0 521 47018 8
15:41
CY535/Garnett
0 521 47018 8
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
15:41
ii
CY535/Garnett
0 521 47018 8
15:41
iii
CY535/Garnett
0 521 47018 8
15:41
iv
CY535/Garnett
0 521 47018 8
15:41
Harmonic Measure
JOHN B. GARNETT
University of California, Los Angeles
DONALD E. MARSHALL
University of Washington
Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or thirdparty 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
15:41
vii
CY535/Garnett
0 521 47018 8
15:41
viii
CY535/Garnett
0 521 47018 8
15:41
Contents
Preface
page xiii
I.
1.
2.
3.
4.
5.
Jordan Domains
The HalfPlane 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.
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
15:41
Contents
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
157
157
159
162
166
173
184
193
194
200
200
203
205
207
212
216
221
221
229
229
232
241
246
253
257
258
CY535/Garnett
0 521 47018 8
15:41
Contents
VIII.
1.
2.
3.
4.
5.
6.
7.
xi
269
269
272
281
286
289
296
302
305
307
IX.
1.
2.
3.
315
315
324
331
341
342
X.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
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
15:41
Contents
H.
I.
J.
K.
L.
M.
484
497
503
512
518
520
Bibliography
Author Index
Symbol Index
Subject Index
531
555
559
561
CY535/Garnett
0 521 47018 8
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 halfplane
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
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
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 outofprint 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. PrezMarco, P. PoggiCorridini, S. Rohde,
I. UriarteTuero, 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.
John B. Garnett
Donald E. Marshall
CY535/Garnett
0 521 47018 8
15:41
xvi
CY535/Garnett
0 521 47018 8
15:41
I
Jordan Domains
a
b
Figure I.1 The harmonic function (z).
1
CY535/Garnett
0 521 47018 8
15:41
I. Jordan Domains
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)
CY535/Garnett
0 521 47018 8
15:41
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 halfplane 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
We claim that
lim sup u (z) f (t)
zt
(1.4)
CY535/Garnett
0 521 47018 8
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
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
CY535/Garnett
0 521 47018 8
15:41
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)
(1.7)
D zei0
u(z) = f (ei0 ).
(1.8)
1 1 z2
2 ei z2
CY535/Garnett
0 521 47018 8
15:41
I. Jordan Domains
1 z2
i
i0 d
( f (e ) f (e )) (z, I, D) .
u(z) u (z) =
i
2
2
I e z
CY535/Garnett
0 521 47018 8
15:41
z
0
( )
2 sec1
( ) 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).
( )
( ) z
u(z) = f ( )
(2.2)
(2.3)
CY535/Garnett
0 521 47018 8
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 ( ).
{ :  f ( ) > }  f 1 ,
so that by (2.3),
{ : W f ( ) > } { : u ( ) > /2} + { :  f ( ) > /2}
8 + 12
 f 1 .
(2.4)
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
15:41
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)
1 1 + z
,
  0 
= 2 1 z
Pz ( )
Pz ( )
Pz .
CY535/Garnett
0 521 47018 8
10
15:41
I. Jordan Domains
(2.6)
Pz ( )
0 0
Pz
by a step function.
(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 +
20  1 + z
(1 + 2).
2 1 z
CY535/Garnett
0 521 47018 8
15:41
11
By Lemma 2.2, the inequality (2.3) will follow from the simpler inequality
3 f
1
,
(2.8)
{ D : M f ( ) > }
In that case, if
(I1 )
1
(I j ),
2
j odd
we take for {Jk } the family of oddnumbered 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
15:41
I. Jordan Domains
 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.
lim
( ) z
u(z)
CY535/Garnett
0 521 47018 8
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 onetoone map from D onto .
Because maps D onto , the continuous extension (also denoted by ) must
map D onto = , and because is onetoone 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 w2 d
(3.1)
u(z) u f (z) =
f (ei ) i
e w2 2
0
is harmonic on , and by Theorem 3.1 and Theorem 1.3,
lim u(z) = f ( )
z
(3.2)
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
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.
L 2 ()
so that for < 1,
0
L 2 ()
d
 (z)2 ds,
DB(,)
 (z)2 d x d y
(3.5)
n
Un
(n )
CY535/Garnett
0 521 47018 8
15:41
3. Carathodorys Theorem
15
(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)
CY535/Garnett
0 521 47018 8
16
15:41
I. Jordan Domains
One can also prove is onetoone 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.
za
, a D,  = 1.
1 az
When T M, we have
T (z)
1
=
,
2
1 T (z)
1 z2
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 z2 ) 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 z2
CY535/Garnett
In general
0 521 47018 8
15:41
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)
r (a, d) =
and euclidean distance to D
dist(B, D) =
t (1 a2 )
1 t 2 a2
1t
(1 a).
1 + at
(4.4)
(4.5)
z1
0
a1
z2
x1
a2
CY535/Garnett
0 521 47018 8
18
15:41
I. Jordan Domains
(4.6)
(4.7)
Note that
(z) =
nn1 z n
n=1
(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))
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
15:41
19
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
nbn 2 1.
(4.12)
n=1
1
nbn 2 r 2n
A(r ) = 2
r
n=1
and
1
n=1
nbn 2 = lim
r 1
A(r )
0,
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 z2 ) dist((z), )  (z)(1 z2 ).
4
(4.13)
CY535/Garnett
20
0 521 47018 8
15:41
I. Jordan Domains
(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 lefthand inequality in (4.13).
To prove the righthand 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)
4 (z 0 )
4
dist(z 0 , 1 )
(4.14)
,
dist(w, )
1 z2
1 z2
CY535/Garnett
0 521 47018 8
15:41
21
(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 .
CY535/Garnett
0 521 47018 8
22
15:41
I. Jordan Domains
(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
2z 0 
.
(z )
2
1
z

1
z 0 2
0
0
(4.20)
zg (z)
Reg
=
z
r
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
15:41
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.
CY535/Garnett
0 521 47018 8
24
15:41
I. Jordan Domains
is simply connected.
k
j
k
k ()
Figure I.9 The map z = ( )
(5.2)
CY535/Garnett
0 521 47018 8
15:41
25
x x
,
D (( ), ( )) = (, ) k (, ) = i H (x, x ) =
x + x
,
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
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].
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
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 z2
and
=
Re
= Im
.
ei z
tz
ei z2
(x t)2 + y 2
CY535/Garnett
0 521 47018 8
15:41
27
u( p)
C p,q .
u(q)
CY535/Garnett
0 521 47018 8
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 z2
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
C(K ).
C(K )
v(z)
7.
8.
9.
10.
11.
CY535/Garnett
0 521 47018 8
15:41
29
CY535/Garnett
0 521 47018 8
30
15:41
I. Jordan Domains
.
2
1  f (z)
1 z2
(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)
1 + z
1 z
w+z
1 + zw
to get
(z)
4
,
(z)
1 z2
and integrate along [0, z].
2
.
CY535/Garnett
0 521 47018 8
15:41
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)
(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
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 .
aD
1 a2
d(z) < .
1 az2
T M
CY535/Garnett
0 521 47018 8
15:41
33
(E.3)
Hint: If r < 1 z or r 0.4 use a linear map and the HaymanWu
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
C2
2 dudv
.
1  1 (w) 3
w2
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
CY535/Garnett
0 521 47018 8
34
15:41
I. Jordan Domains
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 }
1
2
+
8
log
2
.
8
log
2
1 R
R r2
CY535/Garnett
0 521 47018 8
15:41
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
 p = p 2  p2  2 .
(b) If 0 < p <
1
2
then H p and
p
 H p
C
,
1 2p
p2
2r
w p2 d A(w).
({z<r })
w p2 d A(w) =
w<M(r )
p
p
M(r ) ,
r
 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
15:41
I. Jordan Domains
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
15:41
II
Finitely Connected Domains
CY535/Garnett
0 521 47018 8
38
15:41
(1.1)
a
b
CY535/Garnett
0 521 47018 8
15:41
39
(1.1)
D\ 1 ()
1 z2
d
u n (ei ) .
2
ei z2
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
15:41
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)
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
CY535/Garnett
0 521 47018 8
15:41
41
1
+ h(z, w).
z w
(2.1)
on ,
(2.2)
g(, w) = 0
on ,
(2.3)
(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
42
15:41
By definition, an analytic arc is the image ((1, 1)) of the open interval
(1, 1) under a map onetoone 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 )
CY535/Garnett
0 521 47018 8
15:41
43
(2.8)
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
44
15:41
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.
(2.10)
for all .
u=0
D
V
u>0
v=0
v>0
Figure II.3 Straightening an analytic arc in .
CY535/Garnett
0 521 47018 8
15:41
Set
u( 1 (w)),
v(w) = u( 1 (w)),
0,
45
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
+
with an = 0. But if n
2, then
(2.11)
Because of (2.12),
Pz ( ) =
is called the Poisson kernel for .
1 g(, z)
2 n
CY535/Garnett
0 521 47018 8
46
15:41
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
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
15:41
47
g(z, ) ds( )
.
n
2
(2.14)
c1 <
(2.15)
( ) 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
15:41
z0
(2.19)
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( ),
CY535/Garnett
0 521 47018 8
15:41
49
lim
( ) z
v f (z) = 0
( )
(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 weaktype 11, 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
15:41
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
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
15:41
3. Conjugate Functions
51
u (r e
e
i )
d
sec .
2
N
(an cos n + bn sin n )
n=0
CY535/Garnett
0 521 47018 8
52
15:41
N
n=0
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
.
C = f L (D) : sup
t
t>0
f ( + t) f ( )
.
t
(3.2)
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
15:41
3. Conjugate Functions
53
CauchyRiemann equations,
u(z) = F (z) =
eit f (t) dt
.
(eit z)2
(3.4)
(3.5)
4M
.
R
(3.6)
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
15:41
(e
so that
u(z)
=
 f (t) f (t0 ) dt
eit z2
+
tt0 <1z
1ztt0 <
t
it z2
2
e
(1
z)
tt0 <1z
0
2 f
(1 z)1 .
=
(1 + )
When 1 z t t0  < , the inequality t t0 2 ceit z2 ,
valid with c independent of z, gives
 f (t) f (t0 ) dt
dt
2c
f
t 2
it
2
e z
1z<tt0 <
1z
2c f
(1 z)1 .
(1 )
Therefore we obtain (3.8) with
2
C1 = sup
0<<1
(1 )
2(c + 1)
c+
=
.
(1 + )
CY535/Garnett
0 521 47018 8
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 
.
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 nonnegative 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 ) Cz 1 z 2 
for all z 1 , z 2 D.
Corollary 3.3. Assume k is a nonnegative 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
15:41
an ein ,
an z n .
n=1
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 ( ) Mt,
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 nonnegative 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
CY535/Garnett
0 521 47018 8
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, 1z
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 .
(3.12)
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
58
15:41
and
2u
2 (z) = ReF (z)
x
t1z
f Z t
c
dt +
3
(1 z)
1z<t
f Z
dt
t2
C(1 z)1 f Z .
2
Unfortunately this trick does not us help with ImF (z) = xuy . Instead,
+ 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
15:41
4. Boundary Smoothness
59
Because
2
u(r ei(+t) ) + u(r ei(t) ) 2u(r ei ) t2 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 lefthand 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
15:41
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 onetoone 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
15:41
4. Boundary Smoothness
Proof. Set
v(z) = arg
61
(z)
.
z ei
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
((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
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
15:41
h0
r <1
i
 (r e )d b a
 (r e ) d
b a B1 .
a
(4.5)
CY535/Garnett
0 521 47018 8
15:41
4. Boundary Smoothness
63
1
1
g (y1 ) g (y2 ) =
f (g(y1 ))
f (g(y2 ))
s
s ( )
arg
 (ei )
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 )) Cs(1 ) s(2 )
(4.7)
ds
ds
so that by (4.5)
d
(s(1 )) d (s(2 )) C1 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
15:41
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 nonnegative 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
nonnegative 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 nonnegative 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 nth 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
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)
d
< c2
ds
(4.9)
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
15:41
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.
0
z0
z0
0
z ,
CY535/Garnett
0 521 47018 8
15:41
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
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
15:41
n
Hint: If u f is the harmonic extension of f  to , lim (u f f )dn = 0
and thus any weakstar 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
CY535/Garnett
0 521 47018 8
15:41
69
components, and if = (z, E, ), show that for some choice of a > 0,
)
= ea(+i
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 ) = 2u2 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
15:41
2r
cos(
)
+
r
0 >
cot
f (ei )
d
.
2
()
dt.
f
t
t2
0
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
15:41
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.
CY535/Garnett
72
0 521 47018 8
15:41
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,
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
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 weakstar 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
15:41
.
as z .
(1.1)
b1
+ ...
z
(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
15:41
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)
(1.4)
(1.5)
(1.6)
(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
15:41
(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([, ]) =
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
15:41
77
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)
1
 z
CY535/Garnett
0 521 47018 8
78
15:41
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 ( ).
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
15:41
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 weakstar 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 weakstar 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 weakstar limit { E n } for a bounded domain
is given in Exercise 4.
(3.1)
CY535/Garnett
80
0 521 47018 8
15:41
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)
C
z
(3.4)
C
z2
(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
15:41
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)
K K (z) =
and
K (z w)K (w)dudv
K K L(w) =
(3.7)
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 (),
(3.8)
CY535/Garnett
0 521 47018 8
82
15:41
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,
f d = lim
1
f d = lim
0 2
f U d xd y = 0.
Therefore = 0.
C (C)
satisfy
(3.9)
and
Un 2 0.
Then
Un 2 d xd y 0,
(3.10)
L
L
L
2
Un
dy d x + 2C 2 /L
R y
y
4LUn 22 + 2C 2 /L ,
and a second integration gives (3.10) when L is large compared to R.
CY535/Garnett
0 521 47018 8
15:41
83
(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 weakstar 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
CY535/Garnett
0 521 47018 8
84
15:41
= 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 weakstar 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)
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,
(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
15:41
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 weakstar 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)
zC
&
z = C \ E,
& : U E (z) = (E)} = 0.
Cap {z E
(4.9)
(4.10)
(4.11)
CY535/Garnett
0 521 47018 8
86
15:41
U E d = (E)
E 
,
4
(4.12)
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
15:41
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)
(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 weakstar. 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 pointmass at z, such that z (n)
j (E), such that a j > 0
(n)
(n)
(n)
and a j = 1, and such that n weakstar. Pick any w j 1 (z j ),
and set
n =
Nn
j=1
a j(n) w(n) .
j
CY535/Garnett
0 521 47018 8
88
15:41
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
(K n1 ) + log 3
(K n ) 2
4
+
log 3
(K n1 ) + 2 log 3
=
.
2
2
Theorem 4.5 provides another proof of the Koebe onequarter 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
15:41
89
1
( 1z )
= z + b0 +
b1
+ ...
z
1
,
4a
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 weakstar 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
90
15:41
(5.2)
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 weakstar 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
15:41
91
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
15:41
(z, n B (0), n )
Un (z) + log M + n
log( 1 M) + n
log 1 + M + log M + n
(z, n B (0), n )
.
log 1 M + n
on E,
(E)
on n ,
f = (E n )
on z = R,
U E
CY535/Garnett
0 521 47018 8
15:41
6. Regular Points
93
on
nR .
R .
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)
z
CY535/Garnett
0 521 47018 8
94
15:41
{z : z  < }, satisfying
V (z) 0,
(6.2)
(6.3)
(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
(6.5)
CY535/Garnett
0 521 47018 8
15:41
6. Regular Points
95
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
96
15:41
(6.9)
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.
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
15:41
7. Wiener Series
97
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.
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) .
CY535/Garnett
0 521 47018 8
98
15:41
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
N j+1 1
u n (0)
n log 2
1.
W j (0) =
n
n
Nj
(7.5)
Nj
(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
15:41
7. Wiener Series
99
(7.7)
z
n
< .
n
(7.8)
n=N
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)
{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
2
N,
(7.9)
1
M
2 , so that when 2
2
.
N
CY535/Garnett
0 521 47018 8
100
15:41
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)
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
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)
n
k
> 0,
(E Ak )
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 )
CY535/Garnett
0 521 47018 8
102
15:41
(8.1)
(8.2)
u( ),
,
.
z
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
15:41
103
with respect to weakstar 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 ,
(8.4)
(8.5)
V ( )d(z, , ),
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
CY535/Garnett
0 521 47018 8
104
15:41
B(z 0 ,)
Letting 0 we obtain
sup u(z) u(z 0 ).
C \E
(9.1)
CY535/Garnett
0 521 47018 8
15:41
Proof. Let
105
1 z
g(z, ) = log
z
(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 lefthand 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)
1
2
g(z, )
( )
ds
n
(9.3)
g(z, )d ( ),
CY535/Garnett
0 521 47018 8
106
15:41
V (z) =
g(z,  )d .
( E)
z
E
E
z
0
E
E
E
E
CY535/Garnett
0 521 47018 8
15:41
107
(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 halfplane 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)
CY535/Garnett
0 521 47018 8
108
15:41
(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 righthand 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
15:41
and
z 2
4y
g(z, )
,
1=
z
z 2
if z  >
109
.
4
(9.9)
Q
,
1 + (E T (Q)) (T (Q))
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)
CY535/Garnett
0 521 47018 8
110
15:41
and if we can show that c A E for some constant c, then the lefthand 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)
Q
1 + (E T (Q)) (T (Q))
C
by (9.10) and (9.11).
To prove the righthand part of (9.7) let A E and set
=
(T (Q)) Q and = .
Q dyadic
(9.12)
CY535/Garnett
0 521 47018 8
15:41
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
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)
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
15:41
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 
1 p (z) dz
g ds
=
.
(iii) d =
n 2
k p(z) 2i
CY535/Garnett
0 521 47018 8
15:41
113
Using this result, part (a), and the boundedness of {n }, prove that for all
f C(),
lim
f dn exists.
n
CY535/Garnett
0 521 47018 8
114
15:41
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 nth iterate
p (n) (z) = p p . . . p(z)
and write
1
p (n) = p (n) .
and
\ E. It is known that the Julia set E is a nonempty, 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
CY535/Garnett
0 521 47018 8
15:41
115
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 weakstar 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
116
15:41
n=1
n
= ,
(G n \ )
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
15:41
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)
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
(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
CY535/Garnett
0 521 47018 8
118
15:41
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
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
15:41
119
(E.6)
(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
CY535/Garnett
0 521 47018 8
120
15:41
< r () <
3
4
so that
{z  = r ()} \ K 
< 6,
2r ()
and by a comparison,
Jj
j
Jj
j \
sup
z < 3r
2
g (z, 0)
CY535/Garnett
0 521 47018 8
15:41
121
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)
.
(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
15:41
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
(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
(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 ).
then
((0, 1)) =
1
0
 (x)d x K .
(E.11)
CY535/Garnett
0 521 47018 8
15:41
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)
CY535/Garnett
124
0 521 47018 8
15:41
x R.
2
(z , E , H).
3
(E.12)
CY535/Garnett
0 521 47018 8
15:41
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
126
15:41
then
S( f ) D(u) =
D
2 u
2
0
u2 d xd y =
2 r dr d =
n(an 2 + bn 2 ).
2 U
0
(r e ) d r
i
2 U
2
0
2
(r ei ) d.
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
15:41
Then
U (r e ) =
i
log
127
rn
1
d(t) =
(h n cos n + kn sin n ).
r ei eit
n
0
(g) If S( f ) , then
E = {ei : lim sup u(r ei ) > }
r 1
satisfies
Cap(E ) e .
2
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
128
15:41
Prove
(z 0 , E, ) e1L
2 (z
0 ,E)
CY535/Garnett
0 521 47018 8
15:41
IV
Extremal Distance
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
Area () = 1. See Exercise III.22 for an inequality about (1.2). However,
129
CY535/Garnett
130
0 521 47018 8
15:41
L(, )2
.
A(, )
(1.3)
CY535/Garnett
0 521 47018 8
15:41
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
E
L 2 (, )
.
A(R, )
h
(1.6)
CY535/Garnett
0 521 47018 8
132
15:41
(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
15:41
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)
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
CY535/Garnett
134
0 521 47018 8
15:41
1
21
.
2 A(, j )
Then
L 2 (, 3 )
L 2 (, 1 )
.
A(, 3 )
A(, 1 )
(2.1)
(3.1)
(3.2)
CY535/Garnett
0 521 47018 8
15:41
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
CY535/Garnett
0 521 47018 8
136
15:41
1
2
+
.
()
(1 ) (2 )
(3.3)
E2
E1
1
G2
G1
Figure IV.5 Parallel rule.
2
CY535/Garnett
0 521 47018 8
15:41
137
+
.
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 )
T 1 ( )
1 dz.
.
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
138
15:41
R that
connecting E 1 to F1 , then reflecting T () yields a curve
joins E to F and
dz =
dz.
E
R
T (E)
T ()
T (F)
CY535/Garnett
0 521 47018 8
15:41
139
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
15:41
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
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.
CY535/Garnett
0 521 47018 8
15:41
141
=
=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 halfplane 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
142
15:41
1
1
(E, F) = 2 log R,
d
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
15:41
143
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 nonzero constant
c so that
ds = 2.
c
s
e 2 L (0, E L , R L )
8 L
e 2 .
(5.1)
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
CY535/Garnett
0 521 47018 8
144
15:41
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)
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)
,
t tan1 (t) min t,
4
4
valid for 0 t 1, and both equalities in (5.2) are immediate from (5.4).
z0
CY535/Garnett
0 521 47018 8
15:41
145
8 (z 0 ,E)
e
.
Moreover
lim (z 0 , E, )e(z 0 ,E) =
and
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
E1
0 E2
1
E
2
E
CY535/Garnett
0 521 47018 8
146
15:41
8 dD (,E)
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) .
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
6. The
dx
(x)
15:41
Estimate
147
(x)
y = m(x)
x
Figure IV.12 A strip domain.
n
d( j1 , j ).
j=1
1
,
(x)
z = x + i y.
(x)
a
CY535/Garnett
0 521 47018 8
148
15:41
Moreover
2 d yd x =
a
so that
d (0 , n )
1
d x,
(x)
1
d x.
(x)
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
15:41
Notes
149
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 \
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
15:41
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.
CY535/Garnett
0 521 47018 8
15:41
151
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
15:41
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 + z2 )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
15:41
153
(ii) If h is real valued and Borel measurable in and if hdz 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
154
15:41
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.
+
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
15:41
155
yn 1/yn
y0 =
Prove
(0, E L , R L )  Ce L2
n1
Ln.
n=1
CY535/Garnett
0 521 47018 8
156
15:41
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 + .
CY535/Garnett
0 521 47018 8
15:41
V
Applications and Reverse Inequalities
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
157
CY535/Garnett
158
0 521 47018 8
15:41
lim inf
r
logM(r )
r n/2
> 0.
(1.1)
R
dr
8
logM(R).
log  f (z j ) exp
R0 r j (r )
j (r ) = 2,
j=1
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
15:41
2. Lower Bounds
Hence
159
n
R
1 R dr
n
log
,
n
2
R0
R0 r j (r )
j=1
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
,
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
15:41
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)
C() = exp 2
x0 +
x0
(2.3)
(x)d x
B (x, y)
B2 (x, y) + 2
(x, y) =
if x > x0 + ,
if x0 < x < x0 + ,
if x0 < x < x0 ,
elsewhere in ,
CY535/Garnett
0 521 47018 8
15:41
2. Lower Bounds
161
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
m (x)2
+ (x)2
(x)
d x = A < ,
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
udz 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
15:41
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
1
1
1
log =
log ,
2
2
a
CY535/Garnett
0 521 47018 8
15:41
163
We conclude that
(0, E) = lim d ( B , E) d ( B , )
0
= (0, (E)),
and (z 0 , E) is a conformal invariant.
CY535/Garnett
164
0 521 47018 8
15:41
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
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
15:41
165

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
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
1
.
dist (0, z)
CY535/Garnett
0 521 47018 8
166
15:41
 
s
and
2 (z)d xd y
\B
<w<R
1
R
d A(w) = 2 log ,
w2
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
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
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)
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
15:41
As in (3.1),
M +
log
log
M +
,
2
2
(4.4)
log
1
=
log  (0).
2
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)
n=1
an z n
CY535/Garnett
0 521 47018 8
15:41
4. Teichmllers Modulsatz
169
(z) =
b1
+
bn z n
z
n=0
(4.7)
nan 2 = Area(G 1 ) b1 2
nbn 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 z2
nan 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)
b
1
1
1
(z)
+ b0
nbn 2 log
C1 b1 2 log .
z
1 r2
n=1
CY535/Garnett
0 521 47018 8
170
15:41
(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 ,
(4.10)
C2
1
log R +
,
2
log 1
(4.11)
C
2
1
log R +
.
2
log R log log R + log 1
2
1
R2
2
(4.12)
CY535/Garnett
0 521 47018 8
15:41
4. Teichmllers Modulsatz
171
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
15:41
because
(4.15)
(4.16)
By (4.10) we have
0
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
15:41
173
(/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.
(5.1)
CY535/Garnett
0 521 47018 8
174
15:41
( )
( ) z
f (z) = A,
lim
F(z)
arg
F(z) F( )
z
( ) z
lim
( ) z
(5.2)
CY535/Garnett
0 521 47018 8
15:41
175
lim
( ) z
F(z)
and
F ( )
lim
( ) z
F(z) F( )
z
(5.3)
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 .
CY535/Garnett
176
0 521 47018 8
15:41
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 halfplane H onto a domain such that
0 and we assume has nontangential limit
lim
(0) z0
(z) = 0.
(5.4)
(5.5)
defined by
is a conformal map of S onto the region
= {i/2 log w : w }
or
}.
= {iez : z
(0, /2)
CY535/Garnett
0 521 47018 8
15:41
177
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 halfplane version to the strip version.
i 2
R x0
dist(x, )
= 0.
x
(5.7)
CY535/Garnett
178
0 521 47018 8
15:41
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
zEr
r 0
(5.8)
CY535/Garnett
0 521 47018 8
15:41
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+
U
FRe
1 (U )
1
1
log .
2
(5.12)
(5.13)
(5.14)
CY535/Garnett
180
0 521 47018 8
15:41
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)
CY535/Garnett
0 521 47018 8
15:41
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
Rez+
zS
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)
CY535/Garnett
0 521 47018 8
182
15:41
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
(5.19)
whereas
{z : s + < Rez < t , Imz < /2}
( 1 (Fs ) 1 (Ft )) = .
(5.20)
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
Rew+
wS
Rew Re 1 (w)
CY535/Garnett
0 521 47018 8
15:41
183
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
(5.23)
CY535/Garnett
184
0 521 47018 8
15:41
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) =
s,t+ s
1
m (x)2 + 12
(x)2
d x = 0.
(x)
(5.25)
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
15:41
185
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
CY535/Garnett
0 521 47018 8
186
15:41
%
{T : T T M }.
(6.2)
M is the largest domain such that R
M
and
Then
M consists of
two MLipschitz graphs.
\
M
S
M
(6.3)
CY535/Garnett
0 521 47018 8
15:41
187
M \ S) < , then
M ) < and Area(
Proof. If Area(S \
lim dist(z, S) = 0
Rez+
M
z
(6.4)
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
15:41
4
j=1 U j .
S
U1
Fs
U1
U2
U
U4
Ft
U4
U3
Figure V.10 Proof of Sastrys lemma.
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
u2 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
15:41
Thus,
189
u2 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
Thus
d
(Fs , Ft )
u
2 d yd x
(inf udz)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
15:41
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
x Rez lj
Rez rj Rez lj
y /2
M(Rez rj Rez lj )
z j
Imz d x
2
CY535/Garnett
0 521 47018 8
15:41
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)
2
Fs
1
Ft
3
k
CY535/Garnett
0 521 47018 8
192
15:41
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 .
(t s )2
(t s) + (1 + 8
M )Area( M \ S)s,t (1
8
M )Area(S \ M )s,t
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
15:41
193
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 threedimensional 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
15:41
geometric conditions equivalent to the existence of a nonzero 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 halfplane 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 RodinWarschawski [1986] and the other half was
proved by Sastry [1995]. For other proofs of the halfplane 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].
sup
D j {z=R}
 f (z).
R j (R)
R0
1
< 2.
j (R)
(E.1)
CY535/Garnett
0 521 47018 8
15:41
195
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
Using
(z) =
1
2 log z
Aj
gives
L 2j
R 2
log
log R log
Aj
z j 
Aj
R0
(E.4)
1
2
1
1
1
1
V (z) + log
log
.
1 + z
2
z
1 z
where = {z D : V (z) > (log )/2}.
(c) Set
u =
V + 21 log 1
E + 21 log 1
CY535/Garnett
0 521 47018 8
196
15:41
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
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
3
L
+ O(5 ).
27 3
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
15:41
197
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 )
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.
CY535/Garnett
198
0 521 47018 8
15:41
(E.6)
1
1
=
.
2
H
Re d A
CY535/Garnett
0 521 47018 8
15:41
Show that
Re
2d A
199
d (E, F) 1 + C
2
log 1
CY535/Garnett
0 521 47018 8
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 nonzero
angular derivative at with the existence of a cone at ( ).
(1.1)
CY535/Garnett
0 521 47018 8
15:41
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
n
(zei j ) (zei j1 ).
j=1
Thus
r 1 E
CY535/Garnett
0 521 47018 8
202
15:41
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.
(1.4)
CY535/Garnett
0 521 47018 8
15:41
203
h
2 sec1 ()
CY535/Garnett
0 521 47018 8
204
15:41
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
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
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 nonconstant 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 nonzero 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 ( ) .
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
15:41
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
15:41
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).
"
B = D : has a nontangential limit at , and
#
lim inf  (z) = 0, for all > 1 .
(4.1)
( ) z
CY535/Garnett
208
0 521 47018 8
15:41
is conformal 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)
CY535/Garnett
0 521 47018 8
15:41
209
w
U
Tn (w)
Ln
Figure VI.3 nLipschitz 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
CY535/Garnett
0 521 47018 8
210
15:41
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
U \E
(4.3)
U \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
15:41
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 onetoone 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.
CY535/Garnett
0 521 47018 8
212
15:41
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
CY535/Garnett
0 521 47018 8
15:41
213
C log 1 ()
,
1 +  log 1 (E)
(5.1)
Then is singular to h , h .
CY535/Garnett
214
0 521 47018 8
15:41
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
CY535/Garnett
0 521 47018 8
15:41
215
{z k }, so that
lim
( ) z
u(z) c
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
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 z2 < n ,
(5.4)
CY535/Garnett
216
0 521 47018 8
15:41
< 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.
C
h 2 (z n, j )(1 z n, j 2 ) n .
2
j
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
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
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 .
(6.1)
(S (G)) = 1.
(6.2)
and
(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
CY535/Garnett
0 521 47018 8
218
15:41
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
CY535/Garnett
0 521 47018 8
15:41
6. Pommerenkes Extension
219
Because
have
1
dt
8
.
j (B(w, r )) exp
r t j (t)
1
2
Therefore
1
64
2dt
64
1 (B(w, r ))2 (B(w, r )) 2 exp
= 2 r 2.
t
(6.4)
rk j (S),
8 N
and hence
1 (S) j (S).
4 N
(6.5)
CY535/Garnett
220
0 521 47018 8
15:41
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,
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
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 nonzero 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
15:41
221
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.
CY535/Garnett
0 521 47018 8
222
15:41
C log 1 ()
.
1 + log 1 (E)
 ds and
= 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
CY535/Garnett
0 521 47018 8
15:41
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 \
CY535/Garnett
0 521 47018 8
224
15:41
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
J n, j [0, bn+1 ] .
n+1 = n \
j
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 nonconstant 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
15:41
225
z0
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
CY535/Garnett
226
0 521 47018 8
15:41
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,
CY535/Garnett
0 521 47018 8
15:41
227
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,
(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
CY535/Garnett
228
0 521 47018 8
15:41
A(a, r ) =
L(a, s)ds.
0
A(a, r )
1
.
2
r 2
1
L(a, r )
.
2r
2
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
CY535/Garnett
0 521 47018 8
15:41
VII
Bloch Functions and Quasicircles
1. Bloch Functions
An analytic function g(z) on D is a Bloch function if
gB = sup g (z)(1 z2 ) < .
zD
(1.1)
We write B for the space of Bloch functions and we call gB 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
gB = sup (g T ) (0),
T M
(1.3)
z+a
,
1 + az
(1.4)
z2
n=1
1 Actually g is only a seminorm, but g = 0 only if g is constant.
B
B
229
(1.5)
CY535/Garnett
0 521 47018 8
230
15:41
sup
Hence (1.1) yields:
gB = sup
Proof. If
g(z 1 ) g(z 2 ) B
z2
z1
1
dz,
1 z2
(1.6)
CY535/Garnett
0 521 47018 8
15:41
1. Bloch Functions
231
(1.8)
2
1+r
gB (1 r ).
= gB log
(1.9)
ei(+t)
(1 t)ei(+t)
ei
( )
Figure VII.1
(1.10)
CY535/Garnett
0 521 47018 8
232
15:41
Set r = 1 t. Then
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 righthand 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.
g(z) = log (z) , (0) = 0,
or equivalently
(z) =
(2.1)
e g(w) dw.
CY535/Garnett
0 521 47018 8
15:41
w+z
1+zw
(z)
(z)(1 z2 )
233
(z)
(1 z2 ) 2z = g (z)(1 z2 ) 2z,
(z)
it follows that
g (z)(1 z2 ) 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
(2.3)
I
zI
Figure VII.2
CY535/Garnett
0 521 47018 8
234
15:41
g(z)g(z )
I 1 dz
( ) (z I ) (z I )( z I )= (z I )
e
= (z I )
zI
T 1 ( )
ez
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 dw
1 w
0
2
(z I )I ,
2
The inequality e z
2
1 z I 
/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 onetoone map from
D onto a Jordan curve, and it then follows by the argument principle that is
univalent on D.
CY535/Garnett
0 521 47018 8
15:41
235
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
15:41
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 z2 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
CY535/Garnett
0 521 47018 8
15:41
D (w)
w
w0
237
w
w
0
D (w)
w
w0 w1
L (w)
w
w
0
L (w)
sup
n {z:(z,z )M}
n
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)
.
(2.9)
(z n )
1 + z
CY535/Garnett
0 521 47018 8
238
15:41
+ 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)
(2.11)
CY535/Garnett
0 521 47018 8
15:41
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 nonzero 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
Pz (e ) log  (e )
Pz (ei )ds ( ),
2
D
D
where the singular measure s is the weakstar limit
d
d
s = lim log  (ei )
log  (r ei )
.
r 1
2
2
Because H 1 ,
log  (z)
d
2
CY535/Garnett
0 521 47018 8
240
15:41
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 weakstar, 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
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 ) CI .
(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 4adic 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
Kquasiconformal 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
15:41
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
(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
1quasiconformal 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
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 largescale 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
15:41
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 )
CY535/Garnett
0 521 47018 8
15:41
3. Quasicircles
245
CY535/Garnett
246
0 521 47018 8
15:41
(4.1)
(4.2)
CY535/Garnett
0 521 47018 8
15:41
247
.
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 chordarc 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
15:41
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
 
+
1
I 
 
q
q1
q
q1
/q
< ,
CY535/Garnett
0 521 47018 8
15:41
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 chordarc 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)
(4.5)
CY535/Garnett
250
0 521 47018 8
15:41
 
k
.
k+1 
(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)
z/A
(zei ) (zei )
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
15:41
251
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 chordarc 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 chordarc 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 chordarc curve if and only if the following three conditions
all hold:
(i) is a quasicircle,
CY535/Garnett
252
0 521 47018 8
15:41
r
.
C
1
 d C exp
log  d
I  I
I
d
.
log  (ei )Pz Q ( )
C exp
2
D
 d C (z Q )
I
d
,
2
2Cr
,
I 
(4.8)
CY535/Garnett
0 521 47018 8
15:41
5. BMO Domains
and is a chordarc curve.
253
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
CY535/Garnett
0 521 47018 8
254
15:41
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
ebgg I  d < .
I

I
I
Then ebg A2 and ebg d and d are A equivalent by Proposition 4.1.
CY535/Garnett
0 521 47018 8
15:41
5. BMO Domains
255
z0
U
CY535/Garnett
256
0 521 47018 8
15:41
,
F 1 (z)
and Ahlfors [1963] shows that G can be chosen to be bilipschitz.
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 chordarc boundary,
has a nonzero 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
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 gB 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. Chordarc 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 chordarc} 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 chordarc curves and BMO domains.
Theorem 5.3 is from Bishop and Jones [1994].
CY535/Garnett
0 521 47018 8
258
15:41
z1
1
+
z2
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
15:41
259
while by (b)
( f 1 ())
.
4
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) Cz 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
Mcondition (3.5).
(c) Let H : C C be the quasiconformal extension of h guaranteed by (b)
CY535/Garnett
260
0 521 47018 8
15:41
1 (z) =
if z 1;
1 (z),
2 (H (z)), if z > 1.
CY535/Garnett
0 521 47018 8
15:41
261
(E.1)
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) Cz w.
CY535/Garnett
262
0 521 47018 8
15:41
1 K 2
1 K 1 K 2 1 K 1 K 2 2 .
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
15:41
263
whenever is a crosscut of ,
min diam( j ) C diam( ).
j
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
15:41
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
y1x
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
15:41
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  = 2I 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
15:41
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
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
.
A for = 1+
C
1+
I ,
dx
k
kd
x
I
I
I kd x
CY535/Garnett
0 521 47018 8
15:41
267
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 halfplane 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
15:41
such that
lim sup
0 I <
1
I 
g g I ds = 0,
I
sup
0 w1 w2 <
(w1 , w2 )
= 1,
w1 w2 
CY535/Garnett
0 521 47018 8
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.
lim sup
r 1
log
1
1r
1
log log log 1r
CgB .
(1.1)
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) =
15:41
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 gB 1 and if there exist > 0 and M < such that for
all z 0 D,
(1.4)
1 z2 g (z) ,
sup
{z:(z,z 0 )<M}
lim sup
r 1
log
1
1r
1
log log log 1r
c(, M) > 0.
(1.5)
(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
15:41
p!
log
,
1r
1 r2
271
(1.7)
valid for gB = 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
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
15:41
272
satisfies E p 
of gs ,
1
. Therefore if
p2
g(r ei )

1
1
log 1r
log log log 1r
p> p0
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.
t0
h(t)
= .
t
(2.1)
CY535/Garnett
0 521 47018 8
15:41
(1 r 2 ) (r )
> 0 a.e. on D,
h 1 (1 r )
(1 r 2 ) (r )
= 0 a.e. on D, and
h 1 (1 r )
r 1
r 1
273
(1 r 2 ) (r )
< a.e. on D.
h 1 (1 r )
C log
1
t
1
t
(2.2)
c log
1
t
1
t
(2.3)
(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,
gB 6, and by Theorem 1.1 there is a constant C > 0 such that
lim sup
r 1
Reg(r )
1
C
1
C log
1
t
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
15:41
Reg(r )
c3 > 0,
1
1
log 1r log log log 1r
c log
1
t
1
t
Then
lim inf
r 1
(1 r ) (r )
h 1
2 (1 r )
=0
CY535/Garnett
0 521 47018 8
15:41
275
(1 r ) (r )
A
h 1 (1 r )
(1 r ) (r )
>B>0
h 1 (1 r )
(2.6)
B
1
(1 + B) 2
Eq > 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
15:41
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
I (z k ) 4 AE.
(2.7)
C
,
diam( )
CY535/Garnett
0 521 47018 8
15:41
277
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
D
D
1,2
2
1,1
U1
1
w0
CY535/Garnett
278
0 521 47018 8
15:41
8 ( j )
,
e
(2.9)
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
#
"
8
N = # j : j r k0 .
(2.10)
CY535/Garnett
0 521 47018 8
15:41
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
2/3
Ct h(t) ct exp log(1/t)
.
(2.11)
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 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
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 Lfunctions,
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
15:41
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
inf
1
zn ( )
1
(1 z) (z)
(cB + )} E
1
n
h (1 z)
such that j
and
2
1
log( ),
log 2
r
()
(D \
N%
()
()
j ) r
h(r )
.
B
(2.15)
()
()
h 1 (( j ))
2c2 r
.
B
CY535/Garnett
0 521 47018 8
15:41
()
( j ) =
j=1
Because N ()
2
log 2
t j s
h(t j ) =
h(t)
t
h(t j ) +
h(t j ).
t j >s
t
2
2
log(1/t)h
C4 h
.
log 2
log(1/t)
B
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)
(1 + B) 2
E = ((E)) C1
h ((E)),
B
which is assertion (b).
(3.1)
CY535/Garnett
0 521 47018 8
282
15:41
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 ,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 .
(3.3)
1
2
1
2
log .
log 2
() = ( B(, 2))
2 + 2 +2
2
log 2
log
2 +2 .
CY535/Garnett
0 521 47018 8
15:41
283
j=1
Hence
n1
1
1
(n 1)
1
1
log +
+ 2 log ,
Xj 1 +
2
j=1
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 ( )
CY535/Garnett
284
0 521 47018 8
15:41
3
(log 1 )2
k
2
(3.5)
(3.6)
and
n1
X j ( p ) .
(3.7)
j=1
Also let
N (n, , ) = sup N (, n, , ).
{,D}
(3.8)
CY535/Garnett
0 521 47018 8
15:41
285
In this case
n1
X j ( p ) 2 .
j=2
1
2
2 )
Hence
N (, n, ) c
1
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 ,
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 )
CY535/Garnett
286
0 521 47018 8
15:41
< 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
z
(1z)2
3t 1,
k (t) = 0,
t 1,
if t 13 ,
if 1 t < 13 ,
if t < 1.
(4.4)
CY535/Garnett
0 521 47018 8
15:41
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 + 7t3
(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
15:41
are increasing in r ,
 (r ei )2 p d
1
1r
 (z)2 p d xd y
{r <z<1}
(4.8)
(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
15:41
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
2 + C A 1 2 1 +
n t (1 n ) 2 n
C
t
t + K /2
k=0
t1
1
1r
(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
15:41
and
( j )
1
log / , j = 1, 2.
2
Then
0 ( ) ( )
1
+ ( 1 )
log /
2
1
2
log /
+ ( )
2
( ) ( ) + O( 2 + 2 ),
(5.1)
lim ( ) ( )
0
(5.2)
Then
0 (, 1 , 2 ) 1.
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
15:41
291
(z ).
(5.3)
Let x R and assume that (x) = a, (x) = 0, and (x) exists. Then
(, a, ) =
2
 (x)
(5.4)
2
 (x)
2
while
( )
1
1
log 2 ,
2
a2
a1
a5
a4
a7
a6
Figure VIII.5 Polygonal tree.
2
 (x
j )
CY535/Garnett
0 521 47018 8
292
15:41
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
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
15:41
293
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
(, 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
1 z j  1 ( j ) (, j , ) c
2 +
log 1
(5.7)
CY535/Garnett
0 521 47018 8
294
15:41
Therefore
(B j ) B( j , c),
(5.8)
c
,
1 z j 
(5.9)
}.
2
(5.10)
2+q
q
 
d xd y
Bj
cr j2  (z j )q
rj
q
1c 2 z1 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 lefthand side of (5.12) increases with R and 1 1+R
2 is comparable to
1
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
15:41
295
(1) = 1e . Then
1 (z n )
n (z) = 1 (z ) z
zn
n
1
n
1
n
(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 )
3
2 },
CY535/Garnett
0 521 47018 8
296
15:41
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.
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 )
CY535/Garnett
0 521 47018 8
15:41
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 .
J = (n , a J , ),
nJ = n
%
I =n
I = J
B(a I , 2 n ),
CY535/Garnett
298
0 521 47018 8
15:41
and
J = ( j1 , j2 , . . . , jn1 ).
By the extension rule
2J (nJ , a J , ) J ,
(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 multiindex
J of length n,
2
2J (1 )2
(6.2)
jn
J
2
and
2J (1 )n j1 j2 . . . jn .
J
(6.3)
) 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
J
\ B(a J , 2 n1 ).
nJ \ B(a J , 2 n ) n1
3
Set = {z : z a J  = n 2
CY535/Garnett
0 521 47018 8
15:41
299
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
= 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
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
15:41
(6.4)
.
(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 )
D \ VJ
CY535/Garnett
0 521 47018 8
15:41
301
(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)
2J
C n
1
2
(6.10)
 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
15:41
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
2J p+
J =n
N
0
p+ 1n
C (1 ) p+
j
.
j=1
By (6.4)
(1 ) p+
N
p+
>1
j=1
(6.13)
1
Because 0 (1 r )1+ dr < for > 0, (6.13) implies B( p) > p 1 and
(c) (a).
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
15:41
303
T = {(1 + 2 2 )ei : 0 } {i y : y 1 + 2 2 }
and its truncation
1
T0 = {(1 + 2 2 )ei : 0 }
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
15:41
Theorem 7.1. If is sufficiently small there exists p, 0 < p < 2 such that
 f  p d x = .
(7.2)
L
1
$
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
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
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
1 2 (1 )A > 1
CY535/Garnett
0 521 47018 8
15:41
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
CY535/Garnett
306
0 521 47018 8
15:41
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
15:41
307
k=0
z 2 , z D and Sn ( ) =
n
k=0
2k ,
D. Prove
Sn ( )
lim sup $
=1
n log log n
(1.3)
lim sup
r 1
log
1
1r
1
log log log 1r
= 1.
z 15
n=1
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
15:41
pt
>t
( p) + p + 1 t
and
=
ps
<s<1
p+s
Mr t ( 1 (D \
N
%
j )) < Ct .
j=1
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 CI j  p1 (r ei ) p d CI j (( p) p1)
for any > 0. Together, these inequalities yield
I j t C,
CY535/Garnett
0 521 47018 8
15:41
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 closetoconvex if there exists
starlike such that
z (z)
Re
>0
(E.1)
(z)
on D, and is closetoconvex 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 closetoconvex 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 closetoconvex and E D is a Borel set, then
dim(E)
dimE
.
2 dimE
.
dim(D)
2
See Makarov [1987] for (d) and (e).
CY535/Garnett
0 521 47018 8
310
15:41
4 dist (z),
 (z)
= o
2
1 z
1 z
and
1
 (z)
.
1 z
= O
1 2
where
N (a) = #{Ik :  (z k ) a },
and
aj =
j
log 1
CY535/Garnett
0 521 47018 8
15:41
311
Because the sum (E.2) has O log 1 terms, it compares to its largest term
and we obtain
log N (a j )
log N (, , )
log 1
where
N (, , ) = #{disjoint B = B(, ), , + (B) }.
Then for =
1
1a ,
f ()
log N (a)
log 1
(E.4)
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
15:41
centers z k from (c). For 0 < b 2 and 1 < a < 1 be of the form
define
j
log
and
 (r ei )t d
1
(a+2b)t1
(a,b)
N (a, b).
(E.7)
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)
CY535/Garnett
0 521 47018 8
15:41
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)
CY535/Garnett
314
0 521 47018 8
15:41
lim sup
wz
G(z) G(w)
M < ,
z w
2K
K 1 ,
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
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)
315
CY535/Garnett
0 521 47018 8
316
15:41
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
(1.3)
(1.4)
but
CY535/Garnett
0 521 47018 8
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
(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)
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
15:41
(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, )
n
n
(, K 1,1,...,1,1
, ) (1 + )(, K 1,1,...,1,4
, ).
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 halfplane 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)
CY535/Garnett
0 521 47018 8
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
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)
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
15:41
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)
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
15:41
1. Cantor Sets
321
u(z)
v(z)
= 1. Then on j(1) ,
(1)
E1 = j
(1)
j(2) .
con
j(2) ,
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
15:41
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 ).
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
15:41
1. Cantor Sets
323
(K Jn+N
)
4 N +1 N
N +1
4
n
(K J )
4
+
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
"
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
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
15:41
Therefore (F) = 0.
(2.1)
(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
15:41
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)
(2.8)
(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
15:41
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
) = .
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 .
%
%
(Q n,k ) (Q m, j ) .
BM
GM
= g
).
(E) = (, E,
Write G(z)
(z, ) and
(2.12)
CY535/Garnett
0 521 47018 8
15:41
327
(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
Q n,k
2 c
G
( Q n,k ) cM,
n
( Q n,k )
(2.15)
(2.17)
CY535/Garnett
0 521 47018 8
328
15:41
(a) If Q n, j U N then
c
inf G(z)
(Q n, j ).
(2.18)
Q n, j
1 inf G.
sup G
2 Q n, j
Q n+N ,k
(2.19)
(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
15:41
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
GM
(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
330
15:41
and on B M
(Q n,k )
1
c
4
(Q n,k )
by (2.13). Therefore
(Q n,k )
B M
1
.
2c
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
15:41
331
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
CY535/Garnett
0 521 47018 8
332
15:41
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
15:41
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 
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
)
(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
R
2 },
let
CY535/Garnett
0 521 47018 8
334
15:41
C2
(z, B, U \ B),
g(z, ) log 2 +
R
and on U \ E
g E (z, ) = g(z, )
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
15:41
335
N
j=1 j
where
j {g(z) = c j }
is a Jordan curve. Then
I () =
1
2
g
log gds 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 gds +
I () =
2 j n
k
j
cj
=
log gds + ,
g(k ) +
2 j n
k
1
log gdz = j j = 1
2 j n
by Greens theorem. Therefore
I () =
k
g(k )
j
cj + .
CY535/Garnett
0 521 47018 8
336
15:41
and
j
cj =
0
{cc j } (c)dc
c0
n(c)dc,
0
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
15:41
337
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
15:41
(3.7)
%
S
2R Q k
Qj.
(3.8)
CY535/Garnett
0 521 47018 8
15:41
339
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) .
(B)
1
C M 2 log
.
(Q)
(3.9)
C log log
CY535/Garnett
0 521 47018 8
340
15:41
C 1+
(Q k )
C R 1+
.
M
(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
log gds
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
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 selfsimilar 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
SC () =
sup
simply connected
f ().
(N.3)
CY535/Garnett
342
0 521 47018 8
15:41
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 nth 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)
CY535/Garnett
0 521 47018 8
15:41
343
(E.1)
(E.2)
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
CY535/Garnett
344
0 521 47018 8
15:41
 (z)dz
, (z) = w
1 z2
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
15:41
345
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 Cquasidisc. 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
1
.
N2
CY535/Garnett
0 521 47018 8
346
15:41
CY535/Garnett
0 521 47018 8
15:41
X
Rectifiability and Quadratic Expressions
347
CY535/Garnett
0 521 47018 8
348
15:41
( ) = 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
cAF p F F(0) H p c1 AF p .
p
(1.1)
2
AF( ) ds =
F (z)2 { : z ( )}d xd y,
D
AF( ) ds c()
F (z)2 (1 z2 )d xd y
c2 ()
(1.2)
2
AF( ) ds.
CY535/Garnett
0 521 47018 8
15:41
349
an z n . Then
F F(0)2H 2 = 2
an 2
n=1
and
2
F (z) (1 z2 )d xd y =
n=1
n
an 2 .
n+1
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 H p (U ) =
i
p i
F (e )  (e )d.
p
F H p (U )
(1.5)
CY535/Garnett
350
0 521 47018 8
15:41
E
h
cU
E
h ( ).
(, U )
Figure X.2 (, U ) and ( , U ).
CY535/Garnett
0 521 47018 8
15:41
351
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)
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 chordarc 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
15:41
d(z)
(1.10)
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
CY535/Garnett
0 521 47018 8
15:41
353
(AF( ))2 ds =
F (z)2 1 { U : z (, U )} d xd y,
U
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.
F 2  d,
F (z)2  (z)(1 z2 )d xd y,
B=
D
and
D=
D
2
(z)
F 2
(1 z2 )d xd y.
(z)
B8
2
F (z) d(z)d xd y 8B.
2
d
F((z)) (z) (1 z2 )d xd y
A2
dz
D
2
1
(1 z2 )d xd y
(z)
D
=B
D
.
2
CY535/Garnett
0 521 47018 8
354
15:41
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).
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
15:41
355
!1
2
2
1
!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
1
d xd y
z
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
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
15:41
2d(w)
U
U
B(w,d(w))
2
F (w) d(w)dudv,
48
U
B(cU , 2 ).
for z
Therefore
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
z
w
z0
cU
Figure X.5
CY535/Garnett
0 521 47018 8
15:41
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,
CY535/Garnett
358
0 521 47018 8
15:41
so that
( ) (, U ) U
(1.13)
when
=
1+
1
(1.14)
( )
cU = 0
(, U )
CY535/Garnett
0 521 47018 8
15:41
359
> 1. Take > and define by (1.14). Taking K smaller, we can assume
F is bounded on the cone domain
%
U=
( ).
K
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
=
U
F (z)2 K { : z ( )}d xd y
c()
F (z)2 (1 z2 )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
15:41
(1.16)
( )
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 z2 )
2
F (z) (1 z2 )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
15:41
361
Let
U0 =
( ).
E0
F (z)2 (1 z2 )2 { E 0 : z ( )}d xd y
U0
F (z)2 d(z)3 d xd y,
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 z2 )2 d xd y C(, )
F (z) d xd y.
( )
( )
( )
The proof of Lemma 1.5, which is very like the proof of (1.8), is an exercise
for the reader.
CY535/Garnett
0 521 47018 8
362
15:41
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
15:41
363
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)
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
15:41
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
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
15:41
365
(2.9)
A1
M(S)
S1
A0
T
L1
L
S0
(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
15:41
,
(1 + 5 I2m )
Im .
L I ( j) L 0
I =n 1mn
1mn
and by (2.10)
,
Im = 1.
I =n 1mn
Therefore
L I ( j) CeC K
(2.12)
I =n
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
15:41
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)
(2.15)
CY535/Garnett
368
0 521 47018 8
15:41
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),
Q dyadic
The lefthand 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 ) Mx2 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
15:41
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 ,
(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).
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
15:41
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
(2.19)
CY535/Garnett
0 521 47018 8
15:41
By Lemma 2.4,
2f (3Q)(Q) < and
g2 (3Q)(Q) < .
371
(2.20)
But whenever K n Q = ,
(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
ir r
ir r
,
B
,
=
B
2 4
2 4
CY535/Garnett
0 521 47018 8
372
15:41
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)
3. A Decomposition Theorem
To complete the proof of the lefthand inequality in (2.14) we use the following
decomposition theorem, also from Jones [1990]. We call a simply connected
domain an MLipschitz domain if, after a translation and a dilation, 0
and
= {r ( )ei : 0 2 },
where
r (1 ) r (2 ) M1 2 
CY535/Garnett
0 521 47018 8
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
(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 chordarc
curves. The second step is to further decompose each chordarc domain U j into
MLipschitz domains.
(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)
2F 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 MLipschitz domain. Since
H 1 , we also have
1 (U0 ) 1 ().
(3.5)
CY535/Garnett
0 521 47018 8
374
15:41
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 chordarc curve with chordarc 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)
D(Q)
(3.8)
CY535/Garnett
0 521 47018 8
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
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
(Q)
,
2
bDD(Q)
(3.11)
CY535/Garnett
0 521 47018 8
376
15:41
Therefore
1 (U j ) C
type 1
(3.12)
1 (U j ) =
D(Q)
F(z)2 ds C ()
D(Q)
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 chordarc 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)
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
(3.14)
CY535/Garnett
0 521 47018 8
15:41
3. A Decomposition Theorem
377
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
15:41
Write
Dj \
T (Q k ) =
U j,k .
G(Q)
Then by (3.8) each (U j,k ) is an MLipschitz domain, and by (3.15) and (3.8),
1 ((U j,k )) C (z Q )(Q),
k
(3.16)
G(Q)
CY535/Garnett
0 521 47018 8
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)
C2 2 1 (3Q) + 2 2 (3Q) + C1
Area(D j ).
G(Q)
CY535/Garnett
0 521 47018 8
380
15:41
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().
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)
az + b
,
cz + d
(4.2)
CY535/Garnett
0 521 47018 8
15:41
4. Schwarzian Derivatives
381
(z) =
A
,
(a z)2
2
(4.3)
S( )(z) = S((z)) (z) + S(z),
and therefore
S(T )(z) = S(z)
(4.4)
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 z2 S( T )(z) = 1 z2 S(T (z))T (z)
2
= 1 T (z)2 S(T (z)).
Consequently the expression
2
1 z2 S(z)
(4.5)
is a conformal invariant on D.
Theorem 4.1. If is univalent on D, then
2
1 z2 S(z) 6.
(4.6)
2
sup 1 z2 S(z) < 2,
(4.7)
CY535/Garnett
382
0 521 47018 8
15:41
2
S = sup 1 z2 S(z).
D
CY535/Garnett
0 521 47018 8
15:41
4. Schwarzian Derivatives
383
then is univalent. In fact, the interplay between the two conformal invariants
(1 z2 )g (z)
and
2
1 z2 S(z)
2
a hyperbolic ball B, then 1 z2 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 z2 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)
(4.14)
CY535/Garnett
0 521 47018 8
384
15:41
h
,
h
(4.14) becomes
S
h = 0.
2
This differential equation has a unique solution satisfying
h +
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 = ().
CY535/Garnett
0 521 47018 8
15:41
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 ( )
2
1 w0 2 S(w0 ) C
(z , 2k d)2ks .
(5.1)
k=0
CY535/Garnett
386
0 521 47018 8
15:41
Vk = int H0 B(z , 2)
(Hj A j ) (Hk \ Bk ) .
1 jk1
H2 \ B2
H1 A1
L2
z0
L1
z 0
k=0
k 2ks .
(5.2)
CY535/Garnett
0 521 47018 8
15:41
387
N
1
k 2ks + C(s)2N s .
(5.3)
k=0
z 0 w
w1
be
so that on
1 ( )
= {w : w = 1/4},
T 1 ((w))
1
1 + .
w
dist L k B k , L k1 B k C2k (k + k1 ).
(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
15:41
dr
C exp
C2ks .
+ C 1
r
(5.5)
z z
0
G 0 (z) = log
z z0
dist(z, V0 ) C2k
n ,
n=0
and therefore
G 0 (z) C
k
n .
n=0
N
min(k,N
)
k=0
n=0
C(s)
n 2ks
2ns
nk
k 2ks ,
k=0
CY535/Garnett
0 521 47018 8
15:41
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
15:41
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)
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
15:41
391
Therefore, since VN D ,
sup G G N  = sup(G G N )
C2 2N s +
N
1
k 2ks .
k=0
The second estimate of 1 w2 )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).
CY535/Garnett
392
0 521 47018 8
15:41
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
15:41
393
z w
0
is the Mbius transformation from
Define G n = gVn (z, z 0 ). If T (w) = w1
D to the halfplane V0 with T (0) = 0, then
1
G 0 (T (w)) = log .
w
(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.
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
8F 2 + 8F 2 g 2 .
+ 8 36
F 2 (1 z2 )d xd y.
CY535/Garnett
0 521 47018 8
394
15:41
following lemma.
Lemma 6.1. There is a constant C such that if is rectifiable, then
 (z)S(z)2 (1 z2 )3 d xd y C().
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 righthand
inequality of (6.2) holds, even when is not a quasicircle.
To prove the lefthand 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 z2 ) ,
(6.3)
sup S(z)(1 z2 )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 z2 )3 d xd y
(6.5)
CY535/Garnett
0 521 47018 8
15:41
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 z2 )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 z2 )3 d xd y,
2
D
 (0)g (0)2
36 dist((0), ),
4
+ C1
+ C1
D\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 z2 ) 6. Furthermore, if (6.4) fails for Sj , then by
CY535/Garnett
0 521 47018 8
396
15:41
Theorem 4.1 and the Schwarz lemma, there are constants c1 and c2 such that
S(z)(1 z2 ) 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 z2 )3 d xd y
D\B
64 C
2
2
 (z)S(z) (1 z2 )3 d xd y.
Therefore
 (z)ds C1 dist((0), )
D
64
+ C1 1 + 2
+ 64 C1
2
 (z)S(z) (1 z2 )3 d xd y
(Sj ) + C1 2
(6.8)
 (z)ds
2
64
 (z)S(z) (1 z2 )3 d xd y
+ C1 1 + 2
D
4
(Sj ),
+ 6 C1
B,
64 C 1
.
1C2 2
CY535/Garnett
0 521 47018 8
15:41
397
(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)
4
(1 z 0 2 ),
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 ))
(7.1)
CY535/Garnett
0 521 47018 8
398
15:41
H2
C()B.
(7.2)
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 z2 )2 d xd y <
(7.5)
( )
CY535/Garnett
0 521 47018 8
15:41
399
(7.6)
( )
1
F(z) = (z) 2 satisfies F (z)2 =  (z)g (z)2 , Theorem 1.3 and Corollary 7.3 imply has a nonzero 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 = 4F 2 , so that
(7.6) implies
2
F (z) d xd y < ,
( )
2
F (z) (1 z)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
15:41
and
 g 2 (1 z2 )d xd y C,
(7.9)
1
F (z)2 d xd y 2
AF( ) =
( )
1
4
 (z)g (z)2 d xd y
1
2
( )
C 2 p + C 2 ,
t
4
(7.10)
C()
t0
CY535/Garnett
0 521 47018 8
15:41
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 z2 )d xd y C (0)
(7.11)
z< 21
because gB 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 z2 )2 S(z) > .
T (Q)
T (Q)
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
(7.13)
CY535/Garnett
0 521 47018 8
402
15:41
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 gB 6, Theorem 4.1 and the Schwarz lemma yield
 (z)g (z)2 (1 z2 )d xd y
T (Q)
6 2
2
 (z)S(z) (1 z2 )3 d xd y
(7.14)
T (Q)
1
Put D(Q) R and pass to the maximal Q Q. Recall that F = 2
and 4F 2 =  g 2 . To estimate the contribution to (7.9) in Case II we need
the inequality
 (z)g (z)2 (1 z2 )d xd y
D(Q)
=4
(7.15)
D(Q)
+C
F (z)2 (1 z2 )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 2F  = FS + (g )2 
CY535/Garnett
0 521 47018 8
15:41
403
C(Q) (z Q ) + C
2
 (z)S(z) (1 z2 )3 d xd y
(7.16)
D(Q)
+C
D(Q)
2
2 . If
Accept Lemma 7.4 for a moment. Then we can bound the last term in (7.16)
using
 (z)g (z)4 (1 z2 )3 d xd y 2
 (z)g (z)2 (1 z2 )d xd y.
D(Q)
D(Q)
C(Q) (z Q ) + C
2
 (z)S(z) (1 z2 )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 z2 )3 d xd y c 2 (Q).
T (Q)
CY535/Garnett
0 521 47018 8
404
15:41
Because gB 6 it follows that the first term on the righthand side of (7.17)
is majorized by a constant multiple of the second term, and hence that
 (z)g (z)2 (1 z2 )d xd y
D(Q)
2
 (z)S(z) (1 z2 )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 z2 )2 S(z) ,
T (Q)
and if
B(Q) = Q {1 z 2n (Q)} {(1 z2 )g (z) },
then there exists a hyperbolic geodesic such that
inf (z, ) < C,
B(Q)
(7.19)
CY535/Garnett
0 521 47018 8
15:41
405
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 z2 )3 d xd y (Q) (z Q ).
D(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
15:41
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)
Dj
+
D(Q)
 (z)S(z)2 (1z2 )3 d xd y
Dj
(1 z j 2 ) (z j )
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 z2 )d xd y C
(1 w j 2 ) (w j )
D(Q)
C(Q ) (z Q ) CJ.
CY535/Garnett
0 521 47018 8
15:41
407
2
 (z) S(z) (1 z2 )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 z2 )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
D(Q)
CY535/Garnett
0 521 47018 8
408
15:41
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 z2 )3 d xd y
W
J1 ( )
Set
D=
( )
J1
and
V=
%
{T (Q) : T (Q) D}.
CY535/Garnett
0 521 47018 8
15:41
409
(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 z2 )d xd y is a Carleson measure and hence
g (z)4 (1 z2 )3 d xd y is also a Carleson measure, because g B. For any
analytic function we have
g (z)2 (1 z2 )3 d xd y C
g (z)2 (1 z2 )d xd y,
T (Q)
(Q)
T
CY535/Garnett
0 521 47018 8
410
15:41
show
g(ei ) g(0)2 d C.
A=
(8.1)
g (z)2 (1 z2 )d xd y
and
B=
g (z)2 (1 z2 )3 d xd y.
Then since
g (0)
(8.2)
Set U = {z D :
g (z)(1 z2 )
g (z)4 (1 z2 )3 d xd y
A
B
9
+ .
4
12 2
(8.3)
(8.4)
D\(U V )
T (Q j )
9
23
B 2+ 2
S(z)2 (1 z2 )3 d xd y + + 64 C ,
24
4
(8.6)
CY535/Garnett
0 521 47018 8
15:41
9. Angular Derivatives
411
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)
(c)
( )
CY535/Garnett
0 521 47018 8
412
15:41
(d)
2 (( ), t)
(e)
dt
< .
t
S(z)2 (1 z2 )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
( )
Then
{ : z ( )} c (1 z),
2
 (z)S(z) (1 z2 )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 = 4F 2 , where F = 2 , Theorem 1.3 and
Theorem VI.2.3 show has nonzero 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 z2 )2 C 21 ,
(9.4)
CY535/Garnett
0 521 47018 8
15:41
9. Angular Derivatives
on
V =
413
( ) {dist((z), ) 2 4 }.
K
1
2
CY535/Garnett
0 521 47018 8
414
15:41
(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 z2 )3 d xd y < .
(9.6)
U
(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 nonzero 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
For fixed > 0, set U2 = U {z : S(z)(1 z2 )2 }. Then because
gB 6,
4
4
g (z) d(z)3 d xd y
g (z) (1 z2 )3 d xd y
U2
U2
C
2
S(z)2 (1 z2 )3 d xd y.
U2
CY535/Garnett
0 521 47018 8
15:41
415
and we obtain
2
C
B
S (1 z2 )3 d xd y + 64
2+ 2
(Q j ).
2
(10.1)
(10.2)
 (0) c ().
(10.3)
CY535/Garnett
416
0 521 47018 8
15:41
j z+a j
1+a j j z ,
where
1 a j 2
1 + a j j z2
and the fundamental domain F has the form D \ B j where each B j is a disc
orthogonal to D.
 j (z) =
(10.4)
D F 2 .
(10.5)
dist(0, F)
and
( ) F.
(10.6)
(10.7)
,
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
15:41
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
and by Theorem 7.2 there exist N < and K 0 K such that K 0  > K /2
CY535/Garnett
0 521 47018 8
418
15:41
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
U 1 (Q j )
 (z)
d xd y.
(1 z2 )
U 1 (Q j )
 (z)
d xd y C(Q j ),
(1 z2 )
and we obtain
2
 (z)S(z) (1 z2 )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
15:41
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
by Theorem 2.3.
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
15:41
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
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)
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
CY535/Garnett
0 521 47018 8
15:41
421
(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)
1 ()
Figure X.22
CY535/Garnett
0 521 47018 8
422
15:41
(11.4)
Sn
1
8
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
(11.5)
But
inf log
j
1
c4 (1 w j )
z
(11.6)
CY535/Garnett
0 521 47018 8
15:41
423
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
15:41
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)
CY535/Garnett
0 521 47018 8
15:41
425
k )) diamQ j .
%
2j }
{k : z k
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 < .
CY535/Garnett
0 521 47018 8
426
15:41
Cg2  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.
( )
( )
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 ).
CY535/Garnett
0 521 47018 8
15:41
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
15:41
(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
15:41
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)
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
15:41
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)
2 )nk
D(Q)
The construction is simpler in the upper halfplane 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
15:41
431
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)
D(Q)
2
supD 1 z2 S is small, then there is a Mbius transformation T such
that
(T )
1 z2
(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
15:41
and that
F (z)2 d(z)d xd y
and
C F2H 2 (D) F(0)2 + F (0)2 +
1
F2H 2 (D)
C
F (z)2 d(z)3 d xd y
1
F2H 2 (D) ,
C
where C and C depend only on the chordarc 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
CY535/Garnett
0 521 47018 8
15:41
and assume
433
inf w j , D, D \
k > 0,
j
(E.6)
k= 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
(1 )
P0 (M n; Bn j )
= (1 )P0 (M n),
CY535/Garnett
434
0 521 47018 8
15:41
and by induction
2
(1 w j ) 2
(1 )n1 =
.
n=1
CY535/Garnett
0 521 47018 8
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
 f (z) p
(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
436
15:41
Appendices
f (z) = lim f (r z) =
r 1
Pz ( ) f (ei )
d
,
2
 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)
CY535/Garnett
0 521 47018 8
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
438
15:41
Appendices
(A.6)
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)
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
15:41
A. Hardy Spaces
439
and
f 2 = f { f  } .
2
{ 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
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
15:41
Appendices
(A.11)
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
15:41
441
C
.
(B.2)
CY535/Garnett
442
0 521 47018 8
15:41
Appendices
(B.3)
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
CY535/Garnett
0 521 47018 8
15:41
443
Pj (
u ) = 0,
j=1
for j = 1, . . . , m 1.
k, j a j = Pk (
u ),
1 k m 1,
j=1
CY535/Garnett
0 521 47018 8
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
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
m1
j=1
bj j .
CY535/Garnett
0 521 47018 8
15:41
445
Then
j
v
v2 j
=
+
= f1 +
= f,
bj
bj
n
n
n
n
m1
m1
j=1
j=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
15:41
Appendices
CY535/Garnett
0 521 47018 8
15:41
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
m1
Exercise B.3. Prove that the matrix k, j j,k=1 defined by (B.4) is symmetric,
k, j = j,k .
\J
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
448
15:41
Appendices
$
D(u) D(v),
and
D(u v) = D(u) + D(v) 2D(u, v).
(C.1)
(C.2)
f (ei )
an ein ,
(C.3)
+
an r n ein ,
0 r < 1,
and
D(u) = 2
+
nan 2 .
(C.4)
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
15:41
449
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 nN +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
15:41
Appendices
(C.6)
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
15:41
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
W D +
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
CY535/Garnett
0 521 47018 8
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) < = .
(C.8)
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
CY535/Garnett
0 521 47018 8
15:41
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
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
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
15:41
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
on ,
on E
u=1
(C.13)
and
u=0
and
on F.
(C.14)
CY535/Garnett
456
0 521 47018 8
15:41
Appendices
D. Hausdorff Measure
The twodimensional 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
15:41
D. Hausdorff Measure
457
limit
h (E) = lim h (E)
0
(D.3)
(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
(D.4)
and
dim(E) = inf{ : Mr (E) = 0}.
(D.5)
CY535/Garnett
0 521 47018 8
458
15:41
Appendices
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)
(D.7)
CY535/Garnett
0 521 47018 8
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)
CY535/Garnett
0 521 47018 8
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
(E) I
dr + log
,
(E)
(E) 0
r
diam(E)
which yields (D.11).
CY535/Garnett
0 521 47018 8
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
{ z1}
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)
r 0 EB(z,r )
log
1
d E ( ) = 0
 z
CY535/Garnett
0 521 47018 8
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
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 onethird 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 .
CY535/Garnett
0 521 47018 8
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 onethird, 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
(D.14)
1
h (K ) 1.
36
(D.15)
for j = 0, 1, . . . Then
CY535/Garnett
0 521 47018 8
464
15:41
Appendices
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
15:41
D. Hausdorff Measure
465
r
E
E 0
as 0, and
m(z, r )
dr d(z)
r
E 0
2n
dr
I () =
(D.17)
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
m=1
4m log
1
< .
lm
(D.18)
CY535/Garnett
0 521 47018 8
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
1
lm1
4m log
= ,
dr
r
A(r
)
lm
0
m=1
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.
max
z 1 ,...,z n E
Dn (z 1 , . . . , z n ).
CY535/Garnett
0 521 47018 8
15:41
467
(E.2)
Write
log
2
1
1
=
.
log
dn (E)
n(n 1)
z j z i 
(E.3)
i, j:i< j
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
468
15:41
Appendices
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)
1
= lim m n (E),
n
dn (E)
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
15:41
469
1
z
(E.8)
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.
CY535/Garnett
470
0 521 47018 8
15:41
Appendices
CY535/Garnett
0 521 47018 8
15:41
471
N
j Ij ,
j=1
where
1
j =
j j1
f d
Ij
Mn
generates M.
(F.2)
n=1
(F.3)
k > 0.
(F.4)
(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
472
15:41
Appendices
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 weakL1 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
M
.
(F.6)
n1
*
k=1
{x :  f k (x) } Mn .
CY535/Garnett
0 521 47018 8
15:41
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
 f n  = E( f m Mn ) E( f m  Mn ).
(F.7)
So (F.6) is proved.
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
474
15:41
Appendices
 f d P.
(F.8)
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 +
(F.9)
CY535/Garnett
0 521 47018 8
15:41
z0
475
z1
z2
zn
Figure F.1
Because is bounded, the martingale theorem ensures that
z = lim z n
n
(F.10)
4
sinh( )e 2 + O(en )
2
follows from conformally mapping to the upper halfplane.
w(z 0 , E, ) =
CY535/Garnett
0 521 47018 8
476
15:41
Appendices
observation to show
P(z is a regular point of ) = 1.
(F.11)
*
P(B(t j+1 ) B(t j ) Sj ).
P
{B(t j+1 ) B(t j ) Sj }) =
CY535/Garnett
0 521 47018 8
15:41
477
CY535/Garnett
0 521 47018 8
478
15:41
Appendices
f In, j In, j .
Equivalently,
 f BMO(M)
1
= sup
I =In, j I 
 f f I d.
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 
CY535/Garnett
0 521 47018 8
15:41
479
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 < .
u2 log
sup
za
aD
D
CY535/Garnett
0 521 47018 8
480
15:41
Appendices
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)
(g )2 d x
g 2 d x.
(G.2)
ba
a
a
CY535/Garnett
0 521 47018 8
15:41
G. Carlemans Method
481
A (x)
x
(G.3)
2y dy
x
2y dy
By Greens theorem
(x)
2
x
A(x) =
x
x dy,
Thus
A
2 dy.
A2
+
2
x dy
(x)
x
2 dy.
2
x
(G.4)
2 dy
2
A,
(x)
CY535/Garnett
0 521 47018 8
482
15:41
Appendices
By Harnacks inequality
(z 0 )/3
on B(z 0 , r0 /2), so that
9(x0 )
.
r0
2 (z 0 )
(G.5)
.
(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.
(G.6)
CY535/Garnett
0 521 47018 8
15:41
G. Carlemans Method
483
(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
r0
9M
1+
r0
(z 0 )2
9M (x0 )
r0
(b)
t
! 1
b
2
exp
(s)ds dt
M x0
x0
(G.8)
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
484
15:41
Appendices
and thus
(z 0 , E b )
9M 2
2r0 (1 eb )
21
exp
b
x0
dx
.
(x)
(G.10)
Notes
Tsuji [1959] proved (G.9) in polar coordinates. Our presentation follows
Carleman [1933] and Haliste [1965].
ds = 0
j s
and hence
= + i
CY535/Garnett
0 521 47018 8
15:41
485
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
v2 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
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)
CY535/Garnett
0 521 47018 8
15:41
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
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
15:41
489
(H.12)
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
nonzero 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
490
15:41
Appendices
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