garnett

386

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

that H0 B(z , 2) . Then z 0 H0 by the first arc condition (c), because


z 0 = (0) B(z , 2). For k N , let L k be a line with L k B(z , 2k1 ) =
such that L k gives the minimum k . Again L k exists because k . Let Hk be
that component of C \ L k with Hk B(z , 2k+1 ) . Then by condition (c)
Hk Hk1 B(z , 2k1 ) = . Write Bk = B(z , 2k ) and Ak = Bk+1 \ Bk .
Set V0 = H0 and for 1 k N ,

%


Vk = int H0 B(z , 2)
(Hj A j ) (Hk \ Bk )  .
1 jk1

H2 \ B2
H1 A1
L2

z0

L1
z 0

Figure X.17 The region V2 .


Then Vk is a simply connected domain, Vk , and z 0 Vk but z 0
/ Vk if z 0
is the reflection of z 0 through L 0 = V0 .
By the distortion theorem, |  (0)| dist(z 0 , ) 1/2, and if 0 is small
then
*
= ({|w| = 1/4})
Vk .
Set G(z) = g (z, z 0 ) and G k (z) = gVk (z, z 0 ). Then the theorem is an immediate consequence of the next three lemmas:
Lemma 5.2. There is a constant C, independent of , such that






S(0) C supG(z) G 0 (z).

Lemma 5.3. If z , then


|G N (z) G 0 (z)| C(s)


k=0

k 2ks .

(5.2)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Geometric Estimates of Schwarzian Derivatives

387

Lemma 5.4. If z , then


|G(z) G N (z)| C(s)

N
1


k 2ks + C(s)2N s .

(5.3)

k=0

Proof of Lemma 5.2. By (4.6) we can assume




= supG(z) G 0 (z)

is small, because otherwise the lemma holds with C = 6 . Let T (w) =


the conformal map of D onto V0 . Then




sup log | 1 (z)| + log |T 1 (z)| = ,

z 0 w
w1

be

so that on

1 ( )

= {w : |w| = 1/4},
 T 1 ((w)) 


1 
 1 + .
w

Because T 1 (0) = 0, it follows via Taylor series that






sup T 1 ((w)) w  C
|w|1/8

for some constant with || = 1 and therefore that


 


S(0) = S(T 1 )(0) C,


which proves Lemma 5.2.


Proof of Lemma 5.3. First notice that by the definition of (z, t),

dist L k B k , L k1 B k C2k (k + k1 ).

(5.4)

Let r (r ) be the length of VN {|z| = r }. For k 1 and 2k r 2k+1 we


have
{|z| = r } Ak1 Ak Ak+1 ,
so that by (5.4)



(r )  C(k1 + k + k+1 ) C,

and
(r ) + C.
Now choose so that

s.
+ C

CY535/Garnett

0 521 47018 8

388

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Then for z and k 2, Theorem IV.6.1 gives us


)
+
2k1
dr
(z, Ak VN , VN ) C exp
r (r )
1
)
+
2k1

dr
C exp
C2ks .
+ C 1
r

(5.5)

In the same way, we also get


(z, ( VN ) \ B N , VN ) C2N s .
The function

 z z 

0
G 0 (z) = log

z z0

is harmonic on C \ {z 0 , z 0 } and when k 1, sup Ak |G 0 (z)| C2k since


|z 0 z | 2. If z Ak VN , then
min(k,N
 )

dist(z, V0 ) C2k

n ,

n=0

and therefore
G 0 (z) C

k


n .

n=0

But then since G N G 0 is harmonic on VN , (5.5) gives



 



G 0 ( )d(z, )
G N (z) G 0 (z) = 
VN

N min(k,N

 )
k=0

n=0

C(s)


n 2ks

2ns

nk

k 2ks ,

k=0

and that proves (5.2).


Proof of Lemma 5.4. Let  be a curve in VN such that for z  Ak
c1 dist(z, L k ) 2k (k1 + k + k+1 ) c2 dist(z, L k ),

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Geometric Estimates of Schwarzian Derivatives

389

provided k1 + k + k+1 > 0, and such that if k (s) is the arc length parameterization of  Ak , then
|k | c, and |k | c2k .
This curve  exists because of (5.4). If k1 + k + k+1 = 0, then  Ak+1
falls in a line segment and in this case let  Ak be a parallel segment such
that dist( Ak , ) 2k .

B2
B1
z0

L2


L1

z 0




Figure X.18
Let D VN be the simply connected domain such that z 0 D and such that
D consists of two arcs 1 = B N +1  and 2 B N +1 . Partition D into
arcs I j such that
dist(I j , ) diam(I j ).
Then if I j 1 Ak ,
diam(I j ) c2k (k1 + k + k+1 ).
Let z D Ak satisfy dist(z, ) 2k . Then for I j Ak 1 , a simple
comparison using the smoothness of  Ak yields
(z, I j , D) C diam(I j )2k .
By the proof of (5.5),
(z 0 , Ak 1 , D) C2ks .
Therefore,
(z 0 , I j , D) C diam(I j )2k 2ks C2ks (k1 + k + k+1 ).
On the other hand, if I j 2 , then

CY535/Garnett

390

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions


(z 0 , I j , D) (z 0 , 2 , D) C2N s ,

also by the proof of (5.5). Altogether these inequalities give us


(z 0 , I j , D) 2N s +

N


k 2ks .

(5.6)

k=0

Let Jj = 1 (I j ). Then
sup G(z) = sup sup log
D

Jj

1
sup sup(1 |w|).
|w|
j
Jj

(5.7)

Let w 1 (D) satisfy 1 |w | = sup(D) (1 |w|) and say w Jn . Let


Jn1 denote the two arcs adjacent to Jn and set J = Jn1 Jn Jn+1 . Then
if w 1 (D) \ J, we have
|w| |w | and (w, w ) C.
Set

U = 1 (D) {|w| < |w |} {(w, w ) < C} .

1 (D)
U

J
Figure X.19
Then by (5.6) and (5.7),
sup G(z) C1 (1 |w |) C2 (0, J, U)
D

C2 (z 0 , In1 In In+1 , D)
N
1


C2 2N s +
k 2ks .
k=0

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

5. Geometric Estimates of Schwarzian Derivatives

391

Therefore, since  VN D ,
sup |G G N | = sup(G G N )

sup(G G N ) sup G(z)


D


C2 2N s +

N
1


k 2ks .

k=0

That proves (5.3).




The second estimate of 1 |w|2 )2 S(w) uses the numbers from Section 2. It will also be valid in the more general situation in which is replaced
by a covering map. Let E be a compact set having at least three points and let
: D C \ E be a universal covering map, which by definition means that
is analytic, maps D onto C \ E, and is locally univalent, i.e.,  (z) = 0. (It
is a famous theorem of Koebe that such exist when E has at least three points,
see Ahlfors [1973], Fisher [1983], or Gamelin [2001].) Universal covering maps
satisfy a weakened form of Theorem I.4.3.
Lemma 5.5. If (w0 ) = z 0 , and if is univalent on B(w0 , (1 |w0 |2 )),
then
4
dist(z 0 , E) |  (w0 )|(1 |w0 |2 ) dist(z 0 , E).

Proof. We may take z 0 = 0 and w0 = 0. Then the right-hand inequality is the


(w)
1
Koebe estimate (I.2.11) applied to
 (0) . By the monodromy theorem,
has a single valued univalent branch on B(0, dist(0, E)) satisfying 1 (0) = 0,
and the left-hand inequality is the Schwarz lemma for this branch of 1 . 
When z 0
/ E, and let Q = Q 0 be the smallest dyadic square such that
z 0 Q and dist(z 0 , E) (Q). Let Q n be that dyadic square of side 2n (Q)
such that Q Q n and set
n = E (10Q n )
as defined in (2.2). Let Sn be a closed strip of width 10 2n+1 n (Q) such that
E 10Q n Sn ,
let L n be the axis of Sn , and let E n be the orthogonal projection of E 10Q n
onto L n . Let m n be the length of the largest interval in (L n 10Q n ) \ E n and
set
mn
n = E (Q n ) =
, n = n + n .
(Q n )

CY535/Garnett

392

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Clearly n measures the largest gap in E n .


Theorem 5.6. Fix 0 < < 1 and 0 < s < 1. Let : D C \ E be a universal covering map and w0 D satisfy (w0 ) =
z 0 . Then there is C = C(, s)
such that if is univalent on B w0 , (1 |w0 |) , then




2 

n 2ns .
1 |w0 |2 S(w0 ) C

(5.8)

n=0

Proof. The argument resembles the proof of Theorem 5.1. Fix = (s).We
can suppose 0 where 0 is to be determined later, because otherwise (4.6)
gives (5.8) with C = 6/. If N denotes the first integer such that N +1 > then
N > 1 and N is finite, because eventually n is large. We assume that w0 = 0
and that (Q 0 ) = 1.
Let S0 be a closed strip of width 20 such that
E 10Q 0 S0 .
/ S0 . Let H0 be the component of C \ S0 with z 0 H0 . For
Since 0 < , z 0
n < N , let Sn be a closed strip of width 2n+1 n such that
Sn Q n = and E 10Q n Sn ,
and let Hn be that component of C \ Sn such that
Hn Hn1 Q (n1) = .
The component Hn exists because 2n + n1 < 1, and Hn is unique because
diam(E) 2n .
This time write An = 5Q n \ 5Q n1 . Set V0 = H0 and for 1 n N ,


Vn = int (H0 5Q 0 ) (H1 A1 ) . . . (Hn An ) .
Then Vn is a simply connected domain, Vn (C \ E), and z 0 Vn but if z 0
is the reflection of z 0 through V0 , then z 0
/ Vn . The analogue of (5.4) holds
because n + n1 is small.
We continue to assume w0 = z 0 = 0. By the univalence hypothesis and the
distortion theorem, |  (0)| dist(z 0 , E) and
*
({|w| /8})
Vn .
Set
= ({|w| = /8}).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Schwarzian Derivatives and Rectifiable Quasicircles

393

z w

0
is the Mbius transformation from
Define G n = gVn (z, z 0 ). If T (w) = w1
D to the half-plane V0 with T (0) = 0, then
1
G 0 (T (w)) = log .
w

Since  (w) = 0, there exists, by the monodromy theorem, a unique branch


of 1 defined on VN and satisfying 1 (0) = 0. Using 1 , we define the
harmonic function g(z), z VN \ {0} by
1

(5.9)
G (w) = log 
w
on 1 (VN ).
At this point the proofs of Lemmas 5.2, 5.3, and 5.4 can be repeated with
replaced by and with k replaced by k to produce a proof of Theorem 5.6.
Note that (5.7) holds for the new G because of the definition (5.9). We leave

the details to the reader.

6. Schwarzian Derivatives and Rectifiable Quasicircles


Let be a conformal mapping from D onto a bounded Jordan domain . We

1
write F =  2 and, as always, g = log(  ).
Then
1
F  = Fg  ,
2
F  =
and

1  1  2
1
F g + (g ) = F S + (g  )2 ,
2
2
2

(6.1)

 2

2
|  |S  2|  |S + (g  )2  + 2|  ||g  |4
8|F  |2 + 8|F  |2 |g  |2 .

Because ||g||B 6, we therefore have




2

|  (z)|S(z) (1 |z|2 )3 d xd y 8
|F  |2 (1 |z|2 )3 d xd y
D


+ 8 36

|F  |2 (1 |z|2 )d xd y.

By Theorem 1.2 both terms on the right are bounded by C||F||2H 2 = ||  || H 1


and by the F. and M. Riesz theorem, ||  || H 1 = (). Therefore we have the

CY535/Garnett

0 521 47018 8

394

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

following lemma.
Lemma 6.1. There is a constant C such that if  is rectifiable, then

|  (z)||S(z)|2 (1 |z|2 )3 d xd y C().
D

When  is a quasicircle, Bishop and Jones [1994] proved the converse of


Lemma 6.1.
Theorem 6.2. If  is a bounded quasidisc, then  is rectifiable if and only if

|  (z)||S(z)|2 (1 |z|2 )3 d xd y < ,
D

and there is a constant C1 such that



1
C1 () diam() +
|  (z)||S(z)|2 (1 |z|2 )3 d xd y
D

(6.2)

C1 ().
The constant C1 depends only on the quasiconformality constant K of the
quasicircle . See Theorem VII.3.3.
Proof. By Lemma 6.1 and the trivial estimate diam() (), the right-hand
inequality of (6.2) holds, even when  is not a quasicircle.
To prove the left-hand inequality we must come to grips with the set where
|g  (z)| is large but |S(z)| is small. Let {Sj } denote the Whitney squares for 
and fix > 0 and > 0. We write Sj B, and say Sj is a bad square if
sup |g  (z)|(1 |z|2 ) ,

(6.3)



sup S(z)(1 |z|2 )2 .

(6.4)

1 (S

but

j)

1 (Sj )

Lemma 6.3. There is 0 > 0 and C > 0, such that if 0 < < 0 and > 0,
then


|  |ds C dist((0), ) + C
(Sj )
D

+ C(1 + 2 )

B,

2

|  (z)|S(z) (1 |z|2 )3 d xd y

(6.5)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

6. Schwarzian Derivatives and Rectifiable Quasicircles

395

 1
{ (Sj ) : Sj B, }. Using Koebes theorem and using

1
Theorem 1.2 twice with F =  2 , we get


|  (z)|ds dist((0), ) +
|  (z)||g  (z)|2 (1 |z|2 )d xd y, (6.6)

Proof. Write B =

and by (6.1)

|  (z)|ds
D

C0 dist((0), )
+ C0 |F  (0)|2
(6.7)




2
2
(g (z)) 

+ C0
|  (z)|S(z) +
 (1 |z|2 )3 d xd y,
2
D

for some constant C0 . Since ||g||B 6 we have


|F  (0)|2 =

|  (0)||g  (0)|2
36 dist((0), ),
4

so that (6.7) gives



|  (z)|ds C1 dist((0), )
D

2

|  (z)|S(z) (1 |z|2 )3 d xd y
+ C1
D


+ C1

|  (z)||g  (z)|4 (1 |z|2 )3 d xd y


+ C1

|  (z)||g  (z)|4 (1 |z|2 )3 d xd y,

D\B

for some constant C1 .


To estimate the integral over B we use ||g||B 6 and the Koebe theorem to
get

 |  (z)|
|  (z)||g  (z)|4 (1 |z|2 )3 d xd y 64
d xd y
1 |z|2
B

B,

64 C1

1 (Sj )

(Sj ).

B,

For the remaining integral we note that on D \ B either (6.3) fails, or (6.4)
fails but still |g  (z)|(1 |z|2 ) 6. Furthermore, if (6.4) fails for Sj , then by

CY535/Garnett

0 521 47018 8

396

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Theorem 4.1 and the Schwarz lemma, there are constants c1 and c2 such that
|S(z)|(1 |z|2 ) c1
on a disc E j 1 (Sj ) with |E j | c2 | 1 (Sj )|. Because
sup |  (z)| c3 inf |  (z)|

1 (S

1 (Sj )

j)

we consequently have

|  (z)||g  (z)|4 (1 |z|2 )3 d xd y
D\B

|  (z)||g  (z)|2 (1 |z|2 )d xd y

64 C
2

2

|  (z)|S(z) (1 |z|2 )3 d xd y.

Therefore

|  (z)|ds C1 dist((0), )
D


64

+ C1 1 + 2

+ 64 C1

2

|  (z)|S(z) (1 |z|2 )3 d xd y

(Sj ) + C1 2

|  (z)||g  (z)|2 (1 |z|2 )3 d xd y.

Then by (6.6) and (6.8),




|  (z)|ds C1 dist((0), ) + C2 2
D

(6.8)

|  (z)|ds


2


64

|  (z)|S(z) (1 |z|2 )3 d xd y
+ C1 1 + 2

D

4
(Sj ),
+ 6 C1
B,

and whenever C2 2 < 1, (6.5) holds with C =

64 C 1
.
1C2 2

Now assume  is a quasidisc with constant K . By a theorem of Ahlfors


[1963], there is A = A (K ) such that whenever  D is a circle, () satisfies
the Ahlfors three-point condition (VII.2.5) with constant A .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. The BishopJones H 2 Theorem

397

When T M, replacing by T does not change the integral in (6.2)


and so we can assume
dist((0), ) diam().

(6.9)

In that case we claim that if is sufficiently small, then every bad square is a
bounded hyperbolic distance from (0). By (6.9) and Lemma 6.3, that claim
will prove the theorem.
Suppose Sj is a bad square such that (z 0 ) Sj where 1 |z 0 |2 is small
and where (6.3) holds at z 0 . If is small, then by Lemma 4.3 there is a Mbius
transformation
z z1
Tz = A
z z2
such that



T  <

(6.10)

on the hyperbolic ball B = {z : (z, z 0 ) b} and


 T  (z ) 
0 

(1 |z 0 |2 ) 
 .
T (z 0 )
2
Hence by (4.10), the pole z 2 of T satisfies
dist(z 2 , B) |z 2 z 0 |

4
(1 |z 0 |2 ),

while for b = b() fixed so that sinh(2b) = 2 , B has euclidian radius


r 2 (1 |z 0 |) + O (1 |z 0 |)3 .
Consequently, if 1 |z 0 | is small, there are adjacent arcs I B and J B
with (I ) = (J ) 2 (1 |z 0 |) such that
C
(T (I ))
.
(T (J ))

But by (6.10), that contradicts (VII.2.5) for ( B).

7. The BishopJones H 2 Theorem


1

When  is not a quasidisc the condition



2

|  (z)|S(z) (1 |z|2 )3 d xd y <
D

(7.1)

CY535/Garnett

0 521 47018 8

398

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

of Theorem 6.2 no longer implies that  is rectifiable. For example, if  is


a half-plane, then S = 0 but  is not rectifiable. However in [1994] Bishop
and Jones obtained a sharp substitute theorem, and the proof of this theorem is
the key to the deeper results in this chapter.
Theorem 7.1. Let be the conformal mapping from D onto . If

2

B = |  (0)| +
|  (z)|S(z) (1 |z|2 )3 d xd y < ,
D
1

then for any , 0 < < 21 ,  H 2 , and


||  ||

H2

C()B.

(7.2)

In particular, if the BishopJones integral (7.1) is finite, then has non-zero


angular derivative almost everywhere on D, the cone points of  have full
harmonic measure relative to , and  1 by Theorem VI.4.2. Theorem 7.1
is sharp; again the counterexample is the map from D to a half-plane.
For the applications to come, Theorem 7.1 will be less important than its local
versions, Theorem 7.2 and Corollary 7.3. Recall we always write g = log(  ).

Theorem 7.2. Let E D be compact. Let U = E  ( ) be a cone domain, let 1 < < , and let > 0. Let be the conformal mapping from U
onto a simply connected domain  and assume that at every J E,

2

|  (z)|S(z) (1 |z|2 )2 d xd y N < .
(7.3)
 ( )

Then there is C(N , ) < and there exists J0 J such that |J0 | (1 )|J |
and

|  (z)||g  (z)|2 d xd y < C(N , )
(7.4)
 ( )

at every J0 .


Corollary 7.3. Let E D be compact. Let U = E  ( ) be a cone domain, and let 1 < < . Let be the conformal mapping from U onto a simply
connected domain . Then at almost every E for which

2

|  (z)|S(z) (1 |z|2 )2 d xd y <
(7.5)
 ( )

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. The BishopJones H 2 Theorem


we also have

|  (z)||g  (z)|2 d xd y < .

399

(7.6)

 ( )

The corollary follows by sending 0 and N in Theorem 7.2. Since


1
F(z) =  (z) 2 satisfies |F  (z)|2 = |  (z)||g  (z)|2 , Theorem 1.3 and Corollary 7.3 imply has a non-zero angular derivative at almost every point where
(7.5) holds.
The converse assertion, that (7.5) holds almost everywhere (7.6) holds, is

1
very easy. Since g = log(  ) and F =  2 we have |  ||g  |2 = 4|F  |2 , so that
(7.6) implies

  2
 F (z) d xd y < ,
 ( )

and by Lemma 1.5,

  2
 F (z) (1 |z|)2 d xd y <

 ( )

for any , 1 < < . Also recall from Section 6,


|  ||S|2 8|F  |2 + 8|g  |2 |F  |2 .
Since ||g||B 6, we conclude that

2

|  (z)|S(z) (1 |z|2 )2 d xd y <
 ( )

if (7.6) holds at , and a point of density argument then implies (if.5) holds
almost everywhere that (7.6) holds.
The proof of Theorem 7.2 resembles the main argument in the proof of
Theorem 7.1 and we will prove Theorem 7.1 first and then indicate the changes
needed to get Theorem 7.2.


Proof of Theorem 7.1. We take B = 1. Set ( ) = z : |z | < 1 |z| .
Given , 0 < < 21 and given 0 > 1, we construct a region R D and
a compact set E D such that
%
( ) R,
(7.7)
E



D \ E  C1+ ,

(7.8)

CY535/Garnett

0 521 47018 8

400

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

and

|  ||g  |2 (1 |z|2 )d xd y C,

(7.9)

where the constant C in (7.8) and (7.9) depends on but not on .


Assume that we have built sets R and E satisfying (7.7), (7.8), and (7.9).

1
Write p = 21 and as usual take F =  2 . Then  H p if and only if
F H 2 p and
||  || H p = ||F||2H 2 p .
Recall the area function


1

 F  (z)2 d xd y 2
AF( ) =
( )

1
4

|  (z)||g  (z)|2 d xd y

1
2

( )

By Theorem 1.1 F H 2 p if and only if AF L 2 p and


||F|| H 2 p C p ||AF||2 p .


Since { : z ( )} c(1 |z|2 ), we have by (7.7)


(AF)2 d C
|  (z)||g  (z)|2 (1 |z|2 )d xd y,
E

and thus by (7.8) and (7.9),




 

{ : AF( ) > t} D \ E  + 1
(AF)2 d
t2 E
1

C 2 p + C 2 ,
t
4

(7.10)

for all 0 because 21 p = 1 + . Take = t 2 p+3 so that the two terms


on the extreme right of (7.10) are equal. Then by (7.10) there is t0 = t0 (0 ) such
that



2p
t 2 p1 { : AF( ) > t}dt
||AF||2 p = 2 p
0
t0
4 p2 5
2p
2 t 2 p1 dt + C
t 2 p+3 dt
0

C()

t0

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. The BishopJones H 2 Theorem

401

if 0 < < 21 .
That proves (7.2), and the proof of Theorem 7.1 is reduced to constructing
sets R and E that satisfy (7.7), (7.8), and (7.9).
We first construct R. To start put {|z| < 21 } R and notice that we have

|  (z)||g  (z)|2 (1 |z|2 )d xd y C|  (0)|
(7.11)
|z|< 21

because ||g||B 6.
As in Section 3 take the dyadic Carleson boxes
Q = {r ei : 1 2n r < 1, j2(n+1) < ( j + 1)2(n+1) },
0 j < 2n+1 , of sidelength (Q) = 2n and their top halves
T (Q) = Q {1 2n r < 1 2(n+1) }
%
= Q \ {Q  : Q  Q, Q  = Q},
and write z Q for the center of T (Q). Fix and to be determined later. Say
Q L, for large, if


(7.12)
sup (1 |z|2 )2 S(z) > .
T (Q)

When Q is large, define D(Q) = T (Q). Say Q G, for good, if Q


/ L and
sup (1 |z|2 )|g  (z)| < .

T (Q)

Say Q B, for bad, if Q


/ L and Q is not good, i.e.,
sup (1 |z|2 )|g  (z)| .

T (Q)

 Q, ( Q)
 = 2(Q) satisfies
If Q
/ L, we call Q maximal if the next bigger Q
1

/ L. When Q M
Q L or if (Q) = 2 . Write M for the set of maximal Q
define
%
D(Q) = Q \ {Q  Q, Q  L}.
Then


z:

%

1
|z| < 1 =
D(Q),
2
LM

and the sets under this union are disjoint.

(7.13)

CY535/Garnett

0 521 47018 8

402

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions


Q1
Q2
D(Q 2 )

Q3
Figure X.20 Q 1 L; Q 2 , Q 3 , M.
To complete the construction of R we consider four cases and we estimate
the contribution to (7.9) in each case.
Case I: Q L.
Put D(Q) = T (Q) R and pass to the next level of boxes Q  Q with
(Q  ) = 21 (Q). Since ||g||B 6, Theorem 4.1 and the Schwarz lemma yield

|  (z)||g  (z)|2 (1 |z|2 )d xd y
T (Q)

6 2

2

|  (z)|S(z) (1 |z|2 )3 d xd y

(7.14)

T (Q)

for every Case I box.


Case II: Q G M.


1
Put D(Q) R and pass to the maximal Q  Q. Recall that F =  2
and 4|F  |2 = |  ||g  |2 . To estimate the contribution to (7.9) in Case II we need
the inequality

|  (z)|g  (z)|2 (1 |z|2 )d xd y
D(Q)

=4

|F  (z)|2 (1 |z|2 )d xd y C((Q))3 |F  (z Q )|2

(7.15)

D(Q)

+C

|F  (z)|2 (1 |z|2 )3 d xd y.

D(Q)

Because inequality (7.15) very much resembles (1.7) of Theorem 1.2 its proof
is left as an exercise for the reader. Remembering that 2|F  | = |F||S + (g  )2 |

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. The BishopJones H 2 Theorem

403

and that ||g||B 6, we obtain from (7.15)



|  (z)|g  (z)|2 (1 |z|2 )d xd y
D(Q)

C(Q)|  (z Q )| + C

2

|  (z)|S(z) (1 |z|2 )3 d xd y

(7.16)

D(Q)

|  (z)||g  (z)|4 (1 |z|2 )3 d xd y

+C
D(Q)

in Case II. We need the following lemma.


Lemma 7.4. Assume <
Q  G.

2
2 . If

Q G, if Q  Q, and if Q  D(Q) = , then

Accept Lemma 7.4 for a moment. Then we can bound the last term in (7.16)
using


|  (z)||g  (z)|4 (1 |z|2 )3 d xd y 2
|  (z)||g  (z)|2 (1 |z|2 )d xd y.
D(Q)

D(Q)

Therefore if C2 < 1 we have



|  (z)||g  (z)|2 (1 |z|2 )d xd y
D(Q)

C(Q)|  (z Q )| + C

2

|  (z)|S(z) (1 |z|2 )3 d xd y.

(7.17)

D(Q)

Consider the first term C(Q)|  (z Q )| on the right side of (7.17). Because Q
 for which (7.12)
is maximal, either (Q) = 1/2 or T (Q) is adjacent to a T ( Q)
holds. In the first case
(Q)|  (z Q )| c|  (0)|,
and this first case can occur for at most four squares Q. In the second case
Theorem 4.1 and the Schwarz lemma give



S(z)2 (1 |z|2 )3 d xd y c 2 (Q).
T (Q)

CY535/Garnett

0 521 47018 8

404

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Because ||g||B 6 it follows that the first term on the right-hand side of (7.17)
is majorized by a constant multiple of the second term, and hence that

|  (z)||g  (z)|2 (1 |z|2 )d xd y
D(Q)

2

| (z)|S(z) (1 |z|2 )3 d xd y.


(7.18)

D(Q)

Thus for Case II boxes (7.18) holds with at most four exceptions, when we
have the additional term c|  (0)|, and for (7.9) this additional term is harmless
if 0 1.
Proof of Lemma 7.4. Let r1 ei T (Q  ). There is r0 21 with r0 ei T (Q)
and [r0 ei , r1 ei ] D(Q). Set


r = sup s r1 : (1 t 2 )|g  (tei )| on [r0 , s] .
Then since
(1 t 2 )2 |S(tei )|
(4.14) gives
+ 2
d 
|g (te )|
<
dt 1 t 2
(1 t 2 )2
2



on [r0 , r ]. Therefore
(1 r 2 )|g  (r ei )| <
and r = r1 .

The remaining two cases concern the bad boxes Q B. We begin with a
lemma that shows the bad boxes are sparsely distributed.
Lemma 7.5. Given > 0 and an integer n > 0, there exist C = C() and
= (, n) such that if Q is a bad box for which
sup (1 |z|2 )2 |S(z)| ,

T (Q)

and if
B(Q) = Q {1 |z| 2n (Q)} {(1 |z|2 )|g  (z)| },
then there exists a hyperbolic geodesic such that
inf (z, ) < C,

B(Q)

(7.19)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. The BishopJones H 2 Theorem

405

where denotes hyperbolic distance in D. Moreover, given > 0 there is > 0


such that if (7.19) holds for , then
  (z ) 
1 |z | 2+
1 
0

(7.20)
 
C
(z 0 )
1 |z 1 |
whenever z 0 T (Q) and z 1 Q {1 |z 1 | 2n (Q)}.
Proof. If is small, then by Lemma 4.3 there is a Mbius transformation T
such that | T | is small on Q {1 |z| 2n (Q)}, so small in fact that
 T  (z)



g  (z)
(1 |z|2 ) 
T (z)
2
on Q {1 |z| 2n (Q)}. Let z 0 T (Q) {|g  (z)|(1 |z|2 ) }, and let
z be the pole of T . Then by Lemma 7.4 the conclusions of Lemma 7.5 all hold

when is the circular arc orthogonal to D that passes through z 0 and |zz | . 
Now fix = (, ) c 2 .
Case III: Q B M and

2

J=
|  (z)|S(z) (1 |z|2 )3 d xd y (Q)|  (z Q )|.
D(Q)

If possible, we follow the geodesic of Lemma 7.5 from z 0 T (Q) until it


hits a Q Q, Q B, such that
(Q )|  (z Q )| J (Q)|  (z Q )|,

(7.21)



= D(Q). In
and define D(Q)
= D(Q) \ Q . If no such Q exists, take D(Q)

either instance put D(Q)
R, and do not consider any smaller boxes Q  Q.

D(Q)
zQ

Figure X.21

CY535/Garnett

0 521 47018 8

406

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

By Lemmas 7.5 and 7.4 there is a cone  with vertex D such that
if Q j B and Q j Q, then T (Q j ) . Write





2
2
| (z)||g (z)| (1 |z| )d xd y =
+
.
D(Q)

D(Q)\

D(Q)



Let  be a cone with the same vertex and larger aperture. If Q j Q,


if Q j D(Q) = , and if Q j   = , extend the radial edges of Q j to .

This partitions D(Q)
\  into chord-arc domains D j with uniformly bounded
chord-arc constants. Choose z j  D j . Then by the proof for Case II and
Harnacks inequality,


|  (z)||g  (z)|2 (1|z|2 )d xd y =
|  (z)||g  (z)|2 (1 |z|2 )d xd y

D(Q)

Dj

|  (z)||g  (z)|2 (1|z|2 )d xd y

+

D(Q)

|  (z)||S(z)|2 (1|z|2 )3 d xd y

Dj

(1 |z j |2 )|  (z j )|

|  (z)||g  (z)|2 (1|z|2 )d xd y.


D(Q)

The values 1 |z j | decrease geometrically and each z j is a bounded hyperbolic distance from the geodesic . Take w j such that |w j | = |z j |. Then by
Harnacks inequality and (7.20),


(1 |w j |2 )|  (w j )|
(1 |z j |2 )|  (z j )|
C(Q )|  (z Q )| CJ.
Harnacks inequality also gives


|  (z)||g  (z)|(1 |z|2 )d xd y C
(1 |w j |2 )|  (w j )|

D(Q)

C(Q )|  (z Q )| CJ.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

7. The BishopJones H 2 Theorem

407

Since > 1, we therefore have



|  (z)||g  (z)|2 (1 |z|2 )d xd y

D(Q)

2

|  (z)| S(z) (1 |z|2 )3 d xd y

(7.22)


D(Q)

if Q is a Case III box and Q exists. If no box Q exists, we stop the sum at
the last j for which w j Q and we still obtain (7.22). Thus we have (7.22) for
all Case III boxes.
Case IV: Q B M and

2

|  (z)| S(z) (1 |z|2 )3 d xd y < |  (z Q )|(Q).
J=
D(Q)

Since c 2 , this case can only occur if (Q) = 1/2, and thus for at most
four boxes Q. Define Q by Q = and
(Q )|  (z Q )| (Q)|  (z Q )|

(7.23)



= D(Q). Put
and take D(Q)
= D(Q) \ Q . If no such Q exists take D(Q)


D(Q)
R and do not consider any smaller Q Q. The argument used in
Case III yields

|  (z)||g  (z)|2 (1 |z|2 )d xd y C|  (0)|,


D(Q)

which is be good enough because there are at most four such Q.


Proof of (7.7) and (7.9). By (7.13) we have
%
R = D \ {Q : Q B M}.
Since |  (0)| 6, (7.11), (7.14), (7.18), (7.22), and (7.23) give us (7.9) for R
provided B. Let I D be the base of Q and set
%
(7.24)
E = D \ {3I : Q B M}.
Then (7.7) holds on E.
Proof of (7.8). By (7.20), (7.21), and (7.23),
(Q ) C1+ (Q)

CY535/Garnett

0 521 47018 8

408

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

whenever Q B M. On the other hand the squares in B M have disjoint


interiors because they only arise in Cases III and IV. Therefore by (7.24),

|E| 3
(Q ) C1+ ,
QBM

and (7.8) holds.

Proof of Theorem 7.2. A point of density argument shows that there exists

J1 J , |J1 | (1 3 )|J | such that if W = J1  ( ), then

|  (z)||S(z)|2 (1 |z|2 )3 d xd y
W

|  (z)||S(z)|2 (1 |z|2 )2 d xd yds( ) C N .

J1  ( )

Set
D=

 ( )

J1

and
V=

%
{T (Q) : T (Q) D}.

Then V D. Define L, M, G, and B as in the proof of Theorem 7.1, but include


only those T (Q) such that T (Q) D. For such Q define
%
D(Q) = V Q \ {Q  Q D : Q  L}.
Then the proof of Theorem 7.1 yields a set E J1 , defined by (7.24) and a
region R V so that (7.7), (7.8), and (7.9) hold, and so that |E| (1 3 )|J |
and another point of density argument gives J0 E for which |J0 | (1 )|J |

and for which (7.4) holds.

8. Schwarzians and BMO Domains


Recall that a simply connected domain  is called a BMO domain if the mapping
function : D  satisfies
g = log(  ) BMO.
The results in this chapter yield two characterizations of BMO domains that
complement Theorem VII.5.3.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

8. Schwarzians and BMO Domains

409

Theorem 8.1. The following are equivalent.


(a)  is a BMO domain.
(b) There exist > 0 and C > 0 such that for all z 0  there is a subdomain
U  such that
(i) z 0 U,
(ii) U is rectifiable and (U) Cdist(z 0 , ), and
(iii) (z 0 ,  U, U) .
(c) There exist > 0 and C > 0 such that for all z 0  there is a subdomain
U  such that
(i) z 0 U and dist(z 0 , ) C dist(z 0 , U),
(ii) U is chord-arc with constant at most C and (U) C dist(z 0 , ),
and
(iii) ( U) dist(z 0 , ).
(d) |S(z)|2 (1 |z|2 )3 d xd y is a Carleson measure on D.
(e) There exist > 0 and C > 0 such that for every z 0 D there exists a
Lipschitz domain V D such that
(i) z 0 V,
(ii)
(z
0 , V D, V) , and
|  (z)||S(z)|2 (1 |z|2 )3 d xd y C|  (z 0 )|(1 |z 0 |2 ).

(iii)
V

Proof. Theorem VII.5.3 had the implications (a)  (b)  (c)  (a). Here
we use the arguments of the previous section to treat (a) (d), (a)  (e),
and (e)  (b).
(a)  (d): This was first observed by Zinsmeister [1984]. By (4.14) we
have
|S()|

|g  |2
+ |g  |.
2

It follows from (a) that |g  (z)|2 (1 |z|2 )d xd y is a Carleson measure and hence
|g  (z)|4 (1 |z|2 )3 d xd y is also a Carleson measure, because g B. For any
analytic function we have


|g  (z)|2 (1 |z|2 )3 d xd y C
|g  (z)|2 (1 |z|2 )d xd y,
T (Q)

(Q)
T

(Q) = {z : dist(z, T (Q)) (Q)/4}. Thus |g  (z)|2 (1 |z|2 )3 d xd y


where T
is also a Carleson measure and (d) holds.
(d)  (a): This is due to Astala and Zinsmeister [1991]. Because (a) and
(d) are invariant under conformal self maps of D, to prove (a) it is enough to

CY535/Garnett

0 521 47018 8

410

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

show


|g(ei ) g(0)|2 d C.

See Garnett [1981]. Let

A=

(8.1)

|g  (z)|2 (1 |z|2 )d xd y

and


B=

|g  (z)|2 (1 |z|2 )3 d xd y.

Then since

|g  (0)|

6 we have by Fourier series,


B 12 A 3B + 63

(8.2)

and by Theorem 1.2 it will be enough to show B C  .


By (4.14) we have


1
|S(z)|2 (1 |z|2 )3 d xd y +
|g  (z)|4 (1 |z|2 )3 d xd y.
B2
2
D

Set U = {z D :

|g  (z)|(1 |z|2 )

1/2}. Then by (8.2)

|g  (z)|4 (1 |z|2 )3 d xd y

A
B
9

+ .
4
12 2

(8.3)
(8.4)

Set V = {z D : |S(z)|(1 |z|2 ) > } where = (1/4, 10) from Lemma


7.5. Then because ||g||B 6,


64
|g  (z)|4 (1 |z|2 )3 d xd y 2
|S(z)|2 (1 |z|2 )3 d xd y. (8.5)

By Lemma 7.5, D \ (U V ) T (Q j ), where {Q j } is a sequence of Car


leson boxes with (Q j ) C  , with C  independent of . Hence


|g  (z)|4 (1 |z|2 )3 d xd y
64
(1 |z|2 )1 d xd y 64 C  .
j

D\(U V )

T (Q j )

Together (8.3), (8.4), (8.5), and (8.6) give us




64

9
23
B 2+ 2
|S(z)|2 (1 |z|2 )3 d xd y + + 64 C  ,
24
4

(8.6)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

9. Angular Derivatives

411

which establishes (8.1).


(d) and (a)  (e): Because (e) is invariant under Mbius self maps of D,
we can suppose z 0 = 0. Then let V be the Lipschitz region constructed in the
proof of (a)  (b) from Theorem VII.5.3 and note that |  | is bounded above
and below on V. Then use (d).
(e)  (b): We may assume z 0 = (0). We repeat the proof of Theorem 7.1,
with D replaced by the Lipschitz domain V given by (e), just as we did in
the
of Theorem 7.2. We obtain a Lipschitz domain R V such that
 proof


R | |ds < , with 0 R and with (0, D R, R) 2 . Then (b) holds

for U = (R).

9. Angular Derivatives
Let be a conformal mapping from D onto a simply connected domain  and
let
G = { D : has an angular derivative at and |  ( )| = 0}.
In this section we give several almost everywhere characterizations of G. By
Theorem VI.6.1 we already know that G is almost everywhere equivalent
to
( ) is a cone point of .

(9.1)

Furthermore, for any > 1, G is almost everywhere equivalent to



  2
 (z) d xd y < ,
(9.2)
 ( )

by the theorem of Marcinkiewicz, Zygmund, and Spencer, Theorem 1.3 above.


Bishop and Jones [1994] gives several other almost everywhere characterizations of G.
Theorem 9.1. Let be a conformal map from D onto a simply connected domain  and let g = log(  ). Then the following are almost everywhere equivalent on D:
(a) has non-zero angular derivative at .
2

|  (z)|S(z) (1 |z|2 )2 d xd y < .
(b)

( )

(c)
 ( )

|  (z)||g  (z)|2 d xd y < .

CY535/Garnett

0 521 47018 8

412

15:41

X. Rectifiability and Quadratic Expressions


(d)

January 27, 2005

2 (( ), t)

(e)

dt
< .
t



S(z)2 (1 |z|2 )2 d xd y < .

 ( )

Proof. The road map is: (a)  (b)  (c)  (a); then (d)  (a)  (d);
and finally (b)  (e)  (a).
(a)  (b): Suppose K D is a compact set such that
sup |  (z)| M

 ( )

for fixed M < for all ei K . Form the cone domain


%
U=
 (ei ).
K

Then



{ : z  ( )} c (1 |z|),

and Fubinis theorem again gives



2

|  (z)|S(z) (1 |z|2 )2 d xd yd
K  (ei )

2

| (z)|S(z) (1 |z|2 )3 d xd y.


(9.3)

But the integral on the right in (9.3) is finite by Lemma 6.1 because (U ) is
rectifiable. Therefore (b) holds almost everywhere on K .
(b)  (c): This is Corollary 7.3.

1
(c)  (a): Because |  ||g  |2 = 4|F  |2 , where F =  2 , Theorem 1.3 and
Theorem VI.2.3 show  has non-zero nontangential limit almost everywhere
that (c) holds.
(d)  (a): Fix > 1 large. If (d) holds on a set E of positive measure,
then for any > 0 there exists > 0 and compact K E such that

dt
2 (( ), t) <
t
0
and |E \ K | < . Then by Theorem 4.1 and the CauchySchwarz inequality,


S(z)(1 |z|2 )2 C 21 ,
(9.4)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

9. Angular Derivatives

on
V =

413

 ( ) {dist((z), ) 2 4 }.

K
1
2

Let U = V B(, 4 ) where is a point of density of K . If is sufficiently


large, then U is a quasidisc and by Theorem 4.2, (U ) is a quasidisc if is
sufficiently small. But then by Exercise 12, (( ), t) C(( ), t) on K .
Therefore by (d), by Theorem 2.5, and by the F. and M. Riesz theorem, (U )
has a tangent at ( ) for almost every K . Hence (9.1) and (a) hold almost
everywhere on K .
(a)  (d): Now suppose the angular derivative  ( ) exists and is continuous on a compact set K D. It is enough to show (d) holds almost everywhere
on K . Let > 0 be small and cover K by arcs I j = {| j | } and form
the cone domains
%
U j = {|z| 1 }
2 ( ).
K I j

If is small then U j is connected and  is continuous and almost constant on


U j , so that j = (U j ) is a chord-arc curve. By Theorem 2.5
1
dt
(9.5)
2j (z , t) <
t
0
almost everywhere on K I j . Let (9.5) hold at ei K I j and let t be so
small that
1
.
j (z , 2t) <
16
Let S be a strip of width 4t j (z , 2t) containing j B(z , 2t). Then one side
L of S satisfies (a), (b), and (c) in the definitions of (U j ) (z , t) and of  (z , t)
because (U j ) . Let z L B(z , 2t) and let z be the orthogonal (to L)
projection of z onto j . Then
dist(z, ) |z z | + dist(z , j (K )),
and by the definition of U j and the continuity of  on U j ,
dist(z , j (K )) Ct j (z , t).
Therefore
 (z , t) C j (z , t)
and by (9.5), (d) holds at ei .
(b)  (e): This is clear because (b)  (a).

CY535/Garnett

0 521 47018 8

414

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

(e)  (a): We follow the proof of Theorem 8.1, (d)  (a). Assume
(e) holds on a set E. Then there exists a compact set K E such that when

U = K  ( ),



S(z)2 (1 |z|2 )3 d xd y < .
(9.6)
U

Recall that g = log(  ). We will show (9.6) implies



  2
g (z) d(z)3 d xd y <
B=

(9.7)

where d(z) = dist(z, U ). Since |g  (0)| 6, it will then follow from Theorem 1.2 that g, and also  , has a finite non-zero nontangential limit almost
everywhere on K .
By (4.14)




  4
S(z)2 d(z)3 d xd y + 1
g (z) d(z)3 d xd y,
B2
2
U

and by Theorem 1.2



  2
g (z) d(z)d xd y C()(B + |g  (0)|2 ).
A=
U

Set U1 = U {z : |g  (z)|d(z) C()1/2 }. Then




2
  4
1
g (z) d(z)3 d xd y B + |g (0)| .
2
2
2
U1



For fixed > 0, set U2 = U {z : S(z)(1 |z|2 )2 }. Then because
||g||B 6,


  4
  4
g (z) d(z)3 d xd y
g (z) (1 |z|2 )3 d xd y
U2

U2

C
2



S(z)2 (1 |z|2 )3 d xd y.

U2

Finally, set U3 = U \ (U1 U2 ). By Lemma 7.5, there is a sequence {Q j } of




Carleson boxes such that (Q j ) < and U3 T (Q j ). But then


 
g (z)|4 (1 |z|2 )3 d xd y 64
(Q j ) < ,
U3

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

10. A Local F. and M. Riesz Theorem

415

and we obtain


 2
C

B
S  (1 |z|2 )3 d xd y + 64
2+ 2
(Q j ).
2

Therefore (9.7) holds.

10. A Local F. and M. Riesz Theorem


Theorem 10.1. Let  be a simply connected domain, let z 0  satisfy
dist(z 0 , ) 1, and let  be a connected set with 1 () M < .
Given > 0 there exists = (M, ) > 0 such that if E   and
(z 0 , E, ) > , then 1 (E) > .
The F. and M. Riesz theorem is a simple corollary of Theorem 10.1. In
[1980] ksendal conjectured Theorem 10.1 and proved it where  is a line. See
also Exercise III.14 and ksendal [1981]. Shortly thereafter Kaufman and Wu
[1982] proved the theorem when  is a chord-arc curve. But the full result was
not proved until Bishop and Jones [1990]. The difficulty in the general case is
to make effective use of the hypothesis 1 () M, and for this we will use
Theorem 2.3.
Proof. Let : D C \ E be the universal covering map, (0) = z 0 . By the
monodromy theorem there is a simply connected G D such that 0 G, is
univalent on G, and (G) = . By hypothesis
(0, G D, G) > .

(10.1)

Then by Beurlings projection theorem


dist(0, G) c(),

(10.2)

|  (0)| c ().

(10.3)

and hence by Lemma 5.5,

Let G denote the Fuchsian group { M : = } and let I be the


identity of G, I (z) = z. The normal fundamental domain of G is defined to
be
F = {z D : |  (z)| < 1 for all G \ {I }},

CY535/Garnett

416

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

and is one-to-one on F. Write G \ {I } = { j } and j (z) =


| j | = 1 and j (0) = a j . Then

j z+a j
1+a j j z ,

where

1 |a j |2
|1 + a j j z|2

and the fundamental domain F has the form D \ B j where each B j is a disc
orthogonal to D.
| j (z)| =

Lemma 10.2. The normal fundamental domain F satisfies


c()
2

(10.4)

|D F| 2 .

(10.5)

dist(0, F)
and

There exist (  ) > 1 and a compact set K D F such that


|K |
and

 ( ) F.

(10.6)

(10.7)

Proof. By (10.2), |a j | c() for all j, and since



|a j |
1
1


,
1
inf{|z| : | j (z)| = 1} =
|a j |
2
|a j |2
that gives (10.4).
By Lemma I.2 from Appendix I and Plessners theorem, G almost every
point of D G is a cone point for G. Write K G for the set of cone points of
G. Then by (10.1) and a comparison,
|D K G | 2 .
Because D K G j (K G ) = for all j , we have
 

 

D F j 1 (K G \ F)
D F  D F K G  +


D K G ,
because |( j 1 ) | 1 on j (F) (G \ F). That gives (10.5).
Write I j = B j D. By (10.4) we have
|I j |  < ,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

10. A Local F. and M. Riesz Theorem

417

and if > 0 is small there is > 1 such that if Jj is the arc concentric with
I j having |Jj | = (1 + )|I j |, then  ( ) B j = for all D \ Jj . Choose

and set
= 2(1)
%
K = D \
Jj .
Then (10.7) holds and (10.6) follows from (10.5). That proves Lemma 10.2. 
Fix 1 < < and form the cone domain
%
 ( ).
U=
K

If = (, ) is sufficiently small, then B(w, (1 |w|)) F for all w U.


We claim that

2

|  (z)|S(z) (1 |z|2 )3 d xd y C M.
(10.8)
U

If (10.8) holds then by Fubinis theorem



2

|  (z)|S(z) (1 |z|2 )2 d xd yds( ) < C  M,
K  ( )

and by Theorem 7.2 there exist N < and K 0 K such that |K 0 | > |K |/2

and F =  has area function A(F) N at all K 0 . Take


%
U0 =
 ( ).
K0

Then by Theorem 1.2, (U0 ) is a Jordan domain having


1 ((U0 )) C(|  (0)| + N ) C  (M, ),
(z 0 , E (U0 ))  (, )
by (10.6), and
dist(z 0 , (U0 )) c .
Therefore by (10.3) and Lavrentievs theorem, Exercise VI.2,
1 (E) c(, 1 ()),
and Theorem 10.1 is proved, provided (10.8) holds.
Proof of (10.8). We use Theorem 5.6. Let {Q j } denote the Whitney squares in
C \ E. We may assume each Q j is a dyadic square. Let Q nj denote the dyadic

CY535/Garnett

0 521 47018 8

418

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

square Q nj Q j such that (Q nj ) = 2n (Q j ), and let N (Q j ) be the first n such
that
n (Q j ) = E (Q nj ) + E (Q nj ) 103 .
If z U 1 (Q j ) then by Theorem 5.6,
N (Q j )



n (Q j )2ns (Q j ),
(1 |z| ) S(z) Cs

2 2

n=0

for any s < 1. Therefore





2


2 3
2


| (z)| S(z) (1 |z| ) d xd y
(Q j )
Qj

But by Lemma 5.5


U 1 (Q j )

|  (z)|
d xd y.
(1 |z|2 )

U 1 (Q j )

|  (z)|
d xd y C(Q j ),
(1 |z|2 )

and we obtain

2

|  (z)|S(z) (1 |z|2 )3 d xd y
U

2 (Q j )(Q j )

Qj

= 2C

(Q j )
 N
j

+ 2C

E (Q nj )2ns

2

(Q j )

(10.9)

n=0
(Q j )
 N
j

E (Q nj )2ns

2

(Q j ).

n=0

The two sums on the right of (10.9) will be estimated using the next two
lemmas. Write Q for the set of all dyadic squares.
Lemma 10.3. For n 0,



E2 (Q nj )(Q j ) : N (Q j ) n 2n
E2 (Q)(Q) C2n ().
Q

Proof. For each Q Q there are at most 4n dyadic squares Q j such that
Q nj = Q and for each of them (Q j ) = 2n (Q). Therefore

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

10. A Local F. and M. Riesz Theorem

419



E2 (Q nj )(Q j ) : N (Q j ) n
=

E2 (Q)



(Q j ) : Q nj = Q, N (Q j ) n

E2 (Q)


{(Q j ) : Q nj = Q}

2n

E 2 (Q)(Q) C2n (),

by Theorem 2.3.

We take s > 1. Then by Lemma 10.3 and the CauchySchwarz inequality,


(Q j )
 N
j

E (Q nj )2ns

2

(Q j )

n=0


1
s
12
n

E2 (Q nj )2ns (Q j )

{ j:N (Q j )n}



1
2(1s)n
E2 (Q)(Q)
s
12
n

(10.10)

C().
That bounds half of the right side of (10.9).
Lemma 10.4. For n 0,

{ 2 (Q nj )(Q j ) : N (Q j ) n} C2n ().
Proof. We can assume
(Q nj ) 105 (Q nj )

(10.11)

because Lemma 10.3 bounds the sum over the other Q nj . Using the Whitney
squares of C \ E, we partition  \ E into disjoint subsets Ik such that
c1 dist(Ik , E) diam(Ik ) c2 dist(Ik , E)
for absolute constants c1 and c2 . For each Q = Q nj satisfying (10.11) there is
k = k(Q) such that Ik 10Q and
(Q) c

diam(Ik )
.
(Q)

CY535/Garnett

420

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Then for each k and  there is a bounded number of Q Q such that k(Q) = k
and (Q) = . Because (Q) 10 diam(Ik ), that gives

k(Q)=k


1
C
C
2

.
(Q)
diam(Ik )
diam(Ik )

(10.12)

=0

Again because there are at most 4n squares Q j with Q nj = Q Q, we have by


(10.11)

{ 2 (Q nj )(Q j ) : N (Q j ) n and (10.11) holds}


2 (Q)
{(Q j ) : Q nj = Q}
=
(10.11) holds

2n

2 (Q)(Q)

(10.11) holds

C2

(10.11) holds

= C2n

2
diam(Ik )

C2


{Q:k(Q)=k}

2
diam(Ik (Q))
(Q)
1
(Q)

diam(Ik ) C2n (),

because the Ik are disjoint.

To estimate the second sum in (10.9), we keep s > 1 and repeat the proof of
(10.10) to get
(Q j )
 N
j

E (Q nj )2ns

2

(Q j ) C().

n=0

That concludes the proof of (10.9) and therefore (10.8).

11. Ahlfors Regular Sets and the HaymanWu Theorem


Recall that a compact set  is Ahlfors regular if there exists a constant A > 0
such that for every disc B(z, r ),
1 ( B(z, r )) Ar.
Theorem 11.1. Let  be a compact connected set. Then  is Ahlfors regular if

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

11. Ahlfors Regular Sets and the HaymanWu Theorem

421

and only if there is a constant C() such that


1 ( 1 ( )) C()

(11.1)

for every simply connected domain  and every conformal mapping from D
onto .
It is easy to see that (11.1) implies  is Ahlfors regular; take to be a
linear map from D to B(z, r ). The proof of the converse uses Theorem 10.1 and
Lemma 11.2 below.
Lemma 11.2. Let  be an Ahlfors regular compact connected set and let  be
a simply connected domain. Let S = {z j } be a sequence in   such that
sup inf (z, z j ) M.


(11.2)

Set

"
dist(z j , ) #
,
K j = z : |z z j |
2
and assume S has a finite partition S = S1 . . . S N , so that if z j Sn and

%
j =  \
Kk ,
{z k Sn ,k = j}

then
inf (z j ,   j ,  j ) > 0.

(11.3)

Then for any conformal : D ,


1 ( 1 ( )) C(A, N , M, ).



1 ()

Figure X.22

CY535/Garnett

0 521 47018 8

422

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Proof. Write w j = 1 (z j ) and D j = {w D : (w, w j ) M}. Then by


(11.2),

1 (D j 1 ( )),
1 ( 1 ( ))
while by the Koebe estimates in Theorem I.4.3 and the Ahlfors regularity of ,
1 (D j 1 ( )) C(A, M)(1 |w j |).
Thus the lemma will be proved if we show

1 |w j | C.

(11.4)

Sn

We can assume (w j , wk ) = (z j , z k ) 1 whenever z j = z k by taking


M 2 in (11.2) and by removing any unnecessary z j . Then the discs
j = {w : (w, w j ) 1/3}
are disjoint. Take 0 < <

1
8

to be determined and set

j = {w : (w, w j ) }.
Since < 18 we have j 1 (K j ) by the estimates in Section I.4.
When proving (11.4) we may assume that { j : z j Sn } is finite and that
0
/ j . Consider the harmonic function
%
u(z) = u n (z) = (z, D, D \ { j : z j Sn }).
By Greens theorem,
u(0) =

1
2

u(ei )d

 1
1
u
log ds,
2 j n
|z|
Sn

where the normal derivatives are exterior to j . Hence


+
)
+)



1
1
u
1
ds
u(ei )d u(0) 1.
inf log
j
|z|
2 j n
2 D
Sn

(11.5)

But
inf log
j

1
c4 (1 |w j |)
|z|

and in a moment we will prove the inequality



1
u
ds c5 .
2 j n

(11.6)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

11. Ahlfors Regular Sets and the HaymanWu Theorem

423

Therefore (11.4) follows from (11.5) and (11.6).


To prove (11.6) we use the functions
%
U j (z) = (z, D, D \ { k , z k Sn \ {z j }})
and
Vj (z) = (z, j , D \ j ).
By the maximum principle
u(z) U j (z) Vj (z)
if z D \ { k : z k Sn }. All the annuli D \ j have the same modulus, so
that when z j ,
Vj (z) = (),
where lim0 () = 0. By our hypothesis (11.3) and by Harnacks inequality
U j (z) c6 = c6 () on j .
Thus we can choose = () so small that
u(z) c7 = c6 /2 on

j .

Sn

D
Figure X.23 The proof of (11.6).
Let A j be the annulus ( j ) \ j . Then
u(z) c7 (z, j , A j ) on A j ,
so that

(z, j , A j )
u
c7
on j .
n
n

CY535/Garnett

424

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

But since all annuli A j have the same modulus, we have



(z, j , A j )
1
ds = C() > 0.
2 j
n
Therefore (11.6) holds with c5 = C()c7 .

The moment he saw this proof of Lemma 11.2, Burgess Davis came up with
a simple Brownian motion proof. See Exercise 15.
Proof of Theorem 11.1. Let {Q j } be the set of Whitney squares for  such
that Q j  = and fix z j Q j . Given T large, we partition {z j } into
N = N (T ) subsets S1 , . . . , S N such that (z j , z k ) T whenever z j , z k Sn
and j = k. To prove Theorem 11.1 we verify condition (11.3) of Lemma 11.2
for S1 S2 . . . S N .


2j


z j k


Figure X.24 The proof of Theorem 11.1.
2j be the simply connected component of  j B(z j , 10 diamQ j ) that
Let 
contains z j . By the Beurling projection theorem
 \ B(z j , 10 diamQ j ), 
2j ) 1/4.
(z j , 
Therefore (11.3) will hold if
2j ) 1/4,
(z j , ( j \ ) B(z j , 10 diamQ j ), 

(11.7)

because (11.7) implies


2j ) 1/2.
(z j ,   j ,  j ) (z j ,  B(z j , 10 diamQ j ), 
To establish (11.7) we let k be a line segment joining Q k  to  such
that (k ) 4 diam(Q k ). Now since  is Ahlfors regular we can, given > 0,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

11. Ahlfors Regular Sets and the HaymanWu Theorem

take T so large that


1 (B(z j , 10 diamQ j ) ( j
On the other hand,
2j )
 j = ( 

425

k )) diamQ j .

%
2j }
{k : z k 

is a connected set and ( j ) C diamQ j . Therefore Theorem 10.1 implies



inequality (11.7).

Notes
The books of Zygmund [1959] and Stein [1970], [1993] give more complete discussions of the Lusin area function. Our proof of Theorem 1.2 is from Coifman,
Jones, and Semmes [1989]. We have derived Theorem 1.3 from Theorem 1.2,
but Chapter XIV of Zygmund [1959] has a proof more like the arguments in
our Chapter VI. Sections 2 and 3 are based on Jones [1990]. For further results and applications see Jones [1989], [1991]; Bishop and Jones [1997a],
[1997b], [1997c]; David [1991], [1998]; David and Semmes [1991], [1993],
[1997]; Fang [1994]; Garnett, Jones and Marshall [1992]; Mattila, Melnikov
and Verdera [1996]; Okikiolu [1992]; and Pajot [2002].
We have barely touched the Schwarzian derivative. Lehto [1987] has an
excellent introduction. See Thurston [1986] for a completely geometrical explanation and see Chuaqui and Osgood [1998] and Osgood [1997] for more
recent results.
The recent work of Choi [2004] includes a proof of Theorem 10.1 that uses
cone domain constructions instead of numbers. Walden [1994] has the L p
version of Theorem 1.1: For some p > 1,

|( 1 ) (w)| p |dw| < .


Waldens theorem generalizes the Fernndez, Heinonen and Martio theorem


from Exercise VII.7. Lemma 11.2 is from Garnett, Gehring and Jones [1983].
See also Garnett [1986] and Exercise 15.

CY535/Garnett

0 521 47018 8

426

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

Exercises and Further Results


1. (a) If f H 2 (D) and g(z) H (D) then by Greens theorem and the product rule,


1
| f (z)|2 |g  (z)|2 (1 |z|)d xd y
| f (z)|2 |g  (z)|2 log d xd y
|z|
D

C||g||2 || f ||2H 2 .
(b) Use (a) to prove (1.7) without using Carleson measures. See Gamelin
[1980] and Coifman, Jones and Semmes [1989].
2. (a) Prove Lemma 1.5.
(b) More generally, prove that for n 2 and > 1,

n1

  2
 F (z) d xd y C(n, )
|F ( j) (0)|2
1

 ( )

+ C(n, )

 (n) 2
 F (z) (1 |z|)2n2 d xd y.

 ( )

(c) If n 2 and 1 < < , there is C(n, , ) such that




 (n) 2
  2
 F  (1 |z|)2n2 d xd y C(n, , )
 F (z) d xd y.
 ( )

 ( )

3. Let K be the Cantor set obtained from [0, 1] by repeatedly removing middle
halves, and let E = K K . Then 1 (E) < but E is not contained in a
rectifiable curve. Also construct an example with 1 (E) = 0. Hint: Alter
n
the construction so that at stage n there are 4n squares of side 4n .
4. (a) Show that the integral in (2.1) is infinite at z = 0 when E is the union of
two radii of the unit circle that meet at an angle less than .
(b) Prove that if
1
dt
E2 (z, t) < ,
t
0
then limt0 E (z, t) = 0.
(c) Suppose L n are lines through z with
n = E (z, 2n ) =

1
2n

sup

EB(z,2n )

dist(w, L n ).

If L n converges to a line L and n 0, then E has a tangent at z.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

427

1

(d) Prove that if n < (equivalently 0 E (z, t)/tdt < ) then E has
a tangent at z.
5. (a) Construct a Jordan curve  such that 0  and
(0, t) = 0, t < 1
but
 (0, t) 0 > 0 t < 1.
(b) Construct a Jordan curve  such that 0 ,
1
dt
 2 (0, t) <
t
0
but such that  fails to have a tangent at z = 0.
6. Let E be a finite set of points in C. Define a path from z 1 E to z n E to
be a collection {[z i , z i+1 ] : 1 i n 1} of line segments with endpoints
{z i }n1 E, and define a circuit to be a closed path (i.e. z 1 = z n ) such that
z i = z j , for 1 i < j n 1 and with n 4. A spanning tree G for E
is a collection of line segments with endpoints in E such that G contains a
path connecting every pair of points z 1 , z 2 E, and such that G contains no
circuit in E. A tour of E is a closed path containing every point of E.
The greedy algorithm is the following construction of a spanning tree for
E : Choose any z 1 E. Let z 2 E satisfy |z 1 z 2 | = minzE\{z 1 } |z z 1 |,
and put the line segment e1 = [z 1 , z 2 ] into G. Having chosen distinct points
z 1 , . . . , z m1 in E, choose z m E \ {z 1 , . . . , z m1 } so that for some k < m,
|z m z k | = min{|z z j | : z E \ {z 1 , . . . , z m1 } and 1 j m 1},
and put the line segment em = [z k , z m ] into G. Continue this process until
G = E.
(a) Prove that the greedy algorithm constructs a spanning tree of minimal
possible length. Hint: Let M be a minimal spanning tree and let G be the
spanning tree constructed by the greedy algorithm. Let k be the largest integer so that e1 , . . . , ek G M. If we add ek+1 to M, then M contains a
circuit all of whose segments have length no more that |ek+1 |. But by the
construction of G, this circuit must contain and edge e with exactly one
vertex in the collection of endpoints of e1 , . . . , ek and length at least |ek+1 |.
Replacing e with ek+1 , we obtain another minimal spanning tree M  with
e1 , . . . , ek , ek+1 G M  .
(b) Construct a tour T of E with (T ) 21 (G). Hint: Do induction on the
number of edges.
(c) Prove that every tour of minimal length contains a spanning tree.

CY535/Garnett

428

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

(d) If K is a connected set containing E with 1 (K ) minimal, then K


contains a closed rectifiable curve  with () 21 (K ).
(e) Use the above to show that
2 min{1 (K ) : K connected E}
min{() :  closed curve E}
= min{(T ) : T tour of E}

(E.1)

> 1 (G)
min{1 (K ) : K connected E},
where G is the spanning tree constructed by the greedy algorithm. A connected set of minimal length containing E does not necessarily consist of line
segments with vertices in E, but by (E.1) the greedy algorithm constructs
a connected set G E of length no more that twice the minimal possible
length. The greedy algorithm together with the construction in (b) constructs
a tour of length no more than twice the length of a shortest tour.
(f) Show that the method of Section 2 constructs a spanning tree for E. Each
segment comes from applications of Case 2.
(g) If E is a compact set in C contained in a rectifiable curve , then replacing
each arc in  \ E with a straight line segment with endpoints in E, we
obtain a rectifiable curve which is not longer than . Given > 0, choose
z 1 , . . . , z m E so that if z E then |z z j | < for some j. Applying
the greedy algorithm to {z 1 , . . . , z m } we then obtain a tree G such that
1 (G ) (), and dist(z, G ) < for every z E. Construct a tour T of
G with (T ) 1 (G ). Show that there is a limiting tour T = lim j 0 T j
with (T ) 2().
Constructing a tour of minimal length containing a finite set E is called
the traveling salesman problem. The greedy algorithm together with the
construction in (b) gives a tour of length no more than twice the minimal
possible length.
7. Recall from Chapter VII that a curve  is Ahlfors regular if there is a constant
A > 0 such that for every disc B(z, r ), 1 ( B(z, r )) Ar. Let E C
be compact. Prove E is contained in an Ahlfors regular curve  if and only
if there is a constant A such that for any z E and any 0 < R < diamE,

R
dt
E2 (w, t) d1 (w) A R.
t
EB(z,R) 0
See Jones [1990]. Related results can be found in David and Semmes [1991]
and Pajot [2002].

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

429

8. This exercise outlines the proof of the following result from Bishop and
Jones [1997a].
Theorem. Let E be a compact connected set such that
E (3Q) 0 > 0
for all dyadic squares Q such that Q E = and (Q)

(E.2)
diam(E)
.
3

dim(E) 1 + c0 2 ,

Then
(E.3)

where c is a constant independent of 0 .


(a) When E is a compact connected set define


E2 (3Q)(Q) : Q is dyadic and (Q) 2k ,
k2 (E) = diam(E) +
and


1
k (E) = inf () :  is rectifiable and inf dist(z, ) 2k 2 .
E

Then
C1 k2 (E) k (E) C2 k2 (E),
where the constants C1 and C2 do not depend on 0 . Hint: Apply Theorem
2.3 to
%"
#
Q : Q dyadic, (Q) = 2k , and Q E = .
k (E) =
(b) Let E be a compact connected set, let Q N be a dyadic square of side
such that Q E = . Then for n > N , 3Q contains at least
2N diam(E)
3
2nN dyadic squares Q n of side 2n such that Q n E = .
(c) Now let E be a compact connected set such that (E.2) holds for every
. Fix a dyadic square
dyadic Q such that Q E = and (Q) diam(E)
3
such
that
Q

E
=

.
Then
by (b),
Q N of side 2N diam(E)
N
3
N2 +k (Q N E) k0 2 2N .
Apply (a) to E Q = (E 3Q N ) Q to show 3Q N contains at least
Nk = ck0 2 2k
dyadic squares {Q j } such that
(Q j ) = 2N k ,


3Q j 3Q j = , j = j  ,

(E.4)

CY535/Garnett

0 521 47018 8

430

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

and
Q j E = ,
where c is an absolute constant. Hint: Consider  N +k (E Q ).
(d) We continue to assume the compact connected set E satisfies (E.2). Fix
k = [ 12 ] and note that by (E.4)
0

Nk 2(1+c0

2 )k

for some constant c. Assume diam(E) 3. By (c) and induction, there are
generations Gn = {Q nj }, n = 0, 1, . . . of dyadic squares of side 2nk such
that
Gn1 ,
(i) For n 1 each Q nj Gn is a subset of some Q n1
j
(ii) Each Q n1
Gn1 contains exactly Nk distinct squares Q nj Gn ,
j
and
(iii) Q nj E = for all Q nj Gn .
(e) If the compact set E satisfies (i)(iii), then
dim(E) 1 + c0 2 .

(E.5)

Hint: Define the measure by


(Q nj ) = Nk n 2(1+c0

2 )nk

Then has support E and (E) = 1. Moreover, there is a constant C = C(k)


such that Frostmans Theorem D.1 can be applied to C to yield
1+c0 2 (E) > 0.
(f) Except for the value of c in (E.5) the above theorem is sharp. If > 0,
this can be seen by a variation on the von Koch snowflake construction.
9. Let Q be a dyadic Carleson box in D, let z Q be the center of its top half T (Q),
let G(Q) be a family of pairwise disjoint Carleson boxes Q  Q, and let

D(Q) = Q \ G(Q) Q  . Let F H 2 have no zeros and || log F||B 6. This
exercise outlines a proof without using Appendix M of the inequality


|F(z) F(z Q )|2 ds C
|F  (z)|2 dist(z, D(Q))d xd y.
D(Q)

D(Q)

The construction is simpler in the upper half-plane and for this reason we
assume the box Q and its subboxes Q  are dyadic squares in H that have
one side in R. Let {Jk } be the set of maximal vertical line segments in

D(Q) G(Q) Q  . For each, Jk Q  (k) for a unique Q  (k) G(Q).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

431

Let Wk Q  (k) be the triangle bounded by Jk , a segment of slope 2, and


a segment Ik of length (J2k ) on the top edge of Q  (k). Define
%
U = D(Q) (Wk Jk Ik ).
k

Then U is a Lipschitz domain and the proof of Theorem 1.2 can be adapted
to yield (1.7) for U. Now since log F is a Bloch function it is not hard to
show


|F(z) F(z Q )|2 ds C
|F(z) F(z Q )|2 ds.
D(Q)

On the other hand,





2
|F (z)| dist(z, U)d xd y 2
|F  (z)|2 dist(z, D(Q))d xd y.
U

D(Q)

Hint: The reflections of the Wk through the Jk are pairwise disjoint.



10. Let be univalent

2 on D and let g(z) = log( (z)). If ||g||B is small, then


2
supD 1 |z| |S| is small, but not conversely. On the other hand if

2
supD 1 |z|2 |S| is small, then there is a Mbius transformation T such
that

 (T ) 


1 |z|2 
(T )
is small.
11. Suppose D D is a special cone domain and let be univalent on D.
Then Lemma 4.3 holds for : Given > 0, a > 0, and b > 0, there exists
= (, a, b) such that if


2 
dist(z, D) S(z) <
on the hyperbolic ball {z : D (z, z 0 ) < a}, then for some Mbius transformation T ,

z z 
z z 0  
0 

<
 + (z) T 1
T ((z))
1 z0 z
1 z0 z
on the hyperbolic ball {z : D (z, z 0 ) < b}.
12. Let  be a simply connected Jordan domain and let z .
(a) If  is a quasidisc, prove
1
 (z, t)  (z, t) C (z, t)
C
for small t, where the constant C depends only on the quasiconformality of
.

CY535/Garnett

0 521 47018 8

432

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

(b) For any constant M make an example for which


 (z, t) M (z, t).
(c) For any constant M make an example for which
 (z, t) M (z, t).
13. (a) Let D D be a special cone domain and let be a conformal map from
D onto a quasidisc . Write d(z) = dist(z, D). Then

2

1
() diam() +
|  (z)|S(z) d(z)3 d xd y.
C1
D

(b) Let D be a chord-arc domain with 0 D and dist(0, D) c diam(D).


Write d(z) = dist(z, D). Let F(z) be analytic on D. Verify that F H 2 (D)
if and only if

|F  (z)|2 d(z)d xd y < ,
D

and that

C||F||2H 2 (D) |F(0)|2 +

|F  (z)|2 d(z)d xd y

and
C  ||F||2H 2 (D) |F(0)|2 + |F  (0)|2 +

1
||F||2H 2 (D)
C

|F  (z)|2 d(z)3 d xd y

1
 ||F||2H 2 (D) ,
C
where C and C  depend only on the chord-arc constant of D and the constant
c above. Hint: Show (1.9) holds for D, and then use Theorem M.1 from
Appendix M.
(c) Let D be as in part (b) and let be a conformal mapping from D to a
domain  such that  is rectifiable. Then

2

|  (z)|S(z) d(z)3 d xd y C().
D

14. Prove inequality (7.15) without using Theorem M.1.


15. (B. Davis) Let {w j } D satisfy (w j , wk ) 1, let j = {(w, w j ) 18 },

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Exercises and Further Results

and assume

433


inf w j , D, D \
k > 0,
j

(E.6)

k = j

just as in Lemma 11.2. Let Pz denote probability on the space of Brownian


paths Bt starting at z and let E denote integration with respect to P0 . By
Kakutanis theorem the set
E j = {Bt hits j before hitting D}
has
c(1 |w j |)
.
(E.7)
2

Set N = E j . The random variable N counts the number of j visited
by Bt before Bt reaches D. By (E.7),

(1 |w j |) 2 E(N ).
P0 (E j )

In terms of Pz our hypothesis (E.6) says that


%

%

j before hitting D = inf Pz


Ek
inf Pz Bt hits
z j

z j

k = j

k = j

(E.8)

1 < 1.

Let 1 = inf{t : Bt j } be the first time Bt hits j . Assuming that
n has been defined and that Bn j (n) , let
%
n+1 = inf{t > n : Bt
j }


j = j (n)

be the time of the first visit to a different disc j . Then the functions n and
D = inf{t : Bt D} are measurable with respect to P0 .
Now let M = max{n : n D }. Then M 1 is the number of trips from
one j to another taken by Bt before it hits D. Therefore M N and
E(N ) E(M) =

P0 (M n).

n=0

Now by (E.8) and the independence of Brownian increments,



P0 (M n + 1) =
P0 (M n + 1; Bn j )
j

(1 )

P0 (M n; Bn j )

= (1 )P0 (M n),

CY535/Garnett

434

0 521 47018 8

January 27, 2005

15:41

X. Rectifiability and Quadratic Expressions

and by induction



2
(1 |w j |) 2
(1 )n1 =
.

n=1

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

Appendices

A. Hardy Spaces
If 0 < p < , then by definition an analytic function f (z) on the unit disc D
is in the Hardy space H p if

p
(A.1)
| f (r ei )| p d = || f || H p < .
sup
0<r <1

For p = the Hardy space H is defined to be the space of bounded analytic


functions on D with norm || f || = supD | f (z)|.
Write fr (z) = f (r z) and suppose 0 < r < s < 1. Then || fr || || f s || .
If p < , then
2
P r z | f s (ei )| p d
| fr (z)| p
0

| f (z)| p

is subharmonic. It then follows from Fubinis


because the function
theorem that || fr || p is an increasing function of r whenever 0 < p .
Theorem A.4. Assume 1 p < and let f H p . Then
(a) The function f (z) has a nontangential limit f (ei ) almost everywhere d
and
2


 f (r ei ) f (ei ) p d = 0.
(A.2)
lim
r 1 0

(b) If f 0, then



log f (ei )d > .

(A.3)

The case p > 1 of part (a) follows from Exercise I.11 of Chapter I, and the
significance of Theorem A.1 lies in the case p = 1 and the inequality (A.3).

435

CY535/Garnett

0 521 47018 8

January 27, 2005

436

15:41

Appendices

From part (a) it follows that

f (z) = lim f (r z) =
r 1

Pz ( ) f (ei )

d
,
2

and since || fr || p is increasing in r , we obtain


p

|| f || H p = || f || L p ( D)
for all f H p . Theorem A.1 is also true for f H p for all p < 1. See Exercise
A.1.
Part (a) of Theorem A.1 is due to F. and M. Riesz [1916], and part (b) is
from F. Riesz [1923]. Part (b) has the following important corollary.
Corollary A.5. If f H 1 and f 0, then | f (ei )| > 0 almost everywhere.
The proof of Theorem A.1 will use Theorem A.3 below, which is called
HardyLittlewood maximal theorem for H p functions:
Theorem A.3. Let > 1 be fixed, let 0 < p , and let f H p . Then the
nontangential maximal function
f ( ) = sup | f (z)|
 ( )

satisfies
|| f || p A || f || H p ,

(A.4)

where the constant A depends only on .


We first derive Theorem A.1 from Theorem A.3 and then return to the proof
of Theorem A.3.
Proof of Theorem A.1(a). First suppose p = 1. By (A.1) and the Banach
Alaoglu theorem, there are rn 1 so that f (rn ei )d converges weak-star to
a finite complex measure on D. Then


1
d
=
Pz ( ) f (rn ei )
Pz ( )d( ).
f (z) = lim f (rn z) = lim
n
n
2
2
Write
d = f d + d,
where f (ei ) L 1 (d ) and is singular to Lebesgue measure. Then


1
d
+
f (z) =
Pz ( ) f (ei )
Pz ( )d( ),
2
2

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

A. Hardy Spaces

437


d
has nontangential limit f (ei ) almost
and by Fatous theorem Pz ( ) f (ei ) 2
everywhere.
In a moment we will prove that

(A.5)
Pz ( )d = 0
lim
 ( ) z

almost everywhere. That will mean f (z) has nontangential limit f (ei ) almost
everywhere, and (A.2) will then follow from (A.4) and dominated convergence.
To prove (A.5) we can assume > 0. Fix > 1 and > 0. Take a compact
set K D so that |K | < and (D \ K ) < , and write 1 = K and
2 = 1 . Then lim z Pz ( )d1 ( ) = 0 for all D \ K . The proofs
of Lemmas I.2.2 and I.2.4 show that
"

# 3 + 6

(3 + 6)
 D : sup
.
Pz ( )d2 ( ) > 
d2


 ( )
Taking = 1/2 then gives
"

#

1/2 
1/2
 D : lim sup
Pz ( )d( )
 |K | + (3 + 6)

 ( ) z

+ (3 + 6)1/2 ,
so that (A.5) holds almost everywhere.
If 1 < p < , part (a) is easier. There are rn 1 so that frn ( ) = f (rn ei )
converges weakly in L p to some function f L p . Then

f (z) = lim f (rn z) =
n

Pz ( ) f (ei )

d
2

p
, and then Fatous theorem, (A.4), and dominated
because Pz L q , q = p1

convergence yield (A.2).

Proof of Theorem A.1(b). Part (b) follows from part (a). We may suppose
f (0) = 0, by dividing f by z n in case f (0) = 0. Then for r < 1 the familiar
Jensen formula gives

d
log | f (0)| log | f (r ei )|
2


d
+
i d
log | f (r ei )| ,
= log | f (r e )|
2
2

CY535/Garnett

0 521 47018 8

January 27, 2005

438

15:41

Appendices

where log x = (log(1/x))+ = max( log x, 0). By dominated convergence


and part (a),


lim log+ | f (r ei )|d = log+ | f (ei )|d,
r 1

while by Fatous lemma,




log | f (ei )|d lim inf log | f (r ei )|d.
r 1

Therefore we have the important inequality



d
log | f (0)| log | f (ei )| ,
2

(A.6)


which implies (A.3).

Proof of Theorem A.3.The proof will be divided into two cases, p 2 and
0 < p < 2. In the first case, p 2, f is the Poisson integral of an L p function
f (ei ) so that by Lemma I.2.2,
f (ei ) (1 + 2)M f (ei ),

(A.7)

where M f is the HardyLittlewood maximal function of f (ei ). That means


(A.4) is a consequence of the basic inequality
||M f || p C p || f || p ,

1 < p ,

(A.8)

which we will
establish in a moment. In (A.8) the constants C p depend on p,
but C p 2 6 for p 2. Except for its range of p, (A.8) is just (A.4) with
f replaced by M f , and for this reason (A.8) is also known as the Hardy
Littlewood maximal theorem.
The second case, 0 < p < 2, will be reduced to the first case by exploiting
the subharmonicity of log f (z).
Case I: p 2. We prove (A.8). Let f L p . For > 0 set


m() = {M f > }.
The function m() is called the distribution function of M f . By Fubinis
theorem,

2 " M f (ei )

#
p
p1
(M f ) d =
p
d d =
p p1 m()d.
(A.9)
0

We also have
m()

3|| f ||1
,

(A.10)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

A. Hardy Spaces

439

from Lemma I.2.4. Now we use an important trick, due to Marcinkiewicz, to


improve inequality (A.10). Set
f 1 = f {| f |> } ,
2

and
f 2 = f {| f | } .
2

Then M f M( f 1 ) + M( f 2 ), while M( f 2 ) || f 2 || 2 , so that by (A.10),




 6

m()  M( f 1 ) >
| f |d.

2
{| f |> 2 }
Then from (A.9) and Fubinis theorem we get


p
p2
||M f || p 6
p

{| f |> 2 }

=6





 f (ei )


| f |d d

2| f (ei )|

p p2 dd


3 p2 p
| f | p d.
p1
3p
1
p is decreasing we also have the bound
That proves (A.8), and because p1

C p 2 6 for p 2. Together (A.8) and (A.7) prove (A.4) when p 2 with


A 2 6(1 + 2).
The Marcinkiewicz argument is also valid when 1 < p < 2, but it unfortu 3p
1
p which are unbounded as p 1.
nately gives constants 2 p1
=

Case II. 0 < p 2. We can assume f 0. Take 0 < r < 1 so that f has
no zeros on {|z| = r } and set

1
Pz ( ) log | f (r ei )|d.
Ur (z) =
2
Then log | f (r z)| Ur (z) on D. Since Ur is real and harmonic on D, there is
r is analytic and
r such that Ur + i U
a unique harmonic conjugate function U

Ur (0) = 0. Write
p

gr (z) = e 2 (Ur +i Ur ) .

CY535/Garnett

440

0 521 47018 8

January 27, 2005

15:41

Appendices

Then | f (r z)| p |gr (z)|2 and




p
2
|gr | d = | f (r ei )| p d || f || p .

(A.11)

Let r 1, and let g be a weak limit of gr in L 2 . Then g H 2 because Poisson


2/ p
kernels are in L 2 and ||g|| H 2 || f || H p by (A.11). Moreover,
| f (z)| = lim | f (r z)| |g(z)|2/ p .
r 1

Therefore ( f ) p (g )2 L 1 , and
|| f || p ||g ||2 .
2/ p

Thus (A.4) holds for 0 < p < 2 with the same bound on A as for the case

p = 2.
Exercise A.1. (a) Suppose 0 < p < 1, and suppose f H p . Prove f has
nontangential limit f (ei ) almost everywhere and if f (0) = 0,
2
d
log | f (ei )| .
log | f (0)|
2
0
(b) On the other hand, if E D is compact, if |E| = 0, and if p < ,
then there is f (z) H p such that limr 1 | f (r ei )| = almost everywhere
on E.
Exercise A.2. (a) For p > 1 and > 1, there exists C( p, ) > 0 such that if
f L p (D) and u = u f then
||u || p C( p, )|| f || p .
(b) However there is f L 1 (D) such that sup0<r <1 |u(r ei )|
/ L 1 (D).
Exercise A.3. Let be a Carleson measure on the unit disc. By Exercise I.19(a)
it follows that


|u| p d C | f | p d
if p > 1 and if u is the Poisson integral of f L p . It also follows from Exercise
I.19(a) that


| f (z)| p d C | f (ei )| p d
if p > 0 and if f H p . These results are from Carleson [1958] and [1962].
Exercise A.4. If F(z) is analytic on D and if ReF(z) > 0 for all z D, then
F H p for all p < 1. Hint: Use Zygmunds theorem on conjugate functions

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

B. Mixed Boundary Value Problems

441

from Chapter II. Somewhat deeper is the weak-type estimate


|{ : |F | > }|

C
.

See for example Garnett [1981].

B. Mixed Boundary Value Problems


This appendix shows how to solve a mixed boundary value problem on a domain
 by reducing it to the Dirichlet problem on a Riemann surface double of .

Let  be a bounded finitely connected Jordan domain with  = m
j=1  j ,
1+
where each  j is of class C
and  j k = if j = k. Let v represent the
steady-state temperature of a thin homogeneous plate in the shape of . Then
v is harmonic in  and the loss of heat through an arc J on  is given by

v
ds,
J n
where n is the unit outer normal to . Given a function f on , the Neumann
problem is to find v C 1 () such that v is harmonic on  and v/n = f on
. In other words, find a steady-state temperature on the plate that produces a
prescribed heat flow f across the boundary. By Greens theorem f must satisfy
the necessary condition


v
ds = 0.
(B.1)
f ds =

 n
Except for issues of smoothness, (B.1) is the only condition necessary for the
Neumann problem.
Theorem B.1. Suppose  is a finitely connected Jordan domain such that
1+ for some > 0. Let > 0, and suppose f C () satisfies
 C
1
 f ds = 0. Then there is a harmonic function v C () such that
v
= f
n
on , and v is unique up to an additive constant.

(B.2)

If f satisfies (B.1) and is continuous but not necessarily Hlder continuous,


there is an elementary argument that yields a weaker version of the conclusion
v/n = f. See Exercise B.1.
Proof of Theorem B.1.The uniqueness is easy: we may assume f = 0 and,
via a conformal mapping, we may assume  consists of analytic curves. Each

CY535/Garnett

442

0 521 47018 8

January 27, 2005

15:41

Appendices

 has a neighborhood V and conformal map : V D so that


(V ) = R D
and
(V ) = D {Imz > 0}.
Then v 1 = Im H for some function H analytic on D {Imz > 0}. We
can take ReH = 0 on R D, since (ReH )x = (v 1 ) y = 0 on R D, by
Corollary II.4.5 and the CauchyRiemann equations. By the Schwarz reflection
principle, H extends to be analytic on D, with
Im H (z) = Im H (z).

(B.3)

Hence v extends to be harmonic on a neighborhood W of . But then by (B.3),


supW v(z) is attained on  and v is constant.
To prove there exists harmonic v C 1 () satisfying (B.2), first suppose that
 is simply connected and let (s) denote arc length parameterization of 
with positive orientation. Define
s
f ( (t))dt,
F( (s)) =
0

and let u be the solution to the Dirichlet problem with boundary data F. Then
u
s = f on , by Corollary II.4.6, and u = ReH for some analytic function
H , so that by Corollary II.4.5, v = Im H C 1+ (), for some > 0. Then
by continuity and the CauchyRiemann equations,
u
F
v
=
=
= f
n
s
s
on , and (B.2) holds.
When  is not simply connected, this definition may not give a single valued
function F, and not every harmonic function has a single valued conjugate
function. To get around those difficulties, we need two lemmas.
Lemma B.2. Let  be a bounded, finitely connected Jordan domain. Write

 = m
j=1  j , where  j k = for j = k and each  j is a Jordan curve.
Let j (z) = (z,  j , ). If u is a harmonic function on , then there exist
unique a1 , . . . , am1 such that
u(z) +

m1


a j j (z)

j=1

has a single valued conjugate function on .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

B. Mixed Boundary Value Problems

443

Proof. By Lemma II.2.2 we may assume each  j is an analytic curve. If


is a closed C 1 Jordan curve in , then 
u /s is a single valued function on
, where 
u is a locally defined harmonic conjugate of u, and there is a single
valued conjugate function 
u on the full domain  if and only if


u
ds = 0
s
for all such Jordan curves. If 1 and 2 are C 1 Jordan curves homologous in ,
then by Cauchys theorem,




u
ds =
u n ds = Im
(u x iu y )dz = 0,
1 2 s
1 2
1 2
since u x iu y is analytic on . Let j  be homologous to  j and define
the period of 
u over  j by

u
ds.
u) =
Pj (
j n
We have just seen that the Pj (
u ) does not depend on the choice of j . By Greens
theorem,
m


Pj (
u ) = 0,

j=1

so that u has a single valued conjugate if and only if


Pj (
u) = 0

for j = 1, . . . , m 1.

By Schwarz reflection each j extends to be harmonic on a neighborhood of


. Set

j
k, j =
ds.
(B.4)
k n
Then Lemma B.2 is proved if we can find a1 , . . . , am1 such that
m1


k, j a j = Pk (
u ),

1 k m 1,

j=1

and the next lemma ensures that such a1 , . . . , am1 exist.


Lemma B.3. The matrix {k, j }m1
j,k=1 is nonsingular.

CY535/Garnett

0 521 47018 8

January 27, 2005

444

15:41

Appendices


Proof. If k, j b j = 0 then, as we have seen, there is an analytic function F
on  such that
ReF =

m1


bj j .

j=1

But then ImF/n = 0 on , and by the (already proven) uniqueness part of


Theorem B.1, ImF is constant on . Hence F is constant, and since ReF = 0

on m , we have b j = ReF| j = 0, and thus {k, j } is nonsingular.
We return to the proof of the existence of v in Theorem B.1. By Lemma B.3,
there exist b1 , . . . , bm1 so that
m1


f1 = f

j=1

satisfies

bj

j
n


k

f 1 ds = 0,

(B.5)



for k = 1, . . . , m 1. Since  f ds = 0 and  j /n = 0 , the equality
(B.5) also holds when k = m. On each  j define
s
f 1 ( j (t))dt,
F( j (s)) =
0

where j (s) is the arc length parameterization of  j with positive orientation. Then by (B.5), F is well defined on  and F C 1+ () for some
> 0. By Corollary II.4.6 there exists u 1 C 1+ () with u 1 /s = f 1 . Find
a1 , . . . , am1 , by Lemma B.2, so that
u2 = u1 +

m1


a j j = ReH

j=1

for some function H analytic in . Let v2 = Im H . If , there is a


simply connected domain N  bounded by a C 1+ curve, so that N 
is a neighborhood of in . By Corollary II.4.5, v2 C 1+ () for some
> 0. Again continuity and the CauchyRiemann equations yield
v2
u 2
u 1
F
=
=
=
= f1
n
s
s
s
Now let
v = v2 +

m1

j=1

bj j .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

B. Mixed Boundary Value Problems

445

Then
 j
v
v2  j
=
+
= f1 +
= f,
bj
bj
n
n
n
n
m1

m1

j=1

j=1

and Theorem B.1 is proved.

The solutions to the Neumann and Dirichlet problems combine to solve


mixed boundary value problems.
Theorem B.4. Let  be a finitely connected Jordan domain with  C 1+

for
 some > 0. Let f 1 C(), let > 0, and let f 2 C () be such that
 f 2 ds = 0. Let J be a finite union of open arcs in  with pairwise disjoint
closures. Then there is a harmonic function u C 1 ( J ) C() such that
u = f1

on  \ J,

(B.6)

and
u
= f 2 on J.
n
If  \ J =
, the function u is unique.

(B.7)

Proof.
J

J
J

J
J
R

(J )

(J )

d

()
(J )
Figure B.1. The doubled Riemann surface d .
Form the doubled Riemann surface d by attaching two copies of 
along the set J . In other words, assume  {Imz > 0}, let (z)

= z map  to
the reflected domain () and set d =  J ( J ) / , where the
union denotes a disjoint union and where z w if and only if w = (z) = z
and z J. The sets  and () are declared open sets in d . To define a
basic neighborhood of p J , let W be a simply connected neighborhood of

CY535/Garnett

446

0 521 47018 8

January 27, 2005

15:41

Appendices

p in  such that W  J and let : W  D+ = D {Imz > 0} be


a conformal map so that the continuous extension of to W  satisfies
(W ) = (1, 1). The companion neighborhood (W ) in () is
(1, 1) so that (z) corresponds
identified with the reflection D of D+ through

to (z). The full disc V = W (W ) / then defines a neighborhood of p


in d , with local coordinate
'
(z),
z W,
(z) =
(z),
z (W ).
Thus, f is analytic on V if and only if f 1 is analytic on D. For example,
if  is the unit disc D, then d is conformally equivalent to C \ (D \ J ).
Notice that if v is harmonic on V and if v( (z)) = v(z), then v 1 / y = 0
on (1, 1), and since  C 1+ , it then follows from Theorem II.4.3 that
v/n = 0 on J W, where the derivative is now taken with respect to the
original coordinates on W .
We may assume  consists of analytic curves. Choose a domain 1 
with  \ 1 = J, such that d \ (1 ) is (conformally equivalent to) a domain 2  with 2  = J . On d we can apply the Schwarz alternating
method from the proof of Theorem II.1.1, with and  augmented by 1 
and 2 \ , respectively, and therefore the Dirichlet problem can be solved
on d .
By Theorem B.1, there is a u 1 C 1 () with u 1 /n = f 2 on . Let u 2
solve the Dirichlet problem on d for the boundary values f 1 u 1 on d .
Then, by the remarks above, u 2 /n = 0 on J , and therefore u 1 + u 2 satisfies
conditions (B.6) and (B.7).
To prove uniqueness when  \ J = , we can suppose f 1 = f 2 = 0. Then
if u satisfies (B.6) and (B.7), u extends to be harmonic on d with u = 0 on

d . Hence u = 0 by the maximum principle.
We do not know the origin of the idea of reducing mixed boundary value
problems on  to Dirichlet problems on Riemann surface doubles of . Ohtsuka
[1970] credited conversations between Ohtsuka and Strebel.
and continuous on D, and assume
Exercise B.1. (a) Let f be real valued
1
f d = 0. Then there is v(z) C (D), harmonic on D, with v/r = f on
D if and only if 
f is continuous on D. 
2
(b) If f C(D) is real valued and if 0 f (eit )dt = 0, set
2 i

e +z
d
H (z) =
f (eit )dt
.
i
e z 0
2
0

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

C. The Dirichlet Principle

Then

H (z) =

and
H
i
(z) =
r
r



2
log ei z
i

447


f (ei )

d
,
2


ei + z
d

1
f (ei ) .
i
e z
2

Thus v = Im H C(D) and v/r = f on D. This gives an alternate proof


of Theorem B.1 when  = D without the assumption that f C . However,
it does not prove v C 1 and in fact v may not be C 1 in this situation.
Exercise B.2. Use Greens function to find the kernel N (z, ), z , ,
such that under the conditions of Theorem B.1,

N (z, ) f ( )ds.
v(z) =


m1

Exercise B.3. Prove that the matrix k, j j,k=1 defined by (B.4) is symmetric,
k, j = j,k .

C. The Dirichlet Principle


A real valued function on D is piecewise continuous if f is continuous except
on a finite set E but f has finite left and right limits at each point of E. A real
valued function u(z) is piecewise smooth on a domain  if u(z) is continuous
on  and if u(z) is continuously differentiable on  \ J, where J = J (u) is a
finite union of (closed or open) analytic arcs in . The Dirichlet integral of u
or the squared Dirichlet norm of u is


|u|2 d xd y =
|u|2 d xd y.
D(u) = D (u) =


\J

If u = ReF and F is analytic on , then



D(u) =
|F  |2 d xd y = Area F(),


where each value F(z) is counted according to its multiplicity. If D(u) <
and D(v) < , the Dirichlet inner product is


u v d xd y =
(u x vx + u y v y )d xd y.
D(u, v) =


CY535/Garnett

0 521 47018 8

January 27, 2005

448

15:41

Appendices

It satisfies the CauchySchwarz inequality,


D(u, v)

$
D(u) D(v),

and
D(u v) = D(u) + D(v) 2D(u, v).

(C.1)

Given f C(), set




D f = v : v is piecewise smooth in  and lim v(z) = f ( ), for all  .
z

The Dirichlet principle is the assertion that the extremal problem


D(u) = inf{D(v) : v D f },

(C.2)

has unique solution u D f , and that u = u f is the solution of the Dirichlet


problem on  for the boundary value f . Unfortunately, this formulation of the
Dirichlet principle is not true. Here is a counterexample: If f C(D) has
Fourier series
+


f (ei )

an ein ,

(C.3)

then its Poisson integral is


u(r e ) =
i

+


an r |n| ein ,

0 r < 1,

and
D(u) = 2

+


|n||an |2 .

(C.4)

Hence there exist f C(D) such that D(u) = . For example,


f (ei ) =


1
cos(2k )
k2
k=1

is in C(D), but
D(u) =

k

2
k=1

k4

= .

However this example is the only thing wrong with Dirichlet principle. We shall
see that (C.2) holds whenever


D f v : D(v) < = .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

C. The Dirichlet Principle

449

Lemma C.1. Let f be piecewise continuous on D and let u be the Poisson


integral of f . Let E D be the finite set of discontinuities of f , and set

D f = v : v bounded and piecewise smooth on D, lim v(z)
z

= f ( ), for all D \ E .


If D f v : D(v) < = , then D(u) < , and u is the unique element of
D f of minimal Dirichlet norm. That is,
D(u) = inf{D(v) : v D f },
and if v D f satisfies D(v) = D(u), then v = u.
Proof. Let f have Fourier series (C.3) and set
u N (r ei ) =

N


an r |n| ein .

Then D(u N ) < . If v D f and D(v) < , then D(v u N ) < , and by
Greens theorem and dominated convergence,
2
u N
(r ei )d
D(u N , v u N ) = lim
(v u N )(r ei )
r 1 0
r
2
u N i
(e )d.
=
(v u N )(ei )
r
0

But on D, v u N = f u N has Fourier series |n|N +1 an ein , so that by
the Parseval identity,
D(u N , v u N ) = 0.
Consequently
D(v) = D(u N ) + D(v u N )

(C.5)

by (C.1), and
D(u N ) D(v).
Since |u N | |u| on D, Fatous lemma gives
D(u) lim inf D(u N ) D(v)
and we conclude both that D(u) < and that u has minimal Dirichlet norm
in D f .
To prove that u is the unique element of D f of minimal Dirichlet norm, note
that when D(u) < , (C.4) implies that D(u u N ) 0, and consequently

CY535/Garnett

450

0 521 47018 8

January 27, 2005

15:41

Appendices

that D(v u N ) D(v u). Hence by (C.5) we have


D(v) = D(u) + D(v u)

(C.6)

whenever v D f and D(v) < . Therefore if v D f and D(v) = D(u), then


D(v u) = 0 and v u is constant on . But then since v(z) u(z) 0
when z D \ E, that means v = u.

Because the Dirichlet integral is a conformal invariant, Lemma C.1 also
holds with any simply connected Jordan domain in place of D.
Lemma C.2. Let I be an open arc on D, and let f be a bounded continuous
function on I . Set


C f = v : v is piecewise smooth on D and lim v(z) = f ( ), for all I .
z

If



C f v : D(v) < = ,

then C f contains a unique element u(z) of minimal Dirichlet norm and u(z)
solves the mixed boundary value problem:

u = 0, on ,

u = f, on I,
(C.7)
u

= 0, on D \ I .
r
The set D \ I is called a free boundary because nothing is required of a
function in C f on D \ I .
Proof. Conformally map D to D+ = D {Imz > 0} so that D \ I is mapped
to [1, 1] D+ . Because Dirichlet integrals are conformally invariant and
because the map is conformal across D \ I , this does not change the extremal
problem inf{D(v) : v C f } nor the condition (C.7) expected of its solution.
Continue to write f ( ) for the boundary data on D+ \ [1, 1]. Let U (z) be
the Poisson integral on D of

f ( ), Im 0,
g( ) =
f ( ), Im < 0.
Then U satisfies (C.7) on D+ and D(U ) = inf{D(V ) : V Dg } by Lemma C.1.
If v C f were continuous up to (1, 1), its reflection

v(z), Imz 0,
V (z) =
v(z), Imz < 0,

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

C. The Dirichlet Principle

451

would be in Dg and the lemma would follow from Lemma C.1. However,
because v may not be continuous up to (1, 1) we must first make some minor
adjustments.

Figure C.1
For > 0, set = D+ {z : |z + i | = 1 + 12 }, and
" 
i 
i 
1 # * " 
1#

z : z  < 1 + 2 ,
W = z : z +  < 1 + 2

and for z W , let z satisfy Rez = Rez. For small > 0,




2
|v| dsd C
|v|2 d xd y,
0

W D +

and there is , 0 < < , so that |v|2 ds . Set

v(z), z D+ \ W ,
V (z) = v(z), z D+ \ W ;

v(z ), z W .
Then V Dg , so that
D(U ) D(V ).
But
DD+ (v) =

1
lim DD (V ),
2 0

and hence DD+ (U ) DD+ (v).


The proof of uniqueness is the same as for Lemma C.1. If D(v) = DD+ (U ),
then v U is constant on D+ , and because lim z (v(z) U (z)) = 0 for all

D {Imz > 0}, then v = U .

CY535/Garnett

0 521 47018 8

January 27, 2005

452

15:41

Appendices

Theorem C.3. Let  be a finitely connected Jordan domain, let E be a nonempty finite union of open arcs on  and let f C(E). Set


C f = v : v is piecewise smooth on , lim v(z) = f ( ), for all E ,
z

and assume



C f v : D(v) < = .

Then the extremal problem


D(u) = inf{D(v) : v C f }

(C.8)

has unique solution u, and u satisfies


u = 0 on ,
and
u = f on E.
Moreover, if for some > 0,  \ E consists of C 1+ arcs, then
u
= 0 on  \ E.
(C.9)
n
Proof. After a conformal mapping, we may assume  consists of analytic
Jordan curves. By Theorem II.4.3 this assumption does not change (C.9).
Let {u n } C f be a minimizing sequence for the extremal problem (C.8), so
that
D(u n ) inf{D(v) : v C f }.
Then by (C.1),
un + um

un um

D(u n ) + D(u m )
+D
=
,
2
2
2
and D(u n u m ) 0 because C f is a convex set. Hence there is a vector
(, ) L 2 () such that
"



 #
(u n )x 2 + (u n ) y 2 d xd y 0.
(C.10)
D

Moreover, (, ) is independent of the minimizing sequence {u n }. Fix an open


disc D such that D , and let Un be the solution of the Dirichlet problem
on D with boundary value u n , and replace u n C f by
Un D + u n \D C f .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

C. The Dirichlet Principle

453

Then the new sequence {u n } is still a minimizing sequence for (C.8). The
familiar estimate for harmonic functions
1 

1
2
4 2
2
sup
|U (z)|
|U
(z)|
d
xd
y
,
2
r
{|zz 0 |<r/2}
{|zz 0 |<r }

which follows from the CauchySchwarz inequality and the mean value property, now shows that both sequences {(u n )x } and {(u n ) y } converge to harmonic
functions, uniformly on compact subset of D. Since D is any disc in , we
conclude that and are (after alterations on a set of measure zero) harmonic
functions on  and

2un

= lim
=
.
n x y
y
x
We claim there exists a harmonic function u such that
u u 

,
= , .
u =
x y
Now (C.11) holds if

(C.11)

d x + dy = 0

for every smooth simple closed curve . Fix such a curve and take
a simply connected domain W such that W is piecewise analytic, such that
W , and such that \ W is a single point p. See Figure C.2.

p
W

Figure C.2
As before, we can suppose the minimizing functions u n are harmonic on a
neighborhood of p. In W let Un be the solution of the Dirichlet problem for
the boundary value u n . Then by Corollary II.4.6, Un is bounded on W.
Set vn = Un W + u n \W . Then vn is also a minimizing sequence, and by

CY535/Garnett

0 521 47018 8

January 27, 2005

454

15:41

Appendices

dominated convergence


vn
vn
dx +
dy = lim u n ( p) u n ( p) = 0.
d x + dy = lim
n
n
y

x
Therefore there exists a harmonic function u(z) satisfying (C.11). Because
E = and (C.10) holds, an integration shows lim u n (z) exists for all z .
Then since u n u on , we may assume lim u n (z) = u(z) for all z .
We now claim that u C f and that u satisfies (C.9). Let E. There is
> 0 such that
W =  B(, )
is a Jordan domain whose boundary consists of one arc 1 = B(, )  E
and a second arc 2 =  B(, ). On W let vn be the solution of the Dirichlet
problem with boundary value u n C( W ), and let v be the solution with
boundary value f 1 + u 2 . Then vn v and vn v on W . By Lemma
C.1, DW (vn ) DW (u n ), and in fact by (C.6)
DW (u n ) DW (vn ) = DW (u n vn ).
But
DW (vn ) + D (u n ) DW (u n ) = lim D (u n ),
n

because vn W + u n \W C f , and therefore DW (u n vn ) 0. With (C.10)


that gives
DW (u vn ) 0,
so that again by Fatous lemma,
DW (u v) lim inf DW (u vn ) = 0.
n

Hence u v is constant on W , and because v = u on 2 , u = v in W . Consequently


lim u(z) = f ( ),

and u C f .
To verify (C.9) let 1 be a component of  \ E and let 2  be a Jordan
arc connecting the endpoints of 1 so that 1 2 bounds a Jordan domain
W . Because u C f has minimal Dirichlet norm and functions in C f are
unrestricted on 1 ,


DW (u) = inf DW (v) : v piecewise smooth on W, v = u on 2
and Lemma (C.2) now implies u/n = 0 on 1 .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

C. The Dirichlet Principle

455

Corollary C.4. Let  be a finitely connected Jordan domain, and let E and F
be finite unions of open arcs on  such that E F = . Then the extremal
problem
"
inf D (v) : v piecewise smooth on ,
#
(C.12)
lim sup v(z) 0, lim inf v(z) 1
z F

z E

has a unique solution, which satisfies


u(z) = 0

on ,

on E

u=1

(C.13)

and
u=0

and

on F.

(C.14)

If  consists of C 1+ curves, then


u
= 0 on  \ (E F).
(C.15)
n
Proof. This is a corollary of Theorem B.4 and of Theorem C.3. We will give
both proofs; a third proof will be given in Appendix H. By conformal invariance,
 can be assumed to consist of analytic Jordan curves.
To derive the corollary from Theorem C.3, let > 0 and take h C 1 (R) so
that h(x) = x on [0, 1], and so that 0 h  1 and < h < 1 + on R. When
v is in the class (C.12), so also is h v and D(h v) D(v). Therefore the
extremum (C.12) is the same for the subclass in which lim z E v(z) = 0 and
lim z F v(z) = 1, and by Theorem C.3 the unique extremal for the subclass
is the function u satisfying (C.13) and (C.14).
To obtain the corollary from Theorem B.4 we follow Courants book [1950].
Let u(z) be the solution given by Theorem B.4 of the mixed boundary value
problem (C.13), (C.14), and (C.15). Then u(z) is a harmonic measure on a
Riemann surface double of , so that u/n is real analytic on the open set
E F .
At an endpoint of E F, |u| = O(|z |1/2 ). Indeed, assume that
E and that there is a neighborhood W of and a conformal map of
W onto D with ( ) = 0, (E W ) = [0, 1), and (W ) = D+ or D .
The function U (w) = u 1 (w 2 ) can be reflected from a quarter disc first
across the imaginary axis and then the real axis, to yield a function harmonic
in D \ {0}. Because u is bounded, the singularity at w = 0 is removable, and
U (w) = ImF(w) with F analytic on D and F(0) = 0. Because u > 0 on , F
has only a simple zero and F(w) = aw + O(|w|2 ), with a > 0. Thus
$
$
u(z) = ImF( (z)) = a(Im (z)) + O(|z |),

CY535/Garnett

456

0 521 47018 8

January 27, 2005

15:41

Appendices

and |u(z)| = O(|(z)|1/2 ).


Hence for any v in the class (C.12) and any sequence n  of domains
with smooth boundaries such that n , we have

u
D (v u, u) = lim Dn (v u, u) = lim
(v u) ds
n
n 
n
n

u
=
(v u) ds.
n
EF
Because u is a harmonic measure we have 0 u(z) 1 by Exercise H.5(a),
and hence u/n 0 on E and u/n 0 on F. Therefore D (v u, u) 0,
and
D (v) = D (u) + D (v u) + 2D (v u, u) D (u),
with equality holding if and only if v = u.

D. Hausdorff Measure
The two-dimensional Hausdorff measure of a set E C is defined to be
2 (E) = 1 Area(E), where

 2

(D.1)
Area(E) = inf
r j : E B(z j , r j )
is the area of E. To define other Hausdorff measures we replace r 2 in (D.1)
by other functions of r . A measure function is a continuous and increasing
function h(r ) on [0, ). With respect to the measure function h, the Hausdorff
content of a set E is



h(r j ) : E B(z j , r j ) .
(D.2)
Mh (E) = inf
The Hausdorff content Mh assigns finite mass to every bounded set, but the
content Mh is usually not finitely additive. For example if h(r ) = 2r, then
Mh ([0, 1]) = Mh ([0, 1] + i) = 1, while

$
Mh [0, 1] ([0, 1] + i) = 1 + 2 .
To obtain a measure, we change the definition (M.2) by requiring that the cover
of E be comprised of small balls only. Define



h(r j ) : E B(z j , r j ) and r j .
h (E) = inf
If 2 < 1 , then h2 h1 because h2 entails fewer covers. Consequently the

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

D. Hausdorff Measure

457

limit
h (E) = lim h (E)
0

exists in [0, ], and h (E) is the definition of the h-Hausdorff measure of E.


When h(r ) = r , we call h (E) the -dimensional Hausdorff measure of E
and write h (E) =  (E). For example, 2 (E) = 1 Area(E) as was noted
above, 0 (E) is the number of points in E, and if E lies on a rectifiable curve,
then 1 (E) is half the arc length of E. For any measure function h, h is a
metric outer measure in the sense of Carathodory:
h (A B) = h (A) + h (B), if dist(A, B) > 0,
and therefore h is a countably additive measure for which all Borel sets are
measurable. See Wheeden and Zygmund [1977] or Folland [1984].
Note that if < , then
 (E)  (E),
so that
 (E) <   (E) = 0.

(D.3)

The Hausdorff dimension of E is


dim(E) = inf{ :  (E) = 0}.
By (D.3) we have


 (E) =

0,
if > dim(E),
, if < dim(E).

Exercise D.1 shows that for E C, dim(E) can be any number in [0, 2].
Hausdorff content can be used to find dim(E). Clearly Mh (E)  h (E).On
the other hand, if Mh (E) = 0, then for any > 0 there is a cover B(z j , r j ) of

E such that h(r j ) < , and r j inf{t : h(t) } 0 as 0 because
h is increasing. It follows that h (E) = 0. Thus for any measure function h,
Mh (E) = 0

if and only if h (E) = 0,

(D.4)

and
dim(E) = inf{ : Mr (E) = 0}.

(D.5)

There is a dyadic version of Mh that is easier to work with. For integers j, k,

CY535/Garnett

0 521 47018 8

January 27, 2005

458

15:41

Appendices

and n, form the dyadic square



 

j j +1
k k+1
Q = n, n
n, n
.
2
2
2
2
It has side length (Q) = 2n and lower left corner ( j2n , k2n ). For each n,
the dyadic squares Q with (Q) = 2n are a disjoint paving of the plane. Every
 with ( Q)
 = 2(Q).
dyadic square Q is contained in a unique dyadic square Q
Therefore the dyadic squares have this important property:
If Q 1 and Q 2 are two dyadic squares, then

Q1 Q2,
Q 2 Q 1 , or

Q 1 Q 2 = .
Now define
m h (E) = inf



h((Q p )) : E


Q p , and each Q p is a dyadic square .

Every dyadic square of side r is contained in a ball of radius r while each ball of
radius r is contained in 25 dyadic squares Q with r/2 < (Q) r . Therefore
Mh (E) m h (E) 25Mh (E),

(D.6)

and (D.4) and (D.5) also hold with m h in place of Mh .


Using m h yields an easy proof that every E C has Hausdorff dimension
at most two. Partition the unit square Q into 4n smaller squares Q j such that
(Q j ) = 2n . Then if > 2,

(2n ) = lim 2n(2) = 0,
lim
n

and m r (Q) = 0. Thus  (Q) = 0 and since  is countably additive,


 (C) = 0. This shows that dim(E) 2 for all E C.
The following theorem is useful for computing dimensions and estimating
capacities.
Theorem D.1 (Frostman). Let h be a measure function. If is a positive
measure such that
(B(z, r )) h(r )
for all z and r , then
(E) h (E)

(D.7)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

D. Hausdorff Measure

459

for every Borel set E. Conversely, if E is a compact set, and if Mh (E) > 0,
then there exists a positive measure , supported on E, such that (D.7) holds
for all z and r and such that
Mh (E)
.
(D.8)
225


Proof. Suppose (B(z, r )) h(r ) for all z and r . If E B(z j , r j ) = B j
then


%
(B j )
h(r j ),
(E) ( B j )
(E)

and that proves the first assertion of the theorem.


Now suppose E is compact, and without loss of generality suppose E is
contained in the unit square. For each n 0, define the measure n so that for
every dyadic Q n with (Q n ) = 2n ,

h(2n ), if Q n E = ,
n (Q n ) =
0,
if Q n E = ,
and such that on Q n the measure n is absolutely continuous to area measure
and has constant density. Each Q n1 with (Q n1 ) = 2n+1 is the union of 4
squares Q n, j with (Q n, j ) = 2n . If

n (Q n1 ) =
n (Q n, j ) > h(2n+1 ),
j

then reduce the density on each Q n, j so that



n (Q n1 ) =
n (Q n, j ) = h(2n+1 ),
. Now repeat this process for squares Q k with
and call this new measure (n1)
n
(k)
(Q k ) = 2k and obtain measures {n }, for k = n 2, . . . , 0. We arrive at
(0)
(0)
a measure n satisfying n (Q) h((Q)) for all dyadic squares Q with
(Q) 2n .
(0)
(0)
The sequence {n } has a subsequence {n j } which converges weak-star
(0)

to a measure . Then because n has support in {z : dist(z, E) 2 2n },


supp E, Because weak-star convergence implies (U ) lim inf n (U ),
for all open U, we have (Q) 9h((Q)) for every dyadic square Q. Therefore
since each B(z, r ) is contained in 25 dyadic squares Q with r/2 < (Q) r ,
we obtain (D.6) for the new measure /225.
(0)
By construction, the measure n is supported on a union of dyadic squares
n
Q with (Q) = 2 such that either
(0)
n (Q) = h((Q)),

CY535/Garnett

0 521 47018 8

January 27, 2005

460

15:41

Appendices

 with
or Q is contained in a larger dyadic square Q


(0)
n ( Q) = h(( Q)).
Thus
(0)
n (E) inf

h(Q j ),

(D.9)

where the infimum is taken over all disjoint collections of dyadic squares {Q j }

satisfying (Q j ) 2n and Q j E. Then by (D.9) and (D.6)
(E) = lim (0)
n (E) m h (E) Mh (E),
and dividing by 225 yields (D.8).
Corollary D.2. Let h be a measure function such that
1
h(r )
dr < .
r
0
If E is a compact set such that if h (E) > 0, then
+
)
diam(E)
225
h(r )
dr .
Cap(E) diam(E) exp
Mh (E) 0
r

(D.10)

(D.11)

Proof. Suppose (D.10) holds, and suppose h (E) > 0. Then by (D.4),
Mh (E) > 0 and by Theorem D.1 there is a measure , supported on E, for
which (D.7) and (D.8) hold. Fix z E and write m(r ) = (B(z, r )). Then

1
d( )
log
U (z) =
|

z|
E
diam(E)
1
= lim
log dm(r ),
0
r
and by a partial integration,

diam(E)
m(r )
1
+ lim
dr
diam(E) 0
r
diam(E)
1
h(r )
+
dr.
(E) log
diam(E)
r
0

U (z) (E) log

Hence by Theorem III.4.1,


diam(E)

1
h(r )
1

(E) I
dr + log
,
(E)
(E) 0
r
diam(E)
which yields (D.11).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

D. Hausdorff Measure

461

Corollary D.2 shows that Cap(E) is positive if dim(E) > 0. The next corollary is a quantitative version of that result.
Corollary D.3. If 0 < 2, there is a constant C = C > 0 such that if E is
compact then

1

(D.12)
Cap(E) C Mr (E) .
Proof. By a change of scale z z, > 0, we may assume Mr (E) = 1.
Let be a measure on E satisfying (B(z, r )) r for all balls B(z, r ) and
1
. Then by a partial integration
(E) 225

1
d( )
U (z)
log
| z|
{| z|1}
1
((B(z, r ))
=
dr
r
0
1
1

r 1 dr = ,

0
and (E) 225 by Theorem III.4.1. That gives (D.12) with C e225 . 
Theorem D.4. Let

1
1
h(r ) = log
.
r
If Cap(E) > 0, then h (E) = .

(D.13)

Theorem D.4 is due to Erds and Gillis [1937]. It is almost a converse


to Corollary D.2. For example, when p > 1 and 0 < r < 1/3, the measure
function h(r ) = (log 1/r )1 (log log 1/r ) p , satisfies the hypothesis (D.10) of
Corollary D.2.
Proof. If Cap(E) > 0, then

lim

r 0 EB(z,r )

log

1
d E ( ) = 0
| z|

almost everywhere E . Then by Egoroffs theorem there exists compact J E,


such that E (J ) > 0 and there exists a positive decreasing function K (r ) such
that
log 1/r
= 0,
lim
r 0 K (r )

CY535/Garnett

0 521 47018 8

January 27, 2005

462

15:41

Appendices

and


K (| z|)d E ( ) M .

sup
zJ

Suppose r j and
J

N
%

B(z j , r j ).

Then
E (J )

E (B(z j , r j )) M

log 1/r  
1
M sup
h(r j ),
K (r j )
K (r )
r

and sending 0 shows h (J ) = .

Exercise D.1. (a) The Cantor set, defined in Example III.4.7, has Hausdorff
dimension
log 2
=
.
log 3

Recall that K = K n where K 0 = [0, 1] and K n is obtained from K n1 by
removing the middle one-third of each interval in K n1 . If = 3n /2, then we
can cover K n by 2n balls of radius and
 (K )  (K n ) 2n

3n

= 2 .

Letting n , we obtain  (K ) 2 and dim(K ) . For the opposite


inequality, let n be a constant multiple of one-dimensional Lebesgue measure
restricted to K n . Choose the constant so that each of the 2n intervals I in K n
has mass
n (I ) = 2n .
Let be a weak-star limit of n . Then
(K ) = lim (K m ) = lim lim n (K m ) = 1.
m

Let B = B(z, r ) and choose m so that


3m 2r < 3m+1 .
The distance between two intervals in K m1 is at least 3m+1 , so B intersects
at most one interval in K m1 . Then for n > m,
n (B) 2m+1 2+1r .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

D. Hausdorff Measure

463

Letting n , we obtain
(B) 2+1r ,
and by Theorem D.1,
1 = (K )  (K )2+1 .
Thus 21  (K ) 2 and dim(K ) . A more careful estimate shows
that  (K ) = 2 . An alternate proof can be found in Falconer [1985].
(b) If, when we construct K n from K n1 , we remove the middle intervals of
relative length 1 2, 0 < < 1/2, instead of the middle one-third, then we
obtain dim(E) = log 2/ log 1/.
(c) We describe a general construction of planar Cantor sets. Let K 0 be a
square in R2 with edge length l0 = 1 and let l j+1 < l j /2, j = 0, 1, . . . . The
(closed) set K j consists of 4 j squares of side length l j , each sitting in a corner
of a square in K j1 . The sets K 1 , K 2 , and K 3 are shown in Figure D.1.

K1

Let K =

K2
Figure D.1

K3

K j and let h be a measure function with


h(l j ) = 4 j ,

(D.14)

1
h (K ) 1.
36

(D.15)

for j = 0, 1, . . . Then

To prove (D.15), cover K j by 4 j discs of radius l j . Then for = l j ,


h (K ) h (K j ) 4 j h(l j ) = 1.
Letting j , we obtain h (K ) 1. To prove the lower bound on h (K ), let
j be a constant multiple of area measure restricted to K j . Choose the constant
so that each of the 4 j squares S in K j has mass
j (S) = 4 j .

CY535/Garnett

0 521 47018 8

January 27, 2005

464

15:41

Appendices

Let be a weak-star limit of j . Then


(K ) = lim (K m ) = lim lim j (K m ) = 1.
m

Let B = B(z, r ) and choose m so that


lm r < lm1 .
Then B intersects at most nine disjoint squares of side lm1 and hence at most
nine squares in K m1 . Thus for n > m
n (B) 9 4m+1 36h(r ).
Letting n , we obtain (B) 36h(r ) and by Theorem D.1
1 = (K ) 36h (K ),
and (D.15) holds.
(d) If l j = a j , 0 < a < 1/2, then the Hausdorff dimension of K = K a is
= dim(K a ) =

log 4
.
log 1/a

This shows that the Hausdorff dimension can be any number in [0, 2].
(e) More generally, if
= sup{t : lim inf 4 j l jt = } = inf{t : lim inf 4 j l jt = 0},
then dim(K ) = .
(f) Let  = C \ K , with K as in part (d). Prove g (z, ) is Hlder continuous, and find its best Hlder exponent.
Exercise D.2. (a) Assume the compact set E can be covered by A(r ) balls of
radius at most r and
1
1
dr = .
(D.16)
r
A(r
)
0
Then Cap(E) = 0. Hint: Without loss of generality, we may assume that A(r )
is a decreasing function of r , and that E B(0, 1/2). Then log 1/|z | 0
for all z, E. Assume Cap(E) > 0, and let be the equilibrium distribution
for E. Then

1
d( )d(z)
log
I () =
|z

|
E E
1
1
=
log dm(z, r )d(z),
r
E 0

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

D. Hausdorff Measure

465

where m(z, r ) = (B(z, r )). Show




1
1
log m(z, )d(z)
log dm(z, r )d(z) 0,

r
E
E 0
as 0, and

m(z, r )
dr d(z)
r
E 0
2n

dr

m(z, 2n1 )d(z) .


r
2n1 E

I () =

(D.17)

n=0

Now cover E by An+2 = A(2n2 ) balls B(z k , 2n2 ). We may suppose


that no point is in more than 16 balls, by modifying the collection of balls. Then
by (D.17)
An+2
2n

1 
dr
m(z, 2n1 )d(z)
I ()
n1
n2
16
r
2
B(z k ,2
)
n=0

k=1

2n

n1
n=0 2

An+2
1 
dr
m(z k , 2n2 )2
16
r
k=1

An+2
log 2  
m(z k , 2n2 )2 .
16
n=0 k=1

But
n+2


1 = (E)2

(B(z k , 2n2 )

k=1

Hence
> I ()

An+2

An+2

m(z k , 2n2 )2 .

k=1

1
log 2  1
dr,
C
16
An+2
0 r A(r )
n=0

which contradicts (D.15).


(b) Let K be the set constructed in Exercise D.1(c). Then Cap(K ) > 0 if
and only if


m=1

4m log

1
< .
lm

(D.18)

Hint: Suppose (D.17) holds. As in (D.14) let h be a measure function with

CY535/Garnett

0 521 47018 8

January 27, 2005

466

15:41

Appendices

h(lm ) = 4m . Then
1



h(r )
lm1
1
dr
h(lm1 ) log
=3
4m log
< .
r
l
l
m
m
0
m
m=1

Since h (K ) > 0, Corollary A.2 implies Cap(K ) > 0.


But if (D.9) fails we can cover K by 4m balls of radius lm , so let A be
decreasing with A(lm ) = 4m . Then
1


1
lm1
4m log
= ,
dr
r
A(r
)
lm
0
m=1

and by part (a), Cap(K ) = 0.


(c) Cantor sets can also be used to show that (D.10) is necessary in Corollary
D.2. Suppose h is a measure function such that h 1 (t/4) < h 1 (t)/2 and
1
h(r )
dr = ,
r
0
then there exists a compact set K with h (K ) > 0 and Cap(K ) = 0. Hint: Take
lm = h 1 (4m ), construct the Cantor set as in D.1(c) and apply part (b).
Further examples show that no complete description of capacity can be given
in terms of Hausdorff measures. See Carleson [1967a] for (a), (b), and (c).

E. Transfinite Diameter and Evans Functions


Transfinite diameter is a generalization of diameter that gives a geometric interpretation of capacity. Let E be a compact set. For z 1 , . . . , z n E set
)
+ 2
n(n1)
,
Dn (z 1 , . . . , z n ) =
|z i z j |
.
(E.1)
1i< jn

The exponent pn = n(n 1)/2 is chosen because the product Dn has pn factors, so that Dn is homogeneous of degree 1:
Dn (z 1 , . . . , z n ) = Dn (z 1 , . . . , z n ),

> 0.

The diameter of order n is defined by


dn = dn (E) =

max

z 1 ,...,z n E

Dn (z 1 , . . . , z n ).

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

E. Transfinite Diameter and Evans Functions


The usual diameter of E is d2 (E). A second quantity related to dn is
1
1
min
log
m n (E) = max
.
z 1 ,...,z n E zE n
|z z j |

467

(E.2)

Theorem E.1. If E is compact then


Cap(E) = lim dn (E) = lim em n (E) .
n

The limit d (E) = limn dn (E) is called the transfinite diameter of E.


Proof. We first compare (E.1) and (E.2). Choose z 1 , . . . , z n E so that
Dn (z 1 , . . . , z n ) = dn (E) and let
1 
1
Tj (z) =
log
.
n1
|z z i |
i;i = j

Write
log


2
1
1
=
.
log
dn (E)
n(n 1)
|z j z i |

(E.3)

i, j:i< j

If we view {z i : i = j} as fixed, then the terms in (E.3) which entail z j are



1
2
log
.
n(n 1)
|z j z i |
i:i = j

Thus z j is chosen so as to minimize Tj on E, and hence Tj (z j ) m n1 (E).


This implies
log

n
1 1 
1
1
=
log
dn (E)
n
n1
|z j z i |

1
=
n

j=1
n


i:i = j

Tj (z j ) m n1 (E).

j=1

z n . Then


min log
I N (n )

Now let n =

1
n

1
n2

log


1
, N dn (z)dn ( )
|z |

N
1
+
log
|z j z i |
n

i, j:i = j

N
1
+ .
dn (E)
n

(E.4)

CY535/Garnett

0 521 47018 8

January 27, 2005

468

15:41

Appendices

Take a subsequence {n j } converging weak-star to a measure P(E). Then





1
1
, N d(z)d( ) lim inf log
.
I N ()
min log
n
|z |
dn (E)
Letting N we obtain by Theorem III.4.1
(E) I () lim inf log
n

1
.
dn (E)

(E.5)

Suppose now that Cap(E) > 0, and let be the equilibrium distribution for
E. Then for any z 1 , . . . , z n E
n
n
1
1
1
1
inf

d(z)
log
log
zE n
|z z j |
n
|z

zj |
E
j=1

j=1

n
1
=
U (z j ) (E),
n
j=1

and hence
m n (E) (E).

(E.6)

By (E.4), (E.5) and (E.6),


(E) = lim log
n

1
= lim m n (E),
n
dn (E)

which proves the theorem when Cap(E) > 0.


If Cap(E) = 0, take E j decreasing to E such that Cap(E j ) > 0. Then by
(E.4)
em n1 (E) dn (E),
and by (E.5)
lim sup dn (E) lim sup dn (E j ) e (E j ) ,
n

and the Theorem follows.

The extremal (E.2) can be used to construct, for any compact set E of zero
capacity, a potential that is infinite exactly on E.
Theorem E.2 (Evans). Suppose E is compact. If Cap(E) = 0 there exists a
probability measure supported on E such that the potential U = U of
satisfies
lim U (z) = +,

z0

(E.7)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

E. Transfinite Diameter and Evans Functions

469

for all 0 E. Conversely, if there exist a function U harmonic on C \ E and


a constant K > 0 such that (E.7) holds and
U (z) K log

1
|z|

(E.8)

for |z| sufficiently large, then Cap(E) = 0.


See Evans [1933]. A potential U satisfying (E.7) is called an Evans function for E.
Proof. Suppose Cap(E) = 0. Choose z 1 , . . . , z n E so that
min
zE

and let n =

1
n

n
1
1
log
= m n (E),
n
|z z j |
j=1

n

j=1 z j .

Then

Un (z) =
log
E

1
d( ) m n (E) .
|z |

Choose a subsequence {Un j } of {Un } so that Un j (z) 2 j for all z E, and let
=

2 j n j .

j=1

Then
U (z) =

2 j Un j (z)

j=1

satisfies (E.7).
Conversely, suppose there is U harmonic in C \ E satisfying (E.7) and (E.8)
and suppose that Cap(E) > 0. Let be the equilibrium distribution for E and
let U be its potential. Then by the maximum principle
U (z) 2K U (z) +,
for z C \ E, which is impossible. Thus Cap(E) = 0.

Exercise E.1. If f is analytic in a region , then the cluster set of f at 


is


Cl( f, ) = w : there exists {z n } , with lim z n = and lim f (z n ) = w .
If K  then the cluster set of f at K is
%
Cl( f, K ) =
Cl( f, ).
K

CY535/Garnett

470

0 521 47018 8

January 27, 2005

15:41

Appendices

Suppose f is bounded and analytic on D \ K where K is compact and


Cap(Cl( f, K )) = 0. Prove f extends to be analytic and bounded on D. Rad
[1924] proved this result in the case Cl( f, K ) = {0}. Hint: First show we may
suppose f is bounded and analytic on C \ K , by considering

f ( ) d
f 1 (z) =
.
B(0,12) z 2i
By Cauchys integral formula f f 1 extends to be bounded and analytic on
C \ K . When f is bounded and analytic in a neighborhood of , f extends
to be analytic at , and we may suppose that f () Cl( f, K ) by replacing
f with f for some linear fractional transformation if needed. Now let U
be an Evans function for E = Cl( f, K ) and consider V = U f . Then V is
constant by Theorem E.2 and Corollary III.8.4, and hence f is constant.

F. Martingales, Brownian Motion, and Kakutanis


Theorem
Let (X, M, P) be a probability space, that is, a measure space with P(X ) = 1,
and let N M be a sub -algebra. For f L 1 (X, M, P) there is a unique
N -measurable function F such that


(F.1)
S f d P = S Fd P
for all S N . The function F is called the conditional expectation of f given
N , and it is written as
F = E( f |N ).
To prove F exists (and is unique almost everywhere) we use the Radon
Nikodym theorem: The signed measure N S (S) = S f d P is absolutely continuous to P on the -algebra N . Hence there is an N -measurable
F so that

(S) = S Fd P,
and F is unique almost everywhere. We will have F = f , unless f happens to
be N -measurable. If f L 2 (X, M, P), then F is the orthogonal projection of
f onto the closed subspace L 2 (X, N , P) of N -measurable functions.
Example F.1. Let (X, M, P) = ([0, 2 ), B, d/2 ), where B is the algebra

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

F. Martingales, Brownian Motion, and Kakutanis Theorem

471

of Borel sets. Let 0 = 0 < 1 < . . . < n = 2 , set


I j = [ j1 , j ),
and let N be the -algebra generated by {I1 , I2 , . . . , I N }. The N -measurable
functions are the step functions constant on the I j , and
E( f |N ) =

N


j Ij ,

j=1

where
1
j =
j j1


f d
Ij

is the average of f over I j .


Now let
M1 M2 . . . M
be an increasing sequence of -algebras such that

Mn

generates M.

(F.2)

n=1

Suppose f n is Mn -measurable, f n L 1 (X, Mn , P). We say the sequence { f n }


is a martingale (relative to {Mn }) if for all n
f n = E( f n+1 |Mn ).

(F.3)

From the uniqueness of conditional expectations it follows that


f n = E( f n+k |Mn ),

k > 0.

(F.4)

Martingales are ubiquitous in probability theory, see Chung [1974] or Doob


[1974].
Given the sequence Mn , we can always make a martingale by starting with
f L 1 (X, M, P) and setting
f n = E( f |Mn ).

(F.5)

It is a theorem that a martingale { f n } has the form (F.5) if and only if the sequence
{ f n } is uniformly integrable. We need only a weak form of that theorem.
Lemma F.2. Let 1 < p < and let { f n } be a martingale relative to {Mn }
such that
sup  f n  p < .
n

CY535/Garnett

0 521 47018 8

January 27, 2005

472

15:41

Appendices

Then there is f L p (X, M, P) such that


f n = E( f |Mn )
and f n converges to f weakly in L p .
Proof. From { f n } take a subsequence converging weakly in L p to some limit
f . For S Mn , S L q , q = p/( p 1) and hence (F.4) yields



S f n d P = lim S f n j d P = S f d P.
j

Hence f n = E( f |Mn ). By (F.2) the limit point f does not depend on the

subsequence { f n } and that means { f n } converges weakly to f .
Doobs martingale theorem says that an L 1 bounded martingale converges
almost everywhere. We will prove the theorem in the special case of L p bounded
martingales, and for L 1 bounded martingales we only prove lim sup | f n | <
a.e. Both results hinge on a crucial inequality, called the weak-L1 estimate for
the maximal function, which is the martingale variant of the estimate for the
HardyLittlewood maximal function, Lemma I.2.4.
Lemma F.3. Let { f n } be a martingale such that
sup  f n 1 = M < .
n

Set
f (x) = sup | f n (x)|.
n

Then for all > 0,


P( f > )

M
.

(F.6)

Proof. Fix and define


N (x) = inf{n : | f n (x)| > }.
Then
Sn {x : N (x) = n} = {x : | f n (x)| > }

n1
*
k=1

{x : | f k (x)| } Mn .

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

F. Martingales, Brownian Motion, and Kakutanis Theorem

473

Therefore
P( f > ) = lim
m

lim
m

lim
m

m


P(Sn )

n=1
m


Sn | f n |d P

n=1

m


Sn | f m |d P

n=1

lim  f m 1 M,
m

The second inequality above follows from


| f n | = |E( f m |Mn )| E(| f m |  Mn ).

(F.7)


So (F.6) is proved.

Incidentally, the function N (x) is called a stopping time, which by definition


means that {N (x) n} Mn .
Corollary F.4. If { f n } is a martingale such that supn  f n 1 < , then
lim sup | f n | < almost everywhere.
Proof. This is trivial because
P(lim sup | f n | > 1/) P( f > 1/) M
for all > 0.

Theorem F.5. Let p > 1 and let { f n } be a martingale satisfying


sup  f n  p < .
n

Let f be the weak limit provided by Lemma F.3. Then f n (x) f (x) for almost
all x.
Proof. By Hlders inequality we may assume p < . This case of the martingale theorem is easier because we already have
f n = E( f |Mn )
for some f L p . We may suppose f is real. The argument is like the proof of
Fatous theorem, I.2.1. By (F.6)
 f (x) = lim sup f n (x) lim inf f n (x)

CY535/Garnett

0 521 47018 8

January 27, 2005

474

15:41

Appendices

is finite almost everywhere and in fact


P( f (x) > )


| f |d P.

(F.8)

Let D = {g L p : g is Mn measurable for some n}. Then D is dense in L p


by (F.2) and g = 0 when g D. Now take g D such that  f g p < 2 .
Then  f =  f g and so (F.8) yields, via Hlders inequality,
2
2
 f g1 2 < 2.

Consequently  f = 0 almost everywhere and F(x) = lim f n (x) exists almost everywhere. Because { f n } is L p bounded for p > 1 it follows that
 f n F p 0, by Egoroffs theorem and the weak convergence of f n to

f . Hence F = f a.e. and the theorem is proved.
P( f > )

Now let  be a bounded domain in C and for each z  let (z) be the
circle
1
(z) = {w : |w z| = dist(z, )}.
2
For fixed z 0  we define a martingale z 0 , z 1 , . . . , inductively as follows. Take
z n+1 (z n ),

n0

such that z n+1 z n is uniformly distributed, which means that for A (z n ),
P(z n+1 A) =

length(A)
,
length((z n ))

n 0.

One often meets a martingale before seeing its -algebras Mn ; to define them
for {z n } let
X = {(ei1 , ei2 , . . .)}
be an infinite product of unit circles, let M be the Borel -algebra on X and
7
let P be the infinite product measure d j /2 . Set M0 = {, X } and let Mn
be the smallest -algebra for which 1 , 2 , . . . , n are measurable. Then
1
dist(z n , )ein+1 ,
2
and {z n } is a martingale relative to {Mn } because
2
1
ei d = 0.
2 0
z n+1 = z n +

Also note that (F.9) implies z n+1  and z n+1 = z n .

(F.9)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

F. Martingales, Brownian Motion, and Kakutanis Theorem

z0

475

z1
z2

zn

Figure F.1
Because  is bounded, the martingale theorem ensures that
z = lim z n
n

exists almost everywhere on X . Clearly z , and because


dist(z n , ) = 2|z n+1 z n | 0,
we have z  whenever z exists. Kakutanis famous theorem identifies
the hitting distribution of z with the harmonic measure of z 0 on .
Theorem F.6. Let z 0  and let E  be a Borel set. Then
(z 0 , E, ) = P(z E)

(F.10)

where z is the limit of the martingale (F.9) started at z 0 .


For example, let  be the (unbounded) domain {z : Rez > 0, |Imz| < 1}.
Let z 0 = 0 and let E =  {Rez n}. From Theorem F.6 one can see
w(z 0 , E, ) C1 ec2 n
because if k Rez j k + 1, then P(Rez < k + 2) has a positive lower
bound independent of k. Though Theorem F.6 has been stated only for bounded
domains, we can use it on  {|z| < N } and let N . The precise estimate
n

4
sinh( )e 2 + O(en )

2
follows from conformally mapping  to the upper half-plane.

w(z 0 , E, ) =

Proof of Theorem F.6. The key observation is that if u is harmonic in ,


then by the mean value property u n = u(z n ) is a martingale. First we use the

CY535/Garnett

0 521 47018 8

476

January 27, 2005

15:41

Appendices

observation to show
P(z is a regular point of ) = 1.

(F.11)

Suppose (F.11) is false. Then there is a compact set K , consisting only of


irregular points, such that P(z K ) > 0. By Kelloggs theorem Cap(K ) = 0
and by Evans theorem there is u(z) harmonic on C\K such that u(z) if
z K . Adding a constant, we may suppose
 u > 0 on . Then the positive
martingale u n = u(z n ) is L 1 bounded (since |u n |d P = u n d P = u(z 0 )) but
lim u n = on {z K }, contradicting Lemma F.3.
Now let f C() and let u be the solution of the Dirichlet problem for
the boundary data f . Then u n = u(z n ) is a bounded martingale and by (F.11)
u n f (z )
almost everywhere with respect to P. Hence by Lemma F.2 and the uniqueness
of martingale limits, u n = E( f (z )|Mn ) and in particular


f ( )dw(z 0 , ) = u(z 0 ) = u 0 =
f (z )d P.


Approximating characteristic functions by continuous functions now yields



(F.10).
The martingale {z n } is a discrete variant of Brownian motion. Let us give,
without proof, a brief description of Brownian motion. Let W be the set of continuous paths B(t), t 0, in the plane having common initial point B(0) = z 0 .
By a theorem of Wiener there is a probability measure P on W with respect
to which B(t) will almost surely meet any circle r (z 0 ) = {z : |z z 0 | = r }
at some finite time and the location of the first hitting of r (z 0 ) is uniformly
distributed:
length(A)
.
P(first hit of r (z) A) =
2r
All of the increments B(t + s) B(s) have the same distribution as B(t) z 0
and if
0 t 0 < t1 < < t n
then the increments
B(t1 ) B(t0 ), B(t2 ) B(t1 ), . . . , B(tn ) B(tn1 )
are stochastically independent:
,

*
P(B(t j+1 ) B(t j ) Sj ).
P
{B(t j+1 ) B(t j ) Sj }) =

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

F. Martingales, Brownian Motion, and Kakutanis Theorem

477

With an appropriate -algebra M, the measure space (W, M, P) is called


Brownian motion, and a random (with respect to P) function B(t) is called a
Brownian path.
Again let  be a bounded domain and let z 0 . Then
1 = 1 (B) = inf{t : B(t) (z 0 )}
is M-measurable and so is z 1 = B(1 ). Similarly,
2 = 2 (B) = inf{t > 1 : B(t) (z 1 ) = (B(1 ))}
and z 2 = B(2 ) are also measurable. In fact there is an infinite sequence
0 < 1 < of random times such that z n = B(n ). Thus our martingale z n
coincides with B(t) observed at the discrete sequence of times n . Now let
= inf{t : B(t) }. Because almost every path B(t) is continuous,
= lim n
and
B( ) = lim z n = z .
Therefore Theorem F.6 can be restated as
w(z 0 , E, ) = P(B( ) E),
or in words, the harmonic measure at z 0 of the Borel set E is the probability
that Brownian motion, started at z 0 , first exits  in E.
We do not know how to show directly that

f (z )d P
u(z 0 ) =
is a harmonic function of z 0 without using a prior solution to the Dirichlet
problem. Thus we cannot use (F.10) as an alternate definition of the solution of
the Dirichlet problem. Had we worked with Brownian motion B(t) instead of
the discrete version z n , there would have been no such difficulty because then
u(z) would have the mean value property for all discs contained in .
Exercise F.1. Let p > 1 and let f n = E( f, Mn ) be an L p bounded martingale.
(a) Show that limn || f n || p = || f || p .
(b) Use (a) to show
|| f n f || p 0.
(c) Prove the maximal theorem
|| f || p A p || f || p ,

CY535/Garnett

0 521 47018 8

January 27, 2005

478

15:41

Appendices

and use it to give another proof of (b).


Exercise F.2. In the probability space

d 
[0, 2 ), M,
,
2
where M is the Borel sets, let Mn be the -algebra generated by the dyadic
intervals
5 j2 ( j + 1)2 
, 0 j < 2n .
In, j =
,
2n
2n
Thus Mn consists of unions of thedyadic intervals In, j , and Mn Mn+1 .
For f L 1 ([0, 2 )), set f n = E( f Mn ) and for I = In, j ,

1
fI =
f d,
|I | I
so that
fn =

f In, j In, j .

We say f BMO(M), for dyadic bounded mean oscillation, if



sup E(| f f n |Mn ) = || f ||BMO(M) < .
n

Equivalently,
|| f ||BMO(M)

1
= sup
I =In, j |I |


| f f I |d.
I

(a) Prove f BMO(M) if and only



1
| f (ei ) |d < .
sup inf
|I
|

C
I
I =In, j
(b) Assume || f ||BMO(M) = 1, and let > 2. Prove the JohnNirenberg
theorem: There are constants C and c, independent of f , such that for all n,


P({ : | f (ei ) f n (ei )| > Mn }) E( {:| f fn |<} Mn )
(F.12)
Cec .
Equivalently,
sup

I Mn

|{ I : | f (ei ) f I | > }|
Cec .
|I |

Hint: Fix n, set N0 = n, and form the stopping time


N1 ( ) = inf{m > n : | f m ( ) f n ( )| > 2}.

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

F. Martingales, Brownian Motion, and Kakutanis Theorem

479

Then by Chebychevs inequality,




1
P(N1 < Mn ) E( {N1 <} Mn ) .
2
Now define
N2 ( ) = inf{m > N1 ( ) : | f m ( ) f N1 ( )| > 2},
and continue. The BMO assumption yields

P(Nk < Mn ) 2k ,
and when 2k < 2(k + 1) that gives (F.12).
(c) Without {Mn }, we say f L 1 ((0, 2 ]) has bounded mean oscillation,
f BMO, if for all intervals I D,

1
| f f I |d || f ||BMO < .
|I | I
See Exercise 21 of Chapter I. Suppose f BMO. Prove that if I J are
intervals such that |J | = 2n |I |, then
| f J f I | cn|| f ||BMO .
(d) Prove that if || f ||BMO 1, then for > 2 and for any interval I ,
|{ I : | f (ei ) f I | > }|
Cec .
|I |
Repeat the argument of (b), using subintervals of I having lengths 2n |I |.
(e) Let p > 1. There is C p such that if f BMO, then

1
p
sup inf
| f | p d C p || f ||BMO .
|I
|

C
I
I
This is immediate from (d). See John and Nirenberg [1961].
Exercise F.3. Let u(z) be the Poisson integral of f L 1 (D).
(a) Prove f BMO if and only if



 f (ei ) u(a)2 Pa ( )d < .
sup
aD D

Hint: Compare the Poisson kernel to the box kernel |I I| as in Figure I.4 from
Chapter I and use the JohnNirenberg theorem.
(b) Using (a) and Greens theorem, show f BMO if and only if

 1 az 
d xd y < .
|u|2 log
sup
za
aD
D

CY535/Garnett

0 521 47018 8

January 27, 2005

480

15:41

Appendices

See Exercise II.8.


(c) Using (b) and the identity
2
2
 z a 2
 = (1 |a| )(1 |z| ) ,
1
2
1 az
|1 az|

show f BMO if and only if


|u|2 (1 |z|2 )d xd y
is a Carleson measure. It follows that f BMO if and only if its conjugate
function 
f BMO. Fefferman and Stein [1972].
(d) It follows from (c) that BMO is conformally invariant. If f (z) is analytic on D and if f (ei ) BMO, then for every Mbius transformation
za
, a D,
T z = 1az
|| f T ||BMO C|| f ||BMO .
For detailed proofs see, for example, Garnett [1981].

G. Carlemans Method
Let  be a rectangle with sides parallel to the axes, let z 0  and let E be
the right edge of . If finitely many horizontal slits are removed from , then
the right side of inequality (IV.6.3) does not change, but the harmonic measure
(z 0 , E) becomes smaller. This defect in Theorem IV.6.1 is circumvented in
Carlemans [1933] version of the theorem.
Theorem G.1 (Carleman). Let  C be a domain, let x =  {Rez = x}
and let E b =  {Rez b}. Suppose |x | M < , and let (x) denote
the length of the longest interval in x . Assume z 0 = x0 + i y0 B(z 0 , r0 ) .
Then for b > x0

(z 0 , E b , )

2r0
9M 2

b
x0


exp 2

t
x0

!  21
dx
dt
.
(x)

(G.1)

The function (x) is measurable because it is lower semicontinuous. The


main improvement of (G.1) over (IV.6.3) is the replacement of (x) with the
smaller (x). Further comparisons will be given after the proof of Theorem G.1.
The proof uses Wirtingers inequality: If g and g  are real valued continuous
functions on the interval (a, b) and if g(a) = g(b) = 0, then
2 b

b

(g  )2 d x
g 2 d x.
(G.2)
ba
a
a

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

G. Carlemans Method

481

Wirtingers inequality (G.2) is proved by examining the Fourier series of g.


Lemma III.3.2 is very similar to Wirtingers inequality.
Carlemans idea is to find a differential inequality for the Dirichlet integral
of the harmonic measure (z, E b , ). To prove Theorem G.1, we may suppose
that  is bounded, that  consists of finitely many analytic Jordan curves, and
that inf{Rez : z } = 0. Write (z) = (z, E b , ) and define
t
||2 d yd x.
A(t) =
x

By our smoothness assumptions on , the function A(t) is continuously dif


ferentiable. Write x = ix where {ix } are the connected components of
x .
Lemma G.2. (Carlemans differential inequality). For x (0, b),
2
A(x).
(x)

A (x)

Proof. By the assumptions on ,





2
A (x) =
x dy +
x

x

(G.3)

2y dy

for x (0, b). Because = 0 on ix , Wirtingers inequality gives





2
2y dy
2 dy,
|ix |
ix
ix
and hence


x

2y dy

By Greens theorem

(x)

2
x


A(x) =

x

x dy,

and by the CauchySchwarz inequality


8

2
2
x A (x)
x

Thus


A 

2 dy.

A2
+
2
x dy

and that proves the lemma.

(x)

x

2 dy.

2
x

(G.4)

2 dy

2
A,
(x)


CY535/Garnett

0 521 47018 8

January 27, 2005

482

15:41

Appendices

The Dirichlet integral A(t) is connected to the harmonic measure (z 0 ) via


the function

(x) =
2 dy.
x

By Harnacks inequality
(z 0 )/3
on B(z 0 , r0 /2), so that
9(x0 )
.
r0

2 (z 0 )

(G.5)

On the other hand, because = 0 on ,



 (x) = 2
x dy.
x

By (G.4),  = 2 A and so Carlemans differential inequality (G.3) reads


2
 (x)

.
 (x)
(x)
Now set (x) = 2/(x) and

(x) =

exp
0


d dt,

so that

2
.
=


(x)
Then (G.3) can be rewritten as


 


log  =   0.

Therefore  /  is non-decreasing. Because  > 0, we obtain


 (x)  (t)  (t)  (x)
whenever 0 < x < t. Because (0) = (0) = 0, integrating the above inequality from 0 to x gives
(x)  (t)  (t)(x),
and integrating again from x to t then gives
(x)(t) (t)(x),

(G.6)

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

G. Carlemans Method

483

whenever 0 < x < t.


Increasing  in {Rez < x0 } increases but does not change the right side
of (G.1), so we may assume that (x) = 2/(x) = 2/M on x < x0 . Then
M 2 x0 /M
(e
1)
2

(x0 ) =
and



(b) = (x0 ) +

x0

exp
x0

2
(x0 ) 1 +
M

!
t
!
(s)ds exp
(s)ds dt

exp
x0

x0

(G.7)

(s)ds dt .

x0

Now (b) |b | M, so that by (G.5), (G.6), and (G.7)


9(z 0 )

r0

9M
1+

r0

(z 0 )2

9M (x0 )

r0
(b)
t
! 1
b
2
exp
(s)ds dt
M x0
x0

(G.8)

and taking the square root in (G.8) gives us (G.1).


To compare (G.1) and (IV.6.3), note that

b
x0


exp 2

t
x0

!
!
t
b
dx
dx
dt
dt
exp 2
(x)
b
x0 (x)
!
b
dx
,
exp 2
(x)
x0

so that

(z 0 , E b )

9M 2
2r0

 21


exp

b
x0

!
dx
.
(x)

(G.9)

If (x) = (x) inequality (G.9) is weaker than (IV.6.3) because the upper limit
is b rather than b. If (x) 2 > 0 for all x > (1 )b > 0, then we can
make a closer comparison with (IV.6.3), for in that case
!

b
b


dx
dt 1 eb ,
exp 2
x0
t (x)

CY535/Garnett

0 521 47018 8

January 27, 2005

484

15:41

Appendices

and thus

(z 0 , E b )

9M 2
2r0 (1 eb )

 21



exp

b
x0


dx
.
(x)

(G.10)

Of course, either (G.9) or (G.10) is enough to give Theorem V.1.1.

Notes
Tsuji [1959] proved (G.9) in polar coordinates. Our presentation follows
Carleman [1933] and Haliste [1965].

H. Extremal Distance in Finitely Connected Jordan


Domains
Several of the results in Chapter IV for simply connected domains have reformulations valid for finitely connected Jordan domains.
We begin with the description of the extremal distance d (E, F) in Theorem
IV.4.1. Assume  is a finitely connected Jordan domain, let E and F be finite
unions of closed arcs on , and assume E F = . In that case we call
(, E, F) a triple. The extension of Theorem IV.4.1 is easy when there is
one component  of  such that E F 1 , because in that case there
is a comformal mapping of  onto a rectangle with horizontal line segments
removed so that E and F correspond to the vertical ends of the rectangle if and
only if there is a subarc 1 such that E and F = . For the proof
we can assume  is bounded by analytic curves. Take two copies of  and form
the Riemann surface double d by attaching the copies along  \ (E F).
Let (z) = (z, Fd , d ) be the harmonic measure of the double Fd of F in
be the (locally defined) harmonic conjugate of in . Then just
d , and let 
as in the proof of Theorem IV.4.1, 
/s = 0 on each component  j of 
for which  j (E F) = . Therefore



ds = 0
 j s
and hence
= + i

is single valued and analytic on . The remainder of the argument is exactly


like the proof of Theorem IV.4.1. There is a similar finitely connected version

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

H. Extremal Distance in Finitely Connected Jordan Domains

485

of Theorem IV.4.2; see Exercise H.1.


There is also a generalization of Theorem IV.4.1 that holds for all triples
(, E, F), but now the slit rectangle must be replaced by a union of rectangles
pasted together along their edges to form a Riemann surface. We formulate the
result in terms of the real part of the conformal map to the Riemann surface.
For motivation, let us look back at the proof of Theorem IV.4.1 when  is
an analytic Jordan curve. The extremal metric is 0 (z) = |  (z)| = |(z)|,
where = Re and maps  onto a rectangle. By reflection, the function
is continuous on  and harmonic on a neighborhood of  \ P, where P
is the set of endpoints of the arcs in E F. Moreover, solves the mixed
DirichletNeumann problem:
(z) = 0,

(z) =

z ,

(H.1)

z E,
z F,

0,
1,

(H.2)

and
(z)
= 0,
z  \ (E F).
(H.3)
n
Theorem H.1. Suppose  is a finitely connected Jordan domain and suppose
E and F are disjoint closed subsets of , each consisting of a finite many
subarcs of . Then the extremal distance between E and F is given by

||2 d xd y,
d (E, F)1 = D()


where is a solution to (H.1) and (H.2). Extremal metrics exist and are given
by constant multiples of = || a.e. dxdy. If  consists of C 1+ curves
then is the unique solution to (H.1), (H.2), and (H.3). The conjugate extremal
distance satisfies
d  (E, F) = 1/d (E, F).
Theorem H.1 also implies Corollary C.4 of Appendix C.
Corollary H.2. Suppose  is a finitely connected Jordan domain and suppose
E and F are disjoint closed subsets of , each consisting of a finite many
subarcs of . Then is the unique function in C 1 () satisfying

"
||2 d xd y = inf
|v|2 d xd y : v C 1 (),


# (H.4)
lim sup v(z) 0, lim inf v(z) 1 .
z E

z F

CY535/Garnett

486

0 521 47018 8

January 27, 2005

15:41

Appendices

Indeed, for v in the right side of (H.4), |v| is an admissible metric for
the problem d (E, F) and L(, |v|) 1. Since L(, ||) = 1, and since
|| is the extremal metric, we must have A(, |v|) A(, ||), which
is (H.4). Theorem IV.2.1 implies is the unique minimizing function for (H.4).
This proof only needs v C() to be piecewise C 1 .
Proof of Theorem H.1. Before starting, we return to the conformal mapping
constructed in the proof of Theorem IV.4.1. As z traces , each excised
horizontal line segment on the boundary of () is traced twice by (z) and
each vertical line segment is traced once. The points in R which are endpoints
of the horizontal slits correspond to the points in  where = 0; these are
at the critical points.
the critical points of . Let {vi } denote the values of 
If we remove the curves {z : 
(z) = vi }, we partition  into finitely many
is an
smaller regions  j . The image of each  j under the map = + i
uncut rectangle of length 1. The shortest curves for the extremal distance on
the slit rectangle are the horizontal line segments not meeting the excised slits.
In other words, the shortest curves in  are those level sets {
(z) = c} whose
images do not meet the excised slits. In the general case the definition of is
the same. We can assume  consists of analytic Jordan curves. By Corollary
II.4.6, this assumption will not change (H.3) when  consist of C 1+ curves. If
 \ (E F) = , take two copies of  joined along  \ (E F), forming
the Riemann surface double d , and let Fd consist of those boundary curves of
d corresponding to F. If  \ (E F) = , we let d =  and Fd = F.
Let (z) be the harmonic measure of Fd in d . As before, we work with the
restriction of (z) to . From the proof of Theorem IV.4.1 it is clear that
the (locally defined) function = + i
is continuous up to  and that
= Re satisfies (H.1), (H.2), and (H.3). In fact, by the Schwarz reflection
principle,
is the only solution
to (H.1)(H.3). Moreover, except on the set


P = endpoints of E F , is analytic on , and (IV.4.2), (IV.4.3), and
(IV.4.4) hold.
In our more general case, 
is not single valued. For each  \ P there
is a conformal map f defined in a neighborhood U of with f ( ) = 0 and
a constant a so that
+ i
= a + f k

(H.5)

for some k 1. If k 2, we call a critical point of order k1. Near ,


the level set {z : 
(z) = 
( )} consists of 2k analytic arcs which meet at
separated by equal angles of /k, and these level lines do not depend on the
choice of 
that was made. Each  for which || = 0 then lies on a
maximal analytic curve along which + i
can be continued analytically with

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

H. Extremal Distance in Finitely Connected Jordan Domains

487


constant and parameterized so that is increasing. Call these curves level
curves of 
. We claim that each such level curve will either terminate at a
critical point of or at a point of E F. To see this, first note that if is a
level curve then cannot be closed since is continuous and increasing on .
If U has infinitely many components, then we can find an arc in U with
both endpoints on and on which is constant. A subarc of together with
then forms a closed curve, along which either is increasing or constant.
This contradicts the continuity of . By compactness, the level curve must
either terminate at a critical point or terminate at a point of . If terminates
at a point  \ {E F}, then since 
is constant on  U , the local
description (H.5) implies is a critical point of . This proves the claim.
We can also show that cannot terminate at a point P. To see this note
that
$
(H.6)
(z) = ( ) + a1 (z ) + . . . .
Indeed, since  is analytic, there is a neighborhood W of and a conformal
map of the unit disc D onto W so that (0) = and
({z D : Imz > 0}) = W .
On the quarter disc Q = {z D : Imz > 0, Rez > 0}, f (z) ((z 2 )) is analytic and on each segment in Q D, either Re f or Im f is constant. Thus
f extends by two reflections to be analytic in D with Re f 0 or Re f 1 on
a diameter. By the argument used to establish (IV.4.3), f  (0) = 0. Thus has
an analytic extension to W \ (E F), on which (H.6) holds. Moreover, near
the level set {z : 
(z) = 
( )} consists of one analytic arc terminating at ,
namely the subarc of  \ (E F) contained in W . In particular, a level curve
of 
in  cannot terminate at .
We will call the level lines {
= vi } passing through the critical points of
the critical level lines. The critical level lines partition  into smaller regions
 j . See Figure H.1. The boundary of  j consists of finitely many analytic arcs
j,k , on the relative interior of which either

= 0 and
=0
s
n

= vi }
if j,k {

(H.7)

or

= 0, and
= 0, if j,k E F.
(H.8)
s
n
Every endpoint of j,k is either a critical point or a point of E F. We will show
that = + i
is single valued on  j and that ( j ) is either a rectangle or

CY535/Garnett

0 521 47018 8

January 27, 2005

488

15:41

Appendices

an annulus. This will follow from the argument principle applied to the (single
valued) analytic function  = x i y .
E

1

2

R1

3

F 
4

R2

(E)

(F)

R3
R4

5

R5

F
0

Figure H.1 ( j ) = R j .
We shall need an extended version of the argument principle, and we digress briefly to prove it. Suppose W is a bounded, finitely connected Jordan
domain and suppose each boundary curve is piecewise continuously differentiable. Orient W so that it has index 1 about each point of W . We say W has
an interior angle of 0 [0, 2 ] at 0 if the argument of the tangent vector,
m
arg dz
ds , increases by 0 at 0 . Let F = {k }k=1 be a finite subset of W
and let W = W \ k {z : |z k | }. If g is continuous on W \ F then we
define the principal value integral as


PV
g(z)dz = lim
g(z)dz,
W

0 W W

provided the limit exists. For example, the total increase in the argument of the
tangent vector around W is given by




d
dz
P V Im
log
dz +
( k ) = 2(1 (N 1)), (H.9)
ds
W dz
where N is the number of components of W and the sum on the left side is
taken over the finite set of points where the interior angle k is not equal to . If
f is meromorphic on W , then we say f has a singularity at 0 W of order
0 R if
 

f (z)
0
(H.10)

lim
W z0
f (z)
z 0
exists. Thus if f (z) (z 0 )0 then f has a singularity of order 0 at 0 .
Lemma H.3 (extended argument principle). Suppose W is a bounded, finitely
connected Jordan domain and suppose each boundary curve in  is piece-

CY535/Garnett

0 521 47018 8

January 27, 2005

15:41

H. Extremal Distance in Finitely Connected Jordan Domains

489

wise continuously differentiable. Let P = {k }m


k=1 W and suppose W has
interior angle k at k , k = 1, . . . , m, and interior angle at all W \ P.
If f is meromorphic in a neighborhood of W \ P, having no zeros or poles on
W \ P, and if f has a singularity of order k at k , k = 1, . . . , m, then

m

f  (z) dz
k k
= (Nz N p ) +
,
(H.11)
P V Im
2
W f (z) 2
k=1

where Nz and N p respectively are the number of zeros and poles of f in W ,


counting multiplicity.
For example, if W is smooth and if f is meromorphic on W then

f  (z) dz
= (Nz N p ) + (Nzb N pb )/2,
P V Im
W f (z) 2
where Nzb and N pb are the zeros and poles of f on W , counting multiplicity.
Proof. By the usual argument principle, if > 0 is small,

f  (z)
dz = 2i(Nz N p ).
W f (z)

(H.12)

Set k = {z : |z k | = } W . Then by the definition of k



dz
lim Im
= k .

0
z

k
k
By (H.10)


lim Im

k k
f  (z) dz
=
.
f (z) 2
2

(H.13)

Summing (H.13) over k and subtracting the result from (H.12) proves the
lemma.

To apply this lemma on the region W =  j , recall that  is analytic and
non-zero on  j . Moreover on the relative interior of each analytic arc j,k in
d
d
 j , arg d
ds is constant by (H.7) and (H.8). Thus ds arg ds = 0 and



d
d
log
ds
0 = P V Im
ds
 j ds

d
log  (z) dz
= P V Im
(H.14)
 j dz



dz
d
log
dz.
+ P V Im
ds
 j dz

CY535/Garnett

0 521 47018 8

January 27, 2005

490

15:41

Appendices

Let 1 , . . . , m be the endpoints of the arcs j,k in  j , and let i and i be


the corresponding interior angles of  j and singularities of  . Then by (H.6),
(H.9), (H.11), and (H.14)
0=

m


i i + 2(2 N ) +

i=1

m


(i ),

i=1

and hence
2(N 2) =

m


i i + i .

(H.15)

i=1

There are three kinds of endpoints i of j,k :


(i) If i is an endpoint of E F then by (H.6)  has a singularity of order
that contains i is an arc of
i = 1/2. By (H.6) the only level curve of 
 \ (E F) and so  j has an interior angle i = at i . In this case,
then i i + i = /2.
(ii) If i (E F) , then  is analytic and non-zero in a neighborhood of i .
= c2 } to the other and
Thus  j turns from one of the sets {w = c1 } or {
 j has interior angle